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 * @(#)vfs_vnops.c 8.2 (Berkeley) 1/21/94 43 */ 44 45 #include <sys/cdefs.h> 46 __FBSDID("$FreeBSD$"); 47 48 #include "opt_hwpmc_hooks.h" 49 50 #include <sys/param.h> 51 #include <sys/systm.h> 52 #include <sys/disk.h> 53 #include <sys/fail.h> 54 #include <sys/fcntl.h> 55 #include <sys/file.h> 56 #include <sys/kdb.h> 57 #include <sys/ktr.h> 58 #include <sys/stat.h> 59 #include <sys/priv.h> 60 #include <sys/proc.h> 61 #include <sys/limits.h> 62 #include <sys/lock.h> 63 #include <sys/mman.h> 64 #include <sys/mount.h> 65 #include <sys/mutex.h> 66 #include <sys/namei.h> 67 #include <sys/vnode.h> 68 #include <sys/dirent.h> 69 #include <sys/bio.h> 70 #include <sys/buf.h> 71 #include <sys/filio.h> 72 #include <sys/resourcevar.h> 73 #include <sys/rwlock.h> 74 #include <sys/prng.h> 75 #include <sys/sx.h> 76 #include <sys/sleepqueue.h> 77 #include <sys/sysctl.h> 78 #include <sys/ttycom.h> 79 #include <sys/conf.h> 80 #include <sys/syslog.h> 81 #include <sys/unistd.h> 82 #include <sys/user.h> 83 #include <sys/ktrace.h> 84 85 #include <security/audit/audit.h> 86 #include <security/mac/mac_framework.h> 87 88 #include <vm/vm.h> 89 #include <vm/vm_extern.h> 90 #include <vm/pmap.h> 91 #include <vm/vm_map.h> 92 #include <vm/vm_object.h> 93 #include <vm/vm_page.h> 94 #include <vm/vm_pager.h> 95 96 #ifdef HWPMC_HOOKS 97 #include <sys/pmckern.h> 98 #endif 99 100 static fo_rdwr_t vn_read; 101 static fo_rdwr_t vn_write; 102 static fo_rdwr_t vn_io_fault; 103 static fo_truncate_t vn_truncate; 104 static fo_ioctl_t vn_ioctl; 105 static fo_poll_t vn_poll; 106 static fo_kqfilter_t vn_kqfilter; 107 static fo_close_t vn_closefile; 108 static fo_mmap_t vn_mmap; 109 static fo_fallocate_t vn_fallocate; 110 static fo_fspacectl_t vn_fspacectl; 111 112 struct fileops vnops = { 113 .fo_read = vn_io_fault, 114 .fo_write = vn_io_fault, 115 .fo_truncate = vn_truncate, 116 .fo_ioctl = vn_ioctl, 117 .fo_poll = vn_poll, 118 .fo_kqfilter = vn_kqfilter, 119 .fo_stat = vn_statfile, 120 .fo_close = vn_closefile, 121 .fo_chmod = vn_chmod, 122 .fo_chown = vn_chown, 123 .fo_sendfile = vn_sendfile, 124 .fo_seek = vn_seek, 125 .fo_fill_kinfo = vn_fill_kinfo, 126 .fo_mmap = vn_mmap, 127 .fo_fallocate = vn_fallocate, 128 .fo_fspacectl = vn_fspacectl, 129 .fo_flags = DFLAG_PASSABLE | DFLAG_SEEKABLE 130 }; 131 132 const u_int io_hold_cnt = 16; 133 static int vn_io_fault_enable = 1; 134 SYSCTL_INT(_debug, OID_AUTO, vn_io_fault_enable, CTLFLAG_RWTUN, 135 &vn_io_fault_enable, 0, "Enable vn_io_fault lock avoidance"); 136 static int vn_io_fault_prefault = 0; 137 SYSCTL_INT(_debug, OID_AUTO, vn_io_fault_prefault, CTLFLAG_RWTUN, 138 &vn_io_fault_prefault, 0, "Enable vn_io_fault prefaulting"); 139 static int vn_io_pgcache_read_enable = 1; 140 SYSCTL_INT(_debug, OID_AUTO, vn_io_pgcache_read_enable, CTLFLAG_RWTUN, 141 &vn_io_pgcache_read_enable, 0, 142 "Enable copying from page cache for reads, avoiding fs"); 143 static u_long vn_io_faults_cnt; 144 SYSCTL_ULONG(_debug, OID_AUTO, vn_io_faults, CTLFLAG_RD, 145 &vn_io_faults_cnt, 0, "Count of vn_io_fault lock avoidance triggers"); 146 147 static int vfs_allow_read_dir = 0; 148 SYSCTL_INT(_security_bsd, OID_AUTO, allow_read_dir, CTLFLAG_RW, 149 &vfs_allow_read_dir, 0, 150 "Enable read(2) of directory by root for filesystems that support it"); 151 152 /* 153 * Returns true if vn_io_fault mode of handling the i/o request should 154 * be used. 155 */ 156 static bool 157 do_vn_io_fault(struct vnode *vp, struct uio *uio) 158 { 159 struct mount *mp; 160 161 return (uio->uio_segflg == UIO_USERSPACE && vp->v_type == VREG && 162 (mp = vp->v_mount) != NULL && 163 (mp->mnt_kern_flag & MNTK_NO_IOPF) != 0 && vn_io_fault_enable); 164 } 165 166 /* 167 * Structure used to pass arguments to vn_io_fault1(), to do either 168 * file- or vnode-based I/O calls. 169 */ 170 struct vn_io_fault_args { 171 enum { 172 VN_IO_FAULT_FOP, 173 VN_IO_FAULT_VOP 174 } kind; 175 struct ucred *cred; 176 int flags; 177 union { 178 struct fop_args_tag { 179 struct file *fp; 180 fo_rdwr_t *doio; 181 } fop_args; 182 struct vop_args_tag { 183 struct vnode *vp; 184 } vop_args; 185 } args; 186 }; 187 188 static int vn_io_fault1(struct vnode *vp, struct uio *uio, 189 struct vn_io_fault_args *args, struct thread *td); 190 191 int 192 vn_open(struct nameidata *ndp, int *flagp, int cmode, struct file *fp) 193 { 194 struct thread *td = curthread; 195 196 return (vn_open_cred(ndp, flagp, cmode, 0, td->td_ucred, fp)); 197 } 198 199 static uint64_t 200 open2nameif(int fmode, u_int vn_open_flags) 201 { 202 uint64_t res; 203 204 res = ISOPEN | LOCKLEAF; 205 if ((fmode & O_RESOLVE_BENEATH) != 0) 206 res |= RBENEATH; 207 if ((fmode & O_EMPTY_PATH) != 0) 208 res |= EMPTYPATH; 209 if ((fmode & FREAD) != 0) 210 res |= OPENREAD; 211 if ((fmode & FWRITE) != 0) 212 res |= OPENWRITE; 213 if ((vn_open_flags & VN_OPEN_NOAUDIT) == 0) 214 res |= AUDITVNODE1; 215 if ((vn_open_flags & VN_OPEN_NOCAPCHECK) != 0) 216 res |= NOCAPCHECK; 217 if ((vn_open_flags & VN_OPEN_WANTIOCTLCAPS) != 0) 218 res |= WANTIOCTLCAPS; 219 return (res); 220 } 221 222 /* 223 * Common code for vnode open operations via a name lookup. 224 * Lookup the vnode and invoke VOP_CREATE if needed. 225 * Check permissions, and call the VOP_OPEN or VOP_CREATE routine. 226 * 227 * Note that this does NOT free nameidata for the successful case, 228 * due to the NDINIT being done elsewhere. 229 */ 230 int 231 vn_open_cred(struct nameidata *ndp, int *flagp, int cmode, u_int vn_open_flags, 232 struct ucred *cred, struct file *fp) 233 { 234 struct vnode *vp; 235 struct mount *mp; 236 struct vattr vat; 237 struct vattr *vap = &vat; 238 int fmode, error; 239 bool first_open; 240 241 restart: 242 first_open = false; 243 fmode = *flagp; 244 if ((fmode & (O_CREAT | O_EXCL | O_DIRECTORY)) == (O_CREAT | 245 O_EXCL | O_DIRECTORY) || 246 (fmode & (O_CREAT | O_EMPTY_PATH)) == (O_CREAT | O_EMPTY_PATH)) 247 return (EINVAL); 248 else if ((fmode & (O_CREAT | O_DIRECTORY)) == O_CREAT) { 249 ndp->ni_cnd.cn_nameiop = CREATE; 250 ndp->ni_cnd.cn_flags = open2nameif(fmode, vn_open_flags); 251 /* 252 * Set NOCACHE to avoid flushing the cache when 253 * rolling in many files at once. 254 * 255 * Set NC_KEEPPOSENTRY to keep positive entries if they already 256 * exist despite NOCACHE. 257 */ 258 ndp->ni_cnd.cn_flags |= LOCKPARENT | NOCACHE | NC_KEEPPOSENTRY; 259 if ((fmode & O_EXCL) == 0 && (fmode & O_NOFOLLOW) == 0) 260 ndp->ni_cnd.cn_flags |= FOLLOW; 261 if ((vn_open_flags & VN_OPEN_INVFS) == 0) 262 bwillwrite(); 263 if ((error = namei(ndp)) != 0) 264 return (error); 265 if (ndp->ni_vp == NULL) { 266 VATTR_NULL(vap); 267 vap->va_type = VREG; 268 vap->va_mode = cmode; 269 if (fmode & O_EXCL) 270 vap->va_vaflags |= VA_EXCLUSIVE; 271 if (vn_start_write(ndp->ni_dvp, &mp, V_NOWAIT) != 0) { 272 NDFREE_PNBUF(ndp); 273 vput(ndp->ni_dvp); 274 if ((error = vn_start_write(NULL, &mp, 275 V_XSLEEP | V_PCATCH)) != 0) 276 return (error); 277 NDREINIT(ndp); 278 goto restart; 279 } 280 if ((vn_open_flags & VN_OPEN_NAMECACHE) != 0) 281 ndp->ni_cnd.cn_flags |= MAKEENTRY; 282 #ifdef MAC 283 error = mac_vnode_check_create(cred, ndp->ni_dvp, 284 &ndp->ni_cnd, vap); 285 if (error == 0) 286 #endif 287 error = VOP_CREATE(ndp->ni_dvp, &ndp->ni_vp, 288 &ndp->ni_cnd, vap); 289 vp = ndp->ni_vp; 290 if (error == 0 && (fmode & O_EXCL) != 0 && 291 (fmode & (O_EXLOCK | O_SHLOCK)) != 0) { 292 VI_LOCK(vp); 293 vp->v_iflag |= VI_FOPENING; 294 VI_UNLOCK(vp); 295 first_open = true; 296 } 297 VOP_VPUT_PAIR(ndp->ni_dvp, error == 0 ? &vp : NULL, 298 false); 299 vn_finished_write(mp); 300 if (error) { 301 NDFREE_PNBUF(ndp); 302 if (error == ERELOOKUP) { 303 NDREINIT(ndp); 304 goto restart; 305 } 306 return (error); 307 } 308 fmode &= ~O_TRUNC; 309 } else { 310 if (ndp->ni_dvp == ndp->ni_vp) 311 vrele(ndp->ni_dvp); 312 else 313 vput(ndp->ni_dvp); 314 ndp->ni_dvp = NULL; 315 vp = ndp->ni_vp; 316 if (fmode & O_EXCL) { 317 error = EEXIST; 318 goto bad; 319 } 320 if (vp->v_type == VDIR) { 321 error = EISDIR; 322 goto bad; 323 } 324 fmode &= ~O_CREAT; 325 } 326 } else { 327 ndp->ni_cnd.cn_nameiop = LOOKUP; 328 ndp->ni_cnd.cn_flags = open2nameif(fmode, vn_open_flags); 329 ndp->ni_cnd.cn_flags |= (fmode & O_NOFOLLOW) != 0 ? NOFOLLOW : 330 FOLLOW; 331 if ((fmode & FWRITE) == 0) 332 ndp->ni_cnd.cn_flags |= LOCKSHARED; 333 if ((error = namei(ndp)) != 0) 334 return (error); 335 vp = ndp->ni_vp; 336 } 337 error = vn_open_vnode(vp, fmode, cred, curthread, fp); 338 if (first_open) { 339 VI_LOCK(vp); 340 vp->v_iflag &= ~VI_FOPENING; 341 wakeup(vp); 342 VI_UNLOCK(vp); 343 } 344 if (error) 345 goto bad; 346 *flagp = fmode; 347 return (0); 348 bad: 349 NDFREE_PNBUF(ndp); 350 vput(vp); 351 *flagp = fmode; 352 ndp->ni_vp = NULL; 353 return (error); 354 } 355 356 static int 357 vn_open_vnode_advlock(struct vnode *vp, int fmode, struct file *fp) 358 { 359 struct flock lf; 360 int error, lock_flags, type; 361 362 ASSERT_VOP_LOCKED(vp, "vn_open_vnode_advlock"); 363 if ((fmode & (O_EXLOCK | O_SHLOCK)) == 0) 364 return (0); 365 KASSERT(fp != NULL, ("open with flock requires fp")); 366 if (fp->f_type != DTYPE_NONE && fp->f_type != DTYPE_VNODE) 367 return (EOPNOTSUPP); 368 369 lock_flags = VOP_ISLOCKED(vp); 370 VOP_UNLOCK(vp); 371 372 lf.l_whence = SEEK_SET; 373 lf.l_start = 0; 374 lf.l_len = 0; 375 lf.l_type = (fmode & O_EXLOCK) != 0 ? F_WRLCK : F_RDLCK; 376 type = F_FLOCK; 377 if ((fmode & FNONBLOCK) == 0) 378 type |= F_WAIT; 379 if ((fmode & (O_CREAT | O_EXCL)) == (O_CREAT | O_EXCL)) 380 type |= F_FIRSTOPEN; 381 error = VOP_ADVLOCK(vp, (caddr_t)fp, F_SETLK, &lf, type); 382 if (error == 0) 383 fp->f_flag |= FHASLOCK; 384 385 vn_lock(vp, lock_flags | LK_RETRY); 386 return (error); 387 } 388 389 /* 390 * Common code for vnode open operations once a vnode is located. 391 * Check permissions, and call the VOP_OPEN routine. 392 */ 393 int 394 vn_open_vnode(struct vnode *vp, int fmode, struct ucred *cred, 395 struct thread *td, struct file *fp) 396 { 397 accmode_t accmode; 398 int error; 399 400 if (vp->v_type == VLNK) { 401 if ((fmode & O_PATH) == 0 || (fmode & FEXEC) != 0) 402 return (EMLINK); 403 } 404 if (vp->v_type != VDIR && fmode & O_DIRECTORY) 405 return (ENOTDIR); 406 407 accmode = 0; 408 if ((fmode & O_PATH) == 0) { 409 if (vp->v_type == VSOCK) 410 return (EOPNOTSUPP); 411 if ((fmode & (FWRITE | O_TRUNC)) != 0) { 412 if (vp->v_type == VDIR) 413 return (EISDIR); 414 accmode |= VWRITE; 415 } 416 if ((fmode & FREAD) != 0) 417 accmode |= VREAD; 418 if ((fmode & O_APPEND) && (fmode & FWRITE)) 419 accmode |= VAPPEND; 420 #ifdef MAC 421 if ((fmode & O_CREAT) != 0) 422 accmode |= VCREAT; 423 #endif 424 } 425 if ((fmode & FEXEC) != 0) 426 accmode |= VEXEC; 427 #ifdef MAC 428 if ((fmode & O_VERIFY) != 0) 429 accmode |= VVERIFY; 430 error = mac_vnode_check_open(cred, vp, accmode); 431 if (error != 0) 432 return (error); 433 434 accmode &= ~(VCREAT | VVERIFY); 435 #endif 436 if ((fmode & O_CREAT) == 0 && accmode != 0) { 437 error = VOP_ACCESS(vp, accmode, cred, td); 438 if (error != 0) 439 return (error); 440 } 441 if ((fmode & O_PATH) != 0) { 442 if (vp->v_type != VFIFO && vp->v_type != VSOCK && 443 VOP_ACCESS(vp, VREAD, cred, td) == 0) 444 fp->f_flag |= FKQALLOWED; 445 return (0); 446 } 447 448 if (vp->v_type == VFIFO && VOP_ISLOCKED(vp) != LK_EXCLUSIVE) 449 vn_lock(vp, LK_UPGRADE | LK_RETRY); 450 error = VOP_OPEN(vp, fmode, cred, td, fp); 451 if (error != 0) 452 return (error); 453 454 error = vn_open_vnode_advlock(vp, fmode, fp); 455 if (error == 0 && (fmode & FWRITE) != 0) { 456 error = VOP_ADD_WRITECOUNT(vp, 1); 457 if (error == 0) { 458 CTR3(KTR_VFS, "%s: vp %p v_writecount increased to %d", 459 __func__, vp, vp->v_writecount); 460 } 461 } 462 463 /* 464 * Error from advlock or VOP_ADD_WRITECOUNT() still requires 465 * calling VOP_CLOSE() to pair with earlier VOP_OPEN(). 466 */ 467 if (error != 0) { 468 if (fp != NULL) { 469 /* 470 * Arrange the call by having fdrop() to use 471 * vn_closefile(). This is to satisfy 472 * filesystems like devfs or tmpfs, which 473 * override fo_close(). 474 */ 475 fp->f_flag |= FOPENFAILED; 476 fp->f_vnode = vp; 477 if (fp->f_ops == &badfileops) { 478 fp->f_type = DTYPE_VNODE; 479 fp->f_ops = &vnops; 480 } 481 vref(vp); 482 } else { 483 /* 484 * If there is no fp, due to kernel-mode open, 485 * we can call VOP_CLOSE() now. 486 */ 487 if ((vp->v_type == VFIFO || 488 !MNT_EXTENDED_SHARED(vp->v_mount)) && 489 VOP_ISLOCKED(vp) != LK_EXCLUSIVE) 490 vn_lock(vp, LK_UPGRADE | LK_RETRY); 491 (void)VOP_CLOSE(vp, fmode & (FREAD | FWRITE | FEXEC), 492 cred, td); 493 } 494 } 495 496 ASSERT_VOP_LOCKED(vp, "vn_open_vnode"); 497 return (error); 498 499 } 500 501 /* 502 * Check for write permissions on the specified vnode. 503 * Prototype text segments cannot be written. 504 * It is racy. 505 */ 506 int 507 vn_writechk(struct vnode *vp) 508 { 509 510 ASSERT_VOP_LOCKED(vp, "vn_writechk"); 511 /* 512 * If there's shared text associated with 513 * the vnode, try to free it up once. If 514 * we fail, we can't allow writing. 515 */ 516 if (VOP_IS_TEXT(vp)) 517 return (ETXTBSY); 518 519 return (0); 520 } 521 522 /* 523 * Vnode close call 524 */ 525 static int 526 vn_close1(struct vnode *vp, int flags, struct ucred *file_cred, 527 struct thread *td, bool keep_ref) 528 { 529 struct mount *mp; 530 int error, lock_flags; 531 532 lock_flags = vp->v_type != VFIFO && MNT_EXTENDED_SHARED(vp->v_mount) ? 533 LK_SHARED : LK_EXCLUSIVE; 534 535 vn_start_write(vp, &mp, V_WAIT); 536 vn_lock(vp, lock_flags | LK_RETRY); 537 AUDIT_ARG_VNODE1(vp); 538 if ((flags & (FWRITE | FOPENFAILED)) == FWRITE) { 539 VOP_ADD_WRITECOUNT_CHECKED(vp, -1); 540 CTR3(KTR_VFS, "%s: vp %p v_writecount decreased to %d", 541 __func__, vp, vp->v_writecount); 542 } 543 error = VOP_CLOSE(vp, flags, file_cred, td); 544 if (keep_ref) 545 VOP_UNLOCK(vp); 546 else 547 vput(vp); 548 vn_finished_write(mp); 549 return (error); 550 } 551 552 int 553 vn_close(struct vnode *vp, int flags, struct ucred *file_cred, 554 struct thread *td) 555 { 556 557 return (vn_close1(vp, flags, file_cred, td, false)); 558 } 559 560 /* 561 * Heuristic to detect sequential operation. 562 */ 563 static int 564 sequential_heuristic(struct uio *uio, struct file *fp) 565 { 566 enum uio_rw rw; 567 568 ASSERT_VOP_LOCKED(fp->f_vnode, __func__); 569 570 rw = uio->uio_rw; 571 if (fp->f_flag & FRDAHEAD) 572 return (fp->f_seqcount[rw] << IO_SEQSHIFT); 573 574 /* 575 * Offset 0 is handled specially. open() sets f_seqcount to 1 so 576 * that the first I/O is normally considered to be slightly 577 * sequential. Seeking to offset 0 doesn't change sequentiality 578 * unless previous seeks have reduced f_seqcount to 0, in which 579 * case offset 0 is not special. 580 */ 581 if ((uio->uio_offset == 0 && fp->f_seqcount[rw] > 0) || 582 uio->uio_offset == fp->f_nextoff[rw]) { 583 /* 584 * f_seqcount is in units of fixed-size blocks so that it 585 * depends mainly on the amount of sequential I/O and not 586 * much on the number of sequential I/O's. The fixed size 587 * of 16384 is hard-coded here since it is (not quite) just 588 * a magic size that works well here. This size is more 589 * closely related to the best I/O size for real disks than 590 * to any block size used by software. 591 */ 592 if (uio->uio_resid >= IO_SEQMAX * 16384) 593 fp->f_seqcount[rw] = IO_SEQMAX; 594 else { 595 fp->f_seqcount[rw] += howmany(uio->uio_resid, 16384); 596 if (fp->f_seqcount[rw] > IO_SEQMAX) 597 fp->f_seqcount[rw] = IO_SEQMAX; 598 } 599 return (fp->f_seqcount[rw] << IO_SEQSHIFT); 600 } 601 602 /* Not sequential. Quickly draw-down sequentiality. */ 603 if (fp->f_seqcount[rw] > 1) 604 fp->f_seqcount[rw] = 1; 605 else 606 fp->f_seqcount[rw] = 0; 607 return (0); 608 } 609 610 /* 611 * Package up an I/O request on a vnode into a uio and do it. 612 */ 613 int 614 vn_rdwr(enum uio_rw rw, struct vnode *vp, void *base, int len, off_t offset, 615 enum uio_seg segflg, int ioflg, struct ucred *active_cred, 616 struct ucred *file_cred, ssize_t *aresid, struct thread *td) 617 { 618 struct uio auio; 619 struct iovec aiov; 620 struct mount *mp; 621 struct ucred *cred; 622 void *rl_cookie; 623 struct vn_io_fault_args args; 624 int error, lock_flags; 625 626 if (offset < 0 && vp->v_type != VCHR) 627 return (EINVAL); 628 auio.uio_iov = &aiov; 629 auio.uio_iovcnt = 1; 630 aiov.iov_base = base; 631 aiov.iov_len = len; 632 auio.uio_resid = len; 633 auio.uio_offset = offset; 634 auio.uio_segflg = segflg; 635 auio.uio_rw = rw; 636 auio.uio_td = td; 637 error = 0; 638 639 if ((ioflg & IO_NODELOCKED) == 0) { 640 if ((ioflg & IO_RANGELOCKED) == 0) { 641 if (rw == UIO_READ) { 642 rl_cookie = vn_rangelock_rlock(vp, offset, 643 offset + len); 644 } else if ((ioflg & IO_APPEND) != 0) { 645 rl_cookie = vn_rangelock_wlock(vp, 0, OFF_MAX); 646 } else { 647 rl_cookie = vn_rangelock_wlock(vp, offset, 648 offset + len); 649 } 650 } else 651 rl_cookie = NULL; 652 mp = NULL; 653 if (rw == UIO_WRITE) { 654 if (vp->v_type != VCHR && 655 (error = vn_start_write(vp, &mp, V_WAIT | V_PCATCH)) 656 != 0) 657 goto out; 658 lock_flags = vn_lktype_write(mp, vp); 659 } else 660 lock_flags = LK_SHARED; 661 vn_lock(vp, lock_flags | LK_RETRY); 662 } else 663 rl_cookie = NULL; 664 665 ASSERT_VOP_LOCKED(vp, "IO_NODELOCKED with no vp lock held"); 666 #ifdef MAC 667 if ((ioflg & IO_NOMACCHECK) == 0) { 668 if (rw == UIO_READ) 669 error = mac_vnode_check_read(active_cred, file_cred, 670 vp); 671 else 672 error = mac_vnode_check_write(active_cred, file_cred, 673 vp); 674 } 675 #endif 676 if (error == 0) { 677 if (file_cred != NULL) 678 cred = file_cred; 679 else 680 cred = active_cred; 681 if (do_vn_io_fault(vp, &auio)) { 682 args.kind = VN_IO_FAULT_VOP; 683 args.cred = cred; 684 args.flags = ioflg; 685 args.args.vop_args.vp = vp; 686 error = vn_io_fault1(vp, &auio, &args, td); 687 } else if (rw == UIO_READ) { 688 error = VOP_READ(vp, &auio, ioflg, cred); 689 } else /* if (rw == UIO_WRITE) */ { 690 error = VOP_WRITE(vp, &auio, ioflg, cred); 691 } 692 } 693 if (aresid) 694 *aresid = auio.uio_resid; 695 else 696 if (auio.uio_resid && error == 0) 697 error = EIO; 698 if ((ioflg & IO_NODELOCKED) == 0) { 699 VOP_UNLOCK(vp); 700 if (mp != NULL) 701 vn_finished_write(mp); 702 } 703 out: 704 if (rl_cookie != NULL) 705 vn_rangelock_unlock(vp, rl_cookie); 706 return (error); 707 } 708 709 /* 710 * Package up an I/O request on a vnode into a uio and do it. The I/O 711 * request is split up into smaller chunks and we try to avoid saturating 712 * the buffer cache while potentially holding a vnode locked, so we 713 * check bwillwrite() before calling vn_rdwr(). We also call kern_yield() 714 * to give other processes a chance to lock the vnode (either other processes 715 * core'ing the same binary, or unrelated processes scanning the directory). 716 */ 717 int 718 vn_rdwr_inchunks(enum uio_rw rw, struct vnode *vp, void *base, size_t len, 719 off_t offset, enum uio_seg segflg, int ioflg, struct ucred *active_cred, 720 struct ucred *file_cred, size_t *aresid, struct thread *td) 721 { 722 int error = 0; 723 ssize_t iaresid; 724 725 do { 726 int chunk; 727 728 /* 729 * Force `offset' to a multiple of MAXBSIZE except possibly 730 * for the first chunk, so that filesystems only need to 731 * write full blocks except possibly for the first and last 732 * chunks. 733 */ 734 chunk = MAXBSIZE - (uoff_t)offset % MAXBSIZE; 735 736 if (chunk > len) 737 chunk = len; 738 if (rw != UIO_READ && vp->v_type == VREG) 739 bwillwrite(); 740 iaresid = 0; 741 error = vn_rdwr(rw, vp, base, chunk, offset, segflg, 742 ioflg, active_cred, file_cred, &iaresid, td); 743 len -= chunk; /* aresid calc already includes length */ 744 if (error) 745 break; 746 offset += chunk; 747 base = (char *)base + chunk; 748 kern_yield(PRI_USER); 749 } while (len); 750 if (aresid) 751 *aresid = len + iaresid; 752 return (error); 753 } 754 755 #if OFF_MAX <= LONG_MAX 756 off_t 757 foffset_lock(struct file *fp, int flags) 758 { 759 volatile short *flagsp; 760 off_t res; 761 short state; 762 763 KASSERT((flags & FOF_OFFSET) == 0, ("FOF_OFFSET passed")); 764 765 if ((flags & FOF_NOLOCK) != 0) 766 return (atomic_load_long(&fp->f_offset)); 767 768 /* 769 * According to McKusick the vn lock was protecting f_offset here. 770 * It is now protected by the FOFFSET_LOCKED flag. 771 */ 772 flagsp = &fp->f_vnread_flags; 773 if (atomic_cmpset_acq_16(flagsp, 0, FOFFSET_LOCKED)) 774 return (atomic_load_long(&fp->f_offset)); 775 776 sleepq_lock(&fp->f_vnread_flags); 777 state = atomic_load_16(flagsp); 778 for (;;) { 779 if ((state & FOFFSET_LOCKED) == 0) { 780 if (!atomic_fcmpset_acq_16(flagsp, &state, 781 FOFFSET_LOCKED)) 782 continue; 783 break; 784 } 785 if ((state & FOFFSET_LOCK_WAITING) == 0) { 786 if (!atomic_fcmpset_acq_16(flagsp, &state, 787 state | FOFFSET_LOCK_WAITING)) 788 continue; 789 } 790 DROP_GIANT(); 791 sleepq_add(&fp->f_vnread_flags, NULL, "vofflock", 0, 0); 792 sleepq_wait(&fp->f_vnread_flags, PUSER -1); 793 PICKUP_GIANT(); 794 sleepq_lock(&fp->f_vnread_flags); 795 state = atomic_load_16(flagsp); 796 } 797 res = atomic_load_long(&fp->f_offset); 798 sleepq_release(&fp->f_vnread_flags); 799 return (res); 800 } 801 802 void 803 foffset_unlock(struct file *fp, off_t val, int flags) 804 { 805 volatile short *flagsp; 806 short state; 807 808 KASSERT((flags & FOF_OFFSET) == 0, ("FOF_OFFSET passed")); 809 810 if ((flags & FOF_NOUPDATE) == 0) 811 atomic_store_long(&fp->f_offset, val); 812 if ((flags & FOF_NEXTOFF_R) != 0) 813 fp->f_nextoff[UIO_READ] = val; 814 if ((flags & FOF_NEXTOFF_W) != 0) 815 fp->f_nextoff[UIO_WRITE] = val; 816 817 if ((flags & FOF_NOLOCK) != 0) 818 return; 819 820 flagsp = &fp->f_vnread_flags; 821 state = atomic_load_16(flagsp); 822 if ((state & FOFFSET_LOCK_WAITING) == 0 && 823 atomic_cmpset_rel_16(flagsp, state, 0)) 824 return; 825 826 sleepq_lock(&fp->f_vnread_flags); 827 MPASS((fp->f_vnread_flags & FOFFSET_LOCKED) != 0); 828 MPASS((fp->f_vnread_flags & FOFFSET_LOCK_WAITING) != 0); 829 fp->f_vnread_flags = 0; 830 sleepq_broadcast(&fp->f_vnread_flags, SLEEPQ_SLEEP, 0, 0); 831 sleepq_release(&fp->f_vnread_flags); 832 } 833 #else 834 off_t 835 foffset_lock(struct file *fp, int flags) 836 { 837 struct mtx *mtxp; 838 off_t res; 839 840 KASSERT((flags & FOF_OFFSET) == 0, ("FOF_OFFSET passed")); 841 842 mtxp = mtx_pool_find(mtxpool_sleep, fp); 843 mtx_lock(mtxp); 844 if ((flags & FOF_NOLOCK) == 0) { 845 while (fp->f_vnread_flags & FOFFSET_LOCKED) { 846 fp->f_vnread_flags |= FOFFSET_LOCK_WAITING; 847 msleep(&fp->f_vnread_flags, mtxp, PUSER -1, 848 "vofflock", 0); 849 } 850 fp->f_vnread_flags |= FOFFSET_LOCKED; 851 } 852 res = fp->f_offset; 853 mtx_unlock(mtxp); 854 return (res); 855 } 856 857 void 858 foffset_unlock(struct file *fp, off_t val, int flags) 859 { 860 struct mtx *mtxp; 861 862 KASSERT((flags & FOF_OFFSET) == 0, ("FOF_OFFSET passed")); 863 864 mtxp = mtx_pool_find(mtxpool_sleep, fp); 865 mtx_lock(mtxp); 866 if ((flags & FOF_NOUPDATE) == 0) 867 fp->f_offset = val; 868 if ((flags & FOF_NEXTOFF_R) != 0) 869 fp->f_nextoff[UIO_READ] = val; 870 if ((flags & FOF_NEXTOFF_W) != 0) 871 fp->f_nextoff[UIO_WRITE] = val; 872 if ((flags & FOF_NOLOCK) == 0) { 873 KASSERT((fp->f_vnread_flags & FOFFSET_LOCKED) != 0, 874 ("Lost FOFFSET_LOCKED")); 875 if (fp->f_vnread_flags & FOFFSET_LOCK_WAITING) 876 wakeup(&fp->f_vnread_flags); 877 fp->f_vnread_flags = 0; 878 } 879 mtx_unlock(mtxp); 880 } 881 #endif 882 883 void 884 foffset_lock_uio(struct file *fp, struct uio *uio, int flags) 885 { 886 887 if ((flags & FOF_OFFSET) == 0) 888 uio->uio_offset = foffset_lock(fp, flags); 889 } 890 891 void 892 foffset_unlock_uio(struct file *fp, struct uio *uio, int flags) 893 { 894 895 if ((flags & FOF_OFFSET) == 0) 896 foffset_unlock(fp, uio->uio_offset, flags); 897 } 898 899 static int 900 get_advice(struct file *fp, struct uio *uio) 901 { 902 struct mtx *mtxp; 903 int ret; 904 905 ret = POSIX_FADV_NORMAL; 906 if (fp->f_advice == NULL || fp->f_vnode->v_type != VREG) 907 return (ret); 908 909 mtxp = mtx_pool_find(mtxpool_sleep, fp); 910 mtx_lock(mtxp); 911 if (fp->f_advice != NULL && 912 uio->uio_offset >= fp->f_advice->fa_start && 913 uio->uio_offset + uio->uio_resid <= fp->f_advice->fa_end) 914 ret = fp->f_advice->fa_advice; 915 mtx_unlock(mtxp); 916 return (ret); 917 } 918 919 static int 920 get_write_ioflag(struct file *fp) 921 { 922 int ioflag; 923 struct mount *mp; 924 struct vnode *vp; 925 926 ioflag = 0; 927 vp = fp->f_vnode; 928 mp = atomic_load_ptr(&vp->v_mount); 929 930 if ((fp->f_flag & O_DIRECT) != 0) 931 ioflag |= IO_DIRECT; 932 933 if ((fp->f_flag & O_FSYNC) != 0 || 934 (mp != NULL && (mp->mnt_flag & MNT_SYNCHRONOUS) != 0)) 935 ioflag |= IO_SYNC; 936 937 /* 938 * For O_DSYNC we set both IO_SYNC and IO_DATASYNC, so that VOP_WRITE() 939 * or VOP_DEALLOCATE() implementations that don't understand IO_DATASYNC 940 * fall back to full O_SYNC behavior. 941 */ 942 if ((fp->f_flag & O_DSYNC) != 0) 943 ioflag |= IO_SYNC | IO_DATASYNC; 944 945 return (ioflag); 946 } 947 948 int 949 vn_read_from_obj(struct vnode *vp, struct uio *uio) 950 { 951 vm_object_t obj; 952 vm_page_t ma[io_hold_cnt + 2]; 953 off_t off, vsz; 954 ssize_t resid; 955 int error, i, j; 956 957 MPASS(uio->uio_resid <= ptoa(io_hold_cnt + 2)); 958 obj = atomic_load_ptr(&vp->v_object); 959 if (obj == NULL) 960 return (EJUSTRETURN); 961 962 /* 963 * Depends on type stability of vm_objects. 964 */ 965 vm_object_pip_add(obj, 1); 966 if ((obj->flags & OBJ_DEAD) != 0) { 967 /* 968 * Note that object might be already reused from the 969 * vnode, and the OBJ_DEAD flag cleared. This is fine, 970 * we recheck for DOOMED vnode state after all pages 971 * are busied, and retract then. 972 * 973 * But we check for OBJ_DEAD to ensure that we do not 974 * busy pages while vm_object_terminate_pages() 975 * processes the queue. 976 */ 977 error = EJUSTRETURN; 978 goto out_pip; 979 } 980 981 resid = uio->uio_resid; 982 off = uio->uio_offset; 983 for (i = 0; resid > 0; i++) { 984 MPASS(i < io_hold_cnt + 2); 985 ma[i] = vm_page_grab_unlocked(obj, atop(off), 986 VM_ALLOC_NOCREAT | VM_ALLOC_SBUSY | VM_ALLOC_IGN_SBUSY | 987 VM_ALLOC_NOWAIT); 988 if (ma[i] == NULL) 989 break; 990 991 /* 992 * Skip invalid pages. Valid mask can be partial only 993 * at EOF, and we clip later. 994 */ 995 if (vm_page_none_valid(ma[i])) { 996 vm_page_sunbusy(ma[i]); 997 break; 998 } 999 1000 resid -= PAGE_SIZE; 1001 off += PAGE_SIZE; 1002 } 1003 if (i == 0) { 1004 error = EJUSTRETURN; 1005 goto out_pip; 1006 } 1007 1008 /* 1009 * Check VIRF_DOOMED after we busied our pages. Since 1010 * vgonel() terminates the vnode' vm_object, it cannot 1011 * process past pages busied by us. 1012 */ 1013 if (VN_IS_DOOMED(vp)) { 1014 error = EJUSTRETURN; 1015 goto out; 1016 } 1017 1018 resid = PAGE_SIZE - (uio->uio_offset & PAGE_MASK) + ptoa(i - 1); 1019 if (resid > uio->uio_resid) 1020 resid = uio->uio_resid; 1021 1022 /* 1023 * Unlocked read of vnp_size is safe because truncation cannot 1024 * pass busied page. But we load vnp_size into a local 1025 * variable so that possible concurrent extension does not 1026 * break calculation. 1027 */ 1028 #if defined(__powerpc__) && !defined(__powerpc64__) 1029 vsz = obj->un_pager.vnp.vnp_size; 1030 #else 1031 vsz = atomic_load_64(&obj->un_pager.vnp.vnp_size); 1032 #endif 1033 if (uio->uio_offset >= vsz) { 1034 error = EJUSTRETURN; 1035 goto out; 1036 } 1037 if (uio->uio_offset + resid > vsz) 1038 resid = vsz - uio->uio_offset; 1039 1040 error = vn_io_fault_pgmove(ma, uio->uio_offset & PAGE_MASK, resid, uio); 1041 1042 out: 1043 for (j = 0; j < i; j++) { 1044 if (error == 0) 1045 vm_page_reference(ma[j]); 1046 vm_page_sunbusy(ma[j]); 1047 } 1048 out_pip: 1049 vm_object_pip_wakeup(obj); 1050 if (error != 0) 1051 return (error); 1052 return (uio->uio_resid == 0 ? 0 : EJUSTRETURN); 1053 } 1054 1055 /* 1056 * File table vnode read routine. 1057 */ 1058 static int 1059 vn_read(struct file *fp, struct uio *uio, struct ucred *active_cred, int flags, 1060 struct thread *td) 1061 { 1062 struct vnode *vp; 1063 off_t orig_offset; 1064 int error, ioflag; 1065 int advice; 1066 1067 KASSERT(uio->uio_td == td, ("uio_td %p is not td %p", 1068 uio->uio_td, td)); 1069 KASSERT(flags & FOF_OFFSET, ("No FOF_OFFSET")); 1070 vp = fp->f_vnode; 1071 ioflag = 0; 1072 if (fp->f_flag & FNONBLOCK) 1073 ioflag |= IO_NDELAY; 1074 if (fp->f_flag & O_DIRECT) 1075 ioflag |= IO_DIRECT; 1076 1077 /* 1078 * Try to read from page cache. VIRF_DOOMED check is racy but 1079 * allows us to avoid unneeded work outright. 1080 */ 1081 if (vn_io_pgcache_read_enable && !mac_vnode_check_read_enabled() && 1082 (vn_irflag_read(vp) & (VIRF_DOOMED | VIRF_PGREAD)) == VIRF_PGREAD) { 1083 error = VOP_READ_PGCACHE(vp, uio, ioflag, fp->f_cred); 1084 if (error == 0) { 1085 fp->f_nextoff[UIO_READ] = uio->uio_offset; 1086 return (0); 1087 } 1088 if (error != EJUSTRETURN) 1089 return (error); 1090 } 1091 1092 advice = get_advice(fp, uio); 1093 vn_lock(vp, LK_SHARED | LK_RETRY); 1094 1095 switch (advice) { 1096 case POSIX_FADV_NORMAL: 1097 case POSIX_FADV_SEQUENTIAL: 1098 case POSIX_FADV_NOREUSE: 1099 ioflag |= sequential_heuristic(uio, fp); 1100 break; 1101 case POSIX_FADV_RANDOM: 1102 /* Disable read-ahead for random I/O. */ 1103 break; 1104 } 1105 orig_offset = uio->uio_offset; 1106 1107 #ifdef MAC 1108 error = mac_vnode_check_read(active_cred, fp->f_cred, vp); 1109 if (error == 0) 1110 #endif 1111 error = VOP_READ(vp, uio, ioflag, fp->f_cred); 1112 fp->f_nextoff[UIO_READ] = uio->uio_offset; 1113 VOP_UNLOCK(vp); 1114 if (error == 0 && advice == POSIX_FADV_NOREUSE && 1115 orig_offset != uio->uio_offset) 1116 /* 1117 * Use POSIX_FADV_DONTNEED to flush pages and buffers 1118 * for the backing file after a POSIX_FADV_NOREUSE 1119 * read(2). 1120 */ 1121 error = VOP_ADVISE(vp, orig_offset, uio->uio_offset - 1, 1122 POSIX_FADV_DONTNEED); 1123 return (error); 1124 } 1125 1126 /* 1127 * File table vnode write routine. 1128 */ 1129 static int 1130 vn_write(struct file *fp, struct uio *uio, struct ucred *active_cred, int flags, 1131 struct thread *td) 1132 { 1133 struct vnode *vp; 1134 struct mount *mp; 1135 off_t orig_offset; 1136 int error, ioflag; 1137 int advice; 1138 bool need_finished_write; 1139 1140 KASSERT(uio->uio_td == td, ("uio_td %p is not td %p", 1141 uio->uio_td, td)); 1142 KASSERT(flags & FOF_OFFSET, ("No FOF_OFFSET")); 1143 vp = fp->f_vnode; 1144 if (vp->v_type == VREG) 1145 bwillwrite(); 1146 ioflag = IO_UNIT; 1147 if (vp->v_type == VREG && (fp->f_flag & O_APPEND) != 0) 1148 ioflag |= IO_APPEND; 1149 if ((fp->f_flag & FNONBLOCK) != 0) 1150 ioflag |= IO_NDELAY; 1151 ioflag |= get_write_ioflag(fp); 1152 1153 mp = NULL; 1154 need_finished_write = false; 1155 if (vp->v_type != VCHR) { 1156 error = vn_start_write(vp, &mp, V_WAIT | V_PCATCH); 1157 if (error != 0) 1158 goto unlock; 1159 need_finished_write = true; 1160 } 1161 1162 advice = get_advice(fp, uio); 1163 1164 vn_lock(vp, vn_lktype_write(mp, vp) | LK_RETRY); 1165 switch (advice) { 1166 case POSIX_FADV_NORMAL: 1167 case POSIX_FADV_SEQUENTIAL: 1168 case POSIX_FADV_NOREUSE: 1169 ioflag |= sequential_heuristic(uio, fp); 1170 break; 1171 case POSIX_FADV_RANDOM: 1172 /* XXX: Is this correct? */ 1173 break; 1174 } 1175 orig_offset = uio->uio_offset; 1176 1177 #ifdef MAC 1178 error = mac_vnode_check_write(active_cred, fp->f_cred, vp); 1179 if (error == 0) 1180 #endif 1181 error = VOP_WRITE(vp, uio, ioflag, fp->f_cred); 1182 fp->f_nextoff[UIO_WRITE] = uio->uio_offset; 1183 VOP_UNLOCK(vp); 1184 if (need_finished_write) 1185 vn_finished_write(mp); 1186 if (error == 0 && advice == POSIX_FADV_NOREUSE && 1187 orig_offset != uio->uio_offset) 1188 /* 1189 * Use POSIX_FADV_DONTNEED to flush pages and buffers 1190 * for the backing file after a POSIX_FADV_NOREUSE 1191 * write(2). 1192 */ 1193 error = VOP_ADVISE(vp, orig_offset, uio->uio_offset - 1, 1194 POSIX_FADV_DONTNEED); 1195 unlock: 1196 return (error); 1197 } 1198 1199 /* 1200 * The vn_io_fault() is a wrapper around vn_read() and vn_write() to 1201 * prevent the following deadlock: 1202 * 1203 * Assume that the thread A reads from the vnode vp1 into userspace 1204 * buffer buf1 backed by the pages of vnode vp2. If a page in buf1 is 1205 * currently not resident, then system ends up with the call chain 1206 * vn_read() -> VOP_READ(vp1) -> uiomove() -> [Page Fault] -> 1207 * vm_fault(buf1) -> vnode_pager_getpages(vp2) -> VOP_GETPAGES(vp2) 1208 * which establishes lock order vp1->vn_lock, then vp2->vn_lock. 1209 * If, at the same time, thread B reads from vnode vp2 into buffer buf2 1210 * backed by the pages of vnode vp1, and some page in buf2 is not 1211 * resident, we get a reversed order vp2->vn_lock, then vp1->vn_lock. 1212 * 1213 * To prevent the lock order reversal and deadlock, vn_io_fault() does 1214 * not allow page faults to happen during VOP_READ() or VOP_WRITE(). 1215 * Instead, it first tries to do the whole range i/o with pagefaults 1216 * disabled. If all pages in the i/o buffer are resident and mapped, 1217 * VOP will succeed (ignoring the genuine filesystem errors). 1218 * Otherwise, we get back EFAULT, and vn_io_fault() falls back to do 1219 * i/o in chunks, with all pages in the chunk prefaulted and held 1220 * using vm_fault_quick_hold_pages(). 1221 * 1222 * Filesystems using this deadlock avoidance scheme should use the 1223 * array of the held pages from uio, saved in the curthread->td_ma, 1224 * instead of doing uiomove(). A helper function 1225 * vn_io_fault_uiomove() converts uiomove request into 1226 * uiomove_fromphys() over td_ma array. 1227 * 1228 * Since vnode locks do not cover the whole i/o anymore, rangelocks 1229 * make the current i/o request atomic with respect to other i/os and 1230 * truncations. 1231 */ 1232 1233 /* 1234 * Decode vn_io_fault_args and perform the corresponding i/o. 1235 */ 1236 static int 1237 vn_io_fault_doio(struct vn_io_fault_args *args, struct uio *uio, 1238 struct thread *td) 1239 { 1240 int error, save; 1241 1242 error = 0; 1243 save = vm_fault_disable_pagefaults(); 1244 switch (args->kind) { 1245 case VN_IO_FAULT_FOP: 1246 error = (args->args.fop_args.doio)(args->args.fop_args.fp, 1247 uio, args->cred, args->flags, td); 1248 break; 1249 case VN_IO_FAULT_VOP: 1250 if (uio->uio_rw == UIO_READ) { 1251 error = VOP_READ(args->args.vop_args.vp, uio, 1252 args->flags, args->cred); 1253 } else if (uio->uio_rw == UIO_WRITE) { 1254 error = VOP_WRITE(args->args.vop_args.vp, uio, 1255 args->flags, args->cred); 1256 } 1257 break; 1258 default: 1259 panic("vn_io_fault_doio: unknown kind of io %d %d", 1260 args->kind, uio->uio_rw); 1261 } 1262 vm_fault_enable_pagefaults(save); 1263 return (error); 1264 } 1265 1266 static int 1267 vn_io_fault_touch(char *base, const struct uio *uio) 1268 { 1269 int r; 1270 1271 r = fubyte(base); 1272 if (r == -1 || (uio->uio_rw == UIO_READ && subyte(base, r) == -1)) 1273 return (EFAULT); 1274 return (0); 1275 } 1276 1277 static int 1278 vn_io_fault_prefault_user(const struct uio *uio) 1279 { 1280 char *base; 1281 const struct iovec *iov; 1282 size_t len; 1283 ssize_t resid; 1284 int error, i; 1285 1286 KASSERT(uio->uio_segflg == UIO_USERSPACE, 1287 ("vn_io_fault_prefault userspace")); 1288 1289 error = i = 0; 1290 iov = uio->uio_iov; 1291 resid = uio->uio_resid; 1292 base = iov->iov_base; 1293 len = iov->iov_len; 1294 while (resid > 0) { 1295 error = vn_io_fault_touch(base, uio); 1296 if (error != 0) 1297 break; 1298 if (len < PAGE_SIZE) { 1299 if (len != 0) { 1300 error = vn_io_fault_touch(base + len - 1, uio); 1301 if (error != 0) 1302 break; 1303 resid -= len; 1304 } 1305 if (++i >= uio->uio_iovcnt) 1306 break; 1307 iov = uio->uio_iov + i; 1308 base = iov->iov_base; 1309 len = iov->iov_len; 1310 } else { 1311 len -= PAGE_SIZE; 1312 base += PAGE_SIZE; 1313 resid -= PAGE_SIZE; 1314 } 1315 } 1316 return (error); 1317 } 1318 1319 /* 1320 * Common code for vn_io_fault(), agnostic to the kind of i/o request. 1321 * Uses vn_io_fault_doio() to make the call to an actual i/o function. 1322 * Used from vn_rdwr() and vn_io_fault(), which encode the i/o request 1323 * into args and call vn_io_fault1() to handle faults during the user 1324 * mode buffer accesses. 1325 */ 1326 static int 1327 vn_io_fault1(struct vnode *vp, struct uio *uio, struct vn_io_fault_args *args, 1328 struct thread *td) 1329 { 1330 vm_page_t ma[io_hold_cnt + 2]; 1331 struct uio *uio_clone, short_uio; 1332 struct iovec short_iovec[1]; 1333 vm_page_t *prev_td_ma; 1334 vm_prot_t prot; 1335 vm_offset_t addr, end; 1336 size_t len, resid; 1337 ssize_t adv; 1338 int error, cnt, saveheld, prev_td_ma_cnt; 1339 1340 if (vn_io_fault_prefault) { 1341 error = vn_io_fault_prefault_user(uio); 1342 if (error != 0) 1343 return (error); /* Or ignore ? */ 1344 } 1345 1346 prot = uio->uio_rw == UIO_READ ? VM_PROT_WRITE : VM_PROT_READ; 1347 1348 /* 1349 * The UFS follows IO_UNIT directive and replays back both 1350 * uio_offset and uio_resid if an error is encountered during the 1351 * operation. But, since the iovec may be already advanced, 1352 * uio is still in an inconsistent state. 1353 * 1354 * Cache a copy of the original uio, which is advanced to the redo 1355 * point using UIO_NOCOPY below. 1356 */ 1357 uio_clone = cloneuio(uio); 1358 resid = uio->uio_resid; 1359 1360 short_uio.uio_segflg = UIO_USERSPACE; 1361 short_uio.uio_rw = uio->uio_rw; 1362 short_uio.uio_td = uio->uio_td; 1363 1364 error = vn_io_fault_doio(args, uio, td); 1365 if (error != EFAULT) 1366 goto out; 1367 1368 atomic_add_long(&vn_io_faults_cnt, 1); 1369 uio_clone->uio_segflg = UIO_NOCOPY; 1370 uiomove(NULL, resid - uio->uio_resid, uio_clone); 1371 uio_clone->uio_segflg = uio->uio_segflg; 1372 1373 saveheld = curthread_pflags_set(TDP_UIOHELD); 1374 prev_td_ma = td->td_ma; 1375 prev_td_ma_cnt = td->td_ma_cnt; 1376 1377 while (uio_clone->uio_resid != 0) { 1378 len = uio_clone->uio_iov->iov_len; 1379 if (len == 0) { 1380 KASSERT(uio_clone->uio_iovcnt >= 1, 1381 ("iovcnt underflow")); 1382 uio_clone->uio_iov++; 1383 uio_clone->uio_iovcnt--; 1384 continue; 1385 } 1386 if (len > ptoa(io_hold_cnt)) 1387 len = ptoa(io_hold_cnt); 1388 addr = (uintptr_t)uio_clone->uio_iov->iov_base; 1389 end = round_page(addr + len); 1390 if (end < addr) { 1391 error = EFAULT; 1392 break; 1393 } 1394 /* 1395 * A perfectly misaligned address and length could cause 1396 * both the start and the end of the chunk to use partial 1397 * page. +2 accounts for such a situation. 1398 */ 1399 cnt = vm_fault_quick_hold_pages(&td->td_proc->p_vmspace->vm_map, 1400 addr, len, prot, ma, io_hold_cnt + 2); 1401 if (cnt == -1) { 1402 error = EFAULT; 1403 break; 1404 } 1405 short_uio.uio_iov = &short_iovec[0]; 1406 short_iovec[0].iov_base = (void *)addr; 1407 short_uio.uio_iovcnt = 1; 1408 short_uio.uio_resid = short_iovec[0].iov_len = len; 1409 short_uio.uio_offset = uio_clone->uio_offset; 1410 td->td_ma = ma; 1411 td->td_ma_cnt = cnt; 1412 1413 error = vn_io_fault_doio(args, &short_uio, td); 1414 vm_page_unhold_pages(ma, cnt); 1415 adv = len - short_uio.uio_resid; 1416 1417 uio_clone->uio_iov->iov_base = 1418 (char *)uio_clone->uio_iov->iov_base + adv; 1419 uio_clone->uio_iov->iov_len -= adv; 1420 uio_clone->uio_resid -= adv; 1421 uio_clone->uio_offset += adv; 1422 1423 uio->uio_resid -= adv; 1424 uio->uio_offset += adv; 1425 1426 if (error != 0 || adv == 0) 1427 break; 1428 } 1429 td->td_ma = prev_td_ma; 1430 td->td_ma_cnt = prev_td_ma_cnt; 1431 curthread_pflags_restore(saveheld); 1432 out: 1433 free(uio_clone, M_IOV); 1434 return (error); 1435 } 1436 1437 static int 1438 vn_io_fault(struct file *fp, struct uio *uio, struct ucred *active_cred, 1439 int flags, struct thread *td) 1440 { 1441 fo_rdwr_t *doio; 1442 struct vnode *vp; 1443 void *rl_cookie; 1444 struct vn_io_fault_args args; 1445 int error; 1446 bool rl_locked; 1447 1448 doio = uio->uio_rw == UIO_READ ? vn_read : vn_write; 1449 vp = fp->f_vnode; 1450 1451 /* 1452 * The ability to read(2) on a directory has historically been 1453 * allowed for all users, but this can and has been the source of 1454 * at least one security issue in the past. As such, it is now hidden 1455 * away behind a sysctl for those that actually need it to use it, and 1456 * restricted to root when it's turned on to make it relatively safe to 1457 * leave on for longer sessions of need. 1458 */ 1459 if (vp->v_type == VDIR) { 1460 KASSERT(uio->uio_rw == UIO_READ, 1461 ("illegal write attempted on a directory")); 1462 if (!vfs_allow_read_dir) 1463 return (EISDIR); 1464 if ((error = priv_check(td, PRIV_VFS_READ_DIR)) != 0) 1465 return (EISDIR); 1466 } 1467 1468 foffset_lock_uio(fp, uio, flags); 1469 if (vp->v_type == VREG) { 1470 if (uio->uio_rw == UIO_READ) { 1471 rl_cookie = vn_rangelock_rlock(vp, uio->uio_offset, 1472 uio->uio_offset + uio->uio_resid); 1473 } else if ((fp->f_flag & O_APPEND) != 0 || 1474 (flags & FOF_OFFSET) == 0) { 1475 /* For appenders, punt and lock the whole range. */ 1476 rl_cookie = vn_rangelock_wlock(vp, 0, OFF_MAX); 1477 } else { 1478 rl_cookie = vn_rangelock_wlock(vp, uio->uio_offset, 1479 uio->uio_offset + uio->uio_resid); 1480 } 1481 rl_locked = true; 1482 } else { 1483 rl_locked = false; 1484 } 1485 if (do_vn_io_fault(vp, uio)) { 1486 args.kind = VN_IO_FAULT_FOP; 1487 args.args.fop_args.fp = fp; 1488 args.args.fop_args.doio = doio; 1489 args.cred = active_cred; 1490 args.flags = flags | FOF_OFFSET; 1491 error = vn_io_fault1(vp, uio, &args, td); 1492 } else { 1493 error = doio(fp, uio, active_cred, flags | FOF_OFFSET, td); 1494 } 1495 if (rl_locked) 1496 vn_rangelock_unlock(vp, rl_cookie); 1497 foffset_unlock_uio(fp, uio, flags); 1498 return (error); 1499 } 1500 1501 /* 1502 * Helper function to perform the requested uiomove operation using 1503 * the held pages for io->uio_iov[0].iov_base buffer instead of 1504 * copyin/copyout. Access to the pages with uiomove_fromphys() 1505 * instead of iov_base prevents page faults that could occur due to 1506 * pmap_collect() invalidating the mapping created by 1507 * vm_fault_quick_hold_pages(), or pageout daemon, page laundry or 1508 * object cleanup revoking the write access from page mappings. 1509 * 1510 * Filesystems specified MNTK_NO_IOPF shall use vn_io_fault_uiomove() 1511 * instead of plain uiomove(). 1512 */ 1513 int 1514 vn_io_fault_uiomove(char *data, int xfersize, struct uio *uio) 1515 { 1516 struct uio transp_uio; 1517 struct iovec transp_iov[1]; 1518 struct thread *td; 1519 size_t adv; 1520 int error, pgadv; 1521 1522 td = curthread; 1523 if ((td->td_pflags & TDP_UIOHELD) == 0 || 1524 uio->uio_segflg != UIO_USERSPACE) 1525 return (uiomove(data, xfersize, uio)); 1526 1527 KASSERT(uio->uio_iovcnt == 1, ("uio_iovcnt %d", uio->uio_iovcnt)); 1528 transp_iov[0].iov_base = data; 1529 transp_uio.uio_iov = &transp_iov[0]; 1530 transp_uio.uio_iovcnt = 1; 1531 if (xfersize > uio->uio_resid) 1532 xfersize = uio->uio_resid; 1533 transp_uio.uio_resid = transp_iov[0].iov_len = xfersize; 1534 transp_uio.uio_offset = 0; 1535 transp_uio.uio_segflg = UIO_SYSSPACE; 1536 /* 1537 * Since transp_iov points to data, and td_ma page array 1538 * corresponds to original uio->uio_iov, we need to invert the 1539 * direction of the i/o operation as passed to 1540 * uiomove_fromphys(). 1541 */ 1542 switch (uio->uio_rw) { 1543 case UIO_WRITE: 1544 transp_uio.uio_rw = UIO_READ; 1545 break; 1546 case UIO_READ: 1547 transp_uio.uio_rw = UIO_WRITE; 1548 break; 1549 } 1550 transp_uio.uio_td = uio->uio_td; 1551 error = uiomove_fromphys(td->td_ma, 1552 ((vm_offset_t)uio->uio_iov->iov_base) & PAGE_MASK, 1553 xfersize, &transp_uio); 1554 adv = xfersize - transp_uio.uio_resid; 1555 pgadv = 1556 (((vm_offset_t)uio->uio_iov->iov_base + adv) >> PAGE_SHIFT) - 1557 (((vm_offset_t)uio->uio_iov->iov_base) >> PAGE_SHIFT); 1558 td->td_ma += pgadv; 1559 KASSERT(td->td_ma_cnt >= pgadv, ("consumed pages %d %d", td->td_ma_cnt, 1560 pgadv)); 1561 td->td_ma_cnt -= pgadv; 1562 uio->uio_iov->iov_base = (char *)uio->uio_iov->iov_base + adv; 1563 uio->uio_iov->iov_len -= adv; 1564 uio->uio_resid -= adv; 1565 uio->uio_offset += adv; 1566 return (error); 1567 } 1568 1569 int 1570 vn_io_fault_pgmove(vm_page_t ma[], vm_offset_t offset, int xfersize, 1571 struct uio *uio) 1572 { 1573 struct thread *td; 1574 vm_offset_t iov_base; 1575 int cnt, pgadv; 1576 1577 td = curthread; 1578 if ((td->td_pflags & TDP_UIOHELD) == 0 || 1579 uio->uio_segflg != UIO_USERSPACE) 1580 return (uiomove_fromphys(ma, offset, xfersize, uio)); 1581 1582 KASSERT(uio->uio_iovcnt == 1, ("uio_iovcnt %d", uio->uio_iovcnt)); 1583 cnt = xfersize > uio->uio_resid ? uio->uio_resid : xfersize; 1584 iov_base = (vm_offset_t)uio->uio_iov->iov_base; 1585 switch (uio->uio_rw) { 1586 case UIO_WRITE: 1587 pmap_copy_pages(td->td_ma, iov_base & PAGE_MASK, ma, 1588 offset, cnt); 1589 break; 1590 case UIO_READ: 1591 pmap_copy_pages(ma, offset, td->td_ma, iov_base & PAGE_MASK, 1592 cnt); 1593 break; 1594 } 1595 pgadv = ((iov_base + cnt) >> PAGE_SHIFT) - (iov_base >> PAGE_SHIFT); 1596 td->td_ma += pgadv; 1597 KASSERT(td->td_ma_cnt >= pgadv, ("consumed pages %d %d", td->td_ma_cnt, 1598 pgadv)); 1599 td->td_ma_cnt -= pgadv; 1600 uio->uio_iov->iov_base = (char *)(iov_base + cnt); 1601 uio->uio_iov->iov_len -= cnt; 1602 uio->uio_resid -= cnt; 1603 uio->uio_offset += cnt; 1604 return (0); 1605 } 1606 1607 /* 1608 * File table truncate routine. 1609 */ 1610 static int 1611 vn_truncate(struct file *fp, off_t length, struct ucred *active_cred, 1612 struct thread *td) 1613 { 1614 struct mount *mp; 1615 struct vnode *vp; 1616 void *rl_cookie; 1617 int error; 1618 1619 vp = fp->f_vnode; 1620 1621 retry: 1622 /* 1623 * Lock the whole range for truncation. Otherwise split i/o 1624 * might happen partly before and partly after the truncation. 1625 */ 1626 rl_cookie = vn_rangelock_wlock(vp, 0, OFF_MAX); 1627 error = vn_start_write(vp, &mp, V_WAIT | V_PCATCH); 1628 if (error) 1629 goto out1; 1630 vn_lock(vp, LK_EXCLUSIVE | LK_RETRY); 1631 AUDIT_ARG_VNODE1(vp); 1632 if (vp->v_type == VDIR) { 1633 error = EISDIR; 1634 goto out; 1635 } 1636 #ifdef MAC 1637 error = mac_vnode_check_write(active_cred, fp->f_cred, vp); 1638 if (error) 1639 goto out; 1640 #endif 1641 error = vn_truncate_locked(vp, length, (fp->f_flag & O_FSYNC) != 0, 1642 fp->f_cred); 1643 out: 1644 VOP_UNLOCK(vp); 1645 vn_finished_write(mp); 1646 out1: 1647 vn_rangelock_unlock(vp, rl_cookie); 1648 if (error == ERELOOKUP) 1649 goto retry; 1650 return (error); 1651 } 1652 1653 /* 1654 * Truncate a file that is already locked. 1655 */ 1656 int 1657 vn_truncate_locked(struct vnode *vp, off_t length, bool sync, 1658 struct ucred *cred) 1659 { 1660 struct vattr vattr; 1661 int error; 1662 1663 error = VOP_ADD_WRITECOUNT(vp, 1); 1664 if (error == 0) { 1665 VATTR_NULL(&vattr); 1666 vattr.va_size = length; 1667 if (sync) 1668 vattr.va_vaflags |= VA_SYNC; 1669 error = VOP_SETATTR(vp, &vattr, cred); 1670 VOP_ADD_WRITECOUNT_CHECKED(vp, -1); 1671 } 1672 return (error); 1673 } 1674 1675 /* 1676 * File table vnode stat routine. 1677 */ 1678 int 1679 vn_statfile(struct file *fp, struct stat *sb, struct ucred *active_cred) 1680 { 1681 struct vnode *vp = fp->f_vnode; 1682 int error; 1683 1684 vn_lock(vp, LK_SHARED | LK_RETRY); 1685 error = VOP_STAT(vp, sb, active_cred, fp->f_cred); 1686 VOP_UNLOCK(vp); 1687 1688 return (error); 1689 } 1690 1691 /* 1692 * File table vnode ioctl routine. 1693 */ 1694 static int 1695 vn_ioctl(struct file *fp, u_long com, void *data, struct ucred *active_cred, 1696 struct thread *td) 1697 { 1698 struct vnode *vp; 1699 struct fiobmap2_arg *bmarg; 1700 off_t size; 1701 int error; 1702 1703 vp = fp->f_vnode; 1704 switch (vp->v_type) { 1705 case VDIR: 1706 case VREG: 1707 switch (com) { 1708 case FIONREAD: 1709 error = vn_getsize(vp, &size, active_cred); 1710 if (error == 0) 1711 *(int *)data = size - fp->f_offset; 1712 return (error); 1713 case FIOBMAP2: 1714 bmarg = (struct fiobmap2_arg *)data; 1715 vn_lock(vp, LK_SHARED | LK_RETRY); 1716 #ifdef MAC 1717 error = mac_vnode_check_read(active_cred, fp->f_cred, 1718 vp); 1719 if (error == 0) 1720 #endif 1721 error = VOP_BMAP(vp, bmarg->bn, NULL, 1722 &bmarg->bn, &bmarg->runp, &bmarg->runb); 1723 VOP_UNLOCK(vp); 1724 return (error); 1725 case FIONBIO: 1726 case FIOASYNC: 1727 return (0); 1728 default: 1729 return (VOP_IOCTL(vp, com, data, fp->f_flag, 1730 active_cred, td)); 1731 } 1732 break; 1733 case VCHR: 1734 return (VOP_IOCTL(vp, com, data, fp->f_flag, 1735 active_cred, td)); 1736 default: 1737 return (ENOTTY); 1738 } 1739 } 1740 1741 /* 1742 * File table vnode poll routine. 1743 */ 1744 static int 1745 vn_poll(struct file *fp, int events, struct ucred *active_cred, 1746 struct thread *td) 1747 { 1748 struct vnode *vp; 1749 int error; 1750 1751 vp = fp->f_vnode; 1752 #if defined(MAC) || defined(AUDIT) 1753 if (AUDITING_TD(td) || mac_vnode_check_poll_enabled()) { 1754 vn_lock(vp, LK_EXCLUSIVE | LK_RETRY); 1755 AUDIT_ARG_VNODE1(vp); 1756 error = mac_vnode_check_poll(active_cred, fp->f_cred, vp); 1757 VOP_UNLOCK(vp); 1758 if (error != 0) 1759 return (error); 1760 } 1761 #endif 1762 error = VOP_POLL(vp, events, fp->f_cred, td); 1763 return (error); 1764 } 1765 1766 /* 1767 * Acquire the requested lock and then check for validity. LK_RETRY 1768 * permits vn_lock to return doomed vnodes. 1769 */ 1770 static int __noinline 1771 _vn_lock_fallback(struct vnode *vp, int flags, const char *file, int line, 1772 int error) 1773 { 1774 1775 KASSERT((flags & LK_RETRY) == 0 || error == 0, 1776 ("vn_lock: error %d incompatible with flags %#x", error, flags)); 1777 1778 if (error == 0) 1779 VNASSERT(VN_IS_DOOMED(vp), vp, ("vnode not doomed")); 1780 1781 if ((flags & LK_RETRY) == 0) { 1782 if (error == 0) { 1783 VOP_UNLOCK(vp); 1784 error = ENOENT; 1785 } 1786 return (error); 1787 } 1788 1789 /* 1790 * LK_RETRY case. 1791 * 1792 * Nothing to do if we got the lock. 1793 */ 1794 if (error == 0) 1795 return (0); 1796 1797 /* 1798 * Interlock was dropped by the call in _vn_lock. 1799 */ 1800 flags &= ~LK_INTERLOCK; 1801 do { 1802 error = VOP_LOCK1(vp, flags, file, line); 1803 } while (error != 0); 1804 return (0); 1805 } 1806 1807 int 1808 _vn_lock(struct vnode *vp, int flags, const char *file, int line) 1809 { 1810 int error; 1811 1812 VNASSERT((flags & LK_TYPE_MASK) != 0, vp, 1813 ("vn_lock: no locktype (%d passed)", flags)); 1814 VNPASS(vp->v_holdcnt > 0, vp); 1815 error = VOP_LOCK1(vp, flags, file, line); 1816 if (__predict_false(error != 0 || VN_IS_DOOMED(vp))) 1817 return (_vn_lock_fallback(vp, flags, file, line, error)); 1818 return (0); 1819 } 1820 1821 /* 1822 * File table vnode close routine. 1823 */ 1824 static int 1825 vn_closefile(struct file *fp, struct thread *td) 1826 { 1827 struct vnode *vp; 1828 struct flock lf; 1829 int error; 1830 bool ref; 1831 1832 vp = fp->f_vnode; 1833 fp->f_ops = &badfileops; 1834 ref = (fp->f_flag & FHASLOCK) != 0; 1835 1836 error = vn_close1(vp, fp->f_flag, fp->f_cred, td, ref); 1837 1838 if (__predict_false(ref)) { 1839 lf.l_whence = SEEK_SET; 1840 lf.l_start = 0; 1841 lf.l_len = 0; 1842 lf.l_type = F_UNLCK; 1843 (void) VOP_ADVLOCK(vp, fp, F_UNLCK, &lf, F_FLOCK); 1844 vrele(vp); 1845 } 1846 return (error); 1847 } 1848 1849 /* 1850 * Preparing to start a filesystem write operation. If the operation is 1851 * permitted, then we bump the count of operations in progress and 1852 * proceed. If a suspend request is in progress, we wait until the 1853 * suspension is over, and then proceed. 1854 */ 1855 static int 1856 vn_start_write_refed(struct mount *mp, int flags, bool mplocked) 1857 { 1858 struct mount_pcpu *mpcpu; 1859 int error, mflags; 1860 1861 if (__predict_true(!mplocked) && (flags & V_XSLEEP) == 0 && 1862 vfs_op_thread_enter(mp, mpcpu)) { 1863 MPASS((mp->mnt_kern_flag & MNTK_SUSPEND) == 0); 1864 vfs_mp_count_add_pcpu(mpcpu, writeopcount, 1); 1865 vfs_op_thread_exit(mp, mpcpu); 1866 return (0); 1867 } 1868 1869 if (mplocked) 1870 mtx_assert(MNT_MTX(mp), MA_OWNED); 1871 else 1872 MNT_ILOCK(mp); 1873 1874 error = 0; 1875 1876 /* 1877 * Check on status of suspension. 1878 */ 1879 if ((curthread->td_pflags & TDP_IGNSUSP) == 0 || 1880 mp->mnt_susp_owner != curthread) { 1881 mflags = 0; 1882 if ((mp->mnt_vfc->vfc_flags & VFCF_SBDRY) != 0) { 1883 if (flags & V_PCATCH) 1884 mflags |= PCATCH; 1885 } 1886 mflags |= (PUSER - 1); 1887 while ((mp->mnt_kern_flag & MNTK_SUSPEND) != 0) { 1888 if ((flags & V_NOWAIT) != 0) { 1889 error = EWOULDBLOCK; 1890 goto unlock; 1891 } 1892 error = msleep(&mp->mnt_flag, MNT_MTX(mp), mflags, 1893 "suspfs", 0); 1894 if (error != 0) 1895 goto unlock; 1896 } 1897 } 1898 if ((flags & V_XSLEEP) != 0) 1899 goto unlock; 1900 mp->mnt_writeopcount++; 1901 unlock: 1902 if (error != 0 || (flags & V_XSLEEP) != 0) 1903 MNT_REL(mp); 1904 MNT_IUNLOCK(mp); 1905 return (error); 1906 } 1907 1908 int 1909 vn_start_write(struct vnode *vp, struct mount **mpp, int flags) 1910 { 1911 struct mount *mp; 1912 int error; 1913 1914 KASSERT((flags & ~V_VALID_FLAGS) == 0, 1915 ("%s: invalid flags passed %d\n", __func__, flags)); 1916 1917 error = 0; 1918 /* 1919 * If a vnode is provided, get and return the mount point that 1920 * to which it will write. 1921 */ 1922 if (vp != NULL) { 1923 if ((error = VOP_GETWRITEMOUNT(vp, mpp)) != 0) { 1924 *mpp = NULL; 1925 if (error != EOPNOTSUPP) 1926 return (error); 1927 return (0); 1928 } 1929 } 1930 if ((mp = *mpp) == NULL) 1931 return (0); 1932 1933 /* 1934 * VOP_GETWRITEMOUNT() returns with the mp refcount held through 1935 * a vfs_ref(). 1936 * As long as a vnode is not provided we need to acquire a 1937 * refcount for the provided mountpoint too, in order to 1938 * emulate a vfs_ref(). 1939 */ 1940 if (vp == NULL) 1941 vfs_ref(mp); 1942 1943 error = vn_start_write_refed(mp, flags, false); 1944 if (error != 0 && (flags & V_NOWAIT) == 0) 1945 *mpp = NULL; 1946 return (error); 1947 } 1948 1949 /* 1950 * Secondary suspension. Used by operations such as vop_inactive 1951 * routines that are needed by the higher level functions. These 1952 * are allowed to proceed until all the higher level functions have 1953 * completed (indicated by mnt_writeopcount dropping to zero). At that 1954 * time, these operations are halted until the suspension is over. 1955 */ 1956 int 1957 vn_start_secondary_write(struct vnode *vp, struct mount **mpp, int flags) 1958 { 1959 struct mount *mp; 1960 int error, mflags; 1961 1962 KASSERT((flags & (~V_VALID_FLAGS | V_XSLEEP)) == 0, 1963 ("%s: invalid flags passed %d\n", __func__, flags)); 1964 1965 retry: 1966 if (vp != NULL) { 1967 if ((error = VOP_GETWRITEMOUNT(vp, mpp)) != 0) { 1968 *mpp = NULL; 1969 if (error != EOPNOTSUPP) 1970 return (error); 1971 return (0); 1972 } 1973 } 1974 /* 1975 * If we are not suspended or have not yet reached suspended 1976 * mode, then let the operation proceed. 1977 */ 1978 if ((mp = *mpp) == NULL) 1979 return (0); 1980 1981 /* 1982 * VOP_GETWRITEMOUNT() returns with the mp refcount held through 1983 * a vfs_ref(). 1984 * As long as a vnode is not provided we need to acquire a 1985 * refcount for the provided mountpoint too, in order to 1986 * emulate a vfs_ref(). 1987 */ 1988 MNT_ILOCK(mp); 1989 if (vp == NULL) 1990 MNT_REF(mp); 1991 if ((mp->mnt_kern_flag & (MNTK_SUSPENDED | MNTK_SUSPEND2)) == 0) { 1992 mp->mnt_secondary_writes++; 1993 mp->mnt_secondary_accwrites++; 1994 MNT_IUNLOCK(mp); 1995 return (0); 1996 } 1997 if ((flags & V_NOWAIT) != 0) { 1998 MNT_REL(mp); 1999 MNT_IUNLOCK(mp); 2000 *mpp = NULL; 2001 return (EWOULDBLOCK); 2002 } 2003 /* 2004 * Wait for the suspension to finish. 2005 */ 2006 mflags = 0; 2007 if ((mp->mnt_vfc->vfc_flags & VFCF_SBDRY) != 0) { 2008 if ((flags & V_PCATCH) != 0) 2009 mflags |= PCATCH; 2010 } 2011 mflags |= (PUSER - 1) | PDROP; 2012 error = msleep(&mp->mnt_flag, MNT_MTX(mp), mflags, "suspfs", 0); 2013 vfs_rel(mp); 2014 if (error == 0) 2015 goto retry; 2016 *mpp = NULL; 2017 return (error); 2018 } 2019 2020 /* 2021 * Filesystem write operation has completed. If we are suspending and this 2022 * operation is the last one, notify the suspender that the suspension is 2023 * now in effect. 2024 */ 2025 void 2026 vn_finished_write(struct mount *mp) 2027 { 2028 struct mount_pcpu *mpcpu; 2029 int c; 2030 2031 if (mp == NULL) 2032 return; 2033 2034 if (vfs_op_thread_enter(mp, mpcpu)) { 2035 vfs_mp_count_sub_pcpu(mpcpu, writeopcount, 1); 2036 vfs_mp_count_sub_pcpu(mpcpu, ref, 1); 2037 vfs_op_thread_exit(mp, mpcpu); 2038 return; 2039 } 2040 2041 MNT_ILOCK(mp); 2042 vfs_assert_mount_counters(mp); 2043 MNT_REL(mp); 2044 c = --mp->mnt_writeopcount; 2045 if (mp->mnt_vfs_ops == 0) { 2046 MPASS((mp->mnt_kern_flag & MNTK_SUSPEND) == 0); 2047 MNT_IUNLOCK(mp); 2048 return; 2049 } 2050 if (c < 0) 2051 vfs_dump_mount_counters(mp); 2052 if ((mp->mnt_kern_flag & MNTK_SUSPEND) != 0 && c == 0) 2053 wakeup(&mp->mnt_writeopcount); 2054 MNT_IUNLOCK(mp); 2055 } 2056 2057 /* 2058 * Filesystem secondary write operation has completed. If we are 2059 * suspending and this operation is the last one, notify the suspender 2060 * that the suspension is now in effect. 2061 */ 2062 void 2063 vn_finished_secondary_write(struct mount *mp) 2064 { 2065 if (mp == NULL) 2066 return; 2067 MNT_ILOCK(mp); 2068 MNT_REL(mp); 2069 mp->mnt_secondary_writes--; 2070 if (mp->mnt_secondary_writes < 0) 2071 panic("vn_finished_secondary_write: neg cnt"); 2072 if ((mp->mnt_kern_flag & MNTK_SUSPEND) != 0 && 2073 mp->mnt_secondary_writes <= 0) 2074 wakeup(&mp->mnt_secondary_writes); 2075 MNT_IUNLOCK(mp); 2076 } 2077 2078 /* 2079 * Request a filesystem to suspend write operations. 2080 */ 2081 int 2082 vfs_write_suspend(struct mount *mp, int flags) 2083 { 2084 int error; 2085 2086 vfs_op_enter(mp); 2087 2088 MNT_ILOCK(mp); 2089 vfs_assert_mount_counters(mp); 2090 if (mp->mnt_susp_owner == curthread) { 2091 vfs_op_exit_locked(mp); 2092 MNT_IUNLOCK(mp); 2093 return (EALREADY); 2094 } 2095 while (mp->mnt_kern_flag & MNTK_SUSPEND) 2096 msleep(&mp->mnt_flag, MNT_MTX(mp), PUSER - 1, "wsuspfs", 0); 2097 2098 /* 2099 * Unmount holds a write reference on the mount point. If we 2100 * own busy reference and drain for writers, we deadlock with 2101 * the reference draining in the unmount path. Callers of 2102 * vfs_write_suspend() must specify VS_SKIP_UNMOUNT if 2103 * vfs_busy() reference is owned and caller is not in the 2104 * unmount context. 2105 */ 2106 if ((flags & VS_SKIP_UNMOUNT) != 0 && 2107 (mp->mnt_kern_flag & MNTK_UNMOUNT) != 0) { 2108 vfs_op_exit_locked(mp); 2109 MNT_IUNLOCK(mp); 2110 return (EBUSY); 2111 } 2112 2113 mp->mnt_kern_flag |= MNTK_SUSPEND; 2114 mp->mnt_susp_owner = curthread; 2115 if (mp->mnt_writeopcount > 0) 2116 (void) msleep(&mp->mnt_writeopcount, 2117 MNT_MTX(mp), (PUSER - 1)|PDROP, "suspwt", 0); 2118 else 2119 MNT_IUNLOCK(mp); 2120 if ((error = VFS_SYNC(mp, MNT_SUSPEND)) != 0) { 2121 vfs_write_resume(mp, 0); 2122 /* vfs_write_resume does vfs_op_exit() for us */ 2123 } 2124 return (error); 2125 } 2126 2127 /* 2128 * Request a filesystem to resume write operations. 2129 */ 2130 void 2131 vfs_write_resume(struct mount *mp, int flags) 2132 { 2133 2134 MNT_ILOCK(mp); 2135 if ((mp->mnt_kern_flag & MNTK_SUSPEND) != 0) { 2136 KASSERT(mp->mnt_susp_owner == curthread, ("mnt_susp_owner")); 2137 mp->mnt_kern_flag &= ~(MNTK_SUSPEND | MNTK_SUSPEND2 | 2138 MNTK_SUSPENDED); 2139 mp->mnt_susp_owner = NULL; 2140 wakeup(&mp->mnt_writeopcount); 2141 wakeup(&mp->mnt_flag); 2142 curthread->td_pflags &= ~TDP_IGNSUSP; 2143 if ((flags & VR_START_WRITE) != 0) { 2144 MNT_REF(mp); 2145 mp->mnt_writeopcount++; 2146 } 2147 MNT_IUNLOCK(mp); 2148 if ((flags & VR_NO_SUSPCLR) == 0) 2149 VFS_SUSP_CLEAN(mp); 2150 vfs_op_exit(mp); 2151 } else if ((flags & VR_START_WRITE) != 0) { 2152 MNT_REF(mp); 2153 vn_start_write_refed(mp, 0, true); 2154 } else { 2155 MNT_IUNLOCK(mp); 2156 } 2157 } 2158 2159 /* 2160 * Helper loop around vfs_write_suspend() for filesystem unmount VFS 2161 * methods. 2162 */ 2163 int 2164 vfs_write_suspend_umnt(struct mount *mp) 2165 { 2166 int error; 2167 2168 KASSERT((curthread->td_pflags & TDP_IGNSUSP) == 0, 2169 ("vfs_write_suspend_umnt: recursed")); 2170 2171 /* dounmount() already called vn_start_write(). */ 2172 for (;;) { 2173 vn_finished_write(mp); 2174 error = vfs_write_suspend(mp, 0); 2175 if (error != 0) { 2176 vn_start_write(NULL, &mp, V_WAIT); 2177 return (error); 2178 } 2179 MNT_ILOCK(mp); 2180 if ((mp->mnt_kern_flag & MNTK_SUSPENDED) != 0) 2181 break; 2182 MNT_IUNLOCK(mp); 2183 vn_start_write(NULL, &mp, V_WAIT); 2184 } 2185 mp->mnt_kern_flag &= ~(MNTK_SUSPENDED | MNTK_SUSPEND2); 2186 wakeup(&mp->mnt_flag); 2187 MNT_IUNLOCK(mp); 2188 curthread->td_pflags |= TDP_IGNSUSP; 2189 return (0); 2190 } 2191 2192 /* 2193 * Implement kqueues for files by translating it to vnode operation. 2194 */ 2195 static int 2196 vn_kqfilter(struct file *fp, struct knote *kn) 2197 { 2198 2199 return (VOP_KQFILTER(fp->f_vnode, kn)); 2200 } 2201 2202 int 2203 vn_kqfilter_opath(struct file *fp, struct knote *kn) 2204 { 2205 if ((fp->f_flag & FKQALLOWED) == 0) 2206 return (EBADF); 2207 return (vn_kqfilter(fp, kn)); 2208 } 2209 2210 /* 2211 * Simplified in-kernel wrapper calls for extended attribute access. 2212 * Both calls pass in a NULL credential, authorizing as "kernel" access. 2213 * Set IO_NODELOCKED in ioflg if the vnode is already locked. 2214 */ 2215 int 2216 vn_extattr_get(struct vnode *vp, int ioflg, int attrnamespace, 2217 const char *attrname, int *buflen, char *buf, struct thread *td) 2218 { 2219 struct uio auio; 2220 struct iovec iov; 2221 int error; 2222 2223 iov.iov_len = *buflen; 2224 iov.iov_base = buf; 2225 2226 auio.uio_iov = &iov; 2227 auio.uio_iovcnt = 1; 2228 auio.uio_rw = UIO_READ; 2229 auio.uio_segflg = UIO_SYSSPACE; 2230 auio.uio_td = td; 2231 auio.uio_offset = 0; 2232 auio.uio_resid = *buflen; 2233 2234 if ((ioflg & IO_NODELOCKED) == 0) 2235 vn_lock(vp, LK_SHARED | LK_RETRY); 2236 2237 ASSERT_VOP_LOCKED(vp, "IO_NODELOCKED with no vp lock held"); 2238 2239 /* authorize attribute retrieval as kernel */ 2240 error = VOP_GETEXTATTR(vp, attrnamespace, attrname, &auio, NULL, NULL, 2241 td); 2242 2243 if ((ioflg & IO_NODELOCKED) == 0) 2244 VOP_UNLOCK(vp); 2245 2246 if (error == 0) { 2247 *buflen = *buflen - auio.uio_resid; 2248 } 2249 2250 return (error); 2251 } 2252 2253 /* 2254 * XXX failure mode if partially written? 2255 */ 2256 int 2257 vn_extattr_set(struct vnode *vp, int ioflg, int attrnamespace, 2258 const char *attrname, int buflen, char *buf, struct thread *td) 2259 { 2260 struct uio auio; 2261 struct iovec iov; 2262 struct mount *mp; 2263 int error; 2264 2265 iov.iov_len = buflen; 2266 iov.iov_base = buf; 2267 2268 auio.uio_iov = &iov; 2269 auio.uio_iovcnt = 1; 2270 auio.uio_rw = UIO_WRITE; 2271 auio.uio_segflg = UIO_SYSSPACE; 2272 auio.uio_td = td; 2273 auio.uio_offset = 0; 2274 auio.uio_resid = buflen; 2275 2276 if ((ioflg & IO_NODELOCKED) == 0) { 2277 if ((error = vn_start_write(vp, &mp, V_WAIT)) != 0) 2278 return (error); 2279 vn_lock(vp, LK_EXCLUSIVE | LK_RETRY); 2280 } 2281 2282 ASSERT_VOP_LOCKED(vp, "IO_NODELOCKED with no vp lock held"); 2283 2284 /* authorize attribute setting as kernel */ 2285 error = VOP_SETEXTATTR(vp, attrnamespace, attrname, &auio, NULL, td); 2286 2287 if ((ioflg & IO_NODELOCKED) == 0) { 2288 vn_finished_write(mp); 2289 VOP_UNLOCK(vp); 2290 } 2291 2292 return (error); 2293 } 2294 2295 int 2296 vn_extattr_rm(struct vnode *vp, int ioflg, int attrnamespace, 2297 const char *attrname, struct thread *td) 2298 { 2299 struct mount *mp; 2300 int error; 2301 2302 if ((ioflg & IO_NODELOCKED) == 0) { 2303 if ((error = vn_start_write(vp, &mp, V_WAIT)) != 0) 2304 return (error); 2305 vn_lock(vp, LK_EXCLUSIVE | LK_RETRY); 2306 } 2307 2308 ASSERT_VOP_LOCKED(vp, "IO_NODELOCKED with no vp lock held"); 2309 2310 /* authorize attribute removal as kernel */ 2311 error = VOP_DELETEEXTATTR(vp, attrnamespace, attrname, NULL, td); 2312 if (error == EOPNOTSUPP) 2313 error = VOP_SETEXTATTR(vp, attrnamespace, attrname, NULL, 2314 NULL, td); 2315 2316 if ((ioflg & IO_NODELOCKED) == 0) { 2317 vn_finished_write(mp); 2318 VOP_UNLOCK(vp); 2319 } 2320 2321 return (error); 2322 } 2323 2324 static int 2325 vn_get_ino_alloc_vget(struct mount *mp, void *arg, int lkflags, 2326 struct vnode **rvp) 2327 { 2328 2329 return (VFS_VGET(mp, *(ino_t *)arg, lkflags, rvp)); 2330 } 2331 2332 int 2333 vn_vget_ino(struct vnode *vp, ino_t ino, int lkflags, struct vnode **rvp) 2334 { 2335 2336 return (vn_vget_ino_gen(vp, vn_get_ino_alloc_vget, &ino, 2337 lkflags, rvp)); 2338 } 2339 2340 int 2341 vn_vget_ino_gen(struct vnode *vp, vn_get_ino_t alloc, void *alloc_arg, 2342 int lkflags, struct vnode **rvp) 2343 { 2344 struct mount *mp; 2345 int ltype, error; 2346 2347 ASSERT_VOP_LOCKED(vp, "vn_vget_ino_get"); 2348 mp = vp->v_mount; 2349 ltype = VOP_ISLOCKED(vp); 2350 KASSERT(ltype == LK_EXCLUSIVE || ltype == LK_SHARED, 2351 ("vn_vget_ino: vp not locked")); 2352 error = vfs_busy(mp, MBF_NOWAIT); 2353 if (error != 0) { 2354 vfs_ref(mp); 2355 VOP_UNLOCK(vp); 2356 error = vfs_busy(mp, 0); 2357 vn_lock(vp, ltype | LK_RETRY); 2358 vfs_rel(mp); 2359 if (error != 0) 2360 return (ENOENT); 2361 if (VN_IS_DOOMED(vp)) { 2362 vfs_unbusy(mp); 2363 return (ENOENT); 2364 } 2365 } 2366 VOP_UNLOCK(vp); 2367 error = alloc(mp, alloc_arg, lkflags, rvp); 2368 vfs_unbusy(mp); 2369 if (error != 0 || *rvp != vp) 2370 vn_lock(vp, ltype | LK_RETRY); 2371 if (VN_IS_DOOMED(vp)) { 2372 if (error == 0) { 2373 if (*rvp == vp) 2374 vunref(vp); 2375 else 2376 vput(*rvp); 2377 } 2378 error = ENOENT; 2379 } 2380 return (error); 2381 } 2382 2383 static void 2384 vn_send_sigxfsz(struct proc *p) 2385 { 2386 PROC_LOCK(p); 2387 kern_psignal(p, SIGXFSZ); 2388 PROC_UNLOCK(p); 2389 } 2390 2391 int 2392 vn_rlimit_trunc(u_quad_t size, struct thread *td) 2393 { 2394 if (size <= lim_cur(td, RLIMIT_FSIZE)) 2395 return (0); 2396 vn_send_sigxfsz(td->td_proc); 2397 return (EFBIG); 2398 } 2399 2400 static int 2401 vn_rlimit_fsizex1(const struct vnode *vp, struct uio *uio, off_t maxfsz, 2402 bool adj, struct thread *td) 2403 { 2404 off_t lim; 2405 bool ktr_write; 2406 2407 if (vp->v_type != VREG) 2408 return (0); 2409 2410 /* 2411 * Handle file system maximum file size. 2412 */ 2413 if (maxfsz != 0 && uio->uio_offset + uio->uio_resid > maxfsz) { 2414 if (!adj || uio->uio_offset >= maxfsz) 2415 return (EFBIG); 2416 uio->uio_resid = maxfsz - uio->uio_offset; 2417 } 2418 2419 /* 2420 * This is kernel write (e.g. vnode_pager) or accounting 2421 * write, ignore limit. 2422 */ 2423 if (td == NULL || (td->td_pflags2 & TDP2_ACCT) != 0) 2424 return (0); 2425 2426 /* 2427 * Calculate file size limit. 2428 */ 2429 ktr_write = (td->td_pflags & TDP_INKTRACE) != 0; 2430 lim = __predict_false(ktr_write) ? td->td_ktr_io_lim : 2431 lim_cur(td, RLIMIT_FSIZE); 2432 2433 /* 2434 * Is the limit reached? 2435 */ 2436 if (__predict_true((uoff_t)uio->uio_offset + uio->uio_resid <= lim)) 2437 return (0); 2438 2439 /* 2440 * Prepared filesystems can handle writes truncated to the 2441 * file size limit. 2442 */ 2443 if (adj && (uoff_t)uio->uio_offset < lim) { 2444 uio->uio_resid = lim - (uoff_t)uio->uio_offset; 2445 return (0); 2446 } 2447 2448 if (!ktr_write || ktr_filesize_limit_signal) 2449 vn_send_sigxfsz(td->td_proc); 2450 return (EFBIG); 2451 } 2452 2453 /* 2454 * Helper for VOP_WRITE() implementations, the common code to 2455 * handle maximum supported file size on the filesystem, and 2456 * RLIMIT_FSIZE, except for special writes from accounting subsystem 2457 * and ktrace. 2458 * 2459 * For maximum file size (maxfsz argument): 2460 * - return EFBIG if uio_offset is beyond it 2461 * - otherwise, clamp uio_resid if write would extend file beyond maxfsz. 2462 * 2463 * For RLIMIT_FSIZE: 2464 * - return EFBIG and send SIGXFSZ if uio_offset is beyond the limit 2465 * - otherwise, clamp uio_resid if write would extend file beyond limit. 2466 * 2467 * If clamping occured, the adjustment for uio_resid is stored in 2468 * *resid_adj, to be re-applied by vn_rlimit_fsizex_res() on return 2469 * from the VOP. 2470 */ 2471 int 2472 vn_rlimit_fsizex(const struct vnode *vp, struct uio *uio, off_t maxfsz, 2473 ssize_t *resid_adj, struct thread *td) 2474 { 2475 ssize_t resid_orig; 2476 int error; 2477 bool adj; 2478 2479 resid_orig = uio->uio_resid; 2480 adj = resid_adj != NULL; 2481 error = vn_rlimit_fsizex1(vp, uio, maxfsz, adj, td); 2482 if (adj) 2483 *resid_adj = resid_orig - uio->uio_resid; 2484 return (error); 2485 } 2486 2487 void 2488 vn_rlimit_fsizex_res(struct uio *uio, ssize_t resid_adj) 2489 { 2490 uio->uio_resid += resid_adj; 2491 } 2492 2493 int 2494 vn_rlimit_fsize(const struct vnode *vp, const struct uio *uio, 2495 struct thread *td) 2496 { 2497 return (vn_rlimit_fsizex(vp, __DECONST(struct uio *, uio), 0, NULL, 2498 td)); 2499 } 2500 2501 int 2502 vn_chmod(struct file *fp, mode_t mode, struct ucred *active_cred, 2503 struct thread *td) 2504 { 2505 struct vnode *vp; 2506 2507 vp = fp->f_vnode; 2508 #ifdef AUDIT 2509 vn_lock(vp, LK_SHARED | LK_RETRY); 2510 AUDIT_ARG_VNODE1(vp); 2511 VOP_UNLOCK(vp); 2512 #endif 2513 return (setfmode(td, active_cred, vp, mode)); 2514 } 2515 2516 int 2517 vn_chown(struct file *fp, uid_t uid, gid_t gid, struct ucred *active_cred, 2518 struct thread *td) 2519 { 2520 struct vnode *vp; 2521 2522 vp = fp->f_vnode; 2523 #ifdef AUDIT 2524 vn_lock(vp, LK_SHARED | LK_RETRY); 2525 AUDIT_ARG_VNODE1(vp); 2526 VOP_UNLOCK(vp); 2527 #endif 2528 return (setfown(td, active_cred, vp, uid, gid)); 2529 } 2530 2531 /* 2532 * Remove pages in the range ["start", "end") from the vnode's VM object. If 2533 * "end" is 0, then the range extends to the end of the object. 2534 */ 2535 void 2536 vn_pages_remove(struct vnode *vp, vm_pindex_t start, vm_pindex_t end) 2537 { 2538 vm_object_t object; 2539 2540 if ((object = vp->v_object) == NULL) 2541 return; 2542 VM_OBJECT_WLOCK(object); 2543 vm_object_page_remove(object, start, end, 0); 2544 VM_OBJECT_WUNLOCK(object); 2545 } 2546 2547 /* 2548 * Like vn_pages_remove(), but skips invalid pages, which by definition are not 2549 * mapped into any process' address space. Filesystems may use this in 2550 * preference to vn_pages_remove() to avoid blocking on pages busied in 2551 * preparation for a VOP_GETPAGES. 2552 */ 2553 void 2554 vn_pages_remove_valid(struct vnode *vp, vm_pindex_t start, vm_pindex_t end) 2555 { 2556 vm_object_t object; 2557 2558 if ((object = vp->v_object) == NULL) 2559 return; 2560 VM_OBJECT_WLOCK(object); 2561 vm_object_page_remove(object, start, end, OBJPR_VALIDONLY); 2562 VM_OBJECT_WUNLOCK(object); 2563 } 2564 2565 int 2566 vn_bmap_seekhole_locked(struct vnode *vp, u_long cmd, off_t *off, 2567 struct ucred *cred) 2568 { 2569 vm_object_t obj; 2570 off_t size; 2571 daddr_t bn, bnp; 2572 uint64_t bsize; 2573 off_t noff; 2574 int error; 2575 2576 KASSERT(cmd == FIOSEEKHOLE || cmd == FIOSEEKDATA, 2577 ("%s: Wrong command %lu", __func__, cmd)); 2578 ASSERT_VOP_ELOCKED(vp, "vn_bmap_seekhole_locked"); 2579 2580 if (vp->v_type != VREG) { 2581 error = ENOTTY; 2582 goto out; 2583 } 2584 error = vn_getsize_locked(vp, &size, cred); 2585 if (error != 0) 2586 goto out; 2587 noff = *off; 2588 if (noff < 0 || noff >= size) { 2589 error = ENXIO; 2590 goto out; 2591 } 2592 2593 /* See the comment in ufs_bmap_seekdata(). */ 2594 obj = vp->v_object; 2595 if (obj != NULL) { 2596 VM_OBJECT_WLOCK(obj); 2597 vm_object_page_clean(obj, 0, 0, OBJPC_SYNC); 2598 VM_OBJECT_WUNLOCK(obj); 2599 } 2600 2601 bsize = vp->v_mount->mnt_stat.f_iosize; 2602 for (bn = noff / bsize; noff < size; bn++, noff += bsize - 2603 noff % bsize) { 2604 error = VOP_BMAP(vp, bn, NULL, &bnp, NULL, NULL); 2605 if (error == EOPNOTSUPP) { 2606 error = ENOTTY; 2607 goto out; 2608 } 2609 if ((bnp == -1 && cmd == FIOSEEKHOLE) || 2610 (bnp != -1 && cmd == FIOSEEKDATA)) { 2611 noff = bn * bsize; 2612 if (noff < *off) 2613 noff = *off; 2614 goto out; 2615 } 2616 } 2617 if (noff > size) 2618 noff = size; 2619 /* noff == size. There is an implicit hole at the end of file. */ 2620 if (cmd == FIOSEEKDATA) 2621 error = ENXIO; 2622 out: 2623 if (error == 0) 2624 *off = noff; 2625 return (error); 2626 } 2627 2628 int 2629 vn_bmap_seekhole(struct vnode *vp, u_long cmd, off_t *off, struct ucred *cred) 2630 { 2631 int error; 2632 2633 KASSERT(cmd == FIOSEEKHOLE || cmd == FIOSEEKDATA, 2634 ("%s: Wrong command %lu", __func__, cmd)); 2635 2636 if (vn_lock(vp, LK_EXCLUSIVE) != 0) 2637 return (EBADF); 2638 error = vn_bmap_seekhole_locked(vp, cmd, off, cred); 2639 VOP_UNLOCK(vp); 2640 return (error); 2641 } 2642 2643 int 2644 vn_seek(struct file *fp, off_t offset, int whence, struct thread *td) 2645 { 2646 struct ucred *cred; 2647 struct vnode *vp; 2648 off_t foffset, fsize, size; 2649 int error, noneg; 2650 2651 cred = td->td_ucred; 2652 vp = fp->f_vnode; 2653 foffset = foffset_lock(fp, 0); 2654 noneg = (vp->v_type != VCHR); 2655 error = 0; 2656 switch (whence) { 2657 case L_INCR: 2658 if (noneg && 2659 (foffset < 0 || 2660 (offset > 0 && foffset > OFF_MAX - offset))) { 2661 error = EOVERFLOW; 2662 break; 2663 } 2664 offset += foffset; 2665 break; 2666 case L_XTND: 2667 error = vn_getsize(vp, &fsize, cred); 2668 if (error != 0) 2669 break; 2670 2671 /* 2672 * If the file references a disk device, then fetch 2673 * the media size and use that to determine the ending 2674 * offset. 2675 */ 2676 if (fsize == 0 && vp->v_type == VCHR && 2677 fo_ioctl(fp, DIOCGMEDIASIZE, &size, cred, td) == 0) 2678 fsize = size; 2679 if (noneg && offset > 0 && fsize > OFF_MAX - offset) { 2680 error = EOVERFLOW; 2681 break; 2682 } 2683 offset += fsize; 2684 break; 2685 case L_SET: 2686 break; 2687 case SEEK_DATA: 2688 error = fo_ioctl(fp, FIOSEEKDATA, &offset, cred, td); 2689 if (error == ENOTTY) 2690 error = EINVAL; 2691 break; 2692 case SEEK_HOLE: 2693 error = fo_ioctl(fp, FIOSEEKHOLE, &offset, cred, td); 2694 if (error == ENOTTY) 2695 error = EINVAL; 2696 break; 2697 default: 2698 error = EINVAL; 2699 } 2700 if (error == 0 && noneg && offset < 0) 2701 error = EINVAL; 2702 if (error != 0) 2703 goto drop; 2704 VFS_KNOTE_UNLOCKED(vp, 0); 2705 td->td_uretoff.tdu_off = offset; 2706 drop: 2707 foffset_unlock(fp, offset, error != 0 ? FOF_NOUPDATE : 0); 2708 return (error); 2709 } 2710 2711 int 2712 vn_utimes_perm(struct vnode *vp, struct vattr *vap, struct ucred *cred, 2713 struct thread *td) 2714 { 2715 int error; 2716 2717 /* 2718 * Grant permission if the caller is the owner of the file, or 2719 * the super-user, or has ACL_WRITE_ATTRIBUTES permission on 2720 * on the file. If the time pointer is null, then write 2721 * permission on the file is also sufficient. 2722 * 2723 * From NFSv4.1, draft 21, 6.2.1.3.1, Discussion of Mask Attributes: 2724 * A user having ACL_WRITE_DATA or ACL_WRITE_ATTRIBUTES 2725 * will be allowed to set the times [..] to the current 2726 * server time. 2727 */ 2728 error = VOP_ACCESSX(vp, VWRITE_ATTRIBUTES, cred, td); 2729 if (error != 0 && (vap->va_vaflags & VA_UTIMES_NULL) != 0) 2730 error = VOP_ACCESS(vp, VWRITE, cred, td); 2731 return (error); 2732 } 2733 2734 int 2735 vn_fill_kinfo(struct file *fp, struct kinfo_file *kif, struct filedesc *fdp) 2736 { 2737 struct vnode *vp; 2738 int error; 2739 2740 if (fp->f_type == DTYPE_FIFO) 2741 kif->kf_type = KF_TYPE_FIFO; 2742 else 2743 kif->kf_type = KF_TYPE_VNODE; 2744 vp = fp->f_vnode; 2745 vref(vp); 2746 FILEDESC_SUNLOCK(fdp); 2747 error = vn_fill_kinfo_vnode(vp, kif); 2748 vrele(vp); 2749 FILEDESC_SLOCK(fdp); 2750 return (error); 2751 } 2752 2753 static inline void 2754 vn_fill_junk(struct kinfo_file *kif) 2755 { 2756 size_t len, olen; 2757 2758 /* 2759 * Simulate vn_fullpath returning changing values for a given 2760 * vp during e.g. coredump. 2761 */ 2762 len = (arc4random() % (sizeof(kif->kf_path) - 2)) + 1; 2763 olen = strlen(kif->kf_path); 2764 if (len < olen) 2765 strcpy(&kif->kf_path[len - 1], "$"); 2766 else 2767 for (; olen < len; olen++) 2768 strcpy(&kif->kf_path[olen], "A"); 2769 } 2770 2771 int 2772 vn_fill_kinfo_vnode(struct vnode *vp, struct kinfo_file *kif) 2773 { 2774 struct vattr va; 2775 char *fullpath, *freepath; 2776 int error; 2777 2778 kif->kf_un.kf_file.kf_file_type = vntype_to_kinfo(vp->v_type); 2779 freepath = NULL; 2780 fullpath = "-"; 2781 error = vn_fullpath(vp, &fullpath, &freepath); 2782 if (error == 0) { 2783 strlcpy(kif->kf_path, fullpath, sizeof(kif->kf_path)); 2784 } 2785 if (freepath != NULL) 2786 free(freepath, M_TEMP); 2787 2788 KFAIL_POINT_CODE(DEBUG_FP, fill_kinfo_vnode__random_path, 2789 vn_fill_junk(kif); 2790 ); 2791 2792 /* 2793 * Retrieve vnode attributes. 2794 */ 2795 va.va_fsid = VNOVAL; 2796 va.va_rdev = NODEV; 2797 vn_lock(vp, LK_SHARED | LK_RETRY); 2798 error = VOP_GETATTR(vp, &va, curthread->td_ucred); 2799 VOP_UNLOCK(vp); 2800 if (error != 0) 2801 return (error); 2802 if (va.va_fsid != VNOVAL) 2803 kif->kf_un.kf_file.kf_file_fsid = va.va_fsid; 2804 else 2805 kif->kf_un.kf_file.kf_file_fsid = 2806 vp->v_mount->mnt_stat.f_fsid.val[0]; 2807 kif->kf_un.kf_file.kf_file_fsid_freebsd11 = 2808 kif->kf_un.kf_file.kf_file_fsid; /* truncate */ 2809 kif->kf_un.kf_file.kf_file_fileid = va.va_fileid; 2810 kif->kf_un.kf_file.kf_file_mode = MAKEIMODE(va.va_type, va.va_mode); 2811 kif->kf_un.kf_file.kf_file_size = va.va_size; 2812 kif->kf_un.kf_file.kf_file_rdev = va.va_rdev; 2813 kif->kf_un.kf_file.kf_file_rdev_freebsd11 = 2814 kif->kf_un.kf_file.kf_file_rdev; /* truncate */ 2815 kif->kf_un.kf_file.kf_file_nlink = va.va_nlink; 2816 return (0); 2817 } 2818 2819 int 2820 vn_mmap(struct file *fp, vm_map_t map, vm_offset_t *addr, vm_size_t size, 2821 vm_prot_t prot, vm_prot_t cap_maxprot, int flags, vm_ooffset_t foff, 2822 struct thread *td) 2823 { 2824 #ifdef HWPMC_HOOKS 2825 struct pmckern_map_in pkm; 2826 #endif 2827 struct mount *mp; 2828 struct vnode *vp; 2829 vm_object_t object; 2830 vm_prot_t maxprot; 2831 boolean_t writecounted; 2832 int error; 2833 2834 #if defined(COMPAT_FREEBSD7) || defined(COMPAT_FREEBSD6) || \ 2835 defined(COMPAT_FREEBSD5) || defined(COMPAT_FREEBSD4) 2836 /* 2837 * POSIX shared-memory objects are defined to have 2838 * kernel persistence, and are not defined to support 2839 * read(2)/write(2) -- or even open(2). Thus, we can 2840 * use MAP_ASYNC to trade on-disk coherence for speed. 2841 * The shm_open(3) library routine turns on the FPOSIXSHM 2842 * flag to request this behavior. 2843 */ 2844 if ((fp->f_flag & FPOSIXSHM) != 0) 2845 flags |= MAP_NOSYNC; 2846 #endif 2847 vp = fp->f_vnode; 2848 2849 /* 2850 * Ensure that file and memory protections are 2851 * compatible. Note that we only worry about 2852 * writability if mapping is shared; in this case, 2853 * current and max prot are dictated by the open file. 2854 * XXX use the vnode instead? Problem is: what 2855 * credentials do we use for determination? What if 2856 * proc does a setuid? 2857 */ 2858 mp = vp->v_mount; 2859 if (mp != NULL && (mp->mnt_flag & MNT_NOEXEC) != 0) { 2860 maxprot = VM_PROT_NONE; 2861 if ((prot & VM_PROT_EXECUTE) != 0) 2862 return (EACCES); 2863 } else 2864 maxprot = VM_PROT_EXECUTE; 2865 if ((fp->f_flag & FREAD) != 0) 2866 maxprot |= VM_PROT_READ; 2867 else if ((prot & VM_PROT_READ) != 0) 2868 return (EACCES); 2869 2870 /* 2871 * If we are sharing potential changes via MAP_SHARED and we 2872 * are trying to get write permission although we opened it 2873 * without asking for it, bail out. 2874 */ 2875 if ((flags & MAP_SHARED) != 0) { 2876 if ((fp->f_flag & FWRITE) != 0) 2877 maxprot |= VM_PROT_WRITE; 2878 else if ((prot & VM_PROT_WRITE) != 0) 2879 return (EACCES); 2880 } else { 2881 maxprot |= VM_PROT_WRITE; 2882 cap_maxprot |= VM_PROT_WRITE; 2883 } 2884 maxprot &= cap_maxprot; 2885 2886 /* 2887 * For regular files and shared memory, POSIX requires that 2888 * the value of foff be a legitimate offset within the data 2889 * object. In particular, negative offsets are invalid. 2890 * Blocking negative offsets and overflows here avoids 2891 * possible wraparound or user-level access into reserved 2892 * ranges of the data object later. In contrast, POSIX does 2893 * not dictate how offsets are used by device drivers, so in 2894 * the case of a device mapping a negative offset is passed 2895 * on. 2896 */ 2897 if ( 2898 #ifdef _LP64 2899 size > OFF_MAX || 2900 #endif 2901 foff > OFF_MAX - size) 2902 return (EINVAL); 2903 2904 writecounted = FALSE; 2905 error = vm_mmap_vnode(td, size, prot, &maxprot, &flags, vp, 2906 &foff, &object, &writecounted); 2907 if (error != 0) 2908 return (error); 2909 error = vm_mmap_object(map, addr, size, prot, maxprot, flags, object, 2910 foff, writecounted, td); 2911 if (error != 0) { 2912 /* 2913 * If this mapping was accounted for in the vnode's 2914 * writecount, then undo that now. 2915 */ 2916 if (writecounted) 2917 vm_pager_release_writecount(object, 0, size); 2918 vm_object_deallocate(object); 2919 } 2920 #ifdef HWPMC_HOOKS 2921 /* Inform hwpmc(4) if an executable is being mapped. */ 2922 if (PMC_HOOK_INSTALLED(PMC_FN_MMAP)) { 2923 if ((prot & VM_PROT_EXECUTE) != 0 && error == 0) { 2924 pkm.pm_file = vp; 2925 pkm.pm_address = (uintptr_t) *addr; 2926 PMC_CALL_HOOK_UNLOCKED(td, PMC_FN_MMAP, (void *) &pkm); 2927 } 2928 } 2929 #endif 2930 return (error); 2931 } 2932 2933 void 2934 vn_fsid(struct vnode *vp, struct vattr *va) 2935 { 2936 fsid_t *f; 2937 2938 f = &vp->v_mount->mnt_stat.f_fsid; 2939 va->va_fsid = (uint32_t)f->val[1]; 2940 va->va_fsid <<= sizeof(f->val[1]) * NBBY; 2941 va->va_fsid += (uint32_t)f->val[0]; 2942 } 2943 2944 int 2945 vn_fsync_buf(struct vnode *vp, int waitfor) 2946 { 2947 struct buf *bp, *nbp; 2948 struct bufobj *bo; 2949 struct mount *mp; 2950 int error, maxretry; 2951 2952 error = 0; 2953 maxretry = 10000; /* large, arbitrarily chosen */ 2954 mp = NULL; 2955 if (vp->v_type == VCHR) { 2956 VI_LOCK(vp); 2957 mp = vp->v_rdev->si_mountpt; 2958 VI_UNLOCK(vp); 2959 } 2960 bo = &vp->v_bufobj; 2961 BO_LOCK(bo); 2962 loop1: 2963 /* 2964 * MARK/SCAN initialization to avoid infinite loops. 2965 */ 2966 TAILQ_FOREACH(bp, &bo->bo_dirty.bv_hd, b_bobufs) { 2967 bp->b_vflags &= ~BV_SCANNED; 2968 bp->b_error = 0; 2969 } 2970 2971 /* 2972 * Flush all dirty buffers associated with a vnode. 2973 */ 2974 loop2: 2975 TAILQ_FOREACH_SAFE(bp, &bo->bo_dirty.bv_hd, b_bobufs, nbp) { 2976 if ((bp->b_vflags & BV_SCANNED) != 0) 2977 continue; 2978 bp->b_vflags |= BV_SCANNED; 2979 if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT, NULL)) { 2980 if (waitfor != MNT_WAIT) 2981 continue; 2982 if (BUF_LOCK(bp, 2983 LK_EXCLUSIVE | LK_INTERLOCK | LK_SLEEPFAIL, 2984 BO_LOCKPTR(bo)) != 0) { 2985 BO_LOCK(bo); 2986 goto loop1; 2987 } 2988 BO_LOCK(bo); 2989 } 2990 BO_UNLOCK(bo); 2991 KASSERT(bp->b_bufobj == bo, 2992 ("bp %p wrong b_bufobj %p should be %p", 2993 bp, bp->b_bufobj, bo)); 2994 if ((bp->b_flags & B_DELWRI) == 0) 2995 panic("fsync: not dirty"); 2996 if ((vp->v_object != NULL) && (bp->b_flags & B_CLUSTEROK)) { 2997 vfs_bio_awrite(bp); 2998 } else { 2999 bremfree(bp); 3000 bawrite(bp); 3001 } 3002 if (maxretry < 1000) 3003 pause("dirty", hz < 1000 ? 1 : hz / 1000); 3004 BO_LOCK(bo); 3005 goto loop2; 3006 } 3007 3008 /* 3009 * If synchronous the caller expects us to completely resolve all 3010 * dirty buffers in the system. Wait for in-progress I/O to 3011 * complete (which could include background bitmap writes), then 3012 * retry if dirty blocks still exist. 3013 */ 3014 if (waitfor == MNT_WAIT) { 3015 bufobj_wwait(bo, 0, 0); 3016 if (bo->bo_dirty.bv_cnt > 0) { 3017 /* 3018 * If we are unable to write any of these buffers 3019 * then we fail now rather than trying endlessly 3020 * to write them out. 3021 */ 3022 TAILQ_FOREACH(bp, &bo->bo_dirty.bv_hd, b_bobufs) 3023 if ((error = bp->b_error) != 0) 3024 break; 3025 if ((mp != NULL && mp->mnt_secondary_writes > 0) || 3026 (error == 0 && --maxretry >= 0)) 3027 goto loop1; 3028 if (error == 0) 3029 error = EAGAIN; 3030 } 3031 } 3032 BO_UNLOCK(bo); 3033 if (error != 0) 3034 vn_printf(vp, "fsync: giving up on dirty (error = %d) ", error); 3035 3036 return (error); 3037 } 3038 3039 /* 3040 * Copies a byte range from invp to outvp. Calls VOP_COPY_FILE_RANGE() 3041 * or vn_generic_copy_file_range() after rangelocking the byte ranges, 3042 * to do the actual copy. 3043 * vn_generic_copy_file_range() is factored out, so it can be called 3044 * from a VOP_COPY_FILE_RANGE() call as well, but handles vnodes from 3045 * different file systems. 3046 */ 3047 int 3048 vn_copy_file_range(struct vnode *invp, off_t *inoffp, struct vnode *outvp, 3049 off_t *outoffp, size_t *lenp, unsigned int flags, struct ucred *incred, 3050 struct ucred *outcred, struct thread *fsize_td) 3051 { 3052 int error; 3053 size_t len; 3054 uint64_t uval; 3055 3056 len = *lenp; 3057 *lenp = 0; /* For error returns. */ 3058 error = 0; 3059 3060 /* Do some sanity checks on the arguments. */ 3061 if (invp->v_type == VDIR || outvp->v_type == VDIR) 3062 error = EISDIR; 3063 else if (*inoffp < 0 || *outoffp < 0 || 3064 invp->v_type != VREG || outvp->v_type != VREG) 3065 error = EINVAL; 3066 if (error != 0) 3067 goto out; 3068 3069 /* Ensure offset + len does not wrap around. */ 3070 uval = *inoffp; 3071 uval += len; 3072 if (uval > INT64_MAX) 3073 len = INT64_MAX - *inoffp; 3074 uval = *outoffp; 3075 uval += len; 3076 if (uval > INT64_MAX) 3077 len = INT64_MAX - *outoffp; 3078 if (len == 0) 3079 goto out; 3080 3081 /* 3082 * If the two vnode are for the same file system, call 3083 * VOP_COPY_FILE_RANGE(), otherwise call vn_generic_copy_file_range() 3084 * which can handle copies across multiple file systems. 3085 */ 3086 *lenp = len; 3087 if (invp->v_mount == outvp->v_mount) 3088 error = VOP_COPY_FILE_RANGE(invp, inoffp, outvp, outoffp, 3089 lenp, flags, incred, outcred, fsize_td); 3090 else 3091 error = vn_generic_copy_file_range(invp, inoffp, outvp, 3092 outoffp, lenp, flags, incred, outcred, fsize_td); 3093 out: 3094 return (error); 3095 } 3096 3097 /* 3098 * Test len bytes of data starting at dat for all bytes == 0. 3099 * Return true if all bytes are zero, false otherwise. 3100 * Expects dat to be well aligned. 3101 */ 3102 static bool 3103 mem_iszero(void *dat, int len) 3104 { 3105 int i; 3106 const u_int *p; 3107 const char *cp; 3108 3109 for (p = dat; len > 0; len -= sizeof(*p), p++) { 3110 if (len >= sizeof(*p)) { 3111 if (*p != 0) 3112 return (false); 3113 } else { 3114 cp = (const char *)p; 3115 for (i = 0; i < len; i++, cp++) 3116 if (*cp != '\0') 3117 return (false); 3118 } 3119 } 3120 return (true); 3121 } 3122 3123 /* 3124 * Look for a hole in the output file and, if found, adjust *outoffp 3125 * and *xferp to skip past the hole. 3126 * *xferp is the entire hole length to be written and xfer2 is how many bytes 3127 * to be written as 0's upon return. 3128 */ 3129 static off_t 3130 vn_skip_hole(struct vnode *outvp, off_t xfer2, off_t *outoffp, off_t *xferp, 3131 off_t *dataoffp, off_t *holeoffp, struct ucred *cred) 3132 { 3133 int error; 3134 off_t delta; 3135 3136 if (*holeoffp == 0 || *holeoffp <= *outoffp) { 3137 *dataoffp = *outoffp; 3138 error = VOP_IOCTL(outvp, FIOSEEKDATA, dataoffp, 0, cred, 3139 curthread); 3140 if (error == 0) { 3141 *holeoffp = *dataoffp; 3142 error = VOP_IOCTL(outvp, FIOSEEKHOLE, holeoffp, 0, cred, 3143 curthread); 3144 } 3145 if (error != 0 || *holeoffp == *dataoffp) { 3146 /* 3147 * Since outvp is unlocked, it may be possible for 3148 * another thread to do a truncate(), lseek(), write() 3149 * creating a hole at startoff between the above 3150 * VOP_IOCTL() calls, if the other thread does not do 3151 * rangelocking. 3152 * If that happens, *holeoffp == *dataoffp and finding 3153 * the hole has failed, so disable vn_skip_hole(). 3154 */ 3155 *holeoffp = -1; /* Disable use of vn_skip_hole(). */ 3156 return (xfer2); 3157 } 3158 KASSERT(*dataoffp >= *outoffp, 3159 ("vn_skip_hole: dataoff=%jd < outoff=%jd", 3160 (intmax_t)*dataoffp, (intmax_t)*outoffp)); 3161 KASSERT(*holeoffp > *dataoffp, 3162 ("vn_skip_hole: holeoff=%jd <= dataoff=%jd", 3163 (intmax_t)*holeoffp, (intmax_t)*dataoffp)); 3164 } 3165 3166 /* 3167 * If there is a hole before the data starts, advance *outoffp and 3168 * *xferp past the hole. 3169 */ 3170 if (*dataoffp > *outoffp) { 3171 delta = *dataoffp - *outoffp; 3172 if (delta >= *xferp) { 3173 /* Entire *xferp is a hole. */ 3174 *outoffp += *xferp; 3175 *xferp = 0; 3176 return (0); 3177 } 3178 *xferp -= delta; 3179 *outoffp += delta; 3180 xfer2 = MIN(xfer2, *xferp); 3181 } 3182 3183 /* 3184 * If a hole starts before the end of this xfer2, reduce this xfer2 so 3185 * that the write ends at the start of the hole. 3186 * *holeoffp should always be greater than *outoffp, but for the 3187 * non-INVARIANTS case, check this to make sure xfer2 remains a sane 3188 * value. 3189 */ 3190 if (*holeoffp > *outoffp && *holeoffp < *outoffp + xfer2) 3191 xfer2 = *holeoffp - *outoffp; 3192 return (xfer2); 3193 } 3194 3195 /* 3196 * Write an xfer sized chunk to outvp in blksize blocks from dat. 3197 * dat is a maximum of blksize in length and can be written repeatedly in 3198 * the chunk. 3199 * If growfile == true, just grow the file via vn_truncate_locked() instead 3200 * of doing actual writes. 3201 * If checkhole == true, a hole is being punched, so skip over any hole 3202 * already in the output file. 3203 */ 3204 static int 3205 vn_write_outvp(struct vnode *outvp, char *dat, off_t outoff, off_t xfer, 3206 u_long blksize, bool growfile, bool checkhole, struct ucred *cred) 3207 { 3208 struct mount *mp; 3209 off_t dataoff, holeoff, xfer2; 3210 int error; 3211 3212 /* 3213 * Loop around doing writes of blksize until write has been completed. 3214 * Lock/unlock on each loop iteration so that a bwillwrite() can be 3215 * done for each iteration, since the xfer argument can be very 3216 * large if there is a large hole to punch in the output file. 3217 */ 3218 error = 0; 3219 holeoff = 0; 3220 do { 3221 xfer2 = MIN(xfer, blksize); 3222 if (checkhole) { 3223 /* 3224 * Punching a hole. Skip writing if there is 3225 * already a hole in the output file. 3226 */ 3227 xfer2 = vn_skip_hole(outvp, xfer2, &outoff, &xfer, 3228 &dataoff, &holeoff, cred); 3229 if (xfer == 0) 3230 break; 3231 if (holeoff < 0) 3232 checkhole = false; 3233 KASSERT(xfer2 > 0, ("vn_write_outvp: xfer2=%jd", 3234 (intmax_t)xfer2)); 3235 } 3236 bwillwrite(); 3237 mp = NULL; 3238 error = vn_start_write(outvp, &mp, V_WAIT); 3239 if (error != 0) 3240 break; 3241 if (growfile) { 3242 error = vn_lock(outvp, LK_EXCLUSIVE); 3243 if (error == 0) { 3244 error = vn_truncate_locked(outvp, outoff + xfer, 3245 false, cred); 3246 VOP_UNLOCK(outvp); 3247 } 3248 } else { 3249 error = vn_lock(outvp, vn_lktype_write(mp, outvp)); 3250 if (error == 0) { 3251 error = vn_rdwr(UIO_WRITE, outvp, dat, xfer2, 3252 outoff, UIO_SYSSPACE, IO_NODELOCKED, 3253 curthread->td_ucred, cred, NULL, curthread); 3254 outoff += xfer2; 3255 xfer -= xfer2; 3256 VOP_UNLOCK(outvp); 3257 } 3258 } 3259 if (mp != NULL) 3260 vn_finished_write(mp); 3261 } while (!growfile && xfer > 0 && error == 0); 3262 return (error); 3263 } 3264 3265 /* 3266 * Copy a byte range of one file to another. This function can handle the 3267 * case where invp and outvp are on different file systems. 3268 * It can also be called by a VOP_COPY_FILE_RANGE() to do the work, if there 3269 * is no better file system specific way to do it. 3270 */ 3271 int 3272 vn_generic_copy_file_range(struct vnode *invp, off_t *inoffp, 3273 struct vnode *outvp, off_t *outoffp, size_t *lenp, unsigned int flags, 3274 struct ucred *incred, struct ucred *outcred, struct thread *fsize_td) 3275 { 3276 struct mount *mp; 3277 off_t startoff, endoff, xfer, xfer2; 3278 u_long blksize; 3279 int error, interrupted; 3280 bool cantseek, readzeros, eof, lastblock, holetoeof; 3281 ssize_t aresid, r = 0; 3282 size_t copylen, len, savlen; 3283 off_t insize, outsize; 3284 char *dat; 3285 long holein, holeout; 3286 struct timespec curts, endts; 3287 3288 holein = holeout = 0; 3289 savlen = len = *lenp; 3290 error = 0; 3291 interrupted = 0; 3292 dat = NULL; 3293 3294 error = vn_lock(invp, LK_SHARED); 3295 if (error != 0) 3296 goto out; 3297 if (VOP_PATHCONF(invp, _PC_MIN_HOLE_SIZE, &holein) != 0) 3298 holein = 0; 3299 if (holein > 0) 3300 error = vn_getsize_locked(invp, &insize, incred); 3301 VOP_UNLOCK(invp); 3302 if (error != 0) 3303 goto out; 3304 3305 mp = NULL; 3306 error = vn_start_write(outvp, &mp, V_WAIT); 3307 if (error == 0) 3308 error = vn_lock(outvp, LK_EXCLUSIVE); 3309 if (error == 0) { 3310 /* 3311 * If fsize_td != NULL, do a vn_rlimit_fsizex() call, 3312 * now that outvp is locked. 3313 */ 3314 if (fsize_td != NULL) { 3315 struct uio io; 3316 3317 io.uio_offset = *outoffp; 3318 io.uio_resid = len; 3319 error = vn_rlimit_fsizex(outvp, &io, 0, &r, fsize_td); 3320 len = savlen = io.uio_resid; 3321 /* 3322 * No need to call vn_rlimit_fsizex_res before return, 3323 * since the uio is local. 3324 */ 3325 } 3326 if (VOP_PATHCONF(outvp, _PC_MIN_HOLE_SIZE, &holeout) != 0) 3327 holeout = 0; 3328 /* 3329 * Holes that are past EOF do not need to be written as a block 3330 * of zero bytes. So, truncate the output file as far as 3331 * possible and then use size to decide if writing 0 3332 * bytes is necessary in the loop below. 3333 */ 3334 if (error == 0) 3335 error = vn_getsize_locked(outvp, &outsize, outcred); 3336 if (error == 0 && outsize > *outoffp && outsize <= *outoffp + len) { 3337 #ifdef MAC 3338 error = mac_vnode_check_write(curthread->td_ucred, 3339 outcred, outvp); 3340 if (error == 0) 3341 #endif 3342 error = vn_truncate_locked(outvp, *outoffp, 3343 false, outcred); 3344 if (error == 0) 3345 outsize = *outoffp; 3346 } 3347 VOP_UNLOCK(outvp); 3348 } 3349 if (mp != NULL) 3350 vn_finished_write(mp); 3351 if (error != 0) 3352 goto out; 3353 3354 if (holein == 0 && holeout > 0) { 3355 /* 3356 * For this special case, the input data will be scanned 3357 * for blocks of all 0 bytes. For these blocks, the 3358 * write can be skipped for the output file to create 3359 * an unallocated region. 3360 * Therefore, use the appropriate size for the output file. 3361 */ 3362 blksize = holeout; 3363 if (blksize <= 512) { 3364 /* 3365 * Use f_iosize, since ZFS reports a _PC_MIN_HOLE_SIZE 3366 * of 512, although it actually only creates 3367 * unallocated regions for blocks >= f_iosize. 3368 */ 3369 blksize = outvp->v_mount->mnt_stat.f_iosize; 3370 } 3371 } else { 3372 /* 3373 * Use the larger of the two f_iosize values. If they are 3374 * not the same size, one will normally be an exact multiple of 3375 * the other, since they are both likely to be a power of 2. 3376 */ 3377 blksize = MAX(invp->v_mount->mnt_stat.f_iosize, 3378 outvp->v_mount->mnt_stat.f_iosize); 3379 } 3380 3381 /* Clip to sane limits. */ 3382 if (blksize < 4096) 3383 blksize = 4096; 3384 else if (blksize > maxphys) 3385 blksize = maxphys; 3386 dat = malloc(blksize, M_TEMP, M_WAITOK); 3387 3388 /* 3389 * If VOP_IOCTL(FIOSEEKHOLE) works for invp, use it and FIOSEEKDATA 3390 * to find holes. Otherwise, just scan the read block for all 0s 3391 * in the inner loop where the data copying is done. 3392 * Note that some file systems such as NFSv3, NFSv4.0 and NFSv4.1 may 3393 * support holes on the server, but do not support FIOSEEKHOLE. 3394 * The kernel flag COPY_FILE_RANGE_TIMEO1SEC is used to indicate 3395 * that this function should return after 1second with a partial 3396 * completion. 3397 */ 3398 if ((flags & COPY_FILE_RANGE_TIMEO1SEC) != 0) { 3399 getnanouptime(&endts); 3400 endts.tv_sec++; 3401 } else 3402 timespecclear(&endts); 3403 holetoeof = eof = false; 3404 while (len > 0 && error == 0 && !eof && interrupted == 0) { 3405 endoff = 0; /* To shut up compilers. */ 3406 cantseek = true; 3407 startoff = *inoffp; 3408 copylen = len; 3409 3410 /* 3411 * Find the next data area. If there is just a hole to EOF, 3412 * FIOSEEKDATA should fail with ENXIO. 3413 * (I do not know if any file system will report a hole to 3414 * EOF via FIOSEEKHOLE, but I am pretty sure FIOSEEKDATA 3415 * will fail for those file systems.) 3416 * 3417 * For input files that don't support FIOSEEKDATA/FIOSEEKHOLE, 3418 * the code just falls through to the inner copy loop. 3419 */ 3420 error = EINVAL; 3421 if (holein > 0) { 3422 error = VOP_IOCTL(invp, FIOSEEKDATA, &startoff, 0, 3423 incred, curthread); 3424 if (error == ENXIO) { 3425 startoff = endoff = insize; 3426 eof = holetoeof = true; 3427 error = 0; 3428 } 3429 } 3430 if (error == 0 && !holetoeof) { 3431 endoff = startoff; 3432 error = VOP_IOCTL(invp, FIOSEEKHOLE, &endoff, 0, 3433 incred, curthread); 3434 /* 3435 * Since invp is unlocked, it may be possible for 3436 * another thread to do a truncate(), lseek(), write() 3437 * creating a hole at startoff between the above 3438 * VOP_IOCTL() calls, if the other thread does not do 3439 * rangelocking. 3440 * If that happens, startoff == endoff and finding 3441 * the hole has failed, so set an error. 3442 */ 3443 if (error == 0 && startoff == endoff) 3444 error = EINVAL; /* Any error. Reset to 0. */ 3445 } 3446 if (error == 0) { 3447 if (startoff > *inoffp) { 3448 /* Found hole before data block. */ 3449 xfer = MIN(startoff - *inoffp, len); 3450 if (*outoffp < outsize) { 3451 /* Must write 0s to punch hole. */ 3452 xfer2 = MIN(outsize - *outoffp, 3453 xfer); 3454 memset(dat, 0, MIN(xfer2, blksize)); 3455 error = vn_write_outvp(outvp, dat, 3456 *outoffp, xfer2, blksize, false, 3457 holeout > 0, outcred); 3458 } 3459 3460 if (error == 0 && *outoffp + xfer > 3461 outsize && (xfer == len || holetoeof)) { 3462 /* Grow output file (hole at end). */ 3463 error = vn_write_outvp(outvp, dat, 3464 *outoffp, xfer, blksize, true, 3465 false, outcred); 3466 } 3467 if (error == 0) { 3468 *inoffp += xfer; 3469 *outoffp += xfer; 3470 len -= xfer; 3471 if (len < savlen) { 3472 interrupted = sig_intr(); 3473 if (timespecisset(&endts) && 3474 interrupted == 0) { 3475 getnanouptime(&curts); 3476 if (timespeccmp(&curts, 3477 &endts, >=)) 3478 interrupted = 3479 EINTR; 3480 } 3481 } 3482 } 3483 } 3484 copylen = MIN(len, endoff - startoff); 3485 cantseek = false; 3486 } else { 3487 cantseek = true; 3488 startoff = *inoffp; 3489 copylen = len; 3490 error = 0; 3491 } 3492 3493 xfer = blksize; 3494 if (cantseek) { 3495 /* 3496 * Set first xfer to end at a block boundary, so that 3497 * holes are more likely detected in the loop below via 3498 * the for all bytes 0 method. 3499 */ 3500 xfer -= (*inoffp % blksize); 3501 } 3502 /* Loop copying the data block. */ 3503 while (copylen > 0 && error == 0 && !eof && interrupted == 0) { 3504 if (copylen < xfer) 3505 xfer = copylen; 3506 error = vn_lock(invp, LK_SHARED); 3507 if (error != 0) 3508 goto out; 3509 error = vn_rdwr(UIO_READ, invp, dat, xfer, 3510 startoff, UIO_SYSSPACE, IO_NODELOCKED, 3511 curthread->td_ucred, incred, &aresid, 3512 curthread); 3513 VOP_UNLOCK(invp); 3514 lastblock = false; 3515 if (error == 0 && aresid > 0) { 3516 /* Stop the copy at EOF on the input file. */ 3517 xfer -= aresid; 3518 eof = true; 3519 lastblock = true; 3520 } 3521 if (error == 0) { 3522 /* 3523 * Skip the write for holes past the initial EOF 3524 * of the output file, unless this is the last 3525 * write of the output file at EOF. 3526 */ 3527 readzeros = cantseek ? mem_iszero(dat, xfer) : 3528 false; 3529 if (xfer == len) 3530 lastblock = true; 3531 if (!cantseek || *outoffp < outsize || 3532 lastblock || !readzeros) 3533 error = vn_write_outvp(outvp, dat, 3534 *outoffp, xfer, blksize, 3535 readzeros && lastblock && 3536 *outoffp >= outsize, false, 3537 outcred); 3538 if (error == 0) { 3539 *inoffp += xfer; 3540 startoff += xfer; 3541 *outoffp += xfer; 3542 copylen -= xfer; 3543 len -= xfer; 3544 if (len < savlen) { 3545 interrupted = sig_intr(); 3546 if (timespecisset(&endts) && 3547 interrupted == 0) { 3548 getnanouptime(&curts); 3549 if (timespeccmp(&curts, 3550 &endts, >=)) 3551 interrupted = 3552 EINTR; 3553 } 3554 } 3555 } 3556 } 3557 xfer = blksize; 3558 } 3559 } 3560 out: 3561 *lenp = savlen - len; 3562 free(dat, M_TEMP); 3563 return (error); 3564 } 3565 3566 static int 3567 vn_fallocate(struct file *fp, off_t offset, off_t len, struct thread *td) 3568 { 3569 struct mount *mp; 3570 struct vnode *vp; 3571 off_t olen, ooffset; 3572 int error; 3573 #ifdef AUDIT 3574 int audited_vnode1 = 0; 3575 #endif 3576 3577 vp = fp->f_vnode; 3578 if (vp->v_type != VREG) 3579 return (ENODEV); 3580 3581 /* Allocating blocks may take a long time, so iterate. */ 3582 for (;;) { 3583 olen = len; 3584 ooffset = offset; 3585 3586 bwillwrite(); 3587 mp = NULL; 3588 error = vn_start_write(vp, &mp, V_WAIT | V_PCATCH); 3589 if (error != 0) 3590 break; 3591 error = vn_lock(vp, LK_EXCLUSIVE); 3592 if (error != 0) { 3593 vn_finished_write(mp); 3594 break; 3595 } 3596 #ifdef AUDIT 3597 if (!audited_vnode1) { 3598 AUDIT_ARG_VNODE1(vp); 3599 audited_vnode1 = 1; 3600 } 3601 #endif 3602 #ifdef MAC 3603 error = mac_vnode_check_write(td->td_ucred, fp->f_cred, vp); 3604 if (error == 0) 3605 #endif 3606 error = VOP_ALLOCATE(vp, &offset, &len, 0, 3607 td->td_ucred); 3608 VOP_UNLOCK(vp); 3609 vn_finished_write(mp); 3610 3611 if (olen + ooffset != offset + len) { 3612 panic("offset + len changed from %jx/%jx to %jx/%jx", 3613 ooffset, olen, offset, len); 3614 } 3615 if (error != 0 || len == 0) 3616 break; 3617 KASSERT(olen > len, ("Iteration did not make progress?")); 3618 maybe_yield(); 3619 } 3620 3621 return (error); 3622 } 3623 3624 static int 3625 vn_deallocate_impl(struct vnode *vp, off_t *offset, off_t *length, int flags, 3626 int ioflag, struct ucred *cred, struct ucred *active_cred, 3627 struct ucred *file_cred) 3628 { 3629 struct mount *mp; 3630 void *rl_cookie; 3631 off_t off, len; 3632 int error; 3633 #ifdef AUDIT 3634 bool audited_vnode1 = false; 3635 #endif 3636 3637 rl_cookie = NULL; 3638 error = 0; 3639 mp = NULL; 3640 off = *offset; 3641 len = *length; 3642 3643 if ((ioflag & (IO_NODELOCKED | IO_RANGELOCKED)) == 0) 3644 rl_cookie = vn_rangelock_wlock(vp, off, off + len); 3645 while (len > 0 && error == 0) { 3646 /* 3647 * Try to deallocate the longest range in one pass. 3648 * In case a pass takes too long to be executed, it returns 3649 * partial result. The residue will be proceeded in the next 3650 * pass. 3651 */ 3652 3653 if ((ioflag & IO_NODELOCKED) == 0) { 3654 bwillwrite(); 3655 if ((error = vn_start_write(vp, &mp, 3656 V_WAIT | V_PCATCH)) != 0) 3657 goto out; 3658 vn_lock(vp, vn_lktype_write(mp, vp) | LK_RETRY); 3659 } 3660 #ifdef AUDIT 3661 if (!audited_vnode1) { 3662 AUDIT_ARG_VNODE1(vp); 3663 audited_vnode1 = true; 3664 } 3665 #endif 3666 3667 #ifdef MAC 3668 if ((ioflag & IO_NOMACCHECK) == 0) 3669 error = mac_vnode_check_write(active_cred, file_cred, 3670 vp); 3671 #endif 3672 if (error == 0) 3673 error = VOP_DEALLOCATE(vp, &off, &len, flags, ioflag, 3674 cred); 3675 3676 if ((ioflag & IO_NODELOCKED) == 0) { 3677 VOP_UNLOCK(vp); 3678 if (mp != NULL) { 3679 vn_finished_write(mp); 3680 mp = NULL; 3681 } 3682 } 3683 if (error == 0 && len != 0) 3684 maybe_yield(); 3685 } 3686 out: 3687 if (rl_cookie != NULL) 3688 vn_rangelock_unlock(vp, rl_cookie); 3689 *offset = off; 3690 *length = len; 3691 return (error); 3692 } 3693 3694 /* 3695 * This function is supposed to be used in the situations where the deallocation 3696 * is not triggered by a user request. 3697 */ 3698 int 3699 vn_deallocate(struct vnode *vp, off_t *offset, off_t *length, int flags, 3700 int ioflag, struct ucred *active_cred, struct ucred *file_cred) 3701 { 3702 struct ucred *cred; 3703 3704 if (*offset < 0 || *length <= 0 || *length > OFF_MAX - *offset || 3705 flags != 0) 3706 return (EINVAL); 3707 if (vp->v_type != VREG) 3708 return (ENODEV); 3709 3710 cred = file_cred != NOCRED ? file_cred : active_cred; 3711 return (vn_deallocate_impl(vp, offset, length, flags, ioflag, cred, 3712 active_cred, file_cred)); 3713 } 3714 3715 static int 3716 vn_fspacectl(struct file *fp, int cmd, off_t *offset, off_t *length, int flags, 3717 struct ucred *active_cred, struct thread *td) 3718 { 3719 int error; 3720 struct vnode *vp; 3721 int ioflag; 3722 3723 KASSERT(cmd == SPACECTL_DEALLOC, ("vn_fspacectl: Invalid cmd")); 3724 KASSERT((flags & ~SPACECTL_F_SUPPORTED) == 0, 3725 ("vn_fspacectl: non-zero flags")); 3726 KASSERT(*offset >= 0 && *length > 0 && *length <= OFF_MAX - *offset, 3727 ("vn_fspacectl: offset/length overflow or underflow")); 3728 vp = fp->f_vnode; 3729 3730 if (vp->v_type != VREG) 3731 return (ENODEV); 3732 3733 ioflag = get_write_ioflag(fp); 3734 3735 switch (cmd) { 3736 case SPACECTL_DEALLOC: 3737 error = vn_deallocate_impl(vp, offset, length, flags, ioflag, 3738 active_cred, active_cred, fp->f_cred); 3739 break; 3740 default: 3741 panic("vn_fspacectl: unknown cmd %d", cmd); 3742 } 3743 3744 return (error); 3745 } 3746 3747 /* 3748 * Keep this assert as long as sizeof(struct dirent) is used as the maximum 3749 * entry size. 3750 */ 3751 _Static_assert(_GENERIC_MAXDIRSIZ == sizeof(struct dirent), 3752 "'struct dirent' size must be a multiple of its alignment " 3753 "(see _GENERIC_DIRLEN())"); 3754 3755 /* 3756 * Returns successive directory entries through some caller's provided buffer. 3757 * 3758 * This function automatically refills the provided buffer with calls to 3759 * VOP_READDIR() (after MAC permission checks). 3760 * 3761 * 'td' is used for credentials and passed to uiomove(). 'dirbuf' is the 3762 * caller's buffer to fill and 'dirbuflen' its allocated size. 'dirbuf' must 3763 * be properly aligned to access 'struct dirent' structures and 'dirbuflen' 3764 * must be greater than GENERIC_MAXDIRSIZ to avoid VOP_READDIR() returning 3765 * EINVAL (the latter is not a strong guarantee (yet); but EINVAL will always 3766 * be returned if this requirement is not verified). '*dpp' points to the 3767 * current directory entry in the buffer and '*len' contains the remaining 3768 * valid bytes in 'dirbuf' after 'dpp' (including the pointed entry). 3769 * 3770 * At first call (or when restarting the read), '*len' must have been set to 0, 3771 * '*off' to 0 (or any valid start offset) and '*eofflag' to 0. There are no 3772 * more entries as soon as '*len' is 0 after a call that returned 0. Calling 3773 * again this function after such a condition is considered an error and EINVAL 3774 * will be returned. Other possible error codes are those of VOP_READDIR(), 3775 * EINTEGRITY if the returned entries do not pass coherency tests, or EINVAL 3776 * (bad call). All errors are unrecoverable, i.e., the state ('*len', '*off' 3777 * and '*eofflag') must be re-initialized before a subsequent call. On error 3778 * or at end of directory, '*dpp' is reset to NULL. 3779 * 3780 * '*len', '*off' and '*eofflag' are internal state the caller should not 3781 * tamper with except as explained above. '*off' is the next directory offset 3782 * to read from to refill the buffer. '*eofflag' is set to 0 or 1 by the last 3783 * internal call to VOP_READDIR() that returned without error, indicating 3784 * whether it reached the end of the directory, and to 2 by this function after 3785 * all entries have been read. 3786 */ 3787 int 3788 vn_dir_next_dirent(struct vnode *vp, struct thread *td, 3789 char *dirbuf, size_t dirbuflen, 3790 struct dirent **dpp, size_t *len, off_t *off, int *eofflag) 3791 { 3792 struct dirent *dp = NULL; 3793 int reclen; 3794 int error; 3795 struct uio uio; 3796 struct iovec iov; 3797 3798 ASSERT_VOP_LOCKED(vp, "vnode not locked"); 3799 VNASSERT(vp->v_type == VDIR, vp, ("vnode is not a directory")); 3800 MPASS2((uintptr_t)dirbuf < (uintptr_t)dirbuf + dirbuflen, 3801 "Address space overflow"); 3802 3803 if (__predict_false(dirbuflen < GENERIC_MAXDIRSIZ)) { 3804 /* Don't take any chances in this case */ 3805 error = EINVAL; 3806 goto out; 3807 } 3808 3809 if (*len != 0) { 3810 dp = *dpp; 3811 3812 /* 3813 * The caller continued to call us after an error (we set dp to 3814 * NULL in a previous iteration). Bail out right now. 3815 */ 3816 if (__predict_false(dp == NULL)) 3817 return (EINVAL); 3818 3819 MPASS(*len <= dirbuflen); 3820 MPASS2((uintptr_t)dirbuf <= (uintptr_t)dp && 3821 (uintptr_t)dp + *len <= (uintptr_t)dirbuf + dirbuflen, 3822 "Filled range not inside buffer"); 3823 3824 reclen = dp->d_reclen; 3825 if (reclen >= *len) { 3826 /* End of buffer reached */ 3827 *len = 0; 3828 } else { 3829 dp = (struct dirent *)((char *)dp + reclen); 3830 *len -= reclen; 3831 } 3832 } 3833 3834 if (*len == 0) { 3835 dp = NULL; 3836 3837 /* Have to refill. */ 3838 switch (*eofflag) { 3839 case 0: 3840 break; 3841 3842 case 1: 3843 /* Nothing more to read. */ 3844 *eofflag = 2; /* Remember the caller reached EOF. */ 3845 goto success; 3846 3847 default: 3848 /* The caller didn't test for EOF. */ 3849 error = EINVAL; 3850 goto out; 3851 } 3852 3853 iov.iov_base = dirbuf; 3854 iov.iov_len = dirbuflen; 3855 3856 uio.uio_iov = &iov; 3857 uio.uio_iovcnt = 1; 3858 uio.uio_offset = *off; 3859 uio.uio_resid = dirbuflen; 3860 uio.uio_segflg = UIO_SYSSPACE; 3861 uio.uio_rw = UIO_READ; 3862 uio.uio_td = td; 3863 3864 #ifdef MAC 3865 error = mac_vnode_check_readdir(td->td_ucred, vp); 3866 if (error == 0) 3867 #endif 3868 error = VOP_READDIR(vp, &uio, td->td_ucred, eofflag, 3869 NULL, NULL); 3870 if (error != 0) 3871 goto out; 3872 3873 *len = dirbuflen - uio.uio_resid; 3874 *off = uio.uio_offset; 3875 3876 if (*len == 0) { 3877 /* Sanity check on INVARIANTS. */ 3878 MPASS(*eofflag != 0); 3879 *eofflag = 1; 3880 goto success; 3881 } 3882 3883 /* 3884 * Normalize the flag returned by VOP_READDIR(), since we use 2 3885 * as a sentinel value. 3886 */ 3887 if (*eofflag != 0) 3888 *eofflag = 1; 3889 3890 dp = (struct dirent *)dirbuf; 3891 } 3892 3893 if (__predict_false(*len < GENERIC_MINDIRSIZ || 3894 dp->d_reclen < GENERIC_MINDIRSIZ)) { 3895 error = EINTEGRITY; 3896 dp = NULL; 3897 goto out; 3898 } 3899 3900 success: 3901 error = 0; 3902 out: 3903 *dpp = dp; 3904 return (error); 3905 } 3906 3907 /* 3908 * Checks whether a directory is empty or not. 3909 * 3910 * If the directory is empty, returns 0, and if it is not, ENOTEMPTY. Other 3911 * values are genuine errors preventing the check. 3912 */ 3913 int 3914 vn_dir_check_empty(struct vnode *vp) 3915 { 3916 struct thread *const td = curthread; 3917 char *dirbuf; 3918 size_t dirbuflen, len; 3919 off_t off; 3920 int eofflag, error; 3921 struct dirent *dp; 3922 struct vattr va; 3923 3924 ASSERT_VOP_LOCKED(vp, "vfs_emptydir"); 3925 VNPASS(vp->v_type == VDIR, vp); 3926 3927 error = VOP_GETATTR(vp, &va, td->td_ucred); 3928 if (error != 0) 3929 return (error); 3930 3931 dirbuflen = max(DEV_BSIZE, GENERIC_MAXDIRSIZ); 3932 if (dirbuflen < va.va_blocksize) 3933 dirbuflen = va.va_blocksize; 3934 dirbuf = malloc(dirbuflen, M_TEMP, M_WAITOK); 3935 3936 len = 0; 3937 off = 0; 3938 eofflag = 0; 3939 3940 for (;;) { 3941 error = vn_dir_next_dirent(vp, td, dirbuf, dirbuflen, 3942 &dp, &len, &off, &eofflag); 3943 if (error != 0) 3944 goto end; 3945 3946 if (len == 0) { 3947 /* EOF */ 3948 error = 0; 3949 goto end; 3950 } 3951 3952 /* 3953 * Skip whiteouts. Unionfs operates on filesystems only and 3954 * not on hierarchies, so these whiteouts would be shadowed on 3955 * the system hierarchy but not for a union using the 3956 * filesystem of their directories as the upper layer. 3957 * Additionally, unionfs currently transparently exposes 3958 * union-specific metadata of its upper layer, meaning that 3959 * whiteouts can be seen through the union view in empty 3960 * directories. Taking into account these whiteouts would then 3961 * prevent mounting another filesystem on such effectively 3962 * empty directories. 3963 */ 3964 if (dp->d_type == DT_WHT) 3965 continue; 3966 3967 /* 3968 * Any file in the directory which is not '.' or '..' indicates 3969 * the directory is not empty. 3970 */ 3971 switch (dp->d_namlen) { 3972 case 2: 3973 if (dp->d_name[1] != '.') { 3974 /* Can't be '..' (nor '.') */ 3975 error = ENOTEMPTY; 3976 goto end; 3977 } 3978 /* FALLTHROUGH */ 3979 case 1: 3980 if (dp->d_name[0] != '.') { 3981 /* Can't be '..' nor '.' */ 3982 error = ENOTEMPTY; 3983 goto end; 3984 } 3985 break; 3986 3987 default: 3988 error = ENOTEMPTY; 3989 goto end; 3990 } 3991 } 3992 3993 end: 3994 free(dirbuf, M_TEMP); 3995 return (error); 3996 } 3997 3998 3999 static u_long vn_lock_pair_pause_cnt; 4000 SYSCTL_ULONG(_debug, OID_AUTO, vn_lock_pair_pause, CTLFLAG_RD, 4001 &vn_lock_pair_pause_cnt, 0, 4002 "Count of vn_lock_pair deadlocks"); 4003 4004 u_int vn_lock_pair_pause_max; 4005 SYSCTL_UINT(_debug, OID_AUTO, vn_lock_pair_pause_max, CTLFLAG_RW, 4006 &vn_lock_pair_pause_max, 0, 4007 "Max ticks for vn_lock_pair deadlock avoidance sleep"); 4008 4009 static void 4010 vn_lock_pair_pause(const char *wmesg) 4011 { 4012 atomic_add_long(&vn_lock_pair_pause_cnt, 1); 4013 pause(wmesg, prng32_bounded(vn_lock_pair_pause_max)); 4014 } 4015 4016 /* 4017 * Lock pair of vnodes vp1, vp2, avoiding lock order reversal. 4018 * vp1_locked indicates whether vp1 is locked; if not, vp1 must be 4019 * unlocked. Same for vp2 and vp2_locked. One of the vnodes can be 4020 * NULL. 4021 * 4022 * The function returns with both vnodes exclusively or shared locked, 4023 * according to corresponding lkflags, and guarantees that it does not 4024 * create lock order reversal with other threads during its execution. 4025 * Both vnodes could be unlocked temporary (and reclaimed). 4026 * 4027 * If requesting shared locking, locked vnode lock must not be recursed. 4028 * 4029 * Only one of LK_SHARED and LK_EXCLUSIVE must be specified. 4030 * LK_NODDLKTREAT can be optionally passed. 4031 */ 4032 void 4033 vn_lock_pair(struct vnode *vp1, bool vp1_locked, int lkflags1, 4034 struct vnode *vp2, bool vp2_locked, int lkflags2) 4035 { 4036 int error; 4037 4038 MPASS(((lkflags1 & LK_SHARED) != 0) ^ ((lkflags1 & LK_EXCLUSIVE) != 0)); 4039 MPASS((lkflags1 & ~(LK_SHARED | LK_EXCLUSIVE | LK_NODDLKTREAT)) == 0); 4040 MPASS(((lkflags2 & LK_SHARED) != 0) ^ ((lkflags2 & LK_EXCLUSIVE) != 0)); 4041 MPASS((lkflags2 & ~(LK_SHARED | LK_EXCLUSIVE | LK_NODDLKTREAT)) == 0); 4042 4043 if (vp1 == NULL && vp2 == NULL) 4044 return; 4045 4046 if (vp1 != NULL) { 4047 if ((lkflags1 & LK_SHARED) != 0 && 4048 (vp1->v_vnlock->lock_object.lo_flags & LK_NOSHARE) != 0) 4049 lkflags1 = (lkflags1 & ~LK_SHARED) | LK_EXCLUSIVE; 4050 if (vp1_locked && VOP_ISLOCKED(vp1) != LK_EXCLUSIVE) { 4051 ASSERT_VOP_LOCKED(vp1, "vp1"); 4052 if ((lkflags1 & LK_EXCLUSIVE) != 0) { 4053 VOP_UNLOCK(vp1); 4054 ASSERT_VOP_UNLOCKED(vp1, 4055 "vp1 shared recursed"); 4056 vp1_locked = false; 4057 } 4058 } else if (!vp1_locked) 4059 ASSERT_VOP_UNLOCKED(vp1, "vp1"); 4060 } else { 4061 vp1_locked = true; 4062 } 4063 4064 if (vp2 != NULL) { 4065 if ((lkflags2 & LK_SHARED) != 0 && 4066 (vp2->v_vnlock->lock_object.lo_flags & LK_NOSHARE) != 0) 4067 lkflags2 = (lkflags2 & ~LK_SHARED) | LK_EXCLUSIVE; 4068 if (vp2_locked && VOP_ISLOCKED(vp2) != LK_EXCLUSIVE) { 4069 ASSERT_VOP_LOCKED(vp2, "vp2"); 4070 if ((lkflags2 & LK_EXCLUSIVE) != 0) { 4071 VOP_UNLOCK(vp2); 4072 ASSERT_VOP_UNLOCKED(vp2, 4073 "vp2 shared recursed"); 4074 vp2_locked = false; 4075 } 4076 } else if (!vp2_locked) 4077 ASSERT_VOP_UNLOCKED(vp2, "vp2"); 4078 } else { 4079 vp2_locked = true; 4080 } 4081 4082 if (!vp1_locked && !vp2_locked) { 4083 vn_lock(vp1, lkflags1 | LK_RETRY); 4084 vp1_locked = true; 4085 } 4086 4087 while (!vp1_locked || !vp2_locked) { 4088 if (vp1_locked && vp2 != NULL) { 4089 if (vp1 != NULL) { 4090 error = VOP_LOCK1(vp2, lkflags2 | LK_NOWAIT, 4091 __FILE__, __LINE__); 4092 if (error == 0) 4093 break; 4094 VOP_UNLOCK(vp1); 4095 vp1_locked = false; 4096 vn_lock_pair_pause("vlp1"); 4097 } 4098 vn_lock(vp2, lkflags2 | LK_RETRY); 4099 vp2_locked = true; 4100 } 4101 if (vp2_locked && vp1 != NULL) { 4102 if (vp2 != NULL) { 4103 error = VOP_LOCK1(vp1, lkflags1 | LK_NOWAIT, 4104 __FILE__, __LINE__); 4105 if (error == 0) 4106 break; 4107 VOP_UNLOCK(vp2); 4108 vp2_locked = false; 4109 vn_lock_pair_pause("vlp2"); 4110 } 4111 vn_lock(vp1, lkflags1 | LK_RETRY); 4112 vp1_locked = true; 4113 } 4114 } 4115 if (vp1 != NULL) { 4116 if (lkflags1 == LK_EXCLUSIVE) 4117 ASSERT_VOP_ELOCKED(vp1, "vp1 ret"); 4118 else 4119 ASSERT_VOP_LOCKED(vp1, "vp1 ret"); 4120 } 4121 if (vp2 != NULL) { 4122 if (lkflags2 == LK_EXCLUSIVE) 4123 ASSERT_VOP_ELOCKED(vp2, "vp2 ret"); 4124 else 4125 ASSERT_VOP_LOCKED(vp2, "vp2 ret"); 4126 } 4127 } 4128 4129 int 4130 vn_lktype_write(struct mount *mp, struct vnode *vp) 4131 { 4132 if (MNT_SHARED_WRITES(mp) || 4133 (mp == NULL && MNT_SHARED_WRITES(vp->v_mount))) 4134 return (LK_SHARED); 4135 return (LK_EXCLUSIVE); 4136 } 4137