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