1 /*- 2 * SPDX-License-Identifier: BSD-3-Clause 3 * 4 * Copyright (c) 1982, 1986, 1989, 1993 5 * The Regents of the University of California. All rights reserved. 6 * (c) UNIX System Laboratories, Inc. 7 * All or some portions of this file are derived from material licensed 8 * to the University of California by American Telephone and Telegraph 9 * Co. or Unix System Laboratories, Inc. and are reproduced herein with 10 * the permission of UNIX System Laboratories, Inc. 11 * 12 * Copyright (c) 2012 Konstantin Belousov <kib@FreeBSD.org> 13 * Copyright (c) 2013, 2014 The FreeBSD Foundation 14 * 15 * Portions of this software were developed by Konstantin Belousov 16 * under sponsorship from the FreeBSD Foundation. 17 * 18 * Redistribution and use in source and binary forms, with or without 19 * modification, are permitted provided that the following conditions 20 * are met: 21 * 1. Redistributions of source code must retain the above copyright 22 * notice, this list of conditions and the following disclaimer. 23 * 2. Redistributions in binary form must reproduce the above copyright 24 * notice, this list of conditions and the following disclaimer in the 25 * documentation and/or other materials provided with the distribution. 26 * 3. Neither the name of the University nor the names of its contributors 27 * may be used to endorse or promote products derived from this software 28 * without specific prior written permission. 29 * 30 * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND 31 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE 32 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE 33 * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE 34 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL 35 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS 36 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) 37 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT 38 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY 39 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF 40 * SUCH DAMAGE. 41 * 42 * @(#)vfs_vnops.c 8.2 (Berkeley) 1/21/94 43 */ 44 45 #include <sys/cdefs.h> 46 __FBSDID("$FreeBSD$"); 47 48 #include "opt_hwpmc_hooks.h" 49 50 #include <sys/param.h> 51 #include <sys/systm.h> 52 #include <sys/disk.h> 53 #include <sys/fail.h> 54 #include <sys/fcntl.h> 55 #include <sys/file.h> 56 #include <sys/kdb.h> 57 #include <sys/ktr.h> 58 #include <sys/stat.h> 59 #include <sys/priv.h> 60 #include <sys/proc.h> 61 #include <sys/limits.h> 62 #include <sys/lock.h> 63 #include <sys/mman.h> 64 #include <sys/mount.h> 65 #include <sys/mutex.h> 66 #include <sys/namei.h> 67 #include <sys/vnode.h> 68 #include <sys/bio.h> 69 #include <sys/buf.h> 70 #include <sys/filio.h> 71 #include <sys/resourcevar.h> 72 #include <sys/rwlock.h> 73 #include <sys/prng.h> 74 #include <sys/sx.h> 75 #include <sys/sleepqueue.h> 76 #include <sys/sysctl.h> 77 #include <sys/ttycom.h> 78 #include <sys/conf.h> 79 #include <sys/syslog.h> 80 #include <sys/unistd.h> 81 #include <sys/user.h> 82 #include <sys/ktrace.h> 83 84 #include <security/audit/audit.h> 85 #include <security/mac/mac_framework.h> 86 87 #include <vm/vm.h> 88 #include <vm/vm_extern.h> 89 #include <vm/pmap.h> 90 #include <vm/vm_map.h> 91 #include <vm/vm_object.h> 92 #include <vm/vm_page.h> 93 #include <vm/vm_pager.h> 94 95 #ifdef HWPMC_HOOKS 96 #include <sys/pmckern.h> 97 #endif 98 99 static fo_rdwr_t vn_read; 100 static fo_rdwr_t vn_write; 101 static fo_rdwr_t vn_io_fault; 102 static fo_truncate_t vn_truncate; 103 static fo_ioctl_t vn_ioctl; 104 static fo_poll_t vn_poll; 105 static fo_kqfilter_t vn_kqfilter; 106 static fo_close_t vn_closefile; 107 static fo_mmap_t vn_mmap; 108 static fo_fallocate_t vn_fallocate; 109 110 struct fileops vnops = { 111 .fo_read = vn_io_fault, 112 .fo_write = vn_io_fault, 113 .fo_truncate = vn_truncate, 114 .fo_ioctl = vn_ioctl, 115 .fo_poll = vn_poll, 116 .fo_kqfilter = vn_kqfilter, 117 .fo_stat = vn_statfile, 118 .fo_close = vn_closefile, 119 .fo_chmod = vn_chmod, 120 .fo_chown = vn_chown, 121 .fo_sendfile = vn_sendfile, 122 .fo_seek = vn_seek, 123 .fo_fill_kinfo = vn_fill_kinfo, 124 .fo_mmap = vn_mmap, 125 .fo_fallocate = vn_fallocate, 126 .fo_flags = DFLAG_PASSABLE | DFLAG_SEEKABLE 127 }; 128 129 const u_int io_hold_cnt = 16; 130 static int vn_io_fault_enable = 1; 131 SYSCTL_INT(_debug, OID_AUTO, vn_io_fault_enable, CTLFLAG_RWTUN, 132 &vn_io_fault_enable, 0, "Enable vn_io_fault lock avoidance"); 133 static int vn_io_fault_prefault = 0; 134 SYSCTL_INT(_debug, OID_AUTO, vn_io_fault_prefault, CTLFLAG_RWTUN, 135 &vn_io_fault_prefault, 0, "Enable vn_io_fault prefaulting"); 136 static int vn_io_pgcache_read_enable = 1; 137 SYSCTL_INT(_debug, OID_AUTO, vn_io_pgcache_read_enable, CTLFLAG_RWTUN, 138 &vn_io_pgcache_read_enable, 0, 139 "Enable copying from page cache for reads, avoiding fs"); 140 static u_long vn_io_faults_cnt; 141 SYSCTL_ULONG(_debug, OID_AUTO, vn_io_faults, CTLFLAG_RD, 142 &vn_io_faults_cnt, 0, "Count of vn_io_fault lock avoidance triggers"); 143 144 static int vfs_allow_read_dir = 0; 145 SYSCTL_INT(_security_bsd, OID_AUTO, allow_read_dir, CTLFLAG_RW, 146 &vfs_allow_read_dir, 0, 147 "Enable read(2) of directory by root for filesystems that support it"); 148 149 /* 150 * Returns true if vn_io_fault mode of handling the i/o request should 151 * be used. 152 */ 153 static bool 154 do_vn_io_fault(struct vnode *vp, struct uio *uio) 155 { 156 struct mount *mp; 157 158 return (uio->uio_segflg == UIO_USERSPACE && vp->v_type == VREG && 159 (mp = vp->v_mount) != NULL && 160 (mp->mnt_kern_flag & MNTK_NO_IOPF) != 0 && vn_io_fault_enable); 161 } 162 163 /* 164 * Structure used to pass arguments to vn_io_fault1(), to do either 165 * file- or vnode-based I/O calls. 166 */ 167 struct vn_io_fault_args { 168 enum { 169 VN_IO_FAULT_FOP, 170 VN_IO_FAULT_VOP 171 } kind; 172 struct ucred *cred; 173 int flags; 174 union { 175 struct fop_args_tag { 176 struct file *fp; 177 fo_rdwr_t *doio; 178 } fop_args; 179 struct vop_args_tag { 180 struct vnode *vp; 181 } vop_args; 182 } args; 183 }; 184 185 static int vn_io_fault1(struct vnode *vp, struct uio *uio, 186 struct vn_io_fault_args *args, struct thread *td); 187 188 int 189 vn_open(struct nameidata *ndp, int *flagp, int cmode, struct file *fp) 190 { 191 struct thread *td = ndp->ni_cnd.cn_thread; 192 193 return (vn_open_cred(ndp, flagp, cmode, 0, td->td_ucred, fp)); 194 } 195 196 static uint64_t 197 open2nameif(int fmode, u_int vn_open_flags) 198 { 199 uint64_t res; 200 201 res = ISOPEN | LOCKLEAF; 202 if ((fmode & O_RESOLVE_BENEATH) != 0) 203 res |= RBENEATH; 204 if ((fmode & O_EMPTY_PATH) != 0) 205 res |= EMPTYPATH; 206 if ((vn_open_flags & VN_OPEN_NOAUDIT) == 0) 207 res |= AUDITVNODE1; 208 if ((vn_open_flags & VN_OPEN_NOCAPCHECK) != 0) 209 res |= NOCAPCHECK; 210 return (res); 211 } 212 213 /* 214 * Common code for vnode open operations via a name lookup. 215 * Lookup the vnode and invoke VOP_CREATE if needed. 216 * Check permissions, and call the VOP_OPEN or VOP_CREATE routine. 217 * 218 * Note that this does NOT free nameidata for the successful case, 219 * due to the NDINIT being done elsewhere. 220 */ 221 int 222 vn_open_cred(struct nameidata *ndp, int *flagp, int cmode, u_int vn_open_flags, 223 struct ucred *cred, struct file *fp) 224 { 225 struct vnode *vp; 226 struct mount *mp; 227 struct thread *td = ndp->ni_cnd.cn_thread; 228 struct vattr vat; 229 struct vattr *vap = &vat; 230 int fmode, error; 231 bool first_open; 232 233 restart: 234 first_open = false; 235 fmode = *flagp; 236 if ((fmode & (O_CREAT | O_EXCL | O_DIRECTORY)) == (O_CREAT | 237 O_EXCL | O_DIRECTORY) || 238 (fmode & (O_CREAT | O_EMPTY_PATH)) == (O_CREAT | O_EMPTY_PATH)) 239 return (EINVAL); 240 else if ((fmode & (O_CREAT | O_DIRECTORY)) == O_CREAT) { 241 ndp->ni_cnd.cn_nameiop = CREATE; 242 ndp->ni_cnd.cn_flags = open2nameif(fmode, vn_open_flags); 243 /* 244 * Set NOCACHE to avoid flushing the cache when 245 * rolling in many files at once. 246 * 247 * Set NC_KEEPPOSENTRY to keep positive entries if they already 248 * exist despite NOCACHE. 249 */ 250 ndp->ni_cnd.cn_flags |= LOCKPARENT | NOCACHE | NC_KEEPPOSENTRY; 251 if ((fmode & O_EXCL) == 0 && (fmode & O_NOFOLLOW) == 0) 252 ndp->ni_cnd.cn_flags |= FOLLOW; 253 if ((vn_open_flags & VN_OPEN_INVFS) == 0) 254 bwillwrite(); 255 if ((error = namei(ndp)) != 0) 256 return (error); 257 if (ndp->ni_vp == NULL) { 258 VATTR_NULL(vap); 259 vap->va_type = VREG; 260 vap->va_mode = cmode; 261 if (fmode & O_EXCL) 262 vap->va_vaflags |= VA_EXCLUSIVE; 263 if (vn_start_write(ndp->ni_dvp, &mp, V_NOWAIT) != 0) { 264 NDFREE(ndp, NDF_ONLY_PNBUF); 265 vput(ndp->ni_dvp); 266 if ((error = vn_start_write(NULL, &mp, 267 V_XSLEEP | PCATCH)) != 0) 268 return (error); 269 NDREINIT(ndp); 270 goto restart; 271 } 272 if ((vn_open_flags & VN_OPEN_NAMECACHE) != 0) 273 ndp->ni_cnd.cn_flags |= MAKEENTRY; 274 #ifdef MAC 275 error = mac_vnode_check_create(cred, ndp->ni_dvp, 276 &ndp->ni_cnd, vap); 277 if (error == 0) 278 #endif 279 error = VOP_CREATE(ndp->ni_dvp, &ndp->ni_vp, 280 &ndp->ni_cnd, vap); 281 vp = ndp->ni_vp; 282 if (error == 0 && (fmode & O_EXCL) != 0 && 283 (fmode & (O_EXLOCK | O_SHLOCK)) != 0) { 284 VI_LOCK(vp); 285 vp->v_iflag |= VI_FOPENING; 286 VI_UNLOCK(vp); 287 first_open = true; 288 } 289 VOP_VPUT_PAIR(ndp->ni_dvp, error == 0 ? &vp : NULL, 290 false); 291 vn_finished_write(mp); 292 if (error) { 293 NDFREE(ndp, NDF_ONLY_PNBUF); 294 if (error == ERELOOKUP) { 295 NDREINIT(ndp); 296 goto restart; 297 } 298 return (error); 299 } 300 fmode &= ~O_TRUNC; 301 } else { 302 if (ndp->ni_dvp == ndp->ni_vp) 303 vrele(ndp->ni_dvp); 304 else 305 vput(ndp->ni_dvp); 306 ndp->ni_dvp = NULL; 307 vp = ndp->ni_vp; 308 if (fmode & O_EXCL) { 309 error = EEXIST; 310 goto bad; 311 } 312 if (vp->v_type == VDIR) { 313 error = EISDIR; 314 goto bad; 315 } 316 fmode &= ~O_CREAT; 317 } 318 } else { 319 ndp->ni_cnd.cn_nameiop = LOOKUP; 320 ndp->ni_cnd.cn_flags = open2nameif(fmode, vn_open_flags); 321 ndp->ni_cnd.cn_flags |= (fmode & O_NOFOLLOW) != 0 ? NOFOLLOW : 322 FOLLOW; 323 if ((fmode & FWRITE) == 0) 324 ndp->ni_cnd.cn_flags |= LOCKSHARED; 325 if ((error = namei(ndp)) != 0) 326 return (error); 327 vp = ndp->ni_vp; 328 } 329 error = vn_open_vnode(vp, fmode, cred, td, fp); 330 if (first_open) { 331 VI_LOCK(vp); 332 vp->v_iflag &= ~VI_FOPENING; 333 wakeup(vp); 334 VI_UNLOCK(vp); 335 } 336 if (error) 337 goto bad; 338 *flagp = fmode; 339 return (0); 340 bad: 341 NDFREE(ndp, NDF_ONLY_PNBUF); 342 vput(vp); 343 *flagp = fmode; 344 ndp->ni_vp = NULL; 345 return (error); 346 } 347 348 static int 349 vn_open_vnode_advlock(struct vnode *vp, int fmode, struct file *fp) 350 { 351 struct flock lf; 352 int error, lock_flags, type; 353 354 ASSERT_VOP_LOCKED(vp, "vn_open_vnode_advlock"); 355 if ((fmode & (O_EXLOCK | O_SHLOCK)) == 0) 356 return (0); 357 KASSERT(fp != NULL, ("open with flock requires fp")); 358 if (fp->f_type != DTYPE_NONE && fp->f_type != DTYPE_VNODE) 359 return (EOPNOTSUPP); 360 361 lock_flags = VOP_ISLOCKED(vp); 362 VOP_UNLOCK(vp); 363 364 lf.l_whence = SEEK_SET; 365 lf.l_start = 0; 366 lf.l_len = 0; 367 lf.l_type = (fmode & O_EXLOCK) != 0 ? F_WRLCK : F_RDLCK; 368 type = F_FLOCK; 369 if ((fmode & FNONBLOCK) == 0) 370 type |= F_WAIT; 371 if ((fmode & (O_CREAT | O_EXCL)) == (O_CREAT | O_EXCL)) 372 type |= F_FIRSTOPEN; 373 error = VOP_ADVLOCK(vp, (caddr_t)fp, F_SETLK, &lf, type); 374 if (error == 0) 375 fp->f_flag |= FHASLOCK; 376 377 vn_lock(vp, lock_flags | LK_RETRY); 378 return (error); 379 } 380 381 /* 382 * Common code for vnode open operations once a vnode is located. 383 * Check permissions, and call the VOP_OPEN routine. 384 */ 385 int 386 vn_open_vnode(struct vnode *vp, int fmode, struct ucred *cred, 387 struct thread *td, struct file *fp) 388 { 389 accmode_t accmode; 390 int error; 391 392 if (vp->v_type == VLNK) { 393 if ((fmode & O_PATH) == 0 || (fmode & FEXEC) != 0) 394 return (EMLINK); 395 } 396 if (vp->v_type == VSOCK) 397 return (EOPNOTSUPP); 398 if (vp->v_type != VDIR && fmode & O_DIRECTORY) 399 return (ENOTDIR); 400 401 accmode = 0; 402 if ((fmode & O_PATH) == 0) { 403 if ((fmode & (FWRITE | O_TRUNC)) != 0) { 404 if (vp->v_type == VDIR) 405 return (EISDIR); 406 accmode |= VWRITE; 407 } 408 if ((fmode & FREAD) != 0) 409 accmode |= VREAD; 410 if ((fmode & O_APPEND) && (fmode & FWRITE)) 411 accmode |= VAPPEND; 412 #ifdef MAC 413 if ((fmode & O_CREAT) != 0) 414 accmode |= VCREAT; 415 #endif 416 } 417 if ((fmode & FEXEC) != 0) 418 accmode |= VEXEC; 419 #ifdef MAC 420 if ((fmode & O_VERIFY) != 0) 421 accmode |= VVERIFY; 422 error = mac_vnode_check_open(cred, vp, accmode); 423 if (error != 0) 424 return (error); 425 426 accmode &= ~(VCREAT | VVERIFY); 427 #endif 428 if ((fmode & O_CREAT) == 0 && accmode != 0) { 429 error = VOP_ACCESS(vp, accmode, cred, td); 430 if (error != 0) 431 return (error); 432 } 433 if ((fmode & O_PATH) != 0) { 434 if (vp->v_type == VFIFO) 435 error = EPIPE; 436 else 437 error = VOP_ACCESS(vp, VREAD, cred, td); 438 if (error == 0) 439 fp->f_flag |= FKQALLOWED; 440 return (0); 441 } 442 443 if (vp->v_type == VFIFO && VOP_ISLOCKED(vp) != LK_EXCLUSIVE) 444 vn_lock(vp, LK_UPGRADE | LK_RETRY); 445 error = VOP_OPEN(vp, fmode, cred, td, fp); 446 if (error != 0) 447 return (error); 448 449 error = vn_open_vnode_advlock(vp, fmode, fp); 450 if (error == 0 && (fmode & FWRITE) != 0) { 451 error = VOP_ADD_WRITECOUNT(vp, 1); 452 if (error == 0) { 453 CTR3(KTR_VFS, "%s: vp %p v_writecount increased to %d", 454 __func__, vp, vp->v_writecount); 455 } 456 } 457 458 /* 459 * Error from advlock or VOP_ADD_WRITECOUNT() still requires 460 * calling VOP_CLOSE() to pair with earlier VOP_OPEN(). 461 */ 462 if (error != 0) { 463 if (fp != NULL) { 464 /* 465 * Arrange the call by having fdrop() to use 466 * vn_closefile(). This is to satisfy 467 * filesystems like devfs or tmpfs, which 468 * override fo_close(). 469 */ 470 fp->f_flag |= FOPENFAILED; 471 fp->f_vnode = vp; 472 if (fp->f_ops == &badfileops) { 473 fp->f_type = DTYPE_VNODE; 474 fp->f_ops = &vnops; 475 } 476 vref(vp); 477 } else { 478 /* 479 * If there is no fp, due to kernel-mode open, 480 * we can call VOP_CLOSE() now. 481 */ 482 if (vp->v_type != VFIFO && (fmode & FWRITE) != 0 && 483 !MNT_EXTENDED_SHARED(vp->v_mount) && 484 VOP_ISLOCKED(vp) != LK_EXCLUSIVE) 485 vn_lock(vp, LK_UPGRADE | LK_RETRY); 486 (void)VOP_CLOSE(vp, fmode & (FREAD | FWRITE | FEXEC), 487 cred, td); 488 } 489 } 490 491 ASSERT_VOP_LOCKED(vp, "vn_open_vnode"); 492 return (error); 493 494 } 495 496 /* 497 * Check for write permissions on the specified vnode. 498 * Prototype text segments cannot be written. 499 * It is racy. 500 */ 501 int 502 vn_writechk(struct vnode *vp) 503 { 504 505 ASSERT_VOP_LOCKED(vp, "vn_writechk"); 506 /* 507 * If there's shared text associated with 508 * the vnode, try to free it up once. If 509 * we fail, we can't allow writing. 510 */ 511 if (VOP_IS_TEXT(vp)) 512 return (ETXTBSY); 513 514 return (0); 515 } 516 517 /* 518 * Vnode close call 519 */ 520 static int 521 vn_close1(struct vnode *vp, int flags, struct ucred *file_cred, 522 struct thread *td, bool keep_ref) 523 { 524 struct mount *mp; 525 int error, lock_flags; 526 527 if (vp->v_type != VFIFO && (flags & FWRITE) == 0 && 528 MNT_EXTENDED_SHARED(vp->v_mount)) 529 lock_flags = LK_SHARED; 530 else 531 lock_flags = 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 | PCATCH)) 654 != 0) 655 goto out; 656 if (MNT_SHARED_WRITES(mp) || 657 ((mp == NULL) && MNT_SHARED_WRITES(vp->v_mount))) 658 lock_flags = LK_SHARED; 659 else 660 lock_flags = LK_EXCLUSIVE; 661 } else 662 lock_flags = LK_SHARED; 663 vn_lock(vp, lock_flags | LK_RETRY); 664 } else 665 rl_cookie = NULL; 666 667 ASSERT_VOP_LOCKED(vp, "IO_NODELOCKED with no vp lock held"); 668 #ifdef MAC 669 if ((ioflg & IO_NOMACCHECK) == 0) { 670 if (rw == UIO_READ) 671 error = mac_vnode_check_read(active_cred, file_cred, 672 vp); 673 else 674 error = mac_vnode_check_write(active_cred, file_cred, 675 vp); 676 } 677 #endif 678 if (error == 0) { 679 if (file_cred != NULL) 680 cred = file_cred; 681 else 682 cred = active_cred; 683 if (do_vn_io_fault(vp, &auio)) { 684 args.kind = VN_IO_FAULT_VOP; 685 args.cred = cred; 686 args.flags = ioflg; 687 args.args.vop_args.vp = vp; 688 error = vn_io_fault1(vp, &auio, &args, td); 689 } else if (rw == UIO_READ) { 690 error = VOP_READ(vp, &auio, ioflg, cred); 691 } else /* if (rw == UIO_WRITE) */ { 692 error = VOP_WRITE(vp, &auio, ioflg, cred); 693 } 694 } 695 if (aresid) 696 *aresid = auio.uio_resid; 697 else 698 if (auio.uio_resid && error == 0) 699 error = EIO; 700 if ((ioflg & IO_NODELOCKED) == 0) { 701 VOP_UNLOCK(vp); 702 if (mp != NULL) 703 vn_finished_write(mp); 704 } 705 out: 706 if (rl_cookie != NULL) 707 vn_rangelock_unlock(vp, rl_cookie); 708 return (error); 709 } 710 711 /* 712 * Package up an I/O request on a vnode into a uio and do it. The I/O 713 * request is split up into smaller chunks and we try to avoid saturating 714 * the buffer cache while potentially holding a vnode locked, so we 715 * check bwillwrite() before calling vn_rdwr(). We also call kern_yield() 716 * to give other processes a chance to lock the vnode (either other processes 717 * core'ing the same binary, or unrelated processes scanning the directory). 718 */ 719 int 720 vn_rdwr_inchunks(enum uio_rw rw, struct vnode *vp, void *base, size_t len, 721 off_t offset, enum uio_seg segflg, int ioflg, struct ucred *active_cred, 722 struct ucred *file_cred, size_t *aresid, struct thread *td) 723 { 724 int error = 0; 725 ssize_t iaresid; 726 727 do { 728 int chunk; 729 730 /* 731 * Force `offset' to a multiple of MAXBSIZE except possibly 732 * for the first chunk, so that filesystems only need to 733 * write full blocks except possibly for the first and last 734 * chunks. 735 */ 736 chunk = MAXBSIZE - (uoff_t)offset % MAXBSIZE; 737 738 if (chunk > len) 739 chunk = len; 740 if (rw != UIO_READ && vp->v_type == VREG) 741 bwillwrite(); 742 iaresid = 0; 743 error = vn_rdwr(rw, vp, base, chunk, offset, segflg, 744 ioflg, active_cred, file_cred, &iaresid, td); 745 len -= chunk; /* aresid calc already includes length */ 746 if (error) 747 break; 748 offset += chunk; 749 base = (char *)base + chunk; 750 kern_yield(PRI_USER); 751 } while (len); 752 if (aresid) 753 *aresid = len + iaresid; 754 return (error); 755 } 756 757 #if OFF_MAX <= LONG_MAX 758 off_t 759 foffset_lock(struct file *fp, int flags) 760 { 761 volatile short *flagsp; 762 off_t res; 763 short state; 764 765 KASSERT((flags & FOF_OFFSET) == 0, ("FOF_OFFSET passed")); 766 767 if ((flags & FOF_NOLOCK) != 0) 768 return (atomic_load_long(&fp->f_offset)); 769 770 /* 771 * According to McKusick the vn lock was protecting f_offset here. 772 * It is now protected by the FOFFSET_LOCKED flag. 773 */ 774 flagsp = &fp->f_vnread_flags; 775 if (atomic_cmpset_acq_16(flagsp, 0, FOFFSET_LOCKED)) 776 return (atomic_load_long(&fp->f_offset)); 777 778 sleepq_lock(&fp->f_vnread_flags); 779 state = atomic_load_16(flagsp); 780 for (;;) { 781 if ((state & FOFFSET_LOCKED) == 0) { 782 if (!atomic_fcmpset_acq_16(flagsp, &state, 783 FOFFSET_LOCKED)) 784 continue; 785 break; 786 } 787 if ((state & FOFFSET_LOCK_WAITING) == 0) { 788 if (!atomic_fcmpset_acq_16(flagsp, &state, 789 state | FOFFSET_LOCK_WAITING)) 790 continue; 791 } 792 DROP_GIANT(); 793 sleepq_add(&fp->f_vnread_flags, NULL, "vofflock", 0, 0); 794 sleepq_wait(&fp->f_vnread_flags, PUSER -1); 795 PICKUP_GIANT(); 796 sleepq_lock(&fp->f_vnread_flags); 797 state = atomic_load_16(flagsp); 798 } 799 res = atomic_load_long(&fp->f_offset); 800 sleepq_release(&fp->f_vnread_flags); 801 return (res); 802 } 803 804 void 805 foffset_unlock(struct file *fp, off_t val, int flags) 806 { 807 volatile short *flagsp; 808 short state; 809 810 KASSERT((flags & FOF_OFFSET) == 0, ("FOF_OFFSET passed")); 811 812 if ((flags & FOF_NOUPDATE) == 0) 813 atomic_store_long(&fp->f_offset, val); 814 if ((flags & FOF_NEXTOFF_R) != 0) 815 fp->f_nextoff[UIO_READ] = val; 816 if ((flags & FOF_NEXTOFF_W) != 0) 817 fp->f_nextoff[UIO_WRITE] = val; 818 819 if ((flags & FOF_NOLOCK) != 0) 820 return; 821 822 flagsp = &fp->f_vnread_flags; 823 state = atomic_load_16(flagsp); 824 if ((state & FOFFSET_LOCK_WAITING) == 0 && 825 atomic_cmpset_rel_16(flagsp, state, 0)) 826 return; 827 828 sleepq_lock(&fp->f_vnread_flags); 829 MPASS((fp->f_vnread_flags & FOFFSET_LOCKED) != 0); 830 MPASS((fp->f_vnread_flags & FOFFSET_LOCK_WAITING) != 0); 831 fp->f_vnread_flags = 0; 832 sleepq_broadcast(&fp->f_vnread_flags, SLEEPQ_SLEEP, 0, 0); 833 sleepq_release(&fp->f_vnread_flags); 834 } 835 #else 836 off_t 837 foffset_lock(struct file *fp, int flags) 838 { 839 struct mtx *mtxp; 840 off_t res; 841 842 KASSERT((flags & FOF_OFFSET) == 0, ("FOF_OFFSET passed")); 843 844 mtxp = mtx_pool_find(mtxpool_sleep, fp); 845 mtx_lock(mtxp); 846 if ((flags & FOF_NOLOCK) == 0) { 847 while (fp->f_vnread_flags & FOFFSET_LOCKED) { 848 fp->f_vnread_flags |= FOFFSET_LOCK_WAITING; 849 msleep(&fp->f_vnread_flags, mtxp, PUSER -1, 850 "vofflock", 0); 851 } 852 fp->f_vnread_flags |= FOFFSET_LOCKED; 853 } 854 res = fp->f_offset; 855 mtx_unlock(mtxp); 856 return (res); 857 } 858 859 void 860 foffset_unlock(struct file *fp, off_t val, int flags) 861 { 862 struct mtx *mtxp; 863 864 KASSERT((flags & FOF_OFFSET) == 0, ("FOF_OFFSET passed")); 865 866 mtxp = mtx_pool_find(mtxpool_sleep, fp); 867 mtx_lock(mtxp); 868 if ((flags & FOF_NOUPDATE) == 0) 869 fp->f_offset = val; 870 if ((flags & FOF_NEXTOFF_R) != 0) 871 fp->f_nextoff[UIO_READ] = val; 872 if ((flags & FOF_NEXTOFF_W) != 0) 873 fp->f_nextoff[UIO_WRITE] = val; 874 if ((flags & FOF_NOLOCK) == 0) { 875 KASSERT((fp->f_vnread_flags & FOFFSET_LOCKED) != 0, 876 ("Lost FOFFSET_LOCKED")); 877 if (fp->f_vnread_flags & FOFFSET_LOCK_WAITING) 878 wakeup(&fp->f_vnread_flags); 879 fp->f_vnread_flags = 0; 880 } 881 mtx_unlock(mtxp); 882 } 883 #endif 884 885 void 886 foffset_lock_uio(struct file *fp, struct uio *uio, int flags) 887 { 888 889 if ((flags & FOF_OFFSET) == 0) 890 uio->uio_offset = foffset_lock(fp, flags); 891 } 892 893 void 894 foffset_unlock_uio(struct file *fp, struct uio *uio, int flags) 895 { 896 897 if ((flags & FOF_OFFSET) == 0) 898 foffset_unlock(fp, uio->uio_offset, flags); 899 } 900 901 static int 902 get_advice(struct file *fp, struct uio *uio) 903 { 904 struct mtx *mtxp; 905 int ret; 906 907 ret = POSIX_FADV_NORMAL; 908 if (fp->f_advice == NULL || fp->f_vnode->v_type != VREG) 909 return (ret); 910 911 mtxp = mtx_pool_find(mtxpool_sleep, fp); 912 mtx_lock(mtxp); 913 if (fp->f_advice != NULL && 914 uio->uio_offset >= fp->f_advice->fa_start && 915 uio->uio_offset + uio->uio_resid <= fp->f_advice->fa_end) 916 ret = fp->f_advice->fa_advice; 917 mtx_unlock(mtxp); 918 return (ret); 919 } 920 921 int 922 vn_read_from_obj(struct vnode *vp, struct uio *uio) 923 { 924 vm_object_t obj; 925 vm_page_t ma[io_hold_cnt + 2]; 926 off_t off, vsz; 927 ssize_t resid; 928 int error, i, j; 929 930 MPASS(uio->uio_resid <= ptoa(io_hold_cnt + 2)); 931 obj = atomic_load_ptr(&vp->v_object); 932 if (obj == NULL) 933 return (EJUSTRETURN); 934 935 /* 936 * Depends on type stability of vm_objects. 937 */ 938 vm_object_pip_add(obj, 1); 939 if ((obj->flags & OBJ_DEAD) != 0) { 940 /* 941 * Note that object might be already reused from the 942 * vnode, and the OBJ_DEAD flag cleared. This is fine, 943 * we recheck for DOOMED vnode state after all pages 944 * are busied, and retract then. 945 * 946 * But we check for OBJ_DEAD to ensure that we do not 947 * busy pages while vm_object_terminate_pages() 948 * processes the queue. 949 */ 950 error = EJUSTRETURN; 951 goto out_pip; 952 } 953 954 resid = uio->uio_resid; 955 off = uio->uio_offset; 956 for (i = 0; resid > 0; i++) { 957 MPASS(i < io_hold_cnt + 2); 958 ma[i] = vm_page_grab_unlocked(obj, atop(off), 959 VM_ALLOC_NOCREAT | VM_ALLOC_SBUSY | VM_ALLOC_IGN_SBUSY | 960 VM_ALLOC_NOWAIT); 961 if (ma[i] == NULL) 962 break; 963 964 /* 965 * Skip invalid pages. Valid mask can be partial only 966 * at EOF, and we clip later. 967 */ 968 if (vm_page_none_valid(ma[i])) { 969 vm_page_sunbusy(ma[i]); 970 break; 971 } 972 973 resid -= PAGE_SIZE; 974 off += PAGE_SIZE; 975 } 976 if (i == 0) { 977 error = EJUSTRETURN; 978 goto out_pip; 979 } 980 981 /* 982 * Check VIRF_DOOMED after we busied our pages. Since 983 * vgonel() terminates the vnode' vm_object, it cannot 984 * process past pages busied by us. 985 */ 986 if (VN_IS_DOOMED(vp)) { 987 error = EJUSTRETURN; 988 goto out; 989 } 990 991 resid = PAGE_SIZE - (uio->uio_offset & PAGE_MASK) + ptoa(i - 1); 992 if (resid > uio->uio_resid) 993 resid = uio->uio_resid; 994 995 /* 996 * Unlocked read of vnp_size is safe because truncation cannot 997 * pass busied page. But we load vnp_size into a local 998 * variable so that possible concurrent extension does not 999 * break calculation. 1000 */ 1001 #if defined(__powerpc__) && !defined(__powerpc64__) 1002 vsz = obj->un_pager.vnp.vnp_size; 1003 #else 1004 vsz = atomic_load_64(&obj->un_pager.vnp.vnp_size); 1005 #endif 1006 if (uio->uio_offset >= vsz) { 1007 error = EJUSTRETURN; 1008 goto out; 1009 } 1010 if (uio->uio_offset + resid > vsz) 1011 resid = vsz - uio->uio_offset; 1012 1013 error = vn_io_fault_pgmove(ma, uio->uio_offset & PAGE_MASK, resid, uio); 1014 1015 out: 1016 for (j = 0; j < i; j++) { 1017 if (error == 0) 1018 vm_page_reference(ma[j]); 1019 vm_page_sunbusy(ma[j]); 1020 } 1021 out_pip: 1022 vm_object_pip_wakeup(obj); 1023 if (error != 0) 1024 return (error); 1025 return (uio->uio_resid == 0 ? 0 : EJUSTRETURN); 1026 } 1027 1028 /* 1029 * File table vnode read routine. 1030 */ 1031 static int 1032 vn_read(struct file *fp, struct uio *uio, struct ucred *active_cred, int flags, 1033 struct thread *td) 1034 { 1035 struct vnode *vp; 1036 off_t orig_offset; 1037 int error, ioflag; 1038 int advice; 1039 1040 KASSERT(uio->uio_td == td, ("uio_td %p is not td %p", 1041 uio->uio_td, td)); 1042 KASSERT(flags & FOF_OFFSET, ("No FOF_OFFSET")); 1043 vp = fp->f_vnode; 1044 ioflag = 0; 1045 if (fp->f_flag & FNONBLOCK) 1046 ioflag |= IO_NDELAY; 1047 if (fp->f_flag & O_DIRECT) 1048 ioflag |= IO_DIRECT; 1049 1050 /* 1051 * Try to read from page cache. VIRF_DOOMED check is racy but 1052 * allows us to avoid unneeded work outright. 1053 */ 1054 if (vn_io_pgcache_read_enable && !mac_vnode_check_read_enabled() && 1055 (vn_irflag_read(vp) & (VIRF_DOOMED | VIRF_PGREAD)) == VIRF_PGREAD) { 1056 error = VOP_READ_PGCACHE(vp, uio, ioflag, fp->f_cred); 1057 if (error == 0) { 1058 fp->f_nextoff[UIO_READ] = uio->uio_offset; 1059 return (0); 1060 } 1061 if (error != EJUSTRETURN) 1062 return (error); 1063 } 1064 1065 advice = get_advice(fp, uio); 1066 vn_lock(vp, LK_SHARED | LK_RETRY); 1067 1068 switch (advice) { 1069 case POSIX_FADV_NORMAL: 1070 case POSIX_FADV_SEQUENTIAL: 1071 case POSIX_FADV_NOREUSE: 1072 ioflag |= sequential_heuristic(uio, fp); 1073 break; 1074 case POSIX_FADV_RANDOM: 1075 /* Disable read-ahead for random I/O. */ 1076 break; 1077 } 1078 orig_offset = uio->uio_offset; 1079 1080 #ifdef MAC 1081 error = mac_vnode_check_read(active_cred, fp->f_cred, vp); 1082 if (error == 0) 1083 #endif 1084 error = VOP_READ(vp, uio, ioflag, fp->f_cred); 1085 fp->f_nextoff[UIO_READ] = uio->uio_offset; 1086 VOP_UNLOCK(vp); 1087 if (error == 0 && advice == POSIX_FADV_NOREUSE && 1088 orig_offset != uio->uio_offset) 1089 /* 1090 * Use POSIX_FADV_DONTNEED to flush pages and buffers 1091 * for the backing file after a POSIX_FADV_NOREUSE 1092 * read(2). 1093 */ 1094 error = VOP_ADVISE(vp, orig_offset, uio->uio_offset - 1, 1095 POSIX_FADV_DONTNEED); 1096 return (error); 1097 } 1098 1099 /* 1100 * File table vnode write routine. 1101 */ 1102 static int 1103 vn_write(struct file *fp, struct uio *uio, struct ucred *active_cred, int flags, 1104 struct thread *td) 1105 { 1106 struct vnode *vp; 1107 struct mount *mp; 1108 off_t orig_offset; 1109 int error, ioflag, lock_flags; 1110 int advice; 1111 bool need_finished_write; 1112 1113 KASSERT(uio->uio_td == td, ("uio_td %p is not td %p", 1114 uio->uio_td, td)); 1115 KASSERT(flags & FOF_OFFSET, ("No FOF_OFFSET")); 1116 vp = fp->f_vnode; 1117 if (vp->v_type == VREG) 1118 bwillwrite(); 1119 ioflag = IO_UNIT; 1120 if (vp->v_type == VREG && (fp->f_flag & O_APPEND)) 1121 ioflag |= IO_APPEND; 1122 if (fp->f_flag & FNONBLOCK) 1123 ioflag |= IO_NDELAY; 1124 if (fp->f_flag & O_DIRECT) 1125 ioflag |= IO_DIRECT; 1126 if (fp->f_flag & O_FSYNC) { 1127 mp = atomic_load_ptr(&vp->v_mount); 1128 if (mp != NULL && mp->mnt_flag & MNT_SYNCHRONOUS) 1129 ioflag |= IO_SYNC; 1130 } 1131 /* 1132 * For O_DSYNC we set both IO_SYNC and IO_DATASYNC, so that VOP_WRITE() 1133 * implementations that don't understand IO_DATASYNC fall back to full 1134 * O_SYNC behavior. 1135 */ 1136 if (fp->f_flag & O_DSYNC) 1137 ioflag |= IO_SYNC | IO_DATASYNC; 1138 mp = NULL; 1139 need_finished_write = false; 1140 if (vp->v_type != VCHR) { 1141 error = vn_start_write(vp, &mp, V_WAIT | PCATCH); 1142 if (error != 0) 1143 goto unlock; 1144 need_finished_write = true; 1145 } 1146 1147 advice = get_advice(fp, uio); 1148 1149 if (MNT_SHARED_WRITES(mp) || 1150 (mp == NULL && MNT_SHARED_WRITES(vp->v_mount))) { 1151 lock_flags = LK_SHARED; 1152 } else { 1153 lock_flags = LK_EXCLUSIVE; 1154 } 1155 1156 vn_lock(vp, lock_flags | LK_RETRY); 1157 switch (advice) { 1158 case POSIX_FADV_NORMAL: 1159 case POSIX_FADV_SEQUENTIAL: 1160 case POSIX_FADV_NOREUSE: 1161 ioflag |= sequential_heuristic(uio, fp); 1162 break; 1163 case POSIX_FADV_RANDOM: 1164 /* XXX: Is this correct? */ 1165 break; 1166 } 1167 orig_offset = uio->uio_offset; 1168 1169 #ifdef MAC 1170 error = mac_vnode_check_write(active_cred, fp->f_cred, vp); 1171 if (error == 0) 1172 #endif 1173 error = VOP_WRITE(vp, uio, ioflag, fp->f_cred); 1174 fp->f_nextoff[UIO_WRITE] = uio->uio_offset; 1175 VOP_UNLOCK(vp); 1176 if (need_finished_write) 1177 vn_finished_write(mp); 1178 if (error == 0 && advice == POSIX_FADV_NOREUSE && 1179 orig_offset != uio->uio_offset) 1180 /* 1181 * Use POSIX_FADV_DONTNEED to flush pages and buffers 1182 * for the backing file after a POSIX_FADV_NOREUSE 1183 * write(2). 1184 */ 1185 error = VOP_ADVISE(vp, orig_offset, uio->uio_offset - 1, 1186 POSIX_FADV_DONTNEED); 1187 unlock: 1188 return (error); 1189 } 1190 1191 /* 1192 * The vn_io_fault() is a wrapper around vn_read() and vn_write() to 1193 * prevent the following deadlock: 1194 * 1195 * Assume that the thread A reads from the vnode vp1 into userspace 1196 * buffer buf1 backed by the pages of vnode vp2. If a page in buf1 is 1197 * currently not resident, then system ends up with the call chain 1198 * vn_read() -> VOP_READ(vp1) -> uiomove() -> [Page Fault] -> 1199 * vm_fault(buf1) -> vnode_pager_getpages(vp2) -> VOP_GETPAGES(vp2) 1200 * which establishes lock order vp1->vn_lock, then vp2->vn_lock. 1201 * If, at the same time, thread B reads from vnode vp2 into buffer buf2 1202 * backed by the pages of vnode vp1, and some page in buf2 is not 1203 * resident, we get a reversed order vp2->vn_lock, then vp1->vn_lock. 1204 * 1205 * To prevent the lock order reversal and deadlock, vn_io_fault() does 1206 * not allow page faults to happen during VOP_READ() or VOP_WRITE(). 1207 * Instead, it first tries to do the whole range i/o with pagefaults 1208 * disabled. If all pages in the i/o buffer are resident and mapped, 1209 * VOP will succeed (ignoring the genuine filesystem errors). 1210 * Otherwise, we get back EFAULT, and vn_io_fault() falls back to do 1211 * i/o in chunks, with all pages in the chunk prefaulted and held 1212 * using vm_fault_quick_hold_pages(). 1213 * 1214 * Filesystems using this deadlock avoidance scheme should use the 1215 * array of the held pages from uio, saved in the curthread->td_ma, 1216 * instead of doing uiomove(). A helper function 1217 * vn_io_fault_uiomove() converts uiomove request into 1218 * uiomove_fromphys() over td_ma array. 1219 * 1220 * Since vnode locks do not cover the whole i/o anymore, rangelocks 1221 * make the current i/o request atomic with respect to other i/os and 1222 * truncations. 1223 */ 1224 1225 /* 1226 * Decode vn_io_fault_args and perform the corresponding i/o. 1227 */ 1228 static int 1229 vn_io_fault_doio(struct vn_io_fault_args *args, struct uio *uio, 1230 struct thread *td) 1231 { 1232 int error, save; 1233 1234 error = 0; 1235 save = vm_fault_disable_pagefaults(); 1236 switch (args->kind) { 1237 case VN_IO_FAULT_FOP: 1238 error = (args->args.fop_args.doio)(args->args.fop_args.fp, 1239 uio, args->cred, args->flags, td); 1240 break; 1241 case VN_IO_FAULT_VOP: 1242 if (uio->uio_rw == UIO_READ) { 1243 error = VOP_READ(args->args.vop_args.vp, uio, 1244 args->flags, args->cred); 1245 } else if (uio->uio_rw == UIO_WRITE) { 1246 error = VOP_WRITE(args->args.vop_args.vp, uio, 1247 args->flags, args->cred); 1248 } 1249 break; 1250 default: 1251 panic("vn_io_fault_doio: unknown kind of io %d %d", 1252 args->kind, uio->uio_rw); 1253 } 1254 vm_fault_enable_pagefaults(save); 1255 return (error); 1256 } 1257 1258 static int 1259 vn_io_fault_touch(char *base, const struct uio *uio) 1260 { 1261 int r; 1262 1263 r = fubyte(base); 1264 if (r == -1 || (uio->uio_rw == UIO_READ && subyte(base, r) == -1)) 1265 return (EFAULT); 1266 return (0); 1267 } 1268 1269 static int 1270 vn_io_fault_prefault_user(const struct uio *uio) 1271 { 1272 char *base; 1273 const struct iovec *iov; 1274 size_t len; 1275 ssize_t resid; 1276 int error, i; 1277 1278 KASSERT(uio->uio_segflg == UIO_USERSPACE, 1279 ("vn_io_fault_prefault userspace")); 1280 1281 error = i = 0; 1282 iov = uio->uio_iov; 1283 resid = uio->uio_resid; 1284 base = iov->iov_base; 1285 len = iov->iov_len; 1286 while (resid > 0) { 1287 error = vn_io_fault_touch(base, uio); 1288 if (error != 0) 1289 break; 1290 if (len < PAGE_SIZE) { 1291 if (len != 0) { 1292 error = vn_io_fault_touch(base + len - 1, uio); 1293 if (error != 0) 1294 break; 1295 resid -= len; 1296 } 1297 if (++i >= uio->uio_iovcnt) 1298 break; 1299 iov = uio->uio_iov + i; 1300 base = iov->iov_base; 1301 len = iov->iov_len; 1302 } else { 1303 len -= PAGE_SIZE; 1304 base += PAGE_SIZE; 1305 resid -= PAGE_SIZE; 1306 } 1307 } 1308 return (error); 1309 } 1310 1311 /* 1312 * Common code for vn_io_fault(), agnostic to the kind of i/o request. 1313 * Uses vn_io_fault_doio() to make the call to an actual i/o function. 1314 * Used from vn_rdwr() and vn_io_fault(), which encode the i/o request 1315 * into args and call vn_io_fault1() to handle faults during the user 1316 * mode buffer accesses. 1317 */ 1318 static int 1319 vn_io_fault1(struct vnode *vp, struct uio *uio, struct vn_io_fault_args *args, 1320 struct thread *td) 1321 { 1322 vm_page_t ma[io_hold_cnt + 2]; 1323 struct uio *uio_clone, short_uio; 1324 struct iovec short_iovec[1]; 1325 vm_page_t *prev_td_ma; 1326 vm_prot_t prot; 1327 vm_offset_t addr, end; 1328 size_t len, resid; 1329 ssize_t adv; 1330 int error, cnt, saveheld, prev_td_ma_cnt; 1331 1332 if (vn_io_fault_prefault) { 1333 error = vn_io_fault_prefault_user(uio); 1334 if (error != 0) 1335 return (error); /* Or ignore ? */ 1336 } 1337 1338 prot = uio->uio_rw == UIO_READ ? VM_PROT_WRITE : VM_PROT_READ; 1339 1340 /* 1341 * The UFS follows IO_UNIT directive and replays back both 1342 * uio_offset and uio_resid if an error is encountered during the 1343 * operation. But, since the iovec may be already advanced, 1344 * uio is still in an inconsistent state. 1345 * 1346 * Cache a copy of the original uio, which is advanced to the redo 1347 * point using UIO_NOCOPY below. 1348 */ 1349 uio_clone = cloneuio(uio); 1350 resid = uio->uio_resid; 1351 1352 short_uio.uio_segflg = UIO_USERSPACE; 1353 short_uio.uio_rw = uio->uio_rw; 1354 short_uio.uio_td = uio->uio_td; 1355 1356 error = vn_io_fault_doio(args, uio, td); 1357 if (error != EFAULT) 1358 goto out; 1359 1360 atomic_add_long(&vn_io_faults_cnt, 1); 1361 uio_clone->uio_segflg = UIO_NOCOPY; 1362 uiomove(NULL, resid - uio->uio_resid, uio_clone); 1363 uio_clone->uio_segflg = uio->uio_segflg; 1364 1365 saveheld = curthread_pflags_set(TDP_UIOHELD); 1366 prev_td_ma = td->td_ma; 1367 prev_td_ma_cnt = td->td_ma_cnt; 1368 1369 while (uio_clone->uio_resid != 0) { 1370 len = uio_clone->uio_iov->iov_len; 1371 if (len == 0) { 1372 KASSERT(uio_clone->uio_iovcnt >= 1, 1373 ("iovcnt underflow")); 1374 uio_clone->uio_iov++; 1375 uio_clone->uio_iovcnt--; 1376 continue; 1377 } 1378 if (len > ptoa(io_hold_cnt)) 1379 len = ptoa(io_hold_cnt); 1380 addr = (uintptr_t)uio_clone->uio_iov->iov_base; 1381 end = round_page(addr + len); 1382 if (end < addr) { 1383 error = EFAULT; 1384 break; 1385 } 1386 cnt = atop(end - trunc_page(addr)); 1387 /* 1388 * A perfectly misaligned address and length could cause 1389 * both the start and the end of the chunk to use partial 1390 * page. +2 accounts for such a situation. 1391 */ 1392 cnt = vm_fault_quick_hold_pages(&td->td_proc->p_vmspace->vm_map, 1393 addr, len, prot, ma, io_hold_cnt + 2); 1394 if (cnt == -1) { 1395 error = EFAULT; 1396 break; 1397 } 1398 short_uio.uio_iov = &short_iovec[0]; 1399 short_iovec[0].iov_base = (void *)addr; 1400 short_uio.uio_iovcnt = 1; 1401 short_uio.uio_resid = short_iovec[0].iov_len = len; 1402 short_uio.uio_offset = uio_clone->uio_offset; 1403 td->td_ma = ma; 1404 td->td_ma_cnt = cnt; 1405 1406 error = vn_io_fault_doio(args, &short_uio, td); 1407 vm_page_unhold_pages(ma, cnt); 1408 adv = len - short_uio.uio_resid; 1409 1410 uio_clone->uio_iov->iov_base = 1411 (char *)uio_clone->uio_iov->iov_base + adv; 1412 uio_clone->uio_iov->iov_len -= adv; 1413 uio_clone->uio_resid -= adv; 1414 uio_clone->uio_offset += adv; 1415 1416 uio->uio_resid -= adv; 1417 uio->uio_offset += adv; 1418 1419 if (error != 0 || adv == 0) 1420 break; 1421 } 1422 td->td_ma = prev_td_ma; 1423 td->td_ma_cnt = prev_td_ma_cnt; 1424 curthread_pflags_restore(saveheld); 1425 out: 1426 free(uio_clone, M_IOV); 1427 return (error); 1428 } 1429 1430 static int 1431 vn_io_fault(struct file *fp, struct uio *uio, struct ucred *active_cred, 1432 int flags, struct thread *td) 1433 { 1434 fo_rdwr_t *doio; 1435 struct vnode *vp; 1436 void *rl_cookie; 1437 struct vn_io_fault_args args; 1438 int error; 1439 1440 doio = uio->uio_rw == UIO_READ ? vn_read : vn_write; 1441 vp = fp->f_vnode; 1442 1443 /* 1444 * The ability to read(2) on a directory has historically been 1445 * allowed for all users, but this can and has been the source of 1446 * at least one security issue in the past. As such, it is now hidden 1447 * away behind a sysctl for those that actually need it to use it, and 1448 * restricted to root when it's turned on to make it relatively safe to 1449 * leave on for longer sessions of need. 1450 */ 1451 if (vp->v_type == VDIR) { 1452 KASSERT(uio->uio_rw == UIO_READ, 1453 ("illegal write attempted on a directory")); 1454 if (!vfs_allow_read_dir) 1455 return (EISDIR); 1456 if ((error = priv_check(td, PRIV_VFS_READ_DIR)) != 0) 1457 return (EISDIR); 1458 } 1459 1460 foffset_lock_uio(fp, uio, flags); 1461 if (do_vn_io_fault(vp, uio)) { 1462 args.kind = VN_IO_FAULT_FOP; 1463 args.args.fop_args.fp = fp; 1464 args.args.fop_args.doio = doio; 1465 args.cred = active_cred; 1466 args.flags = flags | FOF_OFFSET; 1467 if (uio->uio_rw == UIO_READ) { 1468 rl_cookie = vn_rangelock_rlock(vp, uio->uio_offset, 1469 uio->uio_offset + uio->uio_resid); 1470 } else if ((fp->f_flag & O_APPEND) != 0 || 1471 (flags & FOF_OFFSET) == 0) { 1472 /* For appenders, punt and lock the whole range. */ 1473 rl_cookie = vn_rangelock_wlock(vp, 0, OFF_MAX); 1474 } else { 1475 rl_cookie = vn_rangelock_wlock(vp, uio->uio_offset, 1476 uio->uio_offset + uio->uio_resid); 1477 } 1478 error = vn_io_fault1(vp, uio, &args, td); 1479 vn_rangelock_unlock(vp, rl_cookie); 1480 } else { 1481 error = doio(fp, uio, active_cred, flags | FOF_OFFSET, td); 1482 } 1483 foffset_unlock_uio(fp, uio, flags); 1484 return (error); 1485 } 1486 1487 /* 1488 * Helper function to perform the requested uiomove operation using 1489 * the held pages for io->uio_iov[0].iov_base buffer instead of 1490 * copyin/copyout. Access to the pages with uiomove_fromphys() 1491 * instead of iov_base prevents page faults that could occur due to 1492 * pmap_collect() invalidating the mapping created by 1493 * vm_fault_quick_hold_pages(), or pageout daemon, page laundry or 1494 * object cleanup revoking the write access from page mappings. 1495 * 1496 * Filesystems specified MNTK_NO_IOPF shall use vn_io_fault_uiomove() 1497 * instead of plain uiomove(). 1498 */ 1499 int 1500 vn_io_fault_uiomove(char *data, int xfersize, struct uio *uio) 1501 { 1502 struct uio transp_uio; 1503 struct iovec transp_iov[1]; 1504 struct thread *td; 1505 size_t adv; 1506 int error, pgadv; 1507 1508 td = curthread; 1509 if ((td->td_pflags & TDP_UIOHELD) == 0 || 1510 uio->uio_segflg != UIO_USERSPACE) 1511 return (uiomove(data, xfersize, uio)); 1512 1513 KASSERT(uio->uio_iovcnt == 1, ("uio_iovcnt %d", uio->uio_iovcnt)); 1514 transp_iov[0].iov_base = data; 1515 transp_uio.uio_iov = &transp_iov[0]; 1516 transp_uio.uio_iovcnt = 1; 1517 if (xfersize > uio->uio_resid) 1518 xfersize = uio->uio_resid; 1519 transp_uio.uio_resid = transp_iov[0].iov_len = xfersize; 1520 transp_uio.uio_offset = 0; 1521 transp_uio.uio_segflg = UIO_SYSSPACE; 1522 /* 1523 * Since transp_iov points to data, and td_ma page array 1524 * corresponds to original uio->uio_iov, we need to invert the 1525 * direction of the i/o operation as passed to 1526 * uiomove_fromphys(). 1527 */ 1528 switch (uio->uio_rw) { 1529 case UIO_WRITE: 1530 transp_uio.uio_rw = UIO_READ; 1531 break; 1532 case UIO_READ: 1533 transp_uio.uio_rw = UIO_WRITE; 1534 break; 1535 } 1536 transp_uio.uio_td = uio->uio_td; 1537 error = uiomove_fromphys(td->td_ma, 1538 ((vm_offset_t)uio->uio_iov->iov_base) & PAGE_MASK, 1539 xfersize, &transp_uio); 1540 adv = xfersize - transp_uio.uio_resid; 1541 pgadv = 1542 (((vm_offset_t)uio->uio_iov->iov_base + adv) >> PAGE_SHIFT) - 1543 (((vm_offset_t)uio->uio_iov->iov_base) >> PAGE_SHIFT); 1544 td->td_ma += pgadv; 1545 KASSERT(td->td_ma_cnt >= pgadv, ("consumed pages %d %d", td->td_ma_cnt, 1546 pgadv)); 1547 td->td_ma_cnt -= pgadv; 1548 uio->uio_iov->iov_base = (char *)uio->uio_iov->iov_base + adv; 1549 uio->uio_iov->iov_len -= adv; 1550 uio->uio_resid -= adv; 1551 uio->uio_offset += adv; 1552 return (error); 1553 } 1554 1555 int 1556 vn_io_fault_pgmove(vm_page_t ma[], vm_offset_t offset, int xfersize, 1557 struct uio *uio) 1558 { 1559 struct thread *td; 1560 vm_offset_t iov_base; 1561 int cnt, pgadv; 1562 1563 td = curthread; 1564 if ((td->td_pflags & TDP_UIOHELD) == 0 || 1565 uio->uio_segflg != UIO_USERSPACE) 1566 return (uiomove_fromphys(ma, offset, xfersize, uio)); 1567 1568 KASSERT(uio->uio_iovcnt == 1, ("uio_iovcnt %d", uio->uio_iovcnt)); 1569 cnt = xfersize > uio->uio_resid ? uio->uio_resid : xfersize; 1570 iov_base = (vm_offset_t)uio->uio_iov->iov_base; 1571 switch (uio->uio_rw) { 1572 case UIO_WRITE: 1573 pmap_copy_pages(td->td_ma, iov_base & PAGE_MASK, ma, 1574 offset, cnt); 1575 break; 1576 case UIO_READ: 1577 pmap_copy_pages(ma, offset, td->td_ma, iov_base & PAGE_MASK, 1578 cnt); 1579 break; 1580 } 1581 pgadv = ((iov_base + cnt) >> PAGE_SHIFT) - (iov_base >> PAGE_SHIFT); 1582 td->td_ma += pgadv; 1583 KASSERT(td->td_ma_cnt >= pgadv, ("consumed pages %d %d", td->td_ma_cnt, 1584 pgadv)); 1585 td->td_ma_cnt -= pgadv; 1586 uio->uio_iov->iov_base = (char *)(iov_base + cnt); 1587 uio->uio_iov->iov_len -= cnt; 1588 uio->uio_resid -= cnt; 1589 uio->uio_offset += cnt; 1590 return (0); 1591 } 1592 1593 /* 1594 * File table truncate routine. 1595 */ 1596 static int 1597 vn_truncate(struct file *fp, off_t length, struct ucred *active_cred, 1598 struct thread *td) 1599 { 1600 struct mount *mp; 1601 struct vnode *vp; 1602 void *rl_cookie; 1603 int error; 1604 1605 vp = fp->f_vnode; 1606 1607 retry: 1608 /* 1609 * Lock the whole range for truncation. Otherwise split i/o 1610 * might happen partly before and partly after the truncation. 1611 */ 1612 rl_cookie = vn_rangelock_wlock(vp, 0, OFF_MAX); 1613 error = vn_start_write(vp, &mp, V_WAIT | PCATCH); 1614 if (error) 1615 goto out1; 1616 vn_lock(vp, LK_EXCLUSIVE | LK_RETRY); 1617 AUDIT_ARG_VNODE1(vp); 1618 if (vp->v_type == VDIR) { 1619 error = EISDIR; 1620 goto out; 1621 } 1622 #ifdef MAC 1623 error = mac_vnode_check_write(active_cred, fp->f_cred, vp); 1624 if (error) 1625 goto out; 1626 #endif 1627 error = vn_truncate_locked(vp, length, (fp->f_flag & O_FSYNC) != 0, 1628 fp->f_cred); 1629 out: 1630 VOP_UNLOCK(vp); 1631 vn_finished_write(mp); 1632 out1: 1633 vn_rangelock_unlock(vp, rl_cookie); 1634 if (error == ERELOOKUP) 1635 goto retry; 1636 return (error); 1637 } 1638 1639 /* 1640 * Truncate a file that is already locked. 1641 */ 1642 int 1643 vn_truncate_locked(struct vnode *vp, off_t length, bool sync, 1644 struct ucred *cred) 1645 { 1646 struct vattr vattr; 1647 int error; 1648 1649 error = VOP_ADD_WRITECOUNT(vp, 1); 1650 if (error == 0) { 1651 VATTR_NULL(&vattr); 1652 vattr.va_size = length; 1653 if (sync) 1654 vattr.va_vaflags |= VA_SYNC; 1655 error = VOP_SETATTR(vp, &vattr, cred); 1656 VOP_ADD_WRITECOUNT_CHECKED(vp, -1); 1657 } 1658 return (error); 1659 } 1660 1661 /* 1662 * File table vnode stat routine. 1663 */ 1664 int 1665 vn_statfile(struct file *fp, struct stat *sb, struct ucred *active_cred, 1666 struct thread *td) 1667 { 1668 struct vnode *vp = fp->f_vnode; 1669 int error; 1670 1671 vn_lock(vp, LK_SHARED | LK_RETRY); 1672 error = VOP_STAT(vp, sb, active_cred, fp->f_cred, td); 1673 VOP_UNLOCK(vp); 1674 1675 return (error); 1676 } 1677 1678 /* 1679 * File table vnode ioctl routine. 1680 */ 1681 static int 1682 vn_ioctl(struct file *fp, u_long com, void *data, struct ucred *active_cred, 1683 struct thread *td) 1684 { 1685 struct vattr vattr; 1686 struct vnode *vp; 1687 struct fiobmap2_arg *bmarg; 1688 int error; 1689 1690 vp = fp->f_vnode; 1691 switch (vp->v_type) { 1692 case VDIR: 1693 case VREG: 1694 switch (com) { 1695 case FIONREAD: 1696 vn_lock(vp, LK_SHARED | LK_RETRY); 1697 error = VOP_GETATTR(vp, &vattr, active_cred); 1698 VOP_UNLOCK(vp); 1699 if (error == 0) 1700 *(int *)data = vattr.va_size - fp->f_offset; 1701 return (error); 1702 case FIOBMAP2: 1703 bmarg = (struct fiobmap2_arg *)data; 1704 vn_lock(vp, LK_SHARED | LK_RETRY); 1705 #ifdef MAC 1706 error = mac_vnode_check_read(active_cred, fp->f_cred, 1707 vp); 1708 if (error == 0) 1709 #endif 1710 error = VOP_BMAP(vp, bmarg->bn, NULL, 1711 &bmarg->bn, &bmarg->runp, &bmarg->runb); 1712 VOP_UNLOCK(vp); 1713 return (error); 1714 case FIONBIO: 1715 case FIOASYNC: 1716 return (0); 1717 default: 1718 return (VOP_IOCTL(vp, com, data, fp->f_flag, 1719 active_cred, td)); 1720 } 1721 break; 1722 case VCHR: 1723 return (VOP_IOCTL(vp, com, data, fp->f_flag, 1724 active_cred, td)); 1725 default: 1726 return (ENOTTY); 1727 } 1728 } 1729 1730 /* 1731 * File table vnode poll routine. 1732 */ 1733 static int 1734 vn_poll(struct file *fp, int events, struct ucred *active_cred, 1735 struct thread *td) 1736 { 1737 struct vnode *vp; 1738 int error; 1739 1740 vp = fp->f_vnode; 1741 #if defined(MAC) || defined(AUDIT) 1742 if (AUDITING_TD(td) || mac_vnode_check_poll_enabled()) { 1743 vn_lock(vp, LK_EXCLUSIVE | LK_RETRY); 1744 AUDIT_ARG_VNODE1(vp); 1745 error = mac_vnode_check_poll(active_cred, fp->f_cred, vp); 1746 VOP_UNLOCK(vp); 1747 if (error != 0) 1748 return (error); 1749 } 1750 #endif 1751 error = VOP_POLL(vp, events, fp->f_cred, td); 1752 return (error); 1753 } 1754 1755 /* 1756 * Acquire the requested lock and then check for validity. LK_RETRY 1757 * permits vn_lock to return doomed vnodes. 1758 */ 1759 static int __noinline 1760 _vn_lock_fallback(struct vnode *vp, int flags, const char *file, int line, 1761 int error) 1762 { 1763 1764 KASSERT((flags & LK_RETRY) == 0 || error == 0, 1765 ("vn_lock: error %d incompatible with flags %#x", error, flags)); 1766 1767 if (error == 0) 1768 VNASSERT(VN_IS_DOOMED(vp), vp, ("vnode not doomed")); 1769 1770 if ((flags & LK_RETRY) == 0) { 1771 if (error == 0) { 1772 VOP_UNLOCK(vp); 1773 error = ENOENT; 1774 } 1775 return (error); 1776 } 1777 1778 /* 1779 * LK_RETRY case. 1780 * 1781 * Nothing to do if we got the lock. 1782 */ 1783 if (error == 0) 1784 return (0); 1785 1786 /* 1787 * Interlock was dropped by the call in _vn_lock. 1788 */ 1789 flags &= ~LK_INTERLOCK; 1790 do { 1791 error = VOP_LOCK1(vp, flags, file, line); 1792 } while (error != 0); 1793 return (0); 1794 } 1795 1796 int 1797 _vn_lock(struct vnode *vp, int flags, const char *file, int line) 1798 { 1799 int error; 1800 1801 VNASSERT((flags & LK_TYPE_MASK) != 0, vp, 1802 ("vn_lock: no locktype (%d passed)", flags)); 1803 VNPASS(vp->v_holdcnt > 0, vp); 1804 error = VOP_LOCK1(vp, flags, file, line); 1805 if (__predict_false(error != 0 || VN_IS_DOOMED(vp))) 1806 return (_vn_lock_fallback(vp, flags, file, line, error)); 1807 return (0); 1808 } 1809 1810 /* 1811 * File table vnode close routine. 1812 */ 1813 static int 1814 vn_closefile(struct file *fp, struct thread *td) 1815 { 1816 struct vnode *vp; 1817 struct flock lf; 1818 int error; 1819 bool ref; 1820 1821 vp = fp->f_vnode; 1822 fp->f_ops = &badfileops; 1823 ref = (fp->f_flag & FHASLOCK) != 0 && fp->f_type == DTYPE_VNODE; 1824 1825 error = vn_close1(vp, fp->f_flag, fp->f_cred, td, ref); 1826 1827 if (__predict_false(ref)) { 1828 lf.l_whence = SEEK_SET; 1829 lf.l_start = 0; 1830 lf.l_len = 0; 1831 lf.l_type = F_UNLCK; 1832 (void) VOP_ADVLOCK(vp, fp, F_UNLCK, &lf, F_FLOCK); 1833 vrele(vp); 1834 } 1835 return (error); 1836 } 1837 1838 /* 1839 * Preparing to start a filesystem write operation. If the operation is 1840 * permitted, then we bump the count of operations in progress and 1841 * proceed. If a suspend request is in progress, we wait until the 1842 * suspension is over, and then proceed. 1843 */ 1844 static int 1845 vn_start_write_refed(struct mount *mp, int flags, bool mplocked) 1846 { 1847 struct mount_pcpu *mpcpu; 1848 int error, mflags; 1849 1850 if (__predict_true(!mplocked) && (flags & V_XSLEEP) == 0 && 1851 vfs_op_thread_enter(mp, mpcpu)) { 1852 MPASS((mp->mnt_kern_flag & MNTK_SUSPEND) == 0); 1853 vfs_mp_count_add_pcpu(mpcpu, writeopcount, 1); 1854 vfs_op_thread_exit(mp, mpcpu); 1855 return (0); 1856 } 1857 1858 if (mplocked) 1859 mtx_assert(MNT_MTX(mp), MA_OWNED); 1860 else 1861 MNT_ILOCK(mp); 1862 1863 error = 0; 1864 1865 /* 1866 * Check on status of suspension. 1867 */ 1868 if ((curthread->td_pflags & TDP_IGNSUSP) == 0 || 1869 mp->mnt_susp_owner != curthread) { 1870 mflags = ((mp->mnt_vfc->vfc_flags & VFCF_SBDRY) != 0 ? 1871 (flags & PCATCH) : 0) | (PUSER - 1); 1872 while ((mp->mnt_kern_flag & MNTK_SUSPEND) != 0) { 1873 if (flags & V_NOWAIT) { 1874 error = EWOULDBLOCK; 1875 goto unlock; 1876 } 1877 error = msleep(&mp->mnt_flag, MNT_MTX(mp), mflags, 1878 "suspfs", 0); 1879 if (error) 1880 goto unlock; 1881 } 1882 } 1883 if (flags & V_XSLEEP) 1884 goto unlock; 1885 mp->mnt_writeopcount++; 1886 unlock: 1887 if (error != 0 || (flags & V_XSLEEP) != 0) 1888 MNT_REL(mp); 1889 MNT_IUNLOCK(mp); 1890 return (error); 1891 } 1892 1893 int 1894 vn_start_write(struct vnode *vp, struct mount **mpp, int flags) 1895 { 1896 struct mount *mp; 1897 int error; 1898 1899 KASSERT((flags & V_MNTREF) == 0 || (*mpp != NULL && vp == NULL), 1900 ("V_MNTREF requires mp")); 1901 1902 error = 0; 1903 /* 1904 * If a vnode is provided, get and return the mount point that 1905 * to which it will write. 1906 */ 1907 if (vp != NULL) { 1908 if ((error = VOP_GETWRITEMOUNT(vp, mpp)) != 0) { 1909 *mpp = NULL; 1910 if (error != EOPNOTSUPP) 1911 return (error); 1912 return (0); 1913 } 1914 } 1915 if ((mp = *mpp) == NULL) 1916 return (0); 1917 1918 /* 1919 * VOP_GETWRITEMOUNT() returns with the mp refcount held through 1920 * a vfs_ref(). 1921 * As long as a vnode is not provided we need to acquire a 1922 * refcount for the provided mountpoint too, in order to 1923 * emulate a vfs_ref(). 1924 */ 1925 if (vp == NULL && (flags & V_MNTREF) == 0) 1926 vfs_ref(mp); 1927 1928 return (vn_start_write_refed(mp, flags, false)); 1929 } 1930 1931 /* 1932 * Secondary suspension. Used by operations such as vop_inactive 1933 * routines that are needed by the higher level functions. These 1934 * are allowed to proceed until all the higher level functions have 1935 * completed (indicated by mnt_writeopcount dropping to zero). At that 1936 * time, these operations are halted until the suspension is over. 1937 */ 1938 int 1939 vn_start_secondary_write(struct vnode *vp, struct mount **mpp, int flags) 1940 { 1941 struct mount *mp; 1942 int error; 1943 1944 KASSERT((flags & V_MNTREF) == 0 || (*mpp != NULL && vp == NULL), 1945 ("V_MNTREF requires mp")); 1946 1947 retry: 1948 if (vp != NULL) { 1949 if ((error = VOP_GETWRITEMOUNT(vp, mpp)) != 0) { 1950 *mpp = NULL; 1951 if (error != EOPNOTSUPP) 1952 return (error); 1953 return (0); 1954 } 1955 } 1956 /* 1957 * If we are not suspended or have not yet reached suspended 1958 * mode, then let the operation proceed. 1959 */ 1960 if ((mp = *mpp) == NULL) 1961 return (0); 1962 1963 /* 1964 * VOP_GETWRITEMOUNT() returns with the mp refcount held through 1965 * a vfs_ref(). 1966 * As long as a vnode is not provided we need to acquire a 1967 * refcount for the provided mountpoint too, in order to 1968 * emulate a vfs_ref(). 1969 */ 1970 MNT_ILOCK(mp); 1971 if (vp == NULL && (flags & V_MNTREF) == 0) 1972 MNT_REF(mp); 1973 if ((mp->mnt_kern_flag & (MNTK_SUSPENDED | MNTK_SUSPEND2)) == 0) { 1974 mp->mnt_secondary_writes++; 1975 mp->mnt_secondary_accwrites++; 1976 MNT_IUNLOCK(mp); 1977 return (0); 1978 } 1979 if (flags & V_NOWAIT) { 1980 MNT_REL(mp); 1981 MNT_IUNLOCK(mp); 1982 return (EWOULDBLOCK); 1983 } 1984 /* 1985 * Wait for the suspension to finish. 1986 */ 1987 error = msleep(&mp->mnt_flag, MNT_MTX(mp), (PUSER - 1) | PDROP | 1988 ((mp->mnt_vfc->vfc_flags & VFCF_SBDRY) != 0 ? (flags & PCATCH) : 0), 1989 "suspfs", 0); 1990 vfs_rel(mp); 1991 if (error == 0) 1992 goto retry; 1993 return (error); 1994 } 1995 1996 /* 1997 * Filesystem write operation has completed. If we are suspending and this 1998 * operation is the last one, notify the suspender that the suspension is 1999 * now in effect. 2000 */ 2001 void 2002 vn_finished_write(struct mount *mp) 2003 { 2004 struct mount_pcpu *mpcpu; 2005 int c; 2006 2007 if (mp == NULL) 2008 return; 2009 2010 if (vfs_op_thread_enter(mp, mpcpu)) { 2011 vfs_mp_count_sub_pcpu(mpcpu, writeopcount, 1); 2012 vfs_mp_count_sub_pcpu(mpcpu, ref, 1); 2013 vfs_op_thread_exit(mp, mpcpu); 2014 return; 2015 } 2016 2017 MNT_ILOCK(mp); 2018 vfs_assert_mount_counters(mp); 2019 MNT_REL(mp); 2020 c = --mp->mnt_writeopcount; 2021 if (mp->mnt_vfs_ops == 0) { 2022 MPASS((mp->mnt_kern_flag & MNTK_SUSPEND) == 0); 2023 MNT_IUNLOCK(mp); 2024 return; 2025 } 2026 if (c < 0) 2027 vfs_dump_mount_counters(mp); 2028 if ((mp->mnt_kern_flag & MNTK_SUSPEND) != 0 && c == 0) 2029 wakeup(&mp->mnt_writeopcount); 2030 MNT_IUNLOCK(mp); 2031 } 2032 2033 /* 2034 * Filesystem secondary write operation has completed. If we are 2035 * suspending and this operation is the last one, notify the suspender 2036 * that the suspension is now in effect. 2037 */ 2038 void 2039 vn_finished_secondary_write(struct mount *mp) 2040 { 2041 if (mp == NULL) 2042 return; 2043 MNT_ILOCK(mp); 2044 MNT_REL(mp); 2045 mp->mnt_secondary_writes--; 2046 if (mp->mnt_secondary_writes < 0) 2047 panic("vn_finished_secondary_write: neg cnt"); 2048 if ((mp->mnt_kern_flag & MNTK_SUSPEND) != 0 && 2049 mp->mnt_secondary_writes <= 0) 2050 wakeup(&mp->mnt_secondary_writes); 2051 MNT_IUNLOCK(mp); 2052 } 2053 2054 /* 2055 * Request a filesystem to suspend write operations. 2056 */ 2057 int 2058 vfs_write_suspend(struct mount *mp, int flags) 2059 { 2060 int error; 2061 2062 vfs_op_enter(mp); 2063 2064 MNT_ILOCK(mp); 2065 vfs_assert_mount_counters(mp); 2066 if (mp->mnt_susp_owner == curthread) { 2067 vfs_op_exit_locked(mp); 2068 MNT_IUNLOCK(mp); 2069 return (EALREADY); 2070 } 2071 while (mp->mnt_kern_flag & MNTK_SUSPEND) 2072 msleep(&mp->mnt_flag, MNT_MTX(mp), PUSER - 1, "wsuspfs", 0); 2073 2074 /* 2075 * Unmount holds a write reference on the mount point. If we 2076 * own busy reference and drain for writers, we deadlock with 2077 * the reference draining in the unmount path. Callers of 2078 * vfs_write_suspend() must specify VS_SKIP_UNMOUNT if 2079 * vfs_busy() reference is owned and caller is not in the 2080 * unmount context. 2081 */ 2082 if ((flags & VS_SKIP_UNMOUNT) != 0 && 2083 (mp->mnt_kern_flag & MNTK_UNMOUNT) != 0) { 2084 vfs_op_exit_locked(mp); 2085 MNT_IUNLOCK(mp); 2086 return (EBUSY); 2087 } 2088 2089 mp->mnt_kern_flag |= MNTK_SUSPEND; 2090 mp->mnt_susp_owner = curthread; 2091 if (mp->mnt_writeopcount > 0) 2092 (void) msleep(&mp->mnt_writeopcount, 2093 MNT_MTX(mp), (PUSER - 1)|PDROP, "suspwt", 0); 2094 else 2095 MNT_IUNLOCK(mp); 2096 if ((error = VFS_SYNC(mp, MNT_SUSPEND)) != 0) { 2097 vfs_write_resume(mp, 0); 2098 /* vfs_write_resume does vfs_op_exit() for us */ 2099 } 2100 return (error); 2101 } 2102 2103 /* 2104 * Request a filesystem to resume write operations. 2105 */ 2106 void 2107 vfs_write_resume(struct mount *mp, int flags) 2108 { 2109 2110 MNT_ILOCK(mp); 2111 if ((mp->mnt_kern_flag & MNTK_SUSPEND) != 0) { 2112 KASSERT(mp->mnt_susp_owner == curthread, ("mnt_susp_owner")); 2113 mp->mnt_kern_flag &= ~(MNTK_SUSPEND | MNTK_SUSPEND2 | 2114 MNTK_SUSPENDED); 2115 mp->mnt_susp_owner = NULL; 2116 wakeup(&mp->mnt_writeopcount); 2117 wakeup(&mp->mnt_flag); 2118 curthread->td_pflags &= ~TDP_IGNSUSP; 2119 if ((flags & VR_START_WRITE) != 0) { 2120 MNT_REF(mp); 2121 mp->mnt_writeopcount++; 2122 } 2123 MNT_IUNLOCK(mp); 2124 if ((flags & VR_NO_SUSPCLR) == 0) 2125 VFS_SUSP_CLEAN(mp); 2126 vfs_op_exit(mp); 2127 } else if ((flags & VR_START_WRITE) != 0) { 2128 MNT_REF(mp); 2129 vn_start_write_refed(mp, 0, true); 2130 } else { 2131 MNT_IUNLOCK(mp); 2132 } 2133 } 2134 2135 /* 2136 * Helper loop around vfs_write_suspend() for filesystem unmount VFS 2137 * methods. 2138 */ 2139 int 2140 vfs_write_suspend_umnt(struct mount *mp) 2141 { 2142 int error; 2143 2144 KASSERT((curthread->td_pflags & TDP_IGNSUSP) == 0, 2145 ("vfs_write_suspend_umnt: recursed")); 2146 2147 /* dounmount() already called vn_start_write(). */ 2148 for (;;) { 2149 vn_finished_write(mp); 2150 error = vfs_write_suspend(mp, 0); 2151 if (error != 0) { 2152 vn_start_write(NULL, &mp, V_WAIT); 2153 return (error); 2154 } 2155 MNT_ILOCK(mp); 2156 if ((mp->mnt_kern_flag & MNTK_SUSPENDED) != 0) 2157 break; 2158 MNT_IUNLOCK(mp); 2159 vn_start_write(NULL, &mp, V_WAIT); 2160 } 2161 mp->mnt_kern_flag &= ~(MNTK_SUSPENDED | MNTK_SUSPEND2); 2162 wakeup(&mp->mnt_flag); 2163 MNT_IUNLOCK(mp); 2164 curthread->td_pflags |= TDP_IGNSUSP; 2165 return (0); 2166 } 2167 2168 /* 2169 * Implement kqueues for files by translating it to vnode operation. 2170 */ 2171 static int 2172 vn_kqfilter(struct file *fp, struct knote *kn) 2173 { 2174 2175 return (VOP_KQFILTER(fp->f_vnode, kn)); 2176 } 2177 2178 int 2179 vn_kqfilter_opath(struct file *fp, struct knote *kn) 2180 { 2181 if ((fp->f_flag & FKQALLOWED) == 0) 2182 return (EBADF); 2183 return (vn_kqfilter(fp, kn)); 2184 } 2185 2186 /* 2187 * Simplified in-kernel wrapper calls for extended attribute access. 2188 * Both calls pass in a NULL credential, authorizing as "kernel" access. 2189 * Set IO_NODELOCKED in ioflg if the vnode is already locked. 2190 */ 2191 int 2192 vn_extattr_get(struct vnode *vp, int ioflg, int attrnamespace, 2193 const char *attrname, int *buflen, char *buf, struct thread *td) 2194 { 2195 struct uio auio; 2196 struct iovec iov; 2197 int error; 2198 2199 iov.iov_len = *buflen; 2200 iov.iov_base = buf; 2201 2202 auio.uio_iov = &iov; 2203 auio.uio_iovcnt = 1; 2204 auio.uio_rw = UIO_READ; 2205 auio.uio_segflg = UIO_SYSSPACE; 2206 auio.uio_td = td; 2207 auio.uio_offset = 0; 2208 auio.uio_resid = *buflen; 2209 2210 if ((ioflg & IO_NODELOCKED) == 0) 2211 vn_lock(vp, LK_SHARED | LK_RETRY); 2212 2213 ASSERT_VOP_LOCKED(vp, "IO_NODELOCKED with no vp lock held"); 2214 2215 /* authorize attribute retrieval as kernel */ 2216 error = VOP_GETEXTATTR(vp, attrnamespace, attrname, &auio, NULL, NULL, 2217 td); 2218 2219 if ((ioflg & IO_NODELOCKED) == 0) 2220 VOP_UNLOCK(vp); 2221 2222 if (error == 0) { 2223 *buflen = *buflen - auio.uio_resid; 2224 } 2225 2226 return (error); 2227 } 2228 2229 /* 2230 * XXX failure mode if partially written? 2231 */ 2232 int 2233 vn_extattr_set(struct vnode *vp, int ioflg, int attrnamespace, 2234 const char *attrname, int buflen, char *buf, struct thread *td) 2235 { 2236 struct uio auio; 2237 struct iovec iov; 2238 struct mount *mp; 2239 int error; 2240 2241 iov.iov_len = buflen; 2242 iov.iov_base = buf; 2243 2244 auio.uio_iov = &iov; 2245 auio.uio_iovcnt = 1; 2246 auio.uio_rw = UIO_WRITE; 2247 auio.uio_segflg = UIO_SYSSPACE; 2248 auio.uio_td = td; 2249 auio.uio_offset = 0; 2250 auio.uio_resid = buflen; 2251 2252 if ((ioflg & IO_NODELOCKED) == 0) { 2253 if ((error = vn_start_write(vp, &mp, V_WAIT)) != 0) 2254 return (error); 2255 vn_lock(vp, LK_EXCLUSIVE | LK_RETRY); 2256 } 2257 2258 ASSERT_VOP_LOCKED(vp, "IO_NODELOCKED with no vp lock held"); 2259 2260 /* authorize attribute setting as kernel */ 2261 error = VOP_SETEXTATTR(vp, attrnamespace, attrname, &auio, NULL, td); 2262 2263 if ((ioflg & IO_NODELOCKED) == 0) { 2264 vn_finished_write(mp); 2265 VOP_UNLOCK(vp); 2266 } 2267 2268 return (error); 2269 } 2270 2271 int 2272 vn_extattr_rm(struct vnode *vp, int ioflg, int attrnamespace, 2273 const char *attrname, struct thread *td) 2274 { 2275 struct mount *mp; 2276 int error; 2277 2278 if ((ioflg & IO_NODELOCKED) == 0) { 2279 if ((error = vn_start_write(vp, &mp, V_WAIT)) != 0) 2280 return (error); 2281 vn_lock(vp, LK_EXCLUSIVE | LK_RETRY); 2282 } 2283 2284 ASSERT_VOP_LOCKED(vp, "IO_NODELOCKED with no vp lock held"); 2285 2286 /* authorize attribute removal as kernel */ 2287 error = VOP_DELETEEXTATTR(vp, attrnamespace, attrname, NULL, td); 2288 if (error == EOPNOTSUPP) 2289 error = VOP_SETEXTATTR(vp, attrnamespace, attrname, NULL, 2290 NULL, td); 2291 2292 if ((ioflg & IO_NODELOCKED) == 0) { 2293 vn_finished_write(mp); 2294 VOP_UNLOCK(vp); 2295 } 2296 2297 return (error); 2298 } 2299 2300 static int 2301 vn_get_ino_alloc_vget(struct mount *mp, void *arg, int lkflags, 2302 struct vnode **rvp) 2303 { 2304 2305 return (VFS_VGET(mp, *(ino_t *)arg, lkflags, rvp)); 2306 } 2307 2308 int 2309 vn_vget_ino(struct vnode *vp, ino_t ino, int lkflags, struct vnode **rvp) 2310 { 2311 2312 return (vn_vget_ino_gen(vp, vn_get_ino_alloc_vget, &ino, 2313 lkflags, rvp)); 2314 } 2315 2316 int 2317 vn_vget_ino_gen(struct vnode *vp, vn_get_ino_t alloc, void *alloc_arg, 2318 int lkflags, struct vnode **rvp) 2319 { 2320 struct mount *mp; 2321 int ltype, error; 2322 2323 ASSERT_VOP_LOCKED(vp, "vn_vget_ino_get"); 2324 mp = vp->v_mount; 2325 ltype = VOP_ISLOCKED(vp); 2326 KASSERT(ltype == LK_EXCLUSIVE || ltype == LK_SHARED, 2327 ("vn_vget_ino: vp not locked")); 2328 error = vfs_busy(mp, MBF_NOWAIT); 2329 if (error != 0) { 2330 vfs_ref(mp); 2331 VOP_UNLOCK(vp); 2332 error = vfs_busy(mp, 0); 2333 vn_lock(vp, ltype | LK_RETRY); 2334 vfs_rel(mp); 2335 if (error != 0) 2336 return (ENOENT); 2337 if (VN_IS_DOOMED(vp)) { 2338 vfs_unbusy(mp); 2339 return (ENOENT); 2340 } 2341 } 2342 VOP_UNLOCK(vp); 2343 error = alloc(mp, alloc_arg, lkflags, rvp); 2344 vfs_unbusy(mp); 2345 if (error != 0 || *rvp != vp) 2346 vn_lock(vp, ltype | LK_RETRY); 2347 if (VN_IS_DOOMED(vp)) { 2348 if (error == 0) { 2349 if (*rvp == vp) 2350 vunref(vp); 2351 else 2352 vput(*rvp); 2353 } 2354 error = ENOENT; 2355 } 2356 return (error); 2357 } 2358 2359 int 2360 vn_rlimit_fsize(const struct vnode *vp, const struct uio *uio, 2361 struct thread *td) 2362 { 2363 off_t lim; 2364 bool ktr_write; 2365 2366 if (td == NULL) 2367 return (0); 2368 2369 /* 2370 * There are conditions where the limit is to be ignored. 2371 * However, since it is almost never reached, check it first. 2372 */ 2373 ktr_write = (td->td_pflags & TDP_INKTRACE) != 0; 2374 lim = lim_cur(td, RLIMIT_FSIZE); 2375 if (__predict_false(ktr_write)) 2376 lim = td->td_ktr_io_lim; 2377 if (__predict_true((uoff_t)uio->uio_offset + uio->uio_resid <= lim)) 2378 return (0); 2379 2380 /* 2381 * The limit is reached. 2382 */ 2383 if (vp->v_type != VREG || 2384 (td->td_pflags2 & TDP2_ACCT) != 0) 2385 return (0); 2386 2387 if (!ktr_write || ktr_filesize_limit_signal) { 2388 PROC_LOCK(td->td_proc); 2389 kern_psignal(td->td_proc, SIGXFSZ); 2390 PROC_UNLOCK(td->td_proc); 2391 } 2392 return (EFBIG); 2393 } 2394 2395 int 2396 vn_chmod(struct file *fp, mode_t mode, struct ucred *active_cred, 2397 struct thread *td) 2398 { 2399 struct vnode *vp; 2400 2401 vp = fp->f_vnode; 2402 #ifdef AUDIT 2403 vn_lock(vp, LK_SHARED | LK_RETRY); 2404 AUDIT_ARG_VNODE1(vp); 2405 VOP_UNLOCK(vp); 2406 #endif 2407 return (setfmode(td, active_cred, vp, mode)); 2408 } 2409 2410 int 2411 vn_chown(struct file *fp, uid_t uid, gid_t gid, struct ucred *active_cred, 2412 struct thread *td) 2413 { 2414 struct vnode *vp; 2415 2416 vp = fp->f_vnode; 2417 #ifdef AUDIT 2418 vn_lock(vp, LK_SHARED | LK_RETRY); 2419 AUDIT_ARG_VNODE1(vp); 2420 VOP_UNLOCK(vp); 2421 #endif 2422 return (setfown(td, active_cred, vp, uid, gid)); 2423 } 2424 2425 void 2426 vn_pages_remove(struct vnode *vp, vm_pindex_t start, vm_pindex_t end) 2427 { 2428 vm_object_t object; 2429 2430 if ((object = vp->v_object) == NULL) 2431 return; 2432 VM_OBJECT_WLOCK(object); 2433 vm_object_page_remove(object, start, end, 0); 2434 VM_OBJECT_WUNLOCK(object); 2435 } 2436 2437 int 2438 vn_bmap_seekhole(struct vnode *vp, u_long cmd, off_t *off, struct ucred *cred) 2439 { 2440 struct vattr va; 2441 daddr_t bn, bnp; 2442 uint64_t bsize; 2443 off_t noff; 2444 int error; 2445 2446 KASSERT(cmd == FIOSEEKHOLE || cmd == FIOSEEKDATA, 2447 ("Wrong command %lu", cmd)); 2448 2449 if (vn_lock(vp, LK_SHARED) != 0) 2450 return (EBADF); 2451 if (vp->v_type != VREG) { 2452 error = ENOTTY; 2453 goto unlock; 2454 } 2455 error = VOP_GETATTR(vp, &va, cred); 2456 if (error != 0) 2457 goto unlock; 2458 noff = *off; 2459 if (noff >= va.va_size) { 2460 error = ENXIO; 2461 goto unlock; 2462 } 2463 bsize = vp->v_mount->mnt_stat.f_iosize; 2464 for (bn = noff / bsize; noff < va.va_size; bn++, noff += bsize - 2465 noff % bsize) { 2466 error = VOP_BMAP(vp, bn, NULL, &bnp, NULL, NULL); 2467 if (error == EOPNOTSUPP) { 2468 error = ENOTTY; 2469 goto unlock; 2470 } 2471 if ((bnp == -1 && cmd == FIOSEEKHOLE) || 2472 (bnp != -1 && cmd == FIOSEEKDATA)) { 2473 noff = bn * bsize; 2474 if (noff < *off) 2475 noff = *off; 2476 goto unlock; 2477 } 2478 } 2479 if (noff > va.va_size) 2480 noff = va.va_size; 2481 /* noff == va.va_size. There is an implicit hole at the end of file. */ 2482 if (cmd == FIOSEEKDATA) 2483 error = ENXIO; 2484 unlock: 2485 VOP_UNLOCK(vp); 2486 if (error == 0) 2487 *off = noff; 2488 return (error); 2489 } 2490 2491 int 2492 vn_seek(struct file *fp, off_t offset, int whence, struct thread *td) 2493 { 2494 struct ucred *cred; 2495 struct vnode *vp; 2496 struct vattr vattr; 2497 off_t foffset, size; 2498 int error, noneg; 2499 2500 cred = td->td_ucred; 2501 vp = fp->f_vnode; 2502 foffset = foffset_lock(fp, 0); 2503 noneg = (vp->v_type != VCHR); 2504 error = 0; 2505 switch (whence) { 2506 case L_INCR: 2507 if (noneg && 2508 (foffset < 0 || 2509 (offset > 0 && foffset > OFF_MAX - offset))) { 2510 error = EOVERFLOW; 2511 break; 2512 } 2513 offset += foffset; 2514 break; 2515 case L_XTND: 2516 vn_lock(vp, LK_SHARED | LK_RETRY); 2517 error = VOP_GETATTR(vp, &vattr, cred); 2518 VOP_UNLOCK(vp); 2519 if (error) 2520 break; 2521 2522 /* 2523 * If the file references a disk device, then fetch 2524 * the media size and use that to determine the ending 2525 * offset. 2526 */ 2527 if (vattr.va_size == 0 && vp->v_type == VCHR && 2528 fo_ioctl(fp, DIOCGMEDIASIZE, &size, cred, td) == 0) 2529 vattr.va_size = size; 2530 if (noneg && 2531 (vattr.va_size > OFF_MAX || 2532 (offset > 0 && vattr.va_size > OFF_MAX - offset))) { 2533 error = EOVERFLOW; 2534 break; 2535 } 2536 offset += vattr.va_size; 2537 break; 2538 case L_SET: 2539 break; 2540 case SEEK_DATA: 2541 error = fo_ioctl(fp, FIOSEEKDATA, &offset, cred, td); 2542 if (error == ENOTTY) 2543 error = EINVAL; 2544 break; 2545 case SEEK_HOLE: 2546 error = fo_ioctl(fp, FIOSEEKHOLE, &offset, cred, td); 2547 if (error == ENOTTY) 2548 error = EINVAL; 2549 break; 2550 default: 2551 error = EINVAL; 2552 } 2553 if (error == 0 && noneg && offset < 0) 2554 error = EINVAL; 2555 if (error != 0) 2556 goto drop; 2557 VFS_KNOTE_UNLOCKED(vp, 0); 2558 td->td_uretoff.tdu_off = offset; 2559 drop: 2560 foffset_unlock(fp, offset, error != 0 ? FOF_NOUPDATE : 0); 2561 return (error); 2562 } 2563 2564 int 2565 vn_utimes_perm(struct vnode *vp, struct vattr *vap, struct ucred *cred, 2566 struct thread *td) 2567 { 2568 int error; 2569 2570 /* 2571 * Grant permission if the caller is the owner of the file, or 2572 * the super-user, or has ACL_WRITE_ATTRIBUTES permission on 2573 * on the file. If the time pointer is null, then write 2574 * permission on the file is also sufficient. 2575 * 2576 * From NFSv4.1, draft 21, 6.2.1.3.1, Discussion of Mask Attributes: 2577 * A user having ACL_WRITE_DATA or ACL_WRITE_ATTRIBUTES 2578 * will be allowed to set the times [..] to the current 2579 * server time. 2580 */ 2581 error = VOP_ACCESSX(vp, VWRITE_ATTRIBUTES, cred, td); 2582 if (error != 0 && (vap->va_vaflags & VA_UTIMES_NULL) != 0) 2583 error = VOP_ACCESS(vp, VWRITE, cred, td); 2584 return (error); 2585 } 2586 2587 int 2588 vn_fill_kinfo(struct file *fp, struct kinfo_file *kif, struct filedesc *fdp) 2589 { 2590 struct vnode *vp; 2591 int error; 2592 2593 if (fp->f_type == DTYPE_FIFO) 2594 kif->kf_type = KF_TYPE_FIFO; 2595 else 2596 kif->kf_type = KF_TYPE_VNODE; 2597 vp = fp->f_vnode; 2598 vref(vp); 2599 FILEDESC_SUNLOCK(fdp); 2600 error = vn_fill_kinfo_vnode(vp, kif); 2601 vrele(vp); 2602 FILEDESC_SLOCK(fdp); 2603 return (error); 2604 } 2605 2606 static inline void 2607 vn_fill_junk(struct kinfo_file *kif) 2608 { 2609 size_t len, olen; 2610 2611 /* 2612 * Simulate vn_fullpath returning changing values for a given 2613 * vp during e.g. coredump. 2614 */ 2615 len = (arc4random() % (sizeof(kif->kf_path) - 2)) + 1; 2616 olen = strlen(kif->kf_path); 2617 if (len < olen) 2618 strcpy(&kif->kf_path[len - 1], "$"); 2619 else 2620 for (; olen < len; olen++) 2621 strcpy(&kif->kf_path[olen], "A"); 2622 } 2623 2624 int 2625 vn_fill_kinfo_vnode(struct vnode *vp, struct kinfo_file *kif) 2626 { 2627 struct vattr va; 2628 char *fullpath, *freepath; 2629 int error; 2630 2631 kif->kf_un.kf_file.kf_file_type = vntype_to_kinfo(vp->v_type); 2632 freepath = NULL; 2633 fullpath = "-"; 2634 error = vn_fullpath(vp, &fullpath, &freepath); 2635 if (error == 0) { 2636 strlcpy(kif->kf_path, fullpath, sizeof(kif->kf_path)); 2637 } 2638 if (freepath != NULL) 2639 free(freepath, M_TEMP); 2640 2641 KFAIL_POINT_CODE(DEBUG_FP, fill_kinfo_vnode__random_path, 2642 vn_fill_junk(kif); 2643 ); 2644 2645 /* 2646 * Retrieve vnode attributes. 2647 */ 2648 va.va_fsid = VNOVAL; 2649 va.va_rdev = NODEV; 2650 vn_lock(vp, LK_SHARED | LK_RETRY); 2651 error = VOP_GETATTR(vp, &va, curthread->td_ucred); 2652 VOP_UNLOCK(vp); 2653 if (error != 0) 2654 return (error); 2655 if (va.va_fsid != VNOVAL) 2656 kif->kf_un.kf_file.kf_file_fsid = va.va_fsid; 2657 else 2658 kif->kf_un.kf_file.kf_file_fsid = 2659 vp->v_mount->mnt_stat.f_fsid.val[0]; 2660 kif->kf_un.kf_file.kf_file_fsid_freebsd11 = 2661 kif->kf_un.kf_file.kf_file_fsid; /* truncate */ 2662 kif->kf_un.kf_file.kf_file_fileid = va.va_fileid; 2663 kif->kf_un.kf_file.kf_file_mode = MAKEIMODE(va.va_type, va.va_mode); 2664 kif->kf_un.kf_file.kf_file_size = va.va_size; 2665 kif->kf_un.kf_file.kf_file_rdev = va.va_rdev; 2666 kif->kf_un.kf_file.kf_file_rdev_freebsd11 = 2667 kif->kf_un.kf_file.kf_file_rdev; /* truncate */ 2668 return (0); 2669 } 2670 2671 int 2672 vn_mmap(struct file *fp, vm_map_t map, vm_offset_t *addr, vm_size_t size, 2673 vm_prot_t prot, vm_prot_t cap_maxprot, int flags, vm_ooffset_t foff, 2674 struct thread *td) 2675 { 2676 #ifdef HWPMC_HOOKS 2677 struct pmckern_map_in pkm; 2678 #endif 2679 struct mount *mp; 2680 struct vnode *vp; 2681 vm_object_t object; 2682 vm_prot_t maxprot; 2683 boolean_t writecounted; 2684 int error; 2685 2686 #if defined(COMPAT_FREEBSD7) || defined(COMPAT_FREEBSD6) || \ 2687 defined(COMPAT_FREEBSD5) || defined(COMPAT_FREEBSD4) 2688 /* 2689 * POSIX shared-memory objects are defined to have 2690 * kernel persistence, and are not defined to support 2691 * read(2)/write(2) -- or even open(2). Thus, we can 2692 * use MAP_ASYNC to trade on-disk coherence for speed. 2693 * The shm_open(3) library routine turns on the FPOSIXSHM 2694 * flag to request this behavior. 2695 */ 2696 if ((fp->f_flag & FPOSIXSHM) != 0) 2697 flags |= MAP_NOSYNC; 2698 #endif 2699 vp = fp->f_vnode; 2700 2701 /* 2702 * Ensure that file and memory protections are 2703 * compatible. Note that we only worry about 2704 * writability if mapping is shared; in this case, 2705 * current and max prot are dictated by the open file. 2706 * XXX use the vnode instead? Problem is: what 2707 * credentials do we use for determination? What if 2708 * proc does a setuid? 2709 */ 2710 mp = vp->v_mount; 2711 if (mp != NULL && (mp->mnt_flag & MNT_NOEXEC) != 0) { 2712 maxprot = VM_PROT_NONE; 2713 if ((prot & VM_PROT_EXECUTE) != 0) 2714 return (EACCES); 2715 } else 2716 maxprot = VM_PROT_EXECUTE; 2717 if ((fp->f_flag & FREAD) != 0) 2718 maxprot |= VM_PROT_READ; 2719 else if ((prot & VM_PROT_READ) != 0) 2720 return (EACCES); 2721 2722 /* 2723 * If we are sharing potential changes via MAP_SHARED and we 2724 * are trying to get write permission although we opened it 2725 * without asking for it, bail out. 2726 */ 2727 if ((flags & MAP_SHARED) != 0) { 2728 if ((fp->f_flag & FWRITE) != 0) 2729 maxprot |= VM_PROT_WRITE; 2730 else if ((prot & VM_PROT_WRITE) != 0) 2731 return (EACCES); 2732 } else { 2733 maxprot |= VM_PROT_WRITE; 2734 cap_maxprot |= VM_PROT_WRITE; 2735 } 2736 maxprot &= cap_maxprot; 2737 2738 /* 2739 * For regular files and shared memory, POSIX requires that 2740 * the value of foff be a legitimate offset within the data 2741 * object. In particular, negative offsets are invalid. 2742 * Blocking negative offsets and overflows here avoids 2743 * possible wraparound or user-level access into reserved 2744 * ranges of the data object later. In contrast, POSIX does 2745 * not dictate how offsets are used by device drivers, so in 2746 * the case of a device mapping a negative offset is passed 2747 * on. 2748 */ 2749 if ( 2750 #ifdef _LP64 2751 size > OFF_MAX || 2752 #endif 2753 foff > OFF_MAX - size) 2754 return (EINVAL); 2755 2756 writecounted = FALSE; 2757 error = vm_mmap_vnode(td, size, prot, &maxprot, &flags, vp, 2758 &foff, &object, &writecounted); 2759 if (error != 0) 2760 return (error); 2761 error = vm_mmap_object(map, addr, size, prot, maxprot, flags, object, 2762 foff, writecounted, td); 2763 if (error != 0) { 2764 /* 2765 * If this mapping was accounted for in the vnode's 2766 * writecount, then undo that now. 2767 */ 2768 if (writecounted) 2769 vm_pager_release_writecount(object, 0, size); 2770 vm_object_deallocate(object); 2771 } 2772 #ifdef HWPMC_HOOKS 2773 /* Inform hwpmc(4) if an executable is being mapped. */ 2774 if (PMC_HOOK_INSTALLED(PMC_FN_MMAP)) { 2775 if ((prot & VM_PROT_EXECUTE) != 0 && error == 0) { 2776 pkm.pm_file = vp; 2777 pkm.pm_address = (uintptr_t) *addr; 2778 PMC_CALL_HOOK_UNLOCKED(td, PMC_FN_MMAP, (void *) &pkm); 2779 } 2780 } 2781 #endif 2782 return (error); 2783 } 2784 2785 void 2786 vn_fsid(struct vnode *vp, struct vattr *va) 2787 { 2788 fsid_t *f; 2789 2790 f = &vp->v_mount->mnt_stat.f_fsid; 2791 va->va_fsid = (uint32_t)f->val[1]; 2792 va->va_fsid <<= sizeof(f->val[1]) * NBBY; 2793 va->va_fsid += (uint32_t)f->val[0]; 2794 } 2795 2796 int 2797 vn_fsync_buf(struct vnode *vp, int waitfor) 2798 { 2799 struct buf *bp, *nbp; 2800 struct bufobj *bo; 2801 struct mount *mp; 2802 int error, maxretry; 2803 2804 error = 0; 2805 maxretry = 10000; /* large, arbitrarily chosen */ 2806 mp = NULL; 2807 if (vp->v_type == VCHR) { 2808 VI_LOCK(vp); 2809 mp = vp->v_rdev->si_mountpt; 2810 VI_UNLOCK(vp); 2811 } 2812 bo = &vp->v_bufobj; 2813 BO_LOCK(bo); 2814 loop1: 2815 /* 2816 * MARK/SCAN initialization to avoid infinite loops. 2817 */ 2818 TAILQ_FOREACH(bp, &bo->bo_dirty.bv_hd, b_bobufs) { 2819 bp->b_vflags &= ~BV_SCANNED; 2820 bp->b_error = 0; 2821 } 2822 2823 /* 2824 * Flush all dirty buffers associated with a vnode. 2825 */ 2826 loop2: 2827 TAILQ_FOREACH_SAFE(bp, &bo->bo_dirty.bv_hd, b_bobufs, nbp) { 2828 if ((bp->b_vflags & BV_SCANNED) != 0) 2829 continue; 2830 bp->b_vflags |= BV_SCANNED; 2831 if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT, NULL)) { 2832 if (waitfor != MNT_WAIT) 2833 continue; 2834 if (BUF_LOCK(bp, 2835 LK_EXCLUSIVE | LK_INTERLOCK | LK_SLEEPFAIL, 2836 BO_LOCKPTR(bo)) != 0) { 2837 BO_LOCK(bo); 2838 goto loop1; 2839 } 2840 BO_LOCK(bo); 2841 } 2842 BO_UNLOCK(bo); 2843 KASSERT(bp->b_bufobj == bo, 2844 ("bp %p wrong b_bufobj %p should be %p", 2845 bp, bp->b_bufobj, bo)); 2846 if ((bp->b_flags & B_DELWRI) == 0) 2847 panic("fsync: not dirty"); 2848 if ((vp->v_object != NULL) && (bp->b_flags & B_CLUSTEROK)) { 2849 vfs_bio_awrite(bp); 2850 } else { 2851 bremfree(bp); 2852 bawrite(bp); 2853 } 2854 if (maxretry < 1000) 2855 pause("dirty", hz < 1000 ? 1 : hz / 1000); 2856 BO_LOCK(bo); 2857 goto loop2; 2858 } 2859 2860 /* 2861 * If synchronous the caller expects us to completely resolve all 2862 * dirty buffers in the system. Wait for in-progress I/O to 2863 * complete (which could include background bitmap writes), then 2864 * retry if dirty blocks still exist. 2865 */ 2866 if (waitfor == MNT_WAIT) { 2867 bufobj_wwait(bo, 0, 0); 2868 if (bo->bo_dirty.bv_cnt > 0) { 2869 /* 2870 * If we are unable to write any of these buffers 2871 * then we fail now rather than trying endlessly 2872 * to write them out. 2873 */ 2874 TAILQ_FOREACH(bp, &bo->bo_dirty.bv_hd, b_bobufs) 2875 if ((error = bp->b_error) != 0) 2876 break; 2877 if ((mp != NULL && mp->mnt_secondary_writes > 0) || 2878 (error == 0 && --maxretry >= 0)) 2879 goto loop1; 2880 if (error == 0) 2881 error = EAGAIN; 2882 } 2883 } 2884 BO_UNLOCK(bo); 2885 if (error != 0) 2886 vn_printf(vp, "fsync: giving up on dirty (error = %d) ", error); 2887 2888 return (error); 2889 } 2890 2891 /* 2892 * Copies a byte range from invp to outvp. Calls VOP_COPY_FILE_RANGE() 2893 * or vn_generic_copy_file_range() after rangelocking the byte ranges, 2894 * to do the actual copy. 2895 * vn_generic_copy_file_range() is factored out, so it can be called 2896 * from a VOP_COPY_FILE_RANGE() call as well, but handles vnodes from 2897 * different file systems. 2898 */ 2899 int 2900 vn_copy_file_range(struct vnode *invp, off_t *inoffp, struct vnode *outvp, 2901 off_t *outoffp, size_t *lenp, unsigned int flags, struct ucred *incred, 2902 struct ucred *outcred, struct thread *fsize_td) 2903 { 2904 int error; 2905 size_t len; 2906 uint64_t uval; 2907 2908 len = *lenp; 2909 *lenp = 0; /* For error returns. */ 2910 error = 0; 2911 2912 /* Do some sanity checks on the arguments. */ 2913 if (invp->v_type == VDIR || outvp->v_type == VDIR) 2914 error = EISDIR; 2915 else if (*inoffp < 0 || *outoffp < 0 || 2916 invp->v_type != VREG || outvp->v_type != VREG) 2917 error = EINVAL; 2918 if (error != 0) 2919 goto out; 2920 2921 /* Ensure offset + len does not wrap around. */ 2922 uval = *inoffp; 2923 uval += len; 2924 if (uval > INT64_MAX) 2925 len = INT64_MAX - *inoffp; 2926 uval = *outoffp; 2927 uval += len; 2928 if (uval > INT64_MAX) 2929 len = INT64_MAX - *outoffp; 2930 if (len == 0) 2931 goto out; 2932 2933 /* 2934 * If the two vnode are for the same file system, call 2935 * VOP_COPY_FILE_RANGE(), otherwise call vn_generic_copy_file_range() 2936 * which can handle copies across multiple file systems. 2937 */ 2938 *lenp = len; 2939 if (invp->v_mount == outvp->v_mount) 2940 error = VOP_COPY_FILE_RANGE(invp, inoffp, outvp, outoffp, 2941 lenp, flags, incred, outcred, fsize_td); 2942 else 2943 error = vn_generic_copy_file_range(invp, inoffp, outvp, 2944 outoffp, lenp, flags, incred, outcred, fsize_td); 2945 out: 2946 return (error); 2947 } 2948 2949 /* 2950 * Test len bytes of data starting at dat for all bytes == 0. 2951 * Return true if all bytes are zero, false otherwise. 2952 * Expects dat to be well aligned. 2953 */ 2954 static bool 2955 mem_iszero(void *dat, int len) 2956 { 2957 int i; 2958 const u_int *p; 2959 const char *cp; 2960 2961 for (p = dat; len > 0; len -= sizeof(*p), p++) { 2962 if (len >= sizeof(*p)) { 2963 if (*p != 0) 2964 return (false); 2965 } else { 2966 cp = (const char *)p; 2967 for (i = 0; i < len; i++, cp++) 2968 if (*cp != '\0') 2969 return (false); 2970 } 2971 } 2972 return (true); 2973 } 2974 2975 /* 2976 * Look for a hole in the output file and, if found, adjust *outoffp 2977 * and *xferp to skip past the hole. 2978 * *xferp is the entire hole length to be written and xfer2 is how many bytes 2979 * to be written as 0's upon return. 2980 */ 2981 static off_t 2982 vn_skip_hole(struct vnode *outvp, off_t xfer2, off_t *outoffp, off_t *xferp, 2983 off_t *dataoffp, off_t *holeoffp, struct ucred *cred) 2984 { 2985 int error; 2986 off_t delta; 2987 2988 if (*holeoffp == 0 || *holeoffp <= *outoffp) { 2989 *dataoffp = *outoffp; 2990 error = VOP_IOCTL(outvp, FIOSEEKDATA, dataoffp, 0, cred, 2991 curthread); 2992 if (error == 0) { 2993 *holeoffp = *dataoffp; 2994 error = VOP_IOCTL(outvp, FIOSEEKHOLE, holeoffp, 0, cred, 2995 curthread); 2996 } 2997 if (error != 0 || *holeoffp == *dataoffp) { 2998 /* 2999 * Since outvp is unlocked, it may be possible for 3000 * another thread to do a truncate(), lseek(), write() 3001 * creating a hole at startoff between the above 3002 * VOP_IOCTL() calls, if the other thread does not do 3003 * rangelocking. 3004 * If that happens, *holeoffp == *dataoffp and finding 3005 * the hole has failed, so disable vn_skip_hole(). 3006 */ 3007 *holeoffp = -1; /* Disable use of vn_skip_hole(). */ 3008 return (xfer2); 3009 } 3010 KASSERT(*dataoffp >= *outoffp, 3011 ("vn_skip_hole: dataoff=%jd < outoff=%jd", 3012 (intmax_t)*dataoffp, (intmax_t)*outoffp)); 3013 KASSERT(*holeoffp > *dataoffp, 3014 ("vn_skip_hole: holeoff=%jd <= dataoff=%jd", 3015 (intmax_t)*holeoffp, (intmax_t)*dataoffp)); 3016 } 3017 3018 /* 3019 * If there is a hole before the data starts, advance *outoffp and 3020 * *xferp past the hole. 3021 */ 3022 if (*dataoffp > *outoffp) { 3023 delta = *dataoffp - *outoffp; 3024 if (delta >= *xferp) { 3025 /* Entire *xferp is a hole. */ 3026 *outoffp += *xferp; 3027 *xferp = 0; 3028 return (0); 3029 } 3030 *xferp -= delta; 3031 *outoffp += delta; 3032 xfer2 = MIN(xfer2, *xferp); 3033 } 3034 3035 /* 3036 * If a hole starts before the end of this xfer2, reduce this xfer2 so 3037 * that the write ends at the start of the hole. 3038 * *holeoffp should always be greater than *outoffp, but for the 3039 * non-INVARIANTS case, check this to make sure xfer2 remains a sane 3040 * value. 3041 */ 3042 if (*holeoffp > *outoffp && *holeoffp < *outoffp + xfer2) 3043 xfer2 = *holeoffp - *outoffp; 3044 return (xfer2); 3045 } 3046 3047 /* 3048 * Write an xfer sized chunk to outvp in blksize blocks from dat. 3049 * dat is a maximum of blksize in length and can be written repeatedly in 3050 * the chunk. 3051 * If growfile == true, just grow the file via vn_truncate_locked() instead 3052 * of doing actual writes. 3053 * If checkhole == true, a hole is being punched, so skip over any hole 3054 * already in the output file. 3055 */ 3056 static int 3057 vn_write_outvp(struct vnode *outvp, char *dat, off_t outoff, off_t xfer, 3058 u_long blksize, bool growfile, bool checkhole, struct ucred *cred) 3059 { 3060 struct mount *mp; 3061 off_t dataoff, holeoff, xfer2; 3062 int error, lckf; 3063 3064 /* 3065 * Loop around doing writes of blksize until write has been completed. 3066 * Lock/unlock on each loop iteration so that a bwillwrite() can be 3067 * done for each iteration, since the xfer argument can be very 3068 * large if there is a large hole to punch in the output file. 3069 */ 3070 error = 0; 3071 holeoff = 0; 3072 do { 3073 xfer2 = MIN(xfer, blksize); 3074 if (checkhole) { 3075 /* 3076 * Punching a hole. Skip writing if there is 3077 * already a hole in the output file. 3078 */ 3079 xfer2 = vn_skip_hole(outvp, xfer2, &outoff, &xfer, 3080 &dataoff, &holeoff, cred); 3081 if (xfer == 0) 3082 break; 3083 if (holeoff < 0) 3084 checkhole = false; 3085 KASSERT(xfer2 > 0, ("vn_write_outvp: xfer2=%jd", 3086 (intmax_t)xfer2)); 3087 } 3088 bwillwrite(); 3089 mp = NULL; 3090 error = vn_start_write(outvp, &mp, V_WAIT); 3091 if (error != 0) 3092 break; 3093 if (growfile) { 3094 error = vn_lock(outvp, LK_EXCLUSIVE); 3095 if (error == 0) { 3096 error = vn_truncate_locked(outvp, outoff + xfer, 3097 false, cred); 3098 VOP_UNLOCK(outvp); 3099 } 3100 } else { 3101 if (MNT_SHARED_WRITES(mp)) 3102 lckf = LK_SHARED; 3103 else 3104 lckf = LK_EXCLUSIVE; 3105 error = vn_lock(outvp, lckf); 3106 if (error == 0) { 3107 error = vn_rdwr(UIO_WRITE, outvp, dat, xfer2, 3108 outoff, UIO_SYSSPACE, IO_NODELOCKED, 3109 curthread->td_ucred, cred, NULL, curthread); 3110 outoff += xfer2; 3111 xfer -= xfer2; 3112 VOP_UNLOCK(outvp); 3113 } 3114 } 3115 if (mp != NULL) 3116 vn_finished_write(mp); 3117 } while (!growfile && xfer > 0 && error == 0); 3118 return (error); 3119 } 3120 3121 /* 3122 * Copy a byte range of one file to another. This function can handle the 3123 * case where invp and outvp are on different file systems. 3124 * It can also be called by a VOP_COPY_FILE_RANGE() to do the work, if there 3125 * is no better file system specific way to do it. 3126 */ 3127 int 3128 vn_generic_copy_file_range(struct vnode *invp, off_t *inoffp, 3129 struct vnode *outvp, off_t *outoffp, size_t *lenp, unsigned int flags, 3130 struct ucred *incred, struct ucred *outcred, struct thread *fsize_td) 3131 { 3132 struct vattr va, inva; 3133 struct mount *mp; 3134 struct uio io; 3135 off_t startoff, endoff, xfer, xfer2; 3136 u_long blksize; 3137 int error, interrupted; 3138 bool cantseek, readzeros, eof, lastblock, holetoeof; 3139 ssize_t aresid; 3140 size_t copylen, len, rem, savlen; 3141 char *dat; 3142 long holein, holeout; 3143 3144 holein = holeout = 0; 3145 savlen = len = *lenp; 3146 error = 0; 3147 interrupted = 0; 3148 dat = NULL; 3149 3150 error = vn_lock(invp, LK_SHARED); 3151 if (error != 0) 3152 goto out; 3153 if (VOP_PATHCONF(invp, _PC_MIN_HOLE_SIZE, &holein) != 0) 3154 holein = 0; 3155 if (holein > 0) 3156 error = VOP_GETATTR(invp, &inva, incred); 3157 VOP_UNLOCK(invp); 3158 if (error != 0) 3159 goto out; 3160 3161 mp = NULL; 3162 error = vn_start_write(outvp, &mp, V_WAIT); 3163 if (error == 0) 3164 error = vn_lock(outvp, LK_EXCLUSIVE); 3165 if (error == 0) { 3166 /* 3167 * If fsize_td != NULL, do a vn_rlimit_fsize() call, 3168 * now that outvp is locked. 3169 */ 3170 if (fsize_td != NULL) { 3171 io.uio_offset = *outoffp; 3172 io.uio_resid = len; 3173 error = vn_rlimit_fsize(outvp, &io, fsize_td); 3174 if (error != 0) 3175 error = EFBIG; 3176 } 3177 if (VOP_PATHCONF(outvp, _PC_MIN_HOLE_SIZE, &holeout) != 0) 3178 holeout = 0; 3179 /* 3180 * Holes that are past EOF do not need to be written as a block 3181 * of zero bytes. So, truncate the output file as far as 3182 * possible and then use va.va_size to decide if writing 0 3183 * bytes is necessary in the loop below. 3184 */ 3185 if (error == 0) 3186 error = VOP_GETATTR(outvp, &va, outcred); 3187 if (error == 0 && va.va_size > *outoffp && va.va_size <= 3188 *outoffp + len) { 3189 #ifdef MAC 3190 error = mac_vnode_check_write(curthread->td_ucred, 3191 outcred, outvp); 3192 if (error == 0) 3193 #endif 3194 error = vn_truncate_locked(outvp, *outoffp, 3195 false, outcred); 3196 if (error == 0) 3197 va.va_size = *outoffp; 3198 } 3199 VOP_UNLOCK(outvp); 3200 } 3201 if (mp != NULL) 3202 vn_finished_write(mp); 3203 if (error != 0) 3204 goto out; 3205 3206 /* 3207 * Set the blksize to the larger of the hole sizes for invp and outvp. 3208 * If hole sizes aren't available, set the blksize to the larger 3209 * f_iosize of invp and outvp. 3210 * This code expects the hole sizes and f_iosizes to be powers of 2. 3211 * This value is clipped at 4Kbytes and 1Mbyte. 3212 */ 3213 blksize = MAX(holein, holeout); 3214 3215 /* Clip len to end at an exact multiple of hole size. */ 3216 if (blksize > 1) { 3217 rem = *inoffp % blksize; 3218 if (rem > 0) 3219 rem = blksize - rem; 3220 if (len > rem && len - rem > blksize) 3221 len = savlen = rounddown(len - rem, blksize) + rem; 3222 } 3223 3224 if (blksize <= 1) 3225 blksize = MAX(invp->v_mount->mnt_stat.f_iosize, 3226 outvp->v_mount->mnt_stat.f_iosize); 3227 if (blksize < 4096) 3228 blksize = 4096; 3229 else if (blksize > 1024 * 1024) 3230 blksize = 1024 * 1024; 3231 dat = malloc(blksize, M_TEMP, M_WAITOK); 3232 3233 /* 3234 * If VOP_IOCTL(FIOSEEKHOLE) works for invp, use it and FIOSEEKDATA 3235 * to find holes. Otherwise, just scan the read block for all 0s 3236 * in the inner loop where the data copying is done. 3237 * Note that some file systems such as NFSv3, NFSv4.0 and NFSv4.1 may 3238 * support holes on the server, but do not support FIOSEEKHOLE. 3239 */ 3240 holetoeof = eof = false; 3241 while (len > 0 && error == 0 && !eof && interrupted == 0) { 3242 endoff = 0; /* To shut up compilers. */ 3243 cantseek = true; 3244 startoff = *inoffp; 3245 copylen = len; 3246 3247 /* 3248 * Find the next data area. If there is just a hole to EOF, 3249 * FIOSEEKDATA should fail with ENXIO. 3250 * (I do not know if any file system will report a hole to 3251 * EOF via FIOSEEKHOLE, but I am pretty sure FIOSEEKDATA 3252 * will fail for those file systems.) 3253 * 3254 * For input files that don't support FIOSEEKDATA/FIOSEEKHOLE, 3255 * the code just falls through to the inner copy loop. 3256 */ 3257 error = EINVAL; 3258 if (holein > 0) { 3259 error = VOP_IOCTL(invp, FIOSEEKDATA, &startoff, 0, 3260 incred, curthread); 3261 if (error == ENXIO) { 3262 startoff = endoff = inva.va_size; 3263 eof = holetoeof = true; 3264 error = 0; 3265 } 3266 } 3267 if (error == 0 && !holetoeof) { 3268 endoff = startoff; 3269 error = VOP_IOCTL(invp, FIOSEEKHOLE, &endoff, 0, 3270 incred, curthread); 3271 /* 3272 * Since invp is unlocked, it may be possible for 3273 * another thread to do a truncate(), lseek(), write() 3274 * creating a hole at startoff between the above 3275 * VOP_IOCTL() calls, if the other thread does not do 3276 * rangelocking. 3277 * If that happens, startoff == endoff and finding 3278 * the hole has failed, so set an error. 3279 */ 3280 if (error == 0 && startoff == endoff) 3281 error = EINVAL; /* Any error. Reset to 0. */ 3282 } 3283 if (error == 0) { 3284 if (startoff > *inoffp) { 3285 /* Found hole before data block. */ 3286 xfer = MIN(startoff - *inoffp, len); 3287 if (*outoffp < va.va_size) { 3288 /* Must write 0s to punch hole. */ 3289 xfer2 = MIN(va.va_size - *outoffp, 3290 xfer); 3291 memset(dat, 0, MIN(xfer2, blksize)); 3292 error = vn_write_outvp(outvp, dat, 3293 *outoffp, xfer2, blksize, false, 3294 holeout > 0, outcred); 3295 } 3296 3297 if (error == 0 && *outoffp + xfer > 3298 va.va_size && (xfer == len || holetoeof)) { 3299 /* Grow output file (hole at end). */ 3300 error = vn_write_outvp(outvp, dat, 3301 *outoffp, xfer, blksize, true, 3302 false, outcred); 3303 } 3304 if (error == 0) { 3305 *inoffp += xfer; 3306 *outoffp += xfer; 3307 len -= xfer; 3308 if (len < savlen) 3309 interrupted = sig_intr(); 3310 } 3311 } 3312 copylen = MIN(len, endoff - startoff); 3313 cantseek = false; 3314 } else { 3315 cantseek = true; 3316 startoff = *inoffp; 3317 copylen = len; 3318 error = 0; 3319 } 3320 3321 xfer = blksize; 3322 if (cantseek) { 3323 /* 3324 * Set first xfer to end at a block boundary, so that 3325 * holes are more likely detected in the loop below via 3326 * the for all bytes 0 method. 3327 */ 3328 xfer -= (*inoffp % blksize); 3329 } 3330 /* Loop copying the data block. */ 3331 while (copylen > 0 && error == 0 && !eof && interrupted == 0) { 3332 if (copylen < xfer) 3333 xfer = copylen; 3334 error = vn_lock(invp, LK_SHARED); 3335 if (error != 0) 3336 goto out; 3337 error = vn_rdwr(UIO_READ, invp, dat, xfer, 3338 startoff, UIO_SYSSPACE, IO_NODELOCKED, 3339 curthread->td_ucred, incred, &aresid, 3340 curthread); 3341 VOP_UNLOCK(invp); 3342 lastblock = false; 3343 if (error == 0 && aresid > 0) { 3344 /* Stop the copy at EOF on the input file. */ 3345 xfer -= aresid; 3346 eof = true; 3347 lastblock = true; 3348 } 3349 if (error == 0) { 3350 /* 3351 * Skip the write for holes past the initial EOF 3352 * of the output file, unless this is the last 3353 * write of the output file at EOF. 3354 */ 3355 readzeros = cantseek ? mem_iszero(dat, xfer) : 3356 false; 3357 if (xfer == len) 3358 lastblock = true; 3359 if (!cantseek || *outoffp < va.va_size || 3360 lastblock || !readzeros) 3361 error = vn_write_outvp(outvp, dat, 3362 *outoffp, xfer, blksize, 3363 readzeros && lastblock && 3364 *outoffp >= va.va_size, false, 3365 outcred); 3366 if (error == 0) { 3367 *inoffp += xfer; 3368 startoff += xfer; 3369 *outoffp += xfer; 3370 copylen -= xfer; 3371 len -= xfer; 3372 if (len < savlen) 3373 interrupted = sig_intr(); 3374 } 3375 } 3376 xfer = blksize; 3377 } 3378 } 3379 out: 3380 *lenp = savlen - len; 3381 free(dat, M_TEMP); 3382 return (error); 3383 } 3384 3385 static int 3386 vn_fallocate(struct file *fp, off_t offset, off_t len, struct thread *td) 3387 { 3388 struct mount *mp; 3389 struct vnode *vp; 3390 off_t olen, ooffset; 3391 int error; 3392 #ifdef AUDIT 3393 int audited_vnode1 = 0; 3394 #endif 3395 3396 vp = fp->f_vnode; 3397 if (vp->v_type != VREG) 3398 return (ENODEV); 3399 3400 /* Allocating blocks may take a long time, so iterate. */ 3401 for (;;) { 3402 olen = len; 3403 ooffset = offset; 3404 3405 bwillwrite(); 3406 mp = NULL; 3407 error = vn_start_write(vp, &mp, V_WAIT | PCATCH); 3408 if (error != 0) 3409 break; 3410 error = vn_lock(vp, LK_EXCLUSIVE); 3411 if (error != 0) { 3412 vn_finished_write(mp); 3413 break; 3414 } 3415 #ifdef AUDIT 3416 if (!audited_vnode1) { 3417 AUDIT_ARG_VNODE1(vp); 3418 audited_vnode1 = 1; 3419 } 3420 #endif 3421 #ifdef MAC 3422 error = mac_vnode_check_write(td->td_ucred, fp->f_cred, vp); 3423 if (error == 0) 3424 #endif 3425 error = VOP_ALLOCATE(vp, &offset, &len); 3426 VOP_UNLOCK(vp); 3427 vn_finished_write(mp); 3428 3429 if (olen + ooffset != offset + len) { 3430 panic("offset + len changed from %jx/%jx to %jx/%jx", 3431 ooffset, olen, offset, len); 3432 } 3433 if (error != 0 || len == 0) 3434 break; 3435 KASSERT(olen > len, ("Iteration did not make progress?")); 3436 maybe_yield(); 3437 } 3438 3439 return (error); 3440 } 3441 3442 static u_long vn_lock_pair_pause_cnt; 3443 SYSCTL_ULONG(_debug, OID_AUTO, vn_lock_pair_pause, CTLFLAG_RD, 3444 &vn_lock_pair_pause_cnt, 0, 3445 "Count of vn_lock_pair deadlocks"); 3446 3447 u_int vn_lock_pair_pause_max; 3448 SYSCTL_UINT(_debug, OID_AUTO, vn_lock_pair_pause_max, CTLFLAG_RW, 3449 &vn_lock_pair_pause_max, 0, 3450 "Max ticks for vn_lock_pair deadlock avoidance sleep"); 3451 3452 static void 3453 vn_lock_pair_pause(const char *wmesg) 3454 { 3455 atomic_add_long(&vn_lock_pair_pause_cnt, 1); 3456 pause(wmesg, prng32_bounded(vn_lock_pair_pause_max)); 3457 } 3458 3459 /* 3460 * Lock pair of vnodes vp1, vp2, avoiding lock order reversal. 3461 * vp1_locked indicates whether vp1 is exclusively locked; if not, vp1 3462 * must be unlocked. Same for vp2 and vp2_locked. One of the vnodes 3463 * can be NULL. 3464 * 3465 * The function returns with both vnodes exclusively locked, and 3466 * guarantees that it does not create lock order reversal with other 3467 * threads during its execution. Both vnodes could be unlocked 3468 * temporary (and reclaimed). 3469 */ 3470 void 3471 vn_lock_pair(struct vnode *vp1, bool vp1_locked, struct vnode *vp2, 3472 bool vp2_locked) 3473 { 3474 int error; 3475 3476 if (vp1 == NULL && vp2 == NULL) 3477 return; 3478 if (vp1 != NULL) { 3479 if (vp1_locked) 3480 ASSERT_VOP_ELOCKED(vp1, "vp1"); 3481 else 3482 ASSERT_VOP_UNLOCKED(vp1, "vp1"); 3483 } else { 3484 vp1_locked = true; 3485 } 3486 if (vp2 != NULL) { 3487 if (vp2_locked) 3488 ASSERT_VOP_ELOCKED(vp2, "vp2"); 3489 else 3490 ASSERT_VOP_UNLOCKED(vp2, "vp2"); 3491 } else { 3492 vp2_locked = true; 3493 } 3494 if (!vp1_locked && !vp2_locked) { 3495 vn_lock(vp1, LK_EXCLUSIVE | LK_RETRY); 3496 vp1_locked = true; 3497 } 3498 3499 for (;;) { 3500 if (vp1_locked && vp2_locked) 3501 break; 3502 if (vp1_locked && vp2 != NULL) { 3503 if (vp1 != NULL) { 3504 error = VOP_LOCK1(vp2, LK_EXCLUSIVE | LK_NOWAIT, 3505 __FILE__, __LINE__); 3506 if (error == 0) 3507 break; 3508 VOP_UNLOCK(vp1); 3509 vp1_locked = false; 3510 vn_lock_pair_pause("vlp1"); 3511 } 3512 vn_lock(vp2, LK_EXCLUSIVE | LK_RETRY); 3513 vp2_locked = true; 3514 } 3515 if (vp2_locked && vp1 != NULL) { 3516 if (vp2 != NULL) { 3517 error = VOP_LOCK1(vp1, LK_EXCLUSIVE | LK_NOWAIT, 3518 __FILE__, __LINE__); 3519 if (error == 0) 3520 break; 3521 VOP_UNLOCK(vp2); 3522 vp2_locked = false; 3523 vn_lock_pair_pause("vlp2"); 3524 } 3525 vn_lock(vp1, LK_EXCLUSIVE | LK_RETRY); 3526 vp1_locked = true; 3527 } 3528 } 3529 if (vp1 != NULL) 3530 ASSERT_VOP_ELOCKED(vp1, "vp1 ret"); 3531 if (vp2 != NULL) 3532 ASSERT_VOP_ELOCKED(vp2, "vp2 ret"); 3533 } 3534