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