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