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/sx.h> 74 #include <sys/sleepqueue.h> 75 #include <sys/sysctl.h> 76 #include <sys/ttycom.h> 77 #include <sys/conf.h> 78 #include <sys/syslog.h> 79 #include <sys/unistd.h> 80 #include <sys/user.h> 81 82 #include <security/audit/audit.h> 83 #include <security/mac/mac_framework.h> 84 85 #include <vm/vm.h> 86 #include <vm/vm_extern.h> 87 #include <vm/pmap.h> 88 #include <vm/vm_map.h> 89 #include <vm/vm_object.h> 90 #include <vm/vm_page.h> 91 #include <vm/vm_pager.h> 92 93 #ifdef HWPMC_HOOKS 94 #include <sys/pmckern.h> 95 #endif 96 97 static fo_rdwr_t vn_read; 98 static fo_rdwr_t vn_write; 99 static fo_rdwr_t vn_io_fault; 100 static fo_truncate_t vn_truncate; 101 static fo_ioctl_t vn_ioctl; 102 static fo_poll_t vn_poll; 103 static fo_kqfilter_t vn_kqfilter; 104 static fo_stat_t vn_statfile; 105 static fo_close_t vn_closefile; 106 static fo_mmap_t vn_mmap; 107 static fo_fallocate_t vn_fallocate; 108 109 struct fileops vnops = { 110 .fo_read = vn_io_fault, 111 .fo_write = vn_io_fault, 112 .fo_truncate = vn_truncate, 113 .fo_ioctl = vn_ioctl, 114 .fo_poll = vn_poll, 115 .fo_kqfilter = vn_kqfilter, 116 .fo_stat = vn_statfile, 117 .fo_close = vn_closefile, 118 .fo_chmod = vn_chmod, 119 .fo_chown = vn_chown, 120 .fo_sendfile = vn_sendfile, 121 .fo_seek = vn_seek, 122 .fo_fill_kinfo = vn_fill_kinfo, 123 .fo_mmap = vn_mmap, 124 .fo_fallocate = vn_fallocate, 125 .fo_flags = DFLAG_PASSABLE | DFLAG_SEEKABLE 126 }; 127 128 static const int io_hold_cnt = 16; 129 static int vn_io_fault_enable = 1; 130 SYSCTL_INT(_debug, OID_AUTO, vn_io_fault_enable, CTLFLAG_RW, 131 &vn_io_fault_enable, 0, "Enable vn_io_fault lock avoidance"); 132 static int vn_io_fault_prefault = 0; 133 SYSCTL_INT(_debug, OID_AUTO, vn_io_fault_prefault, CTLFLAG_RW, 134 &vn_io_fault_prefault, 0, "Enable vn_io_fault prefaulting"); 135 static u_long vn_io_faults_cnt; 136 SYSCTL_ULONG(_debug, OID_AUTO, vn_io_faults, CTLFLAG_RD, 137 &vn_io_faults_cnt, 0, "Count of vn_io_fault lock avoidance triggers"); 138 139 static int vfs_allow_read_dir = 0; 140 SYSCTL_INT(_security_bsd, OID_AUTO, allow_read_dir, CTLFLAG_RW, 141 &vfs_allow_read_dir, 0, 142 "Enable read(2) of directory by root for filesystems that support it"); 143 144 /* 145 * Returns true if vn_io_fault mode of handling the i/o request should 146 * be used. 147 */ 148 static bool 149 do_vn_io_fault(struct vnode *vp, struct uio *uio) 150 { 151 struct mount *mp; 152 153 return (uio->uio_segflg == UIO_USERSPACE && vp->v_type == VREG && 154 (mp = vp->v_mount) != NULL && 155 (mp->mnt_kern_flag & MNTK_NO_IOPF) != 0 && vn_io_fault_enable); 156 } 157 158 /* 159 * Structure used to pass arguments to vn_io_fault1(), to do either 160 * file- or vnode-based I/O calls. 161 */ 162 struct vn_io_fault_args { 163 enum { 164 VN_IO_FAULT_FOP, 165 VN_IO_FAULT_VOP 166 } kind; 167 struct ucred *cred; 168 int flags; 169 union { 170 struct fop_args_tag { 171 struct file *fp; 172 fo_rdwr_t *doio; 173 } fop_args; 174 struct vop_args_tag { 175 struct vnode *vp; 176 } vop_args; 177 } args; 178 }; 179 180 static int vn_io_fault1(struct vnode *vp, struct uio *uio, 181 struct vn_io_fault_args *args, struct thread *td); 182 183 int 184 vn_open(struct nameidata *ndp, int *flagp, int cmode, struct file *fp) 185 { 186 struct thread *td = ndp->ni_cnd.cn_thread; 187 188 return (vn_open_cred(ndp, flagp, cmode, 0, td->td_ucred, fp)); 189 } 190 191 /* 192 * Common code for vnode open operations via a name lookup. 193 * Lookup the vnode and invoke VOP_CREATE if needed. 194 * Check permissions, and call the VOP_OPEN or VOP_CREATE routine. 195 * 196 * Note that this does NOT free nameidata for the successful case, 197 * due to the NDINIT being done elsewhere. 198 */ 199 int 200 vn_open_cred(struct nameidata *ndp, int *flagp, int cmode, u_int vn_open_flags, 201 struct ucred *cred, struct file *fp) 202 { 203 struct vnode *vp; 204 struct mount *mp; 205 struct thread *td = ndp->ni_cnd.cn_thread; 206 struct vattr vat; 207 struct vattr *vap = &vat; 208 int fmode, error; 209 210 restart: 211 fmode = *flagp; 212 if ((fmode & (O_CREAT | O_EXCL | O_DIRECTORY)) == (O_CREAT | 213 O_EXCL | O_DIRECTORY)) 214 return (EINVAL); 215 else if ((fmode & (O_CREAT | O_DIRECTORY)) == O_CREAT) { 216 ndp->ni_cnd.cn_nameiop = CREATE; 217 /* 218 * Set NOCACHE to avoid flushing the cache when 219 * rolling in many files at once. 220 */ 221 ndp->ni_cnd.cn_flags = ISOPEN | LOCKPARENT | LOCKLEAF | NOCACHE; 222 if ((fmode & O_EXCL) == 0 && (fmode & O_NOFOLLOW) == 0) 223 ndp->ni_cnd.cn_flags |= FOLLOW; 224 if ((fmode & O_BENEATH) != 0) 225 ndp->ni_cnd.cn_flags |= BENEATH; 226 if (!(vn_open_flags & VN_OPEN_NOAUDIT)) 227 ndp->ni_cnd.cn_flags |= AUDITVNODE1; 228 if (vn_open_flags & VN_OPEN_NOCAPCHECK) 229 ndp->ni_cnd.cn_flags |= NOCAPCHECK; 230 if ((vn_open_flags & VN_OPEN_INVFS) == 0) 231 bwillwrite(); 232 if ((error = namei(ndp)) != 0) 233 return (error); 234 if (ndp->ni_vp == NULL) { 235 VATTR_NULL(vap); 236 vap->va_type = VREG; 237 vap->va_mode = cmode; 238 if (fmode & O_EXCL) 239 vap->va_vaflags |= VA_EXCLUSIVE; 240 if (vn_start_write(ndp->ni_dvp, &mp, V_NOWAIT) != 0) { 241 NDFREE(ndp, NDF_ONLY_PNBUF); 242 vput(ndp->ni_dvp); 243 if ((error = vn_start_write(NULL, &mp, 244 V_XSLEEP | PCATCH)) != 0) 245 return (error); 246 goto restart; 247 } 248 if ((vn_open_flags & VN_OPEN_NAMECACHE) != 0) 249 ndp->ni_cnd.cn_flags |= MAKEENTRY; 250 #ifdef MAC 251 error = mac_vnode_check_create(cred, ndp->ni_dvp, 252 &ndp->ni_cnd, vap); 253 if (error == 0) 254 #endif 255 error = VOP_CREATE(ndp->ni_dvp, &ndp->ni_vp, 256 &ndp->ni_cnd, vap); 257 vput(ndp->ni_dvp); 258 vn_finished_write(mp); 259 if (error) { 260 NDFREE(ndp, NDF_ONLY_PNBUF); 261 return (error); 262 } 263 fmode &= ~O_TRUNC; 264 vp = ndp->ni_vp; 265 } else { 266 if (ndp->ni_dvp == ndp->ni_vp) 267 vrele(ndp->ni_dvp); 268 else 269 vput(ndp->ni_dvp); 270 ndp->ni_dvp = NULL; 271 vp = ndp->ni_vp; 272 if (fmode & O_EXCL) { 273 error = EEXIST; 274 goto bad; 275 } 276 if (vp->v_type == VDIR) { 277 error = EISDIR; 278 goto bad; 279 } 280 fmode &= ~O_CREAT; 281 } 282 } else { 283 ndp->ni_cnd.cn_nameiop = LOOKUP; 284 ndp->ni_cnd.cn_flags = ISOPEN | 285 ((fmode & O_NOFOLLOW) ? NOFOLLOW : FOLLOW) | LOCKLEAF; 286 if (!(fmode & FWRITE)) 287 ndp->ni_cnd.cn_flags |= LOCKSHARED; 288 if ((fmode & O_BENEATH) != 0) 289 ndp->ni_cnd.cn_flags |= BENEATH; 290 if (!(vn_open_flags & VN_OPEN_NOAUDIT)) 291 ndp->ni_cnd.cn_flags |= AUDITVNODE1; 292 if (vn_open_flags & VN_OPEN_NOCAPCHECK) 293 ndp->ni_cnd.cn_flags |= NOCAPCHECK; 294 if ((error = namei(ndp)) != 0) 295 return (error); 296 vp = ndp->ni_vp; 297 } 298 error = vn_open_vnode(vp, fmode, cred, td, fp); 299 if (error) 300 goto bad; 301 *flagp = fmode; 302 return (0); 303 bad: 304 NDFREE(ndp, NDF_ONLY_PNBUF); 305 vput(vp); 306 *flagp = fmode; 307 ndp->ni_vp = NULL; 308 return (error); 309 } 310 311 static int 312 vn_open_vnode_advlock(struct vnode *vp, int fmode, struct file *fp) 313 { 314 struct flock lf; 315 int error, lock_flags, type; 316 317 ASSERT_VOP_LOCKED(vp, "vn_open_vnode_advlock"); 318 if ((fmode & (O_EXLOCK | O_SHLOCK)) == 0) 319 return (0); 320 KASSERT(fp != NULL, ("open with flock requires fp")); 321 if (fp->f_type != DTYPE_NONE && fp->f_type != DTYPE_VNODE) 322 return (EOPNOTSUPP); 323 324 lock_flags = VOP_ISLOCKED(vp); 325 VOP_UNLOCK(vp); 326 327 lf.l_whence = SEEK_SET; 328 lf.l_start = 0; 329 lf.l_len = 0; 330 lf.l_type = (fmode & O_EXLOCK) != 0 ? F_WRLCK : F_RDLCK; 331 type = F_FLOCK; 332 if ((fmode & FNONBLOCK) == 0) 333 type |= F_WAIT; 334 error = VOP_ADVLOCK(vp, (caddr_t)fp, F_SETLK, &lf, type); 335 if (error == 0) 336 fp->f_flag |= FHASLOCK; 337 338 vn_lock(vp, lock_flags | LK_RETRY); 339 if (error == 0 && VN_IS_DOOMED(vp)) 340 error = ENOENT; 341 return (error); 342 } 343 344 /* 345 * Common code for vnode open operations once a vnode is located. 346 * Check permissions, and call the VOP_OPEN routine. 347 */ 348 int 349 vn_open_vnode(struct vnode *vp, int fmode, struct ucred *cred, 350 struct thread *td, struct file *fp) 351 { 352 accmode_t accmode; 353 int error; 354 355 if (vp->v_type == VLNK) 356 return (EMLINK); 357 if (vp->v_type == VSOCK) 358 return (EOPNOTSUPP); 359 if (vp->v_type != VDIR && fmode & O_DIRECTORY) 360 return (ENOTDIR); 361 accmode = 0; 362 if (fmode & (FWRITE | O_TRUNC)) { 363 if (vp->v_type == VDIR) 364 return (EISDIR); 365 accmode |= VWRITE; 366 } 367 if (fmode & FREAD) 368 accmode |= VREAD; 369 if (fmode & FEXEC) 370 accmode |= VEXEC; 371 if ((fmode & O_APPEND) && (fmode & FWRITE)) 372 accmode |= VAPPEND; 373 #ifdef MAC 374 if (fmode & O_CREAT) 375 accmode |= VCREAT; 376 if (fmode & O_VERIFY) 377 accmode |= VVERIFY; 378 error = mac_vnode_check_open(cred, vp, accmode); 379 if (error) 380 return (error); 381 382 accmode &= ~(VCREAT | VVERIFY); 383 #endif 384 if ((fmode & O_CREAT) == 0 && accmode != 0) { 385 error = VOP_ACCESS(vp, accmode, cred, td); 386 if (error != 0) 387 return (error); 388 } 389 if (vp->v_type == VFIFO && VOP_ISLOCKED(vp) != LK_EXCLUSIVE) 390 vn_lock(vp, LK_UPGRADE | LK_RETRY); 391 error = VOP_OPEN(vp, fmode, cred, td, fp); 392 if (error != 0) 393 return (error); 394 395 error = vn_open_vnode_advlock(vp, fmode, fp); 396 if (error == 0 && (fmode & FWRITE) != 0) { 397 error = VOP_ADD_WRITECOUNT(vp, 1); 398 if (error == 0) { 399 CTR3(KTR_VFS, "%s: vp %p v_writecount increased to %d", 400 __func__, vp, vp->v_writecount); 401 } 402 } 403 404 /* 405 * Error from advlock or VOP_ADD_WRITECOUNT() still requires 406 * calling VOP_CLOSE() to pair with earlier VOP_OPEN(). 407 * Arrange for that by having fdrop() to use vn_closefile(). 408 */ 409 if (error != 0) { 410 fp->f_flag |= FOPENFAILED; 411 fp->f_vnode = vp; 412 if (fp->f_ops == &badfileops) { 413 fp->f_type = DTYPE_VNODE; 414 fp->f_ops = &vnops; 415 } 416 vref(vp); 417 } 418 419 ASSERT_VOP_LOCKED(vp, "vn_open_vnode"); 420 return (error); 421 422 } 423 424 /* 425 * Check for write permissions on the specified vnode. 426 * Prototype text segments cannot be written. 427 * It is racy. 428 */ 429 int 430 vn_writechk(struct vnode *vp) 431 { 432 433 ASSERT_VOP_LOCKED(vp, "vn_writechk"); 434 /* 435 * If there's shared text associated with 436 * the vnode, try to free it up once. If 437 * we fail, we can't allow writing. 438 */ 439 if (VOP_IS_TEXT(vp)) 440 return (ETXTBSY); 441 442 return (0); 443 } 444 445 /* 446 * Vnode close call 447 */ 448 static int 449 vn_close1(struct vnode *vp, int flags, struct ucred *file_cred, 450 struct thread *td, bool keep_ref) 451 { 452 struct mount *mp; 453 int error, lock_flags; 454 455 if (vp->v_type != VFIFO && (flags & FWRITE) == 0 && 456 MNT_EXTENDED_SHARED(vp->v_mount)) 457 lock_flags = LK_SHARED; 458 else 459 lock_flags = LK_EXCLUSIVE; 460 461 vn_start_write(vp, &mp, V_WAIT); 462 vn_lock(vp, lock_flags | LK_RETRY); 463 AUDIT_ARG_VNODE1(vp); 464 if ((flags & (FWRITE | FOPENFAILED)) == FWRITE) { 465 VOP_ADD_WRITECOUNT_CHECKED(vp, -1); 466 CTR3(KTR_VFS, "%s: vp %p v_writecount decreased to %d", 467 __func__, vp, vp->v_writecount); 468 } 469 error = VOP_CLOSE(vp, flags, file_cred, td); 470 if (keep_ref) 471 VOP_UNLOCK(vp); 472 else 473 vput(vp); 474 vn_finished_write(mp); 475 return (error); 476 } 477 478 int 479 vn_close(struct vnode *vp, int flags, struct ucred *file_cred, 480 struct thread *td) 481 { 482 483 return (vn_close1(vp, flags, file_cred, td, false)); 484 } 485 486 /* 487 * Heuristic to detect sequential operation. 488 */ 489 static int 490 sequential_heuristic(struct uio *uio, struct file *fp) 491 { 492 enum uio_rw rw; 493 494 ASSERT_VOP_LOCKED(fp->f_vnode, __func__); 495 496 rw = uio->uio_rw; 497 if (fp->f_flag & FRDAHEAD) 498 return (fp->f_seqcount[rw] << IO_SEQSHIFT); 499 500 /* 501 * Offset 0 is handled specially. open() sets f_seqcount to 1 so 502 * that the first I/O is normally considered to be slightly 503 * sequential. Seeking to offset 0 doesn't change sequentiality 504 * unless previous seeks have reduced f_seqcount to 0, in which 505 * case offset 0 is not special. 506 */ 507 if ((uio->uio_offset == 0 && fp->f_seqcount[rw] > 0) || 508 uio->uio_offset == fp->f_nextoff[rw]) { 509 /* 510 * f_seqcount is in units of fixed-size blocks so that it 511 * depends mainly on the amount of sequential I/O and not 512 * much on the number of sequential I/O's. The fixed size 513 * of 16384 is hard-coded here since it is (not quite) just 514 * a magic size that works well here. This size is more 515 * closely related to the best I/O size for real disks than 516 * to any block size used by software. 517 */ 518 if (uio->uio_resid >= IO_SEQMAX * 16384) 519 fp->f_seqcount[rw] = IO_SEQMAX; 520 else { 521 fp->f_seqcount[rw] += howmany(uio->uio_resid, 16384); 522 if (fp->f_seqcount[rw] > IO_SEQMAX) 523 fp->f_seqcount[rw] = IO_SEQMAX; 524 } 525 return (fp->f_seqcount[rw] << IO_SEQSHIFT); 526 } 527 528 /* Not sequential. Quickly draw-down sequentiality. */ 529 if (fp->f_seqcount[rw] > 1) 530 fp->f_seqcount[rw] = 1; 531 else 532 fp->f_seqcount[rw] = 0; 533 return (0); 534 } 535 536 /* 537 * Package up an I/O request on a vnode into a uio and do it. 538 */ 539 int 540 vn_rdwr(enum uio_rw rw, struct vnode *vp, void *base, int len, off_t offset, 541 enum uio_seg segflg, int ioflg, struct ucred *active_cred, 542 struct ucred *file_cred, ssize_t *aresid, struct thread *td) 543 { 544 struct uio auio; 545 struct iovec aiov; 546 struct mount *mp; 547 struct ucred *cred; 548 void *rl_cookie; 549 struct vn_io_fault_args args; 550 int error, lock_flags; 551 552 if (offset < 0 && vp->v_type != VCHR) 553 return (EINVAL); 554 auio.uio_iov = &aiov; 555 auio.uio_iovcnt = 1; 556 aiov.iov_base = base; 557 aiov.iov_len = len; 558 auio.uio_resid = len; 559 auio.uio_offset = offset; 560 auio.uio_segflg = segflg; 561 auio.uio_rw = rw; 562 auio.uio_td = td; 563 error = 0; 564 565 if ((ioflg & IO_NODELOCKED) == 0) { 566 if ((ioflg & IO_RANGELOCKED) == 0) { 567 if (rw == UIO_READ) { 568 rl_cookie = vn_rangelock_rlock(vp, offset, 569 offset + len); 570 } else { 571 rl_cookie = vn_rangelock_wlock(vp, offset, 572 offset + len); 573 } 574 } else 575 rl_cookie = NULL; 576 mp = NULL; 577 if (rw == UIO_WRITE) { 578 if (vp->v_type != VCHR && 579 (error = vn_start_write(vp, &mp, V_WAIT | PCATCH)) 580 != 0) 581 goto out; 582 if (MNT_SHARED_WRITES(mp) || 583 ((mp == NULL) && MNT_SHARED_WRITES(vp->v_mount))) 584 lock_flags = LK_SHARED; 585 else 586 lock_flags = LK_EXCLUSIVE; 587 } else 588 lock_flags = LK_SHARED; 589 vn_lock(vp, lock_flags | LK_RETRY); 590 } else 591 rl_cookie = NULL; 592 593 ASSERT_VOP_LOCKED(vp, "IO_NODELOCKED with no vp lock held"); 594 #ifdef MAC 595 if ((ioflg & IO_NOMACCHECK) == 0) { 596 if (rw == UIO_READ) 597 error = mac_vnode_check_read(active_cred, file_cred, 598 vp); 599 else 600 error = mac_vnode_check_write(active_cred, file_cred, 601 vp); 602 } 603 #endif 604 if (error == 0) { 605 if (file_cred != NULL) 606 cred = file_cred; 607 else 608 cred = active_cred; 609 if (do_vn_io_fault(vp, &auio)) { 610 args.kind = VN_IO_FAULT_VOP; 611 args.cred = cred; 612 args.flags = ioflg; 613 args.args.vop_args.vp = vp; 614 error = vn_io_fault1(vp, &auio, &args, td); 615 } else if (rw == UIO_READ) { 616 error = VOP_READ(vp, &auio, ioflg, cred); 617 } else /* if (rw == UIO_WRITE) */ { 618 error = VOP_WRITE(vp, &auio, ioflg, cred); 619 } 620 } 621 if (aresid) 622 *aresid = auio.uio_resid; 623 else 624 if (auio.uio_resid && error == 0) 625 error = EIO; 626 if ((ioflg & IO_NODELOCKED) == 0) { 627 VOP_UNLOCK(vp); 628 if (mp != NULL) 629 vn_finished_write(mp); 630 } 631 out: 632 if (rl_cookie != NULL) 633 vn_rangelock_unlock(vp, rl_cookie); 634 return (error); 635 } 636 637 /* 638 * Package up an I/O request on a vnode into a uio and do it. The I/O 639 * request is split up into smaller chunks and we try to avoid saturating 640 * the buffer cache while potentially holding a vnode locked, so we 641 * check bwillwrite() before calling vn_rdwr(). We also call kern_yield() 642 * to give other processes a chance to lock the vnode (either other processes 643 * core'ing the same binary, or unrelated processes scanning the directory). 644 */ 645 int 646 vn_rdwr_inchunks(enum uio_rw rw, struct vnode *vp, void *base, size_t len, 647 off_t offset, enum uio_seg segflg, int ioflg, struct ucred *active_cred, 648 struct ucred *file_cred, size_t *aresid, struct thread *td) 649 { 650 int error = 0; 651 ssize_t iaresid; 652 653 do { 654 int chunk; 655 656 /* 657 * Force `offset' to a multiple of MAXBSIZE except possibly 658 * for the first chunk, so that filesystems only need to 659 * write full blocks except possibly for the first and last 660 * chunks. 661 */ 662 chunk = MAXBSIZE - (uoff_t)offset % MAXBSIZE; 663 664 if (chunk > len) 665 chunk = len; 666 if (rw != UIO_READ && vp->v_type == VREG) 667 bwillwrite(); 668 iaresid = 0; 669 error = vn_rdwr(rw, vp, base, chunk, offset, segflg, 670 ioflg, active_cred, file_cred, &iaresid, td); 671 len -= chunk; /* aresid calc already includes length */ 672 if (error) 673 break; 674 offset += chunk; 675 base = (char *)base + chunk; 676 kern_yield(PRI_USER); 677 } while (len); 678 if (aresid) 679 *aresid = len + iaresid; 680 return (error); 681 } 682 683 #if OFF_MAX <= LONG_MAX 684 off_t 685 foffset_lock(struct file *fp, int flags) 686 { 687 volatile short *flagsp; 688 off_t res; 689 short state; 690 691 KASSERT((flags & FOF_OFFSET) == 0, ("FOF_OFFSET passed")); 692 693 if ((flags & FOF_NOLOCK) != 0) 694 return (atomic_load_long(&fp->f_offset)); 695 696 /* 697 * According to McKusick the vn lock was protecting f_offset here. 698 * It is now protected by the FOFFSET_LOCKED flag. 699 */ 700 flagsp = &fp->f_vnread_flags; 701 if (atomic_cmpset_acq_16(flagsp, 0, FOFFSET_LOCKED)) 702 return (atomic_load_long(&fp->f_offset)); 703 704 sleepq_lock(&fp->f_vnread_flags); 705 state = atomic_load_16(flagsp); 706 for (;;) { 707 if ((state & FOFFSET_LOCKED) == 0) { 708 if (!atomic_fcmpset_acq_16(flagsp, &state, 709 FOFFSET_LOCKED)) 710 continue; 711 break; 712 } 713 if ((state & FOFFSET_LOCK_WAITING) == 0) { 714 if (!atomic_fcmpset_acq_16(flagsp, &state, 715 state | FOFFSET_LOCK_WAITING)) 716 continue; 717 } 718 DROP_GIANT(); 719 sleepq_add(&fp->f_vnread_flags, NULL, "vofflock", 0, 0); 720 sleepq_wait(&fp->f_vnread_flags, PUSER -1); 721 PICKUP_GIANT(); 722 sleepq_lock(&fp->f_vnread_flags); 723 state = atomic_load_16(flagsp); 724 } 725 res = atomic_load_long(&fp->f_offset); 726 sleepq_release(&fp->f_vnread_flags); 727 return (res); 728 } 729 730 void 731 foffset_unlock(struct file *fp, off_t val, int flags) 732 { 733 volatile short *flagsp; 734 short state; 735 736 KASSERT((flags & FOF_OFFSET) == 0, ("FOF_OFFSET passed")); 737 738 if ((flags & FOF_NOUPDATE) == 0) 739 atomic_store_long(&fp->f_offset, val); 740 if ((flags & FOF_NEXTOFF_R) != 0) 741 fp->f_nextoff[UIO_READ] = val; 742 if ((flags & FOF_NEXTOFF_W) != 0) 743 fp->f_nextoff[UIO_WRITE] = val; 744 745 if ((flags & FOF_NOLOCK) != 0) 746 return; 747 748 flagsp = &fp->f_vnread_flags; 749 state = atomic_load_16(flagsp); 750 if ((state & FOFFSET_LOCK_WAITING) == 0 && 751 atomic_cmpset_rel_16(flagsp, state, 0)) 752 return; 753 754 sleepq_lock(&fp->f_vnread_flags); 755 MPASS((fp->f_vnread_flags & FOFFSET_LOCKED) != 0); 756 MPASS((fp->f_vnread_flags & FOFFSET_LOCK_WAITING) != 0); 757 fp->f_vnread_flags = 0; 758 sleepq_broadcast(&fp->f_vnread_flags, SLEEPQ_SLEEP, 0, 0); 759 sleepq_release(&fp->f_vnread_flags); 760 } 761 #else 762 off_t 763 foffset_lock(struct file *fp, int flags) 764 { 765 struct mtx *mtxp; 766 off_t res; 767 768 KASSERT((flags & FOF_OFFSET) == 0, ("FOF_OFFSET passed")); 769 770 mtxp = mtx_pool_find(mtxpool_sleep, fp); 771 mtx_lock(mtxp); 772 if ((flags & FOF_NOLOCK) == 0) { 773 while (fp->f_vnread_flags & FOFFSET_LOCKED) { 774 fp->f_vnread_flags |= FOFFSET_LOCK_WAITING; 775 msleep(&fp->f_vnread_flags, mtxp, PUSER -1, 776 "vofflock", 0); 777 } 778 fp->f_vnread_flags |= FOFFSET_LOCKED; 779 } 780 res = fp->f_offset; 781 mtx_unlock(mtxp); 782 return (res); 783 } 784 785 void 786 foffset_unlock(struct file *fp, off_t val, int flags) 787 { 788 struct mtx *mtxp; 789 790 KASSERT((flags & FOF_OFFSET) == 0, ("FOF_OFFSET passed")); 791 792 mtxp = mtx_pool_find(mtxpool_sleep, fp); 793 mtx_lock(mtxp); 794 if ((flags & FOF_NOUPDATE) == 0) 795 fp->f_offset = val; 796 if ((flags & FOF_NEXTOFF_R) != 0) 797 fp->f_nextoff[UIO_READ] = val; 798 if ((flags & FOF_NEXTOFF_W) != 0) 799 fp->f_nextoff[UIO_WRITE] = val; 800 if ((flags & FOF_NOLOCK) == 0) { 801 KASSERT((fp->f_vnread_flags & FOFFSET_LOCKED) != 0, 802 ("Lost FOFFSET_LOCKED")); 803 if (fp->f_vnread_flags & FOFFSET_LOCK_WAITING) 804 wakeup(&fp->f_vnread_flags); 805 fp->f_vnread_flags = 0; 806 } 807 mtx_unlock(mtxp); 808 } 809 #endif 810 811 void 812 foffset_lock_uio(struct file *fp, struct uio *uio, int flags) 813 { 814 815 if ((flags & FOF_OFFSET) == 0) 816 uio->uio_offset = foffset_lock(fp, flags); 817 } 818 819 void 820 foffset_unlock_uio(struct file *fp, struct uio *uio, int flags) 821 { 822 823 if ((flags & FOF_OFFSET) == 0) 824 foffset_unlock(fp, uio->uio_offset, flags); 825 } 826 827 static int 828 get_advice(struct file *fp, struct uio *uio) 829 { 830 struct mtx *mtxp; 831 int ret; 832 833 ret = POSIX_FADV_NORMAL; 834 if (fp->f_advice == NULL || fp->f_vnode->v_type != VREG) 835 return (ret); 836 837 mtxp = mtx_pool_find(mtxpool_sleep, fp); 838 mtx_lock(mtxp); 839 if (fp->f_advice != NULL && 840 uio->uio_offset >= fp->f_advice->fa_start && 841 uio->uio_offset + uio->uio_resid <= fp->f_advice->fa_end) 842 ret = fp->f_advice->fa_advice; 843 mtx_unlock(mtxp); 844 return (ret); 845 } 846 847 /* 848 * File table vnode read routine. 849 */ 850 static int 851 vn_read(struct file *fp, struct uio *uio, struct ucred *active_cred, int flags, 852 struct thread *td) 853 { 854 struct vnode *vp; 855 off_t orig_offset; 856 int error, ioflag; 857 int advice; 858 859 KASSERT(uio->uio_td == td, ("uio_td %p is not td %p", 860 uio->uio_td, td)); 861 KASSERT(flags & FOF_OFFSET, ("No FOF_OFFSET")); 862 vp = fp->f_vnode; 863 ioflag = 0; 864 if (fp->f_flag & FNONBLOCK) 865 ioflag |= IO_NDELAY; 866 if (fp->f_flag & O_DIRECT) 867 ioflag |= IO_DIRECT; 868 advice = get_advice(fp, uio); 869 vn_lock(vp, LK_SHARED | LK_RETRY); 870 871 switch (advice) { 872 case POSIX_FADV_NORMAL: 873 case POSIX_FADV_SEQUENTIAL: 874 case POSIX_FADV_NOREUSE: 875 ioflag |= sequential_heuristic(uio, fp); 876 break; 877 case POSIX_FADV_RANDOM: 878 /* Disable read-ahead for random I/O. */ 879 break; 880 } 881 orig_offset = uio->uio_offset; 882 883 #ifdef MAC 884 error = mac_vnode_check_read(active_cred, fp->f_cred, vp); 885 if (error == 0) 886 #endif 887 error = VOP_READ(vp, uio, ioflag, fp->f_cred); 888 fp->f_nextoff[UIO_READ] = uio->uio_offset; 889 VOP_UNLOCK(vp); 890 if (error == 0 && advice == POSIX_FADV_NOREUSE && 891 orig_offset != uio->uio_offset) 892 /* 893 * Use POSIX_FADV_DONTNEED to flush pages and buffers 894 * for the backing file after a POSIX_FADV_NOREUSE 895 * read(2). 896 */ 897 error = VOP_ADVISE(vp, orig_offset, uio->uio_offset - 1, 898 POSIX_FADV_DONTNEED); 899 return (error); 900 } 901 902 /* 903 * File table vnode write routine. 904 */ 905 static int 906 vn_write(struct file *fp, struct uio *uio, struct ucred *active_cred, int flags, 907 struct thread *td) 908 { 909 struct vnode *vp; 910 struct mount *mp; 911 off_t orig_offset; 912 int error, ioflag, lock_flags; 913 int advice; 914 915 KASSERT(uio->uio_td == td, ("uio_td %p is not td %p", 916 uio->uio_td, td)); 917 KASSERT(flags & FOF_OFFSET, ("No FOF_OFFSET")); 918 vp = fp->f_vnode; 919 if (vp->v_type == VREG) 920 bwillwrite(); 921 ioflag = IO_UNIT; 922 if (vp->v_type == VREG && (fp->f_flag & O_APPEND)) 923 ioflag |= IO_APPEND; 924 if (fp->f_flag & FNONBLOCK) 925 ioflag |= IO_NDELAY; 926 if (fp->f_flag & O_DIRECT) 927 ioflag |= IO_DIRECT; 928 if ((fp->f_flag & O_FSYNC) || 929 (vp->v_mount && (vp->v_mount->mnt_flag & MNT_SYNCHRONOUS))) 930 ioflag |= IO_SYNC; 931 mp = NULL; 932 if (vp->v_type != VCHR && 933 (error = vn_start_write(vp, &mp, V_WAIT | PCATCH)) != 0) 934 goto unlock; 935 936 advice = get_advice(fp, uio); 937 938 if (MNT_SHARED_WRITES(mp) || 939 (mp == NULL && MNT_SHARED_WRITES(vp->v_mount))) { 940 lock_flags = LK_SHARED; 941 } else { 942 lock_flags = LK_EXCLUSIVE; 943 } 944 945 vn_lock(vp, lock_flags | LK_RETRY); 946 switch (advice) { 947 case POSIX_FADV_NORMAL: 948 case POSIX_FADV_SEQUENTIAL: 949 case POSIX_FADV_NOREUSE: 950 ioflag |= sequential_heuristic(uio, fp); 951 break; 952 case POSIX_FADV_RANDOM: 953 /* XXX: Is this correct? */ 954 break; 955 } 956 orig_offset = uio->uio_offset; 957 958 #ifdef MAC 959 error = mac_vnode_check_write(active_cred, fp->f_cred, vp); 960 if (error == 0) 961 #endif 962 error = VOP_WRITE(vp, uio, ioflag, fp->f_cred); 963 fp->f_nextoff[UIO_WRITE] = uio->uio_offset; 964 VOP_UNLOCK(vp); 965 if (vp->v_type != VCHR) 966 vn_finished_write(mp); 967 if (error == 0 && advice == POSIX_FADV_NOREUSE && 968 orig_offset != uio->uio_offset) 969 /* 970 * Use POSIX_FADV_DONTNEED to flush pages and buffers 971 * for the backing file after a POSIX_FADV_NOREUSE 972 * write(2). 973 */ 974 error = VOP_ADVISE(vp, orig_offset, uio->uio_offset - 1, 975 POSIX_FADV_DONTNEED); 976 unlock: 977 return (error); 978 } 979 980 /* 981 * The vn_io_fault() is a wrapper around vn_read() and vn_write() to 982 * prevent the following deadlock: 983 * 984 * Assume that the thread A reads from the vnode vp1 into userspace 985 * buffer buf1 backed by the pages of vnode vp2. If a page in buf1 is 986 * currently not resident, then system ends up with the call chain 987 * vn_read() -> VOP_READ(vp1) -> uiomove() -> [Page Fault] -> 988 * vm_fault(buf1) -> vnode_pager_getpages(vp2) -> VOP_GETPAGES(vp2) 989 * which establishes lock order vp1->vn_lock, then vp2->vn_lock. 990 * If, at the same time, thread B reads from vnode vp2 into buffer buf2 991 * backed by the pages of vnode vp1, and some page in buf2 is not 992 * resident, we get a reversed order vp2->vn_lock, then vp1->vn_lock. 993 * 994 * To prevent the lock order reversal and deadlock, vn_io_fault() does 995 * not allow page faults to happen during VOP_READ() or VOP_WRITE(). 996 * Instead, it first tries to do the whole range i/o with pagefaults 997 * disabled. If all pages in the i/o buffer are resident and mapped, 998 * VOP will succeed (ignoring the genuine filesystem errors). 999 * Otherwise, we get back EFAULT, and vn_io_fault() falls back to do 1000 * i/o in chunks, with all pages in the chunk prefaulted and held 1001 * using vm_fault_quick_hold_pages(). 1002 * 1003 * Filesystems using this deadlock avoidance scheme should use the 1004 * array of the held pages from uio, saved in the curthread->td_ma, 1005 * instead of doing uiomove(). A helper function 1006 * vn_io_fault_uiomove() converts uiomove request into 1007 * uiomove_fromphys() over td_ma array. 1008 * 1009 * Since vnode locks do not cover the whole i/o anymore, rangelocks 1010 * make the current i/o request atomic with respect to other i/os and 1011 * truncations. 1012 */ 1013 1014 /* 1015 * Decode vn_io_fault_args and perform the corresponding i/o. 1016 */ 1017 static int 1018 vn_io_fault_doio(struct vn_io_fault_args *args, struct uio *uio, 1019 struct thread *td) 1020 { 1021 int error, save; 1022 1023 error = 0; 1024 save = vm_fault_disable_pagefaults(); 1025 switch (args->kind) { 1026 case VN_IO_FAULT_FOP: 1027 error = (args->args.fop_args.doio)(args->args.fop_args.fp, 1028 uio, args->cred, args->flags, td); 1029 break; 1030 case VN_IO_FAULT_VOP: 1031 if (uio->uio_rw == UIO_READ) { 1032 error = VOP_READ(args->args.vop_args.vp, uio, 1033 args->flags, args->cred); 1034 } else if (uio->uio_rw == UIO_WRITE) { 1035 error = VOP_WRITE(args->args.vop_args.vp, uio, 1036 args->flags, args->cred); 1037 } 1038 break; 1039 default: 1040 panic("vn_io_fault_doio: unknown kind of io %d %d", 1041 args->kind, uio->uio_rw); 1042 } 1043 vm_fault_enable_pagefaults(save); 1044 return (error); 1045 } 1046 1047 static int 1048 vn_io_fault_touch(char *base, const struct uio *uio) 1049 { 1050 int r; 1051 1052 r = fubyte(base); 1053 if (r == -1 || (uio->uio_rw == UIO_READ && subyte(base, r) == -1)) 1054 return (EFAULT); 1055 return (0); 1056 } 1057 1058 static int 1059 vn_io_fault_prefault_user(const struct uio *uio) 1060 { 1061 char *base; 1062 const struct iovec *iov; 1063 size_t len; 1064 ssize_t resid; 1065 int error, i; 1066 1067 KASSERT(uio->uio_segflg == UIO_USERSPACE, 1068 ("vn_io_fault_prefault userspace")); 1069 1070 error = i = 0; 1071 iov = uio->uio_iov; 1072 resid = uio->uio_resid; 1073 base = iov->iov_base; 1074 len = iov->iov_len; 1075 while (resid > 0) { 1076 error = vn_io_fault_touch(base, uio); 1077 if (error != 0) 1078 break; 1079 if (len < PAGE_SIZE) { 1080 if (len != 0) { 1081 error = vn_io_fault_touch(base + len - 1, uio); 1082 if (error != 0) 1083 break; 1084 resid -= len; 1085 } 1086 if (++i >= uio->uio_iovcnt) 1087 break; 1088 iov = uio->uio_iov + i; 1089 base = iov->iov_base; 1090 len = iov->iov_len; 1091 } else { 1092 len -= PAGE_SIZE; 1093 base += PAGE_SIZE; 1094 resid -= PAGE_SIZE; 1095 } 1096 } 1097 return (error); 1098 } 1099 1100 /* 1101 * Common code for vn_io_fault(), agnostic to the kind of i/o request. 1102 * Uses vn_io_fault_doio() to make the call to an actual i/o function. 1103 * Used from vn_rdwr() and vn_io_fault(), which encode the i/o request 1104 * into args and call vn_io_fault1() to handle faults during the user 1105 * mode buffer accesses. 1106 */ 1107 static int 1108 vn_io_fault1(struct vnode *vp, struct uio *uio, struct vn_io_fault_args *args, 1109 struct thread *td) 1110 { 1111 vm_page_t ma[io_hold_cnt + 2]; 1112 struct uio *uio_clone, short_uio; 1113 struct iovec short_iovec[1]; 1114 vm_page_t *prev_td_ma; 1115 vm_prot_t prot; 1116 vm_offset_t addr, end; 1117 size_t len, resid; 1118 ssize_t adv; 1119 int error, cnt, saveheld, prev_td_ma_cnt; 1120 1121 if (vn_io_fault_prefault) { 1122 error = vn_io_fault_prefault_user(uio); 1123 if (error != 0) 1124 return (error); /* Or ignore ? */ 1125 } 1126 1127 prot = uio->uio_rw == UIO_READ ? VM_PROT_WRITE : VM_PROT_READ; 1128 1129 /* 1130 * The UFS follows IO_UNIT directive and replays back both 1131 * uio_offset and uio_resid if an error is encountered during the 1132 * operation. But, since the iovec may be already advanced, 1133 * uio is still in an inconsistent state. 1134 * 1135 * Cache a copy of the original uio, which is advanced to the redo 1136 * point using UIO_NOCOPY below. 1137 */ 1138 uio_clone = cloneuio(uio); 1139 resid = uio->uio_resid; 1140 1141 short_uio.uio_segflg = UIO_USERSPACE; 1142 short_uio.uio_rw = uio->uio_rw; 1143 short_uio.uio_td = uio->uio_td; 1144 1145 error = vn_io_fault_doio(args, uio, td); 1146 if (error != EFAULT) 1147 goto out; 1148 1149 atomic_add_long(&vn_io_faults_cnt, 1); 1150 uio_clone->uio_segflg = UIO_NOCOPY; 1151 uiomove(NULL, resid - uio->uio_resid, uio_clone); 1152 uio_clone->uio_segflg = uio->uio_segflg; 1153 1154 saveheld = curthread_pflags_set(TDP_UIOHELD); 1155 prev_td_ma = td->td_ma; 1156 prev_td_ma_cnt = td->td_ma_cnt; 1157 1158 while (uio_clone->uio_resid != 0) { 1159 len = uio_clone->uio_iov->iov_len; 1160 if (len == 0) { 1161 KASSERT(uio_clone->uio_iovcnt >= 1, 1162 ("iovcnt underflow")); 1163 uio_clone->uio_iov++; 1164 uio_clone->uio_iovcnt--; 1165 continue; 1166 } 1167 if (len > io_hold_cnt * PAGE_SIZE) 1168 len = io_hold_cnt * PAGE_SIZE; 1169 addr = (uintptr_t)uio_clone->uio_iov->iov_base; 1170 end = round_page(addr + len); 1171 if (end < addr) { 1172 error = EFAULT; 1173 break; 1174 } 1175 cnt = atop(end - trunc_page(addr)); 1176 /* 1177 * A perfectly misaligned address and length could cause 1178 * both the start and the end of the chunk to use partial 1179 * page. +2 accounts for such a situation. 1180 */ 1181 cnt = vm_fault_quick_hold_pages(&td->td_proc->p_vmspace->vm_map, 1182 addr, len, prot, ma, io_hold_cnt + 2); 1183 if (cnt == -1) { 1184 error = EFAULT; 1185 break; 1186 } 1187 short_uio.uio_iov = &short_iovec[0]; 1188 short_iovec[0].iov_base = (void *)addr; 1189 short_uio.uio_iovcnt = 1; 1190 short_uio.uio_resid = short_iovec[0].iov_len = len; 1191 short_uio.uio_offset = uio_clone->uio_offset; 1192 td->td_ma = ma; 1193 td->td_ma_cnt = cnt; 1194 1195 error = vn_io_fault_doio(args, &short_uio, td); 1196 vm_page_unhold_pages(ma, cnt); 1197 adv = len - short_uio.uio_resid; 1198 1199 uio_clone->uio_iov->iov_base = 1200 (char *)uio_clone->uio_iov->iov_base + adv; 1201 uio_clone->uio_iov->iov_len -= adv; 1202 uio_clone->uio_resid -= adv; 1203 uio_clone->uio_offset += adv; 1204 1205 uio->uio_resid -= adv; 1206 uio->uio_offset += adv; 1207 1208 if (error != 0 || adv == 0) 1209 break; 1210 } 1211 td->td_ma = prev_td_ma; 1212 td->td_ma_cnt = prev_td_ma_cnt; 1213 curthread_pflags_restore(saveheld); 1214 out: 1215 free(uio_clone, M_IOV); 1216 return (error); 1217 } 1218 1219 static int 1220 vn_io_fault(struct file *fp, struct uio *uio, struct ucred *active_cred, 1221 int flags, struct thread *td) 1222 { 1223 fo_rdwr_t *doio; 1224 struct vnode *vp; 1225 void *rl_cookie; 1226 struct vn_io_fault_args args; 1227 int error; 1228 1229 doio = uio->uio_rw == UIO_READ ? vn_read : vn_write; 1230 vp = fp->f_vnode; 1231 1232 /* 1233 * The ability to read(2) on a directory has historically been 1234 * allowed for all users, but this can and has been the source of 1235 * at least one security issue in the past. As such, it is now hidden 1236 * away behind a sysctl for those that actually need it to use it, and 1237 * restricted to root when it's turned on to make it relatively safe to 1238 * leave on for longer sessions of need. 1239 */ 1240 if (vp->v_type == VDIR) { 1241 KASSERT(uio->uio_rw == UIO_READ, 1242 ("illegal write attempted on a directory")); 1243 if (!vfs_allow_read_dir) 1244 return (EISDIR); 1245 if ((error = priv_check(td, PRIV_VFS_READ_DIR)) != 0) 1246 return (EISDIR); 1247 } 1248 1249 foffset_lock_uio(fp, uio, flags); 1250 if (do_vn_io_fault(vp, uio)) { 1251 args.kind = VN_IO_FAULT_FOP; 1252 args.args.fop_args.fp = fp; 1253 args.args.fop_args.doio = doio; 1254 args.cred = active_cred; 1255 args.flags = flags | FOF_OFFSET; 1256 if (uio->uio_rw == UIO_READ) { 1257 rl_cookie = vn_rangelock_rlock(vp, uio->uio_offset, 1258 uio->uio_offset + uio->uio_resid); 1259 } else if ((fp->f_flag & O_APPEND) != 0 || 1260 (flags & FOF_OFFSET) == 0) { 1261 /* For appenders, punt and lock the whole range. */ 1262 rl_cookie = vn_rangelock_wlock(vp, 0, OFF_MAX); 1263 } else { 1264 rl_cookie = vn_rangelock_wlock(vp, uio->uio_offset, 1265 uio->uio_offset + uio->uio_resid); 1266 } 1267 error = vn_io_fault1(vp, uio, &args, td); 1268 vn_rangelock_unlock(vp, rl_cookie); 1269 } else { 1270 error = doio(fp, uio, active_cred, flags | FOF_OFFSET, td); 1271 } 1272 foffset_unlock_uio(fp, uio, flags); 1273 return (error); 1274 } 1275 1276 /* 1277 * Helper function to perform the requested uiomove operation using 1278 * the held pages for io->uio_iov[0].iov_base buffer instead of 1279 * copyin/copyout. Access to the pages with uiomove_fromphys() 1280 * instead of iov_base prevents page faults that could occur due to 1281 * pmap_collect() invalidating the mapping created by 1282 * vm_fault_quick_hold_pages(), or pageout daemon, page laundry or 1283 * object cleanup revoking the write access from page mappings. 1284 * 1285 * Filesystems specified MNTK_NO_IOPF shall use vn_io_fault_uiomove() 1286 * instead of plain uiomove(). 1287 */ 1288 int 1289 vn_io_fault_uiomove(char *data, int xfersize, struct uio *uio) 1290 { 1291 struct uio transp_uio; 1292 struct iovec transp_iov[1]; 1293 struct thread *td; 1294 size_t adv; 1295 int error, pgadv; 1296 1297 td = curthread; 1298 if ((td->td_pflags & TDP_UIOHELD) == 0 || 1299 uio->uio_segflg != UIO_USERSPACE) 1300 return (uiomove(data, xfersize, uio)); 1301 1302 KASSERT(uio->uio_iovcnt == 1, ("uio_iovcnt %d", uio->uio_iovcnt)); 1303 transp_iov[0].iov_base = data; 1304 transp_uio.uio_iov = &transp_iov[0]; 1305 transp_uio.uio_iovcnt = 1; 1306 if (xfersize > uio->uio_resid) 1307 xfersize = uio->uio_resid; 1308 transp_uio.uio_resid = transp_iov[0].iov_len = xfersize; 1309 transp_uio.uio_offset = 0; 1310 transp_uio.uio_segflg = UIO_SYSSPACE; 1311 /* 1312 * Since transp_iov points to data, and td_ma page array 1313 * corresponds to original uio->uio_iov, we need to invert the 1314 * direction of the i/o operation as passed to 1315 * uiomove_fromphys(). 1316 */ 1317 switch (uio->uio_rw) { 1318 case UIO_WRITE: 1319 transp_uio.uio_rw = UIO_READ; 1320 break; 1321 case UIO_READ: 1322 transp_uio.uio_rw = UIO_WRITE; 1323 break; 1324 } 1325 transp_uio.uio_td = uio->uio_td; 1326 error = uiomove_fromphys(td->td_ma, 1327 ((vm_offset_t)uio->uio_iov->iov_base) & PAGE_MASK, 1328 xfersize, &transp_uio); 1329 adv = xfersize - transp_uio.uio_resid; 1330 pgadv = 1331 (((vm_offset_t)uio->uio_iov->iov_base + adv) >> PAGE_SHIFT) - 1332 (((vm_offset_t)uio->uio_iov->iov_base) >> PAGE_SHIFT); 1333 td->td_ma += pgadv; 1334 KASSERT(td->td_ma_cnt >= pgadv, ("consumed pages %d %d", td->td_ma_cnt, 1335 pgadv)); 1336 td->td_ma_cnt -= pgadv; 1337 uio->uio_iov->iov_base = (char *)uio->uio_iov->iov_base + adv; 1338 uio->uio_iov->iov_len -= adv; 1339 uio->uio_resid -= adv; 1340 uio->uio_offset += adv; 1341 return (error); 1342 } 1343 1344 int 1345 vn_io_fault_pgmove(vm_page_t ma[], vm_offset_t offset, int xfersize, 1346 struct uio *uio) 1347 { 1348 struct thread *td; 1349 vm_offset_t iov_base; 1350 int cnt, pgadv; 1351 1352 td = curthread; 1353 if ((td->td_pflags & TDP_UIOHELD) == 0 || 1354 uio->uio_segflg != UIO_USERSPACE) 1355 return (uiomove_fromphys(ma, offset, xfersize, uio)); 1356 1357 KASSERT(uio->uio_iovcnt == 1, ("uio_iovcnt %d", uio->uio_iovcnt)); 1358 cnt = xfersize > uio->uio_resid ? uio->uio_resid : xfersize; 1359 iov_base = (vm_offset_t)uio->uio_iov->iov_base; 1360 switch (uio->uio_rw) { 1361 case UIO_WRITE: 1362 pmap_copy_pages(td->td_ma, iov_base & PAGE_MASK, ma, 1363 offset, cnt); 1364 break; 1365 case UIO_READ: 1366 pmap_copy_pages(ma, offset, td->td_ma, iov_base & PAGE_MASK, 1367 cnt); 1368 break; 1369 } 1370 pgadv = ((iov_base + cnt) >> PAGE_SHIFT) - (iov_base >> PAGE_SHIFT); 1371 td->td_ma += pgadv; 1372 KASSERT(td->td_ma_cnt >= pgadv, ("consumed pages %d %d", td->td_ma_cnt, 1373 pgadv)); 1374 td->td_ma_cnt -= pgadv; 1375 uio->uio_iov->iov_base = (char *)(iov_base + cnt); 1376 uio->uio_iov->iov_len -= cnt; 1377 uio->uio_resid -= cnt; 1378 uio->uio_offset += cnt; 1379 return (0); 1380 } 1381 1382 /* 1383 * File table truncate routine. 1384 */ 1385 static int 1386 vn_truncate(struct file *fp, off_t length, struct ucred *active_cred, 1387 struct thread *td) 1388 { 1389 struct mount *mp; 1390 struct vnode *vp; 1391 void *rl_cookie; 1392 int error; 1393 1394 vp = fp->f_vnode; 1395 1396 /* 1397 * Lock the whole range for truncation. Otherwise split i/o 1398 * might happen partly before and partly after the truncation. 1399 */ 1400 rl_cookie = vn_rangelock_wlock(vp, 0, OFF_MAX); 1401 error = vn_start_write(vp, &mp, V_WAIT | PCATCH); 1402 if (error) 1403 goto out1; 1404 vn_lock(vp, LK_EXCLUSIVE | LK_RETRY); 1405 AUDIT_ARG_VNODE1(vp); 1406 if (vp->v_type == VDIR) { 1407 error = EISDIR; 1408 goto out; 1409 } 1410 #ifdef MAC 1411 error = mac_vnode_check_write(active_cred, fp->f_cred, vp); 1412 if (error) 1413 goto out; 1414 #endif 1415 error = vn_truncate_locked(vp, length, (fp->f_flag & O_FSYNC) != 0, 1416 fp->f_cred); 1417 out: 1418 VOP_UNLOCK(vp); 1419 vn_finished_write(mp); 1420 out1: 1421 vn_rangelock_unlock(vp, rl_cookie); 1422 return (error); 1423 } 1424 1425 /* 1426 * Truncate a file that is already locked. 1427 */ 1428 int 1429 vn_truncate_locked(struct vnode *vp, off_t length, bool sync, 1430 struct ucred *cred) 1431 { 1432 struct vattr vattr; 1433 int error; 1434 1435 error = VOP_ADD_WRITECOUNT(vp, 1); 1436 if (error == 0) { 1437 VATTR_NULL(&vattr); 1438 vattr.va_size = length; 1439 if (sync) 1440 vattr.va_vaflags |= VA_SYNC; 1441 error = VOP_SETATTR(vp, &vattr, cred); 1442 VOP_ADD_WRITECOUNT_CHECKED(vp, -1); 1443 } 1444 return (error); 1445 } 1446 1447 /* 1448 * File table vnode stat routine. 1449 */ 1450 static int 1451 vn_statfile(struct file *fp, struct stat *sb, struct ucred *active_cred, 1452 struct thread *td) 1453 { 1454 struct vnode *vp = fp->f_vnode; 1455 int error; 1456 1457 vn_lock(vp, LK_SHARED | LK_RETRY); 1458 error = VOP_STAT(vp, sb, active_cred, fp->f_cred, td); 1459 VOP_UNLOCK(vp); 1460 1461 return (error); 1462 } 1463 1464 /* 1465 * File table vnode ioctl routine. 1466 */ 1467 static int 1468 vn_ioctl(struct file *fp, u_long com, void *data, struct ucred *active_cred, 1469 struct thread *td) 1470 { 1471 struct vattr vattr; 1472 struct vnode *vp; 1473 struct fiobmap2_arg *bmarg; 1474 int error; 1475 1476 vp = fp->f_vnode; 1477 switch (vp->v_type) { 1478 case VDIR: 1479 case VREG: 1480 switch (com) { 1481 case FIONREAD: 1482 vn_lock(vp, LK_SHARED | LK_RETRY); 1483 error = VOP_GETATTR(vp, &vattr, active_cred); 1484 VOP_UNLOCK(vp); 1485 if (error == 0) 1486 *(int *)data = vattr.va_size - fp->f_offset; 1487 return (error); 1488 case FIOBMAP2: 1489 bmarg = (struct fiobmap2_arg *)data; 1490 vn_lock(vp, LK_SHARED | LK_RETRY); 1491 #ifdef MAC 1492 error = mac_vnode_check_read(active_cred, fp->f_cred, 1493 vp); 1494 if (error == 0) 1495 #endif 1496 error = VOP_BMAP(vp, bmarg->bn, NULL, 1497 &bmarg->bn, &bmarg->runp, &bmarg->runb); 1498 VOP_UNLOCK(vp); 1499 return (error); 1500 case FIONBIO: 1501 case FIOASYNC: 1502 return (0); 1503 default: 1504 return (VOP_IOCTL(vp, com, data, fp->f_flag, 1505 active_cred, td)); 1506 } 1507 break; 1508 case VCHR: 1509 return (VOP_IOCTL(vp, com, data, fp->f_flag, 1510 active_cred, td)); 1511 default: 1512 return (ENOTTY); 1513 } 1514 } 1515 1516 /* 1517 * File table vnode poll routine. 1518 */ 1519 static int 1520 vn_poll(struct file *fp, int events, struct ucred *active_cred, 1521 struct thread *td) 1522 { 1523 struct vnode *vp; 1524 int error; 1525 1526 vp = fp->f_vnode; 1527 #if defined(MAC) || defined(AUDIT) 1528 if (AUDITING_TD(td) || mac_vnode_check_poll_enabled()) { 1529 vn_lock(vp, LK_EXCLUSIVE | LK_RETRY); 1530 AUDIT_ARG_VNODE1(vp); 1531 error = mac_vnode_check_poll(active_cred, fp->f_cred, vp); 1532 VOP_UNLOCK(vp); 1533 if (error != 0) 1534 return (error); 1535 } 1536 #endif 1537 error = VOP_POLL(vp, events, fp->f_cred, td); 1538 return (error); 1539 } 1540 1541 /* 1542 * Acquire the requested lock and then check for validity. LK_RETRY 1543 * permits vn_lock to return doomed vnodes. 1544 */ 1545 static int __noinline 1546 _vn_lock_fallback(struct vnode *vp, int flags, const char *file, int line, 1547 int error) 1548 { 1549 1550 KASSERT((flags & LK_RETRY) == 0 || error == 0, 1551 ("vn_lock: error %d incompatible with flags %#x", error, flags)); 1552 1553 if (error == 0) 1554 VNASSERT(VN_IS_DOOMED(vp), vp, ("vnode not doomed")); 1555 1556 if ((flags & LK_RETRY) == 0) { 1557 if (error == 0) { 1558 VOP_UNLOCK(vp); 1559 error = ENOENT; 1560 } 1561 return (error); 1562 } 1563 1564 /* 1565 * LK_RETRY case. 1566 * 1567 * Nothing to do if we got the lock. 1568 */ 1569 if (error == 0) 1570 return (0); 1571 1572 /* 1573 * Interlock was dropped by the call in _vn_lock. 1574 */ 1575 flags &= ~LK_INTERLOCK; 1576 do { 1577 error = VOP_LOCK1(vp, flags, file, line); 1578 } while (error != 0); 1579 return (0); 1580 } 1581 1582 int 1583 _vn_lock(struct vnode *vp, int flags, const char *file, int line) 1584 { 1585 int error; 1586 1587 VNASSERT((flags & LK_TYPE_MASK) != 0, vp, 1588 ("vn_lock: no locktype (%d passed)", flags)); 1589 VNPASS(vp->v_holdcnt > 0, vp); 1590 error = VOP_LOCK1(vp, flags, file, line); 1591 if (__predict_false(error != 0 || VN_IS_DOOMED(vp))) 1592 return (_vn_lock_fallback(vp, flags, file, line, error)); 1593 return (0); 1594 } 1595 1596 /* 1597 * File table vnode close routine. 1598 */ 1599 static int 1600 vn_closefile(struct file *fp, struct thread *td) 1601 { 1602 struct vnode *vp; 1603 struct flock lf; 1604 int error; 1605 bool ref; 1606 1607 vp = fp->f_vnode; 1608 fp->f_ops = &badfileops; 1609 ref= (fp->f_flag & FHASLOCK) != 0 && fp->f_type == DTYPE_VNODE; 1610 1611 error = vn_close1(vp, fp->f_flag, fp->f_cred, td, ref); 1612 1613 if (__predict_false(ref)) { 1614 lf.l_whence = SEEK_SET; 1615 lf.l_start = 0; 1616 lf.l_len = 0; 1617 lf.l_type = F_UNLCK; 1618 (void) VOP_ADVLOCK(vp, fp, F_UNLCK, &lf, F_FLOCK); 1619 vrele(vp); 1620 } 1621 return (error); 1622 } 1623 1624 /* 1625 * Preparing to start a filesystem write operation. If the operation is 1626 * permitted, then we bump the count of operations in progress and 1627 * proceed. If a suspend request is in progress, we wait until the 1628 * suspension is over, and then proceed. 1629 */ 1630 static int 1631 vn_start_write_refed(struct mount *mp, int flags, bool mplocked) 1632 { 1633 int error, mflags; 1634 1635 if (__predict_true(!mplocked) && (flags & V_XSLEEP) == 0 && 1636 vfs_op_thread_enter(mp)) { 1637 MPASS((mp->mnt_kern_flag & MNTK_SUSPEND) == 0); 1638 vfs_mp_count_add_pcpu(mp, writeopcount, 1); 1639 vfs_op_thread_exit(mp); 1640 return (0); 1641 } 1642 1643 if (mplocked) 1644 mtx_assert(MNT_MTX(mp), MA_OWNED); 1645 else 1646 MNT_ILOCK(mp); 1647 1648 error = 0; 1649 1650 /* 1651 * Check on status of suspension. 1652 */ 1653 if ((curthread->td_pflags & TDP_IGNSUSP) == 0 || 1654 mp->mnt_susp_owner != curthread) { 1655 mflags = ((mp->mnt_vfc->vfc_flags & VFCF_SBDRY) != 0 ? 1656 (flags & PCATCH) : 0) | (PUSER - 1); 1657 while ((mp->mnt_kern_flag & MNTK_SUSPEND) != 0) { 1658 if (flags & V_NOWAIT) { 1659 error = EWOULDBLOCK; 1660 goto unlock; 1661 } 1662 error = msleep(&mp->mnt_flag, MNT_MTX(mp), mflags, 1663 "suspfs", 0); 1664 if (error) 1665 goto unlock; 1666 } 1667 } 1668 if (flags & V_XSLEEP) 1669 goto unlock; 1670 mp->mnt_writeopcount++; 1671 unlock: 1672 if (error != 0 || (flags & V_XSLEEP) != 0) 1673 MNT_REL(mp); 1674 MNT_IUNLOCK(mp); 1675 return (error); 1676 } 1677 1678 int 1679 vn_start_write(struct vnode *vp, struct mount **mpp, int flags) 1680 { 1681 struct mount *mp; 1682 int error; 1683 1684 KASSERT((flags & V_MNTREF) == 0 || (*mpp != NULL && vp == NULL), 1685 ("V_MNTREF requires mp")); 1686 1687 error = 0; 1688 /* 1689 * If a vnode is provided, get and return the mount point that 1690 * to which it will write. 1691 */ 1692 if (vp != NULL) { 1693 if ((error = VOP_GETWRITEMOUNT(vp, mpp)) != 0) { 1694 *mpp = NULL; 1695 if (error != EOPNOTSUPP) 1696 return (error); 1697 return (0); 1698 } 1699 } 1700 if ((mp = *mpp) == NULL) 1701 return (0); 1702 1703 /* 1704 * VOP_GETWRITEMOUNT() returns with the mp refcount held through 1705 * a vfs_ref(). 1706 * As long as a vnode is not provided we need to acquire a 1707 * refcount for the provided mountpoint too, in order to 1708 * emulate a vfs_ref(). 1709 */ 1710 if (vp == NULL && (flags & V_MNTREF) == 0) 1711 vfs_ref(mp); 1712 1713 return (vn_start_write_refed(mp, flags, false)); 1714 } 1715 1716 /* 1717 * Secondary suspension. Used by operations such as vop_inactive 1718 * routines that are needed by the higher level functions. These 1719 * are allowed to proceed until all the higher level functions have 1720 * completed (indicated by mnt_writeopcount dropping to zero). At that 1721 * time, these operations are halted until the suspension is over. 1722 */ 1723 int 1724 vn_start_secondary_write(struct vnode *vp, struct mount **mpp, int flags) 1725 { 1726 struct mount *mp; 1727 int error; 1728 1729 KASSERT((flags & V_MNTREF) == 0 || (*mpp != NULL && vp == NULL), 1730 ("V_MNTREF requires mp")); 1731 1732 retry: 1733 if (vp != NULL) { 1734 if ((error = VOP_GETWRITEMOUNT(vp, mpp)) != 0) { 1735 *mpp = NULL; 1736 if (error != EOPNOTSUPP) 1737 return (error); 1738 return (0); 1739 } 1740 } 1741 /* 1742 * If we are not suspended or have not yet reached suspended 1743 * mode, then let the operation proceed. 1744 */ 1745 if ((mp = *mpp) == NULL) 1746 return (0); 1747 1748 /* 1749 * VOP_GETWRITEMOUNT() returns with the mp refcount held through 1750 * a vfs_ref(). 1751 * As long as a vnode is not provided we need to acquire a 1752 * refcount for the provided mountpoint too, in order to 1753 * emulate a vfs_ref(). 1754 */ 1755 MNT_ILOCK(mp); 1756 if (vp == NULL && (flags & V_MNTREF) == 0) 1757 MNT_REF(mp); 1758 if ((mp->mnt_kern_flag & (MNTK_SUSPENDED | MNTK_SUSPEND2)) == 0) { 1759 mp->mnt_secondary_writes++; 1760 mp->mnt_secondary_accwrites++; 1761 MNT_IUNLOCK(mp); 1762 return (0); 1763 } 1764 if (flags & V_NOWAIT) { 1765 MNT_REL(mp); 1766 MNT_IUNLOCK(mp); 1767 return (EWOULDBLOCK); 1768 } 1769 /* 1770 * Wait for the suspension to finish. 1771 */ 1772 error = msleep(&mp->mnt_flag, MNT_MTX(mp), (PUSER - 1) | PDROP | 1773 ((mp->mnt_vfc->vfc_flags & VFCF_SBDRY) != 0 ? (flags & PCATCH) : 0), 1774 "suspfs", 0); 1775 vfs_rel(mp); 1776 if (error == 0) 1777 goto retry; 1778 return (error); 1779 } 1780 1781 /* 1782 * Filesystem write operation has completed. If we are suspending and this 1783 * operation is the last one, notify the suspender that the suspension is 1784 * now in effect. 1785 */ 1786 void 1787 vn_finished_write(struct mount *mp) 1788 { 1789 int c; 1790 1791 if (mp == NULL) 1792 return; 1793 1794 if (vfs_op_thread_enter(mp)) { 1795 vfs_mp_count_sub_pcpu(mp, writeopcount, 1); 1796 vfs_mp_count_sub_pcpu(mp, ref, 1); 1797 vfs_op_thread_exit(mp); 1798 return; 1799 } 1800 1801 MNT_ILOCK(mp); 1802 vfs_assert_mount_counters(mp); 1803 MNT_REL(mp); 1804 c = --mp->mnt_writeopcount; 1805 if (mp->mnt_vfs_ops == 0) { 1806 MPASS((mp->mnt_kern_flag & MNTK_SUSPEND) == 0); 1807 MNT_IUNLOCK(mp); 1808 return; 1809 } 1810 if (c < 0) 1811 vfs_dump_mount_counters(mp); 1812 if ((mp->mnt_kern_flag & MNTK_SUSPEND) != 0 && c == 0) 1813 wakeup(&mp->mnt_writeopcount); 1814 MNT_IUNLOCK(mp); 1815 } 1816 1817 /* 1818 * Filesystem secondary write operation has completed. If we are 1819 * suspending and this operation is the last one, notify the suspender 1820 * that the suspension is now in effect. 1821 */ 1822 void 1823 vn_finished_secondary_write(struct mount *mp) 1824 { 1825 if (mp == NULL) 1826 return; 1827 MNT_ILOCK(mp); 1828 MNT_REL(mp); 1829 mp->mnt_secondary_writes--; 1830 if (mp->mnt_secondary_writes < 0) 1831 panic("vn_finished_secondary_write: neg cnt"); 1832 if ((mp->mnt_kern_flag & MNTK_SUSPEND) != 0 && 1833 mp->mnt_secondary_writes <= 0) 1834 wakeup(&mp->mnt_secondary_writes); 1835 MNT_IUNLOCK(mp); 1836 } 1837 1838 /* 1839 * Request a filesystem to suspend write operations. 1840 */ 1841 int 1842 vfs_write_suspend(struct mount *mp, int flags) 1843 { 1844 int error; 1845 1846 vfs_op_enter(mp); 1847 1848 MNT_ILOCK(mp); 1849 vfs_assert_mount_counters(mp); 1850 if (mp->mnt_susp_owner == curthread) { 1851 vfs_op_exit_locked(mp); 1852 MNT_IUNLOCK(mp); 1853 return (EALREADY); 1854 } 1855 while (mp->mnt_kern_flag & MNTK_SUSPEND) 1856 msleep(&mp->mnt_flag, MNT_MTX(mp), PUSER - 1, "wsuspfs", 0); 1857 1858 /* 1859 * Unmount holds a write reference on the mount point. If we 1860 * own busy reference and drain for writers, we deadlock with 1861 * the reference draining in the unmount path. Callers of 1862 * vfs_write_suspend() must specify VS_SKIP_UNMOUNT if 1863 * vfs_busy() reference is owned and caller is not in the 1864 * unmount context. 1865 */ 1866 if ((flags & VS_SKIP_UNMOUNT) != 0 && 1867 (mp->mnt_kern_flag & MNTK_UNMOUNT) != 0) { 1868 vfs_op_exit_locked(mp); 1869 MNT_IUNLOCK(mp); 1870 return (EBUSY); 1871 } 1872 1873 mp->mnt_kern_flag |= MNTK_SUSPEND; 1874 mp->mnt_susp_owner = curthread; 1875 if (mp->mnt_writeopcount > 0) 1876 (void) msleep(&mp->mnt_writeopcount, 1877 MNT_MTX(mp), (PUSER - 1)|PDROP, "suspwt", 0); 1878 else 1879 MNT_IUNLOCK(mp); 1880 if ((error = VFS_SYNC(mp, MNT_SUSPEND)) != 0) { 1881 vfs_write_resume(mp, 0); 1882 /* vfs_write_resume does vfs_op_exit() for us */ 1883 } 1884 return (error); 1885 } 1886 1887 /* 1888 * Request a filesystem to resume write operations. 1889 */ 1890 void 1891 vfs_write_resume(struct mount *mp, int flags) 1892 { 1893 1894 MNT_ILOCK(mp); 1895 if ((mp->mnt_kern_flag & MNTK_SUSPEND) != 0) { 1896 KASSERT(mp->mnt_susp_owner == curthread, ("mnt_susp_owner")); 1897 mp->mnt_kern_flag &= ~(MNTK_SUSPEND | MNTK_SUSPEND2 | 1898 MNTK_SUSPENDED); 1899 mp->mnt_susp_owner = NULL; 1900 wakeup(&mp->mnt_writeopcount); 1901 wakeup(&mp->mnt_flag); 1902 curthread->td_pflags &= ~TDP_IGNSUSP; 1903 if ((flags & VR_START_WRITE) != 0) { 1904 MNT_REF(mp); 1905 mp->mnt_writeopcount++; 1906 } 1907 MNT_IUNLOCK(mp); 1908 if ((flags & VR_NO_SUSPCLR) == 0) 1909 VFS_SUSP_CLEAN(mp); 1910 vfs_op_exit(mp); 1911 } else if ((flags & VR_START_WRITE) != 0) { 1912 MNT_REF(mp); 1913 vn_start_write_refed(mp, 0, true); 1914 } else { 1915 MNT_IUNLOCK(mp); 1916 } 1917 } 1918 1919 /* 1920 * Helper loop around vfs_write_suspend() for filesystem unmount VFS 1921 * methods. 1922 */ 1923 int 1924 vfs_write_suspend_umnt(struct mount *mp) 1925 { 1926 int error; 1927 1928 KASSERT((curthread->td_pflags & TDP_IGNSUSP) == 0, 1929 ("vfs_write_suspend_umnt: recursed")); 1930 1931 /* dounmount() already called vn_start_write(). */ 1932 for (;;) { 1933 vn_finished_write(mp); 1934 error = vfs_write_suspend(mp, 0); 1935 if (error != 0) { 1936 vn_start_write(NULL, &mp, V_WAIT); 1937 return (error); 1938 } 1939 MNT_ILOCK(mp); 1940 if ((mp->mnt_kern_flag & MNTK_SUSPENDED) != 0) 1941 break; 1942 MNT_IUNLOCK(mp); 1943 vn_start_write(NULL, &mp, V_WAIT); 1944 } 1945 mp->mnt_kern_flag &= ~(MNTK_SUSPENDED | MNTK_SUSPEND2); 1946 wakeup(&mp->mnt_flag); 1947 MNT_IUNLOCK(mp); 1948 curthread->td_pflags |= TDP_IGNSUSP; 1949 return (0); 1950 } 1951 1952 /* 1953 * Implement kqueues for files by translating it to vnode operation. 1954 */ 1955 static int 1956 vn_kqfilter(struct file *fp, struct knote *kn) 1957 { 1958 1959 return (VOP_KQFILTER(fp->f_vnode, kn)); 1960 } 1961 1962 /* 1963 * Simplified in-kernel wrapper calls for extended attribute access. 1964 * Both calls pass in a NULL credential, authorizing as "kernel" access. 1965 * Set IO_NODELOCKED in ioflg if the vnode is already locked. 1966 */ 1967 int 1968 vn_extattr_get(struct vnode *vp, int ioflg, int attrnamespace, 1969 const char *attrname, int *buflen, char *buf, struct thread *td) 1970 { 1971 struct uio auio; 1972 struct iovec iov; 1973 int error; 1974 1975 iov.iov_len = *buflen; 1976 iov.iov_base = buf; 1977 1978 auio.uio_iov = &iov; 1979 auio.uio_iovcnt = 1; 1980 auio.uio_rw = UIO_READ; 1981 auio.uio_segflg = UIO_SYSSPACE; 1982 auio.uio_td = td; 1983 auio.uio_offset = 0; 1984 auio.uio_resid = *buflen; 1985 1986 if ((ioflg & IO_NODELOCKED) == 0) 1987 vn_lock(vp, LK_SHARED | LK_RETRY); 1988 1989 ASSERT_VOP_LOCKED(vp, "IO_NODELOCKED with no vp lock held"); 1990 1991 /* authorize attribute retrieval as kernel */ 1992 error = VOP_GETEXTATTR(vp, attrnamespace, attrname, &auio, NULL, NULL, 1993 td); 1994 1995 if ((ioflg & IO_NODELOCKED) == 0) 1996 VOP_UNLOCK(vp); 1997 1998 if (error == 0) { 1999 *buflen = *buflen - auio.uio_resid; 2000 } 2001 2002 return (error); 2003 } 2004 2005 /* 2006 * XXX failure mode if partially written? 2007 */ 2008 int 2009 vn_extattr_set(struct vnode *vp, int ioflg, int attrnamespace, 2010 const char *attrname, int buflen, char *buf, struct thread *td) 2011 { 2012 struct uio auio; 2013 struct iovec iov; 2014 struct mount *mp; 2015 int error; 2016 2017 iov.iov_len = buflen; 2018 iov.iov_base = buf; 2019 2020 auio.uio_iov = &iov; 2021 auio.uio_iovcnt = 1; 2022 auio.uio_rw = UIO_WRITE; 2023 auio.uio_segflg = UIO_SYSSPACE; 2024 auio.uio_td = td; 2025 auio.uio_offset = 0; 2026 auio.uio_resid = buflen; 2027 2028 if ((ioflg & IO_NODELOCKED) == 0) { 2029 if ((error = vn_start_write(vp, &mp, V_WAIT)) != 0) 2030 return (error); 2031 vn_lock(vp, LK_EXCLUSIVE | LK_RETRY); 2032 } 2033 2034 ASSERT_VOP_LOCKED(vp, "IO_NODELOCKED with no vp lock held"); 2035 2036 /* authorize attribute setting as kernel */ 2037 error = VOP_SETEXTATTR(vp, attrnamespace, attrname, &auio, NULL, td); 2038 2039 if ((ioflg & IO_NODELOCKED) == 0) { 2040 vn_finished_write(mp); 2041 VOP_UNLOCK(vp); 2042 } 2043 2044 return (error); 2045 } 2046 2047 int 2048 vn_extattr_rm(struct vnode *vp, int ioflg, int attrnamespace, 2049 const char *attrname, struct thread *td) 2050 { 2051 struct mount *mp; 2052 int error; 2053 2054 if ((ioflg & IO_NODELOCKED) == 0) { 2055 if ((error = vn_start_write(vp, &mp, V_WAIT)) != 0) 2056 return (error); 2057 vn_lock(vp, LK_EXCLUSIVE | LK_RETRY); 2058 } 2059 2060 ASSERT_VOP_LOCKED(vp, "IO_NODELOCKED with no vp lock held"); 2061 2062 /* authorize attribute removal as kernel */ 2063 error = VOP_DELETEEXTATTR(vp, attrnamespace, attrname, NULL, td); 2064 if (error == EOPNOTSUPP) 2065 error = VOP_SETEXTATTR(vp, attrnamespace, attrname, NULL, 2066 NULL, td); 2067 2068 if ((ioflg & IO_NODELOCKED) == 0) { 2069 vn_finished_write(mp); 2070 VOP_UNLOCK(vp); 2071 } 2072 2073 return (error); 2074 } 2075 2076 static int 2077 vn_get_ino_alloc_vget(struct mount *mp, void *arg, int lkflags, 2078 struct vnode **rvp) 2079 { 2080 2081 return (VFS_VGET(mp, *(ino_t *)arg, lkflags, rvp)); 2082 } 2083 2084 int 2085 vn_vget_ino(struct vnode *vp, ino_t ino, int lkflags, struct vnode **rvp) 2086 { 2087 2088 return (vn_vget_ino_gen(vp, vn_get_ino_alloc_vget, &ino, 2089 lkflags, rvp)); 2090 } 2091 2092 int 2093 vn_vget_ino_gen(struct vnode *vp, vn_get_ino_t alloc, void *alloc_arg, 2094 int lkflags, struct vnode **rvp) 2095 { 2096 struct mount *mp; 2097 int ltype, error; 2098 2099 ASSERT_VOP_LOCKED(vp, "vn_vget_ino_get"); 2100 mp = vp->v_mount; 2101 ltype = VOP_ISLOCKED(vp); 2102 KASSERT(ltype == LK_EXCLUSIVE || ltype == LK_SHARED, 2103 ("vn_vget_ino: vp not locked")); 2104 error = vfs_busy(mp, MBF_NOWAIT); 2105 if (error != 0) { 2106 vfs_ref(mp); 2107 VOP_UNLOCK(vp); 2108 error = vfs_busy(mp, 0); 2109 vn_lock(vp, ltype | LK_RETRY); 2110 vfs_rel(mp); 2111 if (error != 0) 2112 return (ENOENT); 2113 if (VN_IS_DOOMED(vp)) { 2114 vfs_unbusy(mp); 2115 return (ENOENT); 2116 } 2117 } 2118 VOP_UNLOCK(vp); 2119 error = alloc(mp, alloc_arg, lkflags, rvp); 2120 vfs_unbusy(mp); 2121 if (error != 0 || *rvp != vp) 2122 vn_lock(vp, ltype | LK_RETRY); 2123 if (VN_IS_DOOMED(vp)) { 2124 if (error == 0) { 2125 if (*rvp == vp) 2126 vunref(vp); 2127 else 2128 vput(*rvp); 2129 } 2130 error = ENOENT; 2131 } 2132 return (error); 2133 } 2134 2135 int 2136 vn_rlimit_fsize(const struct vnode *vp, const struct uio *uio, 2137 struct thread *td) 2138 { 2139 2140 if (vp->v_type != VREG || td == NULL) 2141 return (0); 2142 if ((uoff_t)uio->uio_offset + uio->uio_resid > 2143 lim_cur(td, RLIMIT_FSIZE)) { 2144 PROC_LOCK(td->td_proc); 2145 kern_psignal(td->td_proc, SIGXFSZ); 2146 PROC_UNLOCK(td->td_proc); 2147 return (EFBIG); 2148 } 2149 return (0); 2150 } 2151 2152 int 2153 vn_chmod(struct file *fp, mode_t mode, struct ucred *active_cred, 2154 struct thread *td) 2155 { 2156 struct vnode *vp; 2157 2158 vp = fp->f_vnode; 2159 #ifdef AUDIT 2160 vn_lock(vp, LK_SHARED | LK_RETRY); 2161 AUDIT_ARG_VNODE1(vp); 2162 VOP_UNLOCK(vp); 2163 #endif 2164 return (setfmode(td, active_cred, vp, mode)); 2165 } 2166 2167 int 2168 vn_chown(struct file *fp, uid_t uid, gid_t gid, struct ucred *active_cred, 2169 struct thread *td) 2170 { 2171 struct vnode *vp; 2172 2173 vp = fp->f_vnode; 2174 #ifdef AUDIT 2175 vn_lock(vp, LK_SHARED | LK_RETRY); 2176 AUDIT_ARG_VNODE1(vp); 2177 VOP_UNLOCK(vp); 2178 #endif 2179 return (setfown(td, active_cred, vp, uid, gid)); 2180 } 2181 2182 void 2183 vn_pages_remove(struct vnode *vp, vm_pindex_t start, vm_pindex_t end) 2184 { 2185 vm_object_t object; 2186 2187 if ((object = vp->v_object) == NULL) 2188 return; 2189 VM_OBJECT_WLOCK(object); 2190 vm_object_page_remove(object, start, end, 0); 2191 VM_OBJECT_WUNLOCK(object); 2192 } 2193 2194 int 2195 vn_bmap_seekhole(struct vnode *vp, u_long cmd, off_t *off, struct ucred *cred) 2196 { 2197 struct vattr va; 2198 daddr_t bn, bnp; 2199 uint64_t bsize; 2200 off_t noff; 2201 int error; 2202 2203 KASSERT(cmd == FIOSEEKHOLE || cmd == FIOSEEKDATA, 2204 ("Wrong command %lu", cmd)); 2205 2206 if (vn_lock(vp, LK_SHARED) != 0) 2207 return (EBADF); 2208 if (vp->v_type != VREG) { 2209 error = ENOTTY; 2210 goto unlock; 2211 } 2212 error = VOP_GETATTR(vp, &va, cred); 2213 if (error != 0) 2214 goto unlock; 2215 noff = *off; 2216 if (noff >= va.va_size) { 2217 error = ENXIO; 2218 goto unlock; 2219 } 2220 bsize = vp->v_mount->mnt_stat.f_iosize; 2221 for (bn = noff / bsize; noff < va.va_size; bn++, noff += bsize - 2222 noff % bsize) { 2223 error = VOP_BMAP(vp, bn, NULL, &bnp, NULL, NULL); 2224 if (error == EOPNOTSUPP) { 2225 error = ENOTTY; 2226 goto unlock; 2227 } 2228 if ((bnp == -1 && cmd == FIOSEEKHOLE) || 2229 (bnp != -1 && cmd == FIOSEEKDATA)) { 2230 noff = bn * bsize; 2231 if (noff < *off) 2232 noff = *off; 2233 goto unlock; 2234 } 2235 } 2236 if (noff > va.va_size) 2237 noff = va.va_size; 2238 /* noff == va.va_size. There is an implicit hole at the end of file. */ 2239 if (cmd == FIOSEEKDATA) 2240 error = ENXIO; 2241 unlock: 2242 VOP_UNLOCK(vp); 2243 if (error == 0) 2244 *off = noff; 2245 return (error); 2246 } 2247 2248 int 2249 vn_seek(struct file *fp, off_t offset, int whence, struct thread *td) 2250 { 2251 struct ucred *cred; 2252 struct vnode *vp; 2253 struct vattr vattr; 2254 off_t foffset, size; 2255 int error, noneg; 2256 2257 cred = td->td_ucred; 2258 vp = fp->f_vnode; 2259 foffset = foffset_lock(fp, 0); 2260 noneg = (vp->v_type != VCHR); 2261 error = 0; 2262 switch (whence) { 2263 case L_INCR: 2264 if (noneg && 2265 (foffset < 0 || 2266 (offset > 0 && foffset > OFF_MAX - offset))) { 2267 error = EOVERFLOW; 2268 break; 2269 } 2270 offset += foffset; 2271 break; 2272 case L_XTND: 2273 vn_lock(vp, LK_SHARED | LK_RETRY); 2274 error = VOP_GETATTR(vp, &vattr, cred); 2275 VOP_UNLOCK(vp); 2276 if (error) 2277 break; 2278 2279 /* 2280 * If the file references a disk device, then fetch 2281 * the media size and use that to determine the ending 2282 * offset. 2283 */ 2284 if (vattr.va_size == 0 && vp->v_type == VCHR && 2285 fo_ioctl(fp, DIOCGMEDIASIZE, &size, cred, td) == 0) 2286 vattr.va_size = size; 2287 if (noneg && 2288 (vattr.va_size > OFF_MAX || 2289 (offset > 0 && vattr.va_size > OFF_MAX - offset))) { 2290 error = EOVERFLOW; 2291 break; 2292 } 2293 offset += vattr.va_size; 2294 break; 2295 case L_SET: 2296 break; 2297 case SEEK_DATA: 2298 error = fo_ioctl(fp, FIOSEEKDATA, &offset, cred, td); 2299 if (error == ENOTTY) 2300 error = EINVAL; 2301 break; 2302 case SEEK_HOLE: 2303 error = fo_ioctl(fp, FIOSEEKHOLE, &offset, cred, td); 2304 if (error == ENOTTY) 2305 error = EINVAL; 2306 break; 2307 default: 2308 error = EINVAL; 2309 } 2310 if (error == 0 && noneg && offset < 0) 2311 error = EINVAL; 2312 if (error != 0) 2313 goto drop; 2314 VFS_KNOTE_UNLOCKED(vp, 0); 2315 td->td_uretoff.tdu_off = offset; 2316 drop: 2317 foffset_unlock(fp, offset, error != 0 ? FOF_NOUPDATE : 0); 2318 return (error); 2319 } 2320 2321 int 2322 vn_utimes_perm(struct vnode *vp, struct vattr *vap, struct ucred *cred, 2323 struct thread *td) 2324 { 2325 int error; 2326 2327 /* 2328 * Grant permission if the caller is the owner of the file, or 2329 * the super-user, or has ACL_WRITE_ATTRIBUTES permission on 2330 * on the file. If the time pointer is null, then write 2331 * permission on the file is also sufficient. 2332 * 2333 * From NFSv4.1, draft 21, 6.2.1.3.1, Discussion of Mask Attributes: 2334 * A user having ACL_WRITE_DATA or ACL_WRITE_ATTRIBUTES 2335 * will be allowed to set the times [..] to the current 2336 * server time. 2337 */ 2338 error = VOP_ACCESSX(vp, VWRITE_ATTRIBUTES, cred, td); 2339 if (error != 0 && (vap->va_vaflags & VA_UTIMES_NULL) != 0) 2340 error = VOP_ACCESS(vp, VWRITE, cred, td); 2341 return (error); 2342 } 2343 2344 int 2345 vn_fill_kinfo(struct file *fp, struct kinfo_file *kif, struct filedesc *fdp) 2346 { 2347 struct vnode *vp; 2348 int error; 2349 2350 if (fp->f_type == DTYPE_FIFO) 2351 kif->kf_type = KF_TYPE_FIFO; 2352 else 2353 kif->kf_type = KF_TYPE_VNODE; 2354 vp = fp->f_vnode; 2355 vref(vp); 2356 FILEDESC_SUNLOCK(fdp); 2357 error = vn_fill_kinfo_vnode(vp, kif); 2358 vrele(vp); 2359 FILEDESC_SLOCK(fdp); 2360 return (error); 2361 } 2362 2363 static inline void 2364 vn_fill_junk(struct kinfo_file *kif) 2365 { 2366 size_t len, olen; 2367 2368 /* 2369 * Simulate vn_fullpath returning changing values for a given 2370 * vp during e.g. coredump. 2371 */ 2372 len = (arc4random() % (sizeof(kif->kf_path) - 2)) + 1; 2373 olen = strlen(kif->kf_path); 2374 if (len < olen) 2375 strcpy(&kif->kf_path[len - 1], "$"); 2376 else 2377 for (; olen < len; olen++) 2378 strcpy(&kif->kf_path[olen], "A"); 2379 } 2380 2381 int 2382 vn_fill_kinfo_vnode(struct vnode *vp, struct kinfo_file *kif) 2383 { 2384 struct vattr va; 2385 char *fullpath, *freepath; 2386 int error; 2387 2388 kif->kf_un.kf_file.kf_file_type = vntype_to_kinfo(vp->v_type); 2389 freepath = NULL; 2390 fullpath = "-"; 2391 error = vn_fullpath(curthread, vp, &fullpath, &freepath); 2392 if (error == 0) { 2393 strlcpy(kif->kf_path, fullpath, sizeof(kif->kf_path)); 2394 } 2395 if (freepath != NULL) 2396 free(freepath, M_TEMP); 2397 2398 KFAIL_POINT_CODE(DEBUG_FP, fill_kinfo_vnode__random_path, 2399 vn_fill_junk(kif); 2400 ); 2401 2402 /* 2403 * Retrieve vnode attributes. 2404 */ 2405 va.va_fsid = VNOVAL; 2406 va.va_rdev = NODEV; 2407 vn_lock(vp, LK_SHARED | LK_RETRY); 2408 error = VOP_GETATTR(vp, &va, curthread->td_ucred); 2409 VOP_UNLOCK(vp); 2410 if (error != 0) 2411 return (error); 2412 if (va.va_fsid != VNOVAL) 2413 kif->kf_un.kf_file.kf_file_fsid = va.va_fsid; 2414 else 2415 kif->kf_un.kf_file.kf_file_fsid = 2416 vp->v_mount->mnt_stat.f_fsid.val[0]; 2417 kif->kf_un.kf_file.kf_file_fsid_freebsd11 = 2418 kif->kf_un.kf_file.kf_file_fsid; /* truncate */ 2419 kif->kf_un.kf_file.kf_file_fileid = va.va_fileid; 2420 kif->kf_un.kf_file.kf_file_mode = MAKEIMODE(va.va_type, va.va_mode); 2421 kif->kf_un.kf_file.kf_file_size = va.va_size; 2422 kif->kf_un.kf_file.kf_file_rdev = va.va_rdev; 2423 kif->kf_un.kf_file.kf_file_rdev_freebsd11 = 2424 kif->kf_un.kf_file.kf_file_rdev; /* truncate */ 2425 return (0); 2426 } 2427 2428 int 2429 vn_mmap(struct file *fp, vm_map_t map, vm_offset_t *addr, vm_size_t size, 2430 vm_prot_t prot, vm_prot_t cap_maxprot, int flags, vm_ooffset_t foff, 2431 struct thread *td) 2432 { 2433 #ifdef HWPMC_HOOKS 2434 struct pmckern_map_in pkm; 2435 #endif 2436 struct mount *mp; 2437 struct vnode *vp; 2438 vm_object_t object; 2439 vm_prot_t maxprot; 2440 boolean_t writecounted; 2441 int error; 2442 2443 #if defined(COMPAT_FREEBSD7) || defined(COMPAT_FREEBSD6) || \ 2444 defined(COMPAT_FREEBSD5) || defined(COMPAT_FREEBSD4) 2445 /* 2446 * POSIX shared-memory objects are defined to have 2447 * kernel persistence, and are not defined to support 2448 * read(2)/write(2) -- or even open(2). Thus, we can 2449 * use MAP_ASYNC to trade on-disk coherence for speed. 2450 * The shm_open(3) library routine turns on the FPOSIXSHM 2451 * flag to request this behavior. 2452 */ 2453 if ((fp->f_flag & FPOSIXSHM) != 0) 2454 flags |= MAP_NOSYNC; 2455 #endif 2456 vp = fp->f_vnode; 2457 2458 /* 2459 * Ensure that file and memory protections are 2460 * compatible. Note that we only worry about 2461 * writability if mapping is shared; in this case, 2462 * current and max prot are dictated by the open file. 2463 * XXX use the vnode instead? Problem is: what 2464 * credentials do we use for determination? What if 2465 * proc does a setuid? 2466 */ 2467 mp = vp->v_mount; 2468 if (mp != NULL && (mp->mnt_flag & MNT_NOEXEC) != 0) { 2469 maxprot = VM_PROT_NONE; 2470 if ((prot & VM_PROT_EXECUTE) != 0) 2471 return (EACCES); 2472 } else 2473 maxprot = VM_PROT_EXECUTE; 2474 if ((fp->f_flag & FREAD) != 0) 2475 maxprot |= VM_PROT_READ; 2476 else if ((prot & VM_PROT_READ) != 0) 2477 return (EACCES); 2478 2479 /* 2480 * If we are sharing potential changes via MAP_SHARED and we 2481 * are trying to get write permission although we opened it 2482 * without asking for it, bail out. 2483 */ 2484 if ((flags & MAP_SHARED) != 0) { 2485 if ((fp->f_flag & FWRITE) != 0) 2486 maxprot |= VM_PROT_WRITE; 2487 else if ((prot & VM_PROT_WRITE) != 0) 2488 return (EACCES); 2489 } else { 2490 maxprot |= VM_PROT_WRITE; 2491 cap_maxprot |= VM_PROT_WRITE; 2492 } 2493 maxprot &= cap_maxprot; 2494 2495 /* 2496 * For regular files and shared memory, POSIX requires that 2497 * the value of foff be a legitimate offset within the data 2498 * object. In particular, negative offsets are invalid. 2499 * Blocking negative offsets and overflows here avoids 2500 * possible wraparound or user-level access into reserved 2501 * ranges of the data object later. In contrast, POSIX does 2502 * not dictate how offsets are used by device drivers, so in 2503 * the case of a device mapping a negative offset is passed 2504 * on. 2505 */ 2506 if ( 2507 #ifdef _LP64 2508 size > OFF_MAX || 2509 #endif 2510 foff < 0 || foff > OFF_MAX - size) 2511 return (EINVAL); 2512 2513 writecounted = FALSE; 2514 error = vm_mmap_vnode(td, size, prot, &maxprot, &flags, vp, 2515 &foff, &object, &writecounted); 2516 if (error != 0) 2517 return (error); 2518 error = vm_mmap_object(map, addr, size, prot, maxprot, flags, object, 2519 foff, writecounted, td); 2520 if (error != 0) { 2521 /* 2522 * If this mapping was accounted for in the vnode's 2523 * writecount, then undo that now. 2524 */ 2525 if (writecounted) 2526 vm_pager_release_writecount(object, 0, size); 2527 vm_object_deallocate(object); 2528 } 2529 #ifdef HWPMC_HOOKS 2530 /* Inform hwpmc(4) if an executable is being mapped. */ 2531 if (PMC_HOOK_INSTALLED(PMC_FN_MMAP)) { 2532 if ((prot & VM_PROT_EXECUTE) != 0 && error == 0) { 2533 pkm.pm_file = vp; 2534 pkm.pm_address = (uintptr_t) *addr; 2535 PMC_CALL_HOOK_UNLOCKED(td, PMC_FN_MMAP, (void *) &pkm); 2536 } 2537 } 2538 #endif 2539 return (error); 2540 } 2541 2542 void 2543 vn_fsid(struct vnode *vp, struct vattr *va) 2544 { 2545 fsid_t *f; 2546 2547 f = &vp->v_mount->mnt_stat.f_fsid; 2548 va->va_fsid = (uint32_t)f->val[1]; 2549 va->va_fsid <<= sizeof(f->val[1]) * NBBY; 2550 va->va_fsid += (uint32_t)f->val[0]; 2551 } 2552 2553 int 2554 vn_fsync_buf(struct vnode *vp, int waitfor) 2555 { 2556 struct buf *bp, *nbp; 2557 struct bufobj *bo; 2558 struct mount *mp; 2559 int error, maxretry; 2560 2561 error = 0; 2562 maxretry = 10000; /* large, arbitrarily chosen */ 2563 mp = NULL; 2564 if (vp->v_type == VCHR) { 2565 VI_LOCK(vp); 2566 mp = vp->v_rdev->si_mountpt; 2567 VI_UNLOCK(vp); 2568 } 2569 bo = &vp->v_bufobj; 2570 BO_LOCK(bo); 2571 loop1: 2572 /* 2573 * MARK/SCAN initialization to avoid infinite loops. 2574 */ 2575 TAILQ_FOREACH(bp, &bo->bo_dirty.bv_hd, b_bobufs) { 2576 bp->b_vflags &= ~BV_SCANNED; 2577 bp->b_error = 0; 2578 } 2579 2580 /* 2581 * Flush all dirty buffers associated with a vnode. 2582 */ 2583 loop2: 2584 TAILQ_FOREACH_SAFE(bp, &bo->bo_dirty.bv_hd, b_bobufs, nbp) { 2585 if ((bp->b_vflags & BV_SCANNED) != 0) 2586 continue; 2587 bp->b_vflags |= BV_SCANNED; 2588 if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT, NULL)) { 2589 if (waitfor != MNT_WAIT) 2590 continue; 2591 if (BUF_LOCK(bp, 2592 LK_EXCLUSIVE | LK_INTERLOCK | LK_SLEEPFAIL, 2593 BO_LOCKPTR(bo)) != 0) { 2594 BO_LOCK(bo); 2595 goto loop1; 2596 } 2597 BO_LOCK(bo); 2598 } 2599 BO_UNLOCK(bo); 2600 KASSERT(bp->b_bufobj == bo, 2601 ("bp %p wrong b_bufobj %p should be %p", 2602 bp, bp->b_bufobj, bo)); 2603 if ((bp->b_flags & B_DELWRI) == 0) 2604 panic("fsync: not dirty"); 2605 if ((vp->v_object != NULL) && (bp->b_flags & B_CLUSTEROK)) { 2606 vfs_bio_awrite(bp); 2607 } else { 2608 bremfree(bp); 2609 bawrite(bp); 2610 } 2611 if (maxretry < 1000) 2612 pause("dirty", hz < 1000 ? 1 : hz / 1000); 2613 BO_LOCK(bo); 2614 goto loop2; 2615 } 2616 2617 /* 2618 * If synchronous the caller expects us to completely resolve all 2619 * dirty buffers in the system. Wait for in-progress I/O to 2620 * complete (which could include background bitmap writes), then 2621 * retry if dirty blocks still exist. 2622 */ 2623 if (waitfor == MNT_WAIT) { 2624 bufobj_wwait(bo, 0, 0); 2625 if (bo->bo_dirty.bv_cnt > 0) { 2626 /* 2627 * If we are unable to write any of these buffers 2628 * then we fail now rather than trying endlessly 2629 * to write them out. 2630 */ 2631 TAILQ_FOREACH(bp, &bo->bo_dirty.bv_hd, b_bobufs) 2632 if ((error = bp->b_error) != 0) 2633 break; 2634 if ((mp != NULL && mp->mnt_secondary_writes > 0) || 2635 (error == 0 && --maxretry >= 0)) 2636 goto loop1; 2637 if (error == 0) 2638 error = EAGAIN; 2639 } 2640 } 2641 BO_UNLOCK(bo); 2642 if (error != 0) 2643 vn_printf(vp, "fsync: giving up on dirty (error = %d) ", error); 2644 2645 return (error); 2646 } 2647 2648 /* 2649 * Copies a byte range from invp to outvp. Calls VOP_COPY_FILE_RANGE() 2650 * or vn_generic_copy_file_range() after rangelocking the byte ranges, 2651 * to do the actual copy. 2652 * vn_generic_copy_file_range() is factored out, so it can be called 2653 * from a VOP_COPY_FILE_RANGE() call as well, but handles vnodes from 2654 * different file systems. 2655 */ 2656 int 2657 vn_copy_file_range(struct vnode *invp, off_t *inoffp, struct vnode *outvp, 2658 off_t *outoffp, size_t *lenp, unsigned int flags, struct ucred *incred, 2659 struct ucred *outcred, struct thread *fsize_td) 2660 { 2661 int error; 2662 size_t len; 2663 uint64_t uvalin, uvalout; 2664 2665 len = *lenp; 2666 *lenp = 0; /* For error returns. */ 2667 error = 0; 2668 2669 /* Do some sanity checks on the arguments. */ 2670 uvalin = *inoffp; 2671 uvalin += len; 2672 uvalout = *outoffp; 2673 uvalout += len; 2674 if (invp->v_type == VDIR || outvp->v_type == VDIR) 2675 error = EISDIR; 2676 else if (*inoffp < 0 || uvalin > INT64_MAX || uvalin < 2677 (uint64_t)*inoffp || *outoffp < 0 || uvalout > INT64_MAX || 2678 uvalout < (uint64_t)*outoffp || invp->v_type != VREG || 2679 outvp->v_type != VREG) 2680 error = EINVAL; 2681 if (error != 0) 2682 goto out; 2683 2684 /* 2685 * If the two vnode are for the same file system, call 2686 * VOP_COPY_FILE_RANGE(), otherwise call vn_generic_copy_file_range() 2687 * which can handle copies across multiple file systems. 2688 */ 2689 *lenp = len; 2690 if (invp->v_mount == outvp->v_mount) 2691 error = VOP_COPY_FILE_RANGE(invp, inoffp, outvp, outoffp, 2692 lenp, flags, incred, outcred, fsize_td); 2693 else 2694 error = vn_generic_copy_file_range(invp, inoffp, outvp, 2695 outoffp, lenp, flags, incred, outcred, fsize_td); 2696 out: 2697 return (error); 2698 } 2699 2700 /* 2701 * Test len bytes of data starting at dat for all bytes == 0. 2702 * Return true if all bytes are zero, false otherwise. 2703 * Expects dat to be well aligned. 2704 */ 2705 static bool 2706 mem_iszero(void *dat, int len) 2707 { 2708 int i; 2709 const u_int *p; 2710 const char *cp; 2711 2712 for (p = dat; len > 0; len -= sizeof(*p), p++) { 2713 if (len >= sizeof(*p)) { 2714 if (*p != 0) 2715 return (false); 2716 } else { 2717 cp = (const char *)p; 2718 for (i = 0; i < len; i++, cp++) 2719 if (*cp != '\0') 2720 return (false); 2721 } 2722 } 2723 return (true); 2724 } 2725 2726 /* 2727 * Look for a hole in the output file and, if found, adjust *outoffp 2728 * and *xferp to skip past the hole. 2729 * *xferp is the entire hole length to be written and xfer2 is how many bytes 2730 * to be written as 0's upon return. 2731 */ 2732 static off_t 2733 vn_skip_hole(struct vnode *outvp, off_t xfer2, off_t *outoffp, off_t *xferp, 2734 off_t *dataoffp, off_t *holeoffp, struct ucred *cred) 2735 { 2736 int error; 2737 off_t delta; 2738 2739 if (*holeoffp == 0 || *holeoffp <= *outoffp) { 2740 *dataoffp = *outoffp; 2741 error = VOP_IOCTL(outvp, FIOSEEKDATA, dataoffp, 0, cred, 2742 curthread); 2743 if (error == 0) { 2744 *holeoffp = *dataoffp; 2745 error = VOP_IOCTL(outvp, FIOSEEKHOLE, holeoffp, 0, cred, 2746 curthread); 2747 } 2748 if (error != 0 || *holeoffp == *dataoffp) { 2749 /* 2750 * Since outvp is unlocked, it may be possible for 2751 * another thread to do a truncate(), lseek(), write() 2752 * creating a hole at startoff between the above 2753 * VOP_IOCTL() calls, if the other thread does not do 2754 * rangelocking. 2755 * If that happens, *holeoffp == *dataoffp and finding 2756 * the hole has failed, so disable vn_skip_hole(). 2757 */ 2758 *holeoffp = -1; /* Disable use of vn_skip_hole(). */ 2759 return (xfer2); 2760 } 2761 KASSERT(*dataoffp >= *outoffp, 2762 ("vn_skip_hole: dataoff=%jd < outoff=%jd", 2763 (intmax_t)*dataoffp, (intmax_t)*outoffp)); 2764 KASSERT(*holeoffp > *dataoffp, 2765 ("vn_skip_hole: holeoff=%jd <= dataoff=%jd", 2766 (intmax_t)*holeoffp, (intmax_t)*dataoffp)); 2767 } 2768 2769 /* 2770 * If there is a hole before the data starts, advance *outoffp and 2771 * *xferp past the hole. 2772 */ 2773 if (*dataoffp > *outoffp) { 2774 delta = *dataoffp - *outoffp; 2775 if (delta >= *xferp) { 2776 /* Entire *xferp is a hole. */ 2777 *outoffp += *xferp; 2778 *xferp = 0; 2779 return (0); 2780 } 2781 *xferp -= delta; 2782 *outoffp += delta; 2783 xfer2 = MIN(xfer2, *xferp); 2784 } 2785 2786 /* 2787 * If a hole starts before the end of this xfer2, reduce this xfer2 so 2788 * that the write ends at the start of the hole. 2789 * *holeoffp should always be greater than *outoffp, but for the 2790 * non-INVARIANTS case, check this to make sure xfer2 remains a sane 2791 * value. 2792 */ 2793 if (*holeoffp > *outoffp && *holeoffp < *outoffp + xfer2) 2794 xfer2 = *holeoffp - *outoffp; 2795 return (xfer2); 2796 } 2797 2798 /* 2799 * Write an xfer sized chunk to outvp in blksize blocks from dat. 2800 * dat is a maximum of blksize in length and can be written repeatedly in 2801 * the chunk. 2802 * If growfile == true, just grow the file via vn_truncate_locked() instead 2803 * of doing actual writes. 2804 * If checkhole == true, a hole is being punched, so skip over any hole 2805 * already in the output file. 2806 */ 2807 static int 2808 vn_write_outvp(struct vnode *outvp, char *dat, off_t outoff, off_t xfer, 2809 u_long blksize, bool growfile, bool checkhole, struct ucred *cred) 2810 { 2811 struct mount *mp; 2812 off_t dataoff, holeoff, xfer2; 2813 int error, lckf; 2814 2815 /* 2816 * Loop around doing writes of blksize until write has been completed. 2817 * Lock/unlock on each loop iteration so that a bwillwrite() can be 2818 * done for each iteration, since the xfer argument can be very 2819 * large if there is a large hole to punch in the output file. 2820 */ 2821 error = 0; 2822 holeoff = 0; 2823 do { 2824 xfer2 = MIN(xfer, blksize); 2825 if (checkhole) { 2826 /* 2827 * Punching a hole. Skip writing if there is 2828 * already a hole in the output file. 2829 */ 2830 xfer2 = vn_skip_hole(outvp, xfer2, &outoff, &xfer, 2831 &dataoff, &holeoff, cred); 2832 if (xfer == 0) 2833 break; 2834 if (holeoff < 0) 2835 checkhole = false; 2836 KASSERT(xfer2 > 0, ("vn_write_outvp: xfer2=%jd", 2837 (intmax_t)xfer2)); 2838 } 2839 bwillwrite(); 2840 mp = NULL; 2841 error = vn_start_write(outvp, &mp, V_WAIT); 2842 if (error == 0) { 2843 if (MNT_SHARED_WRITES(mp)) 2844 lckf = LK_SHARED; 2845 else 2846 lckf = LK_EXCLUSIVE; 2847 error = vn_lock(outvp, lckf); 2848 } 2849 if (error == 0) { 2850 if (growfile) 2851 error = vn_truncate_locked(outvp, outoff + xfer, 2852 false, cred); 2853 else { 2854 error = vn_rdwr(UIO_WRITE, outvp, dat, xfer2, 2855 outoff, UIO_SYSSPACE, IO_NODELOCKED, 2856 curthread->td_ucred, cred, NULL, curthread); 2857 outoff += xfer2; 2858 xfer -= xfer2; 2859 } 2860 VOP_UNLOCK(outvp); 2861 } 2862 if (mp != NULL) 2863 vn_finished_write(mp); 2864 } while (!growfile && xfer > 0 && error == 0); 2865 return (error); 2866 } 2867 2868 /* 2869 * Copy a byte range of one file to another. This function can handle the 2870 * case where invp and outvp are on different file systems. 2871 * It can also be called by a VOP_COPY_FILE_RANGE() to do the work, if there 2872 * is no better file system specific way to do it. 2873 */ 2874 int 2875 vn_generic_copy_file_range(struct vnode *invp, off_t *inoffp, 2876 struct vnode *outvp, off_t *outoffp, size_t *lenp, unsigned int flags, 2877 struct ucred *incred, struct ucred *outcred, struct thread *fsize_td) 2878 { 2879 struct vattr va; 2880 struct mount *mp; 2881 struct uio io; 2882 off_t startoff, endoff, xfer, xfer2; 2883 u_long blksize; 2884 int error; 2885 bool cantseek, readzeros, eof, lastblock; 2886 ssize_t aresid; 2887 size_t copylen, len, savlen; 2888 char *dat; 2889 long holein, holeout; 2890 2891 holein = holeout = 0; 2892 savlen = len = *lenp; 2893 error = 0; 2894 dat = NULL; 2895 2896 error = vn_lock(invp, LK_SHARED); 2897 if (error != 0) 2898 goto out; 2899 if (VOP_PATHCONF(invp, _PC_MIN_HOLE_SIZE, &holein) != 0) 2900 holein = 0; 2901 VOP_UNLOCK(invp); 2902 2903 mp = NULL; 2904 error = vn_start_write(outvp, &mp, V_WAIT); 2905 if (error == 0) 2906 error = vn_lock(outvp, LK_EXCLUSIVE); 2907 if (error == 0) { 2908 /* 2909 * If fsize_td != NULL, do a vn_rlimit_fsize() call, 2910 * now that outvp is locked. 2911 */ 2912 if (fsize_td != NULL) { 2913 io.uio_offset = *outoffp; 2914 io.uio_resid = len; 2915 error = vn_rlimit_fsize(outvp, &io, fsize_td); 2916 if (error != 0) 2917 error = EFBIG; 2918 } 2919 if (VOP_PATHCONF(outvp, _PC_MIN_HOLE_SIZE, &holeout) != 0) 2920 holeout = 0; 2921 /* 2922 * Holes that are past EOF do not need to be written as a block 2923 * of zero bytes. So, truncate the output file as far as 2924 * possible and then use va.va_size to decide if writing 0 2925 * bytes is necessary in the loop below. 2926 */ 2927 if (error == 0) 2928 error = VOP_GETATTR(outvp, &va, outcred); 2929 if (error == 0 && va.va_size > *outoffp && va.va_size <= 2930 *outoffp + len) { 2931 #ifdef MAC 2932 error = mac_vnode_check_write(curthread->td_ucred, 2933 outcred, outvp); 2934 if (error == 0) 2935 #endif 2936 error = vn_truncate_locked(outvp, *outoffp, 2937 false, outcred); 2938 if (error == 0) 2939 va.va_size = *outoffp; 2940 } 2941 VOP_UNLOCK(outvp); 2942 } 2943 if (mp != NULL) 2944 vn_finished_write(mp); 2945 if (error != 0) 2946 goto out; 2947 2948 /* 2949 * Set the blksize to the larger of the hole sizes for invp and outvp. 2950 * If hole sizes aren't available, set the blksize to the larger 2951 * f_iosize of invp and outvp. 2952 * This code expects the hole sizes and f_iosizes to be powers of 2. 2953 * This value is clipped at 4Kbytes and 1Mbyte. 2954 */ 2955 blksize = MAX(holein, holeout); 2956 if (blksize == 0) 2957 blksize = MAX(invp->v_mount->mnt_stat.f_iosize, 2958 outvp->v_mount->mnt_stat.f_iosize); 2959 if (blksize < 4096) 2960 blksize = 4096; 2961 else if (blksize > 1024 * 1024) 2962 blksize = 1024 * 1024; 2963 dat = malloc(blksize, M_TEMP, M_WAITOK); 2964 2965 /* 2966 * If VOP_IOCTL(FIOSEEKHOLE) works for invp, use it and FIOSEEKDATA 2967 * to find holes. Otherwise, just scan the read block for all 0s 2968 * in the inner loop where the data copying is done. 2969 * Note that some file systems such as NFSv3, NFSv4.0 and NFSv4.1 may 2970 * support holes on the server, but do not support FIOSEEKHOLE. 2971 */ 2972 eof = false; 2973 while (len > 0 && error == 0 && !eof) { 2974 endoff = 0; /* To shut up compilers. */ 2975 cantseek = true; 2976 startoff = *inoffp; 2977 copylen = len; 2978 2979 /* 2980 * Find the next data area. If there is just a hole to EOF, 2981 * FIOSEEKDATA should fail and then we drop down into the 2982 * inner loop and create the hole on the outvp file. 2983 * (I do not know if any file system will report a hole to 2984 * EOF via FIOSEEKHOLE, but I am pretty sure FIOSEEKDATA 2985 * will fail for those file systems.) 2986 * 2987 * For input files that don't support FIOSEEKDATA/FIOSEEKHOLE, 2988 * the code just falls through to the inner copy loop. 2989 */ 2990 error = EINVAL; 2991 if (holein > 0) 2992 error = VOP_IOCTL(invp, FIOSEEKDATA, &startoff, 0, 2993 incred, curthread); 2994 if (error == 0) { 2995 endoff = startoff; 2996 error = VOP_IOCTL(invp, FIOSEEKHOLE, &endoff, 0, 2997 incred, curthread); 2998 /* 2999 * Since invp is unlocked, it may be possible for 3000 * another thread to do a truncate(), lseek(), write() 3001 * creating a hole at startoff between the above 3002 * VOP_IOCTL() calls, if the other thread does not do 3003 * rangelocking. 3004 * If that happens, startoff == endoff and finding 3005 * the hole has failed, so set an error. 3006 */ 3007 if (error == 0 && startoff == endoff) 3008 error = EINVAL; /* Any error. Reset to 0. */ 3009 } 3010 if (error == 0) { 3011 if (startoff > *inoffp) { 3012 /* Found hole before data block. */ 3013 xfer = MIN(startoff - *inoffp, len); 3014 if (*outoffp < va.va_size) { 3015 /* Must write 0s to punch hole. */ 3016 xfer2 = MIN(va.va_size - *outoffp, 3017 xfer); 3018 memset(dat, 0, MIN(xfer2, blksize)); 3019 error = vn_write_outvp(outvp, dat, 3020 *outoffp, xfer2, blksize, false, 3021 holeout > 0, outcred); 3022 } 3023 3024 if (error == 0 && *outoffp + xfer > 3025 va.va_size && xfer == len) 3026 /* Grow last block. */ 3027 error = vn_write_outvp(outvp, dat, 3028 *outoffp, xfer, blksize, true, 3029 false, outcred); 3030 if (error == 0) { 3031 *inoffp += xfer; 3032 *outoffp += xfer; 3033 len -= xfer; 3034 } 3035 } 3036 copylen = MIN(len, endoff - startoff); 3037 cantseek = false; 3038 } else { 3039 cantseek = true; 3040 startoff = *inoffp; 3041 copylen = len; 3042 error = 0; 3043 } 3044 3045 xfer = blksize; 3046 if (cantseek) { 3047 /* 3048 * Set first xfer to end at a block boundary, so that 3049 * holes are more likely detected in the loop below via 3050 * the for all bytes 0 method. 3051 */ 3052 xfer -= (*inoffp % blksize); 3053 } 3054 /* Loop copying the data block. */ 3055 while (copylen > 0 && error == 0 && !eof) { 3056 if (copylen < xfer) 3057 xfer = copylen; 3058 error = vn_lock(invp, LK_SHARED); 3059 if (error != 0) 3060 goto out; 3061 error = vn_rdwr(UIO_READ, invp, dat, xfer, 3062 startoff, UIO_SYSSPACE, IO_NODELOCKED, 3063 curthread->td_ucred, incred, &aresid, 3064 curthread); 3065 VOP_UNLOCK(invp); 3066 lastblock = false; 3067 if (error == 0 && aresid > 0) { 3068 /* Stop the copy at EOF on the input file. */ 3069 xfer -= aresid; 3070 eof = true; 3071 lastblock = true; 3072 } 3073 if (error == 0) { 3074 /* 3075 * Skip the write for holes past the initial EOF 3076 * of the output file, unless this is the last 3077 * write of the output file at EOF. 3078 */ 3079 readzeros = cantseek ? mem_iszero(dat, xfer) : 3080 false; 3081 if (xfer == len) 3082 lastblock = true; 3083 if (!cantseek || *outoffp < va.va_size || 3084 lastblock || !readzeros) 3085 error = vn_write_outvp(outvp, dat, 3086 *outoffp, xfer, blksize, 3087 readzeros && lastblock && 3088 *outoffp >= va.va_size, false, 3089 outcred); 3090 if (error == 0) { 3091 *inoffp += xfer; 3092 startoff += xfer; 3093 *outoffp += xfer; 3094 copylen -= xfer; 3095 len -= xfer; 3096 } 3097 } 3098 xfer = blksize; 3099 } 3100 } 3101 out: 3102 *lenp = savlen - len; 3103 free(dat, M_TEMP); 3104 return (error); 3105 } 3106 3107 static int 3108 vn_fallocate(struct file *fp, off_t offset, off_t len, struct thread *td) 3109 { 3110 struct mount *mp; 3111 struct vnode *vp; 3112 off_t olen, ooffset; 3113 int error; 3114 #ifdef AUDIT 3115 int audited_vnode1 = 0; 3116 #endif 3117 3118 vp = fp->f_vnode; 3119 if (vp->v_type != VREG) 3120 return (ENODEV); 3121 3122 /* Allocating blocks may take a long time, so iterate. */ 3123 for (;;) { 3124 olen = len; 3125 ooffset = offset; 3126 3127 bwillwrite(); 3128 mp = NULL; 3129 error = vn_start_write(vp, &mp, V_WAIT | PCATCH); 3130 if (error != 0) 3131 break; 3132 error = vn_lock(vp, LK_EXCLUSIVE); 3133 if (error != 0) { 3134 vn_finished_write(mp); 3135 break; 3136 } 3137 #ifdef AUDIT 3138 if (!audited_vnode1) { 3139 AUDIT_ARG_VNODE1(vp); 3140 audited_vnode1 = 1; 3141 } 3142 #endif 3143 #ifdef MAC 3144 error = mac_vnode_check_write(td->td_ucred, fp->f_cred, vp); 3145 if (error == 0) 3146 #endif 3147 error = VOP_ALLOCATE(vp, &offset, &len); 3148 VOP_UNLOCK(vp); 3149 vn_finished_write(mp); 3150 3151 if (olen + ooffset != offset + len) { 3152 panic("offset + len changed from %jx/%jx to %jx/%jx", 3153 ooffset, olen, offset, len); 3154 } 3155 if (error != 0 || len == 0) 3156 break; 3157 KASSERT(olen > len, ("Iteration did not make progress?")); 3158 maybe_yield(); 3159 } 3160 3161 return (error); 3162 } 3163