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