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/sysctl.h> 75 #include <sys/ttycom.h> 76 #include <sys/conf.h> 77 #include <sys/syslog.h> 78 #include <sys/unistd.h> 79 #include <sys/user.h> 80 81 #include <security/audit/audit.h> 82 #include <security/mac/mac_framework.h> 83 84 #include <vm/vm.h> 85 #include <vm/vm_extern.h> 86 #include <vm/pmap.h> 87 #include <vm/vm_map.h> 88 #include <vm/vm_object.h> 89 #include <vm/vm_page.h> 90 #include <vm/vm_pager.h> 91 92 #ifdef HWPMC_HOOKS 93 #include <sys/pmckern.h> 94 #endif 95 96 static fo_rdwr_t vn_read; 97 static fo_rdwr_t vn_write; 98 static fo_rdwr_t vn_io_fault; 99 static fo_truncate_t vn_truncate; 100 static fo_ioctl_t vn_ioctl; 101 static fo_poll_t vn_poll; 102 static fo_kqfilter_t vn_kqfilter; 103 static fo_stat_t vn_statfile; 104 static fo_close_t vn_closefile; 105 static fo_mmap_t vn_mmap; 106 static fo_fallocate_t vn_fallocate; 107 108 struct fileops vnops = { 109 .fo_read = vn_io_fault, 110 .fo_write = vn_io_fault, 111 .fo_truncate = vn_truncate, 112 .fo_ioctl = vn_ioctl, 113 .fo_poll = vn_poll, 114 .fo_kqfilter = vn_kqfilter, 115 .fo_stat = vn_statfile, 116 .fo_close = vn_closefile, 117 .fo_chmod = vn_chmod, 118 .fo_chown = vn_chown, 119 .fo_sendfile = vn_sendfile, 120 .fo_seek = vn_seek, 121 .fo_fill_kinfo = vn_fill_kinfo, 122 .fo_mmap = vn_mmap, 123 .fo_fallocate = vn_fallocate, 124 .fo_flags = DFLAG_PASSABLE | DFLAG_SEEKABLE 125 }; 126 127 static const int io_hold_cnt = 16; 128 static int vn_io_fault_enable = 1; 129 SYSCTL_INT(_debug, OID_AUTO, vn_io_fault_enable, CTLFLAG_RW, 130 &vn_io_fault_enable, 0, "Enable vn_io_fault lock avoidance"); 131 static int vn_io_fault_prefault = 0; 132 SYSCTL_INT(_debug, OID_AUTO, vn_io_fault_prefault, CTLFLAG_RW, 133 &vn_io_fault_prefault, 0, "Enable vn_io_fault prefaulting"); 134 static u_long vn_io_faults_cnt; 135 SYSCTL_ULONG(_debug, OID_AUTO, vn_io_faults, CTLFLAG_RD, 136 &vn_io_faults_cnt, 0, "Count of vn_io_fault lock avoidance triggers"); 137 138 /* 139 * Returns true if vn_io_fault mode of handling the i/o request should 140 * be used. 141 */ 142 static bool 143 do_vn_io_fault(struct vnode *vp, struct uio *uio) 144 { 145 struct mount *mp; 146 147 return (uio->uio_segflg == UIO_USERSPACE && vp->v_type == VREG && 148 (mp = vp->v_mount) != NULL && 149 (mp->mnt_kern_flag & MNTK_NO_IOPF) != 0 && vn_io_fault_enable); 150 } 151 152 /* 153 * Structure used to pass arguments to vn_io_fault1(), to do either 154 * file- or vnode-based I/O calls. 155 */ 156 struct vn_io_fault_args { 157 enum { 158 VN_IO_FAULT_FOP, 159 VN_IO_FAULT_VOP 160 } kind; 161 struct ucred *cred; 162 int flags; 163 union { 164 struct fop_args_tag { 165 struct file *fp; 166 fo_rdwr_t *doio; 167 } fop_args; 168 struct vop_args_tag { 169 struct vnode *vp; 170 } vop_args; 171 } args; 172 }; 173 174 static int vn_io_fault1(struct vnode *vp, struct uio *uio, 175 struct vn_io_fault_args *args, struct thread *td); 176 177 int 178 vn_open(struct nameidata *ndp, int *flagp, int cmode, struct file *fp) 179 { 180 struct thread *td = ndp->ni_cnd.cn_thread; 181 182 return (vn_open_cred(ndp, flagp, cmode, 0, td->td_ucred, fp)); 183 } 184 185 /* 186 * Common code for vnode open operations via a name lookup. 187 * Lookup the vnode and invoke VOP_CREATE if needed. 188 * Check permissions, and call the VOP_OPEN or VOP_CREATE routine. 189 * 190 * Note that this does NOT free nameidata for the successful case, 191 * due to the NDINIT being done elsewhere. 192 */ 193 int 194 vn_open_cred(struct nameidata *ndp, int *flagp, int cmode, u_int vn_open_flags, 195 struct ucred *cred, struct file *fp) 196 { 197 struct vnode *vp; 198 struct mount *mp; 199 struct thread *td = ndp->ni_cnd.cn_thread; 200 struct vattr vat; 201 struct vattr *vap = &vat; 202 int fmode, error; 203 204 restart: 205 fmode = *flagp; 206 if ((fmode & (O_CREAT | O_EXCL | O_DIRECTORY)) == (O_CREAT | 207 O_EXCL | O_DIRECTORY)) 208 return (EINVAL); 209 else if ((fmode & (O_CREAT | O_DIRECTORY)) == O_CREAT) { 210 ndp->ni_cnd.cn_nameiop = CREATE; 211 /* 212 * Set NOCACHE to avoid flushing the cache when 213 * rolling in many files at once. 214 */ 215 ndp->ni_cnd.cn_flags = ISOPEN | LOCKPARENT | LOCKLEAF | NOCACHE; 216 if ((fmode & O_EXCL) == 0 && (fmode & O_NOFOLLOW) == 0) 217 ndp->ni_cnd.cn_flags |= FOLLOW; 218 if ((fmode & O_BENEATH) != 0) 219 ndp->ni_cnd.cn_flags |= BENEATH; 220 if (!(vn_open_flags & VN_OPEN_NOAUDIT)) 221 ndp->ni_cnd.cn_flags |= AUDITVNODE1; 222 if (vn_open_flags & VN_OPEN_NOCAPCHECK) 223 ndp->ni_cnd.cn_flags |= NOCAPCHECK; 224 if ((vn_open_flags & VN_OPEN_INVFS) == 0) 225 bwillwrite(); 226 if ((error = namei(ndp)) != 0) 227 return (error); 228 if (ndp->ni_vp == NULL) { 229 VATTR_NULL(vap); 230 vap->va_type = VREG; 231 vap->va_mode = cmode; 232 if (fmode & O_EXCL) 233 vap->va_vaflags |= VA_EXCLUSIVE; 234 if (vn_start_write(ndp->ni_dvp, &mp, V_NOWAIT) != 0) { 235 NDFREE(ndp, NDF_ONLY_PNBUF); 236 vput(ndp->ni_dvp); 237 if ((error = vn_start_write(NULL, &mp, 238 V_XSLEEP | PCATCH)) != 0) 239 return (error); 240 goto restart; 241 } 242 if ((vn_open_flags & VN_OPEN_NAMECACHE) != 0) 243 ndp->ni_cnd.cn_flags |= MAKEENTRY; 244 #ifdef MAC 245 error = mac_vnode_check_create(cred, ndp->ni_dvp, 246 &ndp->ni_cnd, vap); 247 if (error == 0) 248 #endif 249 error = VOP_CREATE(ndp->ni_dvp, &ndp->ni_vp, 250 &ndp->ni_cnd, vap); 251 vput(ndp->ni_dvp); 252 vn_finished_write(mp); 253 if (error) { 254 NDFREE(ndp, NDF_ONLY_PNBUF); 255 return (error); 256 } 257 fmode &= ~O_TRUNC; 258 vp = ndp->ni_vp; 259 } else { 260 if (ndp->ni_dvp == ndp->ni_vp) 261 vrele(ndp->ni_dvp); 262 else 263 vput(ndp->ni_dvp); 264 ndp->ni_dvp = NULL; 265 vp = ndp->ni_vp; 266 if (fmode & O_EXCL) { 267 error = EEXIST; 268 goto bad; 269 } 270 if (vp->v_type == VDIR) { 271 error = EISDIR; 272 goto bad; 273 } 274 fmode &= ~O_CREAT; 275 } 276 } else { 277 ndp->ni_cnd.cn_nameiop = LOOKUP; 278 ndp->ni_cnd.cn_flags = ISOPEN | 279 ((fmode & O_NOFOLLOW) ? NOFOLLOW : FOLLOW) | LOCKLEAF; 280 if (!(fmode & FWRITE)) 281 ndp->ni_cnd.cn_flags |= LOCKSHARED; 282 if ((fmode & O_BENEATH) != 0) 283 ndp->ni_cnd.cn_flags |= BENEATH; 284 if (!(vn_open_flags & VN_OPEN_NOAUDIT)) 285 ndp->ni_cnd.cn_flags |= AUDITVNODE1; 286 if (vn_open_flags & VN_OPEN_NOCAPCHECK) 287 ndp->ni_cnd.cn_flags |= NOCAPCHECK; 288 if ((error = namei(ndp)) != 0) 289 return (error); 290 vp = ndp->ni_vp; 291 } 292 error = vn_open_vnode(vp, fmode, cred, td, fp); 293 if (error) 294 goto bad; 295 *flagp = fmode; 296 return (0); 297 bad: 298 NDFREE(ndp, NDF_ONLY_PNBUF); 299 vput(vp); 300 *flagp = fmode; 301 ndp->ni_vp = NULL; 302 return (error); 303 } 304 305 static int 306 vn_open_vnode_advlock(struct vnode *vp, int fmode, struct file *fp) 307 { 308 struct flock lf; 309 int error, lock_flags, type; 310 311 ASSERT_VOP_LOCKED(vp, "vn_open_vnode_advlock"); 312 if ((fmode & (O_EXLOCK | O_SHLOCK)) == 0) 313 return (0); 314 KASSERT(fp != NULL, ("open with flock requires fp")); 315 if (fp->f_type != DTYPE_NONE && fp->f_type != DTYPE_VNODE) 316 return (EOPNOTSUPP); 317 318 lock_flags = VOP_ISLOCKED(vp); 319 VOP_UNLOCK(vp); 320 321 lf.l_whence = SEEK_SET; 322 lf.l_start = 0; 323 lf.l_len = 0; 324 lf.l_type = (fmode & O_EXLOCK) != 0 ? F_WRLCK : F_RDLCK; 325 type = F_FLOCK; 326 if ((fmode & FNONBLOCK) == 0) 327 type |= F_WAIT; 328 error = VOP_ADVLOCK(vp, (caddr_t)fp, F_SETLK, &lf, type); 329 if (error == 0) 330 fp->f_flag |= FHASLOCK; 331 332 vn_lock(vp, lock_flags | LK_RETRY); 333 if (error == 0 && VN_IS_DOOMED(vp)) 334 error = ENOENT; 335 return (error); 336 } 337 338 /* 339 * Common code for vnode open operations once a vnode is located. 340 * Check permissions, and call the VOP_OPEN routine. 341 */ 342 int 343 vn_open_vnode(struct vnode *vp, int fmode, struct ucred *cred, 344 struct thread *td, struct file *fp) 345 { 346 accmode_t accmode; 347 int error; 348 349 if (vp->v_type == VLNK) 350 return (EMLINK); 351 if (vp->v_type == VSOCK) 352 return (EOPNOTSUPP); 353 if (vp->v_type != VDIR && fmode & O_DIRECTORY) 354 return (ENOTDIR); 355 accmode = 0; 356 if (fmode & (FWRITE | O_TRUNC)) { 357 if (vp->v_type == VDIR) 358 return (EISDIR); 359 accmode |= VWRITE; 360 } 361 if (fmode & FREAD) 362 accmode |= VREAD; 363 if (fmode & FEXEC) 364 accmode |= VEXEC; 365 if ((fmode & O_APPEND) && (fmode & FWRITE)) 366 accmode |= VAPPEND; 367 #ifdef MAC 368 if (fmode & O_CREAT) 369 accmode |= VCREAT; 370 if (fmode & O_VERIFY) 371 accmode |= VVERIFY; 372 error = mac_vnode_check_open(cred, vp, accmode); 373 if (error) 374 return (error); 375 376 accmode &= ~(VCREAT | VVERIFY); 377 #endif 378 if ((fmode & O_CREAT) == 0 && accmode != 0) { 379 error = VOP_ACCESS(vp, accmode, cred, td); 380 if (error != 0) 381 return (error); 382 } 383 if (vp->v_type == VFIFO && VOP_ISLOCKED(vp) != LK_EXCLUSIVE) 384 vn_lock(vp, LK_UPGRADE | LK_RETRY); 385 error = VOP_OPEN(vp, fmode, cred, td, fp); 386 if (error != 0) 387 return (error); 388 389 error = vn_open_vnode_advlock(vp, fmode, fp); 390 if (error == 0 && (fmode & FWRITE) != 0) { 391 error = VOP_ADD_WRITECOUNT(vp, 1); 392 if (error == 0) { 393 CTR3(KTR_VFS, "%s: vp %p v_writecount increased to %d", 394 __func__, vp, vp->v_writecount); 395 } 396 } 397 398 /* 399 * Error from advlock or VOP_ADD_WRITECOUNT() still requires 400 * calling VOP_CLOSE() to pair with earlier VOP_OPEN(). 401 * Arrange for that by having fdrop() to use vn_closefile(). 402 */ 403 if (error != 0) { 404 fp->f_flag |= FOPENFAILED; 405 fp->f_vnode = vp; 406 if (fp->f_ops == &badfileops) { 407 fp->f_type = DTYPE_VNODE; 408 fp->f_ops = &vnops; 409 } 410 vref(vp); 411 } 412 413 ASSERT_VOP_LOCKED(vp, "vn_open_vnode"); 414 return (error); 415 416 } 417 418 /* 419 * Check for write permissions on the specified vnode. 420 * Prototype text segments cannot be written. 421 * It is racy. 422 */ 423 int 424 vn_writechk(struct vnode *vp) 425 { 426 427 ASSERT_VOP_LOCKED(vp, "vn_writechk"); 428 /* 429 * If there's shared text associated with 430 * the vnode, try to free it up once. If 431 * we fail, we can't allow writing. 432 */ 433 if (VOP_IS_TEXT(vp)) 434 return (ETXTBSY); 435 436 return (0); 437 } 438 439 /* 440 * Vnode close call 441 */ 442 static int 443 vn_close1(struct vnode *vp, int flags, struct ucred *file_cred, 444 struct thread *td, bool keep_ref) 445 { 446 struct mount *mp; 447 int error, lock_flags; 448 449 if (vp->v_type != VFIFO && (flags & FWRITE) == 0 && 450 MNT_EXTENDED_SHARED(vp->v_mount)) 451 lock_flags = LK_SHARED; 452 else 453 lock_flags = LK_EXCLUSIVE; 454 455 vn_start_write(vp, &mp, V_WAIT); 456 vn_lock(vp, lock_flags | LK_RETRY); 457 AUDIT_ARG_VNODE1(vp); 458 if ((flags & (FWRITE | FOPENFAILED)) == FWRITE) { 459 VOP_ADD_WRITECOUNT_CHECKED(vp, -1); 460 CTR3(KTR_VFS, "%s: vp %p v_writecount decreased to %d", 461 __func__, vp, vp->v_writecount); 462 } 463 error = VOP_CLOSE(vp, flags, file_cred, td); 464 if (keep_ref) 465 VOP_UNLOCK(vp); 466 else 467 vput(vp); 468 vn_finished_write(mp); 469 return (error); 470 } 471 472 int 473 vn_close(struct vnode *vp, int flags, struct ucred *file_cred, 474 struct thread *td) 475 { 476 477 return (vn_close1(vp, flags, file_cred, td, false)); 478 } 479 480 /* 481 * Heuristic to detect sequential operation. 482 */ 483 static int 484 sequential_heuristic(struct uio *uio, struct file *fp) 485 { 486 487 ASSERT_VOP_LOCKED(fp->f_vnode, __func__); 488 if (fp->f_flag & FRDAHEAD) 489 return (fp->f_seqcount << IO_SEQSHIFT); 490 491 /* 492 * Offset 0 is handled specially. open() sets f_seqcount to 1 so 493 * that the first I/O is normally considered to be slightly 494 * sequential. Seeking to offset 0 doesn't change sequentiality 495 * unless previous seeks have reduced f_seqcount to 0, in which 496 * case offset 0 is not special. 497 */ 498 if ((uio->uio_offset == 0 && fp->f_seqcount > 0) || 499 uio->uio_offset == fp->f_nextoff) { 500 /* 501 * f_seqcount is in units of fixed-size blocks so that it 502 * depends mainly on the amount of sequential I/O and not 503 * much on the number of sequential I/O's. The fixed size 504 * of 16384 is hard-coded here since it is (not quite) just 505 * a magic size that works well here. This size is more 506 * closely related to the best I/O size for real disks than 507 * to any block size used by software. 508 */ 509 if (uio->uio_resid >= IO_SEQMAX * 16384) 510 fp->f_seqcount = IO_SEQMAX; 511 else { 512 fp->f_seqcount += howmany(uio->uio_resid, 16384); 513 if (fp->f_seqcount > IO_SEQMAX) 514 fp->f_seqcount = IO_SEQMAX; 515 } 516 return (fp->f_seqcount << IO_SEQSHIFT); 517 } 518 519 /* Not sequential. Quickly draw-down sequentiality. */ 520 if (fp->f_seqcount > 1) 521 fp->f_seqcount = 1; 522 else 523 fp->f_seqcount = 0; 524 return (0); 525 } 526 527 /* 528 * Package up an I/O request on a vnode into a uio and do it. 529 */ 530 int 531 vn_rdwr(enum uio_rw rw, struct vnode *vp, void *base, int len, off_t offset, 532 enum uio_seg segflg, int ioflg, struct ucred *active_cred, 533 struct ucred *file_cred, ssize_t *aresid, struct thread *td) 534 { 535 struct uio auio; 536 struct iovec aiov; 537 struct mount *mp; 538 struct ucred *cred; 539 void *rl_cookie; 540 struct vn_io_fault_args args; 541 int error, lock_flags; 542 543 if (offset < 0 && vp->v_type != VCHR) 544 return (EINVAL); 545 auio.uio_iov = &aiov; 546 auio.uio_iovcnt = 1; 547 aiov.iov_base = base; 548 aiov.iov_len = len; 549 auio.uio_resid = len; 550 auio.uio_offset = offset; 551 auio.uio_segflg = segflg; 552 auio.uio_rw = rw; 553 auio.uio_td = td; 554 error = 0; 555 556 if ((ioflg & IO_NODELOCKED) == 0) { 557 if ((ioflg & IO_RANGELOCKED) == 0) { 558 if (rw == UIO_READ) { 559 rl_cookie = vn_rangelock_rlock(vp, offset, 560 offset + len); 561 } else { 562 rl_cookie = vn_rangelock_wlock(vp, offset, 563 offset + len); 564 } 565 } else 566 rl_cookie = NULL; 567 mp = NULL; 568 if (rw == UIO_WRITE) { 569 if (vp->v_type != VCHR && 570 (error = vn_start_write(vp, &mp, V_WAIT | PCATCH)) 571 != 0) 572 goto out; 573 if (MNT_SHARED_WRITES(mp) || 574 ((mp == NULL) && MNT_SHARED_WRITES(vp->v_mount))) 575 lock_flags = LK_SHARED; 576 else 577 lock_flags = LK_EXCLUSIVE; 578 } else 579 lock_flags = LK_SHARED; 580 vn_lock(vp, lock_flags | LK_RETRY); 581 } else 582 rl_cookie = NULL; 583 584 ASSERT_VOP_LOCKED(vp, "IO_NODELOCKED with no vp lock held"); 585 #ifdef MAC 586 if ((ioflg & IO_NOMACCHECK) == 0) { 587 if (rw == UIO_READ) 588 error = mac_vnode_check_read(active_cred, file_cred, 589 vp); 590 else 591 error = mac_vnode_check_write(active_cred, file_cred, 592 vp); 593 } 594 #endif 595 if (error == 0) { 596 if (file_cred != NULL) 597 cred = file_cred; 598 else 599 cred = active_cred; 600 if (do_vn_io_fault(vp, &auio)) { 601 args.kind = VN_IO_FAULT_VOP; 602 args.cred = cred; 603 args.flags = ioflg; 604 args.args.vop_args.vp = vp; 605 error = vn_io_fault1(vp, &auio, &args, td); 606 } else if (rw == UIO_READ) { 607 error = VOP_READ(vp, &auio, ioflg, cred); 608 } else /* if (rw == UIO_WRITE) */ { 609 error = VOP_WRITE(vp, &auio, ioflg, cred); 610 } 611 } 612 if (aresid) 613 *aresid = auio.uio_resid; 614 else 615 if (auio.uio_resid && error == 0) 616 error = EIO; 617 if ((ioflg & IO_NODELOCKED) == 0) { 618 VOP_UNLOCK(vp); 619 if (mp != NULL) 620 vn_finished_write(mp); 621 } 622 out: 623 if (rl_cookie != NULL) 624 vn_rangelock_unlock(vp, rl_cookie); 625 return (error); 626 } 627 628 /* 629 * Package up an I/O request on a vnode into a uio and do it. The I/O 630 * request is split up into smaller chunks and we try to avoid saturating 631 * the buffer cache while potentially holding a vnode locked, so we 632 * check bwillwrite() before calling vn_rdwr(). We also call kern_yield() 633 * to give other processes a chance to lock the vnode (either other processes 634 * core'ing the same binary, or unrelated processes scanning the directory). 635 */ 636 int 637 vn_rdwr_inchunks(enum uio_rw rw, struct vnode *vp, void *base, size_t len, 638 off_t offset, enum uio_seg segflg, int ioflg, struct ucred *active_cred, 639 struct ucred *file_cred, size_t *aresid, struct thread *td) 640 { 641 int error = 0; 642 ssize_t iaresid; 643 644 do { 645 int chunk; 646 647 /* 648 * Force `offset' to a multiple of MAXBSIZE except possibly 649 * for the first chunk, so that filesystems only need to 650 * write full blocks except possibly for the first and last 651 * chunks. 652 */ 653 chunk = MAXBSIZE - (uoff_t)offset % MAXBSIZE; 654 655 if (chunk > len) 656 chunk = len; 657 if (rw != UIO_READ && vp->v_type == VREG) 658 bwillwrite(); 659 iaresid = 0; 660 error = vn_rdwr(rw, vp, base, chunk, offset, segflg, 661 ioflg, active_cred, file_cred, &iaresid, td); 662 len -= chunk; /* aresid calc already includes length */ 663 if (error) 664 break; 665 offset += chunk; 666 base = (char *)base + chunk; 667 kern_yield(PRI_USER); 668 } while (len); 669 if (aresid) 670 *aresid = len + iaresid; 671 return (error); 672 } 673 674 off_t 675 foffset_lock(struct file *fp, int flags) 676 { 677 struct mtx *mtxp; 678 off_t res; 679 680 KASSERT((flags & FOF_OFFSET) == 0, ("FOF_OFFSET passed")); 681 682 #if OFF_MAX <= LONG_MAX 683 /* 684 * Caller only wants the current f_offset value. Assume that 685 * the long and shorter integer types reads are atomic. 686 */ 687 if ((flags & FOF_NOLOCK) != 0) 688 return (fp->f_offset); 689 #endif 690 691 /* 692 * According to McKusick the vn lock was protecting f_offset here. 693 * It is now protected by the FOFFSET_LOCKED flag. 694 */ 695 mtxp = mtx_pool_find(mtxpool_sleep, fp); 696 mtx_lock(mtxp); 697 if ((flags & FOF_NOLOCK) == 0) { 698 while (fp->f_vnread_flags & FOFFSET_LOCKED) { 699 fp->f_vnread_flags |= FOFFSET_LOCK_WAITING; 700 msleep(&fp->f_vnread_flags, mtxp, PUSER -1, 701 "vofflock", 0); 702 } 703 fp->f_vnread_flags |= FOFFSET_LOCKED; 704 } 705 res = fp->f_offset; 706 mtx_unlock(mtxp); 707 return (res); 708 } 709 710 void 711 foffset_unlock(struct file *fp, off_t val, int flags) 712 { 713 struct mtx *mtxp; 714 715 KASSERT((flags & FOF_OFFSET) == 0, ("FOF_OFFSET passed")); 716 717 #if OFF_MAX <= LONG_MAX 718 if ((flags & FOF_NOLOCK) != 0) { 719 if ((flags & FOF_NOUPDATE) == 0) 720 fp->f_offset = val; 721 if ((flags & FOF_NEXTOFF) != 0) 722 fp->f_nextoff = val; 723 return; 724 } 725 #endif 726 727 mtxp = mtx_pool_find(mtxpool_sleep, fp); 728 mtx_lock(mtxp); 729 if ((flags & FOF_NOUPDATE) == 0) 730 fp->f_offset = val; 731 if ((flags & FOF_NEXTOFF) != 0) 732 fp->f_nextoff = val; 733 if ((flags & FOF_NOLOCK) == 0) { 734 KASSERT((fp->f_vnread_flags & FOFFSET_LOCKED) != 0, 735 ("Lost FOFFSET_LOCKED")); 736 if (fp->f_vnread_flags & FOFFSET_LOCK_WAITING) 737 wakeup(&fp->f_vnread_flags); 738 fp->f_vnread_flags = 0; 739 } 740 mtx_unlock(mtxp); 741 } 742 743 void 744 foffset_lock_uio(struct file *fp, struct uio *uio, int flags) 745 { 746 747 if ((flags & FOF_OFFSET) == 0) 748 uio->uio_offset = foffset_lock(fp, flags); 749 } 750 751 void 752 foffset_unlock_uio(struct file *fp, struct uio *uio, int flags) 753 { 754 755 if ((flags & FOF_OFFSET) == 0) 756 foffset_unlock(fp, uio->uio_offset, flags); 757 } 758 759 static int 760 get_advice(struct file *fp, struct uio *uio) 761 { 762 struct mtx *mtxp; 763 int ret; 764 765 ret = POSIX_FADV_NORMAL; 766 if (fp->f_advice == NULL || fp->f_vnode->v_type != VREG) 767 return (ret); 768 769 mtxp = mtx_pool_find(mtxpool_sleep, fp); 770 mtx_lock(mtxp); 771 if (fp->f_advice != NULL && 772 uio->uio_offset >= fp->f_advice->fa_start && 773 uio->uio_offset + uio->uio_resid <= fp->f_advice->fa_end) 774 ret = fp->f_advice->fa_advice; 775 mtx_unlock(mtxp); 776 return (ret); 777 } 778 779 /* 780 * File table vnode read routine. 781 */ 782 static int 783 vn_read(struct file *fp, struct uio *uio, struct ucred *active_cred, int flags, 784 struct thread *td) 785 { 786 struct vnode *vp; 787 off_t orig_offset; 788 int error, ioflag; 789 int advice; 790 791 KASSERT(uio->uio_td == td, ("uio_td %p is not td %p", 792 uio->uio_td, td)); 793 KASSERT(flags & FOF_OFFSET, ("No FOF_OFFSET")); 794 vp = fp->f_vnode; 795 ioflag = 0; 796 if (fp->f_flag & FNONBLOCK) 797 ioflag |= IO_NDELAY; 798 if (fp->f_flag & O_DIRECT) 799 ioflag |= IO_DIRECT; 800 advice = get_advice(fp, uio); 801 vn_lock(vp, LK_SHARED | LK_RETRY); 802 803 switch (advice) { 804 case POSIX_FADV_NORMAL: 805 case POSIX_FADV_SEQUENTIAL: 806 case POSIX_FADV_NOREUSE: 807 ioflag |= sequential_heuristic(uio, fp); 808 break; 809 case POSIX_FADV_RANDOM: 810 /* Disable read-ahead for random I/O. */ 811 break; 812 } 813 orig_offset = uio->uio_offset; 814 815 #ifdef MAC 816 error = mac_vnode_check_read(active_cred, fp->f_cred, vp); 817 if (error == 0) 818 #endif 819 error = VOP_READ(vp, uio, ioflag, fp->f_cred); 820 fp->f_nextoff = uio->uio_offset; 821 VOP_UNLOCK(vp); 822 if (error == 0 && advice == POSIX_FADV_NOREUSE && 823 orig_offset != uio->uio_offset) 824 /* 825 * Use POSIX_FADV_DONTNEED to flush pages and buffers 826 * for the backing file after a POSIX_FADV_NOREUSE 827 * read(2). 828 */ 829 error = VOP_ADVISE(vp, orig_offset, uio->uio_offset - 1, 830 POSIX_FADV_DONTNEED); 831 return (error); 832 } 833 834 /* 835 * File table vnode write routine. 836 */ 837 static int 838 vn_write(struct file *fp, struct uio *uio, struct ucred *active_cred, int flags, 839 struct thread *td) 840 { 841 struct vnode *vp; 842 struct mount *mp; 843 off_t orig_offset; 844 int error, ioflag, lock_flags; 845 int advice; 846 847 KASSERT(uio->uio_td == td, ("uio_td %p is not td %p", 848 uio->uio_td, td)); 849 KASSERT(flags & FOF_OFFSET, ("No FOF_OFFSET")); 850 vp = fp->f_vnode; 851 if (vp->v_type == VREG) 852 bwillwrite(); 853 ioflag = IO_UNIT; 854 if (vp->v_type == VREG && (fp->f_flag & O_APPEND)) 855 ioflag |= IO_APPEND; 856 if (fp->f_flag & FNONBLOCK) 857 ioflag |= IO_NDELAY; 858 if (fp->f_flag & O_DIRECT) 859 ioflag |= IO_DIRECT; 860 if ((fp->f_flag & O_FSYNC) || 861 (vp->v_mount && (vp->v_mount->mnt_flag & MNT_SYNCHRONOUS))) 862 ioflag |= IO_SYNC; 863 mp = NULL; 864 if (vp->v_type != VCHR && 865 (error = vn_start_write(vp, &mp, V_WAIT | PCATCH)) != 0) 866 goto unlock; 867 868 advice = get_advice(fp, uio); 869 870 if (MNT_SHARED_WRITES(mp) || 871 (mp == NULL && MNT_SHARED_WRITES(vp->v_mount))) { 872 lock_flags = LK_SHARED; 873 } else { 874 lock_flags = LK_EXCLUSIVE; 875 } 876 877 vn_lock(vp, lock_flags | LK_RETRY); 878 switch (advice) { 879 case POSIX_FADV_NORMAL: 880 case POSIX_FADV_SEQUENTIAL: 881 case POSIX_FADV_NOREUSE: 882 ioflag |= sequential_heuristic(uio, fp); 883 break; 884 case POSIX_FADV_RANDOM: 885 /* XXX: Is this correct? */ 886 break; 887 } 888 orig_offset = uio->uio_offset; 889 890 #ifdef MAC 891 error = mac_vnode_check_write(active_cred, fp->f_cred, vp); 892 if (error == 0) 893 #endif 894 error = VOP_WRITE(vp, uio, ioflag, fp->f_cred); 895 fp->f_nextoff = uio->uio_offset; 896 VOP_UNLOCK(vp); 897 if (vp->v_type != VCHR) 898 vn_finished_write(mp); 899 if (error == 0 && advice == POSIX_FADV_NOREUSE && 900 orig_offset != uio->uio_offset) 901 /* 902 * Use POSIX_FADV_DONTNEED to flush pages and buffers 903 * for the backing file after a POSIX_FADV_NOREUSE 904 * write(2). 905 */ 906 error = VOP_ADVISE(vp, orig_offset, uio->uio_offset - 1, 907 POSIX_FADV_DONTNEED); 908 unlock: 909 return (error); 910 } 911 912 /* 913 * The vn_io_fault() is a wrapper around vn_read() and vn_write() to 914 * prevent the following deadlock: 915 * 916 * Assume that the thread A reads from the vnode vp1 into userspace 917 * buffer buf1 backed by the pages of vnode vp2. If a page in buf1 is 918 * currently not resident, then system ends up with the call chain 919 * vn_read() -> VOP_READ(vp1) -> uiomove() -> [Page Fault] -> 920 * vm_fault(buf1) -> vnode_pager_getpages(vp2) -> VOP_GETPAGES(vp2) 921 * which establishes lock order vp1->vn_lock, then vp2->vn_lock. 922 * If, at the same time, thread B reads from vnode vp2 into buffer buf2 923 * backed by the pages of vnode vp1, and some page in buf2 is not 924 * resident, we get a reversed order vp2->vn_lock, then vp1->vn_lock. 925 * 926 * To prevent the lock order reversal and deadlock, vn_io_fault() does 927 * not allow page faults to happen during VOP_READ() or VOP_WRITE(). 928 * Instead, it first tries to do the whole range i/o with pagefaults 929 * disabled. If all pages in the i/o buffer are resident and mapped, 930 * VOP will succeed (ignoring the genuine filesystem errors). 931 * Otherwise, we get back EFAULT, and vn_io_fault() falls back to do 932 * i/o in chunks, with all pages in the chunk prefaulted and held 933 * using vm_fault_quick_hold_pages(). 934 * 935 * Filesystems using this deadlock avoidance scheme should use the 936 * array of the held pages from uio, saved in the curthread->td_ma, 937 * instead of doing uiomove(). A helper function 938 * vn_io_fault_uiomove() converts uiomove request into 939 * uiomove_fromphys() over td_ma array. 940 * 941 * Since vnode locks do not cover the whole i/o anymore, rangelocks 942 * make the current i/o request atomic with respect to other i/os and 943 * truncations. 944 */ 945 946 /* 947 * Decode vn_io_fault_args and perform the corresponding i/o. 948 */ 949 static int 950 vn_io_fault_doio(struct vn_io_fault_args *args, struct uio *uio, 951 struct thread *td) 952 { 953 int error, save; 954 955 error = 0; 956 save = vm_fault_disable_pagefaults(); 957 switch (args->kind) { 958 case VN_IO_FAULT_FOP: 959 error = (args->args.fop_args.doio)(args->args.fop_args.fp, 960 uio, args->cred, args->flags, td); 961 break; 962 case VN_IO_FAULT_VOP: 963 if (uio->uio_rw == UIO_READ) { 964 error = VOP_READ(args->args.vop_args.vp, uio, 965 args->flags, args->cred); 966 } else if (uio->uio_rw == UIO_WRITE) { 967 error = VOP_WRITE(args->args.vop_args.vp, uio, 968 args->flags, args->cred); 969 } 970 break; 971 default: 972 panic("vn_io_fault_doio: unknown kind of io %d %d", 973 args->kind, uio->uio_rw); 974 } 975 vm_fault_enable_pagefaults(save); 976 return (error); 977 } 978 979 static int 980 vn_io_fault_touch(char *base, const struct uio *uio) 981 { 982 int r; 983 984 r = fubyte(base); 985 if (r == -1 || (uio->uio_rw == UIO_READ && subyte(base, r) == -1)) 986 return (EFAULT); 987 return (0); 988 } 989 990 static int 991 vn_io_fault_prefault_user(const struct uio *uio) 992 { 993 char *base; 994 const struct iovec *iov; 995 size_t len; 996 ssize_t resid; 997 int error, i; 998 999 KASSERT(uio->uio_segflg == UIO_USERSPACE, 1000 ("vn_io_fault_prefault userspace")); 1001 1002 error = i = 0; 1003 iov = uio->uio_iov; 1004 resid = uio->uio_resid; 1005 base = iov->iov_base; 1006 len = iov->iov_len; 1007 while (resid > 0) { 1008 error = vn_io_fault_touch(base, uio); 1009 if (error != 0) 1010 break; 1011 if (len < PAGE_SIZE) { 1012 if (len != 0) { 1013 error = vn_io_fault_touch(base + len - 1, uio); 1014 if (error != 0) 1015 break; 1016 resid -= len; 1017 } 1018 if (++i >= uio->uio_iovcnt) 1019 break; 1020 iov = uio->uio_iov + i; 1021 base = iov->iov_base; 1022 len = iov->iov_len; 1023 } else { 1024 len -= PAGE_SIZE; 1025 base += PAGE_SIZE; 1026 resid -= PAGE_SIZE; 1027 } 1028 } 1029 return (error); 1030 } 1031 1032 /* 1033 * Common code for vn_io_fault(), agnostic to the kind of i/o request. 1034 * Uses vn_io_fault_doio() to make the call to an actual i/o function. 1035 * Used from vn_rdwr() and vn_io_fault(), which encode the i/o request 1036 * into args and call vn_io_fault1() to handle faults during the user 1037 * mode buffer accesses. 1038 */ 1039 static int 1040 vn_io_fault1(struct vnode *vp, struct uio *uio, struct vn_io_fault_args *args, 1041 struct thread *td) 1042 { 1043 vm_page_t ma[io_hold_cnt + 2]; 1044 struct uio *uio_clone, short_uio; 1045 struct iovec short_iovec[1]; 1046 vm_page_t *prev_td_ma; 1047 vm_prot_t prot; 1048 vm_offset_t addr, end; 1049 size_t len, resid; 1050 ssize_t adv; 1051 int error, cnt, saveheld, prev_td_ma_cnt; 1052 1053 if (vn_io_fault_prefault) { 1054 error = vn_io_fault_prefault_user(uio); 1055 if (error != 0) 1056 return (error); /* Or ignore ? */ 1057 } 1058 1059 prot = uio->uio_rw == UIO_READ ? VM_PROT_WRITE : VM_PROT_READ; 1060 1061 /* 1062 * The UFS follows IO_UNIT directive and replays back both 1063 * uio_offset and uio_resid if an error is encountered during the 1064 * operation. But, since the iovec may be already advanced, 1065 * uio is still in an inconsistent state. 1066 * 1067 * Cache a copy of the original uio, which is advanced to the redo 1068 * point using UIO_NOCOPY below. 1069 */ 1070 uio_clone = cloneuio(uio); 1071 resid = uio->uio_resid; 1072 1073 short_uio.uio_segflg = UIO_USERSPACE; 1074 short_uio.uio_rw = uio->uio_rw; 1075 short_uio.uio_td = uio->uio_td; 1076 1077 error = vn_io_fault_doio(args, uio, td); 1078 if (error != EFAULT) 1079 goto out; 1080 1081 atomic_add_long(&vn_io_faults_cnt, 1); 1082 uio_clone->uio_segflg = UIO_NOCOPY; 1083 uiomove(NULL, resid - uio->uio_resid, uio_clone); 1084 uio_clone->uio_segflg = uio->uio_segflg; 1085 1086 saveheld = curthread_pflags_set(TDP_UIOHELD); 1087 prev_td_ma = td->td_ma; 1088 prev_td_ma_cnt = td->td_ma_cnt; 1089 1090 while (uio_clone->uio_resid != 0) { 1091 len = uio_clone->uio_iov->iov_len; 1092 if (len == 0) { 1093 KASSERT(uio_clone->uio_iovcnt >= 1, 1094 ("iovcnt underflow")); 1095 uio_clone->uio_iov++; 1096 uio_clone->uio_iovcnt--; 1097 continue; 1098 } 1099 if (len > io_hold_cnt * PAGE_SIZE) 1100 len = io_hold_cnt * PAGE_SIZE; 1101 addr = (uintptr_t)uio_clone->uio_iov->iov_base; 1102 end = round_page(addr + len); 1103 if (end < addr) { 1104 error = EFAULT; 1105 break; 1106 } 1107 cnt = atop(end - trunc_page(addr)); 1108 /* 1109 * A perfectly misaligned address and length could cause 1110 * both the start and the end of the chunk to use partial 1111 * page. +2 accounts for such a situation. 1112 */ 1113 cnt = vm_fault_quick_hold_pages(&td->td_proc->p_vmspace->vm_map, 1114 addr, len, prot, ma, io_hold_cnt + 2); 1115 if (cnt == -1) { 1116 error = EFAULT; 1117 break; 1118 } 1119 short_uio.uio_iov = &short_iovec[0]; 1120 short_iovec[0].iov_base = (void *)addr; 1121 short_uio.uio_iovcnt = 1; 1122 short_uio.uio_resid = short_iovec[0].iov_len = len; 1123 short_uio.uio_offset = uio_clone->uio_offset; 1124 td->td_ma = ma; 1125 td->td_ma_cnt = cnt; 1126 1127 error = vn_io_fault_doio(args, &short_uio, td); 1128 vm_page_unhold_pages(ma, cnt); 1129 adv = len - short_uio.uio_resid; 1130 1131 uio_clone->uio_iov->iov_base = 1132 (char *)uio_clone->uio_iov->iov_base + adv; 1133 uio_clone->uio_iov->iov_len -= adv; 1134 uio_clone->uio_resid -= adv; 1135 uio_clone->uio_offset += adv; 1136 1137 uio->uio_resid -= adv; 1138 uio->uio_offset += adv; 1139 1140 if (error != 0 || adv == 0) 1141 break; 1142 } 1143 td->td_ma = prev_td_ma; 1144 td->td_ma_cnt = prev_td_ma_cnt; 1145 curthread_pflags_restore(saveheld); 1146 out: 1147 free(uio_clone, M_IOV); 1148 return (error); 1149 } 1150 1151 static int 1152 vn_io_fault(struct file *fp, struct uio *uio, struct ucred *active_cred, 1153 int flags, struct thread *td) 1154 { 1155 fo_rdwr_t *doio; 1156 struct vnode *vp; 1157 void *rl_cookie; 1158 struct vn_io_fault_args args; 1159 int error; 1160 1161 doio = uio->uio_rw == UIO_READ ? vn_read : vn_write; 1162 vp = fp->f_vnode; 1163 foffset_lock_uio(fp, uio, flags); 1164 if (do_vn_io_fault(vp, uio)) { 1165 args.kind = VN_IO_FAULT_FOP; 1166 args.args.fop_args.fp = fp; 1167 args.args.fop_args.doio = doio; 1168 args.cred = active_cred; 1169 args.flags = flags | FOF_OFFSET; 1170 if (uio->uio_rw == UIO_READ) { 1171 rl_cookie = vn_rangelock_rlock(vp, uio->uio_offset, 1172 uio->uio_offset + uio->uio_resid); 1173 } else if ((fp->f_flag & O_APPEND) != 0 || 1174 (flags & FOF_OFFSET) == 0) { 1175 /* For appenders, punt and lock the whole range. */ 1176 rl_cookie = vn_rangelock_wlock(vp, 0, OFF_MAX); 1177 } else { 1178 rl_cookie = vn_rangelock_wlock(vp, uio->uio_offset, 1179 uio->uio_offset + uio->uio_resid); 1180 } 1181 error = vn_io_fault1(vp, uio, &args, td); 1182 vn_rangelock_unlock(vp, rl_cookie); 1183 } else { 1184 error = doio(fp, uio, active_cred, flags | FOF_OFFSET, td); 1185 } 1186 foffset_unlock_uio(fp, uio, flags); 1187 return (error); 1188 } 1189 1190 /* 1191 * Helper function to perform the requested uiomove operation using 1192 * the held pages for io->uio_iov[0].iov_base buffer instead of 1193 * copyin/copyout. Access to the pages with uiomove_fromphys() 1194 * instead of iov_base prevents page faults that could occur due to 1195 * pmap_collect() invalidating the mapping created by 1196 * vm_fault_quick_hold_pages(), or pageout daemon, page laundry or 1197 * object cleanup revoking the write access from page mappings. 1198 * 1199 * Filesystems specified MNTK_NO_IOPF shall use vn_io_fault_uiomove() 1200 * instead of plain uiomove(). 1201 */ 1202 int 1203 vn_io_fault_uiomove(char *data, int xfersize, struct uio *uio) 1204 { 1205 struct uio transp_uio; 1206 struct iovec transp_iov[1]; 1207 struct thread *td; 1208 size_t adv; 1209 int error, pgadv; 1210 1211 td = curthread; 1212 if ((td->td_pflags & TDP_UIOHELD) == 0 || 1213 uio->uio_segflg != UIO_USERSPACE) 1214 return (uiomove(data, xfersize, uio)); 1215 1216 KASSERT(uio->uio_iovcnt == 1, ("uio_iovcnt %d", uio->uio_iovcnt)); 1217 transp_iov[0].iov_base = data; 1218 transp_uio.uio_iov = &transp_iov[0]; 1219 transp_uio.uio_iovcnt = 1; 1220 if (xfersize > uio->uio_resid) 1221 xfersize = uio->uio_resid; 1222 transp_uio.uio_resid = transp_iov[0].iov_len = xfersize; 1223 transp_uio.uio_offset = 0; 1224 transp_uio.uio_segflg = UIO_SYSSPACE; 1225 /* 1226 * Since transp_iov points to data, and td_ma page array 1227 * corresponds to original uio->uio_iov, we need to invert the 1228 * direction of the i/o operation as passed to 1229 * uiomove_fromphys(). 1230 */ 1231 switch (uio->uio_rw) { 1232 case UIO_WRITE: 1233 transp_uio.uio_rw = UIO_READ; 1234 break; 1235 case UIO_READ: 1236 transp_uio.uio_rw = UIO_WRITE; 1237 break; 1238 } 1239 transp_uio.uio_td = uio->uio_td; 1240 error = uiomove_fromphys(td->td_ma, 1241 ((vm_offset_t)uio->uio_iov->iov_base) & PAGE_MASK, 1242 xfersize, &transp_uio); 1243 adv = xfersize - transp_uio.uio_resid; 1244 pgadv = 1245 (((vm_offset_t)uio->uio_iov->iov_base + adv) >> PAGE_SHIFT) - 1246 (((vm_offset_t)uio->uio_iov->iov_base) >> PAGE_SHIFT); 1247 td->td_ma += pgadv; 1248 KASSERT(td->td_ma_cnt >= pgadv, ("consumed pages %d %d", td->td_ma_cnt, 1249 pgadv)); 1250 td->td_ma_cnt -= pgadv; 1251 uio->uio_iov->iov_base = (char *)uio->uio_iov->iov_base + adv; 1252 uio->uio_iov->iov_len -= adv; 1253 uio->uio_resid -= adv; 1254 uio->uio_offset += adv; 1255 return (error); 1256 } 1257 1258 int 1259 vn_io_fault_pgmove(vm_page_t ma[], vm_offset_t offset, int xfersize, 1260 struct uio *uio) 1261 { 1262 struct thread *td; 1263 vm_offset_t iov_base; 1264 int cnt, pgadv; 1265 1266 td = curthread; 1267 if ((td->td_pflags & TDP_UIOHELD) == 0 || 1268 uio->uio_segflg != UIO_USERSPACE) 1269 return (uiomove_fromphys(ma, offset, xfersize, uio)); 1270 1271 KASSERT(uio->uio_iovcnt == 1, ("uio_iovcnt %d", uio->uio_iovcnt)); 1272 cnt = xfersize > uio->uio_resid ? uio->uio_resid : xfersize; 1273 iov_base = (vm_offset_t)uio->uio_iov->iov_base; 1274 switch (uio->uio_rw) { 1275 case UIO_WRITE: 1276 pmap_copy_pages(td->td_ma, iov_base & PAGE_MASK, ma, 1277 offset, cnt); 1278 break; 1279 case UIO_READ: 1280 pmap_copy_pages(ma, offset, td->td_ma, iov_base & PAGE_MASK, 1281 cnt); 1282 break; 1283 } 1284 pgadv = ((iov_base + cnt) >> PAGE_SHIFT) - (iov_base >> PAGE_SHIFT); 1285 td->td_ma += pgadv; 1286 KASSERT(td->td_ma_cnt >= pgadv, ("consumed pages %d %d", td->td_ma_cnt, 1287 pgadv)); 1288 td->td_ma_cnt -= pgadv; 1289 uio->uio_iov->iov_base = (char *)(iov_base + cnt); 1290 uio->uio_iov->iov_len -= cnt; 1291 uio->uio_resid -= cnt; 1292 uio->uio_offset += cnt; 1293 return (0); 1294 } 1295 1296 /* 1297 * File table truncate routine. 1298 */ 1299 static int 1300 vn_truncate(struct file *fp, off_t length, struct ucred *active_cred, 1301 struct thread *td) 1302 { 1303 struct mount *mp; 1304 struct vnode *vp; 1305 void *rl_cookie; 1306 int error; 1307 1308 vp = fp->f_vnode; 1309 1310 /* 1311 * Lock the whole range for truncation. Otherwise split i/o 1312 * might happen partly before and partly after the truncation. 1313 */ 1314 rl_cookie = vn_rangelock_wlock(vp, 0, OFF_MAX); 1315 error = vn_start_write(vp, &mp, V_WAIT | PCATCH); 1316 if (error) 1317 goto out1; 1318 vn_lock(vp, LK_EXCLUSIVE | LK_RETRY); 1319 AUDIT_ARG_VNODE1(vp); 1320 if (vp->v_type == VDIR) { 1321 error = EISDIR; 1322 goto out; 1323 } 1324 #ifdef MAC 1325 error = mac_vnode_check_write(active_cred, fp->f_cred, vp); 1326 if (error) 1327 goto out; 1328 #endif 1329 error = vn_truncate_locked(vp, length, (fp->f_flag & O_FSYNC) != 0, 1330 fp->f_cred); 1331 out: 1332 VOP_UNLOCK(vp); 1333 vn_finished_write(mp); 1334 out1: 1335 vn_rangelock_unlock(vp, rl_cookie); 1336 return (error); 1337 } 1338 1339 /* 1340 * Truncate a file that is already locked. 1341 */ 1342 int 1343 vn_truncate_locked(struct vnode *vp, off_t length, bool sync, 1344 struct ucred *cred) 1345 { 1346 struct vattr vattr; 1347 int error; 1348 1349 error = VOP_ADD_WRITECOUNT(vp, 1); 1350 if (error == 0) { 1351 VATTR_NULL(&vattr); 1352 vattr.va_size = length; 1353 if (sync) 1354 vattr.va_vaflags |= VA_SYNC; 1355 error = VOP_SETATTR(vp, &vattr, cred); 1356 VOP_ADD_WRITECOUNT_CHECKED(vp, -1); 1357 } 1358 return (error); 1359 } 1360 1361 /* 1362 * File table vnode stat routine. 1363 */ 1364 static int 1365 vn_statfile(struct file *fp, struct stat *sb, struct ucred *active_cred, 1366 struct thread *td) 1367 { 1368 struct vnode *vp = fp->f_vnode; 1369 int error; 1370 1371 vn_lock(vp, LK_SHARED | LK_RETRY); 1372 error = vn_stat(vp, sb, active_cred, fp->f_cred, td); 1373 VOP_UNLOCK(vp); 1374 1375 return (error); 1376 } 1377 1378 /* 1379 * Stat a vnode; implementation for the stat syscall 1380 */ 1381 int 1382 vn_stat(struct vnode *vp, struct stat *sb, struct ucred *active_cred, 1383 struct ucred *file_cred, struct thread *td) 1384 { 1385 struct vattr vattr; 1386 struct vattr *vap; 1387 int error; 1388 u_short mode; 1389 1390 AUDIT_ARG_VNODE1(vp); 1391 #ifdef MAC 1392 error = mac_vnode_check_stat(active_cred, file_cred, vp); 1393 if (error) 1394 return (error); 1395 #endif 1396 1397 vap = &vattr; 1398 1399 /* 1400 * Initialize defaults for new and unusual fields, so that file 1401 * systems which don't support these fields don't need to know 1402 * about them. 1403 */ 1404 vap->va_birthtime.tv_sec = -1; 1405 vap->va_birthtime.tv_nsec = 0; 1406 vap->va_fsid = VNOVAL; 1407 vap->va_rdev = NODEV; 1408 1409 error = VOP_GETATTR(vp, vap, active_cred); 1410 if (error) 1411 return (error); 1412 1413 /* 1414 * Zero the spare stat fields 1415 */ 1416 bzero(sb, sizeof *sb); 1417 1418 /* 1419 * Copy from vattr table 1420 */ 1421 if (vap->va_fsid != VNOVAL) 1422 sb->st_dev = vap->va_fsid; 1423 else 1424 sb->st_dev = vp->v_mount->mnt_stat.f_fsid.val[0]; 1425 sb->st_ino = vap->va_fileid; 1426 mode = vap->va_mode; 1427 switch (vap->va_type) { 1428 case VREG: 1429 mode |= S_IFREG; 1430 break; 1431 case VDIR: 1432 mode |= S_IFDIR; 1433 break; 1434 case VBLK: 1435 mode |= S_IFBLK; 1436 break; 1437 case VCHR: 1438 mode |= S_IFCHR; 1439 break; 1440 case VLNK: 1441 mode |= S_IFLNK; 1442 break; 1443 case VSOCK: 1444 mode |= S_IFSOCK; 1445 break; 1446 case VFIFO: 1447 mode |= S_IFIFO; 1448 break; 1449 default: 1450 return (EBADF); 1451 } 1452 sb->st_mode = mode; 1453 sb->st_nlink = vap->va_nlink; 1454 sb->st_uid = vap->va_uid; 1455 sb->st_gid = vap->va_gid; 1456 sb->st_rdev = vap->va_rdev; 1457 if (vap->va_size > OFF_MAX) 1458 return (EOVERFLOW); 1459 sb->st_size = vap->va_size; 1460 sb->st_atim.tv_sec = vap->va_atime.tv_sec; 1461 sb->st_atim.tv_nsec = vap->va_atime.tv_nsec; 1462 sb->st_mtim.tv_sec = vap->va_mtime.tv_sec; 1463 sb->st_mtim.tv_nsec = vap->va_mtime.tv_nsec; 1464 sb->st_ctim.tv_sec = vap->va_ctime.tv_sec; 1465 sb->st_ctim.tv_nsec = vap->va_ctime.tv_nsec; 1466 sb->st_birthtim.tv_sec = vap->va_birthtime.tv_sec; 1467 sb->st_birthtim.tv_nsec = vap->va_birthtime.tv_nsec; 1468 1469 /* 1470 * According to www.opengroup.org, the meaning of st_blksize is 1471 * "a filesystem-specific preferred I/O block size for this 1472 * object. In some filesystem types, this may vary from file 1473 * to file" 1474 * Use miminum/default of PAGE_SIZE (e.g. for VCHR). 1475 */ 1476 1477 sb->st_blksize = max(PAGE_SIZE, vap->va_blocksize); 1478 1479 sb->st_flags = vap->va_flags; 1480 if (priv_check_cred_vfs_generation(td->td_ucred)) 1481 sb->st_gen = 0; 1482 else 1483 sb->st_gen = vap->va_gen; 1484 1485 sb->st_blocks = vap->va_bytes / S_BLKSIZE; 1486 return (0); 1487 } 1488 1489 /* 1490 * File table vnode ioctl routine. 1491 */ 1492 static int 1493 vn_ioctl(struct file *fp, u_long com, void *data, struct ucred *active_cred, 1494 struct thread *td) 1495 { 1496 struct vattr vattr; 1497 struct vnode *vp; 1498 struct fiobmap2_arg *bmarg; 1499 int error; 1500 1501 vp = fp->f_vnode; 1502 switch (vp->v_type) { 1503 case VDIR: 1504 case VREG: 1505 switch (com) { 1506 case FIONREAD: 1507 vn_lock(vp, LK_SHARED | LK_RETRY); 1508 error = VOP_GETATTR(vp, &vattr, active_cred); 1509 VOP_UNLOCK(vp); 1510 if (error == 0) 1511 *(int *)data = vattr.va_size - fp->f_offset; 1512 return (error); 1513 case FIOBMAP2: 1514 bmarg = (struct fiobmap2_arg *)data; 1515 vn_lock(vp, LK_SHARED | LK_RETRY); 1516 #ifdef MAC 1517 error = mac_vnode_check_read(active_cred, fp->f_cred, 1518 vp); 1519 if (error == 0) 1520 #endif 1521 error = VOP_BMAP(vp, bmarg->bn, NULL, 1522 &bmarg->bn, &bmarg->runp, &bmarg->runb); 1523 VOP_UNLOCK(vp); 1524 return (error); 1525 case FIONBIO: 1526 case FIOASYNC: 1527 return (0); 1528 default: 1529 return (VOP_IOCTL(vp, com, data, fp->f_flag, 1530 active_cred, td)); 1531 } 1532 break; 1533 case VCHR: 1534 return (VOP_IOCTL(vp, com, data, fp->f_flag, 1535 active_cred, td)); 1536 default: 1537 return (ENOTTY); 1538 } 1539 } 1540 1541 /* 1542 * File table vnode poll routine. 1543 */ 1544 static int 1545 vn_poll(struct file *fp, int events, struct ucred *active_cred, 1546 struct thread *td) 1547 { 1548 struct vnode *vp; 1549 int error; 1550 1551 vp = fp->f_vnode; 1552 #ifdef MAC 1553 vn_lock(vp, LK_EXCLUSIVE | LK_RETRY); 1554 AUDIT_ARG_VNODE1(vp); 1555 error = mac_vnode_check_poll(active_cred, fp->f_cred, vp); 1556 VOP_UNLOCK(vp); 1557 if (!error) 1558 #endif 1559 1560 error = VOP_POLL(vp, events, fp->f_cred, td); 1561 return (error); 1562 } 1563 1564 /* 1565 * Acquire the requested lock and then check for validity. LK_RETRY 1566 * permits vn_lock to return doomed vnodes. 1567 */ 1568 static int __noinline 1569 _vn_lock_fallback(struct vnode *vp, int flags, const char *file, int line, 1570 int error) 1571 { 1572 1573 KASSERT((flags & LK_RETRY) == 0 || error == 0, 1574 ("vn_lock: error %d incompatible with flags %#x", error, flags)); 1575 1576 if (error == 0) 1577 VNASSERT(VN_IS_DOOMED(vp), vp, ("vnode not doomed")); 1578 1579 if ((flags & LK_RETRY) == 0) { 1580 if (error == 0) { 1581 VOP_UNLOCK(vp); 1582 error = ENOENT; 1583 } 1584 return (error); 1585 } 1586 1587 /* 1588 * LK_RETRY case. 1589 * 1590 * Nothing to do if we got the lock. 1591 */ 1592 if (error == 0) 1593 return (0); 1594 1595 /* 1596 * Interlock was dropped by the call in _vn_lock. 1597 */ 1598 flags &= ~LK_INTERLOCK; 1599 do { 1600 error = VOP_LOCK1(vp, flags, file, line); 1601 } while (error != 0); 1602 return (0); 1603 } 1604 1605 int 1606 _vn_lock(struct vnode *vp, int flags, const char *file, int line) 1607 { 1608 int error; 1609 1610 VNASSERT((flags & LK_TYPE_MASK) != 0, vp, 1611 ("vn_lock: no locktype (%d passed)", flags)); 1612 VNPASS(vp->v_holdcnt > 0, vp); 1613 error = VOP_LOCK1(vp, flags, file, line); 1614 if (__predict_false(error != 0 || VN_IS_DOOMED(vp))) 1615 return (_vn_lock_fallback(vp, flags, file, line, error)); 1616 return (0); 1617 } 1618 1619 /* 1620 * File table vnode close routine. 1621 */ 1622 static int 1623 vn_closefile(struct file *fp, struct thread *td) 1624 { 1625 struct vnode *vp; 1626 struct flock lf; 1627 int error; 1628 bool ref; 1629 1630 vp = fp->f_vnode; 1631 fp->f_ops = &badfileops; 1632 ref= (fp->f_flag & FHASLOCK) != 0 && fp->f_type == DTYPE_VNODE; 1633 1634 error = vn_close1(vp, fp->f_flag, fp->f_cred, td, ref); 1635 1636 if (__predict_false(ref)) { 1637 lf.l_whence = SEEK_SET; 1638 lf.l_start = 0; 1639 lf.l_len = 0; 1640 lf.l_type = F_UNLCK; 1641 (void) VOP_ADVLOCK(vp, fp, F_UNLCK, &lf, F_FLOCK); 1642 vrele(vp); 1643 } 1644 return (error); 1645 } 1646 1647 /* 1648 * Preparing to start a filesystem write operation. If the operation is 1649 * permitted, then we bump the count of operations in progress and 1650 * proceed. If a suspend request is in progress, we wait until the 1651 * suspension is over, and then proceed. 1652 */ 1653 static int 1654 vn_start_write_refed(struct mount *mp, int flags, bool mplocked) 1655 { 1656 int error, mflags; 1657 1658 if (__predict_true(!mplocked) && (flags & V_XSLEEP) == 0 && 1659 vfs_op_thread_enter(mp)) { 1660 MPASS((mp->mnt_kern_flag & MNTK_SUSPEND) == 0); 1661 vfs_mp_count_add_pcpu(mp, writeopcount, 1); 1662 vfs_op_thread_exit(mp); 1663 return (0); 1664 } 1665 1666 if (mplocked) 1667 mtx_assert(MNT_MTX(mp), MA_OWNED); 1668 else 1669 MNT_ILOCK(mp); 1670 1671 error = 0; 1672 1673 /* 1674 * Check on status of suspension. 1675 */ 1676 if ((curthread->td_pflags & TDP_IGNSUSP) == 0 || 1677 mp->mnt_susp_owner != curthread) { 1678 mflags = ((mp->mnt_vfc->vfc_flags & VFCF_SBDRY) != 0 ? 1679 (flags & PCATCH) : 0) | (PUSER - 1); 1680 while ((mp->mnt_kern_flag & MNTK_SUSPEND) != 0) { 1681 if (flags & V_NOWAIT) { 1682 error = EWOULDBLOCK; 1683 goto unlock; 1684 } 1685 error = msleep(&mp->mnt_flag, MNT_MTX(mp), mflags, 1686 "suspfs", 0); 1687 if (error) 1688 goto unlock; 1689 } 1690 } 1691 if (flags & V_XSLEEP) 1692 goto unlock; 1693 mp->mnt_writeopcount++; 1694 unlock: 1695 if (error != 0 || (flags & V_XSLEEP) != 0) 1696 MNT_REL(mp); 1697 MNT_IUNLOCK(mp); 1698 return (error); 1699 } 1700 1701 int 1702 vn_start_write(struct vnode *vp, struct mount **mpp, int flags) 1703 { 1704 struct mount *mp; 1705 int error; 1706 1707 KASSERT((flags & V_MNTREF) == 0 || (*mpp != NULL && vp == NULL), 1708 ("V_MNTREF requires mp")); 1709 1710 error = 0; 1711 /* 1712 * If a vnode is provided, get and return the mount point that 1713 * to which it will write. 1714 */ 1715 if (vp != NULL) { 1716 if ((error = VOP_GETWRITEMOUNT(vp, mpp)) != 0) { 1717 *mpp = NULL; 1718 if (error != EOPNOTSUPP) 1719 return (error); 1720 return (0); 1721 } 1722 } 1723 if ((mp = *mpp) == NULL) 1724 return (0); 1725 1726 /* 1727 * VOP_GETWRITEMOUNT() returns with the mp refcount held through 1728 * a vfs_ref(). 1729 * As long as a vnode is not provided we need to acquire a 1730 * refcount for the provided mountpoint too, in order to 1731 * emulate a vfs_ref(). 1732 */ 1733 if (vp == NULL && (flags & V_MNTREF) == 0) 1734 vfs_ref(mp); 1735 1736 return (vn_start_write_refed(mp, flags, false)); 1737 } 1738 1739 /* 1740 * Secondary suspension. Used by operations such as vop_inactive 1741 * routines that are needed by the higher level functions. These 1742 * are allowed to proceed until all the higher level functions have 1743 * completed (indicated by mnt_writeopcount dropping to zero). At that 1744 * time, these operations are halted until the suspension is over. 1745 */ 1746 int 1747 vn_start_secondary_write(struct vnode *vp, struct mount **mpp, int flags) 1748 { 1749 struct mount *mp; 1750 int error; 1751 1752 KASSERT((flags & V_MNTREF) == 0 || (*mpp != NULL && vp == NULL), 1753 ("V_MNTREF requires mp")); 1754 1755 retry: 1756 if (vp != NULL) { 1757 if ((error = VOP_GETWRITEMOUNT(vp, mpp)) != 0) { 1758 *mpp = NULL; 1759 if (error != EOPNOTSUPP) 1760 return (error); 1761 return (0); 1762 } 1763 } 1764 /* 1765 * If we are not suspended or have not yet reached suspended 1766 * mode, then let the operation proceed. 1767 */ 1768 if ((mp = *mpp) == NULL) 1769 return (0); 1770 1771 /* 1772 * VOP_GETWRITEMOUNT() returns with the mp refcount held through 1773 * a vfs_ref(). 1774 * As long as a vnode is not provided we need to acquire a 1775 * refcount for the provided mountpoint too, in order to 1776 * emulate a vfs_ref(). 1777 */ 1778 MNT_ILOCK(mp); 1779 if (vp == NULL && (flags & V_MNTREF) == 0) 1780 MNT_REF(mp); 1781 if ((mp->mnt_kern_flag & (MNTK_SUSPENDED | MNTK_SUSPEND2)) == 0) { 1782 mp->mnt_secondary_writes++; 1783 mp->mnt_secondary_accwrites++; 1784 MNT_IUNLOCK(mp); 1785 return (0); 1786 } 1787 if (flags & V_NOWAIT) { 1788 MNT_REL(mp); 1789 MNT_IUNLOCK(mp); 1790 return (EWOULDBLOCK); 1791 } 1792 /* 1793 * Wait for the suspension to finish. 1794 */ 1795 error = msleep(&mp->mnt_flag, MNT_MTX(mp), (PUSER - 1) | PDROP | 1796 ((mp->mnt_vfc->vfc_flags & VFCF_SBDRY) != 0 ? (flags & PCATCH) : 0), 1797 "suspfs", 0); 1798 vfs_rel(mp); 1799 if (error == 0) 1800 goto retry; 1801 return (error); 1802 } 1803 1804 /* 1805 * Filesystem write operation has completed. If we are suspending and this 1806 * operation is the last one, notify the suspender that the suspension is 1807 * now in effect. 1808 */ 1809 void 1810 vn_finished_write(struct mount *mp) 1811 { 1812 int c; 1813 1814 if (mp == NULL) 1815 return; 1816 1817 if (vfs_op_thread_enter(mp)) { 1818 vfs_mp_count_sub_pcpu(mp, writeopcount, 1); 1819 vfs_mp_count_sub_pcpu(mp, ref, 1); 1820 vfs_op_thread_exit(mp); 1821 return; 1822 } 1823 1824 MNT_ILOCK(mp); 1825 vfs_assert_mount_counters(mp); 1826 MNT_REL(mp); 1827 c = --mp->mnt_writeopcount; 1828 if (mp->mnt_vfs_ops == 0) { 1829 MPASS((mp->mnt_kern_flag & MNTK_SUSPEND) == 0); 1830 MNT_IUNLOCK(mp); 1831 return; 1832 } 1833 if (c < 0) 1834 vfs_dump_mount_counters(mp); 1835 if ((mp->mnt_kern_flag & MNTK_SUSPEND) != 0 && c == 0) 1836 wakeup(&mp->mnt_writeopcount); 1837 MNT_IUNLOCK(mp); 1838 } 1839 1840 /* 1841 * Filesystem secondary write operation has completed. If we are 1842 * suspending and this operation is the last one, notify the suspender 1843 * that the suspension is now in effect. 1844 */ 1845 void 1846 vn_finished_secondary_write(struct mount *mp) 1847 { 1848 if (mp == NULL) 1849 return; 1850 MNT_ILOCK(mp); 1851 MNT_REL(mp); 1852 mp->mnt_secondary_writes--; 1853 if (mp->mnt_secondary_writes < 0) 1854 panic("vn_finished_secondary_write: neg cnt"); 1855 if ((mp->mnt_kern_flag & MNTK_SUSPEND) != 0 && 1856 mp->mnt_secondary_writes <= 0) 1857 wakeup(&mp->mnt_secondary_writes); 1858 MNT_IUNLOCK(mp); 1859 } 1860 1861 /* 1862 * Request a filesystem to suspend write operations. 1863 */ 1864 int 1865 vfs_write_suspend(struct mount *mp, int flags) 1866 { 1867 int error; 1868 1869 vfs_op_enter(mp); 1870 1871 MNT_ILOCK(mp); 1872 vfs_assert_mount_counters(mp); 1873 if (mp->mnt_susp_owner == curthread) { 1874 vfs_op_exit_locked(mp); 1875 MNT_IUNLOCK(mp); 1876 return (EALREADY); 1877 } 1878 while (mp->mnt_kern_flag & MNTK_SUSPEND) 1879 msleep(&mp->mnt_flag, MNT_MTX(mp), PUSER - 1, "wsuspfs", 0); 1880 1881 /* 1882 * Unmount holds a write reference on the mount point. If we 1883 * own busy reference and drain for writers, we deadlock with 1884 * the reference draining in the unmount path. Callers of 1885 * vfs_write_suspend() must specify VS_SKIP_UNMOUNT if 1886 * vfs_busy() reference is owned and caller is not in the 1887 * unmount context. 1888 */ 1889 if ((flags & VS_SKIP_UNMOUNT) != 0 && 1890 (mp->mnt_kern_flag & MNTK_UNMOUNT) != 0) { 1891 vfs_op_exit_locked(mp); 1892 MNT_IUNLOCK(mp); 1893 return (EBUSY); 1894 } 1895 1896 mp->mnt_kern_flag |= MNTK_SUSPEND; 1897 mp->mnt_susp_owner = curthread; 1898 if (mp->mnt_writeopcount > 0) 1899 (void) msleep(&mp->mnt_writeopcount, 1900 MNT_MTX(mp), (PUSER - 1)|PDROP, "suspwt", 0); 1901 else 1902 MNT_IUNLOCK(mp); 1903 if ((error = VFS_SYNC(mp, MNT_SUSPEND)) != 0) { 1904 vfs_write_resume(mp, 0); 1905 vfs_op_exit(mp); 1906 } 1907 return (error); 1908 } 1909 1910 /* 1911 * Request a filesystem to resume write operations. 1912 */ 1913 void 1914 vfs_write_resume(struct mount *mp, int flags) 1915 { 1916 1917 MNT_ILOCK(mp); 1918 if ((mp->mnt_kern_flag & MNTK_SUSPEND) != 0) { 1919 KASSERT(mp->mnt_susp_owner == curthread, ("mnt_susp_owner")); 1920 mp->mnt_kern_flag &= ~(MNTK_SUSPEND | MNTK_SUSPEND2 | 1921 MNTK_SUSPENDED); 1922 mp->mnt_susp_owner = NULL; 1923 wakeup(&mp->mnt_writeopcount); 1924 wakeup(&mp->mnt_flag); 1925 curthread->td_pflags &= ~TDP_IGNSUSP; 1926 if ((flags & VR_START_WRITE) != 0) { 1927 MNT_REF(mp); 1928 mp->mnt_writeopcount++; 1929 } 1930 MNT_IUNLOCK(mp); 1931 if ((flags & VR_NO_SUSPCLR) == 0) 1932 VFS_SUSP_CLEAN(mp); 1933 vfs_op_exit(mp); 1934 } else if ((flags & VR_START_WRITE) != 0) { 1935 MNT_REF(mp); 1936 vn_start_write_refed(mp, 0, true); 1937 } else { 1938 MNT_IUNLOCK(mp); 1939 } 1940 } 1941 1942 /* 1943 * Helper loop around vfs_write_suspend() for filesystem unmount VFS 1944 * methods. 1945 */ 1946 int 1947 vfs_write_suspend_umnt(struct mount *mp) 1948 { 1949 int error; 1950 1951 KASSERT((curthread->td_pflags & TDP_IGNSUSP) == 0, 1952 ("vfs_write_suspend_umnt: recursed")); 1953 1954 /* dounmount() already called vn_start_write(). */ 1955 for (;;) { 1956 vn_finished_write(mp); 1957 error = vfs_write_suspend(mp, 0); 1958 if (error != 0) { 1959 vn_start_write(NULL, &mp, V_WAIT); 1960 return (error); 1961 } 1962 MNT_ILOCK(mp); 1963 if ((mp->mnt_kern_flag & MNTK_SUSPENDED) != 0) 1964 break; 1965 MNT_IUNLOCK(mp); 1966 vn_start_write(NULL, &mp, V_WAIT); 1967 } 1968 mp->mnt_kern_flag &= ~(MNTK_SUSPENDED | MNTK_SUSPEND2); 1969 wakeup(&mp->mnt_flag); 1970 MNT_IUNLOCK(mp); 1971 curthread->td_pflags |= TDP_IGNSUSP; 1972 return (0); 1973 } 1974 1975 /* 1976 * Implement kqueues for files by translating it to vnode operation. 1977 */ 1978 static int 1979 vn_kqfilter(struct file *fp, struct knote *kn) 1980 { 1981 1982 return (VOP_KQFILTER(fp->f_vnode, kn)); 1983 } 1984 1985 /* 1986 * Simplified in-kernel wrapper calls for extended attribute access. 1987 * Both calls pass in a NULL credential, authorizing as "kernel" access. 1988 * Set IO_NODELOCKED in ioflg if the vnode is already locked. 1989 */ 1990 int 1991 vn_extattr_get(struct vnode *vp, int ioflg, int attrnamespace, 1992 const char *attrname, int *buflen, char *buf, struct thread *td) 1993 { 1994 struct uio auio; 1995 struct iovec iov; 1996 int error; 1997 1998 iov.iov_len = *buflen; 1999 iov.iov_base = buf; 2000 2001 auio.uio_iov = &iov; 2002 auio.uio_iovcnt = 1; 2003 auio.uio_rw = UIO_READ; 2004 auio.uio_segflg = UIO_SYSSPACE; 2005 auio.uio_td = td; 2006 auio.uio_offset = 0; 2007 auio.uio_resid = *buflen; 2008 2009 if ((ioflg & IO_NODELOCKED) == 0) 2010 vn_lock(vp, LK_SHARED | LK_RETRY); 2011 2012 ASSERT_VOP_LOCKED(vp, "IO_NODELOCKED with no vp lock held"); 2013 2014 /* authorize attribute retrieval as kernel */ 2015 error = VOP_GETEXTATTR(vp, attrnamespace, attrname, &auio, NULL, NULL, 2016 td); 2017 2018 if ((ioflg & IO_NODELOCKED) == 0) 2019 VOP_UNLOCK(vp); 2020 2021 if (error == 0) { 2022 *buflen = *buflen - auio.uio_resid; 2023 } 2024 2025 return (error); 2026 } 2027 2028 /* 2029 * XXX failure mode if partially written? 2030 */ 2031 int 2032 vn_extattr_set(struct vnode *vp, int ioflg, int attrnamespace, 2033 const char *attrname, int buflen, char *buf, struct thread *td) 2034 { 2035 struct uio auio; 2036 struct iovec iov; 2037 struct mount *mp; 2038 int error; 2039 2040 iov.iov_len = buflen; 2041 iov.iov_base = buf; 2042 2043 auio.uio_iov = &iov; 2044 auio.uio_iovcnt = 1; 2045 auio.uio_rw = UIO_WRITE; 2046 auio.uio_segflg = UIO_SYSSPACE; 2047 auio.uio_td = td; 2048 auio.uio_offset = 0; 2049 auio.uio_resid = buflen; 2050 2051 if ((ioflg & IO_NODELOCKED) == 0) { 2052 if ((error = vn_start_write(vp, &mp, V_WAIT)) != 0) 2053 return (error); 2054 vn_lock(vp, LK_EXCLUSIVE | LK_RETRY); 2055 } 2056 2057 ASSERT_VOP_LOCKED(vp, "IO_NODELOCKED with no vp lock held"); 2058 2059 /* authorize attribute setting as kernel */ 2060 error = VOP_SETEXTATTR(vp, attrnamespace, attrname, &auio, NULL, td); 2061 2062 if ((ioflg & IO_NODELOCKED) == 0) { 2063 vn_finished_write(mp); 2064 VOP_UNLOCK(vp); 2065 } 2066 2067 return (error); 2068 } 2069 2070 int 2071 vn_extattr_rm(struct vnode *vp, int ioflg, int attrnamespace, 2072 const char *attrname, struct thread *td) 2073 { 2074 struct mount *mp; 2075 int error; 2076 2077 if ((ioflg & IO_NODELOCKED) == 0) { 2078 if ((error = vn_start_write(vp, &mp, V_WAIT)) != 0) 2079 return (error); 2080 vn_lock(vp, LK_EXCLUSIVE | LK_RETRY); 2081 } 2082 2083 ASSERT_VOP_LOCKED(vp, "IO_NODELOCKED with no vp lock held"); 2084 2085 /* authorize attribute removal as kernel */ 2086 error = VOP_DELETEEXTATTR(vp, attrnamespace, attrname, NULL, td); 2087 if (error == EOPNOTSUPP) 2088 error = VOP_SETEXTATTR(vp, attrnamespace, attrname, NULL, 2089 NULL, td); 2090 2091 if ((ioflg & IO_NODELOCKED) == 0) { 2092 vn_finished_write(mp); 2093 VOP_UNLOCK(vp); 2094 } 2095 2096 return (error); 2097 } 2098 2099 static int 2100 vn_get_ino_alloc_vget(struct mount *mp, void *arg, int lkflags, 2101 struct vnode **rvp) 2102 { 2103 2104 return (VFS_VGET(mp, *(ino_t *)arg, lkflags, rvp)); 2105 } 2106 2107 int 2108 vn_vget_ino(struct vnode *vp, ino_t ino, int lkflags, struct vnode **rvp) 2109 { 2110 2111 return (vn_vget_ino_gen(vp, vn_get_ino_alloc_vget, &ino, 2112 lkflags, rvp)); 2113 } 2114 2115 int 2116 vn_vget_ino_gen(struct vnode *vp, vn_get_ino_t alloc, void *alloc_arg, 2117 int lkflags, struct vnode **rvp) 2118 { 2119 struct mount *mp; 2120 int ltype, error; 2121 2122 ASSERT_VOP_LOCKED(vp, "vn_vget_ino_get"); 2123 mp = vp->v_mount; 2124 ltype = VOP_ISLOCKED(vp); 2125 KASSERT(ltype == LK_EXCLUSIVE || ltype == LK_SHARED, 2126 ("vn_vget_ino: vp not locked")); 2127 error = vfs_busy(mp, MBF_NOWAIT); 2128 if (error != 0) { 2129 vfs_ref(mp); 2130 VOP_UNLOCK(vp); 2131 error = vfs_busy(mp, 0); 2132 vn_lock(vp, ltype | LK_RETRY); 2133 vfs_rel(mp); 2134 if (error != 0) 2135 return (ENOENT); 2136 if (VN_IS_DOOMED(vp)) { 2137 vfs_unbusy(mp); 2138 return (ENOENT); 2139 } 2140 } 2141 VOP_UNLOCK(vp); 2142 error = alloc(mp, alloc_arg, lkflags, rvp); 2143 vfs_unbusy(mp); 2144 if (error != 0 || *rvp != vp) 2145 vn_lock(vp, ltype | LK_RETRY); 2146 if (VN_IS_DOOMED(vp)) { 2147 if (error == 0) { 2148 if (*rvp == vp) 2149 vunref(vp); 2150 else 2151 vput(*rvp); 2152 } 2153 error = ENOENT; 2154 } 2155 return (error); 2156 } 2157 2158 int 2159 vn_rlimit_fsize(const struct vnode *vp, const struct uio *uio, 2160 struct thread *td) 2161 { 2162 2163 if (vp->v_type != VREG || td == NULL) 2164 return (0); 2165 if ((uoff_t)uio->uio_offset + uio->uio_resid > 2166 lim_cur(td, RLIMIT_FSIZE)) { 2167 PROC_LOCK(td->td_proc); 2168 kern_psignal(td->td_proc, SIGXFSZ); 2169 PROC_UNLOCK(td->td_proc); 2170 return (EFBIG); 2171 } 2172 return (0); 2173 } 2174 2175 int 2176 vn_chmod(struct file *fp, mode_t mode, struct ucred *active_cred, 2177 struct thread *td) 2178 { 2179 struct vnode *vp; 2180 2181 vp = fp->f_vnode; 2182 #ifdef AUDIT 2183 vn_lock(vp, LK_SHARED | LK_RETRY); 2184 AUDIT_ARG_VNODE1(vp); 2185 VOP_UNLOCK(vp); 2186 #endif 2187 return (setfmode(td, active_cred, vp, mode)); 2188 } 2189 2190 int 2191 vn_chown(struct file *fp, uid_t uid, gid_t gid, struct ucred *active_cred, 2192 struct thread *td) 2193 { 2194 struct vnode *vp; 2195 2196 vp = fp->f_vnode; 2197 #ifdef AUDIT 2198 vn_lock(vp, LK_SHARED | LK_RETRY); 2199 AUDIT_ARG_VNODE1(vp); 2200 VOP_UNLOCK(vp); 2201 #endif 2202 return (setfown(td, active_cred, vp, uid, gid)); 2203 } 2204 2205 void 2206 vn_pages_remove(struct vnode *vp, vm_pindex_t start, vm_pindex_t end) 2207 { 2208 vm_object_t object; 2209 2210 if ((object = vp->v_object) == NULL) 2211 return; 2212 VM_OBJECT_WLOCK(object); 2213 vm_object_page_remove(object, start, end, 0); 2214 VM_OBJECT_WUNLOCK(object); 2215 } 2216 2217 int 2218 vn_bmap_seekhole(struct vnode *vp, u_long cmd, off_t *off, struct ucred *cred) 2219 { 2220 struct vattr va; 2221 daddr_t bn, bnp; 2222 uint64_t bsize; 2223 off_t noff; 2224 int error; 2225 2226 KASSERT(cmd == FIOSEEKHOLE || cmd == FIOSEEKDATA, 2227 ("Wrong command %lu", cmd)); 2228 2229 if (vn_lock(vp, LK_SHARED) != 0) 2230 return (EBADF); 2231 if (vp->v_type != VREG) { 2232 error = ENOTTY; 2233 goto unlock; 2234 } 2235 error = VOP_GETATTR(vp, &va, cred); 2236 if (error != 0) 2237 goto unlock; 2238 noff = *off; 2239 if (noff >= va.va_size) { 2240 error = ENXIO; 2241 goto unlock; 2242 } 2243 bsize = vp->v_mount->mnt_stat.f_iosize; 2244 for (bn = noff / bsize; noff < va.va_size; bn++, noff += bsize - 2245 noff % bsize) { 2246 error = VOP_BMAP(vp, bn, NULL, &bnp, NULL, NULL); 2247 if (error == EOPNOTSUPP) { 2248 error = ENOTTY; 2249 goto unlock; 2250 } 2251 if ((bnp == -1 && cmd == FIOSEEKHOLE) || 2252 (bnp != -1 && cmd == FIOSEEKDATA)) { 2253 noff = bn * bsize; 2254 if (noff < *off) 2255 noff = *off; 2256 goto unlock; 2257 } 2258 } 2259 if (noff > va.va_size) 2260 noff = va.va_size; 2261 /* noff == va.va_size. There is an implicit hole at the end of file. */ 2262 if (cmd == FIOSEEKDATA) 2263 error = ENXIO; 2264 unlock: 2265 VOP_UNLOCK(vp); 2266 if (error == 0) 2267 *off = noff; 2268 return (error); 2269 } 2270 2271 int 2272 vn_seek(struct file *fp, off_t offset, int whence, struct thread *td) 2273 { 2274 struct ucred *cred; 2275 struct vnode *vp; 2276 struct vattr vattr; 2277 off_t foffset, size; 2278 int error, noneg; 2279 2280 cred = td->td_ucred; 2281 vp = fp->f_vnode; 2282 foffset = foffset_lock(fp, 0); 2283 noneg = (vp->v_type != VCHR); 2284 error = 0; 2285 switch (whence) { 2286 case L_INCR: 2287 if (noneg && 2288 (foffset < 0 || 2289 (offset > 0 && foffset > OFF_MAX - offset))) { 2290 error = EOVERFLOW; 2291 break; 2292 } 2293 offset += foffset; 2294 break; 2295 case L_XTND: 2296 vn_lock(vp, LK_SHARED | LK_RETRY); 2297 error = VOP_GETATTR(vp, &vattr, cred); 2298 VOP_UNLOCK(vp); 2299 if (error) 2300 break; 2301 2302 /* 2303 * If the file references a disk device, then fetch 2304 * the media size and use that to determine the ending 2305 * offset. 2306 */ 2307 if (vattr.va_size == 0 && vp->v_type == VCHR && 2308 fo_ioctl(fp, DIOCGMEDIASIZE, &size, cred, td) == 0) 2309 vattr.va_size = size; 2310 if (noneg && 2311 (vattr.va_size > OFF_MAX || 2312 (offset > 0 && vattr.va_size > OFF_MAX - offset))) { 2313 error = EOVERFLOW; 2314 break; 2315 } 2316 offset += vattr.va_size; 2317 break; 2318 case L_SET: 2319 break; 2320 case SEEK_DATA: 2321 error = fo_ioctl(fp, FIOSEEKDATA, &offset, cred, td); 2322 if (error == ENOTTY) 2323 error = EINVAL; 2324 break; 2325 case SEEK_HOLE: 2326 error = fo_ioctl(fp, FIOSEEKHOLE, &offset, cred, td); 2327 if (error == ENOTTY) 2328 error = EINVAL; 2329 break; 2330 default: 2331 error = EINVAL; 2332 } 2333 if (error == 0 && noneg && offset < 0) 2334 error = EINVAL; 2335 if (error != 0) 2336 goto drop; 2337 VFS_KNOTE_UNLOCKED(vp, 0); 2338 td->td_uretoff.tdu_off = offset; 2339 drop: 2340 foffset_unlock(fp, offset, error != 0 ? FOF_NOUPDATE : 0); 2341 return (error); 2342 } 2343 2344 int 2345 vn_utimes_perm(struct vnode *vp, struct vattr *vap, struct ucred *cred, 2346 struct thread *td) 2347 { 2348 int error; 2349 2350 /* 2351 * Grant permission if the caller is the owner of the file, or 2352 * the super-user, or has ACL_WRITE_ATTRIBUTES permission on 2353 * on the file. If the time pointer is null, then write 2354 * permission on the file is also sufficient. 2355 * 2356 * From NFSv4.1, draft 21, 6.2.1.3.1, Discussion of Mask Attributes: 2357 * A user having ACL_WRITE_DATA or ACL_WRITE_ATTRIBUTES 2358 * will be allowed to set the times [..] to the current 2359 * server time. 2360 */ 2361 error = VOP_ACCESSX(vp, VWRITE_ATTRIBUTES, cred, td); 2362 if (error != 0 && (vap->va_vaflags & VA_UTIMES_NULL) != 0) 2363 error = VOP_ACCESS(vp, VWRITE, cred, td); 2364 return (error); 2365 } 2366 2367 int 2368 vn_fill_kinfo(struct file *fp, struct kinfo_file *kif, struct filedesc *fdp) 2369 { 2370 struct vnode *vp; 2371 int error; 2372 2373 if (fp->f_type == DTYPE_FIFO) 2374 kif->kf_type = KF_TYPE_FIFO; 2375 else 2376 kif->kf_type = KF_TYPE_VNODE; 2377 vp = fp->f_vnode; 2378 vref(vp); 2379 FILEDESC_SUNLOCK(fdp); 2380 error = vn_fill_kinfo_vnode(vp, kif); 2381 vrele(vp); 2382 FILEDESC_SLOCK(fdp); 2383 return (error); 2384 } 2385 2386 static inline void 2387 vn_fill_junk(struct kinfo_file *kif) 2388 { 2389 size_t len, olen; 2390 2391 /* 2392 * Simulate vn_fullpath returning changing values for a given 2393 * vp during e.g. coredump. 2394 */ 2395 len = (arc4random() % (sizeof(kif->kf_path) - 2)) + 1; 2396 olen = strlen(kif->kf_path); 2397 if (len < olen) 2398 strcpy(&kif->kf_path[len - 1], "$"); 2399 else 2400 for (; olen < len; olen++) 2401 strcpy(&kif->kf_path[olen], "A"); 2402 } 2403 2404 int 2405 vn_fill_kinfo_vnode(struct vnode *vp, struct kinfo_file *kif) 2406 { 2407 struct vattr va; 2408 char *fullpath, *freepath; 2409 int error; 2410 2411 kif->kf_un.kf_file.kf_file_type = vntype_to_kinfo(vp->v_type); 2412 freepath = NULL; 2413 fullpath = "-"; 2414 error = vn_fullpath(curthread, vp, &fullpath, &freepath); 2415 if (error == 0) { 2416 strlcpy(kif->kf_path, fullpath, sizeof(kif->kf_path)); 2417 } 2418 if (freepath != NULL) 2419 free(freepath, M_TEMP); 2420 2421 KFAIL_POINT_CODE(DEBUG_FP, fill_kinfo_vnode__random_path, 2422 vn_fill_junk(kif); 2423 ); 2424 2425 /* 2426 * Retrieve vnode attributes. 2427 */ 2428 va.va_fsid = VNOVAL; 2429 va.va_rdev = NODEV; 2430 vn_lock(vp, LK_SHARED | LK_RETRY); 2431 error = VOP_GETATTR(vp, &va, curthread->td_ucred); 2432 VOP_UNLOCK(vp); 2433 if (error != 0) 2434 return (error); 2435 if (va.va_fsid != VNOVAL) 2436 kif->kf_un.kf_file.kf_file_fsid = va.va_fsid; 2437 else 2438 kif->kf_un.kf_file.kf_file_fsid = 2439 vp->v_mount->mnt_stat.f_fsid.val[0]; 2440 kif->kf_un.kf_file.kf_file_fsid_freebsd11 = 2441 kif->kf_un.kf_file.kf_file_fsid; /* truncate */ 2442 kif->kf_un.kf_file.kf_file_fileid = va.va_fileid; 2443 kif->kf_un.kf_file.kf_file_mode = MAKEIMODE(va.va_type, va.va_mode); 2444 kif->kf_un.kf_file.kf_file_size = va.va_size; 2445 kif->kf_un.kf_file.kf_file_rdev = va.va_rdev; 2446 kif->kf_un.kf_file.kf_file_rdev_freebsd11 = 2447 kif->kf_un.kf_file.kf_file_rdev; /* truncate */ 2448 return (0); 2449 } 2450 2451 int 2452 vn_mmap(struct file *fp, vm_map_t map, vm_offset_t *addr, vm_size_t size, 2453 vm_prot_t prot, vm_prot_t cap_maxprot, int flags, vm_ooffset_t foff, 2454 struct thread *td) 2455 { 2456 #ifdef HWPMC_HOOKS 2457 struct pmckern_map_in pkm; 2458 #endif 2459 struct mount *mp; 2460 struct vnode *vp; 2461 vm_object_t object; 2462 vm_prot_t maxprot; 2463 boolean_t writecounted; 2464 int error; 2465 2466 #if defined(COMPAT_FREEBSD7) || defined(COMPAT_FREEBSD6) || \ 2467 defined(COMPAT_FREEBSD5) || defined(COMPAT_FREEBSD4) 2468 /* 2469 * POSIX shared-memory objects are defined to have 2470 * kernel persistence, and are not defined to support 2471 * read(2)/write(2) -- or even open(2). Thus, we can 2472 * use MAP_ASYNC to trade on-disk coherence for speed. 2473 * The shm_open(3) library routine turns on the FPOSIXSHM 2474 * flag to request this behavior. 2475 */ 2476 if ((fp->f_flag & FPOSIXSHM) != 0) 2477 flags |= MAP_NOSYNC; 2478 #endif 2479 vp = fp->f_vnode; 2480 2481 /* 2482 * Ensure that file and memory protections are 2483 * compatible. Note that we only worry about 2484 * writability if mapping is shared; in this case, 2485 * current and max prot are dictated by the open file. 2486 * XXX use the vnode instead? Problem is: what 2487 * credentials do we use for determination? What if 2488 * proc does a setuid? 2489 */ 2490 mp = vp->v_mount; 2491 if (mp != NULL && (mp->mnt_flag & MNT_NOEXEC) != 0) { 2492 maxprot = VM_PROT_NONE; 2493 if ((prot & VM_PROT_EXECUTE) != 0) 2494 return (EACCES); 2495 } else 2496 maxprot = VM_PROT_EXECUTE; 2497 if ((fp->f_flag & FREAD) != 0) 2498 maxprot |= VM_PROT_READ; 2499 else if ((prot & VM_PROT_READ) != 0) 2500 return (EACCES); 2501 2502 /* 2503 * If we are sharing potential changes via MAP_SHARED and we 2504 * are trying to get write permission although we opened it 2505 * without asking for it, bail out. 2506 */ 2507 if ((flags & MAP_SHARED) != 0) { 2508 if ((fp->f_flag & FWRITE) != 0) 2509 maxprot |= VM_PROT_WRITE; 2510 else if ((prot & VM_PROT_WRITE) != 0) 2511 return (EACCES); 2512 } else { 2513 maxprot |= VM_PROT_WRITE; 2514 cap_maxprot |= VM_PROT_WRITE; 2515 } 2516 maxprot &= cap_maxprot; 2517 2518 /* 2519 * For regular files and shared memory, POSIX requires that 2520 * the value of foff be a legitimate offset within the data 2521 * object. In particular, negative offsets are invalid. 2522 * Blocking negative offsets and overflows here avoids 2523 * possible wraparound or user-level access into reserved 2524 * ranges of the data object later. In contrast, POSIX does 2525 * not dictate how offsets are used by device drivers, so in 2526 * the case of a device mapping a negative offset is passed 2527 * on. 2528 */ 2529 if ( 2530 #ifdef _LP64 2531 size > OFF_MAX || 2532 #endif 2533 foff < 0 || foff > OFF_MAX - size) 2534 return (EINVAL); 2535 2536 writecounted = FALSE; 2537 error = vm_mmap_vnode(td, size, prot, &maxprot, &flags, vp, 2538 &foff, &object, &writecounted); 2539 if (error != 0) 2540 return (error); 2541 error = vm_mmap_object(map, addr, size, prot, maxprot, flags, object, 2542 foff, writecounted, td); 2543 if (error != 0) { 2544 /* 2545 * If this mapping was accounted for in the vnode's 2546 * writecount, then undo that now. 2547 */ 2548 if (writecounted) 2549 vm_pager_release_writecount(object, 0, size); 2550 vm_object_deallocate(object); 2551 } 2552 #ifdef HWPMC_HOOKS 2553 /* Inform hwpmc(4) if an executable is being mapped. */ 2554 if (PMC_HOOK_INSTALLED(PMC_FN_MMAP)) { 2555 if ((prot & VM_PROT_EXECUTE) != 0 && error == 0) { 2556 pkm.pm_file = vp; 2557 pkm.pm_address = (uintptr_t) *addr; 2558 PMC_CALL_HOOK_UNLOCKED(td, PMC_FN_MMAP, (void *) &pkm); 2559 } 2560 } 2561 #endif 2562 return (error); 2563 } 2564 2565 void 2566 vn_fsid(struct vnode *vp, struct vattr *va) 2567 { 2568 fsid_t *f; 2569 2570 f = &vp->v_mount->mnt_stat.f_fsid; 2571 va->va_fsid = (uint32_t)f->val[1]; 2572 va->va_fsid <<= sizeof(f->val[1]) * NBBY; 2573 va->va_fsid += (uint32_t)f->val[0]; 2574 } 2575 2576 int 2577 vn_fsync_buf(struct vnode *vp, int waitfor) 2578 { 2579 struct buf *bp, *nbp; 2580 struct bufobj *bo; 2581 struct mount *mp; 2582 int error, maxretry; 2583 2584 error = 0; 2585 maxretry = 10000; /* large, arbitrarily chosen */ 2586 mp = NULL; 2587 if (vp->v_type == VCHR) { 2588 VI_LOCK(vp); 2589 mp = vp->v_rdev->si_mountpt; 2590 VI_UNLOCK(vp); 2591 } 2592 bo = &vp->v_bufobj; 2593 BO_LOCK(bo); 2594 loop1: 2595 /* 2596 * MARK/SCAN initialization to avoid infinite loops. 2597 */ 2598 TAILQ_FOREACH(bp, &bo->bo_dirty.bv_hd, b_bobufs) { 2599 bp->b_vflags &= ~BV_SCANNED; 2600 bp->b_error = 0; 2601 } 2602 2603 /* 2604 * Flush all dirty buffers associated with a vnode. 2605 */ 2606 loop2: 2607 TAILQ_FOREACH_SAFE(bp, &bo->bo_dirty.bv_hd, b_bobufs, nbp) { 2608 if ((bp->b_vflags & BV_SCANNED) != 0) 2609 continue; 2610 bp->b_vflags |= BV_SCANNED; 2611 if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT, NULL)) { 2612 if (waitfor != MNT_WAIT) 2613 continue; 2614 if (BUF_LOCK(bp, 2615 LK_EXCLUSIVE | LK_INTERLOCK | LK_SLEEPFAIL, 2616 BO_LOCKPTR(bo)) != 0) { 2617 BO_LOCK(bo); 2618 goto loop1; 2619 } 2620 BO_LOCK(bo); 2621 } 2622 BO_UNLOCK(bo); 2623 KASSERT(bp->b_bufobj == bo, 2624 ("bp %p wrong b_bufobj %p should be %p", 2625 bp, bp->b_bufobj, bo)); 2626 if ((bp->b_flags & B_DELWRI) == 0) 2627 panic("fsync: not dirty"); 2628 if ((vp->v_object != NULL) && (bp->b_flags & B_CLUSTEROK)) { 2629 vfs_bio_awrite(bp); 2630 } else { 2631 bremfree(bp); 2632 bawrite(bp); 2633 } 2634 if (maxretry < 1000) 2635 pause("dirty", hz < 1000 ? 1 : hz / 1000); 2636 BO_LOCK(bo); 2637 goto loop2; 2638 } 2639 2640 /* 2641 * If synchronous the caller expects us to completely resolve all 2642 * dirty buffers in the system. Wait for in-progress I/O to 2643 * complete (which could include background bitmap writes), then 2644 * retry if dirty blocks still exist. 2645 */ 2646 if (waitfor == MNT_WAIT) { 2647 bufobj_wwait(bo, 0, 0); 2648 if (bo->bo_dirty.bv_cnt > 0) { 2649 /* 2650 * If we are unable to write any of these buffers 2651 * then we fail now rather than trying endlessly 2652 * to write them out. 2653 */ 2654 TAILQ_FOREACH(bp, &bo->bo_dirty.bv_hd, b_bobufs) 2655 if ((error = bp->b_error) != 0) 2656 break; 2657 if ((mp != NULL && mp->mnt_secondary_writes > 0) || 2658 (error == 0 && --maxretry >= 0)) 2659 goto loop1; 2660 if (error == 0) 2661 error = EAGAIN; 2662 } 2663 } 2664 BO_UNLOCK(bo); 2665 if (error != 0) 2666 vn_printf(vp, "fsync: giving up on dirty (error = %d) ", error); 2667 2668 return (error); 2669 } 2670 2671 /* 2672 * Copies a byte range from invp to outvp. Calls VOP_COPY_FILE_RANGE() 2673 * or vn_generic_copy_file_range() after rangelocking the byte ranges, 2674 * to do the actual copy. 2675 * vn_generic_copy_file_range() is factored out, so it can be called 2676 * from a VOP_COPY_FILE_RANGE() call as well, but handles vnodes from 2677 * different file systems. 2678 */ 2679 int 2680 vn_copy_file_range(struct vnode *invp, off_t *inoffp, struct vnode *outvp, 2681 off_t *outoffp, size_t *lenp, unsigned int flags, struct ucred *incred, 2682 struct ucred *outcred, struct thread *fsize_td) 2683 { 2684 int error; 2685 size_t len; 2686 uint64_t uvalin, uvalout; 2687 2688 len = *lenp; 2689 *lenp = 0; /* For error returns. */ 2690 error = 0; 2691 2692 /* Do some sanity checks on the arguments. */ 2693 uvalin = *inoffp; 2694 uvalin += len; 2695 uvalout = *outoffp; 2696 uvalout += len; 2697 if (invp->v_type == VDIR || outvp->v_type == VDIR) 2698 error = EISDIR; 2699 else if (*inoffp < 0 || uvalin > INT64_MAX || uvalin < 2700 (uint64_t)*inoffp || *outoffp < 0 || uvalout > INT64_MAX || 2701 uvalout < (uint64_t)*outoffp || invp->v_type != VREG || 2702 outvp->v_type != VREG) 2703 error = EINVAL; 2704 if (error != 0) 2705 goto out; 2706 2707 /* 2708 * If the two vnode are for the same file system, call 2709 * VOP_COPY_FILE_RANGE(), otherwise call vn_generic_copy_file_range() 2710 * which can handle copies across multiple file systems. 2711 */ 2712 *lenp = len; 2713 if (invp->v_mount == outvp->v_mount) 2714 error = VOP_COPY_FILE_RANGE(invp, inoffp, outvp, outoffp, 2715 lenp, flags, incred, outcred, fsize_td); 2716 else 2717 error = vn_generic_copy_file_range(invp, inoffp, outvp, 2718 outoffp, lenp, flags, incred, outcred, fsize_td); 2719 out: 2720 return (error); 2721 } 2722 2723 /* 2724 * Test len bytes of data starting at dat for all bytes == 0. 2725 * Return true if all bytes are zero, false otherwise. 2726 * Expects dat to be well aligned. 2727 */ 2728 static bool 2729 mem_iszero(void *dat, int len) 2730 { 2731 int i; 2732 const u_int *p; 2733 const char *cp; 2734 2735 for (p = dat; len > 0; len -= sizeof(*p), p++) { 2736 if (len >= sizeof(*p)) { 2737 if (*p != 0) 2738 return (false); 2739 } else { 2740 cp = (const char *)p; 2741 for (i = 0; i < len; i++, cp++) 2742 if (*cp != '\0') 2743 return (false); 2744 } 2745 } 2746 return (true); 2747 } 2748 2749 /* 2750 * Look for a hole in the output file and, if found, adjust *outoffp 2751 * and *xferp to skip past the hole. 2752 * *xferp is the entire hole length to be written and xfer2 is how many bytes 2753 * to be written as 0's upon return. 2754 */ 2755 static off_t 2756 vn_skip_hole(struct vnode *outvp, off_t xfer2, off_t *outoffp, off_t *xferp, 2757 off_t *dataoffp, off_t *holeoffp, struct ucred *cred) 2758 { 2759 int error; 2760 off_t delta; 2761 2762 if (*holeoffp == 0 || *holeoffp <= *outoffp) { 2763 *dataoffp = *outoffp; 2764 error = VOP_IOCTL(outvp, FIOSEEKDATA, dataoffp, 0, cred, 2765 curthread); 2766 if (error == 0) { 2767 *holeoffp = *dataoffp; 2768 error = VOP_IOCTL(outvp, FIOSEEKHOLE, holeoffp, 0, cred, 2769 curthread); 2770 } 2771 if (error != 0 || *holeoffp == *dataoffp) { 2772 /* 2773 * Since outvp is unlocked, it may be possible for 2774 * another thread to do a truncate(), lseek(), write() 2775 * creating a hole at startoff between the above 2776 * VOP_IOCTL() calls, if the other thread does not do 2777 * rangelocking. 2778 * If that happens, *holeoffp == *dataoffp and finding 2779 * the hole has failed, so disable vn_skip_hole(). 2780 */ 2781 *holeoffp = -1; /* Disable use of vn_skip_hole(). */ 2782 return (xfer2); 2783 } 2784 KASSERT(*dataoffp >= *outoffp, 2785 ("vn_skip_hole: dataoff=%jd < outoff=%jd", 2786 (intmax_t)*dataoffp, (intmax_t)*outoffp)); 2787 KASSERT(*holeoffp > *dataoffp, 2788 ("vn_skip_hole: holeoff=%jd <= dataoff=%jd", 2789 (intmax_t)*holeoffp, (intmax_t)*dataoffp)); 2790 } 2791 2792 /* 2793 * If there is a hole before the data starts, advance *outoffp and 2794 * *xferp past the hole. 2795 */ 2796 if (*dataoffp > *outoffp) { 2797 delta = *dataoffp - *outoffp; 2798 if (delta >= *xferp) { 2799 /* Entire *xferp is a hole. */ 2800 *outoffp += *xferp; 2801 *xferp = 0; 2802 return (0); 2803 } 2804 *xferp -= delta; 2805 *outoffp += delta; 2806 xfer2 = MIN(xfer2, *xferp); 2807 } 2808 2809 /* 2810 * If a hole starts before the end of this xfer2, reduce this xfer2 so 2811 * that the write ends at the start of the hole. 2812 * *holeoffp should always be greater than *outoffp, but for the 2813 * non-INVARIANTS case, check this to make sure xfer2 remains a sane 2814 * value. 2815 */ 2816 if (*holeoffp > *outoffp && *holeoffp < *outoffp + xfer2) 2817 xfer2 = *holeoffp - *outoffp; 2818 return (xfer2); 2819 } 2820 2821 /* 2822 * Write an xfer sized chunk to outvp in blksize blocks from dat. 2823 * dat is a maximum of blksize in length and can be written repeatedly in 2824 * the chunk. 2825 * If growfile == true, just grow the file via vn_truncate_locked() instead 2826 * of doing actual writes. 2827 * If checkhole == true, a hole is being punched, so skip over any hole 2828 * already in the output file. 2829 */ 2830 static int 2831 vn_write_outvp(struct vnode *outvp, char *dat, off_t outoff, off_t xfer, 2832 u_long blksize, bool growfile, bool checkhole, struct ucred *cred) 2833 { 2834 struct mount *mp; 2835 off_t dataoff, holeoff, xfer2; 2836 int error, lckf; 2837 2838 /* 2839 * Loop around doing writes of blksize until write has been completed. 2840 * Lock/unlock on each loop iteration so that a bwillwrite() can be 2841 * done for each iteration, since the xfer argument can be very 2842 * large if there is a large hole to punch in the output file. 2843 */ 2844 error = 0; 2845 holeoff = 0; 2846 do { 2847 xfer2 = MIN(xfer, blksize); 2848 if (checkhole) { 2849 /* 2850 * Punching a hole. Skip writing if there is 2851 * already a hole in the output file. 2852 */ 2853 xfer2 = vn_skip_hole(outvp, xfer2, &outoff, &xfer, 2854 &dataoff, &holeoff, cred); 2855 if (xfer == 0) 2856 break; 2857 if (holeoff < 0) 2858 checkhole = false; 2859 KASSERT(xfer2 > 0, ("vn_write_outvp: xfer2=%jd", 2860 (intmax_t)xfer2)); 2861 } 2862 bwillwrite(); 2863 mp = NULL; 2864 error = vn_start_write(outvp, &mp, V_WAIT); 2865 if (error == 0) { 2866 if (MNT_SHARED_WRITES(mp)) 2867 lckf = LK_SHARED; 2868 else 2869 lckf = LK_EXCLUSIVE; 2870 error = vn_lock(outvp, lckf); 2871 } 2872 if (error == 0) { 2873 if (growfile) 2874 error = vn_truncate_locked(outvp, outoff + xfer, 2875 false, cred); 2876 else { 2877 error = vn_rdwr(UIO_WRITE, outvp, dat, xfer2, 2878 outoff, UIO_SYSSPACE, IO_NODELOCKED, 2879 curthread->td_ucred, cred, NULL, curthread); 2880 outoff += xfer2; 2881 xfer -= xfer2; 2882 } 2883 VOP_UNLOCK(outvp); 2884 } 2885 if (mp != NULL) 2886 vn_finished_write(mp); 2887 } while (!growfile && xfer > 0 && error == 0); 2888 return (error); 2889 } 2890 2891 /* 2892 * Copy a byte range of one file to another. This function can handle the 2893 * case where invp and outvp are on different file systems. 2894 * It can also be called by a VOP_COPY_FILE_RANGE() to do the work, if there 2895 * is no better file system specific way to do it. 2896 */ 2897 int 2898 vn_generic_copy_file_range(struct vnode *invp, off_t *inoffp, 2899 struct vnode *outvp, off_t *outoffp, size_t *lenp, unsigned int flags, 2900 struct ucred *incred, struct ucred *outcred, struct thread *fsize_td) 2901 { 2902 struct vattr va; 2903 struct mount *mp; 2904 struct uio io; 2905 off_t startoff, endoff, xfer, xfer2; 2906 u_long blksize; 2907 int error; 2908 bool cantseek, readzeros, eof, lastblock; 2909 ssize_t aresid; 2910 size_t copylen, len, savlen; 2911 char *dat; 2912 long holein, holeout; 2913 2914 holein = holeout = 0; 2915 savlen = len = *lenp; 2916 error = 0; 2917 dat = NULL; 2918 2919 error = vn_lock(invp, LK_SHARED); 2920 if (error != 0) 2921 goto out; 2922 if (VOP_PATHCONF(invp, _PC_MIN_HOLE_SIZE, &holein) != 0) 2923 holein = 0; 2924 VOP_UNLOCK(invp); 2925 2926 mp = NULL; 2927 error = vn_start_write(outvp, &mp, V_WAIT); 2928 if (error == 0) 2929 error = vn_lock(outvp, LK_EXCLUSIVE); 2930 if (error == 0) { 2931 /* 2932 * If fsize_td != NULL, do a vn_rlimit_fsize() call, 2933 * now that outvp is locked. 2934 */ 2935 if (fsize_td != NULL) { 2936 io.uio_offset = *outoffp; 2937 io.uio_resid = len; 2938 error = vn_rlimit_fsize(outvp, &io, fsize_td); 2939 if (error != 0) 2940 error = EFBIG; 2941 } 2942 if (VOP_PATHCONF(outvp, _PC_MIN_HOLE_SIZE, &holeout) != 0) 2943 holeout = 0; 2944 /* 2945 * Holes that are past EOF do not need to be written as a block 2946 * of zero bytes. So, truncate the output file as far as 2947 * possible and then use va.va_size to decide if writing 0 2948 * bytes is necessary in the loop below. 2949 */ 2950 if (error == 0) 2951 error = VOP_GETATTR(outvp, &va, outcred); 2952 if (error == 0 && va.va_size > *outoffp && va.va_size <= 2953 *outoffp + len) { 2954 #ifdef MAC 2955 error = mac_vnode_check_write(curthread->td_ucred, 2956 outcred, outvp); 2957 if (error == 0) 2958 #endif 2959 error = vn_truncate_locked(outvp, *outoffp, 2960 false, outcred); 2961 if (error == 0) 2962 va.va_size = *outoffp; 2963 } 2964 VOP_UNLOCK(outvp); 2965 } 2966 if (mp != NULL) 2967 vn_finished_write(mp); 2968 if (error != 0) 2969 goto out; 2970 2971 /* 2972 * Set the blksize to the larger of the hole sizes for invp and outvp. 2973 * If hole sizes aren't available, set the blksize to the larger 2974 * f_iosize of invp and outvp. 2975 * This code expects the hole sizes and f_iosizes to be powers of 2. 2976 * This value is clipped at 4Kbytes and 1Mbyte. 2977 */ 2978 blksize = MAX(holein, holeout); 2979 if (blksize == 0) 2980 blksize = MAX(invp->v_mount->mnt_stat.f_iosize, 2981 outvp->v_mount->mnt_stat.f_iosize); 2982 if (blksize < 4096) 2983 blksize = 4096; 2984 else if (blksize > 1024 * 1024) 2985 blksize = 1024 * 1024; 2986 dat = malloc(blksize, M_TEMP, M_WAITOK); 2987 2988 /* 2989 * If VOP_IOCTL(FIOSEEKHOLE) works for invp, use it and FIOSEEKDATA 2990 * to find holes. Otherwise, just scan the read block for all 0s 2991 * in the inner loop where the data copying is done. 2992 * Note that some file systems such as NFSv3, NFSv4.0 and NFSv4.1 may 2993 * support holes on the server, but do not support FIOSEEKHOLE. 2994 */ 2995 eof = false; 2996 while (len > 0 && error == 0 && !eof) { 2997 endoff = 0; /* To shut up compilers. */ 2998 cantseek = true; 2999 startoff = *inoffp; 3000 copylen = len; 3001 3002 /* 3003 * Find the next data area. If there is just a hole to EOF, 3004 * FIOSEEKDATA should fail and then we drop down into the 3005 * inner loop and create the hole on the outvp file. 3006 * (I do not know if any file system will report a hole to 3007 * EOF via FIOSEEKHOLE, but I am pretty sure FIOSEEKDATA 3008 * will fail for those file systems.) 3009 * 3010 * For input files that don't support FIOSEEKDATA/FIOSEEKHOLE, 3011 * the code just falls through to the inner copy loop. 3012 */ 3013 error = EINVAL; 3014 if (holein > 0) 3015 error = VOP_IOCTL(invp, FIOSEEKDATA, &startoff, 0, 3016 incred, curthread); 3017 if (error == 0) { 3018 endoff = startoff; 3019 error = VOP_IOCTL(invp, FIOSEEKHOLE, &endoff, 0, 3020 incred, curthread); 3021 /* 3022 * Since invp is unlocked, it may be possible for 3023 * another thread to do a truncate(), lseek(), write() 3024 * creating a hole at startoff between the above 3025 * VOP_IOCTL() calls, if the other thread does not do 3026 * rangelocking. 3027 * If that happens, startoff == endoff and finding 3028 * the hole has failed, so set an error. 3029 */ 3030 if (error == 0 && startoff == endoff) 3031 error = EINVAL; /* Any error. Reset to 0. */ 3032 } 3033 if (error == 0) { 3034 if (startoff > *inoffp) { 3035 /* Found hole before data block. */ 3036 xfer = MIN(startoff - *inoffp, len); 3037 if (*outoffp < va.va_size) { 3038 /* Must write 0s to punch hole. */ 3039 xfer2 = MIN(va.va_size - *outoffp, 3040 xfer); 3041 memset(dat, 0, MIN(xfer2, blksize)); 3042 error = vn_write_outvp(outvp, dat, 3043 *outoffp, xfer2, blksize, false, 3044 holeout > 0, outcred); 3045 } 3046 3047 if (error == 0 && *outoffp + xfer > 3048 va.va_size && xfer == len) 3049 /* Grow last block. */ 3050 error = vn_write_outvp(outvp, dat, 3051 *outoffp, xfer, blksize, true, 3052 false, outcred); 3053 if (error == 0) { 3054 *inoffp += xfer; 3055 *outoffp += xfer; 3056 len -= xfer; 3057 } 3058 } 3059 copylen = MIN(len, endoff - startoff); 3060 cantseek = false; 3061 } else { 3062 cantseek = true; 3063 startoff = *inoffp; 3064 copylen = len; 3065 error = 0; 3066 } 3067 3068 xfer = blksize; 3069 if (cantseek) { 3070 /* 3071 * Set first xfer to end at a block boundary, so that 3072 * holes are more likely detected in the loop below via 3073 * the for all bytes 0 method. 3074 */ 3075 xfer -= (*inoffp % blksize); 3076 } 3077 /* Loop copying the data block. */ 3078 while (copylen > 0 && error == 0 && !eof) { 3079 if (copylen < xfer) 3080 xfer = copylen; 3081 error = vn_lock(invp, LK_SHARED); 3082 if (error != 0) 3083 goto out; 3084 error = vn_rdwr(UIO_READ, invp, dat, xfer, 3085 startoff, UIO_SYSSPACE, IO_NODELOCKED, 3086 curthread->td_ucred, incred, &aresid, 3087 curthread); 3088 VOP_UNLOCK(invp); 3089 lastblock = false; 3090 if (error == 0 && aresid > 0) { 3091 /* Stop the copy at EOF on the input file. */ 3092 xfer -= aresid; 3093 eof = true; 3094 lastblock = true; 3095 } 3096 if (error == 0) { 3097 /* 3098 * Skip the write for holes past the initial EOF 3099 * of the output file, unless this is the last 3100 * write of the output file at EOF. 3101 */ 3102 readzeros = cantseek ? mem_iszero(dat, xfer) : 3103 false; 3104 if (xfer == len) 3105 lastblock = true; 3106 if (!cantseek || *outoffp < va.va_size || 3107 lastblock || !readzeros) 3108 error = vn_write_outvp(outvp, dat, 3109 *outoffp, xfer, blksize, 3110 readzeros && lastblock && 3111 *outoffp >= va.va_size, false, 3112 outcred); 3113 if (error == 0) { 3114 *inoffp += xfer; 3115 startoff += xfer; 3116 *outoffp += xfer; 3117 copylen -= xfer; 3118 len -= xfer; 3119 } 3120 } 3121 xfer = blksize; 3122 } 3123 } 3124 out: 3125 *lenp = savlen - len; 3126 free(dat, M_TEMP); 3127 return (error); 3128 } 3129 3130 static int 3131 vn_fallocate(struct file *fp, off_t offset, off_t len, struct thread *td) 3132 { 3133 struct mount *mp; 3134 struct vnode *vp; 3135 off_t olen, ooffset; 3136 int error; 3137 #ifdef AUDIT 3138 int audited_vnode1 = 0; 3139 #endif 3140 3141 vp = fp->f_vnode; 3142 if (vp->v_type != VREG) 3143 return (ENODEV); 3144 3145 /* Allocating blocks may take a long time, so iterate. */ 3146 for (;;) { 3147 olen = len; 3148 ooffset = offset; 3149 3150 bwillwrite(); 3151 mp = NULL; 3152 error = vn_start_write(vp, &mp, V_WAIT | PCATCH); 3153 if (error != 0) 3154 break; 3155 error = vn_lock(vp, LK_EXCLUSIVE); 3156 if (error != 0) { 3157 vn_finished_write(mp); 3158 break; 3159 } 3160 #ifdef AUDIT 3161 if (!audited_vnode1) { 3162 AUDIT_ARG_VNODE1(vp); 3163 audited_vnode1 = 1; 3164 } 3165 #endif 3166 #ifdef MAC 3167 error = mac_vnode_check_write(td->td_ucred, fp->f_cred, vp); 3168 if (error == 0) 3169 #endif 3170 error = VOP_ALLOCATE(vp, &offset, &len); 3171 VOP_UNLOCK(vp); 3172 vn_finished_write(mp); 3173 3174 if (olen + ooffset != offset + len) { 3175 panic("offset + len changed from %jx/%jx to %jx/%jx", 3176 ooffset, olen, offset, len); 3177 } 3178 if (error != 0 || len == 0) 3179 break; 3180 KASSERT(olen > len, ("Iteration did not make progress?")); 3181 maybe_yield(); 3182 } 3183 3184 return (error); 3185 } 3186