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/vnode_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; 106 107 struct fileops vnops = { 108 .fo_read = vn_io_fault, 109 .fo_write = vn_io_fault, 110 .fo_truncate = vn_truncate, 111 .fo_ioctl = vn_ioctl, 112 .fo_poll = vn_poll, 113 .fo_kqfilter = vn_kqfilter, 114 .fo_stat = vn_statfile, 115 .fo_close = vn_closefile, 116 .fo_chmod = vn_chmod, 117 .fo_chown = vn_chown, 118 .fo_sendfile = vn_sendfile, 119 .fo_seek = vn_seek, 120 .fo_fill_kinfo = vn_fill_kinfo, 121 .fo_mmap = vn_mmap, 122 .fo_flags = DFLAG_PASSABLE | DFLAG_SEEKABLE 123 }; 124 125 static const int io_hold_cnt = 16; 126 static int vn_io_fault_enable = 1; 127 SYSCTL_INT(_debug, OID_AUTO, vn_io_fault_enable, CTLFLAG_RW, 128 &vn_io_fault_enable, 0, "Enable vn_io_fault lock avoidance"); 129 static int vn_io_fault_prefault = 0; 130 SYSCTL_INT(_debug, OID_AUTO, vn_io_fault_prefault, CTLFLAG_RW, 131 &vn_io_fault_prefault, 0, "Enable vn_io_fault prefaulting"); 132 static u_long vn_io_faults_cnt; 133 SYSCTL_ULONG(_debug, OID_AUTO, vn_io_faults, CTLFLAG_RD, 134 &vn_io_faults_cnt, 0, "Count of vn_io_fault lock avoidance triggers"); 135 136 /* 137 * Returns true if vn_io_fault mode of handling the i/o request should 138 * be used. 139 */ 140 static bool 141 do_vn_io_fault(struct vnode *vp, struct uio *uio) 142 { 143 struct mount *mp; 144 145 return (uio->uio_segflg == UIO_USERSPACE && vp->v_type == VREG && 146 (mp = vp->v_mount) != NULL && 147 (mp->mnt_kern_flag & MNTK_NO_IOPF) != 0 && vn_io_fault_enable); 148 } 149 150 /* 151 * Structure used to pass arguments to vn_io_fault1(), to do either 152 * file- or vnode-based I/O calls. 153 */ 154 struct vn_io_fault_args { 155 enum { 156 VN_IO_FAULT_FOP, 157 VN_IO_FAULT_VOP 158 } kind; 159 struct ucred *cred; 160 int flags; 161 union { 162 struct fop_args_tag { 163 struct file *fp; 164 fo_rdwr_t *doio; 165 } fop_args; 166 struct vop_args_tag { 167 struct vnode *vp; 168 } vop_args; 169 } args; 170 }; 171 172 static int vn_io_fault1(struct vnode *vp, struct uio *uio, 173 struct vn_io_fault_args *args, struct thread *td); 174 175 int 176 vn_open(struct nameidata *ndp, int *flagp, int cmode, struct file *fp) 177 { 178 struct thread *td = ndp->ni_cnd.cn_thread; 179 180 return (vn_open_cred(ndp, flagp, cmode, 0, td->td_ucred, fp)); 181 } 182 183 /* 184 * Common code for vnode open operations via a name lookup. 185 * Lookup the vnode and invoke VOP_CREATE if needed. 186 * Check permissions, and call the VOP_OPEN or VOP_CREATE routine. 187 * 188 * Note that this does NOT free nameidata for the successful case, 189 * due to the NDINIT being done elsewhere. 190 */ 191 int 192 vn_open_cred(struct nameidata *ndp, int *flagp, int cmode, u_int vn_open_flags, 193 struct ucred *cred, struct file *fp) 194 { 195 struct vnode *vp; 196 struct mount *mp; 197 struct thread *td = ndp->ni_cnd.cn_thread; 198 struct vattr vat; 199 struct vattr *vap = &vat; 200 int fmode, error; 201 202 restart: 203 fmode = *flagp; 204 if ((fmode & (O_CREAT | O_EXCL | O_DIRECTORY)) == (O_CREAT | 205 O_EXCL | O_DIRECTORY)) 206 return (EINVAL); 207 else if ((fmode & (O_CREAT | O_DIRECTORY)) == O_CREAT) { 208 ndp->ni_cnd.cn_nameiop = CREATE; 209 /* 210 * Set NOCACHE to avoid flushing the cache when 211 * rolling in many files at once. 212 */ 213 ndp->ni_cnd.cn_flags = ISOPEN | LOCKPARENT | LOCKLEAF | NOCACHE; 214 if ((fmode & O_EXCL) == 0 && (fmode & O_NOFOLLOW) == 0) 215 ndp->ni_cnd.cn_flags |= FOLLOW; 216 if ((fmode & O_BENEATH) != 0) 217 ndp->ni_cnd.cn_flags |= BENEATH; 218 if (!(vn_open_flags & VN_OPEN_NOAUDIT)) 219 ndp->ni_cnd.cn_flags |= AUDITVNODE1; 220 if (vn_open_flags & VN_OPEN_NOCAPCHECK) 221 ndp->ni_cnd.cn_flags |= NOCAPCHECK; 222 bwillwrite(); 223 if ((error = namei(ndp)) != 0) 224 return (error); 225 if (ndp->ni_vp == NULL) { 226 VATTR_NULL(vap); 227 vap->va_type = VREG; 228 vap->va_mode = cmode; 229 if (fmode & O_EXCL) 230 vap->va_vaflags |= VA_EXCLUSIVE; 231 if (vn_start_write(ndp->ni_dvp, &mp, V_NOWAIT) != 0) { 232 NDFREE(ndp, NDF_ONLY_PNBUF); 233 vput(ndp->ni_dvp); 234 if ((error = vn_start_write(NULL, &mp, 235 V_XSLEEP | PCATCH)) != 0) 236 return (error); 237 goto restart; 238 } 239 if ((vn_open_flags & VN_OPEN_NAMECACHE) != 0) 240 ndp->ni_cnd.cn_flags |= MAKEENTRY; 241 #ifdef MAC 242 error = mac_vnode_check_create(cred, ndp->ni_dvp, 243 &ndp->ni_cnd, vap); 244 if (error == 0) 245 #endif 246 error = VOP_CREATE(ndp->ni_dvp, &ndp->ni_vp, 247 &ndp->ni_cnd, vap); 248 vput(ndp->ni_dvp); 249 vn_finished_write(mp); 250 if (error) { 251 NDFREE(ndp, NDF_ONLY_PNBUF); 252 return (error); 253 } 254 fmode &= ~O_TRUNC; 255 vp = ndp->ni_vp; 256 } else { 257 if (ndp->ni_dvp == ndp->ni_vp) 258 vrele(ndp->ni_dvp); 259 else 260 vput(ndp->ni_dvp); 261 ndp->ni_dvp = NULL; 262 vp = ndp->ni_vp; 263 if (fmode & O_EXCL) { 264 error = EEXIST; 265 goto bad; 266 } 267 fmode &= ~O_CREAT; 268 } 269 } else { 270 ndp->ni_cnd.cn_nameiop = LOOKUP; 271 ndp->ni_cnd.cn_flags = ISOPEN | 272 ((fmode & O_NOFOLLOW) ? NOFOLLOW : FOLLOW) | LOCKLEAF; 273 if (!(fmode & FWRITE)) 274 ndp->ni_cnd.cn_flags |= LOCKSHARED; 275 if ((fmode & O_BENEATH) != 0) 276 ndp->ni_cnd.cn_flags |= BENEATH; 277 if (!(vn_open_flags & VN_OPEN_NOAUDIT)) 278 ndp->ni_cnd.cn_flags |= AUDITVNODE1; 279 if (vn_open_flags & VN_OPEN_NOCAPCHECK) 280 ndp->ni_cnd.cn_flags |= NOCAPCHECK; 281 if ((error = namei(ndp)) != 0) 282 return (error); 283 vp = ndp->ni_vp; 284 } 285 error = vn_open_vnode(vp, fmode, cred, td, fp); 286 if (error) 287 goto bad; 288 *flagp = fmode; 289 return (0); 290 bad: 291 NDFREE(ndp, NDF_ONLY_PNBUF); 292 vput(vp); 293 *flagp = fmode; 294 ndp->ni_vp = NULL; 295 return (error); 296 } 297 298 static int 299 vn_open_vnode_advlock(struct vnode *vp, int fmode, struct file *fp) 300 { 301 struct flock lf; 302 int error, lock_flags, type; 303 304 ASSERT_VOP_LOCKED(vp, "vn_open_vnode_advlock"); 305 if ((fmode & (O_EXLOCK | O_SHLOCK)) == 0) 306 return (0); 307 KASSERT(fp != NULL, ("open with flock requires fp")); 308 if (fp->f_type != DTYPE_NONE && fp->f_type != DTYPE_VNODE) 309 return (EOPNOTSUPP); 310 311 lock_flags = VOP_ISLOCKED(vp); 312 VOP_UNLOCK(vp, 0); 313 314 lf.l_whence = SEEK_SET; 315 lf.l_start = 0; 316 lf.l_len = 0; 317 lf.l_type = (fmode & O_EXLOCK) != 0 ? F_WRLCK : F_RDLCK; 318 type = F_FLOCK; 319 if ((fmode & FNONBLOCK) == 0) 320 type |= F_WAIT; 321 error = VOP_ADVLOCK(vp, (caddr_t)fp, F_SETLK, &lf, type); 322 if (error == 0) 323 fp->f_flag |= FHASLOCK; 324 325 vn_lock(vp, lock_flags | LK_RETRY); 326 if (error == 0 && (vp->v_iflag & VI_DOOMED) != 0) 327 error = ENOENT; 328 return (error); 329 } 330 331 /* 332 * Common code for vnode open operations once a vnode is located. 333 * Check permissions, and call the VOP_OPEN routine. 334 */ 335 int 336 vn_open_vnode(struct vnode *vp, int fmode, struct ucred *cred, 337 struct thread *td, struct file *fp) 338 { 339 accmode_t accmode; 340 int error; 341 342 if (vp->v_type == VLNK) 343 return (EMLINK); 344 if (vp->v_type == VSOCK) 345 return (EOPNOTSUPP); 346 if (vp->v_type != VDIR && fmode & O_DIRECTORY) 347 return (ENOTDIR); 348 accmode = 0; 349 if (fmode & (FWRITE | O_TRUNC)) { 350 if (vp->v_type == VDIR) 351 return (EISDIR); 352 accmode |= VWRITE; 353 } 354 if (fmode & FREAD) 355 accmode |= VREAD; 356 if (fmode & FEXEC) 357 accmode |= VEXEC; 358 if ((fmode & O_APPEND) && (fmode & FWRITE)) 359 accmode |= VAPPEND; 360 #ifdef MAC 361 if (fmode & O_CREAT) 362 accmode |= VCREAT; 363 if (fmode & O_VERIFY) 364 accmode |= VVERIFY; 365 error = mac_vnode_check_open(cred, vp, accmode); 366 if (error) 367 return (error); 368 369 accmode &= ~(VCREAT | VVERIFY); 370 #endif 371 if ((fmode & O_CREAT) == 0 && accmode != 0) { 372 error = VOP_ACCESS(vp, accmode, cred, td); 373 if (error != 0) 374 return (error); 375 } 376 if (vp->v_type == VFIFO && VOP_ISLOCKED(vp) != LK_EXCLUSIVE) 377 vn_lock(vp, LK_UPGRADE | LK_RETRY); 378 error = VOP_OPEN(vp, fmode, cred, td, fp); 379 if (error != 0) 380 return (error); 381 382 error = vn_open_vnode_advlock(vp, fmode, fp); 383 if (error == 0 && (fmode & FWRITE) != 0) { 384 error = VOP_ADD_WRITECOUNT(vp, 1); 385 if (error == 0) { 386 CTR3(KTR_VFS, "%s: vp %p v_writecount increased to %d", 387 __func__, vp, vp->v_writecount); 388 } 389 } 390 391 /* 392 * Error from advlock or VOP_ADD_WRITECOUNT() still requires 393 * calling VOP_CLOSE() to pair with earlier VOP_OPEN(). 394 * Arrange for that by having fdrop() to use vn_closefile(). 395 */ 396 if (error != 0) { 397 fp->f_flag |= FOPENFAILED; 398 fp->f_vnode = vp; 399 if (fp->f_ops == &badfileops) { 400 fp->f_type = DTYPE_VNODE; 401 fp->f_ops = &vnops; 402 } 403 vref(vp); 404 } 405 406 ASSERT_VOP_LOCKED(vp, "vn_open_vnode"); 407 return (error); 408 409 } 410 411 /* 412 * Check for write permissions on the specified vnode. 413 * Prototype text segments cannot be written. 414 * It is racy. 415 */ 416 int 417 vn_writechk(struct vnode *vp) 418 { 419 420 ASSERT_VOP_LOCKED(vp, "vn_writechk"); 421 /* 422 * If there's shared text associated with 423 * the vnode, try to free it up once. If 424 * we fail, we can't allow writing. 425 */ 426 if (VOP_IS_TEXT(vp)) 427 return (ETXTBSY); 428 429 return (0); 430 } 431 432 /* 433 * Vnode close call 434 */ 435 static int 436 vn_close1(struct vnode *vp, int flags, struct ucred *file_cred, 437 struct thread *td, bool keep_ref) 438 { 439 struct mount *mp; 440 int error, lock_flags; 441 442 if (vp->v_type != VFIFO && (flags & FWRITE) == 0 && 443 MNT_EXTENDED_SHARED(vp->v_mount)) 444 lock_flags = LK_SHARED; 445 else 446 lock_flags = LK_EXCLUSIVE; 447 448 vn_start_write(vp, &mp, V_WAIT); 449 vn_lock(vp, lock_flags | LK_RETRY); 450 AUDIT_ARG_VNODE1(vp); 451 if ((flags & (FWRITE | FOPENFAILED)) == FWRITE) { 452 VOP_ADD_WRITECOUNT_CHECKED(vp, -1); 453 CTR3(KTR_VFS, "%s: vp %p v_writecount decreased to %d", 454 __func__, vp, vp->v_writecount); 455 } 456 error = VOP_CLOSE(vp, flags, file_cred, td); 457 if (keep_ref) 458 VOP_UNLOCK(vp, 0); 459 else 460 vput(vp); 461 vn_finished_write(mp); 462 return (error); 463 } 464 465 int 466 vn_close(struct vnode *vp, int flags, struct ucred *file_cred, 467 struct thread *td) 468 { 469 470 return (vn_close1(vp, flags, file_cred, td, false)); 471 } 472 473 /* 474 * Heuristic to detect sequential operation. 475 */ 476 static int 477 sequential_heuristic(struct uio *uio, struct file *fp) 478 { 479 480 ASSERT_VOP_LOCKED(fp->f_vnode, __func__); 481 if (fp->f_flag & FRDAHEAD) 482 return (fp->f_seqcount << IO_SEQSHIFT); 483 484 /* 485 * Offset 0 is handled specially. open() sets f_seqcount to 1 so 486 * that the first I/O is normally considered to be slightly 487 * sequential. Seeking to offset 0 doesn't change sequentiality 488 * unless previous seeks have reduced f_seqcount to 0, in which 489 * case offset 0 is not special. 490 */ 491 if ((uio->uio_offset == 0 && fp->f_seqcount > 0) || 492 uio->uio_offset == fp->f_nextoff) { 493 /* 494 * f_seqcount is in units of fixed-size blocks so that it 495 * depends mainly on the amount of sequential I/O and not 496 * much on the number of sequential I/O's. The fixed size 497 * of 16384 is hard-coded here since it is (not quite) just 498 * a magic size that works well here. This size is more 499 * closely related to the best I/O size for real disks than 500 * to any block size used by software. 501 */ 502 fp->f_seqcount += MIN(IO_SEQMAX, 503 howmany(uio->uio_resid, 16384)); 504 if (fp->f_seqcount > IO_SEQMAX) 505 fp->f_seqcount = IO_SEQMAX; 506 return (fp->f_seqcount << IO_SEQSHIFT); 507 } 508 509 /* Not sequential. Quickly draw-down sequentiality. */ 510 if (fp->f_seqcount > 1) 511 fp->f_seqcount = 1; 512 else 513 fp->f_seqcount = 0; 514 return (0); 515 } 516 517 /* 518 * Package up an I/O request on a vnode into a uio and do it. 519 */ 520 int 521 vn_rdwr(enum uio_rw rw, struct vnode *vp, void *base, int len, off_t offset, 522 enum uio_seg segflg, int ioflg, struct ucred *active_cred, 523 struct ucred *file_cred, ssize_t *aresid, struct thread *td) 524 { 525 struct uio auio; 526 struct iovec aiov; 527 struct mount *mp; 528 struct ucred *cred; 529 void *rl_cookie; 530 struct vn_io_fault_args args; 531 int error, lock_flags; 532 533 if (offset < 0 && vp->v_type != VCHR) 534 return (EINVAL); 535 auio.uio_iov = &aiov; 536 auio.uio_iovcnt = 1; 537 aiov.iov_base = base; 538 aiov.iov_len = len; 539 auio.uio_resid = len; 540 auio.uio_offset = offset; 541 auio.uio_segflg = segflg; 542 auio.uio_rw = rw; 543 auio.uio_td = td; 544 error = 0; 545 546 if ((ioflg & IO_NODELOCKED) == 0) { 547 if ((ioflg & IO_RANGELOCKED) == 0) { 548 if (rw == UIO_READ) { 549 rl_cookie = vn_rangelock_rlock(vp, offset, 550 offset + len); 551 } else { 552 rl_cookie = vn_rangelock_wlock(vp, offset, 553 offset + len); 554 } 555 } else 556 rl_cookie = NULL; 557 mp = NULL; 558 if (rw == UIO_WRITE) { 559 if (vp->v_type != VCHR && 560 (error = vn_start_write(vp, &mp, V_WAIT | PCATCH)) 561 != 0) 562 goto out; 563 if (MNT_SHARED_WRITES(mp) || 564 ((mp == NULL) && MNT_SHARED_WRITES(vp->v_mount))) 565 lock_flags = LK_SHARED; 566 else 567 lock_flags = LK_EXCLUSIVE; 568 } else 569 lock_flags = LK_SHARED; 570 vn_lock(vp, lock_flags | LK_RETRY); 571 } else 572 rl_cookie = NULL; 573 574 ASSERT_VOP_LOCKED(vp, "IO_NODELOCKED with no vp lock held"); 575 #ifdef MAC 576 if ((ioflg & IO_NOMACCHECK) == 0) { 577 if (rw == UIO_READ) 578 error = mac_vnode_check_read(active_cred, file_cred, 579 vp); 580 else 581 error = mac_vnode_check_write(active_cred, file_cred, 582 vp); 583 } 584 #endif 585 if (error == 0) { 586 if (file_cred != NULL) 587 cred = file_cred; 588 else 589 cred = active_cred; 590 if (do_vn_io_fault(vp, &auio)) { 591 args.kind = VN_IO_FAULT_VOP; 592 args.cred = cred; 593 args.flags = ioflg; 594 args.args.vop_args.vp = vp; 595 error = vn_io_fault1(vp, &auio, &args, td); 596 } else if (rw == UIO_READ) { 597 error = VOP_READ(vp, &auio, ioflg, cred); 598 } else /* if (rw == UIO_WRITE) */ { 599 error = VOP_WRITE(vp, &auio, ioflg, cred); 600 } 601 } 602 if (aresid) 603 *aresid = auio.uio_resid; 604 else 605 if (auio.uio_resid && error == 0) 606 error = EIO; 607 if ((ioflg & IO_NODELOCKED) == 0) { 608 VOP_UNLOCK(vp, 0); 609 if (mp != NULL) 610 vn_finished_write(mp); 611 } 612 out: 613 if (rl_cookie != NULL) 614 vn_rangelock_unlock(vp, rl_cookie); 615 return (error); 616 } 617 618 /* 619 * Package up an I/O request on a vnode into a uio and do it. The I/O 620 * request is split up into smaller chunks and we try to avoid saturating 621 * the buffer cache while potentially holding a vnode locked, so we 622 * check bwillwrite() before calling vn_rdwr(). We also call kern_yield() 623 * to give other processes a chance to lock the vnode (either other processes 624 * core'ing the same binary, or unrelated processes scanning the directory). 625 */ 626 int 627 vn_rdwr_inchunks(enum uio_rw rw, struct vnode *vp, void *base, size_t len, 628 off_t offset, enum uio_seg segflg, int ioflg, struct ucred *active_cred, 629 struct ucred *file_cred, size_t *aresid, struct thread *td) 630 { 631 int error = 0; 632 ssize_t iaresid; 633 634 do { 635 int chunk; 636 637 /* 638 * Force `offset' to a multiple of MAXBSIZE except possibly 639 * for the first chunk, so that filesystems only need to 640 * write full blocks except possibly for the first and last 641 * chunks. 642 */ 643 chunk = MAXBSIZE - (uoff_t)offset % MAXBSIZE; 644 645 if (chunk > len) 646 chunk = len; 647 if (rw != UIO_READ && vp->v_type == VREG) 648 bwillwrite(); 649 iaresid = 0; 650 error = vn_rdwr(rw, vp, base, chunk, offset, segflg, 651 ioflg, active_cred, file_cred, &iaresid, td); 652 len -= chunk; /* aresid calc already includes length */ 653 if (error) 654 break; 655 offset += chunk; 656 base = (char *)base + chunk; 657 kern_yield(PRI_USER); 658 } while (len); 659 if (aresid) 660 *aresid = len + iaresid; 661 return (error); 662 } 663 664 off_t 665 foffset_lock(struct file *fp, int flags) 666 { 667 struct mtx *mtxp; 668 off_t res; 669 670 KASSERT((flags & FOF_OFFSET) == 0, ("FOF_OFFSET passed")); 671 672 #if OFF_MAX <= LONG_MAX 673 /* 674 * Caller only wants the current f_offset value. Assume that 675 * the long and shorter integer types reads are atomic. 676 */ 677 if ((flags & FOF_NOLOCK) != 0) 678 return (fp->f_offset); 679 #endif 680 681 /* 682 * According to McKusick the vn lock was protecting f_offset here. 683 * It is now protected by the FOFFSET_LOCKED flag. 684 */ 685 mtxp = mtx_pool_find(mtxpool_sleep, fp); 686 mtx_lock(mtxp); 687 if ((flags & FOF_NOLOCK) == 0) { 688 while (fp->f_vnread_flags & FOFFSET_LOCKED) { 689 fp->f_vnread_flags |= FOFFSET_LOCK_WAITING; 690 msleep(&fp->f_vnread_flags, mtxp, PUSER -1, 691 "vofflock", 0); 692 } 693 fp->f_vnread_flags |= FOFFSET_LOCKED; 694 } 695 res = fp->f_offset; 696 mtx_unlock(mtxp); 697 return (res); 698 } 699 700 void 701 foffset_unlock(struct file *fp, off_t val, int flags) 702 { 703 struct mtx *mtxp; 704 705 KASSERT((flags & FOF_OFFSET) == 0, ("FOF_OFFSET passed")); 706 707 #if OFF_MAX <= LONG_MAX 708 if ((flags & FOF_NOLOCK) != 0) { 709 if ((flags & FOF_NOUPDATE) == 0) 710 fp->f_offset = val; 711 if ((flags & FOF_NEXTOFF) != 0) 712 fp->f_nextoff = val; 713 return; 714 } 715 #endif 716 717 mtxp = mtx_pool_find(mtxpool_sleep, fp); 718 mtx_lock(mtxp); 719 if ((flags & FOF_NOUPDATE) == 0) 720 fp->f_offset = val; 721 if ((flags & FOF_NEXTOFF) != 0) 722 fp->f_nextoff = val; 723 if ((flags & FOF_NOLOCK) == 0) { 724 KASSERT((fp->f_vnread_flags & FOFFSET_LOCKED) != 0, 725 ("Lost FOFFSET_LOCKED")); 726 if (fp->f_vnread_flags & FOFFSET_LOCK_WAITING) 727 wakeup(&fp->f_vnread_flags); 728 fp->f_vnread_flags = 0; 729 } 730 mtx_unlock(mtxp); 731 } 732 733 void 734 foffset_lock_uio(struct file *fp, struct uio *uio, int flags) 735 { 736 737 if ((flags & FOF_OFFSET) == 0) 738 uio->uio_offset = foffset_lock(fp, flags); 739 } 740 741 void 742 foffset_unlock_uio(struct file *fp, struct uio *uio, int flags) 743 { 744 745 if ((flags & FOF_OFFSET) == 0) 746 foffset_unlock(fp, uio->uio_offset, flags); 747 } 748 749 static int 750 get_advice(struct file *fp, struct uio *uio) 751 { 752 struct mtx *mtxp; 753 int ret; 754 755 ret = POSIX_FADV_NORMAL; 756 if (fp->f_advice == NULL || fp->f_vnode->v_type != VREG) 757 return (ret); 758 759 mtxp = mtx_pool_find(mtxpool_sleep, fp); 760 mtx_lock(mtxp); 761 if (fp->f_advice != NULL && 762 uio->uio_offset >= fp->f_advice->fa_start && 763 uio->uio_offset + uio->uio_resid <= fp->f_advice->fa_end) 764 ret = fp->f_advice->fa_advice; 765 mtx_unlock(mtxp); 766 return (ret); 767 } 768 769 /* 770 * File table vnode read routine. 771 */ 772 static int 773 vn_read(struct file *fp, struct uio *uio, struct ucred *active_cred, int flags, 774 struct thread *td) 775 { 776 struct vnode *vp; 777 off_t orig_offset; 778 int error, ioflag; 779 int advice; 780 781 KASSERT(uio->uio_td == td, ("uio_td %p is not td %p", 782 uio->uio_td, td)); 783 KASSERT(flags & FOF_OFFSET, ("No FOF_OFFSET")); 784 vp = fp->f_vnode; 785 ioflag = 0; 786 if (fp->f_flag & FNONBLOCK) 787 ioflag |= IO_NDELAY; 788 if (fp->f_flag & O_DIRECT) 789 ioflag |= IO_DIRECT; 790 advice = get_advice(fp, uio); 791 vn_lock(vp, LK_SHARED | LK_RETRY); 792 793 switch (advice) { 794 case POSIX_FADV_NORMAL: 795 case POSIX_FADV_SEQUENTIAL: 796 case POSIX_FADV_NOREUSE: 797 ioflag |= sequential_heuristic(uio, fp); 798 break; 799 case POSIX_FADV_RANDOM: 800 /* Disable read-ahead for random I/O. */ 801 break; 802 } 803 orig_offset = uio->uio_offset; 804 805 #ifdef MAC 806 error = mac_vnode_check_read(active_cred, fp->f_cred, vp); 807 if (error == 0) 808 #endif 809 error = VOP_READ(vp, uio, ioflag, fp->f_cred); 810 fp->f_nextoff = uio->uio_offset; 811 VOP_UNLOCK(vp, 0); 812 if (error == 0 && advice == POSIX_FADV_NOREUSE && 813 orig_offset != uio->uio_offset) 814 /* 815 * Use POSIX_FADV_DONTNEED to flush pages and buffers 816 * for the backing file after a POSIX_FADV_NOREUSE 817 * read(2). 818 */ 819 error = VOP_ADVISE(vp, orig_offset, uio->uio_offset - 1, 820 POSIX_FADV_DONTNEED); 821 return (error); 822 } 823 824 /* 825 * File table vnode write routine. 826 */ 827 static int 828 vn_write(struct file *fp, struct uio *uio, struct ucred *active_cred, int flags, 829 struct thread *td) 830 { 831 struct vnode *vp; 832 struct mount *mp; 833 off_t orig_offset; 834 int error, ioflag, lock_flags; 835 int advice; 836 837 KASSERT(uio->uio_td == td, ("uio_td %p is not td %p", 838 uio->uio_td, td)); 839 KASSERT(flags & FOF_OFFSET, ("No FOF_OFFSET")); 840 vp = fp->f_vnode; 841 if (vp->v_type == VREG) 842 bwillwrite(); 843 ioflag = IO_UNIT; 844 if (vp->v_type == VREG && (fp->f_flag & O_APPEND)) 845 ioflag |= IO_APPEND; 846 if (fp->f_flag & FNONBLOCK) 847 ioflag |= IO_NDELAY; 848 if (fp->f_flag & O_DIRECT) 849 ioflag |= IO_DIRECT; 850 if ((fp->f_flag & O_FSYNC) || 851 (vp->v_mount && (vp->v_mount->mnt_flag & MNT_SYNCHRONOUS))) 852 ioflag |= IO_SYNC; 853 mp = NULL; 854 if (vp->v_type != VCHR && 855 (error = vn_start_write(vp, &mp, V_WAIT | PCATCH)) != 0) 856 goto unlock; 857 858 advice = get_advice(fp, uio); 859 860 if (MNT_SHARED_WRITES(mp) || 861 (mp == NULL && MNT_SHARED_WRITES(vp->v_mount))) { 862 lock_flags = LK_SHARED; 863 } else { 864 lock_flags = LK_EXCLUSIVE; 865 } 866 867 vn_lock(vp, lock_flags | LK_RETRY); 868 switch (advice) { 869 case POSIX_FADV_NORMAL: 870 case POSIX_FADV_SEQUENTIAL: 871 case POSIX_FADV_NOREUSE: 872 ioflag |= sequential_heuristic(uio, fp); 873 break; 874 case POSIX_FADV_RANDOM: 875 /* XXX: Is this correct? */ 876 break; 877 } 878 orig_offset = uio->uio_offset; 879 880 #ifdef MAC 881 error = mac_vnode_check_write(active_cred, fp->f_cred, vp); 882 if (error == 0) 883 #endif 884 error = VOP_WRITE(vp, uio, ioflag, fp->f_cred); 885 fp->f_nextoff = uio->uio_offset; 886 VOP_UNLOCK(vp, 0); 887 if (vp->v_type != VCHR) 888 vn_finished_write(mp); 889 if (error == 0 && advice == POSIX_FADV_NOREUSE && 890 orig_offset != uio->uio_offset) 891 /* 892 * Use POSIX_FADV_DONTNEED to flush pages and buffers 893 * for the backing file after a POSIX_FADV_NOREUSE 894 * write(2). 895 */ 896 error = VOP_ADVISE(vp, orig_offset, uio->uio_offset - 1, 897 POSIX_FADV_DONTNEED); 898 unlock: 899 return (error); 900 } 901 902 /* 903 * The vn_io_fault() is a wrapper around vn_read() and vn_write() to 904 * prevent the following deadlock: 905 * 906 * Assume that the thread A reads from the vnode vp1 into userspace 907 * buffer buf1 backed by the pages of vnode vp2. If a page in buf1 is 908 * currently not resident, then system ends up with the call chain 909 * vn_read() -> VOP_READ(vp1) -> uiomove() -> [Page Fault] -> 910 * vm_fault(buf1) -> vnode_pager_getpages(vp2) -> VOP_GETPAGES(vp2) 911 * which establishes lock order vp1->vn_lock, then vp2->vn_lock. 912 * If, at the same time, thread B reads from vnode vp2 into buffer buf2 913 * backed by the pages of vnode vp1, and some page in buf2 is not 914 * resident, we get a reversed order vp2->vn_lock, then vp1->vn_lock. 915 * 916 * To prevent the lock order reversal and deadlock, vn_io_fault() does 917 * not allow page faults to happen during VOP_READ() or VOP_WRITE(). 918 * Instead, it first tries to do the whole range i/o with pagefaults 919 * disabled. If all pages in the i/o buffer are resident and mapped, 920 * VOP will succeed (ignoring the genuine filesystem errors). 921 * Otherwise, we get back EFAULT, and vn_io_fault() falls back to do 922 * i/o in chunks, with all pages in the chunk prefaulted and held 923 * using vm_fault_quick_hold_pages(). 924 * 925 * Filesystems using this deadlock avoidance scheme should use the 926 * array of the held pages from uio, saved in the curthread->td_ma, 927 * instead of doing uiomove(). A helper function 928 * vn_io_fault_uiomove() converts uiomove request into 929 * uiomove_fromphys() over td_ma array. 930 * 931 * Since vnode locks do not cover the whole i/o anymore, rangelocks 932 * make the current i/o request atomic with respect to other i/os and 933 * truncations. 934 */ 935 936 /* 937 * Decode vn_io_fault_args and perform the corresponding i/o. 938 */ 939 static int 940 vn_io_fault_doio(struct vn_io_fault_args *args, struct uio *uio, 941 struct thread *td) 942 { 943 int error, save; 944 945 error = 0; 946 save = vm_fault_disable_pagefaults(); 947 switch (args->kind) { 948 case VN_IO_FAULT_FOP: 949 error = (args->args.fop_args.doio)(args->args.fop_args.fp, 950 uio, args->cred, args->flags, td); 951 break; 952 case VN_IO_FAULT_VOP: 953 if (uio->uio_rw == UIO_READ) { 954 error = VOP_READ(args->args.vop_args.vp, uio, 955 args->flags, args->cred); 956 } else if (uio->uio_rw == UIO_WRITE) { 957 error = VOP_WRITE(args->args.vop_args.vp, uio, 958 args->flags, args->cred); 959 } 960 break; 961 default: 962 panic("vn_io_fault_doio: unknown kind of io %d %d", 963 args->kind, uio->uio_rw); 964 } 965 vm_fault_enable_pagefaults(save); 966 return (error); 967 } 968 969 static int 970 vn_io_fault_touch(char *base, const struct uio *uio) 971 { 972 int r; 973 974 r = fubyte(base); 975 if (r == -1 || (uio->uio_rw == UIO_READ && subyte(base, r) == -1)) 976 return (EFAULT); 977 return (0); 978 } 979 980 static int 981 vn_io_fault_prefault_user(const struct uio *uio) 982 { 983 char *base; 984 const struct iovec *iov; 985 size_t len; 986 ssize_t resid; 987 int error, i; 988 989 KASSERT(uio->uio_segflg == UIO_USERSPACE, 990 ("vn_io_fault_prefault userspace")); 991 992 error = i = 0; 993 iov = uio->uio_iov; 994 resid = uio->uio_resid; 995 base = iov->iov_base; 996 len = iov->iov_len; 997 while (resid > 0) { 998 error = vn_io_fault_touch(base, uio); 999 if (error != 0) 1000 break; 1001 if (len < PAGE_SIZE) { 1002 if (len != 0) { 1003 error = vn_io_fault_touch(base + len - 1, uio); 1004 if (error != 0) 1005 break; 1006 resid -= len; 1007 } 1008 if (++i >= uio->uio_iovcnt) 1009 break; 1010 iov = uio->uio_iov + i; 1011 base = iov->iov_base; 1012 len = iov->iov_len; 1013 } else { 1014 len -= PAGE_SIZE; 1015 base += PAGE_SIZE; 1016 resid -= PAGE_SIZE; 1017 } 1018 } 1019 return (error); 1020 } 1021 1022 /* 1023 * Common code for vn_io_fault(), agnostic to the kind of i/o request. 1024 * Uses vn_io_fault_doio() to make the call to an actual i/o function. 1025 * Used from vn_rdwr() and vn_io_fault(), which encode the i/o request 1026 * into args and call vn_io_fault1() to handle faults during the user 1027 * mode buffer accesses. 1028 */ 1029 static int 1030 vn_io_fault1(struct vnode *vp, struct uio *uio, struct vn_io_fault_args *args, 1031 struct thread *td) 1032 { 1033 vm_page_t ma[io_hold_cnt + 2]; 1034 struct uio *uio_clone, short_uio; 1035 struct iovec short_iovec[1]; 1036 vm_page_t *prev_td_ma; 1037 vm_prot_t prot; 1038 vm_offset_t addr, end; 1039 size_t len, resid; 1040 ssize_t adv; 1041 int error, cnt, saveheld, prev_td_ma_cnt; 1042 1043 if (vn_io_fault_prefault) { 1044 error = vn_io_fault_prefault_user(uio); 1045 if (error != 0) 1046 return (error); /* Or ignore ? */ 1047 } 1048 1049 prot = uio->uio_rw == UIO_READ ? VM_PROT_WRITE : VM_PROT_READ; 1050 1051 /* 1052 * The UFS follows IO_UNIT directive and replays back both 1053 * uio_offset and uio_resid if an error is encountered during the 1054 * operation. But, since the iovec may be already advanced, 1055 * uio is still in an inconsistent state. 1056 * 1057 * Cache a copy of the original uio, which is advanced to the redo 1058 * point using UIO_NOCOPY below. 1059 */ 1060 uio_clone = cloneuio(uio); 1061 resid = uio->uio_resid; 1062 1063 short_uio.uio_segflg = UIO_USERSPACE; 1064 short_uio.uio_rw = uio->uio_rw; 1065 short_uio.uio_td = uio->uio_td; 1066 1067 error = vn_io_fault_doio(args, uio, td); 1068 if (error != EFAULT) 1069 goto out; 1070 1071 atomic_add_long(&vn_io_faults_cnt, 1); 1072 uio_clone->uio_segflg = UIO_NOCOPY; 1073 uiomove(NULL, resid - uio->uio_resid, uio_clone); 1074 uio_clone->uio_segflg = uio->uio_segflg; 1075 1076 saveheld = curthread_pflags_set(TDP_UIOHELD); 1077 prev_td_ma = td->td_ma; 1078 prev_td_ma_cnt = td->td_ma_cnt; 1079 1080 while (uio_clone->uio_resid != 0) { 1081 len = uio_clone->uio_iov->iov_len; 1082 if (len == 0) { 1083 KASSERT(uio_clone->uio_iovcnt >= 1, 1084 ("iovcnt underflow")); 1085 uio_clone->uio_iov++; 1086 uio_clone->uio_iovcnt--; 1087 continue; 1088 } 1089 if (len > io_hold_cnt * PAGE_SIZE) 1090 len = io_hold_cnt * PAGE_SIZE; 1091 addr = (uintptr_t)uio_clone->uio_iov->iov_base; 1092 end = round_page(addr + len); 1093 if (end < addr) { 1094 error = EFAULT; 1095 break; 1096 } 1097 cnt = atop(end - trunc_page(addr)); 1098 /* 1099 * A perfectly misaligned address and length could cause 1100 * both the start and the end of the chunk to use partial 1101 * page. +2 accounts for such a situation. 1102 */ 1103 cnt = vm_fault_quick_hold_pages(&td->td_proc->p_vmspace->vm_map, 1104 addr, len, prot, ma, io_hold_cnt + 2); 1105 if (cnt == -1) { 1106 error = EFAULT; 1107 break; 1108 } 1109 short_uio.uio_iov = &short_iovec[0]; 1110 short_iovec[0].iov_base = (void *)addr; 1111 short_uio.uio_iovcnt = 1; 1112 short_uio.uio_resid = short_iovec[0].iov_len = len; 1113 short_uio.uio_offset = uio_clone->uio_offset; 1114 td->td_ma = ma; 1115 td->td_ma_cnt = cnt; 1116 1117 error = vn_io_fault_doio(args, &short_uio, td); 1118 vm_page_unhold_pages(ma, cnt); 1119 adv = len - short_uio.uio_resid; 1120 1121 uio_clone->uio_iov->iov_base = 1122 (char *)uio_clone->uio_iov->iov_base + adv; 1123 uio_clone->uio_iov->iov_len -= adv; 1124 uio_clone->uio_resid -= adv; 1125 uio_clone->uio_offset += adv; 1126 1127 uio->uio_resid -= adv; 1128 uio->uio_offset += adv; 1129 1130 if (error != 0 || adv == 0) 1131 break; 1132 } 1133 td->td_ma = prev_td_ma; 1134 td->td_ma_cnt = prev_td_ma_cnt; 1135 curthread_pflags_restore(saveheld); 1136 out: 1137 free(uio_clone, M_IOV); 1138 return (error); 1139 } 1140 1141 static int 1142 vn_io_fault(struct file *fp, struct uio *uio, struct ucred *active_cred, 1143 int flags, struct thread *td) 1144 { 1145 fo_rdwr_t *doio; 1146 struct vnode *vp; 1147 void *rl_cookie; 1148 struct vn_io_fault_args args; 1149 int error; 1150 1151 doio = uio->uio_rw == UIO_READ ? vn_read : vn_write; 1152 vp = fp->f_vnode; 1153 foffset_lock_uio(fp, uio, flags); 1154 if (do_vn_io_fault(vp, uio)) { 1155 args.kind = VN_IO_FAULT_FOP; 1156 args.args.fop_args.fp = fp; 1157 args.args.fop_args.doio = doio; 1158 args.cred = active_cred; 1159 args.flags = flags | FOF_OFFSET; 1160 if (uio->uio_rw == UIO_READ) { 1161 rl_cookie = vn_rangelock_rlock(vp, uio->uio_offset, 1162 uio->uio_offset + uio->uio_resid); 1163 } else if ((fp->f_flag & O_APPEND) != 0 || 1164 (flags & FOF_OFFSET) == 0) { 1165 /* For appenders, punt and lock the whole range. */ 1166 rl_cookie = vn_rangelock_wlock(vp, 0, OFF_MAX); 1167 } else { 1168 rl_cookie = vn_rangelock_wlock(vp, uio->uio_offset, 1169 uio->uio_offset + uio->uio_resid); 1170 } 1171 error = vn_io_fault1(vp, uio, &args, td); 1172 vn_rangelock_unlock(vp, rl_cookie); 1173 } else { 1174 error = doio(fp, uio, active_cred, flags | FOF_OFFSET, td); 1175 } 1176 foffset_unlock_uio(fp, uio, flags); 1177 return (error); 1178 } 1179 1180 /* 1181 * Helper function to perform the requested uiomove operation using 1182 * the held pages for io->uio_iov[0].iov_base buffer instead of 1183 * copyin/copyout. Access to the pages with uiomove_fromphys() 1184 * instead of iov_base prevents page faults that could occur due to 1185 * pmap_collect() invalidating the mapping created by 1186 * vm_fault_quick_hold_pages(), or pageout daemon, page laundry or 1187 * object cleanup revoking the write access from page mappings. 1188 * 1189 * Filesystems specified MNTK_NO_IOPF shall use vn_io_fault_uiomove() 1190 * instead of plain uiomove(). 1191 */ 1192 int 1193 vn_io_fault_uiomove(char *data, int xfersize, struct uio *uio) 1194 { 1195 struct uio transp_uio; 1196 struct iovec transp_iov[1]; 1197 struct thread *td; 1198 size_t adv; 1199 int error, pgadv; 1200 1201 td = curthread; 1202 if ((td->td_pflags & TDP_UIOHELD) == 0 || 1203 uio->uio_segflg != UIO_USERSPACE) 1204 return (uiomove(data, xfersize, uio)); 1205 1206 KASSERT(uio->uio_iovcnt == 1, ("uio_iovcnt %d", uio->uio_iovcnt)); 1207 transp_iov[0].iov_base = data; 1208 transp_uio.uio_iov = &transp_iov[0]; 1209 transp_uio.uio_iovcnt = 1; 1210 if (xfersize > uio->uio_resid) 1211 xfersize = uio->uio_resid; 1212 transp_uio.uio_resid = transp_iov[0].iov_len = xfersize; 1213 transp_uio.uio_offset = 0; 1214 transp_uio.uio_segflg = UIO_SYSSPACE; 1215 /* 1216 * Since transp_iov points to data, and td_ma page array 1217 * corresponds to original uio->uio_iov, we need to invert the 1218 * direction of the i/o operation as passed to 1219 * uiomove_fromphys(). 1220 */ 1221 switch (uio->uio_rw) { 1222 case UIO_WRITE: 1223 transp_uio.uio_rw = UIO_READ; 1224 break; 1225 case UIO_READ: 1226 transp_uio.uio_rw = UIO_WRITE; 1227 break; 1228 } 1229 transp_uio.uio_td = uio->uio_td; 1230 error = uiomove_fromphys(td->td_ma, 1231 ((vm_offset_t)uio->uio_iov->iov_base) & PAGE_MASK, 1232 xfersize, &transp_uio); 1233 adv = xfersize - transp_uio.uio_resid; 1234 pgadv = 1235 (((vm_offset_t)uio->uio_iov->iov_base + adv) >> PAGE_SHIFT) - 1236 (((vm_offset_t)uio->uio_iov->iov_base) >> PAGE_SHIFT); 1237 td->td_ma += pgadv; 1238 KASSERT(td->td_ma_cnt >= pgadv, ("consumed pages %d %d", td->td_ma_cnt, 1239 pgadv)); 1240 td->td_ma_cnt -= pgadv; 1241 uio->uio_iov->iov_base = (char *)uio->uio_iov->iov_base + adv; 1242 uio->uio_iov->iov_len -= adv; 1243 uio->uio_resid -= adv; 1244 uio->uio_offset += adv; 1245 return (error); 1246 } 1247 1248 int 1249 vn_io_fault_pgmove(vm_page_t ma[], vm_offset_t offset, int xfersize, 1250 struct uio *uio) 1251 { 1252 struct thread *td; 1253 vm_offset_t iov_base; 1254 int cnt, pgadv; 1255 1256 td = curthread; 1257 if ((td->td_pflags & TDP_UIOHELD) == 0 || 1258 uio->uio_segflg != UIO_USERSPACE) 1259 return (uiomove_fromphys(ma, offset, xfersize, uio)); 1260 1261 KASSERT(uio->uio_iovcnt == 1, ("uio_iovcnt %d", uio->uio_iovcnt)); 1262 cnt = xfersize > uio->uio_resid ? uio->uio_resid : xfersize; 1263 iov_base = (vm_offset_t)uio->uio_iov->iov_base; 1264 switch (uio->uio_rw) { 1265 case UIO_WRITE: 1266 pmap_copy_pages(td->td_ma, iov_base & PAGE_MASK, ma, 1267 offset, cnt); 1268 break; 1269 case UIO_READ: 1270 pmap_copy_pages(ma, offset, td->td_ma, iov_base & PAGE_MASK, 1271 cnt); 1272 break; 1273 } 1274 pgadv = ((iov_base + cnt) >> PAGE_SHIFT) - (iov_base >> PAGE_SHIFT); 1275 td->td_ma += pgadv; 1276 KASSERT(td->td_ma_cnt >= pgadv, ("consumed pages %d %d", td->td_ma_cnt, 1277 pgadv)); 1278 td->td_ma_cnt -= pgadv; 1279 uio->uio_iov->iov_base = (char *)(iov_base + cnt); 1280 uio->uio_iov->iov_len -= cnt; 1281 uio->uio_resid -= cnt; 1282 uio->uio_offset += cnt; 1283 return (0); 1284 } 1285 1286 1287 /* 1288 * File table truncate routine. 1289 */ 1290 static int 1291 vn_truncate(struct file *fp, off_t length, struct ucred *active_cred, 1292 struct thread *td) 1293 { 1294 struct vattr vattr; 1295 struct mount *mp; 1296 struct vnode *vp; 1297 void *rl_cookie; 1298 int error; 1299 1300 vp = fp->f_vnode; 1301 1302 /* 1303 * Lock the whole range for truncation. Otherwise split i/o 1304 * might happen partly before and partly after the truncation. 1305 */ 1306 rl_cookie = vn_rangelock_wlock(vp, 0, OFF_MAX); 1307 error = vn_start_write(vp, &mp, V_WAIT | PCATCH); 1308 if (error) 1309 goto out1; 1310 vn_lock(vp, LK_EXCLUSIVE | LK_RETRY); 1311 AUDIT_ARG_VNODE1(vp); 1312 if (vp->v_type == VDIR) { 1313 error = EISDIR; 1314 goto out; 1315 } 1316 #ifdef MAC 1317 error = mac_vnode_check_write(active_cred, fp->f_cred, vp); 1318 if (error) 1319 goto out; 1320 #endif 1321 error = VOP_ADD_WRITECOUNT(vp, 1); 1322 if (error == 0) { 1323 VATTR_NULL(&vattr); 1324 vattr.va_size = length; 1325 if ((fp->f_flag & O_FSYNC) != 0) 1326 vattr.va_vaflags |= VA_SYNC; 1327 error = VOP_SETATTR(vp, &vattr, fp->f_cred); 1328 VOP_ADD_WRITECOUNT_CHECKED(vp, -1); 1329 } 1330 out: 1331 VOP_UNLOCK(vp, 0); 1332 vn_finished_write(mp); 1333 out1: 1334 vn_rangelock_unlock(vp, rl_cookie); 1335 return (error); 1336 } 1337 1338 /* 1339 * File table vnode stat routine. 1340 */ 1341 static int 1342 vn_statfile(struct file *fp, struct stat *sb, struct ucred *active_cred, 1343 struct thread *td) 1344 { 1345 struct vnode *vp = fp->f_vnode; 1346 int error; 1347 1348 vn_lock(vp, LK_SHARED | LK_RETRY); 1349 error = vn_stat(vp, sb, active_cred, fp->f_cred, td); 1350 VOP_UNLOCK(vp, 0); 1351 1352 return (error); 1353 } 1354 1355 /* 1356 * Stat a vnode; implementation for the stat syscall 1357 */ 1358 int 1359 vn_stat(struct vnode *vp, struct stat *sb, struct ucred *active_cred, 1360 struct ucred *file_cred, struct thread *td) 1361 { 1362 struct vattr vattr; 1363 struct vattr *vap; 1364 int error; 1365 u_short mode; 1366 1367 AUDIT_ARG_VNODE1(vp); 1368 #ifdef MAC 1369 error = mac_vnode_check_stat(active_cred, file_cred, vp); 1370 if (error) 1371 return (error); 1372 #endif 1373 1374 vap = &vattr; 1375 1376 /* 1377 * Initialize defaults for new and unusual fields, so that file 1378 * systems which don't support these fields don't need to know 1379 * about them. 1380 */ 1381 vap->va_birthtime.tv_sec = -1; 1382 vap->va_birthtime.tv_nsec = 0; 1383 vap->va_fsid = VNOVAL; 1384 vap->va_rdev = NODEV; 1385 1386 error = VOP_GETATTR(vp, vap, active_cred); 1387 if (error) 1388 return (error); 1389 1390 /* 1391 * Zero the spare stat fields 1392 */ 1393 bzero(sb, sizeof *sb); 1394 1395 /* 1396 * Copy from vattr table 1397 */ 1398 if (vap->va_fsid != VNOVAL) 1399 sb->st_dev = vap->va_fsid; 1400 else 1401 sb->st_dev = vp->v_mount->mnt_stat.f_fsid.val[0]; 1402 sb->st_ino = vap->va_fileid; 1403 mode = vap->va_mode; 1404 switch (vap->va_type) { 1405 case VREG: 1406 mode |= S_IFREG; 1407 break; 1408 case VDIR: 1409 mode |= S_IFDIR; 1410 break; 1411 case VBLK: 1412 mode |= S_IFBLK; 1413 break; 1414 case VCHR: 1415 mode |= S_IFCHR; 1416 break; 1417 case VLNK: 1418 mode |= S_IFLNK; 1419 break; 1420 case VSOCK: 1421 mode |= S_IFSOCK; 1422 break; 1423 case VFIFO: 1424 mode |= S_IFIFO; 1425 break; 1426 default: 1427 return (EBADF); 1428 } 1429 sb->st_mode = mode; 1430 sb->st_nlink = vap->va_nlink; 1431 sb->st_uid = vap->va_uid; 1432 sb->st_gid = vap->va_gid; 1433 sb->st_rdev = vap->va_rdev; 1434 if (vap->va_size > OFF_MAX) 1435 return (EOVERFLOW); 1436 sb->st_size = vap->va_size; 1437 sb->st_atim = vap->va_atime; 1438 sb->st_mtim = vap->va_mtime; 1439 sb->st_ctim = vap->va_ctime; 1440 sb->st_birthtim = vap->va_birthtime; 1441 1442 /* 1443 * According to www.opengroup.org, the meaning of st_blksize is 1444 * "a filesystem-specific preferred I/O block size for this 1445 * object. In some filesystem types, this may vary from file 1446 * to file" 1447 * Use miminum/default of PAGE_SIZE (e.g. for VCHR). 1448 */ 1449 1450 sb->st_blksize = max(PAGE_SIZE, vap->va_blocksize); 1451 1452 sb->st_flags = vap->va_flags; 1453 if (priv_check(td, PRIV_VFS_GENERATION)) 1454 sb->st_gen = 0; 1455 else 1456 sb->st_gen = vap->va_gen; 1457 1458 sb->st_blocks = vap->va_bytes / S_BLKSIZE; 1459 return (0); 1460 } 1461 1462 /* generic FIOBMAP2 implementation */ 1463 static int 1464 vn_ioc_bmap2(struct file *fp, struct fiobmap2_arg *arg, struct ucred *cred) 1465 { 1466 struct vnode *vp = fp->f_vnode; 1467 daddr_t lbn = arg->bn; 1468 int error; 1469 1470 vn_lock(vp, LK_SHARED | LK_RETRY); 1471 #ifdef MAC 1472 error = mac_vnode_check_read(cred, fp->f_cred, vp); 1473 if (error == 0) 1474 #endif 1475 error = VOP_BMAP(vp, lbn, NULL, &arg->bn, &arg->runp, 1476 &arg->runb); 1477 VOP_UNLOCK(vp, 0); 1478 return (error); 1479 } 1480 1481 /* 1482 * File table vnode ioctl routine. 1483 */ 1484 static int 1485 vn_ioctl(struct file *fp, u_long com, void *data, struct ucred *active_cred, 1486 struct thread *td) 1487 { 1488 struct vattr vattr; 1489 struct vnode *vp; 1490 int error; 1491 1492 vp = fp->f_vnode; 1493 switch (vp->v_type) { 1494 case VDIR: 1495 case VREG: 1496 switch (com) { 1497 case FIONREAD: 1498 vn_lock(vp, LK_SHARED | LK_RETRY); 1499 error = VOP_GETATTR(vp, &vattr, active_cred); 1500 VOP_UNLOCK(vp, 0); 1501 if (error == 0) 1502 *(int *)data = vattr.va_size - fp->f_offset; 1503 return (error); 1504 case FIOBMAP2: 1505 return (vn_ioc_bmap2(fp, (struct fiobmap2_arg*)data, 1506 active_cred)); 1507 case FIONBIO: 1508 case FIOASYNC: 1509 return (0); 1510 default: 1511 return (VOP_IOCTL(vp, com, data, fp->f_flag, 1512 active_cred, td)); 1513 } 1514 break; 1515 case VCHR: 1516 return (VOP_IOCTL(vp, com, data, fp->f_flag, 1517 active_cred, td)); 1518 default: 1519 return (ENOTTY); 1520 } 1521 } 1522 1523 /* 1524 * File table vnode poll routine. 1525 */ 1526 static int 1527 vn_poll(struct file *fp, int events, struct ucred *active_cred, 1528 struct thread *td) 1529 { 1530 struct vnode *vp; 1531 int error; 1532 1533 vp = fp->f_vnode; 1534 #ifdef MAC 1535 vn_lock(vp, LK_EXCLUSIVE | LK_RETRY); 1536 AUDIT_ARG_VNODE1(vp); 1537 error = mac_vnode_check_poll(active_cred, fp->f_cred, vp); 1538 VOP_UNLOCK(vp, 0); 1539 if (!error) 1540 #endif 1541 1542 error = VOP_POLL(vp, events, fp->f_cred, td); 1543 return (error); 1544 } 1545 1546 /* 1547 * Acquire the requested lock and then check for validity. LK_RETRY 1548 * permits vn_lock to return doomed vnodes. 1549 */ 1550 int 1551 _vn_lock(struct vnode *vp, int flags, char *file, int line) 1552 { 1553 int error; 1554 1555 VNASSERT((flags & LK_TYPE_MASK) != 0, vp, 1556 ("vn_lock: no locktype")); 1557 VNASSERT(vp->v_holdcnt != 0, vp, ("vn_lock: zero hold count")); 1558 retry: 1559 error = VOP_LOCK1(vp, flags, file, line); 1560 flags &= ~LK_INTERLOCK; /* Interlock is always dropped. */ 1561 KASSERT((flags & LK_RETRY) == 0 || error == 0, 1562 ("vn_lock: error %d incompatible with flags %#x", error, flags)); 1563 1564 if ((flags & LK_RETRY) == 0) { 1565 if (error == 0 && (vp->v_iflag & VI_DOOMED) != 0) { 1566 VOP_UNLOCK(vp, 0); 1567 error = ENOENT; 1568 } 1569 } else if (error != 0) 1570 goto retry; 1571 return (error); 1572 } 1573 1574 /* 1575 * File table vnode close routine. 1576 */ 1577 static int 1578 vn_closefile(struct file *fp, struct thread *td) 1579 { 1580 struct vnode *vp; 1581 struct flock lf; 1582 int error; 1583 bool ref; 1584 1585 vp = fp->f_vnode; 1586 fp->f_ops = &badfileops; 1587 ref= (fp->f_flag & FHASLOCK) != 0 && fp->f_type == DTYPE_VNODE; 1588 1589 error = vn_close1(vp, fp->f_flag, fp->f_cred, td, ref); 1590 1591 if (__predict_false(ref)) { 1592 lf.l_whence = SEEK_SET; 1593 lf.l_start = 0; 1594 lf.l_len = 0; 1595 lf.l_type = F_UNLCK; 1596 (void) VOP_ADVLOCK(vp, fp, F_UNLCK, &lf, F_FLOCK); 1597 vrele(vp); 1598 } 1599 return (error); 1600 } 1601 1602 static bool 1603 vn_suspendable(struct mount *mp) 1604 { 1605 1606 return (mp->mnt_op->vfs_susp_clean != NULL); 1607 } 1608 1609 /* 1610 * Preparing to start a filesystem write operation. If the operation is 1611 * permitted, then we bump the count of operations in progress and 1612 * proceed. If a suspend request is in progress, we wait until the 1613 * suspension is over, and then proceed. 1614 */ 1615 static int 1616 vn_start_write_locked(struct mount *mp, int flags) 1617 { 1618 int error, mflags; 1619 1620 mtx_assert(MNT_MTX(mp), MA_OWNED); 1621 error = 0; 1622 1623 /* 1624 * Check on status of suspension. 1625 */ 1626 if ((curthread->td_pflags & TDP_IGNSUSP) == 0 || 1627 mp->mnt_susp_owner != curthread) { 1628 mflags = ((mp->mnt_vfc->vfc_flags & VFCF_SBDRY) != 0 ? 1629 (flags & PCATCH) : 0) | (PUSER - 1); 1630 while ((mp->mnt_kern_flag & MNTK_SUSPEND) != 0) { 1631 if (flags & V_NOWAIT) { 1632 error = EWOULDBLOCK; 1633 goto unlock; 1634 } 1635 error = msleep(&mp->mnt_flag, MNT_MTX(mp), mflags, 1636 "suspfs", 0); 1637 if (error) 1638 goto unlock; 1639 } 1640 } 1641 if (flags & V_XSLEEP) 1642 goto unlock; 1643 mp->mnt_writeopcount++; 1644 unlock: 1645 if (error != 0 || (flags & V_XSLEEP) != 0) 1646 MNT_REL(mp); 1647 MNT_IUNLOCK(mp); 1648 return (error); 1649 } 1650 1651 int 1652 vn_start_write(struct vnode *vp, struct mount **mpp, int flags) 1653 { 1654 struct mount *mp; 1655 int error; 1656 1657 KASSERT((flags & V_MNTREF) == 0 || (*mpp != NULL && vp == NULL), 1658 ("V_MNTREF requires mp")); 1659 1660 error = 0; 1661 /* 1662 * If a vnode is provided, get and return the mount point that 1663 * to which it will write. 1664 */ 1665 if (vp != NULL) { 1666 if ((error = VOP_GETWRITEMOUNT(vp, mpp)) != 0) { 1667 *mpp = NULL; 1668 if (error != EOPNOTSUPP) 1669 return (error); 1670 return (0); 1671 } 1672 } 1673 if ((mp = *mpp) == NULL) 1674 return (0); 1675 1676 if (!vn_suspendable(mp)) { 1677 if (vp != NULL || (flags & V_MNTREF) != 0) 1678 vfs_rel(mp); 1679 return (0); 1680 } 1681 1682 /* 1683 * VOP_GETWRITEMOUNT() returns with the mp refcount held through 1684 * a vfs_ref(). 1685 * As long as a vnode is not provided we need to acquire a 1686 * refcount for the provided mountpoint too, in order to 1687 * emulate a vfs_ref(). 1688 */ 1689 MNT_ILOCK(mp); 1690 if (vp == NULL && (flags & V_MNTREF) == 0) 1691 MNT_REF(mp); 1692 1693 return (vn_start_write_locked(mp, flags)); 1694 } 1695 1696 /* 1697 * Secondary suspension. Used by operations such as vop_inactive 1698 * routines that are needed by the higher level functions. These 1699 * are allowed to proceed until all the higher level functions have 1700 * completed (indicated by mnt_writeopcount dropping to zero). At that 1701 * time, these operations are halted until the suspension is over. 1702 */ 1703 int 1704 vn_start_secondary_write(struct vnode *vp, struct mount **mpp, int flags) 1705 { 1706 struct mount *mp; 1707 int error; 1708 1709 KASSERT((flags & V_MNTREF) == 0 || (*mpp != NULL && vp == NULL), 1710 ("V_MNTREF requires mp")); 1711 1712 retry: 1713 if (vp != NULL) { 1714 if ((error = VOP_GETWRITEMOUNT(vp, mpp)) != 0) { 1715 *mpp = NULL; 1716 if (error != EOPNOTSUPP) 1717 return (error); 1718 return (0); 1719 } 1720 } 1721 /* 1722 * If we are not suspended or have not yet reached suspended 1723 * mode, then let the operation proceed. 1724 */ 1725 if ((mp = *mpp) == NULL) 1726 return (0); 1727 1728 if (!vn_suspendable(mp)) { 1729 if (vp != NULL || (flags & V_MNTREF) != 0) 1730 vfs_rel(mp); 1731 return (0); 1732 } 1733 1734 /* 1735 * VOP_GETWRITEMOUNT() returns with the mp refcount held through 1736 * a vfs_ref(). 1737 * As long as a vnode is not provided we need to acquire a 1738 * refcount for the provided mountpoint too, in order to 1739 * emulate a vfs_ref(). 1740 */ 1741 MNT_ILOCK(mp); 1742 if (vp == NULL && (flags & V_MNTREF) == 0) 1743 MNT_REF(mp); 1744 if ((mp->mnt_kern_flag & (MNTK_SUSPENDED | MNTK_SUSPEND2)) == 0) { 1745 mp->mnt_secondary_writes++; 1746 mp->mnt_secondary_accwrites++; 1747 MNT_IUNLOCK(mp); 1748 return (0); 1749 } 1750 if (flags & V_NOWAIT) { 1751 MNT_REL(mp); 1752 MNT_IUNLOCK(mp); 1753 return (EWOULDBLOCK); 1754 } 1755 /* 1756 * Wait for the suspension to finish. 1757 */ 1758 error = msleep(&mp->mnt_flag, MNT_MTX(mp), (PUSER - 1) | PDROP | 1759 ((mp->mnt_vfc->vfc_flags & VFCF_SBDRY) != 0 ? (flags & PCATCH) : 0), 1760 "suspfs", 0); 1761 vfs_rel(mp); 1762 if (error == 0) 1763 goto retry; 1764 return (error); 1765 } 1766 1767 /* 1768 * Filesystem write operation has completed. If we are suspending and this 1769 * operation is the last one, notify the suspender that the suspension is 1770 * now in effect. 1771 */ 1772 void 1773 vn_finished_write(struct mount *mp) 1774 { 1775 if (mp == NULL || !vn_suspendable(mp)) 1776 return; 1777 MNT_ILOCK(mp); 1778 MNT_REL(mp); 1779 mp->mnt_writeopcount--; 1780 if (mp->mnt_writeopcount < 0) 1781 panic("vn_finished_write: neg cnt"); 1782 if ((mp->mnt_kern_flag & MNTK_SUSPEND) != 0 && 1783 mp->mnt_writeopcount <= 0) 1784 wakeup(&mp->mnt_writeopcount); 1785 MNT_IUNLOCK(mp); 1786 } 1787 1788 1789 /* 1790 * Filesystem secondary write operation has completed. If we are 1791 * suspending and this operation is the last one, notify the suspender 1792 * that the suspension is now in effect. 1793 */ 1794 void 1795 vn_finished_secondary_write(struct mount *mp) 1796 { 1797 if (mp == NULL || !vn_suspendable(mp)) 1798 return; 1799 MNT_ILOCK(mp); 1800 MNT_REL(mp); 1801 mp->mnt_secondary_writes--; 1802 if (mp->mnt_secondary_writes < 0) 1803 panic("vn_finished_secondary_write: neg cnt"); 1804 if ((mp->mnt_kern_flag & MNTK_SUSPEND) != 0 && 1805 mp->mnt_secondary_writes <= 0) 1806 wakeup(&mp->mnt_secondary_writes); 1807 MNT_IUNLOCK(mp); 1808 } 1809 1810 1811 1812 /* 1813 * Request a filesystem to suspend write operations. 1814 */ 1815 int 1816 vfs_write_suspend(struct mount *mp, int flags) 1817 { 1818 int error; 1819 1820 MPASS(vn_suspendable(mp)); 1821 1822 MNT_ILOCK(mp); 1823 if (mp->mnt_susp_owner == curthread) { 1824 MNT_IUNLOCK(mp); 1825 return (EALREADY); 1826 } 1827 while (mp->mnt_kern_flag & MNTK_SUSPEND) 1828 msleep(&mp->mnt_flag, MNT_MTX(mp), PUSER - 1, "wsuspfs", 0); 1829 1830 /* 1831 * Unmount holds a write reference on the mount point. If we 1832 * own busy reference and drain for writers, we deadlock with 1833 * the reference draining in the unmount path. Callers of 1834 * vfs_write_suspend() must specify VS_SKIP_UNMOUNT if 1835 * vfs_busy() reference is owned and caller is not in the 1836 * unmount context. 1837 */ 1838 if ((flags & VS_SKIP_UNMOUNT) != 0 && 1839 (mp->mnt_kern_flag & MNTK_UNMOUNT) != 0) { 1840 MNT_IUNLOCK(mp); 1841 return (EBUSY); 1842 } 1843 1844 mp->mnt_kern_flag |= MNTK_SUSPEND; 1845 mp->mnt_susp_owner = curthread; 1846 if (mp->mnt_writeopcount > 0) 1847 (void) msleep(&mp->mnt_writeopcount, 1848 MNT_MTX(mp), (PUSER - 1)|PDROP, "suspwt", 0); 1849 else 1850 MNT_IUNLOCK(mp); 1851 if ((error = VFS_SYNC(mp, MNT_SUSPEND)) != 0) 1852 vfs_write_resume(mp, 0); 1853 return (error); 1854 } 1855 1856 /* 1857 * Request a filesystem to resume write operations. 1858 */ 1859 void 1860 vfs_write_resume(struct mount *mp, int flags) 1861 { 1862 1863 MPASS(vn_suspendable(mp)); 1864 1865 MNT_ILOCK(mp); 1866 if ((mp->mnt_kern_flag & MNTK_SUSPEND) != 0) { 1867 KASSERT(mp->mnt_susp_owner == curthread, ("mnt_susp_owner")); 1868 mp->mnt_kern_flag &= ~(MNTK_SUSPEND | MNTK_SUSPEND2 | 1869 MNTK_SUSPENDED); 1870 mp->mnt_susp_owner = NULL; 1871 wakeup(&mp->mnt_writeopcount); 1872 wakeup(&mp->mnt_flag); 1873 curthread->td_pflags &= ~TDP_IGNSUSP; 1874 if ((flags & VR_START_WRITE) != 0) { 1875 MNT_REF(mp); 1876 mp->mnt_writeopcount++; 1877 } 1878 MNT_IUNLOCK(mp); 1879 if ((flags & VR_NO_SUSPCLR) == 0) 1880 VFS_SUSP_CLEAN(mp); 1881 } else if ((flags & VR_START_WRITE) != 0) { 1882 MNT_REF(mp); 1883 vn_start_write_locked(mp, 0); 1884 } else { 1885 MNT_IUNLOCK(mp); 1886 } 1887 } 1888 1889 /* 1890 * Helper loop around vfs_write_suspend() for filesystem unmount VFS 1891 * methods. 1892 */ 1893 int 1894 vfs_write_suspend_umnt(struct mount *mp) 1895 { 1896 int error; 1897 1898 MPASS(vn_suspendable(mp)); 1899 KASSERT((curthread->td_pflags & TDP_IGNSUSP) == 0, 1900 ("vfs_write_suspend_umnt: recursed")); 1901 1902 /* dounmount() already called vn_start_write(). */ 1903 for (;;) { 1904 vn_finished_write(mp); 1905 error = vfs_write_suspend(mp, 0); 1906 if (error != 0) { 1907 vn_start_write(NULL, &mp, V_WAIT); 1908 return (error); 1909 } 1910 MNT_ILOCK(mp); 1911 if ((mp->mnt_kern_flag & MNTK_SUSPENDED) != 0) 1912 break; 1913 MNT_IUNLOCK(mp); 1914 vn_start_write(NULL, &mp, V_WAIT); 1915 } 1916 mp->mnt_kern_flag &= ~(MNTK_SUSPENDED | MNTK_SUSPEND2); 1917 wakeup(&mp->mnt_flag); 1918 MNT_IUNLOCK(mp); 1919 curthread->td_pflags |= TDP_IGNSUSP; 1920 return (0); 1921 } 1922 1923 /* 1924 * Implement kqueues for files by translating it to vnode operation. 1925 */ 1926 static int 1927 vn_kqfilter(struct file *fp, struct knote *kn) 1928 { 1929 1930 return (VOP_KQFILTER(fp->f_vnode, kn)); 1931 } 1932 1933 /* 1934 * Simplified in-kernel wrapper calls for extended attribute access. 1935 * Both calls pass in a NULL credential, authorizing as "kernel" access. 1936 * Set IO_NODELOCKED in ioflg if the vnode is already locked. 1937 */ 1938 int 1939 vn_extattr_get(struct vnode *vp, int ioflg, int attrnamespace, 1940 const char *attrname, int *buflen, char *buf, struct thread *td) 1941 { 1942 struct uio auio; 1943 struct iovec iov; 1944 int error; 1945 1946 iov.iov_len = *buflen; 1947 iov.iov_base = buf; 1948 1949 auio.uio_iov = &iov; 1950 auio.uio_iovcnt = 1; 1951 auio.uio_rw = UIO_READ; 1952 auio.uio_segflg = UIO_SYSSPACE; 1953 auio.uio_td = td; 1954 auio.uio_offset = 0; 1955 auio.uio_resid = *buflen; 1956 1957 if ((ioflg & IO_NODELOCKED) == 0) 1958 vn_lock(vp, LK_SHARED | LK_RETRY); 1959 1960 ASSERT_VOP_LOCKED(vp, "IO_NODELOCKED with no vp lock held"); 1961 1962 /* authorize attribute retrieval as kernel */ 1963 error = VOP_GETEXTATTR(vp, attrnamespace, attrname, &auio, NULL, NULL, 1964 td); 1965 1966 if ((ioflg & IO_NODELOCKED) == 0) 1967 VOP_UNLOCK(vp, 0); 1968 1969 if (error == 0) { 1970 *buflen = *buflen - auio.uio_resid; 1971 } 1972 1973 return (error); 1974 } 1975 1976 /* 1977 * XXX failure mode if partially written? 1978 */ 1979 int 1980 vn_extattr_set(struct vnode *vp, int ioflg, int attrnamespace, 1981 const char *attrname, int buflen, char *buf, struct thread *td) 1982 { 1983 struct uio auio; 1984 struct iovec iov; 1985 struct mount *mp; 1986 int error; 1987 1988 iov.iov_len = buflen; 1989 iov.iov_base = buf; 1990 1991 auio.uio_iov = &iov; 1992 auio.uio_iovcnt = 1; 1993 auio.uio_rw = UIO_WRITE; 1994 auio.uio_segflg = UIO_SYSSPACE; 1995 auio.uio_td = td; 1996 auio.uio_offset = 0; 1997 auio.uio_resid = buflen; 1998 1999 if ((ioflg & IO_NODELOCKED) == 0) { 2000 if ((error = vn_start_write(vp, &mp, V_WAIT)) != 0) 2001 return (error); 2002 vn_lock(vp, LK_EXCLUSIVE | LK_RETRY); 2003 } 2004 2005 ASSERT_VOP_LOCKED(vp, "IO_NODELOCKED with no vp lock held"); 2006 2007 /* authorize attribute setting as kernel */ 2008 error = VOP_SETEXTATTR(vp, attrnamespace, attrname, &auio, NULL, td); 2009 2010 if ((ioflg & IO_NODELOCKED) == 0) { 2011 vn_finished_write(mp); 2012 VOP_UNLOCK(vp, 0); 2013 } 2014 2015 return (error); 2016 } 2017 2018 int 2019 vn_extattr_rm(struct vnode *vp, int ioflg, int attrnamespace, 2020 const char *attrname, struct thread *td) 2021 { 2022 struct mount *mp; 2023 int error; 2024 2025 if ((ioflg & IO_NODELOCKED) == 0) { 2026 if ((error = vn_start_write(vp, &mp, V_WAIT)) != 0) 2027 return (error); 2028 vn_lock(vp, LK_EXCLUSIVE | LK_RETRY); 2029 } 2030 2031 ASSERT_VOP_LOCKED(vp, "IO_NODELOCKED with no vp lock held"); 2032 2033 /* authorize attribute removal as kernel */ 2034 error = VOP_DELETEEXTATTR(vp, attrnamespace, attrname, NULL, td); 2035 if (error == EOPNOTSUPP) 2036 error = VOP_SETEXTATTR(vp, attrnamespace, attrname, NULL, 2037 NULL, td); 2038 2039 if ((ioflg & IO_NODELOCKED) == 0) { 2040 vn_finished_write(mp); 2041 VOP_UNLOCK(vp, 0); 2042 } 2043 2044 return (error); 2045 } 2046 2047 static int 2048 vn_get_ino_alloc_vget(struct mount *mp, void *arg, int lkflags, 2049 struct vnode **rvp) 2050 { 2051 2052 return (VFS_VGET(mp, *(ino_t *)arg, lkflags, rvp)); 2053 } 2054 2055 int 2056 vn_vget_ino(struct vnode *vp, ino_t ino, int lkflags, struct vnode **rvp) 2057 { 2058 2059 return (vn_vget_ino_gen(vp, vn_get_ino_alloc_vget, &ino, 2060 lkflags, rvp)); 2061 } 2062 2063 int 2064 vn_vget_ino_gen(struct vnode *vp, vn_get_ino_t alloc, void *alloc_arg, 2065 int lkflags, struct vnode **rvp) 2066 { 2067 struct mount *mp; 2068 int ltype, error; 2069 2070 ASSERT_VOP_LOCKED(vp, "vn_vget_ino_get"); 2071 mp = vp->v_mount; 2072 ltype = VOP_ISLOCKED(vp); 2073 KASSERT(ltype == LK_EXCLUSIVE || ltype == LK_SHARED, 2074 ("vn_vget_ino: vp not locked")); 2075 error = vfs_busy(mp, MBF_NOWAIT); 2076 if (error != 0) { 2077 vfs_ref(mp); 2078 VOP_UNLOCK(vp, 0); 2079 error = vfs_busy(mp, 0); 2080 vn_lock(vp, ltype | LK_RETRY); 2081 vfs_rel(mp); 2082 if (error != 0) 2083 return (ENOENT); 2084 if (vp->v_iflag & VI_DOOMED) { 2085 vfs_unbusy(mp); 2086 return (ENOENT); 2087 } 2088 } 2089 VOP_UNLOCK(vp, 0); 2090 error = alloc(mp, alloc_arg, lkflags, rvp); 2091 vfs_unbusy(mp); 2092 if (*rvp != vp) 2093 vn_lock(vp, ltype | LK_RETRY); 2094 if (vp->v_iflag & VI_DOOMED) { 2095 if (error == 0) { 2096 if (*rvp == vp) 2097 vunref(vp); 2098 else 2099 vput(*rvp); 2100 } 2101 error = ENOENT; 2102 } 2103 return (error); 2104 } 2105 2106 int 2107 vn_rlimit_fsize(const struct vnode *vp, const struct uio *uio, 2108 struct thread *td) 2109 { 2110 2111 if (vp->v_type != VREG || td == NULL) 2112 return (0); 2113 if ((uoff_t)uio->uio_offset + uio->uio_resid > 2114 lim_cur(td, RLIMIT_FSIZE)) { 2115 PROC_LOCK(td->td_proc); 2116 kern_psignal(td->td_proc, SIGXFSZ); 2117 PROC_UNLOCK(td->td_proc); 2118 return (EFBIG); 2119 } 2120 return (0); 2121 } 2122 2123 int 2124 vn_chmod(struct file *fp, mode_t mode, struct ucred *active_cred, 2125 struct thread *td) 2126 { 2127 struct vnode *vp; 2128 2129 vp = fp->f_vnode; 2130 #ifdef AUDIT 2131 vn_lock(vp, LK_SHARED | LK_RETRY); 2132 AUDIT_ARG_VNODE1(vp); 2133 VOP_UNLOCK(vp, 0); 2134 #endif 2135 return (setfmode(td, active_cred, vp, mode)); 2136 } 2137 2138 int 2139 vn_chown(struct file *fp, uid_t uid, gid_t gid, struct ucred *active_cred, 2140 struct thread *td) 2141 { 2142 struct vnode *vp; 2143 2144 vp = fp->f_vnode; 2145 #ifdef AUDIT 2146 vn_lock(vp, LK_SHARED | LK_RETRY); 2147 AUDIT_ARG_VNODE1(vp); 2148 VOP_UNLOCK(vp, 0); 2149 #endif 2150 return (setfown(td, active_cred, vp, uid, gid)); 2151 } 2152 2153 void 2154 vn_pages_remove(struct vnode *vp, vm_pindex_t start, vm_pindex_t end) 2155 { 2156 vm_object_t object; 2157 2158 if ((object = vp->v_object) == NULL) 2159 return; 2160 VM_OBJECT_WLOCK(object); 2161 vm_object_page_remove(object, start, end, 0); 2162 VM_OBJECT_WUNLOCK(object); 2163 } 2164 2165 int 2166 vn_bmap_seekhole(struct vnode *vp, u_long cmd, off_t *off, struct ucred *cred) 2167 { 2168 struct vattr va; 2169 daddr_t bn, bnp; 2170 uint64_t bsize; 2171 off_t noff; 2172 int error; 2173 2174 KASSERT(cmd == FIOSEEKHOLE || cmd == FIOSEEKDATA, 2175 ("Wrong command %lu", cmd)); 2176 2177 if (vn_lock(vp, LK_SHARED) != 0) 2178 return (EBADF); 2179 if (vp->v_type != VREG) { 2180 error = ENOTTY; 2181 goto unlock; 2182 } 2183 error = VOP_GETATTR(vp, &va, cred); 2184 if (error != 0) 2185 goto unlock; 2186 noff = *off; 2187 if (noff >= va.va_size) { 2188 error = ENXIO; 2189 goto unlock; 2190 } 2191 bsize = vp->v_mount->mnt_stat.f_iosize; 2192 for (bn = noff / bsize; noff < va.va_size; bn++, noff += bsize - 2193 noff % bsize) { 2194 error = VOP_BMAP(vp, bn, NULL, &bnp, NULL, NULL); 2195 if (error == EOPNOTSUPP) { 2196 error = ENOTTY; 2197 goto unlock; 2198 } 2199 if ((bnp == -1 && cmd == FIOSEEKHOLE) || 2200 (bnp != -1 && cmd == FIOSEEKDATA)) { 2201 noff = bn * bsize; 2202 if (noff < *off) 2203 noff = *off; 2204 goto unlock; 2205 } 2206 } 2207 if (noff > va.va_size) 2208 noff = va.va_size; 2209 /* noff == va.va_size. There is an implicit hole at the end of file. */ 2210 if (cmd == FIOSEEKDATA) 2211 error = ENXIO; 2212 unlock: 2213 VOP_UNLOCK(vp, 0); 2214 if (error == 0) 2215 *off = noff; 2216 return (error); 2217 } 2218 2219 int 2220 vn_seek(struct file *fp, off_t offset, int whence, struct thread *td) 2221 { 2222 struct ucred *cred; 2223 struct vnode *vp; 2224 struct vattr vattr; 2225 off_t foffset, size; 2226 int error, noneg; 2227 2228 cred = td->td_ucred; 2229 vp = fp->f_vnode; 2230 foffset = foffset_lock(fp, 0); 2231 noneg = (vp->v_type != VCHR); 2232 error = 0; 2233 switch (whence) { 2234 case L_INCR: 2235 if (noneg && 2236 (foffset < 0 || 2237 (offset > 0 && foffset > OFF_MAX - offset))) { 2238 error = EOVERFLOW; 2239 break; 2240 } 2241 offset += foffset; 2242 break; 2243 case L_XTND: 2244 vn_lock(vp, LK_SHARED | LK_RETRY); 2245 error = VOP_GETATTR(vp, &vattr, cred); 2246 VOP_UNLOCK(vp, 0); 2247 if (error) 2248 break; 2249 2250 /* 2251 * If the file references a disk device, then fetch 2252 * the media size and use that to determine the ending 2253 * offset. 2254 */ 2255 if (vattr.va_size == 0 && vp->v_type == VCHR && 2256 fo_ioctl(fp, DIOCGMEDIASIZE, &size, cred, td) == 0) 2257 vattr.va_size = size; 2258 if (noneg && 2259 (vattr.va_size > OFF_MAX || 2260 (offset > 0 && vattr.va_size > OFF_MAX - offset))) { 2261 error = EOVERFLOW; 2262 break; 2263 } 2264 offset += vattr.va_size; 2265 break; 2266 case L_SET: 2267 break; 2268 case SEEK_DATA: 2269 error = fo_ioctl(fp, FIOSEEKDATA, &offset, cred, td); 2270 break; 2271 case SEEK_HOLE: 2272 error = fo_ioctl(fp, FIOSEEKHOLE, &offset, cred, td); 2273 break; 2274 default: 2275 error = EINVAL; 2276 } 2277 if (error == 0 && noneg && offset < 0) 2278 error = EINVAL; 2279 if (error != 0) 2280 goto drop; 2281 VFS_KNOTE_UNLOCKED(vp, 0); 2282 td->td_uretoff.tdu_off = offset; 2283 drop: 2284 foffset_unlock(fp, offset, error != 0 ? FOF_NOUPDATE : 0); 2285 return (error); 2286 } 2287 2288 int 2289 vn_utimes_perm(struct vnode *vp, struct vattr *vap, struct ucred *cred, 2290 struct thread *td) 2291 { 2292 int error; 2293 2294 /* 2295 * Grant permission if the caller is the owner of the file, or 2296 * the super-user, or has ACL_WRITE_ATTRIBUTES permission on 2297 * on the file. If the time pointer is null, then write 2298 * permission on the file is also sufficient. 2299 * 2300 * From NFSv4.1, draft 21, 6.2.1.3.1, Discussion of Mask Attributes: 2301 * A user having ACL_WRITE_DATA or ACL_WRITE_ATTRIBUTES 2302 * will be allowed to set the times [..] to the current 2303 * server time. 2304 */ 2305 error = VOP_ACCESSX(vp, VWRITE_ATTRIBUTES, cred, td); 2306 if (error != 0 && (vap->va_vaflags & VA_UTIMES_NULL) != 0) 2307 error = VOP_ACCESS(vp, VWRITE, cred, td); 2308 return (error); 2309 } 2310 2311 int 2312 vn_fill_kinfo(struct file *fp, struct kinfo_file *kif, struct filedesc *fdp) 2313 { 2314 struct vnode *vp; 2315 int error; 2316 2317 if (fp->f_type == DTYPE_FIFO) 2318 kif->kf_type = KF_TYPE_FIFO; 2319 else 2320 kif->kf_type = KF_TYPE_VNODE; 2321 vp = fp->f_vnode; 2322 vref(vp); 2323 FILEDESC_SUNLOCK(fdp); 2324 error = vn_fill_kinfo_vnode(vp, kif); 2325 vrele(vp); 2326 FILEDESC_SLOCK(fdp); 2327 return (error); 2328 } 2329 2330 static inline void 2331 vn_fill_junk(struct kinfo_file *kif) 2332 { 2333 size_t len, olen; 2334 2335 /* 2336 * Simulate vn_fullpath returning changing values for a given 2337 * vp during e.g. coredump. 2338 */ 2339 len = (arc4random() % (sizeof(kif->kf_path) - 2)) + 1; 2340 olen = strlen(kif->kf_path); 2341 if (len < olen) 2342 strcpy(&kif->kf_path[len - 1], "$"); 2343 else 2344 for (; olen < len; olen++) 2345 strcpy(&kif->kf_path[olen], "A"); 2346 } 2347 2348 int 2349 vn_fill_kinfo_vnode(struct vnode *vp, struct kinfo_file *kif) 2350 { 2351 struct vattr va; 2352 char *fullpath, *freepath; 2353 int error; 2354 2355 kif->kf_un.kf_file.kf_file_type = vntype_to_kinfo(vp->v_type); 2356 freepath = NULL; 2357 fullpath = "-"; 2358 error = vn_fullpath(curthread, vp, &fullpath, &freepath); 2359 if (error == 0) { 2360 strlcpy(kif->kf_path, fullpath, sizeof(kif->kf_path)); 2361 } 2362 if (freepath != NULL) 2363 free(freepath, M_TEMP); 2364 2365 KFAIL_POINT_CODE(DEBUG_FP, fill_kinfo_vnode__random_path, 2366 vn_fill_junk(kif); 2367 ); 2368 2369 /* 2370 * Retrieve vnode attributes. 2371 */ 2372 va.va_fsid = VNOVAL; 2373 va.va_rdev = NODEV; 2374 vn_lock(vp, LK_SHARED | LK_RETRY); 2375 error = VOP_GETATTR(vp, &va, curthread->td_ucred); 2376 VOP_UNLOCK(vp, 0); 2377 if (error != 0) 2378 return (error); 2379 if (va.va_fsid != VNOVAL) 2380 kif->kf_un.kf_file.kf_file_fsid = va.va_fsid; 2381 else 2382 kif->kf_un.kf_file.kf_file_fsid = 2383 vp->v_mount->mnt_stat.f_fsid.val[0]; 2384 kif->kf_un.kf_file.kf_file_fsid_freebsd11 = 2385 kif->kf_un.kf_file.kf_file_fsid; /* truncate */ 2386 kif->kf_un.kf_file.kf_file_fileid = va.va_fileid; 2387 kif->kf_un.kf_file.kf_file_mode = MAKEIMODE(va.va_type, va.va_mode); 2388 kif->kf_un.kf_file.kf_file_size = va.va_size; 2389 kif->kf_un.kf_file.kf_file_rdev = va.va_rdev; 2390 kif->kf_un.kf_file.kf_file_rdev_freebsd11 = 2391 kif->kf_un.kf_file.kf_file_rdev; /* truncate */ 2392 return (0); 2393 } 2394 2395 int 2396 vn_mmap(struct file *fp, vm_map_t map, vm_offset_t *addr, vm_size_t size, 2397 vm_prot_t prot, vm_prot_t cap_maxprot, int flags, vm_ooffset_t foff, 2398 struct thread *td) 2399 { 2400 #ifdef HWPMC_HOOKS 2401 struct pmckern_map_in pkm; 2402 #endif 2403 struct mount *mp; 2404 struct vnode *vp; 2405 vm_object_t object; 2406 vm_prot_t maxprot; 2407 boolean_t writecounted; 2408 int error; 2409 2410 #if defined(COMPAT_FREEBSD7) || defined(COMPAT_FREEBSD6) || \ 2411 defined(COMPAT_FREEBSD5) || defined(COMPAT_FREEBSD4) 2412 /* 2413 * POSIX shared-memory objects are defined to have 2414 * kernel persistence, and are not defined to support 2415 * read(2)/write(2) -- or even open(2). Thus, we can 2416 * use MAP_ASYNC to trade on-disk coherence for speed. 2417 * The shm_open(3) library routine turns on the FPOSIXSHM 2418 * flag to request this behavior. 2419 */ 2420 if ((fp->f_flag & FPOSIXSHM) != 0) 2421 flags |= MAP_NOSYNC; 2422 #endif 2423 vp = fp->f_vnode; 2424 2425 /* 2426 * Ensure that file and memory protections are 2427 * compatible. Note that we only worry about 2428 * writability if mapping is shared; in this case, 2429 * current and max prot are dictated by the open file. 2430 * XXX use the vnode instead? Problem is: what 2431 * credentials do we use for determination? What if 2432 * proc does a setuid? 2433 */ 2434 mp = vp->v_mount; 2435 if (mp != NULL && (mp->mnt_flag & MNT_NOEXEC) != 0) { 2436 maxprot = VM_PROT_NONE; 2437 if ((prot & VM_PROT_EXECUTE) != 0) 2438 return (EACCES); 2439 } else 2440 maxprot = VM_PROT_EXECUTE; 2441 if ((fp->f_flag & FREAD) != 0) 2442 maxprot |= VM_PROT_READ; 2443 else if ((prot & VM_PROT_READ) != 0) 2444 return (EACCES); 2445 2446 /* 2447 * If we are sharing potential changes via MAP_SHARED and we 2448 * are trying to get write permission although we opened it 2449 * without asking for it, bail out. 2450 */ 2451 if ((flags & MAP_SHARED) != 0) { 2452 if ((fp->f_flag & FWRITE) != 0) 2453 maxprot |= VM_PROT_WRITE; 2454 else if ((prot & VM_PROT_WRITE) != 0) 2455 return (EACCES); 2456 } else { 2457 maxprot |= VM_PROT_WRITE; 2458 cap_maxprot |= VM_PROT_WRITE; 2459 } 2460 maxprot &= cap_maxprot; 2461 2462 /* 2463 * For regular files and shared memory, POSIX requires that 2464 * the value of foff be a legitimate offset within the data 2465 * object. In particular, negative offsets are invalid. 2466 * Blocking negative offsets and overflows here avoids 2467 * possible wraparound or user-level access into reserved 2468 * ranges of the data object later. In contrast, POSIX does 2469 * not dictate how offsets are used by device drivers, so in 2470 * the case of a device mapping a negative offset is passed 2471 * on. 2472 */ 2473 if ( 2474 #ifdef _LP64 2475 size > OFF_MAX || 2476 #endif 2477 foff < 0 || foff > OFF_MAX - size) 2478 return (EINVAL); 2479 2480 writecounted = FALSE; 2481 error = vm_mmap_vnode(td, size, prot, &maxprot, &flags, vp, 2482 &foff, &object, &writecounted); 2483 if (error != 0) 2484 return (error); 2485 error = vm_mmap_object(map, addr, size, prot, maxprot, flags, object, 2486 foff, writecounted, td); 2487 if (error != 0) { 2488 /* 2489 * If this mapping was accounted for in the vnode's 2490 * writecount, then undo that now. 2491 */ 2492 if (writecounted) 2493 vnode_pager_release_writecount(object, 0, size); 2494 vm_object_deallocate(object); 2495 } 2496 #ifdef HWPMC_HOOKS 2497 /* Inform hwpmc(4) if an executable is being mapped. */ 2498 if (PMC_HOOK_INSTALLED(PMC_FN_MMAP)) { 2499 if ((prot & VM_PROT_EXECUTE) != 0 && error == 0) { 2500 pkm.pm_file = vp; 2501 pkm.pm_address = (uintptr_t) *addr; 2502 PMC_CALL_HOOK_UNLOCKED(td, PMC_FN_MMAP, (void *) &pkm); 2503 } 2504 } 2505 #endif 2506 return (error); 2507 } 2508 2509 void 2510 vn_fsid(struct vnode *vp, struct vattr *va) 2511 { 2512 fsid_t *f; 2513 2514 f = &vp->v_mount->mnt_stat.f_fsid; 2515 va->va_fsid = (uint32_t)f->val[1]; 2516 va->va_fsid <<= sizeof(f->val[1]) * NBBY; 2517 va->va_fsid += (uint32_t)f->val[0]; 2518 } 2519 2520 int 2521 vn_fsync_buf(struct vnode *vp, int waitfor) 2522 { 2523 struct buf *bp, *nbp; 2524 struct bufobj *bo; 2525 struct mount *mp; 2526 int error, maxretry; 2527 2528 error = 0; 2529 maxretry = 10000; /* large, arbitrarily chosen */ 2530 mp = NULL; 2531 if (vp->v_type == VCHR) { 2532 VI_LOCK(vp); 2533 mp = vp->v_rdev->si_mountpt; 2534 VI_UNLOCK(vp); 2535 } 2536 bo = &vp->v_bufobj; 2537 BO_LOCK(bo); 2538 loop1: 2539 /* 2540 * MARK/SCAN initialization to avoid infinite loops. 2541 */ 2542 TAILQ_FOREACH(bp, &bo->bo_dirty.bv_hd, b_bobufs) { 2543 bp->b_vflags &= ~BV_SCANNED; 2544 bp->b_error = 0; 2545 } 2546 2547 /* 2548 * Flush all dirty buffers associated with a vnode. 2549 */ 2550 loop2: 2551 TAILQ_FOREACH_SAFE(bp, &bo->bo_dirty.bv_hd, b_bobufs, nbp) { 2552 if ((bp->b_vflags & BV_SCANNED) != 0) 2553 continue; 2554 bp->b_vflags |= BV_SCANNED; 2555 if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT, NULL)) { 2556 if (waitfor != MNT_WAIT) 2557 continue; 2558 if (BUF_LOCK(bp, 2559 LK_EXCLUSIVE | LK_INTERLOCK | LK_SLEEPFAIL, 2560 BO_LOCKPTR(bo)) != 0) { 2561 BO_LOCK(bo); 2562 goto loop1; 2563 } 2564 BO_LOCK(bo); 2565 } 2566 BO_UNLOCK(bo); 2567 KASSERT(bp->b_bufobj == bo, 2568 ("bp %p wrong b_bufobj %p should be %p", 2569 bp, bp->b_bufobj, bo)); 2570 if ((bp->b_flags & B_DELWRI) == 0) 2571 panic("fsync: not dirty"); 2572 if ((vp->v_object != NULL) && (bp->b_flags & B_CLUSTEROK)) { 2573 vfs_bio_awrite(bp); 2574 } else { 2575 bremfree(bp); 2576 bawrite(bp); 2577 } 2578 if (maxretry < 1000) 2579 pause("dirty", hz < 1000 ? 1 : hz / 1000); 2580 BO_LOCK(bo); 2581 goto loop2; 2582 } 2583 2584 /* 2585 * If synchronous the caller expects us to completely resolve all 2586 * dirty buffers in the system. Wait for in-progress I/O to 2587 * complete (which could include background bitmap writes), then 2588 * retry if dirty blocks still exist. 2589 */ 2590 if (waitfor == MNT_WAIT) { 2591 bufobj_wwait(bo, 0, 0); 2592 if (bo->bo_dirty.bv_cnt > 0) { 2593 /* 2594 * If we are unable to write any of these buffers 2595 * then we fail now rather than trying endlessly 2596 * to write them out. 2597 */ 2598 TAILQ_FOREACH(bp, &bo->bo_dirty.bv_hd, b_bobufs) 2599 if ((error = bp->b_error) != 0) 2600 break; 2601 if ((mp != NULL && mp->mnt_secondary_writes > 0) || 2602 (error == 0 && --maxretry >= 0)) 2603 goto loop1; 2604 if (error == 0) 2605 error = EAGAIN; 2606 } 2607 } 2608 BO_UNLOCK(bo); 2609 if (error != 0) 2610 vn_printf(vp, "fsync: giving up on dirty (error = %d) ", error); 2611 2612 return (error); 2613 } 2614