1 /* 2 * Copyright (c) 2004-2020 The DragonFly Project. All rights reserved. 3 * 4 * This code is derived from software contributed to The DragonFly Project 5 * by Matthew Dillon <dillon@backplane.com> 6 * 7 * Redistribution and use in source and binary forms, with or without 8 * modification, are permitted provided that the following conditions 9 * are met: 10 * 11 * 1. Redistributions of source code must retain the above copyright 12 * notice, this list of conditions and the following disclaimer. 13 * 2. Redistributions in binary form must reproduce the above copyright 14 * notice, this list of conditions and the following disclaimer in 15 * the documentation and/or other materials provided with the 16 * distribution. 17 * 3. Neither the name of The DragonFly Project nor the names of its 18 * contributors may be used to endorse or promote products derived 19 * from this software without specific, prior written permission. 20 * 21 * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS 22 * ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT 23 * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS 24 * FOR A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE 25 * COPYRIGHT HOLDERS OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, 26 * INCIDENTAL, SPECIAL, EXEMPLARY OR CONSEQUENTIAL DAMAGES (INCLUDING, 27 * BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; 28 * LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED 29 * AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, 30 * OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT 31 * OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF 32 * SUCH DAMAGE. 33 */ 34 /* 35 * nlookup() is the 'new' namei interface. Rather then return directory and 36 * leaf vnodes (in various lock states) the new interface instead deals in 37 * namecache records. Namecache records may represent both a positive or 38 * a negative hit. The namespace is locked via the namecache record instead 39 * of via the vnode, and only the leaf namecache record (representing the 40 * filename) needs to be locked. 41 * 42 * This greatly improves filesystem parallelism and is a huge simplification 43 * of the API verses the old vnode locking / namei scheme. 44 * 45 * Filesystems must actively control the caching aspects of the namecache, 46 * and since namecache pointers are used as handles they are non-optional 47 * even for filesystems which do not generally wish to cache things. It is 48 * intended that a separate cache coherency API will be constructed to handle 49 * these issues. 50 */ 51 52 #include "opt_ktrace.h" 53 54 #include <sys/param.h> 55 #include <sys/systm.h> 56 #include <sys/uio.h> 57 #include <sys/kernel.h> 58 #include <sys/vnode.h> 59 #include <sys/mount.h> 60 #include <sys/filedesc.h> 61 #include <sys/proc.h> 62 #include <sys/namei.h> 63 #include <sys/nlookup.h> 64 #include <sys/malloc.h> 65 #include <sys/stat.h> 66 #include <sys/objcache.h> 67 #include <sys/file.h> 68 #include <sys/kcollect.h> 69 #include <sys/sysctl.h> 70 71 #ifdef KTRACE 72 #include <sys/ktrace.h> 73 #endif 74 75 __read_mostly static int nlookup_debug; 76 SYSCTL_INT(_debug, OID_AUTO, nlookup_debug, CTLFLAG_RW, &nlookup_debug, 0, 77 "Force retry test"); 78 79 static int naccess(struct nchandle *nch, int vmode, struct ucred *cred, 80 int *stickyp, int nchislocked); 81 82 /* 83 * unmount operations flag NLC_IGNBADDIR in order to allow the 84 * umount to successfully issue a nlookup() on the path in order 85 * to extract the mount point. Allow certain errors through. 86 */ 87 static __inline 88 int 89 keeperror(struct nlookupdata *nd, int error) 90 { 91 if (error) { 92 if ((nd->nl_flags & NLC_IGNBADDIR) == 0 || 93 (error != EIO && error != EBADRPC && error != ESTALE)) { 94 return 1; 95 } 96 } 97 return 0; 98 } 99 100 /* 101 * Initialize a nlookup() structure, early error return for copyin faults 102 * or a degenerate empty string (which is not allowed). 103 * 104 * The first process proc0's credentials are used if the calling thread 105 * is not associated with a process context. 106 * 107 * MPSAFE 108 */ 109 int 110 nlookup_init(struct nlookupdata *nd, 111 const char *path, enum uio_seg seg, int flags) 112 { 113 size_t pathlen; 114 struct proc *p; 115 thread_t td; 116 int error; 117 118 td = curthread; 119 p = td->td_proc; 120 121 /* 122 * note: the pathlen set by copy*str() includes the terminating \0. 123 */ 124 bzero(nd, sizeof(struct nlookupdata)); 125 nd->nl_path = objcache_get(namei_oc, M_WAITOK); 126 nd->nl_flags |= NLC_HASBUF; 127 if (seg == UIO_SYSSPACE) 128 error = copystr(path, nd->nl_path, MAXPATHLEN, &pathlen); 129 else 130 error = copyinstr(path, nd->nl_path, MAXPATHLEN, &pathlen); 131 132 /* 133 * Don't allow empty pathnames. 134 * POSIX.1 requirement: "" is not a vaild file name. 135 */ 136 if (error == 0 && pathlen <= 1) 137 error = ENOENT; 138 139 if (error == 0) { 140 if (p && p->p_fd) { 141 if (nd->nl_path[0] == '/') { 142 if ((flags & NLC_NLNCH_NOINIT) == 0) { 143 nd->nl_basench = &p->p_fd->fd_nrdir; 144 cache_copy(nd->nl_basench, &nd->nl_nch); 145 } 146 cache_copy(&p->p_fd->fd_nrdir, &nd->nl_rootnch); 147 if (p->p_fd->fd_njdir.ncp) 148 cache_copy(&p->p_fd->fd_njdir, &nd->nl_jailnch); 149 nd->nl_cred = td->td_ucred; 150 nd->nl_flags |= NLC_BORROWCRED; 151 } else { 152 if ((flags & NLC_NLNCH_NOINIT) == 0) { 153 nd->nl_basench = &p->p_fd->fd_ncdir; 154 cache_copy(nd->nl_basench, &nd->nl_nch); 155 } 156 cache_copy(&p->p_fd->fd_nrdir, &nd->nl_rootnch); 157 if (p->p_fd->fd_njdir.ncp) 158 cache_copy(&p->p_fd->fd_njdir, &nd->nl_jailnch); 159 nd->nl_cred = td->td_ucred; 160 nd->nl_flags |= NLC_BORROWCRED; 161 } 162 } else { 163 if ((flags & NLC_NLNCH_NOINIT) == 0) { 164 nd->nl_basench = &rootnch; 165 cache_copy(nd->nl_basench, &nd->nl_nch); 166 } 167 cache_copy(&rootnch, &nd->nl_rootnch); 168 cache_copy(&rootnch, &nd->nl_jailnch); 169 nd->nl_cred = proc0.p_ucred; 170 nd->nl_flags |= NLC_BORROWCRED; 171 } 172 nd->nl_td = td; 173 nd->nl_flags |= flags & ~NLC_NLNCH_NOINIT; 174 } else { 175 nlookup_done(nd); 176 } 177 return(error); 178 } 179 180 181 /* 182 * nlookup_init() for "at" family of syscalls. 183 * 184 * Similar to nlookup_init() but if the path is relative and fd is not 185 * AT_FDCWD, the path will be interpreted relative to the directory pointed 186 * to by fd. In this case, the file entry pointed to by fd is ref'ed and 187 * returned in *fpp. 188 * 189 * If the call succeeds, nlookup_done_at() must be called to clean-up the nd 190 * and release the ref to the file entry. 191 */ 192 int 193 nlookup_init_at(struct nlookupdata *nd, struct file **fpp, int fd, 194 const char *path, enum uio_seg seg, int flags) 195 { 196 struct thread *td = curthread; 197 struct file* fp; 198 struct vnode *vp; 199 int error; 200 201 *fpp = NULL; 202 203 /* 204 * Resolve the path, we might have to copy it in from userland, 205 * but don't initialize nl_basench, or nl_nch. 206 */ 207 error = nlookup_init(nd, path, seg, flags | NLC_NLNCH_NOINIT); 208 if (__predict_false(error)) 209 return (error); 210 211 /* 212 * Setup nl_basench (a pointer only not refd), and copy+ref 213 * to initialize nl_nch. Only applicable to relative paths. 214 * For absolute paths, or if (fd) is degenerate, just use the 215 * normal path. 216 */ 217 if (nd->nl_path[0] == '/') { 218 struct proc *p = curproc; 219 nd->nl_basench = &p->p_fd->fd_nrdir; 220 } else if (fd == AT_FDCWD) { 221 struct proc *p = curproc; 222 nd->nl_basench = &p->p_fd->fd_ncdir; 223 } else { 224 if ((error = holdvnode(td, fd, &fp)) != 0) 225 goto done; 226 vp = (struct vnode*)fp->f_data; 227 if (vp->v_type != VDIR || fp->f_nchandle.ncp == NULL) { 228 fdrop(fp); 229 fp = NULL; 230 error = ENOTDIR; 231 goto done; 232 } 233 nd->nl_basench = &fp->f_nchandle; 234 *fpp = fp; 235 } 236 cache_copy(nd->nl_basench, &nd->nl_nch); 237 done: 238 if (error) 239 nlookup_done(nd); 240 return (error); 241 } 242 243 /* 244 * This works similarly to nlookup_init() but does not assume a process 245 * context. rootnch is always chosen for the root directory and the cred 246 * and starting directory are supplied in arguments. 247 */ 248 int 249 nlookup_init_raw(struct nlookupdata *nd, 250 const char *path, enum uio_seg seg, int flags, 251 struct ucred *cred, struct nchandle *ncstart) 252 { 253 size_t pathlen; 254 thread_t td; 255 int error; 256 257 td = curthread; 258 259 bzero(nd, sizeof(struct nlookupdata)); 260 nd->nl_path = objcache_get(namei_oc, M_WAITOK); 261 nd->nl_flags |= NLC_HASBUF; 262 if (seg == UIO_SYSSPACE) 263 error = copystr(path, nd->nl_path, MAXPATHLEN, &pathlen); 264 else 265 error = copyinstr(path, nd->nl_path, MAXPATHLEN, &pathlen); 266 267 /* 268 * Don't allow empty pathnames. 269 * POSIX.1 requirement: "" is not a vaild file name. 270 */ 271 if (error == 0 && pathlen <= 1) 272 error = ENOENT; 273 274 if (error == 0) { 275 cache_copy(ncstart, &nd->nl_nch); 276 cache_copy(&rootnch, &nd->nl_rootnch); 277 cache_copy(&rootnch, &nd->nl_jailnch); 278 nd->nl_cred = crhold(cred); 279 nd->nl_td = td; 280 nd->nl_flags |= flags; 281 } else { 282 nlookup_done(nd); 283 } 284 return(error); 285 } 286 287 /* 288 * This works similarly to nlookup_init_raw() but does not rely 289 * on rootnch being initialized yet. 290 */ 291 int 292 nlookup_init_root(struct nlookupdata *nd, 293 const char *path, enum uio_seg seg, int flags, 294 struct ucred *cred, struct nchandle *ncstart, 295 struct nchandle *ncroot) 296 { 297 size_t pathlen; 298 thread_t td; 299 int error; 300 301 td = curthread; 302 303 bzero(nd, sizeof(struct nlookupdata)); 304 nd->nl_path = objcache_get(namei_oc, M_WAITOK); 305 nd->nl_flags |= NLC_HASBUF; 306 if (seg == UIO_SYSSPACE) 307 error = copystr(path, nd->nl_path, MAXPATHLEN, &pathlen); 308 else 309 error = copyinstr(path, nd->nl_path, MAXPATHLEN, &pathlen); 310 311 /* 312 * Don't allow empty pathnames. 313 * POSIX.1 requirement: "" is not a vaild file name. 314 */ 315 if (error == 0 && pathlen <= 1) 316 error = ENOENT; 317 318 if (error == 0) { 319 cache_copy(ncstart, &nd->nl_nch); 320 cache_copy(ncroot, &nd->nl_rootnch); 321 cache_copy(ncroot, &nd->nl_jailnch); 322 nd->nl_cred = crhold(cred); 323 nd->nl_td = td; 324 nd->nl_flags |= flags; 325 } else { 326 nlookup_done(nd); 327 } 328 return(error); 329 } 330 331 #if 0 332 /* 333 * Set a different credential; this credential will be used by future 334 * operations performed on nd.nl_open_vp and nlookupdata structure. 335 */ 336 void 337 nlookup_set_cred(struct nlookupdata *nd, struct ucred *cred) 338 { 339 KKASSERT(nd->nl_cred != NULL); 340 341 if (nd->nl_cred != cred) { 342 cred = crhold(cred); 343 if ((nd->nl_flags & NLC_BORROWCRED) == 0) 344 crfree(nd->nl_cred); 345 nd->nl_flags &= ~NLC_BORROWCRED; 346 nd->nl_cred = cred; 347 } 348 } 349 #endif 350 351 /* 352 * Cleanup a nlookupdata structure after we are through with it. This may 353 * be called on any nlookupdata structure initialized with nlookup_init(). 354 * Calling nlookup_done() is mandatory in all cases except where nlookup_init() 355 * returns an error, even if as a consumer you believe you have taken all 356 * dynamic elements out of the nlookupdata structure. 357 */ 358 void 359 nlookup_done(struct nlookupdata *nd) 360 { 361 if (nd->nl_nch.ncp) { 362 if (nd->nl_flags & NLC_NCPISLOCKED) 363 cache_unlock(&nd->nl_nch); 364 cache_drop_and_cache(&nd->nl_nch, nd->nl_elmno); 365 } 366 nd->nl_flags &= ~NLC_NCPISLOCKED; 367 if (nd->nl_rootnch.ncp) 368 cache_drop_and_cache(&nd->nl_rootnch, 0); 369 if (nd->nl_jailnch.ncp) 370 cache_drop_and_cache(&nd->nl_jailnch, 0); 371 if ((nd->nl_flags & NLC_HASBUF) && nd->nl_path) { 372 objcache_put(namei_oc, nd->nl_path); 373 nd->nl_path = NULL; 374 } 375 if (nd->nl_cred) { 376 if ((nd->nl_flags & NLC_BORROWCRED) == 0) 377 crfree(nd->nl_cred); 378 nd->nl_cred = NULL; 379 nd->nl_flags &= ~NLC_BORROWCRED; 380 } 381 if (nd->nl_open_vp) { 382 if (nd->nl_flags & NLC_LOCKVP) { 383 vn_unlock(nd->nl_open_vp); 384 nd->nl_flags &= ~NLC_LOCKVP; 385 } 386 vn_close(nd->nl_open_vp, nd->nl_vp_fmode, NULL); 387 nd->nl_open_vp = NULL; 388 } 389 if (nd->nl_dvp) { 390 vrele(nd->nl_dvp); 391 nd->nl_dvp = NULL; 392 } 393 nd->nl_flags = 0; /* clear remaining flags (just clear everything) */ 394 nd->nl_basench = NULL; 395 } 396 397 /* 398 * Works similarly to nlookup_done() when nd initialized with 399 * nlookup_init_at(). 400 */ 401 void 402 nlookup_done_at(struct nlookupdata *nd, struct file *fp) 403 { 404 nlookup_done(nd); 405 if (fp != NULL) 406 fdrop(fp); 407 } 408 409 void 410 nlookup_zero(struct nlookupdata *nd) 411 { 412 bzero(nd, sizeof(struct nlookupdata)); 413 } 414 415 /* 416 * Simple all-in-one nlookup. Returns a locked namecache structure or NULL 417 * if an error occured. 418 * 419 * Note that the returned ncp is not checked for permissions, though VEXEC 420 * is checked on the directory path leading up to the result. The caller 421 * must call naccess() to check the permissions of the returned leaf. 422 */ 423 struct nchandle 424 nlookup_simple(const char *str, enum uio_seg seg, 425 int niflags, int *error) 426 { 427 struct nlookupdata nd; 428 struct nchandle nch; 429 430 *error = nlookup_init(&nd, str, seg, niflags); 431 if (*error == 0) { 432 if ((*error = nlookup(&nd)) == 0) { 433 nch = nd.nl_nch; /* keep hold ref from structure */ 434 cache_zero(&nd.nl_nch); /* and NULL out */ 435 } else { 436 cache_zero(&nch); 437 } 438 nlookup_done(&nd); 439 } else { 440 cache_zero(&nch); 441 } 442 return(nch); 443 } 444 445 /* 446 * Returns non-zero if the path element is the last element 447 */ 448 static 449 int 450 islastelement(const char *ptr) 451 { 452 while (*ptr == '/') 453 ++ptr; 454 return (*ptr == 0); 455 } 456 457 /* 458 * Returns non-zero if we need to lock the namecache element 459 * exclusively. Unless otherwise requested by NLC_SHAREDLOCK, 460 * the last element of the namecache lookup will be locked 461 * exclusively. 462 * 463 * O_CREAT or O_TRUNC need the last element to be locked exlcusively. 464 * Intermediate elements are always locked shared. 465 * 466 * NOTE: Even if we return on-zero, an unresolved namecache record 467 * will always be locked exclusively. 468 */ 469 static __inline 470 int 471 wantsexcllock(struct nlookupdata *nd, int last_element) 472 { 473 if ((nd->nl_flags & NLC_SHAREDLOCK) == 0) 474 return(last_element); 475 return 0; 476 } 477 478 479 /* 480 * Do a generic nlookup. Note that the passed nd is not nlookup_done()'d 481 * on return, even if an error occurs. If no error occurs or NLC_CREATE 482 * is flagged and ENOENT is returned, then the returned nl_nch is always 483 * referenced and locked exclusively. 484 * 485 * WARNING: For any general error other than ENOENT w/NLC_CREATE, the 486 * the resulting nl_nch may or may not be locked and if locked 487 * might be locked either shared or exclusive. 488 * 489 * Intermediate directory elements, including the current directory, require 490 * execute (search) permission. nlookup does not examine the access 491 * permissions on the returned element. 492 * 493 * If NLC_CREATE is set the last directory must allow node creation, 494 * and an error code of 0 will be returned for a non-existant 495 * target (not ENOENT). 496 * 497 * If NLC_RENAME_DST is set the last directory mut allow node deletion, 498 * plus the sticky check is made, and an error code of 0 will be returned 499 * for a non-existant target (not ENOENT). 500 * 501 * If NLC_DELETE is set the last directory mut allow node deletion, 502 * plus the sticky check is made. 503 * 504 * If NLC_REFDVP is set nd->nl_dvp will be set to the directory vnode 505 * of the returned entry. The vnode will be referenced but not locked. 506 * 507 * IF THE PATH REPRESENTS A MOUNT POINT CROSSING THEN NLC_REFDVP WILL SET 508 * NL_DVP TO NULL AND RETURN NO ERROR (ERROR == 0), allowing the operation 509 * to return up the stack. The nch will only be referenced and not locked. 510 * High level code must check this case and do the right thing since, 511 * typically, it means things like 'mkdir' should fail with EEXIST. For 512 * example 'mkdir /var/cache' where /var/cache is a null-mount from 513 * /build/var.cache, needs to return EEXIST rather than a mount-crossing 514 * failure. 515 * 516 * NOTE: As an optimization we attempt to obtain a shared namecache lock 517 * on any intermediate elements. On success, the returned element 518 * is ALWAYS locked exclusively. 519 */ 520 int 521 nlookup(struct nlookupdata *nd) 522 { 523 globaldata_t gd = mycpu; 524 struct nlcomponent nlc; 525 struct nchandle nch; 526 struct nchandle nctmp; 527 struct mount *mp; 528 int wasdotordotdot; 529 char *path_reset; 530 char *ptr; 531 char *nptr; 532 int error; 533 int len; 534 int dflags; 535 int hit = 1; 536 int saveflag = nd->nl_flags; 537 boolean_t doretry = FALSE; 538 boolean_t inretry = FALSE; 539 540 if (nlookup_debug > 0) { 541 --nlookup_debug; 542 doretry = 1; 543 } 544 path_reset = NULL; 545 546 nlookup_start: 547 548 #ifdef KTRACE 549 if (KTRPOINT(nd->nl_td, KTR_NAMEI)) 550 ktrnamei(nd->nl_td->td_lwp, nd->nl_path); 551 #endif 552 bzero(&nlc, sizeof(nlc)); 553 554 /* 555 * Setup for the loop. The current working namecache element is 556 * always at least referenced. We lock it as required, but always 557 * return a locked, resolved namecache entry. 558 */ 559 nd->nl_loopcnt = 0; 560 if (nd->nl_dvp) { 561 vrele(nd->nl_dvp); 562 nd->nl_dvp = NULL; 563 } 564 ptr = nd->nl_path; 565 566 /* 567 * Loop on the path components. At the top of the loop nd->nl_nch 568 * is ref'd and unlocked and represents our current position. 569 */ 570 for (;;) { 571 int last_element; 572 573 ++nd->nl_elmno; 574 KKASSERT((nd->nl_flags & NLC_NCPISLOCKED) == 0); 575 576 /* 577 * Check if the root directory should replace the current 578 * directory. This is done at the start of a translation 579 * or after a symbolic link has been found. In other cases 580 * ptr will never be pointing at a '/'. 581 */ 582 if (*ptr == '/') { 583 do { 584 ++ptr; 585 } while (*ptr == '/'); 586 587 /* 588 * We might already be at the root as a pre-optimization 589 */ 590 if (nd->nl_nch.mount != nd->nl_rootnch.mount || 591 nd->nl_nch.ncp != nd->nl_rootnch.ncp) { 592 cache_drop_and_cache(&nd->nl_nch, 0); 593 cache_copy(&nd->nl_rootnch, &nd->nl_nch); 594 } 595 596 /* 597 * Fast-track termination. There is no parent directory of 598 * the root in the same mount from the point of view of 599 * the caller so return EACCES if NLC_REFDVP is specified, 600 * and EEXIST if NLC_CREATE is also specified. 601 * e.g. 'rmdir /' or 'mkdir /' are not allowed. 602 */ 603 if (*ptr == 0) { 604 if (nd->nl_flags & NLC_REFDVP) 605 error = (nd->nl_flags & NLC_CREATE) ? EEXIST : EACCES; 606 else 607 error = 0; 608 nd->nl_flags |= NLC_NCPISLOCKED; 609 cache_lock_maybe_shared(&nd->nl_nch, 610 wantsexcllock(nd, islastelement(ptr))); 611 break; 612 } 613 continue; 614 } 615 616 /* 617 * Pre-calculate next path component so we can check whether the 618 * current component directory is the last directory in the path 619 * or not. 620 */ 621 for (nptr = ptr; *nptr && *nptr != '/'; ++nptr) 622 ; 623 624 /* 625 * nd->nl_nch is referenced and not locked here. 626 * 627 * Check directory search permissions. This will load dflags to 628 * obtain directory-special permissions to be checked along with the 629 * last component. 630 * 631 * We only need to pass-in &dflags for the second-to-last component. 632 * Optimize by passing-in NULL for any prior components, which may 633 * allow the code to bypass the naccess() call. 634 * 635 * naccess() is optimized to avoid having to lock the nch or get 636 * the related vnode if cached perms are sufficient. 637 */ 638 dflags = 0; 639 if (*nptr == '/' || (saveflag & NLC_MODIFYING_MASK) == 0) 640 error = naccess(&nd->nl_nch, NLC_EXEC, nd->nl_cred, NULL, 0); 641 else 642 error = naccess(&nd->nl_nch, NLC_EXEC, nd->nl_cred, &dflags, 0); 643 if (error) { 644 if (keeperror(nd, error)) 645 break; 646 error = 0; 647 } 648 649 /* 650 * Extract the next (or last) path component. Path components are 651 * limited to 255 characters. 652 */ 653 nlc.nlc_nameptr = ptr; 654 nlc.nlc_namelen = nptr - ptr; 655 ptr = nptr; 656 if (nlc.nlc_namelen >= 256) { 657 error = ENAMETOOLONG; 658 break; 659 } 660 last_element = islastelement(nptr); 661 662 /* 663 * Lookup the path component in the cache, creating an unresolved 664 * entry if necessary. We have to handle "." and ".." as special 665 * cases. 666 * 667 * When handling ".." we have to detect a traversal back through a 668 * mount point. If we are at the root, ".." just returns the root. 669 * 670 * When handling "." or ".." we also have to recalculate dflags 671 * since our dflags will be for some sub-directory instead of the 672 * parent dir. 673 * 674 * This subsection returns a referenced and possibly locked 'nch'. 675 * The locking status is based on the last_element flag. 676 * 677 * The namecache topology is not allowed to be disconnected, so 678 * encountering a NULL parent will generate EINVAL. This typically 679 * occurs when a directory is removed out from under a process. 680 * 681 * WARNING! The unlocking of nd->nl_nch is sensitive code. 682 */ 683 KKASSERT((nd->nl_flags & NLC_NCPISLOCKED) == 0); 684 685 if (nlc.nlc_namelen == 1 && nlc.nlc_nameptr[0] == '.') { 686 if (last_element) { 687 cache_get_maybe_shared(&nd->nl_nch, &nch, 688 wantsexcllock(nd, 1)); 689 } else { 690 cache_copy(&nd->nl_nch, &nch); 691 } 692 wasdotordotdot = 1; 693 } else if (nlc.nlc_namelen == 2 && 694 nlc.nlc_nameptr[0] == '.' && nlc.nlc_nameptr[1] == '.') { 695 if (nd->nl_nch.mount == nd->nl_rootnch.mount && 696 nd->nl_nch.ncp == nd->nl_rootnch.ncp 697 ) { 698 /* 699 * ".." at the root returns the root 700 */ 701 if (last_element) { 702 cache_get_maybe_shared(&nd->nl_nch, &nch, 703 wantsexcllock(nd, 1)); 704 } else { 705 cache_copy(&nd->nl_nch, &nch); 706 } 707 } else { 708 /* 709 * Locate the parent ncp. If we are at the root of a 710 * filesystem mount we have to skip to the mounted-on 711 * point in the underlying filesystem. 712 * 713 * Expect the parent to always be good since the 714 * mountpoint doesn't go away. XXX hack. cache_get() 715 * requires the ncp to already have a ref as a safety. 716 * 717 * However, a process which has been broken out of a chroot 718 * will wind up with a NULL parent if it tries to '..' above 719 * the real root, deal with the case. Note that this does 720 * not protect us from a jail breakout, it just stops a panic 721 * if the jail-broken process tries to '..' past the real 722 * root. 723 */ 724 nctmp = nd->nl_nch; 725 while (nctmp.ncp == nctmp.mount->mnt_ncmountpt.ncp) { 726 nctmp = nctmp.mount->mnt_ncmounton; 727 if (nctmp.ncp == NULL) 728 break; 729 } 730 if (nctmp.ncp == NULL) { 731 if (curthread->td_proc) { 732 kprintf("vfs_nlookup: '..' traverse broke " 733 "jail: pid %d (%s)\n", 734 curthread->td_proc->p_pid, 735 curthread->td_comm); 736 } 737 nctmp = nd->nl_rootnch; 738 } else { 739 nctmp.ncp = nctmp.ncp->nc_parent; 740 } 741 if (last_element) { 742 cache_get_maybe_shared(&nctmp, &nch, 743 wantsexcllock(nd, 1)); 744 } else { 745 cache_copy(&nctmp, &nch); 746 } 747 } 748 wasdotordotdot = 2; 749 } else { 750 /* 751 * Quickly lookup the component. If we can't find it, then 752 * slowly lookup and resolve the component. 753 */ 754 if (last_element) { 755 error = cache_nlookup_maybe_shared(&nd->nl_nch, &nlc, 756 wantsexcllock(nd, 1), 757 &nch); 758 } else { 759 nch = cache_nlookup_nonlocked(&nd->nl_nch, &nlc); 760 if (nch.ncp == NULL) 761 error = EWOULDBLOCK; 762 } 763 764 /* 765 * At this point the only possible error is EWOULDBLOCK. 766 * 767 * If no error nch is set and referenced, and then also locked 768 * according to last_element. For EWOULDBLOCK nch is not set. 769 * For any other error nch is set and referenced, but not locked. 770 * 771 * On EWOULDBLOCK the ncp may be unresolved (if not locked it can 772 * become unresolved at any time, but we don't care at this time). 773 */ 774 if (error == EWOULDBLOCK) { 775 nch = cache_nlookup(&nd->nl_nch, &nlc); 776 if (nch.ncp->nc_flag & NCF_UNRESOLVED) 777 hit = 0; 778 for (;;) { 779 error = cache_resolve(&nch, nd->nl_cred); 780 if (error != EAGAIN && 781 (nch.ncp->nc_flag & NCF_DESTROYED) == 0) { 782 if (error == ESTALE) { 783 if (!inretry) 784 error = ENOENT; 785 doretry = TRUE; 786 } 787 if (last_element == 0) 788 cache_unlock(&nch); 789 break; 790 } 791 kprintf("[diagnostic] nlookup: relookup %*.*s\n", 792 nch.ncp->nc_nlen, nch.ncp->nc_nlen, 793 nch.ncp->nc_name); 794 cache_put(&nch); 795 nch = cache_nlookup(&nd->nl_nch, &nlc); 796 } 797 } 798 wasdotordotdot = 0; 799 } 800 801 /* 802 * If the component is "." or ".." our dflags no longer represents 803 * the parent directory and we have to explicitly look it up. 804 * 805 * Expect the parent to be good since nch is locked. 806 * 807 * nch will continue to be valid even if an error occurs after this 808 * point. 809 */ 810 if (wasdotordotdot && error == 0) { 811 struct nchandle par; 812 813 dflags = 0; 814 if (last_element == 0) 815 cache_lock_maybe_shared(&nch, wantsexcllock(nd, 0)); 816 817 if ((par.ncp = nch.ncp->nc_parent) != NULL) { 818 par.mount = nch.mount; 819 cache_hold(&par); 820 error = naccess(&par, 0, nd->nl_cred, &dflags, 0); 821 cache_drop_and_cache(&par, nd->nl_elmno - 1); 822 if (error) { 823 if (!keeperror(nd, error)) 824 error = 0; 825 if (error == EINVAL) { 826 kprintf("nlookup (%s): trailing . or .. retry on %s\n", 827 curthread->td_comm, nd->nl_path); 828 doretry = TRUE; 829 } 830 } 831 } 832 833 if (last_element == 0) 834 cache_unlock(&nch); 835 } 836 837 /* 838 * [end of subsection] 839 * 840 * nch is referenced and locked according to (last_element). 841 * nd->nl_nch is unlocked and referenced. 842 */ 843 KKASSERT((nd->nl_flags & NLC_NCPISLOCKED) == 0); 844 845 /* 846 * Resolve the namespace if necessary. The ncp returned by 847 * cache_nlookup() is referenced, and also locked according 848 * to last_element. 849 * 850 * XXX neither '.' nor '..' should return EAGAIN since they were 851 * previously resolved and thus cannot be newly created ncp's. 852 */ 853 if (nch.ncp->nc_flag & NCF_UNRESOLVED) { 854 if (last_element == 0) 855 cache_lock(&nch); 856 hit = 0; 857 error = cache_resolve(&nch, nd->nl_cred); 858 if (error == ESTALE) { 859 if (!inretry) 860 error = ENOENT; 861 doretry = TRUE; 862 } 863 if (last_element == 0) 864 cache_unlock(&nch); 865 KKASSERT(error != EAGAIN); 866 } else { 867 error = nch.ncp->nc_error; 868 } 869 870 /* 871 * Early completion. ENOENT is not an error if this is the last 872 * component and NLC_CREATE or NLC_RENAME (rename target) was 873 * requested. Note that ncp->nc_error is left as ENOENT in that 874 * case, which we check later on. 875 * 876 * Also handle invalid '.' or '..' components terminating a path 877 * for a create/rename/delete. The standard requires this and pax 878 * pretty stupidly depends on it. 879 */ 880 if (last_element) { 881 if (error == ENOENT && 882 (nd->nl_flags & (NLC_CREATE | NLC_RENAME_DST)) 883 ) { 884 if (nd->nl_flags & NLC_NFS_RDONLY) { 885 error = EROFS; 886 } else { 887 error = naccess(&nch, nd->nl_flags | dflags, 888 nd->nl_cred, NULL, last_element); 889 } 890 } 891 if (error == 0 && wasdotordotdot && 892 (nd->nl_flags & (NLC_CREATE | NLC_DELETE | 893 NLC_RENAME_SRC | NLC_RENAME_DST))) { 894 /* 895 * POSIX junk 896 */ 897 if (nd->nl_flags & NLC_CREATE) 898 error = EEXIST; 899 else if (nd->nl_flags & NLC_DELETE) 900 error = (wasdotordotdot == 1) ? EINVAL : ENOTEMPTY; 901 else 902 error = EINVAL; 903 } 904 } 905 906 /* 907 * Early completion on error. 908 */ 909 if (error) { 910 if (last_element) 911 cache_unlock(&nch); 912 cache_drop_and_cache(&nch, nd->nl_elmno); 913 break; 914 } 915 916 /* 917 * If the element is a symlink and it is either not the last 918 * element or it is the last element and we are allowed to 919 * follow symlinks, resolve the symlink. 920 */ 921 if ((nch.ncp->nc_flag & NCF_ISSYMLINK) && 922 (*ptr || (nd->nl_flags & NLC_FOLLOW)) 923 ) { 924 if (nd->nl_loopcnt++ >= MAXSYMLINKS) { 925 error = ELOOP; 926 if (last_element) 927 cache_unlock(&nch); 928 cache_drop_and_cache(&nch, nd->nl_elmno); 929 break; 930 } 931 if (last_element == 0) 932 cache_lock_maybe_shared(&nch, 1); 933 934 error = nreadsymlink(nd, &nch, &nlc); 935 cache_put(&nch); 936 if (error) 937 break; 938 939 /* 940 * Concatenate trailing path elements onto the returned symlink. 941 * Note that if the path component (ptr) is not exhausted, it 942 * will being with a '/', so we do not have to add another one. 943 * 944 * The symlink may not be empty. 945 */ 946 len = strlen(ptr); 947 if (nlc.nlc_namelen == 0 || nlc.nlc_namelen + len >= MAXPATHLEN) { 948 error = nlc.nlc_namelen ? ENAMETOOLONG : ENOENT; 949 objcache_put(namei_oc, nlc.nlc_nameptr); 950 break; 951 } 952 bcopy(ptr, nlc.nlc_nameptr + nlc.nlc_namelen, len + 1); 953 954 if (path_reset) { 955 if (nd->nl_flags & NLC_HASBUF) 956 objcache_put(namei_oc, nd->nl_path); 957 } else { 958 path_reset = nd->nl_path; 959 } 960 nd->nl_path = nlc.nlc_nameptr; 961 nd->nl_flags |= NLC_HASBUF; 962 ptr = nd->nl_path; 963 964 /* 965 * Go back up to the top to resolve any initial '/'s in the 966 * symlink. 967 */ 968 continue; 969 } 970 971 /* 972 * If the element is a directory and we are crossing a mount point, 973 * Locate the mount. 974 */ 975 while ((nch.ncp->nc_flag & NCF_ISMOUNTPT) && 976 (nd->nl_flags & NLC_NOCROSSMOUNT) == 0 && 977 (mp = cache_findmount(&nch)) != NULL 978 ) { 979 struct vnode *tdp; 980 int vfs_do_busy = 0; 981 982 /* 983 * VFS must be busied before the namecache entry is locked, 984 * but we don't want to waste time calling vfs_busy() if the 985 * mount point is already resolved. 986 */ 987 again: 988 if (last_element) 989 cache_unlock(&nch); 990 cache_drop_and_cache(&nch, nd->nl_elmno); 991 992 if (vfs_do_busy) { 993 while (vfs_busy(mp, 0)) { 994 if (mp->mnt_kern_flag & MNTK_UNMOUNT) { 995 kprintf("nlookup: warning umount race avoided\n"); 996 cache_dropmount(mp); 997 error = EBUSY; 998 vfs_do_busy = 0; 999 goto double_break; 1000 } 1001 } 1002 } 1003 1004 /* 1005 * We don't need to lock the nch unless the entry is unresolved 1006 * or this is the last element. 1007 */ 1008 if (last_element) 1009 cache_get_maybe_shared(&mp->mnt_ncmountpt, &nch, 1010 wantsexcllock(nd, 1)); 1011 else 1012 cache_copy(&mp->mnt_ncmountpt, &nch); 1013 1014 if (nch.ncp->nc_flag & NCF_UNRESOLVED) { 1015 if (last_element == 0) 1016 cache_lock(&nch); 1017 if (nch.ncp->nc_flag & NCF_UNRESOLVED) { 1018 if (vfs_do_busy == 0) { 1019 vfs_do_busy = 1; 1020 if (last_element == 0) 1021 cache_unlock(&nch); 1022 goto again; 1023 } 1024 error = VFS_ROOT(mp, &tdp); 1025 vfs_unbusy(mp); 1026 vfs_do_busy = 0; 1027 if (keeperror(nd, error)) { 1028 cache_dropmount(mp); 1029 if (last_element == 0) 1030 cache_unlock(&nch); 1031 break; 1032 } 1033 if (error == 0) { 1034 cache_setvp(&nch, tdp); 1035 vput(tdp); 1036 } 1037 } 1038 if (last_element == 0) 1039 cache_unlock(&nch); 1040 } 1041 if (vfs_do_busy) 1042 vfs_unbusy(mp); 1043 cache_dropmount(mp); 1044 } 1045 1046 /* 1047 * Break out on error 1048 */ 1049 if (keeperror(nd, error)) { 1050 if (last_element) 1051 cache_unlock(&nch); 1052 cache_drop_and_cache(&nch, nd->nl_elmno); 1053 double_break: 1054 break; 1055 } 1056 1057 /* 1058 * Skip any slashes to get to the next element. If there 1059 * are any slashes at all the current element must be a 1060 * directory or, in the create case, intended to become a directory. 1061 * If it isn't we break without incrementing ptr and fall through 1062 * to the failure case below. 1063 */ 1064 while (*ptr == '/') { 1065 if ((nch.ncp->nc_flag & NCF_ISDIR) == 0 && 1066 !(nd->nl_flags & NLC_WILLBEDIR) 1067 ) { 1068 break; 1069 } 1070 ++ptr; 1071 } 1072 1073 /* 1074 * Continuation case: additional elements and the current 1075 * element is a directory. 1076 */ 1077 if (*ptr && (nch.ncp->nc_flag & NCF_ISDIR)) { 1078 cache_drop_and_cache(&nd->nl_nch, nd->nl_elmno); 1079 if (last_element) 1080 cache_unlock(&nch); 1081 /*nchislocked = 0; not needed */ 1082 KKASSERT((nd->nl_flags & NLC_NCPISLOCKED) == 0); 1083 nd->nl_nch = nch; 1084 continue; 1085 } 1086 1087 /* 1088 * Failure case: additional elements and the current element 1089 * is not a directory 1090 */ 1091 if (*ptr) { 1092 if (last_element) 1093 cache_unlock(&nch); 1094 cache_drop_and_cache(&nch, nd->nl_elmno); 1095 error = ENOTDIR; 1096 break; 1097 } 1098 1099 /* 1100 * Successful lookup of last element. 1101 * 1102 * Check permissions if the target exists. If the target does not 1103 * exist directory permissions were already tested in the early 1104 * completion code above. 1105 * 1106 * nd->nl_flags will be adjusted on return with NLC_APPENDONLY 1107 * if the file is marked append-only, and NLC_STICKY if the directory 1108 * containing the file is sticky. 1109 */ 1110 KKASSERT(last_element); 1111 1112 if (nch.ncp->nc_vp && (nd->nl_flags & NLC_ALLCHKS)) { 1113 error = naccess(&nch, nd->nl_flags | dflags, 1114 nd->nl_cred, NULL, 1); 1115 if (keeperror(nd, error)) { 1116 cache_put(&nch); 1117 break; 1118 } 1119 } 1120 1121 /* 1122 * Termination: no more elements. 1123 * 1124 * Check to see if the immediate parent has been destroyed. This race 1125 * can occur because the element lookup must temporarily unlock 1126 * the parent. If so, do a retry. 1127 */ 1128 if (nch.ncp->nc_parent && 1129 (nch.ncp->nc_parent->nc_flag & NCF_DESTROYED)) { 1130 doretry = TRUE; 1131 } 1132 1133 /* 1134 * Termination: no more elements. 1135 * 1136 * If NLC_REFDVP is set acquire a referenced parent dvp. Typically 1137 * used for mkdir/mknod/ncreate/nremove/unlink/rename. 1138 * 1139 * If a mount-point transition occurs due to ncp being a mount point, 1140 * or a null-mount, nl_dvp will be set to NULL and an error code of 1141 * 0 will be returned. A NULL nc_parent is not necessarily the only 1142 * indication of a mount-point as null-mounts will also tend to have 1143 * a non-null nc_parent. 1144 * 1145 * nch is locked, standard lock order for the namecache is 1146 * child-to-parent so we can safely lock its parent. We can 1147 * just use cache_dvpref(). 1148 */ 1149 if ((nd->nl_flags & NLC_REFDVP) && 1150 (doretry == FALSE || inretry == TRUE)) { 1151 if (nch.ncp->nc_parent) { 1152 error = cache_resolve_dvp(&nch, nd->nl_cred, 1153 &nd->nl_dvp); 1154 if (error) { 1155 kprintf("Parent directory lost during " 1156 "nlookup: %s/%s (%08x/%08x)\n", 1157 nch.ncp->nc_parent->nc_name, 1158 nch.ncp->nc_name, 1159 nch.ncp->nc_parent->nc_flag, 1160 nch.ncp->nc_flag); 1161 cache_put(&nch); 1162 error = EINVAL; 1163 break; 1164 } 1165 1166 /* 1167 * Mount-point, nl_dvp should remain NULL, error 0, 1168 * caller won't be able to use the results so leave 1169 * the ncp referenced but unlocked. 1170 */ 1171 if (nd->nl_dvp == NULL) { 1172 cache_put(&nch); 1173 break; 1174 } 1175 1176 /* 1177 * Good directory, fall through to drop-and-cache 1178 * below 1179 */ 1180 /* */ 1181 } else { 1182 /* 1183 * Mount-point, nl_dvp should remain NULL, error 0, 1184 * caller won't be able to use the results so leave 1185 * the ncp referenced but unlocked. 1186 */ 1187 error = 0; 1188 cache_put(&nch); 1189 break; 1190 } 1191 } 1192 1193 /* 1194 * ncp left with lock+ref on break, set NLC_NCPISLOCKED flag 1195 */ 1196 cache_drop_and_cache(&nd->nl_nch, nd->nl_elmno); 1197 nd->nl_nch = nch; 1198 nd->nl_flags |= NLC_NCPISLOCKED; 1199 error = 0; 1200 break; 1201 } 1202 1203 /* 1204 * We are done / or possibly retry 1205 */ 1206 if (hit) 1207 ++gd->gd_nchstats->ncs_longhits; 1208 else 1209 ++gd->gd_nchstats->ncs_longmiss; 1210 1211 if (nd->nl_flags & NLC_NCPISLOCKED) 1212 KKASSERT(cache_lockstatus(&nd->nl_nch) > 0); 1213 1214 /* 1215 * Reset nd->nl_path if necessary (due to softlinks). We want to return 1216 * nl_path to its original state before retrying or returning. 1217 */ 1218 if (path_reset) { 1219 if (nd->nl_flags & NLC_HASBUF) { 1220 objcache_put(namei_oc, nd->nl_path); 1221 nd->nl_flags &= ~NLC_HASBUF; 1222 } 1223 nd->nl_path = path_reset; 1224 nd->nl_flags |= saveflag & NLC_HASBUF; 1225 path_reset = NULL; 1226 } 1227 1228 /* 1229 * Retry the whole thing if doretry flag is set, but only once. 1230 * 1231 * autofs(5) may mount another filesystem under its root directory 1232 * while resolving a path. 1233 * 1234 * NFS might return ESTALE 1235 */ 1236 if (doretry && !inretry) { 1237 kprintf("nlookup: errno %d retry %s\n", error, nd->nl_path); 1238 inretry = TRUE; 1239 1240 /* 1241 * Clean up nd->nl_nch and reset to base directory 1242 */ 1243 if (nd->nl_flags & NLC_NCPISLOCKED) { 1244 cache_unlock(&nd->nl_nch); 1245 nd->nl_flags &= ~NLC_NCPISLOCKED; 1246 } 1247 cache_drop(&nd->nl_nch); 1248 cache_copy(nd->nl_basench, &nd->nl_nch); 1249 1250 nd->nl_elmno = 0; 1251 nd->nl_flags |= saveflag; 1252 1253 goto nlookup_start; 1254 } 1255 1256 /* 1257 * NOTE: If NLC_CREATE was set the ncp may represent a negative hit 1258 * (ncp->nc_error will be ENOENT), but we will still return an error 1259 * code of 0. 1260 */ 1261 return(error); 1262 } 1263 1264 /* 1265 * Resolve a mount point's glue ncp. This ncp connects creates the illusion 1266 * of continuity in the namecache tree by connecting the ncp related to the 1267 * vnode under the mount to the ncp related to the mount's root vnode. 1268 * 1269 * If no error occured a locked, ref'd ncp is stored in *ncpp. 1270 */ 1271 int 1272 nlookup_mp(struct mount *mp, struct nchandle *nch) 1273 { 1274 struct vnode *vp; 1275 int error; 1276 1277 error = 0; 1278 cache_get(&mp->mnt_ncmountpt, nch); 1279 if (nch->ncp->nc_flag & NCF_UNRESOLVED) { 1280 while (vfs_busy(mp, 0)) 1281 ; 1282 error = VFS_ROOT(mp, &vp); 1283 vfs_unbusy(mp); 1284 if (error) { 1285 cache_put(nch); 1286 } else { 1287 cache_setvp(nch, vp); 1288 vput(vp); 1289 } 1290 } 1291 return(error); 1292 } 1293 1294 /* 1295 * Read the contents of a symlink, allocate a path buffer out of the 1296 * namei_oc and initialize the supplied nlcomponent with the result. 1297 * 1298 * If an error occurs no buffer will be allocated or returned in the nlc. 1299 */ 1300 int 1301 nreadsymlink(struct nlookupdata *nd, struct nchandle *nch, 1302 struct nlcomponent *nlc) 1303 { 1304 struct vnode *vp; 1305 struct iovec aiov; 1306 struct uio auio; 1307 int linklen; 1308 int error; 1309 char *cp; 1310 1311 nlc->nlc_nameptr = NULL; 1312 nlc->nlc_namelen = 0; 1313 if (nch->ncp->nc_vp == NULL) 1314 return(ENOENT); 1315 if ((error = cache_vget(nch, nd->nl_cred, LK_SHARED, &vp)) != 0) 1316 return(error); 1317 cp = objcache_get(namei_oc, M_WAITOK); 1318 aiov.iov_base = cp; 1319 aiov.iov_len = MAXPATHLEN; 1320 auio.uio_iov = &aiov; 1321 auio.uio_iovcnt = 1; 1322 auio.uio_offset = 0; 1323 auio.uio_rw = UIO_READ; 1324 auio.uio_segflg = UIO_SYSSPACE; 1325 auio.uio_td = nd->nl_td; 1326 auio.uio_resid = MAXPATHLEN - 1; 1327 error = VOP_READLINK(vp, &auio, nd->nl_cred); 1328 if (error) 1329 goto fail; 1330 linklen = MAXPATHLEN - 1 - auio.uio_resid; 1331 if (varsym_enable) { 1332 linklen = varsymreplace(cp, linklen, MAXPATHLEN - 1); 1333 if (linklen < 0) { 1334 error = ENAMETOOLONG; 1335 goto fail; 1336 } 1337 } 1338 cp[linklen] = 0; 1339 nlc->nlc_nameptr = cp; 1340 nlc->nlc_namelen = linklen; 1341 vput(vp); 1342 return(0); 1343 fail: 1344 objcache_put(namei_oc, cp); 1345 vput(vp); 1346 return(error); 1347 } 1348 1349 /* 1350 * Check access [XXX cache vattr!] [XXX quota] 1351 * 1352 * Generally check the NLC_* access bits. All specified bits must pass 1353 * for this function to return 0. 1354 * 1355 * The file does not have to exist when checking NLC_CREATE or NLC_RENAME_DST 1356 * access, otherwise it must exist. No error is returned in this case. 1357 * 1358 * The file must not exist if NLC_EXCL is specified. 1359 * 1360 * Directory permissions in general are tested for NLC_CREATE if the file 1361 * does not exist, NLC_DELETE if the file does exist, and NLC_RENAME_DST 1362 * whether the file exists or not. 1363 * 1364 * The directory sticky bit is tested for NLC_DELETE and NLC_RENAME_DST, 1365 * the latter is only tested if the target exists. 1366 * 1367 * The passed ncp must be referenced and locked. If it is already resolved 1368 * it may be locked shared but otherwise should be locked exclusively. 1369 */ 1370 1371 #define S_WXOK_MASK (S_IRUSR|S_IXUSR|S_IRGRP|S_IXGRP|S_IROTH|S_IXOTH) 1372 #define S_XOK_MASK (S_IXUSR|S_IXGRP|S_IXOTH) 1373 1374 static int 1375 naccess(struct nchandle *nch, int nflags, struct ucred *cred, int *nflagsp, 1376 int nchislocked) 1377 { 1378 struct vnode *vp; 1379 struct vattr_lite lva; 1380 struct namecache *ncp; 1381 int error; 1382 int cflags; 1383 1384 KKASSERT(nchislocked == 0 || cache_lockstatus(nch) > 0); 1385 1386 ncp = nch->ncp; 1387 again: 1388 if (ncp->nc_flag & NCF_UNRESOLVED) { 1389 if (nchislocked == 0) { 1390 cache_lock(nch); 1391 nchislocked = 2; 1392 } 1393 cache_resolve(nch, cred); 1394 ncp = nch->ncp; 1395 } 1396 error = ncp->nc_error; 1397 1398 /* 1399 * Directory permissions checks. Silently ignore ENOENT if these 1400 * tests pass. It isn't an error. 1401 * 1402 * We can safely resolve ncp->nc_parent because ncp is currently 1403 * locked. 1404 */ 1405 if (nflags & (NLC_CREATE | NLC_DELETE | NLC_RENAME_SRC | NLC_RENAME_DST)) { 1406 if (((nflags & NLC_CREATE) && ncp->nc_vp == NULL) || 1407 ((nflags & NLC_DELETE) && ncp->nc_vp != NULL) || 1408 ((nflags & NLC_RENAME_SRC) && ncp->nc_vp != NULL) || 1409 (nflags & NLC_RENAME_DST) 1410 ) { 1411 struct nchandle par; 1412 1413 if (nchislocked == 0) { 1414 cache_lock_maybe_shared(nch, 0); 1415 nchislocked = 2; 1416 if (ncp->nc_flag & NCF_UNRESOLVED) 1417 goto again; 1418 } 1419 if ((par.ncp = ncp->nc_parent) == NULL) { 1420 if (error != EAGAIN) 1421 error = EINVAL; 1422 } else if (error == 0 || error == ENOENT) { 1423 par.mount = nch->mount; 1424 cache_hold(&par); 1425 cache_lock_maybe_shared(&par, 0); 1426 error = naccess(&par, NLC_WRITE, cred, NULL, 1); 1427 cache_put(&par); 1428 } 1429 } 1430 } 1431 1432 /* 1433 * NLC_EXCL check. Target file must not exist. 1434 */ 1435 if (error == 0 && (nflags & NLC_EXCL) && ncp->nc_vp != NULL) 1436 error = EEXIST; 1437 1438 /* 1439 * Try to short-cut the vnode operation for intermediate directory 1440 * components. This is a major SMP win because it avoids having 1441 * to execute a lot of code for intermediate directory components, 1442 * including shared refs and locks on intermediate directory vnodes. 1443 * 1444 * We can only do this if the caller does not need nflagsp. 1445 */ 1446 if (error == 0 && nflagsp == NULL && 1447 nflags == NLC_EXEC && (ncp->nc_flag & NCF_WXOK)) { 1448 if (nchislocked == 2) 1449 cache_unlock(nch); 1450 return 0; 1451 } 1452 1453 /* 1454 * Get the vnode attributes so we can do the rest of our checks. 1455 * 1456 * NOTE: We only call naccess_lva() if the target exists. 1457 */ 1458 if (error == 0) { 1459 if (nchislocked == 0) { 1460 cache_lock_maybe_shared(nch, 0); 1461 nchislocked = 2; 1462 } 1463 #if 0 1464 error = cache_vget(nch, cred, LK_SHARED, &vp); 1465 #else 1466 error = cache_vref(nch, cred, &vp); 1467 #endif 1468 if (error == ENOENT) { 1469 /* 1470 * Silently zero-out ENOENT if creating or renaming 1471 * (rename target). It isn't an error. 1472 */ 1473 if (nflags & (NLC_CREATE | NLC_RENAME_DST)) 1474 error = 0; 1475 } else if (error == 0) { 1476 /* 1477 * Get the vnode attributes and check for illegal O_TRUNC 1478 * requests and read-only mounts. 1479 * 1480 * NOTE: You can still open devices on read-only mounts for 1481 * writing. 1482 * 1483 * NOTE: creates/deletes/renames are handled by the NLC_WRITE 1484 * check on the parent directory above. 1485 * 1486 * XXX cache the va in the namecache or in the vnode 1487 */ 1488 error = VOP_GETATTR_LITE(vp, &lva); 1489 if (error == 0 && (nflags & NLC_TRUNCATE)) { 1490 switch(lva.va_type) { 1491 case VREG: 1492 case VDATABASE: 1493 case VCHR: 1494 case VBLK: 1495 case VFIFO: 1496 break; 1497 case VDIR: 1498 error = EISDIR; 1499 break; 1500 default: 1501 error = EINVAL; 1502 break; 1503 } 1504 } 1505 if (error == 0 && (nflags & NLC_WRITE) && vp->v_mount && 1506 (vp->v_mount->mnt_flag & MNT_RDONLY) 1507 ) { 1508 switch(lva.va_type) { 1509 case VDIR: 1510 case VLNK: 1511 case VREG: 1512 case VDATABASE: 1513 error = EROFS; 1514 break; 1515 default: 1516 break; 1517 } 1518 } 1519 #if 0 1520 vput(vp); 1521 #else 1522 vrele(vp); 1523 #endif 1524 1525 /* 1526 * Check permissions based on file attributes. The passed 1527 * flags (*nflagsp) are modified with feedback based on 1528 * special attributes and requirements. 1529 */ 1530 if (error == 0) { 1531 /* 1532 * Adjust the returned (*nflagsp) if non-NULL. 1533 */ 1534 if (nflagsp) { 1535 if ((lva.va_mode & VSVTX) && lva.va_uid != cred->cr_uid) 1536 *nflagsp |= NLC_STICKY; 1537 if (lva.va_flags & APPEND) 1538 *nflagsp |= NLC_APPENDONLY; 1539 if (lva.va_flags & IMMUTABLE) 1540 *nflagsp |= NLC_IMMUTABLE; 1541 } 1542 1543 /* 1544 * NCF_WXOK can be set for world-searchable directories. 1545 * 1546 * XXX When we implement capabilities this code would also 1547 * need a cap check, or only set the flag if there are no 1548 * capabilities. 1549 */ 1550 cflags = 0; 1551 if (lva.va_type == VDIR && 1552 (lva.va_mode & S_WXOK_MASK) == S_WXOK_MASK) { 1553 cflags |= NCF_WXOK; 1554 } 1555 if ((lva.va_mode & S_XOK_MASK) == 0) 1556 cflags |= NCF_NOTX; 1557 1558 /* 1559 * Track swapcache management flags in the namecache. 1560 * 1561 * Calculate the flags based on the current vattr_lite info 1562 * and recalculate the inherited flags from the parent 1563 * (the original cache linkage may have occurred without 1564 * getattrs and thus have stale flags). 1565 */ 1566 if (lva.va_flags & SF_NOCACHE) 1567 cflags |= NCF_SF_NOCACHE; 1568 if (lva.va_flags & UF_CACHE) 1569 cflags |= NCF_UF_CACHE; 1570 if (ncp->nc_parent) { 1571 if (ncp->nc_parent->nc_flag & 1572 (NCF_SF_NOCACHE | NCF_SF_PNOCACHE)) { 1573 cflags |= NCF_SF_PNOCACHE; 1574 } 1575 if (ncp->nc_parent->nc_flag & 1576 (NCF_UF_CACHE | NCF_UF_PCACHE)) { 1577 cflags |= NCF_UF_PCACHE; 1578 } 1579 } 1580 1581 /* 1582 * We're not supposed to update nc_flag when holding a shared 1583 * lock, but we allow the case for certain flags. Note that 1584 * holding an exclusive lock allows updating nc_flag without 1585 * atomics. nc_flag is not allowe to be updated at all unless 1586 * a shared or exclusive lock is held. 1587 */ 1588 atomic_clear_short(&ncp->nc_flag, 1589 (NCF_SF_NOCACHE | NCF_UF_CACHE | 1590 NCF_SF_PNOCACHE | NCF_UF_PCACHE | 1591 NCF_WXOK | NCF_NOTX) & ~cflags); 1592 atomic_set_short(&ncp->nc_flag, cflags); 1593 1594 /* 1595 * Process general access. 1596 */ 1597 error = naccess_lva(&lva, nflags, cred); 1598 } 1599 } 1600 } 1601 if (nchislocked == 2) 1602 cache_unlock(nch); 1603 return(error); 1604 } 1605 1606 /* 1607 * Check the requested access against the given vattr using cred. 1608 */ 1609 int 1610 naccess_lva(struct vattr_lite *lvap, int nflags, struct ucred *cred) 1611 { 1612 int i; 1613 int vmode; 1614 1615 /* 1616 * Test the immutable bit. Creations, deletions, renames (source 1617 * or destination) are not allowed. chown/chmod/other is also not 1618 * allowed but is handled by SETATTR. Hardlinks to the immutable 1619 * file are allowed. 1620 * 1621 * If the directory is set to immutable then creations, deletions, 1622 * renames (source or dest) and hardlinks to files within the directory 1623 * are not allowed, and regular files opened through the directory may 1624 * not be written to or truncated (unless a special device). 1625 * 1626 * NOTE! New hardlinks to immutable files work but new hardlinks to 1627 * files, immutable or not, sitting inside an immutable directory are 1628 * not allowed. As always if the file is hardlinked via some other 1629 * path additional hardlinks may be possible even if the file is marked 1630 * immutable. The sysop needs to create a closure by checking the hard 1631 * link count. Once closure is achieved you are good, and security 1632 * scripts should check link counts anyway. 1633 * 1634 * Writes and truncations are only allowed on special devices. 1635 */ 1636 if ((lvap->va_flags & IMMUTABLE) || (nflags & NLC_IMMUTABLE)) { 1637 if ((nflags & NLC_IMMUTABLE) && (nflags & NLC_HLINK)) 1638 return (EPERM); 1639 if (nflags & (NLC_CREATE | NLC_DELETE | 1640 NLC_RENAME_SRC | NLC_RENAME_DST)) { 1641 return (EPERM); 1642 } 1643 if (nflags & (NLC_WRITE | NLC_TRUNCATE)) { 1644 switch(lvap->va_type) { 1645 case VDIR: 1646 return (EISDIR); 1647 case VLNK: 1648 case VREG: 1649 case VDATABASE: 1650 return (EPERM); 1651 default: 1652 break; 1653 } 1654 } 1655 } 1656 1657 /* 1658 * Test the no-unlink and append-only bits for opens, rename targets, 1659 * and deletions. These bits are not tested for creations or 1660 * rename sources. 1661 * 1662 * Unlike FreeBSD we allow a file with APPEND set to be renamed. 1663 * If you do not wish this you must also set NOUNLINK. 1664 * 1665 * If the governing directory is marked APPEND-only it implies 1666 * NOUNLINK for all entries in the directory. 1667 */ 1668 if (((lvap->va_flags & NOUNLINK) || (nflags & NLC_APPENDONLY)) && 1669 (nflags & (NLC_DELETE | NLC_RENAME_SRC | NLC_RENAME_DST)) 1670 ) { 1671 return (EPERM); 1672 } 1673 1674 /* 1675 * A file marked append-only may not be deleted but can be renamed. 1676 */ 1677 if ((lvap->va_flags & APPEND) && 1678 (nflags & (NLC_DELETE | NLC_RENAME_DST)) 1679 ) { 1680 return (EPERM); 1681 } 1682 1683 /* 1684 * A file marked append-only which is opened for writing must also 1685 * be opened O_APPEND. 1686 */ 1687 if ((lvap->va_flags & APPEND) && (nflags & (NLC_OPEN | NLC_TRUNCATE))) { 1688 if (nflags & NLC_TRUNCATE) 1689 return (EPERM); 1690 if ((nflags & (NLC_OPEN | NLC_WRITE)) == (NLC_OPEN | NLC_WRITE)) { 1691 if ((nflags & NLC_APPEND) == 0) 1692 return (EPERM); 1693 } 1694 } 1695 1696 /* 1697 * root gets universal access 1698 */ 1699 if (cred->cr_uid == 0) 1700 return(0); 1701 1702 /* 1703 * Check owner perms. 1704 * 1705 * If NLC_OWN is set the owner of the file is allowed no matter when 1706 * the owner-mode bits say (utimes). 1707 */ 1708 vmode = 0; 1709 if (nflags & NLC_READ) 1710 vmode |= S_IRUSR; 1711 if (nflags & NLC_WRITE) 1712 vmode |= S_IWUSR; 1713 if (nflags & NLC_EXEC) 1714 vmode |= S_IXUSR; 1715 1716 if (cred->cr_uid == lvap->va_uid) { 1717 if ((nflags & NLC_OWN) == 0) { 1718 if ((vmode & lvap->va_mode) != vmode) 1719 return(EACCES); 1720 } 1721 return(0); 1722 } 1723 1724 /* 1725 * If NLC_STICKY is set only the owner may delete or rename a file. 1726 * This bit is typically set on /tmp. 1727 * 1728 * Note that the NLC_READ/WRITE/EXEC bits are not typically set in 1729 * the specific delete or rename case. For deletions and renames we 1730 * usually just care about directory permissions, not file permissions. 1731 */ 1732 if ((nflags & NLC_STICKY) && 1733 (nflags & (NLC_RENAME_SRC | NLC_RENAME_DST | NLC_DELETE))) { 1734 return(EACCES); 1735 } 1736 1737 /* 1738 * Check group perms 1739 */ 1740 vmode >>= 3; 1741 for (i = 0; i < cred->cr_ngroups; ++i) { 1742 if (lvap->va_gid == cred->cr_groups[i]) { 1743 if ((vmode & lvap->va_mode) != vmode) 1744 return(EACCES); 1745 return(0); 1746 } 1747 } 1748 1749 /* 1750 * Check world perms 1751 */ 1752 vmode >>= 3; 1753 if ((vmode & lvap->va_mode) != vmode) 1754 return(EACCES); 1755 return(0); 1756 } 1757 1758 /* 1759 * Long-term (10-second interval) statistics collection 1760 */ 1761 static 1762 uint64_t 1763 collect_nlookup_callback(int n) 1764 { 1765 static uint64_t last_total; 1766 uint64_t save; 1767 uint64_t total; 1768 1769 total = 0; 1770 for (n = 0; n < ncpus; ++n) { 1771 globaldata_t gd = globaldata_find(n); 1772 struct nchstats *sp; 1773 1774 if ((sp = gd->gd_nchstats) != NULL) 1775 total += sp->ncs_longhits + sp->ncs_longmiss; 1776 } 1777 save = total; 1778 total = total - last_total; 1779 last_total = save; 1780 1781 return total; 1782 } 1783 1784 static 1785 void 1786 nlookup_collect_init(void *dummy __unused) 1787 { 1788 kcollect_register(KCOLLECT_NLOOKUP, "nlookup", collect_nlookup_callback, 1789 KCOLLECT_SCALE(KCOLLECT_NLOOKUP_FORMAT, 0)); 1790 } 1791 SYSINIT(collect_nlookup, SI_SUB_PROP, SI_ORDER_ANY, nlookup_collect_init, 0); 1792