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 /* 896 * POSIX junk 897 */ 898 if (nd->nl_flags & NLC_CREATE) 899 error = EEXIST; 900 else if (nd->nl_flags & NLC_DELETE) 901 error = (wasdotordotdot == 1) ? EINVAL : ENOTEMPTY; 902 else 903 error = EINVAL; 904 } 905 } 906 907 /* 908 * Early completion on error. 909 */ 910 if (error) { 911 if (last_element) 912 cache_unlock(&nch); 913 cache_drop_and_cache(&nch, nd->nl_elmno); 914 break; 915 } 916 917 /* 918 * If the element is a symlink and it is either not the last 919 * element or it is the last element and we are allowed to 920 * follow symlinks, resolve the symlink. 921 */ 922 if ((nch.ncp->nc_flag & NCF_ISSYMLINK) && 923 (*ptr || (nd->nl_flags & NLC_FOLLOW)) 924 ) { 925 if (nd->nl_loopcnt++ >= MAXSYMLINKS) { 926 error = ELOOP; 927 if (last_element) 928 cache_unlock(&nch); 929 cache_drop_and_cache(&nch, nd->nl_elmno); 930 break; 931 } 932 if (last_element == 0) 933 cache_lock_maybe_shared(&nch, 1); 934 935 error = nreadsymlink(nd, &nch, &nlc); 936 cache_put(&nch); 937 if (error) 938 break; 939 940 /* 941 * Concatenate trailing path elements onto the returned symlink. 942 * Note that if the path component (ptr) is not exhausted, it 943 * will being with a '/', so we do not have to add another one. 944 * 945 * The symlink may not be empty. 946 */ 947 len = strlen(ptr); 948 if (nlc.nlc_namelen == 0 || nlc.nlc_namelen + len >= MAXPATHLEN) { 949 error = nlc.nlc_namelen ? ENAMETOOLONG : ENOENT; 950 objcache_put(namei_oc, nlc.nlc_nameptr); 951 break; 952 } 953 bcopy(ptr, nlc.nlc_nameptr + nlc.nlc_namelen, len + 1); 954 955 if (path_reset) { 956 if (nd->nl_flags & NLC_HASBUF) 957 objcache_put(namei_oc, nd->nl_path); 958 } else { 959 path_reset = nd->nl_path; 960 } 961 nd->nl_path = nlc.nlc_nameptr; 962 nd->nl_flags |= NLC_HASBUF; 963 ptr = nd->nl_path; 964 965 /* 966 * Go back up to the top to resolve any initial '/'s in the 967 * symlink. 968 */ 969 continue; 970 } 971 972 /* 973 * If the element is a directory and we are crossing a mount point, 974 * Locate the mount. 975 */ 976 while ((nch.ncp->nc_flag & NCF_ISMOUNTPT) && 977 (nd->nl_flags & NLC_NOCROSSMOUNT) == 0 && 978 (mp = cache_findmount(&nch)) != NULL 979 ) { 980 struct vnode *tdp; 981 int vfs_do_busy = 0; 982 983 /* 984 * VFS must be busied before the namecache entry is locked, 985 * but we don't want to waste time calling vfs_busy() if the 986 * mount point is already resolved. 987 */ 988 again: 989 if (last_element) 990 cache_unlock(&nch); 991 cache_drop_and_cache(&nch, nd->nl_elmno); 992 993 if (vfs_do_busy) { 994 while (vfs_busy(mp, 0)) { 995 if (mp->mnt_kern_flag & MNTK_UNMOUNT) { 996 kprintf("nlookup: warning umount race avoided\n"); 997 cache_dropmount(mp); 998 error = EBUSY; 999 vfs_do_busy = 0; 1000 goto double_break; 1001 } 1002 } 1003 } 1004 1005 /* 1006 * We don't need to lock the nch unless the entry is unresolved 1007 * or this is the last element. 1008 */ 1009 if (last_element) 1010 cache_get_maybe_shared(&mp->mnt_ncmountpt, &nch, 1011 wantsexcllock(nd, 1)); 1012 else 1013 cache_copy(&mp->mnt_ncmountpt, &nch); 1014 1015 if (nch.ncp->nc_flag & NCF_UNRESOLVED) { 1016 if (last_element == 0) 1017 cache_lock(&nch); 1018 if (nch.ncp->nc_flag & NCF_UNRESOLVED) { 1019 if (vfs_do_busy == 0) { 1020 vfs_do_busy = 1; 1021 if (last_element == 0) 1022 cache_unlock(&nch); 1023 goto again; 1024 } 1025 error = VFS_ROOT(mp, &tdp); 1026 vfs_unbusy(mp); 1027 vfs_do_busy = 0; 1028 if (keeperror(nd, error)) { 1029 cache_dropmount(mp); 1030 if (last_element == 0) 1031 cache_unlock(&nch); 1032 break; 1033 } 1034 if (error == 0) { 1035 cache_setvp(&nch, tdp); 1036 vput(tdp); 1037 } 1038 } 1039 if (last_element == 0) 1040 cache_unlock(&nch); 1041 } 1042 if (vfs_do_busy) 1043 vfs_unbusy(mp); 1044 cache_dropmount(mp); 1045 } 1046 1047 /* 1048 * Break out on error 1049 */ 1050 if (keeperror(nd, error)) { 1051 if (last_element) 1052 cache_unlock(&nch); 1053 cache_drop_and_cache(&nch, nd->nl_elmno); 1054 double_break: 1055 break; 1056 } 1057 1058 /* 1059 * Skip any slashes to get to the next element. If there 1060 * are any slashes at all the current element must be a 1061 * directory or, in the create case, intended to become a directory. 1062 * If it isn't we break without incrementing ptr and fall through 1063 * to the failure case below. 1064 */ 1065 while (*ptr == '/') { 1066 if ((nch.ncp->nc_flag & NCF_ISDIR) == 0 && 1067 !(nd->nl_flags & NLC_WILLBEDIR) 1068 ) { 1069 break; 1070 } 1071 ++ptr; 1072 } 1073 1074 /* 1075 * Continuation case: additional elements and the current 1076 * element is a directory. 1077 */ 1078 if (*ptr && (nch.ncp->nc_flag & NCF_ISDIR)) { 1079 cache_drop_and_cache(&nd->nl_nch, nd->nl_elmno); 1080 if (last_element) 1081 cache_unlock(&nch); 1082 /*nchislocked = 0; not needed */ 1083 KKASSERT((nd->nl_flags & NLC_NCPISLOCKED) == 0); 1084 nd->nl_nch = nch; 1085 continue; 1086 } 1087 1088 /* 1089 * Failure case: additional elements and the current element 1090 * is not a directory 1091 */ 1092 if (*ptr) { 1093 if (last_element) 1094 cache_unlock(&nch); 1095 cache_drop_and_cache(&nch, nd->nl_elmno); 1096 error = ENOTDIR; 1097 break; 1098 } 1099 1100 /* 1101 * Successful lookup of last element. 1102 * 1103 * Check permissions if the target exists. If the target does not 1104 * exist directory permissions were already tested in the early 1105 * completion code above. 1106 * 1107 * nd->nl_flags will be adjusted on return with NLC_APPENDONLY 1108 * if the file is marked append-only, and NLC_STICKY if the directory 1109 * containing the file is sticky. 1110 */ 1111 KKASSERT(last_element); 1112 1113 if (nch.ncp->nc_vp && (nd->nl_flags & NLC_ALLCHKS)) { 1114 error = naccess(&nch, nd->nl_flags | dflags, 1115 nd->nl_cred, NULL, 1); 1116 if (keeperror(nd, error)) { 1117 cache_put(&nch); 1118 break; 1119 } 1120 } 1121 1122 /* 1123 * Termination: no more elements. 1124 * 1125 * Check to see if the immediate parent has been destroyed. This race 1126 * can occur because the element lookup must temporarily unlock 1127 * the parent. If so, do a retry. 1128 */ 1129 if (nch.ncp->nc_parent && 1130 (nch.ncp->nc_parent->nc_flag & NCF_DESTROYED)) { 1131 doretry = TRUE; 1132 } 1133 1134 /* 1135 * Termination: no more elements. 1136 * 1137 * If NLC_REFDVP is set acquire a referenced parent dvp. Typically 1138 * used for mkdir/mknod/ncreate/nremove/unlink/rename. 1139 * 1140 * If a mount-point transition occurs due to ncp being a mount point, 1141 * or a null-mount, nl_dvp will be set to NULL and an error code of 1142 * 0 will be returned. A NULL nc_parent is not necessarily the only 1143 * indication of a mount-point as null-mounts will also tend to have 1144 * a non-null nc_parent. 1145 * 1146 * nch is locked, standard lock order for the namecache is 1147 * child-to-parent so we can safely lock its parent. We can 1148 * just use cache_dvpref(). 1149 */ 1150 if ((nd->nl_flags & NLC_REFDVP) && 1151 (doretry == FALSE || inretry == TRUE)) { 1152 if (nch.ncp->nc_parent) { 1153 error = cache_resolve_dvp(&nch, nd->nl_cred, 1154 &nd->nl_dvp); 1155 if (error) { 1156 kprintf("Parent directory lost during " 1157 "nlookup: %s/%s (%08x/%08x)\n", 1158 nch.ncp->nc_parent->nc_name, 1159 nch.ncp->nc_name, 1160 nch.ncp->nc_parent->nc_flag, 1161 nch.ncp->nc_flag); 1162 cache_put(&nch); 1163 error = EINVAL; 1164 break; 1165 } 1166 1167 /* 1168 * Mount-point, nl_dvp should remain NULL, error 0, 1169 * caller won't be able to use the results so leave 1170 * the ncp referenced but unlocked. 1171 */ 1172 if (nd->nl_dvp == NULL) { 1173 cache_put(&nch); 1174 break; 1175 } 1176 1177 /* 1178 * Good directory, fall through to drop-and-cache 1179 * below 1180 */ 1181 /* */ 1182 } else { 1183 /* 1184 * Mount-point, nl_dvp should remain NULL, error 0, 1185 * caller won't be able to use the results so leave 1186 * the ncp referenced but unlocked. 1187 */ 1188 error = 0; 1189 cache_put(&nch); 1190 break; 1191 } 1192 } 1193 1194 /* 1195 * ncp left with lock+ref on break, set NLC_NCPISLOCKED flag 1196 */ 1197 cache_drop_and_cache(&nd->nl_nch, nd->nl_elmno); 1198 nd->nl_nch = nch; 1199 nd->nl_flags |= NLC_NCPISLOCKED; 1200 error = 0; 1201 break; 1202 } 1203 1204 /* 1205 * We are done / or possibly retry 1206 */ 1207 if (hit) 1208 ++gd->gd_nchstats->ncs_longhits; 1209 else 1210 ++gd->gd_nchstats->ncs_longmiss; 1211 1212 if (nd->nl_flags & NLC_NCPISLOCKED) 1213 KKASSERT(cache_lockstatus(&nd->nl_nch) > 0); 1214 1215 /* 1216 * Reset nd->nl_path if necessary (due to softlinks). We want to return 1217 * nl_path to its original state before retrying or returning. 1218 */ 1219 if (path_reset) { 1220 if (nd->nl_flags & NLC_HASBUF) { 1221 objcache_put(namei_oc, nd->nl_path); 1222 nd->nl_flags &= ~NLC_HASBUF; 1223 } 1224 nd->nl_path = path_reset; 1225 nd->nl_flags |= saveflag & NLC_HASBUF; 1226 path_reset = NULL; 1227 } 1228 1229 /* 1230 * Retry the whole thing if doretry flag is set, but only once. 1231 * 1232 * autofs(5) may mount another filesystem under its root directory 1233 * while resolving a path. 1234 * 1235 * NFS might return ESTALE 1236 */ 1237 if (doretry && !inretry) { 1238 kprintf("nlookup: errno %d retry %s\n", error, nd->nl_path); 1239 inretry = TRUE; 1240 1241 /* 1242 * Clean up nd->nl_nch and reset to base directory 1243 */ 1244 if (nd->nl_flags & NLC_NCPISLOCKED) { 1245 cache_unlock(&nd->nl_nch); 1246 nd->nl_flags &= ~NLC_NCPISLOCKED; 1247 } 1248 cache_drop(&nd->nl_nch); 1249 cache_copy(nd->nl_basench, &nd->nl_nch); 1250 1251 nd->nl_elmno = 0; 1252 nd->nl_flags |= saveflag; 1253 1254 goto nlookup_start; 1255 } 1256 1257 /* 1258 * NOTE: If NLC_CREATE was set the ncp may represent a negative hit 1259 * (ncp->nc_error will be ENOENT), but we will still return an error 1260 * code of 0. 1261 */ 1262 return(error); 1263 } 1264 1265 /* 1266 * Resolve a mount point's glue ncp. This ncp connects creates the illusion 1267 * of continuity in the namecache tree by connecting the ncp related to the 1268 * vnode under the mount to the ncp related to the mount's root vnode. 1269 * 1270 * If no error occured a locked, ref'd ncp is stored in *ncpp. 1271 */ 1272 int 1273 nlookup_mp(struct mount *mp, struct nchandle *nch) 1274 { 1275 struct vnode *vp; 1276 int error; 1277 1278 error = 0; 1279 cache_get(&mp->mnt_ncmountpt, nch); 1280 if (nch->ncp->nc_flag & NCF_UNRESOLVED) { 1281 while (vfs_busy(mp, 0)) 1282 ; 1283 error = VFS_ROOT(mp, &vp); 1284 vfs_unbusy(mp); 1285 if (error) { 1286 cache_put(nch); 1287 } else { 1288 cache_setvp(nch, vp); 1289 vput(vp); 1290 } 1291 } 1292 return(error); 1293 } 1294 1295 /* 1296 * Read the contents of a symlink, allocate a path buffer out of the 1297 * namei_oc and initialize the supplied nlcomponent with the result. 1298 * 1299 * If an error occurs no buffer will be allocated or returned in the nlc. 1300 */ 1301 int 1302 nreadsymlink(struct nlookupdata *nd, struct nchandle *nch, 1303 struct nlcomponent *nlc) 1304 { 1305 struct vnode *vp; 1306 struct iovec aiov; 1307 struct uio auio; 1308 int linklen; 1309 int error; 1310 char *cp; 1311 1312 nlc->nlc_nameptr = NULL; 1313 nlc->nlc_namelen = 0; 1314 if (nch->ncp->nc_vp == NULL) 1315 return(ENOENT); 1316 if ((error = cache_vget(nch, nd->nl_cred, LK_SHARED, &vp)) != 0) 1317 return(error); 1318 cp = objcache_get(namei_oc, M_WAITOK); 1319 aiov.iov_base = cp; 1320 aiov.iov_len = MAXPATHLEN; 1321 auio.uio_iov = &aiov; 1322 auio.uio_iovcnt = 1; 1323 auio.uio_offset = 0; 1324 auio.uio_rw = UIO_READ; 1325 auio.uio_segflg = UIO_SYSSPACE; 1326 auio.uio_td = nd->nl_td; 1327 auio.uio_resid = MAXPATHLEN - 1; 1328 error = VOP_READLINK(vp, &auio, nd->nl_cred); 1329 if (error) 1330 goto fail; 1331 linklen = MAXPATHLEN - 1 - auio.uio_resid; 1332 if (varsym_enable) { 1333 linklen = varsymreplace(cp, linklen, MAXPATHLEN - 1); 1334 if (linklen < 0) { 1335 error = ENAMETOOLONG; 1336 goto fail; 1337 } 1338 } 1339 cp[linklen] = 0; 1340 nlc->nlc_nameptr = cp; 1341 nlc->nlc_namelen = linklen; 1342 vput(vp); 1343 return(0); 1344 fail: 1345 objcache_put(namei_oc, cp); 1346 vput(vp); 1347 return(error); 1348 } 1349 1350 /* 1351 * Check access [XXX cache vattr!] [XXX quota] 1352 * 1353 * Generally check the NLC_* access bits. All specified bits must pass 1354 * for this function to return 0. 1355 * 1356 * The file does not have to exist when checking NLC_CREATE or NLC_RENAME_DST 1357 * access, otherwise it must exist. No error is returned in this case. 1358 * 1359 * The file must not exist if NLC_EXCL is specified. 1360 * 1361 * Directory permissions in general are tested for NLC_CREATE if the file 1362 * does not exist, NLC_DELETE if the file does exist, and NLC_RENAME_DST 1363 * whether the file exists or not. 1364 * 1365 * The directory sticky bit is tested for NLC_DELETE and NLC_RENAME_DST, 1366 * the latter is only tested if the target exists. 1367 * 1368 * The passed ncp must be referenced and locked. If it is already resolved 1369 * it may be locked shared but otherwise should be locked exclusively. 1370 */ 1371 1372 #define S_WXOK_MASK (S_IRUSR|S_IXUSR|S_IRGRP|S_IXGRP|S_IROTH|S_IXOTH) 1373 #define S_XOK_MASK (S_IXUSR|S_IXGRP|S_IXOTH) 1374 1375 static int 1376 naccess(struct nchandle *nch, int nflags, struct ucred *cred, int *nflagsp, 1377 int nchislocked) 1378 { 1379 struct vnode *vp; 1380 struct vattr_lite lva; 1381 struct namecache *ncp; 1382 int error; 1383 int cflags; 1384 int ls = cache_lockstatus(nch); 1385 1386 /* 1387 * nchislocked: 0 not locked by caller, might not be resolved 1388 * 1 locked by caller shared or exclusive & resolved 1389 * 2 locked by us shared or exclusive & resolved 1390 */ 1391 KKASSERT(nchislocked == 0 || ls > 0); 1392 1393 ncp = nch->ncp; 1394 1395 /* 1396 * If the ncp is exclusively held by another process our ref may 1397 * have raced an eviction and we cannot safely test the fields 1398 * unlocked. 1399 */ 1400 if (ls < 0 && nchislocked == 0) { 1401 cache_lock(nch); 1402 nchislocked = 2; 1403 ls = LK_EXCLUSIVE; 1404 } 1405 1406 again: 1407 /* 1408 * We need a resolved entry. If the entry is not resolved we need 1409 * to lock and resolve it. If it is already resolved, our ref should 1410 * prevent normal evictions (as long as we tested the lock race above). 1411 * 1412 * If the ncp was locked by the caller and left unresolved, it must 1413 * have been locked exclusively. 1414 */ 1415 if (ncp->nc_flag & NCF_UNRESOLVED) { 1416 if (nchislocked == 0) { 1417 cache_lock(nch); 1418 nchislocked = 2; 1419 ls = LK_EXCLUSIVE; 1420 } 1421 KKASSERT(ls == LK_EXCLUSIVE); 1422 cache_resolve(nch, cred); 1423 ncp = nch->ncp; 1424 } 1425 error = ncp->nc_error; 1426 1427 /* 1428 * Only unresolved entries should return this error (though maybe 1429 * in-filesystem sockets can too). XXX check filetype for VSOCK. 1430 */ 1431 if (error == ENOTCONN) { 1432 if (nchislocked == 0) { 1433 kprintf("ncp %p %08x %d %s: Warning, unexpected state, " 1434 "forcing lock\n", 1435 ncp, ncp->nc_flag, ncp->nc_error, ncp->nc_name); 1436 print_backtrace(-1); 1437 cache_lock(nch); 1438 nchislocked = 2; 1439 ls = LK_EXCLUSIVE; 1440 goto again; 1441 } 1442 kprintf("ncp %p %08x %d %s: Warning, unexpected state\n", 1443 ncp, ncp->nc_flag, ncp->nc_error, ncp->nc_name); 1444 print_backtrace(-1); 1445 } 1446 1447 /* 1448 * Directory permissions checks. Silently ignore ENOENT if these 1449 * tests pass. It isn't an error. 1450 * 1451 * We can safely resolve ncp->nc_parent because ncp is currently 1452 * locked. 1453 */ 1454 if (nflags & (NLC_CREATE | NLC_DELETE | NLC_RENAME_SRC | NLC_RENAME_DST)) { 1455 if (((nflags & NLC_CREATE) && ncp->nc_vp == NULL) || 1456 ((nflags & NLC_DELETE) && ncp->nc_vp != NULL) || 1457 ((nflags & NLC_RENAME_SRC) && ncp->nc_vp != NULL) || 1458 (nflags & NLC_RENAME_DST) 1459 ) { 1460 struct nchandle par; 1461 1462 if (nchislocked == 0) { 1463 cache_lock_maybe_shared(nch, 0); 1464 nchislocked = 2; 1465 ls = LK_EXCLUSIVE; 1466 goto again; 1467 } 1468 if ((par.ncp = ncp->nc_parent) == NULL) { 1469 if (error != EAGAIN) 1470 error = EINVAL; 1471 } else if (error == 0 || error == ENOENT) { 1472 par.mount = nch->mount; 1473 cache_hold(&par); 1474 cache_lock_maybe_shared(&par, 0); 1475 error = naccess(&par, NLC_WRITE, cred, NULL, 1); 1476 cache_put(&par); 1477 } 1478 } 1479 } 1480 1481 /* 1482 * NLC_EXCL check. Target file must not exist. 1483 */ 1484 if (error == 0 && (nflags & NLC_EXCL) && ncp->nc_vp != NULL) 1485 error = EEXIST; 1486 1487 /* 1488 * Try to short-cut the vnode operation for intermediate directory 1489 * components. This is a major SMP win because it avoids having 1490 * to execute a lot of code for intermediate directory components, 1491 * including shared refs and locks on intermediate directory vnodes. 1492 * 1493 * We can only do this if the caller does not need nflagsp. 1494 */ 1495 if (error == 0 && nflagsp == NULL && 1496 nflags == NLC_EXEC && (ncp->nc_flag & NCF_WXOK)) { 1497 if (nchislocked == 2) 1498 cache_unlock(nch); 1499 return 0; 1500 } 1501 1502 /* 1503 * Get the vnode attributes so we can do the rest of our checks. 1504 * 1505 * NOTE: We only call naccess_lva() if the target exists. 1506 */ 1507 if (error == 0) { 1508 if (nchislocked == 0) { 1509 cache_lock_maybe_shared(nch, 0); 1510 nchislocked = 2; 1511 } 1512 #if 0 1513 error = cache_vget(nch, cred, LK_SHARED, &vp); 1514 #else 1515 error = cache_vref(nch, cred, &vp); 1516 #endif 1517 if (error == ENOENT) { 1518 /* 1519 * Silently zero-out ENOENT if creating or renaming 1520 * (rename target). It isn't an error. 1521 */ 1522 if (nflags & (NLC_CREATE | NLC_RENAME_DST)) 1523 error = 0; 1524 } else if (error == 0) { 1525 /* 1526 * Get the vnode attributes and check for illegal O_TRUNC 1527 * requests and read-only mounts. 1528 * 1529 * NOTE: You can still open devices on read-only mounts for 1530 * writing. 1531 * 1532 * NOTE: creates/deletes/renames are handled by the NLC_WRITE 1533 * check on the parent directory above. 1534 * 1535 * XXX cache the va in the namecache or in the vnode 1536 */ 1537 error = VOP_GETATTR_LITE(vp, &lva); 1538 if (error == 0 && (nflags & NLC_TRUNCATE)) { 1539 switch(lva.va_type) { 1540 case VREG: 1541 case VDATABASE: 1542 case VCHR: 1543 case VBLK: 1544 case VFIFO: 1545 break; 1546 case VDIR: 1547 error = EISDIR; 1548 break; 1549 default: 1550 error = EINVAL; 1551 break; 1552 } 1553 } 1554 if (error == 0 && (nflags & NLC_WRITE) && vp->v_mount && 1555 (vp->v_mount->mnt_flag & MNT_RDONLY) 1556 ) { 1557 switch(lva.va_type) { 1558 case VDIR: 1559 case VLNK: 1560 case VREG: 1561 case VDATABASE: 1562 error = EROFS; 1563 break; 1564 default: 1565 break; 1566 } 1567 } 1568 #if 0 1569 vput(vp); 1570 #else 1571 vrele(vp); 1572 #endif 1573 1574 /* 1575 * Check permissions based on file attributes. The passed 1576 * flags (*nflagsp) are modified with feedback based on 1577 * special attributes and requirements. 1578 */ 1579 if (error == 0) { 1580 /* 1581 * Adjust the returned (*nflagsp) if non-NULL. 1582 */ 1583 if (nflagsp) { 1584 if ((lva.va_mode & VSVTX) && lva.va_uid != cred->cr_uid) 1585 *nflagsp |= NLC_STICKY; 1586 if (lva.va_flags & APPEND) 1587 *nflagsp |= NLC_APPENDONLY; 1588 if (lva.va_flags & IMMUTABLE) 1589 *nflagsp |= NLC_IMMUTABLE; 1590 } 1591 1592 /* 1593 * NCF_WXOK can be set for world-searchable directories. 1594 * 1595 * XXX When we implement capabilities this code would also 1596 * need a cap check, or only set the flag if there are no 1597 * capabilities. 1598 */ 1599 cflags = 0; 1600 if (lva.va_type == VDIR && 1601 (lva.va_mode & S_WXOK_MASK) == S_WXOK_MASK) { 1602 cflags |= NCF_WXOK; 1603 } 1604 if ((lva.va_mode & S_XOK_MASK) == 0) 1605 cflags |= NCF_NOTX; 1606 1607 /* 1608 * Track swapcache management flags in the namecache. 1609 * 1610 * Calculate the flags based on the current vattr_lite info 1611 * and recalculate the inherited flags from the parent 1612 * (the original cache linkage may have occurred without 1613 * getattrs and thus have stale flags). 1614 */ 1615 if (lva.va_flags & SF_NOCACHE) 1616 cflags |= NCF_SF_NOCACHE; 1617 if (lva.va_flags & UF_CACHE) 1618 cflags |= NCF_UF_CACHE; 1619 if (ncp->nc_parent) { 1620 if (ncp->nc_parent->nc_flag & 1621 (NCF_SF_NOCACHE | NCF_SF_PNOCACHE)) { 1622 cflags |= NCF_SF_PNOCACHE; 1623 } 1624 if (ncp->nc_parent->nc_flag & 1625 (NCF_UF_CACHE | NCF_UF_PCACHE)) { 1626 cflags |= NCF_UF_PCACHE; 1627 } 1628 } 1629 1630 /* 1631 * We're not supposed to update nc_flag when holding a shared 1632 * lock, but we allow the case for certain flags. Note that 1633 * holding an exclusive lock allows updating nc_flag without 1634 * atomics. nc_flag is not allowe to be updated at all unless 1635 * a shared or exclusive lock is held. 1636 */ 1637 atomic_clear_short(&ncp->nc_flag, 1638 (NCF_SF_NOCACHE | NCF_UF_CACHE | 1639 NCF_SF_PNOCACHE | NCF_UF_PCACHE | 1640 NCF_WXOK | NCF_NOTX) & ~cflags); 1641 atomic_set_short(&ncp->nc_flag, cflags); 1642 1643 /* 1644 * Process general access. 1645 */ 1646 error = naccess_lva(&lva, nflags, cred); 1647 } 1648 } 1649 } 1650 if (nchislocked == 2) 1651 cache_unlock(nch); 1652 return(error); 1653 } 1654 1655 /* 1656 * Check the requested access against the given vattr using cred. 1657 */ 1658 int 1659 naccess_lva(struct vattr_lite *lvap, int nflags, struct ucred *cred) 1660 { 1661 int i; 1662 int vmode; 1663 1664 /* 1665 * Test the immutable bit. Creations, deletions, renames (source 1666 * or destination) are not allowed. chown/chmod/other is also not 1667 * allowed but is handled by SETATTR. Hardlinks to the immutable 1668 * file are allowed. 1669 * 1670 * If the directory is set to immutable then creations, deletions, 1671 * renames (source or dest) and hardlinks to files within the directory 1672 * are not allowed, and regular files opened through the directory may 1673 * not be written to or truncated (unless a special device). 1674 * 1675 * NOTE! New hardlinks to immutable files work but new hardlinks to 1676 * files, immutable or not, sitting inside an immutable directory are 1677 * not allowed. As always if the file is hardlinked via some other 1678 * path additional hardlinks may be possible even if the file is marked 1679 * immutable. The sysop needs to create a closure by checking the hard 1680 * link count. Once closure is achieved you are good, and security 1681 * scripts should check link counts anyway. 1682 * 1683 * Writes and truncations are only allowed on special devices. 1684 */ 1685 if ((lvap->va_flags & IMMUTABLE) || (nflags & NLC_IMMUTABLE)) { 1686 if ((nflags & NLC_IMMUTABLE) && (nflags & NLC_HLINK)) 1687 return (EPERM); 1688 if (nflags & (NLC_CREATE | NLC_DELETE | 1689 NLC_RENAME_SRC | NLC_RENAME_DST)) { 1690 return (EPERM); 1691 } 1692 if (nflags & (NLC_WRITE | NLC_TRUNCATE)) { 1693 switch(lvap->va_type) { 1694 case VDIR: 1695 return (EISDIR); 1696 case VLNK: 1697 case VREG: 1698 case VDATABASE: 1699 return (EPERM); 1700 default: 1701 break; 1702 } 1703 } 1704 } 1705 1706 /* 1707 * Test the no-unlink and append-only bits for opens, rename targets, 1708 * and deletions. These bits are not tested for creations or 1709 * rename sources. 1710 * 1711 * Unlike FreeBSD we allow a file with APPEND set to be renamed. 1712 * If you do not wish this you must also set NOUNLINK. 1713 * 1714 * If the governing directory is marked APPEND-only it implies 1715 * NOUNLINK for all entries in the directory. 1716 */ 1717 if (((lvap->va_flags & NOUNLINK) || (nflags & NLC_APPENDONLY)) && 1718 (nflags & (NLC_DELETE | NLC_RENAME_SRC | NLC_RENAME_DST)) 1719 ) { 1720 return (EPERM); 1721 } 1722 1723 /* 1724 * A file marked append-only may not be deleted but can be renamed. 1725 */ 1726 if ((lvap->va_flags & APPEND) && 1727 (nflags & (NLC_DELETE | NLC_RENAME_DST)) 1728 ) { 1729 return (EPERM); 1730 } 1731 1732 /* 1733 * A file marked append-only which is opened for writing must also 1734 * be opened O_APPEND. 1735 */ 1736 if ((lvap->va_flags & APPEND) && (nflags & (NLC_OPEN | NLC_TRUNCATE))) { 1737 if (nflags & NLC_TRUNCATE) 1738 return (EPERM); 1739 if ((nflags & (NLC_OPEN | NLC_WRITE)) == (NLC_OPEN | NLC_WRITE)) { 1740 if ((nflags & NLC_APPEND) == 0) 1741 return (EPERM); 1742 } 1743 } 1744 1745 /* 1746 * root gets universal access 1747 */ 1748 if (cred->cr_uid == 0) 1749 return(0); 1750 1751 /* 1752 * Check owner perms. 1753 * 1754 * If NLC_OWN is set the owner of the file is allowed no matter when 1755 * the owner-mode bits say (utimes). 1756 */ 1757 vmode = 0; 1758 if (nflags & NLC_READ) 1759 vmode |= S_IRUSR; 1760 if (nflags & NLC_WRITE) 1761 vmode |= S_IWUSR; 1762 if (nflags & NLC_EXEC) 1763 vmode |= S_IXUSR; 1764 1765 if (cred->cr_uid == lvap->va_uid) { 1766 if ((nflags & NLC_OWN) == 0) { 1767 if ((vmode & lvap->va_mode) != vmode) 1768 return(EACCES); 1769 } 1770 return(0); 1771 } 1772 1773 /* 1774 * If NLC_STICKY is set only the owner may delete or rename a file. 1775 * This bit is typically set on /tmp. 1776 * 1777 * Note that the NLC_READ/WRITE/EXEC bits are not typically set in 1778 * the specific delete or rename case. For deletions and renames we 1779 * usually just care about directory permissions, not file permissions. 1780 */ 1781 if ((nflags & NLC_STICKY) && 1782 (nflags & (NLC_RENAME_SRC | NLC_RENAME_DST | NLC_DELETE))) { 1783 return(EACCES); 1784 } 1785 1786 /* 1787 * Check group perms 1788 */ 1789 vmode >>= 3; 1790 for (i = 0; i < cred->cr_ngroups; ++i) { 1791 if (lvap->va_gid == cred->cr_groups[i]) { 1792 if ((vmode & lvap->va_mode) != vmode) 1793 return(EACCES); 1794 return(0); 1795 } 1796 } 1797 1798 /* 1799 * Check world perms 1800 */ 1801 vmode >>= 3; 1802 if ((vmode & lvap->va_mode) != vmode) 1803 return(EACCES); 1804 return(0); 1805 } 1806 1807 /* 1808 * Long-term (10-second interval) statistics collection 1809 */ 1810 static 1811 uint64_t 1812 collect_nlookup_callback(int n) 1813 { 1814 static uint64_t last_total; 1815 uint64_t save; 1816 uint64_t total; 1817 1818 total = 0; 1819 for (n = 0; n < ncpus; ++n) { 1820 globaldata_t gd = globaldata_find(n); 1821 struct nchstats *sp; 1822 1823 if ((sp = gd->gd_nchstats) != NULL) 1824 total += sp->ncs_longhits + sp->ncs_longmiss; 1825 } 1826 save = total; 1827 total = total - last_total; 1828 last_total = save; 1829 1830 return total; 1831 } 1832 1833 static 1834 void 1835 nlookup_collect_init(void *dummy __unused) 1836 { 1837 kcollect_register(KCOLLECT_NLOOKUP, "nlookup", collect_nlookup_callback, 1838 KCOLLECT_SCALE(KCOLLECT_NLOOKUP_FORMAT, 0)); 1839 } 1840 SYSINIT(collect_nlookup, SI_SUB_PROP, SI_ORDER_ANY, nlookup_collect_init, 0); 1841