1 /* 2 * Copyright (c) 2004 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/kernel.h> 57 #include <sys/vnode.h> 58 #include <sys/mount.h> 59 #include <sys/filedesc.h> 60 #include <sys/proc.h> 61 #include <sys/namei.h> 62 #include <sys/nlookup.h> 63 #include <sys/malloc.h> 64 #include <sys/stat.h> 65 #include <sys/objcache.h> 66 #include <sys/file.h> 67 68 #ifdef KTRACE 69 #include <sys/ktrace.h> 70 #endif 71 72 static int naccess(struct nchandle *nch, int vmode, struct ucred *cred, 73 int *stickyp); 74 75 /* 76 * Initialize a nlookup() structure, early error return for copyin faults 77 * or a degenerate empty string (which is not allowed). 78 * 79 * The first process proc0's credentials are used if the calling thread 80 * is not associated with a process context. 81 * 82 * MPSAFE 83 */ 84 int 85 nlookup_init(struct nlookupdata *nd, 86 const char *path, enum uio_seg seg, int flags) 87 { 88 size_t pathlen; 89 struct proc *p; 90 thread_t td; 91 int error; 92 93 td = curthread; 94 p = td->td_proc; 95 96 /* 97 * note: the pathlen set by copy*str() includes the terminating \0. 98 */ 99 bzero(nd, sizeof(struct nlookupdata)); 100 nd->nl_path = objcache_get(namei_oc, M_WAITOK); 101 nd->nl_flags |= NLC_HASBUF; 102 if (seg == UIO_SYSSPACE) 103 error = copystr(path, nd->nl_path, MAXPATHLEN, &pathlen); 104 else 105 error = copyinstr(path, nd->nl_path, MAXPATHLEN, &pathlen); 106 107 /* 108 * Don't allow empty pathnames. 109 * POSIX.1 requirement: "" is not a vaild file name. 110 */ 111 if (error == 0 && pathlen <= 1) 112 error = ENOENT; 113 114 if (error == 0) { 115 if (p && p->p_fd) { 116 cache_copy(&p->p_fd->fd_ncdir, &nd->nl_nch); 117 cache_copy(&p->p_fd->fd_nrdir, &nd->nl_rootnch); 118 if (p->p_fd->fd_njdir.ncp) 119 cache_copy(&p->p_fd->fd_njdir, &nd->nl_jailnch); 120 nd->nl_cred = crhold(p->p_ucred); 121 } else { 122 cache_copy(&rootnch, &nd->nl_nch); 123 cache_copy(&nd->nl_nch, &nd->nl_rootnch); 124 cache_copy(&nd->nl_nch, &nd->nl_jailnch); 125 nd->nl_cred = crhold(proc0.p_ucred); 126 } 127 nd->nl_td = td; 128 nd->nl_flags |= flags; 129 } else { 130 nlookup_done(nd); 131 } 132 return(error); 133 } 134 135 136 /* 137 * nlookup_init() for "at" family of syscalls. 138 * 139 * Works similarly to nlookup_init() but if path is relative and fd is not 140 * AT_FDCWD, path is interpreted relative to the directory pointed to by fd. 141 * In this case, the file entry pointed to by fd is ref'ed and returned in 142 * *fpp. 143 * 144 * If the call succeeds, nlookup_done_at() must be called to clean-up the nd 145 * and release the ref to the file entry. 146 */ 147 int 148 nlookup_init_at(struct nlookupdata *nd, struct file **fpp, int fd, 149 const char *path, enum uio_seg seg, int flags) 150 { 151 struct thread *td = curthread; 152 struct proc *p = td->td_proc; 153 struct file* fp; 154 struct vnode *vp; 155 int error; 156 157 *fpp = NULL; 158 159 if ((error = nlookup_init(nd, path, seg, flags)) != 0) { 160 return (error); 161 } 162 163 if (nd->nl_path[0] != '/' && fd != AT_FDCWD) { 164 if ((error = holdvnode(p->p_fd, fd, &fp)) != 0) 165 goto done; 166 vp = (struct vnode*)fp->f_data; 167 if (vp->v_type != VDIR || fp->f_nchandle.ncp == NULL) { 168 fdrop(fp); 169 fp = NULL; 170 error = ENOTDIR; 171 goto done; 172 } 173 cache_drop(&nd->nl_nch); 174 cache_copy(&fp->f_nchandle, &nd->nl_nch); 175 *fpp = fp; 176 } 177 178 179 done: 180 if (error) 181 nlookup_done(nd); 182 return (error); 183 184 } 185 186 /* 187 * This works similarly to nlookup_init() but does not assume a process 188 * context. rootnch is always chosen for the root directory and the cred 189 * and starting directory are supplied in arguments. 190 */ 191 int 192 nlookup_init_raw(struct nlookupdata *nd, 193 const char *path, enum uio_seg seg, int flags, 194 struct ucred *cred, struct nchandle *ncstart) 195 { 196 size_t pathlen; 197 thread_t td; 198 int error; 199 200 td = curthread; 201 202 bzero(nd, sizeof(struct nlookupdata)); 203 nd->nl_path = objcache_get(namei_oc, M_WAITOK); 204 nd->nl_flags |= NLC_HASBUF; 205 if (seg == UIO_SYSSPACE) 206 error = copystr(path, nd->nl_path, MAXPATHLEN, &pathlen); 207 else 208 error = copyinstr(path, nd->nl_path, MAXPATHLEN, &pathlen); 209 210 /* 211 * Don't allow empty pathnames. 212 * POSIX.1 requirement: "" is not a vaild file name. 213 */ 214 if (error == 0 && pathlen <= 1) 215 error = ENOENT; 216 217 if (error == 0) { 218 cache_copy(ncstart, &nd->nl_nch); 219 cache_copy(&rootnch, &nd->nl_rootnch); 220 cache_copy(&rootnch, &nd->nl_jailnch); 221 nd->nl_cred = crhold(cred); 222 nd->nl_td = td; 223 nd->nl_flags |= flags; 224 } else { 225 nlookup_done(nd); 226 } 227 return(error); 228 } 229 230 /* 231 * This works similarly to nlookup_init_raw() but does not rely 232 * on rootnch being initialized yet. 233 */ 234 int 235 nlookup_init_root(struct nlookupdata *nd, 236 const char *path, enum uio_seg seg, int flags, 237 struct ucred *cred, struct nchandle *ncstart, 238 struct nchandle *ncroot) 239 { 240 size_t pathlen; 241 thread_t td; 242 int error; 243 244 td = curthread; 245 246 bzero(nd, sizeof(struct nlookupdata)); 247 nd->nl_path = objcache_get(namei_oc, M_WAITOK); 248 nd->nl_flags |= NLC_HASBUF; 249 if (seg == UIO_SYSSPACE) 250 error = copystr(path, nd->nl_path, MAXPATHLEN, &pathlen); 251 else 252 error = copyinstr(path, nd->nl_path, MAXPATHLEN, &pathlen); 253 254 /* 255 * Don't allow empty pathnames. 256 * POSIX.1 requirement: "" is not a vaild file name. 257 */ 258 if (error == 0 && pathlen <= 1) 259 error = ENOENT; 260 261 if (error == 0) { 262 cache_copy(ncstart, &nd->nl_nch); 263 cache_copy(ncroot, &nd->nl_rootnch); 264 cache_copy(ncroot, &nd->nl_jailnch); 265 nd->nl_cred = crhold(cred); 266 nd->nl_td = td; 267 nd->nl_flags |= flags; 268 } else { 269 nlookup_done(nd); 270 } 271 return(error); 272 } 273 274 /* 275 * Set a different credential; this credential will be used by future 276 * operations performed on nd.nl_open_vp and nlookupdata structure. 277 */ 278 void 279 nlookup_set_cred(struct nlookupdata *nd, struct ucred *cred) 280 { 281 KKASSERT(nd->nl_cred != NULL); 282 283 if (nd->nl_cred != cred) { 284 cred = crhold(cred); 285 crfree(nd->nl_cred); 286 nd->nl_cred = cred; 287 } 288 } 289 290 /* 291 * Cleanup a nlookupdata structure after we are through with it. This may 292 * be called on any nlookupdata structure initialized with nlookup_init(). 293 * Calling nlookup_done() is mandatory in all cases except where nlookup_init() 294 * returns an error, even if as a consumer you believe you have taken all 295 * dynamic elements out of the nlookupdata structure. 296 */ 297 void 298 nlookup_done(struct nlookupdata *nd) 299 { 300 if (nd->nl_nch.ncp) { 301 if (nd->nl_flags & NLC_NCPISLOCKED) { 302 nd->nl_flags &= ~NLC_NCPISLOCKED; 303 cache_unlock(&nd->nl_nch); 304 } 305 cache_drop(&nd->nl_nch); /* NULL's out the nch */ 306 } 307 if (nd->nl_rootnch.ncp) 308 cache_drop(&nd->nl_rootnch); 309 if (nd->nl_jailnch.ncp) 310 cache_drop(&nd->nl_jailnch); 311 if ((nd->nl_flags & NLC_HASBUF) && nd->nl_path) { 312 objcache_put(namei_oc, nd->nl_path); 313 nd->nl_path = NULL; 314 } 315 if (nd->nl_cred) { 316 crfree(nd->nl_cred); 317 nd->nl_cred = NULL; 318 } 319 if (nd->nl_open_vp) { 320 if (nd->nl_flags & NLC_LOCKVP) { 321 vn_unlock(nd->nl_open_vp); 322 nd->nl_flags &= ~NLC_LOCKVP; 323 } 324 vn_close(nd->nl_open_vp, nd->nl_vp_fmode); 325 nd->nl_open_vp = NULL; 326 } 327 if (nd->nl_dvp) { 328 vrele(nd->nl_dvp); 329 nd->nl_dvp = NULL; 330 } 331 nd->nl_flags = 0; /* clear remaining flags (just clear everything) */ 332 } 333 334 /* 335 * Works similarly to nlookup_done() when nd initialized with 336 * nlookup_init_at(). 337 */ 338 void 339 nlookup_done_at(struct nlookupdata *nd, struct file *fp) 340 { 341 nlookup_done(nd); 342 if (fp != NULL) 343 fdrop(fp); 344 } 345 346 void 347 nlookup_zero(struct nlookupdata *nd) 348 { 349 bzero(nd, sizeof(struct nlookupdata)); 350 } 351 352 /* 353 * Simple all-in-one nlookup. Returns a locked namecache structure or NULL 354 * if an error occured. 355 * 356 * Note that the returned ncp is not checked for permissions, though VEXEC 357 * is checked on the directory path leading up to the result. The caller 358 * must call naccess() to check the permissions of the returned leaf. 359 */ 360 struct nchandle 361 nlookup_simple(const char *str, enum uio_seg seg, 362 int niflags, int *error) 363 { 364 struct nlookupdata nd; 365 struct nchandle nch; 366 367 *error = nlookup_init(&nd, str, seg, niflags); 368 if (*error == 0) { 369 if ((*error = nlookup(&nd)) == 0) { 370 nch = nd.nl_nch; /* keep hold ref from structure */ 371 cache_zero(&nd.nl_nch); /* and NULL out */ 372 } else { 373 cache_zero(&nch); 374 } 375 nlookup_done(&nd); 376 } else { 377 cache_zero(&nch); 378 } 379 return(nch); 380 } 381 382 /* 383 * Do a generic nlookup. Note that the passed nd is not nlookup_done()'d 384 * on return, even if an error occurs. If no error occurs or NLC_CREATE 385 * is flagged and ENOENT is returned, then the returned nl_nch is always 386 * referenced and locked exclusively. 387 * 388 * WARNING: For any general error other than ENOENT w/NLC_CREATE, the 389 * the resulting nl_nch may or may not be locked and if locked 390 * might be locked either shared or exclusive. 391 * 392 * Intermediate directory elements, including the current directory, require 393 * execute (search) permission. nlookup does not examine the access 394 * permissions on the returned element. 395 * 396 * If NLC_CREATE is set the last directory must allow node creation, 397 * and an error code of 0 will be returned for a non-existant 398 * target (not ENOENT). 399 * 400 * If NLC_RENAME_DST is set the last directory mut allow node deletion, 401 * plus the sticky check is made, and an error code of 0 will be returned 402 * for a non-existant target (not ENOENT). 403 * 404 * If NLC_DELETE is set the last directory mut allow node deletion, 405 * plus the sticky check is made. 406 * 407 * If NLC_REFDVP is set nd->nl_dvp will be set to the directory vnode 408 * of the returned entry. The vnode will be referenced, but not locked, 409 * and will be released by nlookup_done() along with everything else. 410 * 411 * NOTE: As an optimization we attempt to obtain a shared namecache lock 412 * on any intermediate elements. On success, the returned element 413 * is ALWAYS locked exclusively. 414 */ 415 static 416 int 417 islastelement(const char *ptr) 418 { 419 while (*ptr == '/') 420 ++ptr; 421 return (*ptr == 0); 422 } 423 424 int 425 nlookup(struct nlookupdata *nd) 426 { 427 globaldata_t gd = mycpu; 428 struct nlcomponent nlc; 429 struct nchandle nch; 430 struct nchandle par; 431 struct nchandle nctmp; 432 struct mount *mp; 433 struct vnode *hvp; /* hold to prevent recyclement */ 434 int wasdotordotdot; 435 char *ptr; 436 int error; 437 int len; 438 int dflags; 439 int hit = 1; 440 441 #ifdef KTRACE 442 if (KTRPOINT(nd->nl_td, KTR_NAMEI)) 443 ktrnamei(nd->nl_td->td_lwp, nd->nl_path); 444 #endif 445 bzero(&nlc, sizeof(nlc)); 446 447 /* 448 * Setup for the loop. The current working namecache element is 449 * always at least referenced. We lock it as required, but always 450 * return a locked, resolved namecache entry. 451 */ 452 nd->nl_loopcnt = 0; 453 if (nd->nl_dvp) { 454 vrele(nd->nl_dvp); 455 nd->nl_dvp = NULL; 456 } 457 ptr = nd->nl_path; 458 459 /* 460 * Loop on the path components. At the top of the loop nd->nl_nch 461 * is ref'd and unlocked and represents our current position. 462 */ 463 for (;;) { 464 /* 465 * Make sure nl_nch is locked so we can access the vnode, resolution 466 * state, etc. 467 */ 468 if ((nd->nl_flags & NLC_NCPISLOCKED) == 0) { 469 nd->nl_flags |= NLC_NCPISLOCKED; 470 cache_lock_maybe_shared(&nd->nl_nch, islastelement(ptr)); 471 } 472 473 /* 474 * Check if the root directory should replace the current 475 * directory. This is done at the start of a translation 476 * or after a symbolic link has been found. In other cases 477 * ptr will never be pointing at a '/'. 478 */ 479 if (*ptr == '/') { 480 do { 481 ++ptr; 482 } while (*ptr == '/'); 483 cache_unlock(&nd->nl_nch); 484 cache_get_maybe_shared(&nd->nl_rootnch, &nch, 485 islastelement(ptr)); 486 cache_drop(&nd->nl_nch); 487 nd->nl_nch = nch; /* remains locked */ 488 489 /* 490 * Fast-track termination. There is no parent directory of 491 * the root in the same mount from the point of view of 492 * the caller so return EACCES if NLC_REFDVP is specified, 493 * and EEXIST if NLC_CREATE is also specified. 494 * e.g. 'rmdir /' or 'mkdir /' are not allowed. 495 */ 496 if (*ptr == 0) { 497 if (nd->nl_flags & NLC_REFDVP) 498 error = (nd->nl_flags & NLC_CREATE) ? EEXIST : EACCES; 499 else 500 error = 0; 501 break; 502 } 503 continue; 504 } 505 506 /* 507 * Check directory search permissions (nd->nl_nch is locked & refd) 508 */ 509 dflags = 0; 510 error = naccess(&nd->nl_nch, NLC_EXEC, nd->nl_cred, &dflags); 511 if (error) 512 break; 513 514 /* 515 * Extract the path component. Path components are limited to 516 * 255 characters. 517 */ 518 nlc.nlc_nameptr = ptr; 519 while (*ptr && *ptr != '/') 520 ++ptr; 521 nlc.nlc_namelen = ptr - nlc.nlc_nameptr; 522 if (nlc.nlc_namelen >= 256) { 523 error = ENAMETOOLONG; 524 break; 525 } 526 527 /* 528 * Lookup the path component in the cache, creating an unresolved 529 * entry if necessary. We have to handle "." and ".." as special 530 * cases. 531 * 532 * When handling ".." we have to detect a traversal back through a 533 * mount point. If we are at the root, ".." just returns the root. 534 * 535 * When handling "." or ".." we also have to recalculate dflags 536 * since our dflags will be for some sub-directory instead of the 537 * parent dir. 538 * 539 * This subsection returns a locked, refd 'nch' unless it errors out, 540 * and an unlocked but still ref'd nd->nl_nch. 541 * 542 * The namecache topology is not allowed to be disconnected, so 543 * encountering a NULL parent will generate EINVAL. This typically 544 * occurs when a directory is removed out from under a process. 545 * 546 * WARNING! The unlocking of nd->nl_nch is sensitive code. 547 */ 548 KKASSERT(nd->nl_flags & NLC_NCPISLOCKED); 549 550 if (nlc.nlc_namelen == 1 && nlc.nlc_nameptr[0] == '.') { 551 cache_unlock(&nd->nl_nch); 552 nd->nl_flags &= ~NLC_NCPISLOCKED; 553 cache_get_maybe_shared(&nd->nl_nch, &nch, islastelement(ptr)); 554 wasdotordotdot = 1; 555 } else if (nlc.nlc_namelen == 2 && 556 nlc.nlc_nameptr[0] == '.' && nlc.nlc_nameptr[1] == '.') { 557 if (nd->nl_nch.mount == nd->nl_rootnch.mount && 558 nd->nl_nch.ncp == nd->nl_rootnch.ncp 559 ) { 560 /* 561 * ".." at the root returns the root 562 */ 563 cache_unlock(&nd->nl_nch); 564 nd->nl_flags &= ~NLC_NCPISLOCKED; 565 cache_get_maybe_shared(&nd->nl_nch, &nch, islastelement(ptr)); 566 } else { 567 /* 568 * Locate the parent ncp. If we are at the root of a 569 * filesystem mount we have to skip to the mounted-on 570 * point in the underlying filesystem. 571 * 572 * Expect the parent to always be good since the 573 * mountpoint doesn't go away. XXX hack. cache_get() 574 * requires the ncp to already have a ref as a safety. 575 */ 576 nctmp = nd->nl_nch; 577 while (nctmp.ncp == nctmp.mount->mnt_ncmountpt.ncp) 578 nctmp = nctmp.mount->mnt_ncmounton; 579 nctmp.ncp = nctmp.ncp->nc_parent; 580 KKASSERT(nctmp.ncp != NULL); 581 cache_hold(&nctmp); 582 cache_unlock(&nd->nl_nch); 583 nd->nl_flags &= ~NLC_NCPISLOCKED; 584 cache_get_maybe_shared(&nctmp, &nch, islastelement(ptr)); 585 cache_drop(&nctmp); /* NOTE: zero's nctmp */ 586 } 587 wasdotordotdot = 2; 588 } else { 589 /* 590 * Must unlock nl_nch when traversing down the path. However, 591 * the child ncp has not yet been found/created and the parent's 592 * child list might be empty. Thus releasing the lock can 593 * allow a race whereby the parent ncp's vnode is recycled. 594 * This case can occur especially when maxvnodes is set very low. 595 * 596 * We need the parent's ncp to remain resolved for all normal 597 * filesystem activities, so we vhold() the vp during the lookup 598 * to prevent recyclement due to vnlru / maxvnodes. 599 * 600 * If we race an unlink or rename the ncp might be marked 601 * DESTROYED after resolution, requiring a retry. 602 */ 603 if ((hvp = nd->nl_nch.ncp->nc_vp) != NULL) 604 vhold(hvp); 605 cache_unlock(&nd->nl_nch); 606 nd->nl_flags &= ~NLC_NCPISLOCKED; 607 error = cache_nlookup_maybe_shared(&nd->nl_nch, &nlc, 608 islastelement(ptr), &nch); 609 if (error == EWOULDBLOCK) { 610 nch = cache_nlookup(&nd->nl_nch, &nlc); 611 if (nch.ncp->nc_flag & NCF_UNRESOLVED) 612 hit = 0; 613 for (;;) { 614 error = cache_resolve(&nch, nd->nl_cred); 615 if (error != EAGAIN && 616 (nch.ncp->nc_flag & NCF_DESTROYED) == 0) { 617 break; 618 } 619 kprintf("[diagnostic] nlookup: relookup %*.*s\n", 620 nch.ncp->nc_nlen, nch.ncp->nc_nlen, 621 nch.ncp->nc_name); 622 cache_put(&nch); 623 nch = cache_nlookup(&nd->nl_nch, &nlc); 624 } 625 } 626 if (hvp) 627 vdrop(hvp); 628 wasdotordotdot = 0; 629 } 630 631 /* 632 * If the last component was "." or ".." our dflags no longer 633 * represents the parent directory and we have to explicitly 634 * look it up. 635 * 636 * Expect the parent to be good since nch is locked. 637 */ 638 if (wasdotordotdot && error == 0) { 639 dflags = 0; 640 if ((par.ncp = nch.ncp->nc_parent) != NULL) { 641 par.mount = nch.mount; 642 cache_hold(&par); 643 cache_lock_maybe_shared(&par, islastelement(ptr)); 644 error = naccess(&par, 0, nd->nl_cred, &dflags); 645 cache_put(&par); 646 } 647 } 648 649 /* 650 * [end of subsection] 651 * 652 * nch is locked and referenced. 653 * nd->nl_nch is unlocked and referenced. 654 * 655 * nl_nch must be unlocked or we could chain lock to the root 656 * if a resolve gets stuck (e.g. in NFS). 657 */ 658 KKASSERT((nd->nl_flags & NLC_NCPISLOCKED) == 0); 659 660 /* 661 * Resolve the namespace if necessary. The ncp returned by 662 * cache_nlookup() is referenced and locked. 663 * 664 * XXX neither '.' nor '..' should return EAGAIN since they were 665 * previously resolved and thus cannot be newly created ncp's. 666 */ 667 if (nch.ncp->nc_flag & NCF_UNRESOLVED) { 668 hit = 0; 669 error = cache_resolve(&nch, nd->nl_cred); 670 KKASSERT(error != EAGAIN); 671 } else { 672 error = nch.ncp->nc_error; 673 } 674 675 /* 676 * Early completion. ENOENT is not an error if this is the last 677 * component and NLC_CREATE or NLC_RENAME (rename target) was 678 * requested. Note that ncp->nc_error is left as ENOENT in that 679 * case, which we check later on. 680 * 681 * Also handle invalid '.' or '..' components terminating a path 682 * for a create/rename/delete. The standard requires this and pax 683 * pretty stupidly depends on it. 684 */ 685 if (islastelement(ptr)) { 686 if (error == ENOENT && 687 (nd->nl_flags & (NLC_CREATE | NLC_RENAME_DST)) 688 ) { 689 if (nd->nl_flags & NLC_NFS_RDONLY) { 690 error = EROFS; 691 } else { 692 error = naccess(&nch, nd->nl_flags | dflags, 693 nd->nl_cred, NULL); 694 } 695 } 696 if (error == 0 && wasdotordotdot && 697 (nd->nl_flags & (NLC_CREATE | NLC_DELETE | 698 NLC_RENAME_SRC | NLC_RENAME_DST))) { 699 /* 700 * POSIX junk 701 */ 702 if (nd->nl_flags & NLC_CREATE) 703 error = EEXIST; 704 else if (nd->nl_flags & NLC_DELETE) 705 error = (wasdotordotdot == 1) ? EINVAL : ENOTEMPTY; 706 else 707 error = EINVAL; 708 } 709 } 710 711 /* 712 * Early completion on error. 713 */ 714 if (error) { 715 cache_put(&nch); 716 break; 717 } 718 719 /* 720 * If the element is a symlink and it is either not the last 721 * element or it is the last element and we are allowed to 722 * follow symlinks, resolve the symlink. 723 */ 724 if ((nch.ncp->nc_flag & NCF_ISSYMLINK) && 725 (*ptr || (nd->nl_flags & NLC_FOLLOW)) 726 ) { 727 if (nd->nl_loopcnt++ >= MAXSYMLINKS) { 728 error = ELOOP; 729 cache_put(&nch); 730 break; 731 } 732 error = nreadsymlink(nd, &nch, &nlc); 733 cache_put(&nch); 734 if (error) 735 break; 736 737 /* 738 * Concatenate trailing path elements onto the returned symlink. 739 * Note that if the path component (ptr) is not exhausted, it 740 * will being with a '/', so we do not have to add another one. 741 * 742 * The symlink may not be empty. 743 */ 744 len = strlen(ptr); 745 if (nlc.nlc_namelen == 0 || nlc.nlc_namelen + len >= MAXPATHLEN) { 746 error = nlc.nlc_namelen ? ENAMETOOLONG : ENOENT; 747 objcache_put(namei_oc, nlc.nlc_nameptr); 748 break; 749 } 750 bcopy(ptr, nlc.nlc_nameptr + nlc.nlc_namelen, len + 1); 751 if (nd->nl_flags & NLC_HASBUF) 752 objcache_put(namei_oc, nd->nl_path); 753 nd->nl_path = nlc.nlc_nameptr; 754 nd->nl_flags |= NLC_HASBUF; 755 ptr = nd->nl_path; 756 757 /* 758 * Go back up to the top to resolve any initial '/'s in the 759 * symlink. 760 */ 761 continue; 762 } 763 764 /* 765 * If the element is a directory and we are crossing a mount point, 766 * Locate the mount. 767 */ 768 while ((nch.ncp->nc_flag & NCF_ISMOUNTPT) && 769 (nd->nl_flags & NLC_NOCROSSMOUNT) == 0 && 770 (mp = cache_findmount(&nch)) != NULL 771 ) { 772 struct vnode *tdp; 773 int vfs_do_busy = 0; 774 775 /* 776 * VFS must be busied before the namecache entry is locked, 777 * but we don't want to waste time calling vfs_busy() if the 778 * mount point is already resolved. 779 */ 780 again: 781 cache_put(&nch); 782 if (vfs_do_busy) { 783 while (vfs_busy(mp, 0)) { 784 if (mp->mnt_kern_flag & MNTK_UNMOUNT) { 785 kprintf("nlookup: warning umount race avoided\n"); 786 cache_dropmount(mp); 787 error = EBUSY; 788 vfs_do_busy = 0; 789 goto double_break; 790 } 791 } 792 } 793 cache_get_maybe_shared(&mp->mnt_ncmountpt, &nch, 794 islastelement(ptr)); 795 796 if (nch.ncp->nc_flag & NCF_UNRESOLVED) { 797 if (vfs_do_busy == 0) { 798 vfs_do_busy = 1; 799 goto again; 800 } 801 error = VFS_ROOT(mp, &tdp); 802 vfs_unbusy(mp); 803 vfs_do_busy = 0; 804 if (error) { 805 cache_dropmount(mp); 806 break; 807 } 808 cache_setvp(&nch, tdp); 809 vput(tdp); 810 } 811 if (vfs_do_busy) 812 vfs_unbusy(mp); 813 cache_dropmount(mp); 814 } 815 816 if (error) { 817 cache_put(&nch); 818 double_break: 819 break; 820 } 821 822 /* 823 * Skip any slashes to get to the next element. If there 824 * are any slashes at all the current element must be a 825 * directory or, in the create case, intended to become a directory. 826 * If it isn't we break without incrementing ptr and fall through 827 * to the failure case below. 828 */ 829 while (*ptr == '/') { 830 if ((nch.ncp->nc_flag & NCF_ISDIR) == 0 && 831 !(nd->nl_flags & NLC_WILLBEDIR) 832 ) { 833 break; 834 } 835 ++ptr; 836 } 837 838 /* 839 * Continuation case: additional elements and the current 840 * element is a directory. 841 */ 842 if (*ptr && (nch.ncp->nc_flag & NCF_ISDIR)) { 843 cache_drop(&nd->nl_nch); 844 cache_unlock(&nch); 845 KKASSERT((nd->nl_flags & NLC_NCPISLOCKED) == 0); 846 nd->nl_nch = nch; 847 continue; 848 } 849 850 /* 851 * Failure case: additional elements and the current element 852 * is not a directory 853 */ 854 if (*ptr) { 855 cache_put(&nch); 856 error = ENOTDIR; 857 break; 858 } 859 860 /* 861 * Successful lookup of last element. 862 * 863 * Check permissions if the target exists. If the target does not 864 * exist directory permissions were already tested in the early 865 * completion code above. 866 * 867 * nd->nl_flags will be adjusted on return with NLC_APPENDONLY 868 * if the file is marked append-only, and NLC_STICKY if the directory 869 * containing the file is sticky. 870 */ 871 if (nch.ncp->nc_vp && (nd->nl_flags & NLC_ALLCHKS)) { 872 error = naccess(&nch, nd->nl_flags | dflags, 873 nd->nl_cred, NULL); 874 if (error) { 875 cache_put(&nch); 876 break; 877 } 878 } 879 880 /* 881 * Termination: no more elements. 882 * 883 * If NLC_REFDVP is set acquire a referenced parent dvp. 884 */ 885 if (nd->nl_flags & NLC_REFDVP) { 886 cache_lock(&nd->nl_nch); 887 error = cache_vref(&nd->nl_nch, nd->nl_cred, &nd->nl_dvp); 888 cache_unlock(&nd->nl_nch); 889 if (error) { 890 kprintf("NLC_REFDVP: Cannot ref dvp of %p\n", nch.ncp); 891 cache_put(&nch); 892 break; 893 } 894 } 895 cache_drop(&nd->nl_nch); 896 nd->nl_nch = nch; 897 nd->nl_flags |= NLC_NCPISLOCKED; 898 error = 0; 899 break; 900 } 901 902 if (hit) 903 ++gd->gd_nchstats->ncs_longhits; 904 else 905 ++gd->gd_nchstats->ncs_longmiss; 906 907 if (nd->nl_flags & NLC_NCPISLOCKED) 908 KKASSERT(cache_lockstatus(&nd->nl_nch) > 0); 909 910 /* 911 * NOTE: If NLC_CREATE was set the ncp may represent a negative hit 912 * (ncp->nc_error will be ENOENT), but we will still return an error 913 * code of 0. 914 */ 915 return(error); 916 } 917 918 /* 919 * Resolve a mount point's glue ncp. This ncp connects creates the illusion 920 * of continuity in the namecache tree by connecting the ncp related to the 921 * vnode under the mount to the ncp related to the mount's root vnode. 922 * 923 * If no error occured a locked, ref'd ncp is stored in *ncpp. 924 */ 925 int 926 nlookup_mp(struct mount *mp, struct nchandle *nch) 927 { 928 struct vnode *vp; 929 int error; 930 931 error = 0; 932 cache_get(&mp->mnt_ncmountpt, nch); 933 if (nch->ncp->nc_flag & NCF_UNRESOLVED) { 934 while (vfs_busy(mp, 0)) 935 ; 936 error = VFS_ROOT(mp, &vp); 937 vfs_unbusy(mp); 938 if (error) { 939 cache_put(nch); 940 } else { 941 cache_setvp(nch, vp); 942 vput(vp); 943 } 944 } 945 return(error); 946 } 947 948 /* 949 * Read the contents of a symlink, allocate a path buffer out of the 950 * namei_oc and initialize the supplied nlcomponent with the result. 951 * 952 * If an error occurs no buffer will be allocated or returned in the nlc. 953 */ 954 int 955 nreadsymlink(struct nlookupdata *nd, struct nchandle *nch, 956 struct nlcomponent *nlc) 957 { 958 struct vnode *vp; 959 struct iovec aiov; 960 struct uio auio; 961 int linklen; 962 int error; 963 char *cp; 964 965 nlc->nlc_nameptr = NULL; 966 nlc->nlc_namelen = 0; 967 if (nch->ncp->nc_vp == NULL) 968 return(ENOENT); 969 if ((error = cache_vget(nch, nd->nl_cred, LK_SHARED, &vp)) != 0) 970 return(error); 971 cp = objcache_get(namei_oc, M_WAITOK); 972 aiov.iov_base = cp; 973 aiov.iov_len = MAXPATHLEN; 974 auio.uio_iov = &aiov; 975 auio.uio_iovcnt = 1; 976 auio.uio_offset = 0; 977 auio.uio_rw = UIO_READ; 978 auio.uio_segflg = UIO_SYSSPACE; 979 auio.uio_td = nd->nl_td; 980 auio.uio_resid = MAXPATHLEN - 1; 981 error = VOP_READLINK(vp, &auio, nd->nl_cred); 982 if (error) 983 goto fail; 984 linklen = MAXPATHLEN - 1 - auio.uio_resid; 985 if (varsym_enable) { 986 linklen = varsymreplace(cp, linklen, MAXPATHLEN - 1); 987 if (linklen < 0) { 988 error = ENAMETOOLONG; 989 goto fail; 990 } 991 } 992 cp[linklen] = 0; 993 nlc->nlc_nameptr = cp; 994 nlc->nlc_namelen = linklen; 995 vput(vp); 996 return(0); 997 fail: 998 objcache_put(namei_oc, cp); 999 vput(vp); 1000 return(error); 1001 } 1002 1003 /* 1004 * Check access [XXX cache vattr!] [XXX quota] 1005 * 1006 * Generally check the NLC_* access bits. All specified bits must pass 1007 * for this function to return 0. 1008 * 1009 * The file does not have to exist when checking NLC_CREATE or NLC_RENAME_DST 1010 * access, otherwise it must exist. No error is returned in this case. 1011 * 1012 * The file must not exist if NLC_EXCL is specified. 1013 * 1014 * Directory permissions in general are tested for NLC_CREATE if the file 1015 * does not exist, NLC_DELETE if the file does exist, and NLC_RENAME_DST 1016 * whether the file exists or not. 1017 * 1018 * The directory sticky bit is tested for NLC_DELETE and NLC_RENAME_DST, 1019 * the latter is only tested if the target exists. 1020 * 1021 * The passed ncp must be referenced and locked. If it is already resolved 1022 * it may be locked shared but otherwise should be locked exclusively. 1023 */ 1024 static int 1025 naccess(struct nchandle *nch, int nflags, struct ucred *cred, int *nflagsp) 1026 { 1027 struct vnode *vp; 1028 struct vattr va; 1029 struct namecache *ncp; 1030 int error; 1031 int cflags; 1032 1033 KKASSERT(cache_lockstatus(nch) > 0); 1034 1035 ncp = nch->ncp; 1036 if (ncp->nc_flag & NCF_UNRESOLVED) { 1037 cache_resolve(nch, cred); 1038 ncp = nch->ncp; 1039 } 1040 error = ncp->nc_error; 1041 1042 /* 1043 * Directory permissions checks. Silently ignore ENOENT if these 1044 * tests pass. It isn't an error. 1045 * 1046 * We can safely resolve ncp->nc_parent because ncp is currently 1047 * locked. 1048 */ 1049 if (nflags & (NLC_CREATE | NLC_DELETE | NLC_RENAME_SRC | NLC_RENAME_DST)) { 1050 if (((nflags & NLC_CREATE) && ncp->nc_vp == NULL) || 1051 ((nflags & NLC_DELETE) && ncp->nc_vp != NULL) || 1052 ((nflags & NLC_RENAME_SRC) && ncp->nc_vp != NULL) || 1053 (nflags & NLC_RENAME_DST) 1054 ) { 1055 struct nchandle par; 1056 1057 if ((par.ncp = ncp->nc_parent) == NULL) { 1058 if (error != EAGAIN) 1059 error = EINVAL; 1060 } else if (error == 0 || error == ENOENT) { 1061 par.mount = nch->mount; 1062 cache_hold(&par); 1063 cache_lock_maybe_shared(&par, 0); 1064 error = naccess(&par, NLC_WRITE, cred, NULL); 1065 cache_put(&par); 1066 } 1067 } 1068 } 1069 1070 /* 1071 * NLC_EXCL check. Target file must not exist. 1072 */ 1073 if (error == 0 && (nflags & NLC_EXCL) && ncp->nc_vp != NULL) 1074 error = EEXIST; 1075 1076 /* 1077 * Get the vnode attributes so we can do the rest of our checks. 1078 * 1079 * NOTE: We only call naccess_va() if the target exists. 1080 */ 1081 if (error == 0) { 1082 error = cache_vget(nch, cred, LK_SHARED, &vp); 1083 if (error == ENOENT) { 1084 /* 1085 * Silently zero-out ENOENT if creating or renaming 1086 * (rename target). It isn't an error. 1087 */ 1088 if (nflags & (NLC_CREATE | NLC_RENAME_DST)) 1089 error = 0; 1090 } else if (error == 0) { 1091 /* 1092 * Get the vnode attributes and check for illegal O_TRUNC 1093 * requests and read-only mounts. 1094 * 1095 * NOTE: You can still open devices on read-only mounts for 1096 * writing. 1097 * 1098 * NOTE: creates/deletes/renames are handled by the NLC_WRITE 1099 * check on the parent directory above. 1100 * 1101 * XXX cache the va in the namecache or in the vnode 1102 */ 1103 error = VOP_GETATTR(vp, &va); 1104 if (error == 0 && (nflags & NLC_TRUNCATE)) { 1105 switch(va.va_type) { 1106 case VREG: 1107 case VDATABASE: 1108 case VCHR: 1109 case VBLK: 1110 case VFIFO: 1111 break; 1112 case VDIR: 1113 error = EISDIR; 1114 break; 1115 default: 1116 error = EINVAL; 1117 break; 1118 } 1119 } 1120 if (error == 0 && (nflags & NLC_WRITE) && vp->v_mount && 1121 (vp->v_mount->mnt_flag & MNT_RDONLY) 1122 ) { 1123 switch(va.va_type) { 1124 case VDIR: 1125 case VLNK: 1126 case VREG: 1127 case VDATABASE: 1128 error = EROFS; 1129 break; 1130 default: 1131 break; 1132 } 1133 } 1134 vput(vp); 1135 1136 /* 1137 * Check permissions based on file attributes. The passed 1138 * flags (*nflagsp) are modified with feedback based on 1139 * special attributes and requirements. 1140 */ 1141 if (error == 0) { 1142 /* 1143 * Adjust the returned (*nflagsp) if non-NULL. 1144 */ 1145 if (nflagsp) { 1146 if ((va.va_mode & VSVTX) && va.va_uid != cred->cr_uid) 1147 *nflagsp |= NLC_STICKY; 1148 if (va.va_flags & APPEND) 1149 *nflagsp |= NLC_APPENDONLY; 1150 if (va.va_flags & IMMUTABLE) 1151 *nflagsp |= NLC_IMMUTABLE; 1152 } 1153 1154 /* 1155 * Track swapcache management flags in the namecache. 1156 * 1157 * Calculate the flags based on the current vattr info 1158 * and recalculate the inherited flags from the parent 1159 * (the original cache linkage may have occurred without 1160 * getattrs and thus have stale flags). 1161 */ 1162 cflags = 0; 1163 if (va.va_flags & SF_NOCACHE) 1164 cflags |= NCF_SF_NOCACHE; 1165 if (va.va_flags & UF_CACHE) 1166 cflags |= NCF_UF_CACHE; 1167 if (ncp->nc_parent) { 1168 if (ncp->nc_parent->nc_flag & 1169 (NCF_SF_NOCACHE | NCF_SF_PNOCACHE)) { 1170 cflags |= NCF_SF_PNOCACHE; 1171 } 1172 if (ncp->nc_parent->nc_flag & 1173 (NCF_UF_CACHE | NCF_UF_PCACHE)) { 1174 cflags |= NCF_UF_PCACHE; 1175 } 1176 } 1177 1178 /* 1179 * XXX we're not supposed to update nc_flag when holding 1180 * a shared lock. 1181 */ 1182 atomic_clear_short(&ncp->nc_flag, 1183 (NCF_SF_NOCACHE | NCF_UF_CACHE | 1184 NCF_SF_PNOCACHE | NCF_UF_PCACHE) & ~cflags); 1185 atomic_set_short(&ncp->nc_flag, cflags); 1186 1187 /* 1188 * Process general access. 1189 */ 1190 error = naccess_va(&va, nflags, cred); 1191 } 1192 } 1193 } 1194 return(error); 1195 } 1196 1197 /* 1198 * Check the requested access against the given vattr using cred. 1199 */ 1200 int 1201 naccess_va(struct vattr *va, int nflags, struct ucred *cred) 1202 { 1203 int i; 1204 int vmode; 1205 1206 /* 1207 * Test the immutable bit. Creations, deletions, renames (source 1208 * or destination) are not allowed. chown/chmod/other is also not 1209 * allowed but is handled by SETATTR. Hardlinks to the immutable 1210 * file are allowed. 1211 * 1212 * If the directory is set to immutable then creations, deletions, 1213 * renames (source or dest) and hardlinks to files within the directory 1214 * are not allowed, and regular files opened through the directory may 1215 * not be written to or truncated (unless a special device). 1216 * 1217 * NOTE! New hardlinks to immutable files work but new hardlinks to 1218 * files, immutable or not, sitting inside an immutable directory are 1219 * not allowed. As always if the file is hardlinked via some other 1220 * path additional hardlinks may be possible even if the file is marked 1221 * immutable. The sysop needs to create a closure by checking the hard 1222 * link count. Once closure is achieved you are good, and security 1223 * scripts should check link counts anyway. 1224 * 1225 * Writes and truncations are only allowed on special devices. 1226 */ 1227 if ((va->va_flags & IMMUTABLE) || (nflags & NLC_IMMUTABLE)) { 1228 if ((nflags & NLC_IMMUTABLE) && (nflags & NLC_HLINK)) 1229 return (EPERM); 1230 if (nflags & (NLC_CREATE | NLC_DELETE | 1231 NLC_RENAME_SRC | NLC_RENAME_DST)) { 1232 return (EPERM); 1233 } 1234 if (nflags & (NLC_WRITE | NLC_TRUNCATE)) { 1235 switch(va->va_type) { 1236 case VDIR: 1237 return (EISDIR); 1238 case VLNK: 1239 case VREG: 1240 case VDATABASE: 1241 return (EPERM); 1242 default: 1243 break; 1244 } 1245 } 1246 } 1247 1248 /* 1249 * Test the no-unlink and append-only bits for opens, rename targets, 1250 * and deletions. These bits are not tested for creations or 1251 * rename sources. 1252 * 1253 * Unlike FreeBSD we allow a file with APPEND set to be renamed. 1254 * If you do not wish this you must also set NOUNLINK. 1255 * 1256 * If the governing directory is marked APPEND-only it implies 1257 * NOUNLINK for all entries in the directory. 1258 */ 1259 if (((va->va_flags & NOUNLINK) || (nflags & NLC_APPENDONLY)) && 1260 (nflags & (NLC_DELETE | NLC_RENAME_SRC | NLC_RENAME_DST)) 1261 ) { 1262 return (EPERM); 1263 } 1264 1265 /* 1266 * A file marked append-only may not be deleted but can be renamed. 1267 */ 1268 if ((va->va_flags & APPEND) && 1269 (nflags & (NLC_DELETE | NLC_RENAME_DST)) 1270 ) { 1271 return (EPERM); 1272 } 1273 1274 /* 1275 * A file marked append-only which is opened for writing must also 1276 * be opened O_APPEND. 1277 */ 1278 if ((va->va_flags & APPEND) && (nflags & (NLC_OPEN | NLC_TRUNCATE))) { 1279 if (nflags & NLC_TRUNCATE) 1280 return (EPERM); 1281 if ((nflags & (NLC_OPEN | NLC_WRITE)) == (NLC_OPEN | NLC_WRITE)) { 1282 if ((nflags & NLC_APPEND) == 0) 1283 return (EPERM); 1284 } 1285 } 1286 1287 /* 1288 * root gets universal access 1289 */ 1290 if (cred->cr_uid == 0) 1291 return(0); 1292 1293 /* 1294 * Check owner perms. 1295 * 1296 * If NLC_OWN is set the owner of the file is allowed no matter when 1297 * the owner-mode bits say (utimes). 1298 */ 1299 vmode = 0; 1300 if (nflags & NLC_READ) 1301 vmode |= S_IRUSR; 1302 if (nflags & NLC_WRITE) 1303 vmode |= S_IWUSR; 1304 if (nflags & NLC_EXEC) 1305 vmode |= S_IXUSR; 1306 1307 if (cred->cr_uid == va->va_uid) { 1308 if ((nflags & NLC_OWN) == 0) { 1309 if ((vmode & va->va_mode) != vmode) 1310 return(EACCES); 1311 } 1312 return(0); 1313 } 1314 1315 /* 1316 * If NLC_STICKY is set only the owner may delete or rename a file. 1317 * This bit is typically set on /tmp. 1318 * 1319 * Note that the NLC_READ/WRITE/EXEC bits are not typically set in 1320 * the specific delete or rename case. For deletions and renames we 1321 * usually just care about directory permissions, not file permissions. 1322 */ 1323 if ((nflags & NLC_STICKY) && 1324 (nflags & (NLC_RENAME_SRC | NLC_RENAME_DST | NLC_DELETE))) { 1325 return(EACCES); 1326 } 1327 1328 /* 1329 * Check group perms 1330 */ 1331 vmode >>= 3; 1332 for (i = 0; i < cred->cr_ngroups; ++i) { 1333 if (va->va_gid == cred->cr_groups[i]) { 1334 if ((vmode & va->va_mode) != vmode) 1335 return(EACCES); 1336 return(0); 1337 } 1338 } 1339 1340 /* 1341 * Check world perms 1342 */ 1343 vmode >>= 3; 1344 if ((vmode & va->va_mode) != vmode) 1345 return(EACCES); 1346 return(0); 1347 } 1348 1349