FreeBSD/Linux Kernel Cross Reference
sys/kern/vfs_lookup.c
1 /*-
2 * SPDX-License-Identifier: BSD-3-Clause
3 *
4 * Copyright (c) 1982, 1986, 1989, 1993
5 * The Regents of the University of California. All rights reserved.
6 * (c) UNIX System Laboratories, Inc.
7 * All or some portions of this file are derived from material licensed
8 * to the University of California by American Telephone and Telegraph
9 * Co. or Unix System Laboratories, Inc. and are reproduced herein with
10 * the permission of UNIX System Laboratories, Inc.
11 *
12 * Redistribution and use in source and binary forms, with or without
13 * modification, are permitted provided that the following conditions
14 * are met:
15 * 1. Redistributions of source code must retain the above copyright
16 * notice, this list of conditions and the following disclaimer.
17 * 2. Redistributions in binary form must reproduce the above copyright
18 * notice, this list of conditions and the following disclaimer in the
19 * documentation and/or other materials provided with the distribution.
20 * 3. Neither the name of the University nor the names of its contributors
21 * may be used to endorse or promote products derived from this software
22 * without specific prior written permission.
23 *
24 * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
25 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
26 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
27 * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
28 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
29 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
30 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
31 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
32 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
33 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
34 * SUCH DAMAGE.
35 *
36 * @(#)vfs_lookup.c 8.4 (Berkeley) 2/16/94
37 */
38
39 #include <sys/cdefs.h>
40 __FBSDID("$FreeBSD$");
41
42 #include "opt_capsicum.h"
43 #include "opt_ktrace.h"
44
45 #include <sys/param.h>
46 #include <sys/systm.h>
47 #include <sys/kernel.h>
48 #include <sys/capsicum.h>
49 #include <sys/fcntl.h>
50 #include <sys/jail.h>
51 #include <sys/lock.h>
52 #include <sys/mutex.h>
53 #include <sys/namei.h>
54 #include <sys/vnode.h>
55 #include <sys/mount.h>
56 #include <sys/filedesc.h>
57 #include <sys/proc.h>
58 #include <sys/sdt.h>
59 #include <sys/syscallsubr.h>
60 #include <sys/sysctl.h>
61 #ifdef KTRACE
62 #include <sys/ktrace.h>
63 #endif
64 #ifdef INVARIANTS
65 #include <machine/_inttypes.h>
66 #endif
67
68 #include <security/audit/audit.h>
69 #include <security/mac/mac_framework.h>
70
71 #include <vm/uma.h>
72
73 #define NAMEI_DIAGNOSTIC 1
74 #undef NAMEI_DIAGNOSTIC
75
76 SDT_PROVIDER_DEFINE(vfs);
77 SDT_PROBE_DEFINE4(vfs, namei, lookup, entry, "struct vnode *", "char *",
78 "unsigned long", "bool");
79 SDT_PROBE_DEFINE4(vfs, namei, lookup, return, "int", "struct vnode *", "bool",
80 "struct nameidata");
81
82 /* Allocation zone for namei. */
83 uma_zone_t namei_zone;
84
85 /* Placeholder vnode for mp traversal. */
86 static struct vnode *vp_crossmp;
87
88 static int
89 crossmp_vop_islocked(struct vop_islocked_args *ap)
90 {
91
92 return (LK_SHARED);
93 }
94
95 static int
96 crossmp_vop_lock1(struct vop_lock1_args *ap)
97 {
98 struct vnode *vp;
99 struct lock *lk __unused;
100 const char *file __unused;
101 int flags, line __unused;
102
103 vp = ap->a_vp;
104 lk = vp->v_vnlock;
105 flags = ap->a_flags;
106 file = ap->a_file;
107 line = ap->a_line;
108
109 if ((flags & LK_SHARED) == 0)
110 panic("invalid lock request for crossmp");
111
112 WITNESS_CHECKORDER(&lk->lock_object, LOP_NEWORDER, file, line,
113 flags & LK_INTERLOCK ? &VI_MTX(vp)->lock_object : NULL);
114 WITNESS_LOCK(&lk->lock_object, 0, file, line);
115 if ((flags & LK_INTERLOCK) != 0)
116 VI_UNLOCK(vp);
117 LOCK_LOG_LOCK("SLOCK", &lk->lock_object, 0, 0, ap->a_file, line);
118 return (0);
119 }
120
121 static int
122 crossmp_vop_unlock(struct vop_unlock_args *ap)
123 {
124 struct vnode *vp;
125 struct lock *lk __unused;
126
127 vp = ap->a_vp;
128 lk = vp->v_vnlock;
129
130 WITNESS_UNLOCK(&lk->lock_object, 0, LOCK_FILE, LOCK_LINE);
131 LOCK_LOG_LOCK("SUNLOCK", &lk->lock_object, 0, 0, LOCK_FILE,
132 LOCK_LINE);
133 return (0);
134 }
135
136 static struct vop_vector crossmp_vnodeops = {
137 .vop_default = &default_vnodeops,
138 .vop_islocked = crossmp_vop_islocked,
139 .vop_lock1 = crossmp_vop_lock1,
140 .vop_unlock = crossmp_vop_unlock,
141 };
142 /*
143 * VFS_VOP_VECTOR_REGISTER(crossmp_vnodeops) is not used here since the vnode
144 * gets allocated early. See nameiinit for the direct call below.
145 */
146
147 struct nameicap_tracker {
148 struct vnode *dp;
149 TAILQ_ENTRY(nameicap_tracker) nm_link;
150 };
151
152 /* Zone for cap mode tracker elements used for dotdot capability checks. */
153 MALLOC_DEFINE(M_NAMEITRACKER, "namei_tracker", "namei tracking for dotdot");
154
155 static void
156 nameiinit(void *dummy __unused)
157 {
158
159 namei_zone = uma_zcreate("NAMEI", MAXPATHLEN, NULL, NULL, NULL, NULL,
160 UMA_ALIGN_PTR, 0);
161 vfs_vector_op_register(&crossmp_vnodeops);
162 getnewvnode("crossmp", NULL, &crossmp_vnodeops, &vp_crossmp);
163 }
164 SYSINIT(vfs, SI_SUB_VFS, SI_ORDER_SECOND, nameiinit, NULL);
165
166 static int lookup_cap_dotdot = 1;
167 SYSCTL_INT(_vfs, OID_AUTO, lookup_cap_dotdot, CTLFLAG_RWTUN,
168 &lookup_cap_dotdot, 0,
169 "enables \"..\" components in path lookup in capability mode");
170 static int lookup_cap_dotdot_nonlocal = 1;
171 SYSCTL_INT(_vfs, OID_AUTO, lookup_cap_dotdot_nonlocal, CTLFLAG_RWTUN,
172 &lookup_cap_dotdot_nonlocal, 0,
173 "enables \"..\" components in path lookup in capability mode "
174 "on non-local mount");
175
176 static void
177 nameicap_tracker_add(struct nameidata *ndp, struct vnode *dp)
178 {
179 struct nameicap_tracker *nt;
180 struct componentname *cnp;
181
182 if ((ndp->ni_lcf & NI_LCF_CAP_DOTDOT) == 0 || dp->v_type != VDIR)
183 return;
184 cnp = &ndp->ni_cnd;
185 nt = TAILQ_LAST(&ndp->ni_cap_tracker, nameicap_tracker_head);
186 if (nt != NULL && nt->dp == dp)
187 return;
188 nt = malloc(sizeof(*nt), M_NAMEITRACKER, M_WAITOK);
189 vhold(dp);
190 nt->dp = dp;
191 TAILQ_INSERT_TAIL(&ndp->ni_cap_tracker, nt, nm_link);
192 }
193
194 static void
195 nameicap_cleanup_from(struct nameidata *ndp, struct nameicap_tracker *first)
196 {
197 struct nameicap_tracker *nt, *nt1;
198
199 nt = first;
200 TAILQ_FOREACH_FROM_SAFE(nt, &ndp->ni_cap_tracker, nm_link, nt1) {
201 TAILQ_REMOVE(&ndp->ni_cap_tracker, nt, nm_link);
202 vdrop(nt->dp);
203 free(nt, M_NAMEITRACKER);
204 }
205 }
206
207 static void
208 nameicap_cleanup(struct nameidata *ndp)
209 {
210 KASSERT(TAILQ_EMPTY(&ndp->ni_cap_tracker) ||
211 (ndp->ni_lcf & NI_LCF_CAP_DOTDOT) != 0, ("not strictrelative"));
212 nameicap_cleanup_from(ndp, NULL);
213 }
214
215 /*
216 * For dotdot lookups in capability mode, only allow the component
217 * lookup to succeed if the resulting directory was already traversed
218 * during the operation. This catches situations where already
219 * traversed directory is moved to different parent, and then we walk
220 * over it with dotdots.
221 *
222 * Also allow to force failure of dotdot lookups for non-local
223 * filesystems, where external agents might assist local lookups to
224 * escape the compartment.
225 */
226 static int
227 nameicap_check_dotdot(struct nameidata *ndp, struct vnode *dp)
228 {
229 struct nameicap_tracker *nt;
230 struct mount *mp;
231
232 if (dp == NULL || dp->v_type != VDIR || (ndp->ni_lcf &
233 NI_LCF_STRICTRELATIVE) == 0)
234 return (0);
235 if ((ndp->ni_lcf & NI_LCF_CAP_DOTDOT) == 0)
236 return (ENOTCAPABLE);
237 mp = dp->v_mount;
238 if (lookup_cap_dotdot_nonlocal == 0 && mp != NULL &&
239 (mp->mnt_flag & MNT_LOCAL) == 0)
240 return (ENOTCAPABLE);
241 TAILQ_FOREACH_REVERSE(nt, &ndp->ni_cap_tracker, nameicap_tracker_head,
242 nm_link) {
243 if (dp == nt->dp) {
244 nt = TAILQ_NEXT(nt, nm_link);
245 if (nt != NULL)
246 nameicap_cleanup_from(ndp, nt);
247 return (0);
248 }
249 }
250 return (ENOTCAPABLE);
251 }
252
253 static void
254 namei_cleanup_cnp(struct componentname *cnp)
255 {
256
257 uma_zfree(namei_zone, cnp->cn_pnbuf);
258 #ifdef DIAGNOSTIC
259 cnp->cn_pnbuf = NULL;
260 cnp->cn_nameptr = NULL;
261 #endif
262 }
263
264 static int
265 namei_handle_root(struct nameidata *ndp, struct vnode **dpp)
266 {
267 struct componentname *cnp;
268
269 cnp = &ndp->ni_cnd;
270 if ((ndp->ni_lcf & NI_LCF_STRICTRELATIVE) != 0) {
271 #ifdef KTRACE
272 if (KTRPOINT(curthread, KTR_CAPFAIL))
273 ktrcapfail(CAPFAIL_LOOKUP, NULL, NULL);
274 #endif
275 return (ENOTCAPABLE);
276 }
277 while (*(cnp->cn_nameptr) == '/') {
278 cnp->cn_nameptr++;
279 ndp->ni_pathlen--;
280 }
281 *dpp = ndp->ni_rootdir;
282 vrefact(*dpp);
283 return (0);
284 }
285
286 static int
287 namei_setup(struct nameidata *ndp, struct vnode **dpp, struct pwd **pwdp)
288 {
289 struct componentname *cnp;
290 struct file *dfp;
291 struct thread *td;
292 struct pwd *pwd;
293 cap_rights_t rights;
294 int error;
295 bool startdir_used;
296
297 cnp = &ndp->ni_cnd;
298 td = cnp->cn_thread;
299
300 startdir_used = false;
301 *pwdp = NULL;
302 *dpp = NULL;
303
304 #ifdef CAPABILITY_MODE
305 /*
306 * In capability mode, lookups must be restricted to happen in
307 * the subtree with the root specified by the file descriptor:
308 * - The root must be real file descriptor, not the pseudo-descriptor
309 * AT_FDCWD.
310 * - The passed path must be relative and not absolute.
311 * - If lookup_cap_dotdot is disabled, path must not contain the
312 * '..' components.
313 * - If lookup_cap_dotdot is enabled, we verify that all '..'
314 * components lookups result in the directories which were
315 * previously walked by us, which prevents an escape from
316 * the relative root.
317 */
318 if (IN_CAPABILITY_MODE(td) && (cnp->cn_flags & NOCAPCHECK) == 0) {
319 ndp->ni_lcf |= NI_LCF_STRICTRELATIVE;
320 ndp->ni_resflags |= NIRES_STRICTREL;
321 if (ndp->ni_dirfd == AT_FDCWD) {
322 #ifdef KTRACE
323 if (KTRPOINT(td, KTR_CAPFAIL))
324 ktrcapfail(CAPFAIL_LOOKUP, NULL, NULL);
325 #endif
326 return (ECAPMODE);
327 }
328 }
329 #endif
330 error = 0;
331
332 /*
333 * Get starting point for the translation.
334 */
335 pwd = pwd_hold(td);
336 /*
337 * The reference on ni_rootdir is acquired in the block below to avoid
338 * back-to-back atomics for absolute lookups.
339 */
340 ndp->ni_rootdir = pwd->pwd_rdir;
341 ndp->ni_topdir = pwd->pwd_jdir;
342
343 if (cnp->cn_pnbuf[0] == '/') {
344 ndp->ni_resflags |= NIRES_ABS;
345 error = namei_handle_root(ndp, dpp);
346 } else {
347 if (ndp->ni_startdir != NULL) {
348 *dpp = ndp->ni_startdir;
349 startdir_used = true;
350 } else if (ndp->ni_dirfd == AT_FDCWD) {
351 *dpp = pwd->pwd_cdir;
352 vrefact(*dpp);
353 } else {
354 rights = *ndp->ni_rightsneeded;
355 cap_rights_set_one(&rights, CAP_LOOKUP);
356
357 if (cnp->cn_flags & AUDITVNODE1)
358 AUDIT_ARG_ATFD1(ndp->ni_dirfd);
359 if (cnp->cn_flags & AUDITVNODE2)
360 AUDIT_ARG_ATFD2(ndp->ni_dirfd);
361 /*
362 * Effectively inlined fgetvp_rights, because we need to
363 * inspect the file as well as grabbing the vnode.
364 */
365 error = fget_cap(td, ndp->ni_dirfd, &rights,
366 &dfp, &ndp->ni_filecaps);
367 if (error != 0) {
368 /*
369 * Preserve the error; it should either be EBADF
370 * or capability-related, both of which can be
371 * safely returned to the caller.
372 */
373 } else {
374 if (dfp->f_ops == &badfileops) {
375 error = EBADF;
376 } else if (dfp->f_vnode == NULL) {
377 error = ENOTDIR;
378 } else {
379 *dpp = dfp->f_vnode;
380 vrefact(*dpp);
381
382 if ((dfp->f_flag & FSEARCH) != 0)
383 cnp->cn_flags |= NOEXECCHECK;
384 }
385 fdrop(dfp, td);
386 }
387 #ifdef CAPABILITIES
388 /*
389 * If file descriptor doesn't have all rights,
390 * all lookups relative to it must also be
391 * strictly relative.
392 */
393 CAP_ALL(&rights);
394 if (!cap_rights_contains(&ndp->ni_filecaps.fc_rights,
395 &rights) ||
396 ndp->ni_filecaps.fc_fcntls != CAP_FCNTL_ALL ||
397 ndp->ni_filecaps.fc_nioctls != -1) {
398 ndp->ni_lcf |= NI_LCF_STRICTRELATIVE;
399 ndp->ni_resflags |= NIRES_STRICTREL;
400 }
401 #endif
402 }
403 if (error == 0 && (*dpp)->v_type != VDIR)
404 error = ENOTDIR;
405 }
406 if (error == 0 && (cnp->cn_flags & RBENEATH) != 0) {
407 if (cnp->cn_pnbuf[0] == '/') {
408 error = ENOTCAPABLE;
409 } else if ((ndp->ni_lcf & NI_LCF_STRICTRELATIVE) == 0) {
410 ndp->ni_lcf |= NI_LCF_STRICTRELATIVE |
411 NI_LCF_CAP_DOTDOT;
412 }
413 }
414
415 /*
416 * If we are auditing the kernel pathname, save the user pathname.
417 */
418 if (cnp->cn_flags & AUDITVNODE1)
419 AUDIT_ARG_UPATH1_VP(td, ndp->ni_rootdir, *dpp, cnp->cn_pnbuf);
420 if (cnp->cn_flags & AUDITVNODE2)
421 AUDIT_ARG_UPATH2_VP(td, ndp->ni_rootdir, *dpp, cnp->cn_pnbuf);
422 if (ndp->ni_startdir != NULL && !startdir_used)
423 vrele(ndp->ni_startdir);
424 if (error != 0) {
425 if (*dpp != NULL)
426 vrele(*dpp);
427 pwd_drop(pwd);
428 return (error);
429 }
430 if ((ndp->ni_lcf & NI_LCF_STRICTRELATIVE) != 0 &&
431 lookup_cap_dotdot != 0)
432 ndp->ni_lcf |= NI_LCF_CAP_DOTDOT;
433 SDT_PROBE4(vfs, namei, lookup, entry, *dpp, cnp->cn_pnbuf,
434 cnp->cn_flags, false);
435 *pwdp = pwd;
436 return (0);
437 }
438
439 static int
440 namei_getpath(struct nameidata *ndp)
441 {
442 struct componentname *cnp;
443 int error;
444
445 cnp = &ndp->ni_cnd;
446
447 /*
448 * Get a buffer for the name to be translated, and copy the
449 * name into the buffer.
450 */
451 cnp->cn_pnbuf = uma_zalloc(namei_zone, M_WAITOK);
452 if (ndp->ni_segflg == UIO_SYSSPACE) {
453 error = copystr(ndp->ni_dirp, cnp->cn_pnbuf, MAXPATHLEN,
454 &ndp->ni_pathlen);
455 } else {
456 error = copyinstr(ndp->ni_dirp, cnp->cn_pnbuf, MAXPATHLEN,
457 &ndp->ni_pathlen);
458 }
459
460 if (__predict_false(error != 0)) {
461 namei_cleanup_cnp(cnp);
462 return (error);
463 }
464
465 /*
466 * Don't allow empty pathnames.
467 */
468 if (__predict_false(*cnp->cn_pnbuf == '\0')) {
469 namei_cleanup_cnp(cnp);
470 return (ENOENT);
471 }
472
473 cnp->cn_nameptr = cnp->cn_pnbuf;
474 return (0);
475 }
476
477 /*
478 * Convert a pathname into a pointer to a locked vnode.
479 *
480 * The FOLLOW flag is set when symbolic links are to be followed
481 * when they occur at the end of the name translation process.
482 * Symbolic links are always followed for all other pathname
483 * components other than the last.
484 *
485 * The segflg defines whether the name is to be copied from user
486 * space or kernel space.
487 *
488 * Overall outline of namei:
489 *
490 * copy in name
491 * get starting directory
492 * while (!done && !error) {
493 * call lookup to search path.
494 * if symbolic link, massage name in buffer and continue
495 * }
496 */
497 int
498 namei(struct nameidata *ndp)
499 {
500 char *cp; /* pointer into pathname argument */
501 struct vnode *dp; /* the directory we are searching */
502 struct iovec aiov; /* uio for reading symbolic links */
503 struct componentname *cnp;
504 struct thread *td;
505 struct pwd *pwd;
506 struct uio auio;
507 int error, linklen;
508 enum cache_fpl_status status;
509
510 cnp = &ndp->ni_cnd;
511 td = cnp->cn_thread;
512 #ifdef INVARIANTS
513 KASSERT((ndp->ni_debugflags & NAMEI_DBG_CALLED) == 0,
514 ("%s: repeated call to namei without NDREINIT", __func__));
515 KASSERT(ndp->ni_debugflags == NAMEI_DBG_INITED,
516 ("%s: bad debugflags %d", __func__, ndp->ni_debugflags));
517 ndp->ni_debugflags |= NAMEI_DBG_CALLED;
518 if (ndp->ni_startdir != NULL)
519 ndp->ni_debugflags |= NAMEI_DBG_HADSTARTDIR;
520 if (cnp->cn_flags & FAILIFEXISTS) {
521 KASSERT(cnp->cn_nameiop == CREATE,
522 ("%s: FAILIFEXISTS passed for op %d", __func__, cnp->cn_nameiop));
523 /*
524 * The limitation below is to restrict hairy corner cases.
525 */
526 KASSERT((cnp->cn_flags & (LOCKPARENT | LOCKLEAF)) == LOCKPARENT,
527 ("%s: FAILIFEXISTS must be passed with LOCKPARENT and without LOCKLEAF",
528 __func__));
529 }
530 /*
531 * For NDVALIDATE.
532 *
533 * While NDINIT may seem like a more natural place to do it, there are
534 * callers which directly modify flags past invoking init.
535 */
536 cnp->cn_origflags = cnp->cn_flags;
537 #endif
538 ndp->ni_cnd.cn_cred = ndp->ni_cnd.cn_thread->td_ucred;
539 KASSERT(ndp->ni_resflags == 0, ("%s: garbage in ni_resflags: %x\n",
540 __func__, ndp->ni_resflags));
541 KASSERT(cnp->cn_cred && td->td_proc, ("namei: bad cred/proc"));
542 KASSERT((cnp->cn_flags & NAMEI_INTERNAL_FLAGS) == 0,
543 ("namei: unexpected flags: %" PRIx64 "\n",
544 cnp->cn_flags & NAMEI_INTERNAL_FLAGS));
545 if (cnp->cn_flags & NOCACHE)
546 KASSERT(cnp->cn_nameiop != LOOKUP,
547 ("%s: NOCACHE passed with LOOKUP", __func__));
548 MPASS(ndp->ni_startdir == NULL || ndp->ni_startdir->v_type == VDIR ||
549 ndp->ni_startdir->v_type == VBAD);
550
551 ndp->ni_lcf = 0;
552 ndp->ni_loopcnt = 0;
553 ndp->ni_vp = NULL;
554
555 error = namei_getpath(ndp);
556 if (__predict_false(error != 0)) {
557 return (error);
558 }
559
560 #ifdef KTRACE
561 if (KTRPOINT(td, KTR_NAMEI)) {
562 KASSERT(cnp->cn_thread == curthread,
563 ("namei not using curthread"));
564 ktrnamei(cnp->cn_pnbuf);
565 }
566 #endif
567
568 /*
569 * First try looking up the target without locking any vnodes.
570 *
571 * We may need to start from scratch or pick up where it left off.
572 */
573 error = cache_fplookup(ndp, &status, &pwd);
574 switch (status) {
575 case CACHE_FPL_STATUS_UNSET:
576 __assert_unreachable();
577 break;
578 case CACHE_FPL_STATUS_HANDLED:
579 if (error == 0)
580 NDVALIDATE(ndp);
581 return (error);
582 case CACHE_FPL_STATUS_PARTIAL:
583 TAILQ_INIT(&ndp->ni_cap_tracker);
584 dp = ndp->ni_startdir;
585 break;
586 case CACHE_FPL_STATUS_DESTROYED:
587 ndp->ni_loopcnt = 0;
588 error = namei_getpath(ndp);
589 if (__predict_false(error != 0)) {
590 return (error);
591 }
592 /* FALLTHROUGH */
593 case CACHE_FPL_STATUS_ABORTED:
594 TAILQ_INIT(&ndp->ni_cap_tracker);
595 MPASS(ndp->ni_lcf == 0);
596 error = namei_setup(ndp, &dp, &pwd);
597 if (error != 0) {
598 namei_cleanup_cnp(cnp);
599 return (error);
600 }
601 break;
602 }
603
604 /*
605 * Locked lookup.
606 */
607 for (;;) {
608 ndp->ni_startdir = dp;
609 error = lookup(ndp);
610 if (error != 0)
611 goto out;
612
613 /*
614 * If not a symbolic link, we're done.
615 */
616 if ((cnp->cn_flags & ISSYMLINK) == 0) {
617 SDT_PROBE4(vfs, namei, lookup, return, error,
618 (error == 0 ? ndp->ni_vp : NULL), false, ndp);
619 if ((cnp->cn_flags & (SAVENAME | SAVESTART)) == 0) {
620 namei_cleanup_cnp(cnp);
621 } else
622 cnp->cn_flags |= HASBUF;
623 nameicap_cleanup(ndp);
624 pwd_drop(pwd);
625 if (error == 0)
626 NDVALIDATE(ndp);
627 return (error);
628 }
629 if (ndp->ni_loopcnt++ >= MAXSYMLINKS) {
630 error = ELOOP;
631 break;
632 }
633 #ifdef MAC
634 if ((cnp->cn_flags & NOMACCHECK) == 0) {
635 error = mac_vnode_check_readlink(td->td_ucred,
636 ndp->ni_vp);
637 if (error != 0)
638 break;
639 }
640 #endif
641 if (ndp->ni_pathlen > 1)
642 cp = uma_zalloc(namei_zone, M_WAITOK);
643 else
644 cp = cnp->cn_pnbuf;
645 aiov.iov_base = cp;
646 aiov.iov_len = MAXPATHLEN;
647 auio.uio_iov = &aiov;
648 auio.uio_iovcnt = 1;
649 auio.uio_offset = 0;
650 auio.uio_rw = UIO_READ;
651 auio.uio_segflg = UIO_SYSSPACE;
652 auio.uio_td = td;
653 auio.uio_resid = MAXPATHLEN;
654 error = VOP_READLINK(ndp->ni_vp, &auio, cnp->cn_cred);
655 if (error != 0) {
656 if (ndp->ni_pathlen > 1)
657 uma_zfree(namei_zone, cp);
658 break;
659 }
660 linklen = MAXPATHLEN - auio.uio_resid;
661 if (linklen == 0) {
662 if (ndp->ni_pathlen > 1)
663 uma_zfree(namei_zone, cp);
664 error = ENOENT;
665 break;
666 }
667 if (linklen + ndp->ni_pathlen > MAXPATHLEN) {
668 if (ndp->ni_pathlen > 1)
669 uma_zfree(namei_zone, cp);
670 error = ENAMETOOLONG;
671 break;
672 }
673 if (ndp->ni_pathlen > 1) {
674 bcopy(ndp->ni_next, cp + linklen, ndp->ni_pathlen);
675 uma_zfree(namei_zone, cnp->cn_pnbuf);
676 cnp->cn_pnbuf = cp;
677 } else
678 cnp->cn_pnbuf[linklen] = '\0';
679 ndp->ni_pathlen += linklen;
680 vput(ndp->ni_vp);
681 dp = ndp->ni_dvp;
682 /*
683 * Check if root directory should replace current directory.
684 */
685 cnp->cn_nameptr = cnp->cn_pnbuf;
686 if (*(cnp->cn_nameptr) == '/') {
687 vrele(dp);
688 error = namei_handle_root(ndp, &dp);
689 if (error != 0)
690 goto out;
691 }
692 }
693 vput(ndp->ni_vp);
694 ndp->ni_vp = NULL;
695 vrele(ndp->ni_dvp);
696 out:
697 MPASS(error != 0);
698 SDT_PROBE4(vfs, namei, lookup, return, error, NULL, false, ndp);
699 namei_cleanup_cnp(cnp);
700 nameicap_cleanup(ndp);
701 pwd_drop(pwd);
702 return (error);
703 }
704
705 static int
706 compute_cn_lkflags(struct mount *mp, int lkflags, int cnflags)
707 {
708
709 if (mp == NULL || ((lkflags & LK_SHARED) &&
710 (!(mp->mnt_kern_flag & MNTK_LOOKUP_SHARED) ||
711 ((cnflags & ISDOTDOT) &&
712 (mp->mnt_kern_flag & MNTK_LOOKUP_EXCL_DOTDOT))))) {
713 lkflags &= ~LK_SHARED;
714 lkflags |= LK_EXCLUSIVE;
715 }
716 lkflags |= LK_NODDLKTREAT;
717 return (lkflags);
718 }
719
720 static __inline int
721 needs_exclusive_leaf(struct mount *mp, int flags)
722 {
723
724 /*
725 * Intermediate nodes can use shared locks, we only need to
726 * force an exclusive lock for leaf nodes.
727 */
728 if ((flags & (ISLASTCN | LOCKLEAF)) != (ISLASTCN | LOCKLEAF))
729 return (0);
730
731 /* Always use exclusive locks if LOCKSHARED isn't set. */
732 if (!(flags & LOCKSHARED))
733 return (1);
734
735 /*
736 * For lookups during open(), if the mount point supports
737 * extended shared operations, then use a shared lock for the
738 * leaf node, otherwise use an exclusive lock.
739 */
740 if ((flags & ISOPEN) != 0)
741 return (!MNT_EXTENDED_SHARED(mp));
742
743 /*
744 * Lookup requests outside of open() that specify LOCKSHARED
745 * only need a shared lock on the leaf vnode.
746 */
747 return (0);
748 }
749
750 /*
751 * Search a pathname.
752 * This is a very central and rather complicated routine.
753 *
754 * The pathname is pointed to by ni_ptr and is of length ni_pathlen.
755 * The starting directory is taken from ni_startdir. The pathname is
756 * descended until done, or a symbolic link is encountered. The variable
757 * ni_more is clear if the path is completed; it is set to one if a
758 * symbolic link needing interpretation is encountered.
759 *
760 * The flag argument is LOOKUP, CREATE, RENAME, or DELETE depending on
761 * whether the name is to be looked up, created, renamed, or deleted.
762 * When CREATE, RENAME, or DELETE is specified, information usable in
763 * creating, renaming, or deleting a directory entry may be calculated.
764 * If flag has LOCKPARENT or'ed into it, the parent directory is returned
765 * locked. If flag has WANTPARENT or'ed into it, the parent directory is
766 * returned unlocked. Otherwise the parent directory is not returned. If
767 * the target of the pathname exists and LOCKLEAF is or'ed into the flag
768 * the target is returned locked, otherwise it is returned unlocked.
769 * When creating or renaming and LOCKPARENT is specified, the target may not
770 * be ".". When deleting and LOCKPARENT is specified, the target may be ".".
771 *
772 * Overall outline of lookup:
773 *
774 * dirloop:
775 * identify next component of name at ndp->ni_ptr
776 * handle degenerate case where name is null string
777 * if .. and crossing mount points and on mounted filesys, find parent
778 * call VOP_LOOKUP routine for next component name
779 * directory vnode returned in ni_dvp, unlocked unless LOCKPARENT set
780 * component vnode returned in ni_vp (if it exists), locked.
781 * if result vnode is mounted on and crossing mount points,
782 * find mounted on vnode
783 * if more components of name, do next level at dirloop
784 * return the answer in ni_vp, locked if LOCKLEAF set
785 * if LOCKPARENT set, return locked parent in ni_dvp
786 * if WANTPARENT set, return unlocked parent in ni_dvp
787 */
788 int
789 lookup(struct nameidata *ndp)
790 {
791 char *cp; /* pointer into pathname argument */
792 char *prev_ni_next; /* saved ndp->ni_next */
793 struct vnode *dp = NULL; /* the directory we are searching */
794 struct vnode *tdp; /* saved dp */
795 struct mount *mp; /* mount table entry */
796 struct prison *pr;
797 size_t prev_ni_pathlen; /* saved ndp->ni_pathlen */
798 int docache; /* == 0 do not cache last component */
799 int wantparent; /* 1 => wantparent or lockparent flag */
800 int rdonly; /* lookup read-only flag bit */
801 int error = 0;
802 int dpunlocked = 0; /* dp has already been unlocked */
803 int relookup = 0; /* do not consume the path component */
804 struct componentname *cnp = &ndp->ni_cnd;
805 int lkflags_save;
806 int ni_dvp_unlocked;
807
808 /*
809 * Setup: break out flag bits into variables.
810 */
811 ni_dvp_unlocked = 0;
812 wantparent = cnp->cn_flags & (LOCKPARENT | WANTPARENT);
813 KASSERT(cnp->cn_nameiop == LOOKUP || wantparent,
814 ("CREATE, DELETE, RENAME require LOCKPARENT or WANTPARENT."));
815 /*
816 * When set to zero, docache causes the last component of the
817 * pathname to be deleted from the cache and the full lookup
818 * of the name to be done (via VOP_CACHEDLOOKUP()). Often
819 * filesystems need some pre-computed values that are made
820 * during the full lookup, for instance UFS sets dp->i_offset.
821 *
822 * The docache variable is set to zero when requested by the
823 * NOCACHE flag and for all modifying operations except CREATE.
824 */
825 docache = (cnp->cn_flags & NOCACHE) ^ NOCACHE;
826 if (cnp->cn_nameiop == DELETE ||
827 (wantparent && cnp->cn_nameiop != CREATE &&
828 cnp->cn_nameiop != LOOKUP))
829 docache = 0;
830 rdonly = cnp->cn_flags & RDONLY;
831 cnp->cn_flags &= ~ISSYMLINK;
832 ndp->ni_dvp = NULL;
833 /*
834 * We use shared locks until we hit the parent of the last cn then
835 * we adjust based on the requesting flags.
836 */
837 cnp->cn_lkflags = LK_SHARED;
838 dp = ndp->ni_startdir;
839 ndp->ni_startdir = NULLVP;
840 vn_lock(dp,
841 compute_cn_lkflags(dp->v_mount, cnp->cn_lkflags | LK_RETRY,
842 cnp->cn_flags));
843
844 dirloop:
845 /*
846 * Search a new directory.
847 *
848 * The last component of the filename is left accessible via
849 * cnp->cn_nameptr for callers that need the name. Callers needing
850 * the name set the SAVENAME flag. When done, they assume
851 * responsibility for freeing the pathname buffer.
852 */
853 for (cp = cnp->cn_nameptr; *cp != 0 && *cp != '/'; cp++)
854 continue;
855 cnp->cn_namelen = cp - cnp->cn_nameptr;
856 if (cnp->cn_namelen > NAME_MAX) {
857 error = ENAMETOOLONG;
858 goto bad;
859 }
860 #ifdef NAMEI_DIAGNOSTIC
861 { char c = *cp;
862 *cp = '\0';
863 printf("{%s}: ", cnp->cn_nameptr);
864 *cp = c; }
865 #endif
866 prev_ni_pathlen = ndp->ni_pathlen;
867 ndp->ni_pathlen -= cnp->cn_namelen;
868 KASSERT(ndp->ni_pathlen <= PATH_MAX,
869 ("%s: ni_pathlen underflow to %zd\n", __func__, ndp->ni_pathlen));
870 prev_ni_next = ndp->ni_next;
871 ndp->ni_next = cp;
872
873 /*
874 * Replace multiple slashes by a single slash and trailing slashes
875 * by a null. This must be done before VOP_LOOKUP() because some
876 * fs's don't know about trailing slashes. Remember if there were
877 * trailing slashes to handle symlinks, existing non-directories
878 * and non-existing files that won't be directories specially later.
879 */
880 while (*cp == '/' && (cp[1] == '/' || cp[1] == '\0')) {
881 cp++;
882 ndp->ni_pathlen--;
883 if (*cp == '\0') {
884 *ndp->ni_next = '\0';
885 cnp->cn_flags |= TRAILINGSLASH;
886 }
887 }
888 ndp->ni_next = cp;
889
890 cnp->cn_flags |= MAKEENTRY;
891 if (*cp == '\0' && docache == 0)
892 cnp->cn_flags &= ~MAKEENTRY;
893 if (cnp->cn_namelen == 2 &&
894 cnp->cn_nameptr[1] == '.' && cnp->cn_nameptr[0] == '.')
895 cnp->cn_flags |= ISDOTDOT;
896 else
897 cnp->cn_flags &= ~ISDOTDOT;
898 if (*ndp->ni_next == 0)
899 cnp->cn_flags |= ISLASTCN;
900 else
901 cnp->cn_flags &= ~ISLASTCN;
902
903 if ((cnp->cn_flags & ISLASTCN) != 0 &&
904 cnp->cn_namelen == 1 && cnp->cn_nameptr[0] == '.' &&
905 (cnp->cn_nameiop == DELETE || cnp->cn_nameiop == RENAME)) {
906 error = EINVAL;
907 goto bad;
908 }
909
910 nameicap_tracker_add(ndp, dp);
911
912 /*
913 * Check for degenerate name (e.g. / or "")
914 * which is a way of talking about a directory,
915 * e.g. like "/." or ".".
916 */
917 if (cnp->cn_nameptr[0] == '\0') {
918 if (dp->v_type != VDIR) {
919 error = ENOTDIR;
920 goto bad;
921 }
922 if (cnp->cn_nameiop != LOOKUP) {
923 error = EISDIR;
924 goto bad;
925 }
926 if (wantparent) {
927 ndp->ni_dvp = dp;
928 VREF(dp);
929 }
930 ndp->ni_vp = dp;
931
932 if (cnp->cn_flags & AUDITVNODE1)
933 AUDIT_ARG_VNODE1(dp);
934 else if (cnp->cn_flags & AUDITVNODE2)
935 AUDIT_ARG_VNODE2(dp);
936
937 if (!(cnp->cn_flags & (LOCKPARENT | LOCKLEAF)))
938 VOP_UNLOCK(dp);
939 /* XXX This should probably move to the top of function. */
940 if (cnp->cn_flags & SAVESTART)
941 panic("lookup: SAVESTART");
942 goto success;
943 }
944
945 /*
946 * Handle "..": five special cases.
947 * 0. If doing a capability lookup and lookup_cap_dotdot is
948 * disabled, return ENOTCAPABLE.
949 * 1. Return an error if this is the last component of
950 * the name and the operation is DELETE or RENAME.
951 * 2. If at root directory (e.g. after chroot)
952 * or at absolute root directory
953 * then ignore it so can't get out.
954 * 3. If this vnode is the root of a mounted
955 * filesystem, then replace it with the
956 * vnode which was mounted on so we take the
957 * .. in the other filesystem.
958 * 4. If the vnode is the top directory of
959 * the jail or chroot, don't let them out.
960 * 5. If doing a capability lookup and lookup_cap_dotdot is
961 * enabled, return ENOTCAPABLE if the lookup would escape
962 * from the initial file descriptor directory. Checks are
963 * done by ensuring that namei() already traversed the
964 * result of dotdot lookup.
965 */
966 if (cnp->cn_flags & ISDOTDOT) {
967 if ((ndp->ni_lcf & (NI_LCF_STRICTRELATIVE | NI_LCF_CAP_DOTDOT))
968 == NI_LCF_STRICTRELATIVE) {
969 #ifdef KTRACE
970 if (KTRPOINT(curthread, KTR_CAPFAIL))
971 ktrcapfail(CAPFAIL_LOOKUP, NULL, NULL);
972 #endif
973 error = ENOTCAPABLE;
974 goto bad;
975 }
976 if ((cnp->cn_flags & ISLASTCN) != 0 &&
977 (cnp->cn_nameiop == DELETE || cnp->cn_nameiop == RENAME)) {
978 error = EINVAL;
979 goto bad;
980 }
981 for (;;) {
982 for (pr = cnp->cn_cred->cr_prison; pr != NULL;
983 pr = pr->pr_parent)
984 if (dp == pr->pr_root)
985 break;
986 if (dp == ndp->ni_rootdir ||
987 dp == ndp->ni_topdir ||
988 dp == rootvnode ||
989 pr != NULL ||
990 ((dp->v_vflag & VV_ROOT) != 0 &&
991 (cnp->cn_flags & NOCROSSMOUNT) != 0)) {
992 ndp->ni_dvp = dp;
993 ndp->ni_vp = dp;
994 VREF(dp);
995 goto nextname;
996 }
997 if ((dp->v_vflag & VV_ROOT) == 0)
998 break;
999 if (VN_IS_DOOMED(dp)) { /* forced unmount */
1000 error = ENOENT;
1001 goto bad;
1002 }
1003 tdp = dp;
1004 dp = dp->v_mount->mnt_vnodecovered;
1005 VREF(dp);
1006 vput(tdp);
1007 vn_lock(dp,
1008 compute_cn_lkflags(dp->v_mount, cnp->cn_lkflags |
1009 LK_RETRY, ISDOTDOT));
1010 error = nameicap_check_dotdot(ndp, dp);
1011 if (error != 0) {
1012 #ifdef KTRACE
1013 if (KTRPOINT(curthread, KTR_CAPFAIL))
1014 ktrcapfail(CAPFAIL_LOOKUP, NULL, NULL);
1015 #endif
1016 goto bad;
1017 }
1018 }
1019 }
1020
1021 /*
1022 * We now have a segment name to search for, and a directory to search.
1023 */
1024 unionlookup:
1025 #ifdef MAC
1026 error = mac_vnode_check_lookup(cnp->cn_thread->td_ucred, dp, cnp);
1027 if (error)
1028 goto bad;
1029 #endif
1030 ndp->ni_dvp = dp;
1031 ndp->ni_vp = NULL;
1032 ASSERT_VOP_LOCKED(dp, "lookup");
1033 /*
1034 * If we have a shared lock we may need to upgrade the lock for the
1035 * last operation.
1036 */
1037 if ((cnp->cn_flags & LOCKPARENT) && (cnp->cn_flags & ISLASTCN) &&
1038 dp != vp_crossmp && VOP_ISLOCKED(dp) == LK_SHARED)
1039 vn_lock(dp, LK_UPGRADE|LK_RETRY);
1040 if (VN_IS_DOOMED(dp)) {
1041 error = ENOENT;
1042 goto bad;
1043 }
1044 /*
1045 * If we're looking up the last component and we need an exclusive
1046 * lock, adjust our lkflags.
1047 */
1048 if (needs_exclusive_leaf(dp->v_mount, cnp->cn_flags))
1049 cnp->cn_lkflags = LK_EXCLUSIVE;
1050 #ifdef NAMEI_DIAGNOSTIC
1051 vn_printf(dp, "lookup in ");
1052 #endif
1053 lkflags_save = cnp->cn_lkflags;
1054 cnp->cn_lkflags = compute_cn_lkflags(dp->v_mount, cnp->cn_lkflags,
1055 cnp->cn_flags);
1056 error = VOP_LOOKUP(dp, &ndp->ni_vp, cnp);
1057 cnp->cn_lkflags = lkflags_save;
1058 if (error != 0) {
1059 KASSERT(ndp->ni_vp == NULL, ("leaf should be empty"));
1060 #ifdef NAMEI_DIAGNOSTIC
1061 printf("not found\n");
1062 #endif
1063 if ((error == ENOENT) &&
1064 (dp->v_vflag & VV_ROOT) && (dp->v_mount != NULL) &&
1065 (dp->v_mount->mnt_flag & MNT_UNION)) {
1066 tdp = dp;
1067 dp = dp->v_mount->mnt_vnodecovered;
1068 VREF(dp);
1069 vput(tdp);
1070 vn_lock(dp,
1071 compute_cn_lkflags(dp->v_mount, cnp->cn_lkflags |
1072 LK_RETRY, cnp->cn_flags));
1073 nameicap_tracker_add(ndp, dp);
1074 goto unionlookup;
1075 }
1076
1077 if (error == ERELOOKUP) {
1078 vref(dp);
1079 ndp->ni_vp = dp;
1080 error = 0;
1081 relookup = 1;
1082 goto good;
1083 }
1084
1085 if (error != EJUSTRETURN)
1086 goto bad;
1087 /*
1088 * At this point, we know we're at the end of the
1089 * pathname. If creating / renaming, we can consider
1090 * allowing the file or directory to be created / renamed,
1091 * provided we're not on a read-only filesystem.
1092 */
1093 if (rdonly) {
1094 error = EROFS;
1095 goto bad;
1096 }
1097 /* trailing slash only allowed for directories */
1098 if ((cnp->cn_flags & TRAILINGSLASH) &&
1099 !(cnp->cn_flags & WILLBEDIR)) {
1100 error = ENOENT;
1101 goto bad;
1102 }
1103 if ((cnp->cn_flags & LOCKPARENT) == 0)
1104 VOP_UNLOCK(dp);
1105 /*
1106 * We return with ni_vp NULL to indicate that the entry
1107 * doesn't currently exist, leaving a pointer to the
1108 * (possibly locked) directory vnode in ndp->ni_dvp.
1109 */
1110 if (cnp->cn_flags & SAVESTART) {
1111 ndp->ni_startdir = ndp->ni_dvp;
1112 VREF(ndp->ni_startdir);
1113 }
1114 goto success;
1115 }
1116
1117 good:
1118 #ifdef NAMEI_DIAGNOSTIC
1119 printf("found\n");
1120 #endif
1121 dp = ndp->ni_vp;
1122
1123 /*
1124 * Check to see if the vnode has been mounted on;
1125 * if so find the root of the mounted filesystem.
1126 */
1127 while (dp->v_type == VDIR && (mp = dp->v_mountedhere) &&
1128 (cnp->cn_flags & NOCROSSMOUNT) == 0) {
1129 if (vfs_busy(mp, 0))
1130 continue;
1131 vput(dp);
1132 if (dp != ndp->ni_dvp)
1133 vput(ndp->ni_dvp);
1134 else
1135 vrele(ndp->ni_dvp);
1136 vrefact(vp_crossmp);
1137 ndp->ni_dvp = vp_crossmp;
1138 error = VFS_ROOT(mp, compute_cn_lkflags(mp, cnp->cn_lkflags,
1139 cnp->cn_flags), &tdp);
1140 vfs_unbusy(mp);
1141 if (vn_lock(vp_crossmp, LK_SHARED | LK_NOWAIT))
1142 panic("vp_crossmp exclusively locked or reclaimed");
1143 if (error) {
1144 dpunlocked = 1;
1145 goto bad2;
1146 }
1147 ndp->ni_vp = dp = tdp;
1148 }
1149
1150 /*
1151 * Check for symbolic link
1152 */
1153 if ((dp->v_type == VLNK) &&
1154 ((cnp->cn_flags & FOLLOW) || (cnp->cn_flags & TRAILINGSLASH) ||
1155 *ndp->ni_next == '/')) {
1156 cnp->cn_flags |= ISSYMLINK;
1157 if (VN_IS_DOOMED(dp)) {
1158 /*
1159 * We can't know whether the directory was mounted with
1160 * NOSYMFOLLOW, so we can't follow safely.
1161 */
1162 error = ENOENT;
1163 goto bad2;
1164 }
1165 if (dp->v_mount->mnt_flag & MNT_NOSYMFOLLOW) {
1166 error = EACCES;
1167 goto bad2;
1168 }
1169 /*
1170 * Symlink code always expects an unlocked dvp.
1171 */
1172 if (ndp->ni_dvp != ndp->ni_vp) {
1173 VOP_UNLOCK(ndp->ni_dvp);
1174 ni_dvp_unlocked = 1;
1175 }
1176 goto success;
1177 }
1178
1179 nextname:
1180 /*
1181 * Not a symbolic link that we will follow. Continue with the
1182 * next component if there is any; otherwise, we're done.
1183 */
1184 KASSERT((cnp->cn_flags & ISLASTCN) || *ndp->ni_next == '/',
1185 ("lookup: invalid path state."));
1186 if (relookup) {
1187 relookup = 0;
1188 ndp->ni_pathlen = prev_ni_pathlen;
1189 ndp->ni_next = prev_ni_next;
1190 if (ndp->ni_dvp != dp)
1191 vput(ndp->ni_dvp);
1192 else
1193 vrele(ndp->ni_dvp);
1194 goto dirloop;
1195 }
1196 if (cnp->cn_flags & ISDOTDOT) {
1197 error = nameicap_check_dotdot(ndp, ndp->ni_vp);
1198 if (error != 0) {
1199 #ifdef KTRACE
1200 if (KTRPOINT(curthread, KTR_CAPFAIL))
1201 ktrcapfail(CAPFAIL_LOOKUP, NULL, NULL);
1202 #endif
1203 goto bad2;
1204 }
1205 }
1206 if (*ndp->ni_next == '/') {
1207 cnp->cn_nameptr = ndp->ni_next;
1208 while (*cnp->cn_nameptr == '/') {
1209 cnp->cn_nameptr++;
1210 ndp->ni_pathlen--;
1211 }
1212 if (ndp->ni_dvp != dp)
1213 vput(ndp->ni_dvp);
1214 else
1215 vrele(ndp->ni_dvp);
1216 goto dirloop;
1217 }
1218 /*
1219 * If we're processing a path with a trailing slash,
1220 * check that the end result is a directory.
1221 */
1222 if ((cnp->cn_flags & TRAILINGSLASH) && dp->v_type != VDIR) {
1223 error = ENOTDIR;
1224 goto bad2;
1225 }
1226 /*
1227 * Disallow directory write attempts on read-only filesystems.
1228 */
1229 if (rdonly &&
1230 (cnp->cn_nameiop == DELETE || cnp->cn_nameiop == RENAME)) {
1231 error = EROFS;
1232 goto bad2;
1233 }
1234 if (cnp->cn_flags & SAVESTART) {
1235 ndp->ni_startdir = ndp->ni_dvp;
1236 VREF(ndp->ni_startdir);
1237 }
1238 if (!wantparent) {
1239 ni_dvp_unlocked = 2;
1240 if (ndp->ni_dvp != dp)
1241 vput(ndp->ni_dvp);
1242 else
1243 vrele(ndp->ni_dvp);
1244 } else if ((cnp->cn_flags & LOCKPARENT) == 0 && ndp->ni_dvp != dp) {
1245 VOP_UNLOCK(ndp->ni_dvp);
1246 ni_dvp_unlocked = 1;
1247 }
1248
1249 if (cnp->cn_flags & AUDITVNODE1)
1250 AUDIT_ARG_VNODE1(dp);
1251 else if (cnp->cn_flags & AUDITVNODE2)
1252 AUDIT_ARG_VNODE2(dp);
1253
1254 if ((cnp->cn_flags & LOCKLEAF) == 0)
1255 VOP_UNLOCK(dp);
1256 success:
1257 /*
1258 * FIXME: for lookups which only cross a mount point to fetch the
1259 * root vnode, ni_dvp will be set to vp_crossmp. This can be a problem
1260 * if either WANTPARENT or LOCKPARENT is set.
1261 */
1262 /*
1263 * Because of shared lookup we may have the vnode shared locked, but
1264 * the caller may want it to be exclusively locked.
1265 */
1266 if (needs_exclusive_leaf(dp->v_mount, cnp->cn_flags) &&
1267 VOP_ISLOCKED(dp) != LK_EXCLUSIVE) {
1268 vn_lock(dp, LK_UPGRADE | LK_RETRY);
1269 if (VN_IS_DOOMED(dp)) {
1270 error = ENOENT;
1271 goto bad2;
1272 }
1273 }
1274 if (ndp->ni_vp != NULL) {
1275 if ((cnp->cn_flags & ISDOTDOT) == 0)
1276 nameicap_tracker_add(ndp, ndp->ni_vp);
1277 if ((cnp->cn_flags & (FAILIFEXISTS | ISSYMLINK)) == FAILIFEXISTS)
1278 goto bad_eexist;
1279 }
1280 return (0);
1281
1282 bad2:
1283 if (ni_dvp_unlocked != 2) {
1284 if (dp != ndp->ni_dvp && !ni_dvp_unlocked)
1285 vput(ndp->ni_dvp);
1286 else
1287 vrele(ndp->ni_dvp);
1288 }
1289 bad:
1290 if (!dpunlocked)
1291 vput(dp);
1292 ndp->ni_vp = NULL;
1293 return (error);
1294 bad_eexist:
1295 /*
1296 * FAILIFEXISTS handling.
1297 *
1298 * XXX namei called with LOCKPARENT but not LOCKLEAF has the strange
1299 * behaviour of leaving the vnode unlocked if the target is the same
1300 * vnode as the parent.
1301 */
1302 MPASS((cnp->cn_flags & ISSYMLINK) == 0);
1303 if (ndp->ni_vp == ndp->ni_dvp)
1304 vrele(ndp->ni_dvp);
1305 else
1306 vput(ndp->ni_dvp);
1307 vrele(ndp->ni_vp);
1308 ndp->ni_dvp = NULL;
1309 ndp->ni_vp = NULL;
1310 NDFREE(ndp, NDF_ONLY_PNBUF);
1311 return (EEXIST);
1312 }
1313
1314 /*
1315 * relookup - lookup a path name component
1316 * Used by lookup to re-acquire things.
1317 */
1318 int
1319 relookup(struct vnode *dvp, struct vnode **vpp, struct componentname *cnp)
1320 {
1321 struct vnode *dp = NULL; /* the directory we are searching */
1322 int rdonly; /* lookup read-only flag bit */
1323 int error = 0;
1324
1325 KASSERT(cnp->cn_flags & ISLASTCN,
1326 ("relookup: Not given last component."));
1327 /*
1328 * Setup: break out flag bits into variables.
1329 */
1330 KASSERT((cnp->cn_flags & (LOCKPARENT | WANTPARENT)) != 0,
1331 ("relookup: parent not wanted"));
1332 rdonly = cnp->cn_flags & RDONLY;
1333 cnp->cn_flags &= ~ISSYMLINK;
1334 dp = dvp;
1335 cnp->cn_lkflags = LK_EXCLUSIVE;
1336 vn_lock(dp, LK_EXCLUSIVE | LK_RETRY);
1337
1338 /*
1339 * Search a new directory.
1340 *
1341 * The last component of the filename is left accessible via
1342 * cnp->cn_nameptr for callers that need the name. Callers needing
1343 * the name set the SAVENAME flag. When done, they assume
1344 * responsibility for freeing the pathname buffer.
1345 */
1346 #ifdef NAMEI_DIAGNOSTIC
1347 printf("{%s}: ", cnp->cn_nameptr);
1348 #endif
1349
1350 /*
1351 * Check for "" which represents the root directory after slash
1352 * removal.
1353 */
1354 if (cnp->cn_nameptr[0] == '\0') {
1355 /*
1356 * Support only LOOKUP for "/" because lookup()
1357 * can't succeed for CREATE, DELETE and RENAME.
1358 */
1359 KASSERT(cnp->cn_nameiop == LOOKUP, ("nameiop must be LOOKUP"));
1360 KASSERT(dp->v_type == VDIR, ("dp is not a directory"));
1361
1362 if (!(cnp->cn_flags & LOCKLEAF))
1363 VOP_UNLOCK(dp);
1364 *vpp = dp;
1365 /* XXX This should probably move to the top of function. */
1366 if (cnp->cn_flags & SAVESTART)
1367 panic("lookup: SAVESTART");
1368 return (0);
1369 }
1370
1371 if (cnp->cn_flags & ISDOTDOT)
1372 panic ("relookup: lookup on dot-dot");
1373
1374 /*
1375 * We now have a segment name to search for, and a directory to search.
1376 */
1377 #ifdef NAMEI_DIAGNOSTIC
1378 vn_printf(dp, "search in ");
1379 #endif
1380 if ((error = VOP_LOOKUP(dp, vpp, cnp)) != 0) {
1381 KASSERT(*vpp == NULL, ("leaf should be empty"));
1382 if (error != EJUSTRETURN)
1383 goto bad;
1384 /*
1385 * If creating and at end of pathname, then can consider
1386 * allowing file to be created.
1387 */
1388 if (rdonly) {
1389 error = EROFS;
1390 goto bad;
1391 }
1392 /* ASSERT(dvp == ndp->ni_startdir) */
1393 if (cnp->cn_flags & SAVESTART)
1394 VREF(dvp);
1395 if ((cnp->cn_flags & LOCKPARENT) == 0)
1396 VOP_UNLOCK(dp);
1397 /*
1398 * We return with ni_vp NULL to indicate that the entry
1399 * doesn't currently exist, leaving a pointer to the
1400 * (possibly locked) directory vnode in ndp->ni_dvp.
1401 */
1402 return (0);
1403 }
1404
1405 dp = *vpp;
1406
1407 /*
1408 * Disallow directory write attempts on read-only filesystems.
1409 */
1410 if (rdonly &&
1411 (cnp->cn_nameiop == DELETE || cnp->cn_nameiop == RENAME)) {
1412 if (dvp == dp)
1413 vrele(dvp);
1414 else
1415 vput(dvp);
1416 error = EROFS;
1417 goto bad;
1418 }
1419 /*
1420 * Set the parent lock/ref state to the requested state.
1421 */
1422 if ((cnp->cn_flags & LOCKPARENT) == 0 && dvp != dp)
1423 VOP_UNLOCK(dvp);
1424 /*
1425 * Check for symbolic link
1426 */
1427 KASSERT(dp->v_type != VLNK || !(cnp->cn_flags & FOLLOW),
1428 ("relookup: symlink found.\n"));
1429
1430 /* ASSERT(dvp == ndp->ni_startdir) */
1431 if (cnp->cn_flags & SAVESTART)
1432 VREF(dvp);
1433
1434 if ((cnp->cn_flags & LOCKLEAF) == 0)
1435 VOP_UNLOCK(dp);
1436 return (0);
1437 bad:
1438 vput(dp);
1439 *vpp = NULL;
1440 return (error);
1441 }
1442
1443 /*
1444 * Free data allocated by namei(); see namei(9) for details.
1445 */
1446 void
1447 NDFREE_PNBUF(struct nameidata *ndp)
1448 {
1449
1450 if ((ndp->ni_cnd.cn_flags & HASBUF) != 0) {
1451 MPASS((ndp->ni_cnd.cn_flags & (SAVENAME | SAVESTART)) != 0);
1452 uma_zfree(namei_zone, ndp->ni_cnd.cn_pnbuf);
1453 ndp->ni_cnd.cn_flags &= ~HASBUF;
1454 }
1455 }
1456
1457 /*
1458 * NDFREE_PNBUF replacement for callers that know there is no buffer.
1459 *
1460 * This is a hack. Preferably the VFS layer would not produce anything more
1461 * than it was asked to do. Unfortunately several non-LOOKUP cases can add the
1462 * HASBUF flag to the result. Even then an interface could be implemented where
1463 * the caller specifies what they expected to see in the result and what they
1464 * are going to take care of.
1465 *
1466 * In the meantime provide this kludge as a trivial replacement for NDFREE_PNBUF
1467 * calls scattered throughout the kernel where we know for a fact the flag must not
1468 * be seen.
1469 */
1470 #ifdef INVARIANTS
1471 void
1472 NDFREE_NOTHING(struct nameidata *ndp)
1473 {
1474 struct componentname *cnp;
1475
1476 cnp = &ndp->ni_cnd;
1477 KASSERT(cnp->cn_nameiop == LOOKUP, ("%s: got non-LOOKUP op %d\n",
1478 __func__, cnp->cn_nameiop));
1479 KASSERT((cnp->cn_flags & (SAVENAME | HASBUF)) == 0,
1480 ("%s: bad flags \%" PRIx64 "\n", __func__, cnp->cn_flags));
1481 }
1482 #endif
1483
1484 void
1485 (NDFREE)(struct nameidata *ndp, const u_int flags)
1486 {
1487 int unlock_dvp;
1488 int unlock_vp;
1489
1490 unlock_dvp = 0;
1491 unlock_vp = 0;
1492
1493 if (!(flags & NDF_NO_FREE_PNBUF)) {
1494 NDFREE_PNBUF(ndp);
1495 }
1496 if (!(flags & NDF_NO_VP_UNLOCK) &&
1497 (ndp->ni_cnd.cn_flags & LOCKLEAF) && ndp->ni_vp)
1498 unlock_vp = 1;
1499 if (!(flags & NDF_NO_DVP_UNLOCK) &&
1500 (ndp->ni_cnd.cn_flags & LOCKPARENT) &&
1501 ndp->ni_dvp != ndp->ni_vp)
1502 unlock_dvp = 1;
1503 if (!(flags & NDF_NO_VP_RELE) && ndp->ni_vp) {
1504 if (unlock_vp) {
1505 vput(ndp->ni_vp);
1506 unlock_vp = 0;
1507 } else
1508 vrele(ndp->ni_vp);
1509 ndp->ni_vp = NULL;
1510 }
1511 if (unlock_vp)
1512 VOP_UNLOCK(ndp->ni_vp);
1513 if (!(flags & NDF_NO_DVP_RELE) &&
1514 (ndp->ni_cnd.cn_flags & (LOCKPARENT|WANTPARENT))) {
1515 if (unlock_dvp) {
1516 vput(ndp->ni_dvp);
1517 unlock_dvp = 0;
1518 } else
1519 vrele(ndp->ni_dvp);
1520 ndp->ni_dvp = NULL;
1521 }
1522 if (unlock_dvp)
1523 VOP_UNLOCK(ndp->ni_dvp);
1524 if (!(flags & NDF_NO_STARTDIR_RELE) &&
1525 (ndp->ni_cnd.cn_flags & SAVESTART)) {
1526 vrele(ndp->ni_startdir);
1527 ndp->ni_startdir = NULL;
1528 }
1529 }
1530
1531 #ifdef INVARIANTS
1532 /*
1533 * Validate the final state of ndp after the lookup.
1534 *
1535 * Historically filesystems were allowed to modify cn_flags. Most notably they
1536 * can add SAVENAME to the request, resulting in HASBUF and pushing subsequent
1537 * clean up to the consumer. In practice this seems to only concern != LOOKUP
1538 * operations.
1539 *
1540 * As a step towards stricter API contract this routine validates the state to
1541 * clean up. Note validation is a work in progress with the intent of becoming
1542 * stricter over time.
1543 */
1544 #define NDMODIFYINGFLAGS (LOCKLEAF | LOCKPARENT | WANTPARENT | SAVENAME | SAVESTART | HASBUF)
1545 void
1546 NDVALIDATE(struct nameidata *ndp)
1547 {
1548 struct componentname *cnp;
1549 u_int64_t used, orig;
1550
1551 cnp = &ndp->ni_cnd;
1552 orig = cnp->cn_origflags;
1553 used = cnp->cn_flags;
1554 switch (cnp->cn_nameiop) {
1555 case LOOKUP:
1556 /*
1557 * For plain lookup we require strict conformance -- nothing
1558 * to clean up if it was not requested by the caller.
1559 */
1560 orig &= NDMODIFYINGFLAGS;
1561 used &= NDMODIFYINGFLAGS;
1562 if ((orig & (SAVENAME | SAVESTART)) != 0)
1563 orig |= HASBUF;
1564 if (orig != used) {
1565 goto out_mismatch;
1566 }
1567 break;
1568 case CREATE:
1569 case DELETE:
1570 case RENAME:
1571 /*
1572 * Some filesystems set SAVENAME to provoke HASBUF, accomodate
1573 * for it until it gets fixed.
1574 */
1575 orig &= NDMODIFYINGFLAGS;
1576 orig |= (SAVENAME | HASBUF);
1577 used &= NDMODIFYINGFLAGS;
1578 used |= (SAVENAME | HASBUF);
1579 if (orig != used) {
1580 goto out_mismatch;
1581 }
1582 break;
1583 }
1584 return;
1585 out_mismatch:
1586 panic("%s: mismatched flags for op %d: added %" PRIx64 ", "
1587 "removed %" PRIx64" (%" PRIx64" != %" PRIx64"; stored %" PRIx64" != %" PRIx64")",
1588 __func__, cnp->cn_nameiop, used & ~orig, orig &~ used,
1589 orig, used, cnp->cn_origflags, cnp->cn_flags);
1590 }
1591 #endif
1592
1593 /*
1594 * Determine if there is a suitable alternate filename under the specified
1595 * prefix for the specified path. If the create flag is set, then the
1596 * alternate prefix will be used so long as the parent directory exists.
1597 * This is used by the various compatibility ABIs so that Linux binaries prefer
1598 * files under /compat/linux for example. The chosen path (whether under
1599 * the prefix or under /) is returned in a kernel malloc'd buffer pointed
1600 * to by pathbuf. The caller is responsible for free'ing the buffer from
1601 * the M_TEMP bucket if one is returned.
1602 */
1603 int
1604 kern_alternate_path(struct thread *td, const char *prefix, const char *path,
1605 enum uio_seg pathseg, char **pathbuf, int create, int dirfd)
1606 {
1607 struct nameidata nd, ndroot;
1608 char *ptr, *buf, *cp;
1609 size_t len, sz;
1610 int error;
1611
1612 buf = (char *) malloc(MAXPATHLEN, M_TEMP, M_WAITOK);
1613 *pathbuf = buf;
1614
1615 /* Copy the prefix into the new pathname as a starting point. */
1616 len = strlcpy(buf, prefix, MAXPATHLEN);
1617 if (len >= MAXPATHLEN) {
1618 *pathbuf = NULL;
1619 free(buf, M_TEMP);
1620 return (EINVAL);
1621 }
1622 sz = MAXPATHLEN - len;
1623 ptr = buf + len;
1624
1625 /* Append the filename to the prefix. */
1626 if (pathseg == UIO_SYSSPACE)
1627 error = copystr(path, ptr, sz, &len);
1628 else
1629 error = copyinstr(path, ptr, sz, &len);
1630
1631 if (error) {
1632 *pathbuf = NULL;
1633 free(buf, M_TEMP);
1634 return (error);
1635 }
1636
1637 /* Only use a prefix with absolute pathnames. */
1638 if (*ptr != '/') {
1639 error = EINVAL;
1640 goto keeporig;
1641 }
1642
1643 if (dirfd != AT_FDCWD) {
1644 /*
1645 * We want the original because the "prefix" is
1646 * included in the already opened dirfd.
1647 */
1648 bcopy(ptr, buf, len);
1649 return (0);
1650 }
1651
1652 /*
1653 * We know that there is a / somewhere in this pathname.
1654 * Search backwards for it, to find the file's parent dir
1655 * to see if it exists in the alternate tree. If it does,
1656 * and we want to create a file (cflag is set). We don't
1657 * need to worry about the root comparison in this case.
1658 */
1659
1660 if (create) {
1661 for (cp = &ptr[len] - 1; *cp != '/'; cp--);
1662 *cp = '\0';
1663
1664 NDINIT(&nd, LOOKUP, NOFOLLOW, UIO_SYSSPACE, buf, td);
1665 error = namei(&nd);
1666 *cp = '/';
1667 if (error != 0)
1668 goto keeporig;
1669 } else {
1670 NDINIT(&nd, LOOKUP, NOFOLLOW, UIO_SYSSPACE, buf, td);
1671
1672 error = namei(&nd);
1673 if (error != 0)
1674 goto keeporig;
1675
1676 /*
1677 * We now compare the vnode of the prefix to the one
1678 * vnode asked. If they resolve to be the same, then we
1679 * ignore the match so that the real root gets used.
1680 * This avoids the problem of traversing "../.." to find the
1681 * root directory and never finding it, because "/" resolves
1682 * to the emulation root directory. This is expensive :-(
1683 */
1684 NDINIT(&ndroot, LOOKUP, FOLLOW, UIO_SYSSPACE, prefix,
1685 td);
1686
1687 /* We shouldn't ever get an error from this namei(). */
1688 error = namei(&ndroot);
1689 if (error == 0) {
1690 if (nd.ni_vp == ndroot.ni_vp)
1691 error = ENOENT;
1692
1693 NDFREE(&ndroot, NDF_ONLY_PNBUF);
1694 vrele(ndroot.ni_vp);
1695 }
1696 }
1697
1698 NDFREE(&nd, NDF_ONLY_PNBUF);
1699 vrele(nd.ni_vp);
1700
1701 keeporig:
1702 /* If there was an error, use the original path name. */
1703 if (error)
1704 bcopy(ptr, buf, len);
1705 return (error);
1706 }
Cache object: eaa6225a17a541f84b42acc5d47e7e40
|