xref: /freebsd/sys/kern/vfs_lookup.c (revision f61e92ca5a23450bc28169bbdd71d7674df98c19)
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 	if ((cnp->cn_flags & BENEATH) != 0 &&
186 	    (ndp->ni_lcf & NI_LCF_BENEATH_LATCHED) == 0) {
187 		MPASS((ndp->ni_lcf & NI_LCF_LATCH) != 0);
188 		if (dp != ndp->ni_beneath_latch)
189 			return;
190 		ndp->ni_lcf |= NI_LCF_BENEATH_LATCHED;
191 	}
192 	nt = malloc(sizeof(*nt), M_NAMEITRACKER, M_WAITOK);
193 	vhold(dp);
194 	nt->dp = dp;
195 	TAILQ_INSERT_TAIL(&ndp->ni_cap_tracker, nt, nm_link);
196 }
197 
198 static void
199 nameicap_cleanup(struct nameidata *ndp, bool clean_latch)
200 {
201 	struct nameicap_tracker *nt, *nt1;
202 
203 	KASSERT(TAILQ_EMPTY(&ndp->ni_cap_tracker) ||
204 	    (ndp->ni_lcf & NI_LCF_CAP_DOTDOT) != 0, ("not strictrelative"));
205 	TAILQ_FOREACH_SAFE(nt, &ndp->ni_cap_tracker, nm_link, nt1) {
206 		TAILQ_REMOVE(&ndp->ni_cap_tracker, nt, nm_link);
207 		vdrop(nt->dp);
208 		free(nt, M_NAMEITRACKER);
209 	}
210 	if (clean_latch && (ndp->ni_lcf & NI_LCF_LATCH) != 0) {
211 		ndp->ni_lcf &= ~NI_LCF_LATCH;
212 		vrele(ndp->ni_beneath_latch);
213 	}
214 }
215 
216 /*
217  * For dotdot lookups in capability mode, only allow the component
218  * lookup to succeed if the resulting directory was already traversed
219  * during the operation.  This catches situations where already
220  * traversed directory is moved to different parent, and then we walk
221  * over it with dotdots.
222  *
223  * Also allow to force failure of dotdot lookups for non-local
224  * filesystems, where external agents might assist local lookups to
225  * escape the compartment.
226  */
227 static int
228 nameicap_check_dotdot(struct nameidata *ndp, struct vnode *dp)
229 {
230 	struct nameicap_tracker *nt;
231 	struct mount *mp;
232 
233 	if ((ndp->ni_lcf & NI_LCF_CAP_DOTDOT) == 0 || dp == NULL ||
234 	    dp->v_type != VDIR)
235 		return (0);
236 	mp = dp->v_mount;
237 	if (lookup_cap_dotdot_nonlocal == 0 && mp != NULL &&
238 	    (mp->mnt_flag & MNT_LOCAL) == 0)
239 		return (ENOTCAPABLE);
240 	TAILQ_FOREACH_REVERSE(nt, &ndp->ni_cap_tracker, nameicap_tracker_head,
241 	    nm_link) {
242 		if ((ndp->ni_lcf & NI_LCF_LATCH) != 0 &&
243 		    ndp->ni_beneath_latch == nt->dp) {
244 			ndp->ni_lcf &= ~NI_LCF_BENEATH_LATCHED;
245 			nameicap_cleanup(ndp, false);
246 			return (0);
247 		}
248 		if (dp == nt->dp)
249 			return (0);
250 	}
251 	return (ENOTCAPABLE);
252 }
253 
254 static void
255 namei_cleanup_cnp(struct componentname *cnp)
256 {
257 
258 	uma_zfree(namei_zone, cnp->cn_pnbuf);
259 #ifdef DIAGNOSTIC
260 	cnp->cn_pnbuf = NULL;
261 	cnp->cn_nameptr = NULL;
262 #endif
263 }
264 
265 static int
266 namei_handle_root(struct nameidata *ndp, struct vnode **dpp)
267 {
268 	struct componentname *cnp;
269 
270 	cnp = &ndp->ni_cnd;
271 	if ((ndp->ni_lcf & NI_LCF_STRICTRELATIVE) != 0) {
272 #ifdef KTRACE
273 		if (KTRPOINT(curthread, KTR_CAPFAIL))
274 			ktrcapfail(CAPFAIL_LOOKUP, NULL, NULL);
275 #endif
276 		return (ENOTCAPABLE);
277 	}
278 	if ((cnp->cn_flags & BENEATH) != 0) {
279 		ndp->ni_lcf |= NI_LCF_BENEATH_ABS;
280 		ndp->ni_lcf &= ~NI_LCF_BENEATH_LATCHED;
281 		nameicap_cleanup(ndp, false);
282 	}
283 	while (*(cnp->cn_nameptr) == '/') {
284 		cnp->cn_nameptr++;
285 		ndp->ni_pathlen--;
286 	}
287 	*dpp = ndp->ni_rootdir;
288 	vrefact(*dpp);
289 	return (0);
290 }
291 
292 static int
293 namei_setup(struct nameidata *ndp, struct vnode **dpp, struct pwd **pwdp)
294 {
295 	struct componentname *cnp;
296 	struct file *dfp;
297 	struct thread *td;
298 	struct pwd *pwd;
299 	cap_rights_t rights;
300 	struct filecaps dirfd_caps;
301 	int error;
302 	bool startdir_used;
303 
304 	cnp = &ndp->ni_cnd;
305 	td = cnp->cn_thread;
306 
307 	startdir_used = false;
308 	*pwdp = NULL;
309 	*dpp = NULL;
310 
311 #ifdef CAPABILITY_MODE
312 	/*
313 	 * In capability mode, lookups must be restricted to happen in
314 	 * the subtree with the root specified by the file descriptor:
315 	 * - The root must be real file descriptor, not the pseudo-descriptor
316 	 *   AT_FDCWD.
317 	 * - The passed path must be relative and not absolute.
318 	 * - If lookup_cap_dotdot is disabled, path must not contain the
319 	 *   '..' components.
320 	 * - If lookup_cap_dotdot is enabled, we verify that all '..'
321 	 *   components lookups result in the directories which were
322 	 *   previously walked by us, which prevents an escape from
323 	 *   the relative root.
324 	 */
325 	if (IN_CAPABILITY_MODE(td) && (cnp->cn_flags & NOCAPCHECK) == 0) {
326 		ndp->ni_lcf |= NI_LCF_STRICTRELATIVE;
327 		ndp->ni_resflags |= NIRES_STRICTREL;
328 		if (ndp->ni_dirfd == AT_FDCWD) {
329 #ifdef KTRACE
330 			if (KTRPOINT(td, KTR_CAPFAIL))
331 				ktrcapfail(CAPFAIL_LOOKUP, NULL, NULL);
332 #endif
333 			return (ECAPMODE);
334 		}
335 	}
336 #endif
337 	error = 0;
338 
339 	/*
340 	 * Get starting point for the translation.
341 	 */
342 	pwd = pwd_hold(td);
343 	/*
344 	 * The reference on ni_rootdir is acquired in the block below to avoid
345 	 * back-to-back atomics for absolute lookups.
346 	 */
347 	ndp->ni_rootdir = pwd->pwd_rdir;
348 	ndp->ni_topdir = pwd->pwd_jdir;
349 
350 	if (cnp->cn_pnbuf[0] == '/') {
351 		ndp->ni_resflags |= NIRES_ABS;
352 		error = namei_handle_root(ndp, dpp);
353 	} else {
354 		if (ndp->ni_startdir != NULL) {
355 			*dpp = ndp->ni_startdir;
356 			startdir_used = true;
357 		} else if (ndp->ni_dirfd == AT_FDCWD) {
358 			*dpp = pwd->pwd_cdir;
359 			vrefact(*dpp);
360 		} else {
361 			rights = *ndp->ni_rightsneeded;
362 			cap_rights_set_one(&rights, CAP_LOOKUP);
363 
364 			if (cnp->cn_flags & AUDITVNODE1)
365 				AUDIT_ARG_ATFD1(ndp->ni_dirfd);
366 			if (cnp->cn_flags & AUDITVNODE2)
367 				AUDIT_ARG_ATFD2(ndp->ni_dirfd);
368 			/*
369 			 * Effectively inlined fgetvp_rights, because we need to
370 			 * inspect the file as well as grabbing the vnode.
371 			 */
372 			error = fget_cap(td, ndp->ni_dirfd, &rights,
373 			    &dfp, &ndp->ni_filecaps);
374 			if (error != 0) {
375 				/*
376 				 * Preserve the error; it should either be EBADF
377 				 * or capability-related, both of which can be
378 				 * safely returned to the caller.
379 				 */
380 			} else {
381 				if (dfp->f_ops == &badfileops) {
382 					error = EBADF;
383 				} else if (dfp->f_vnode == NULL) {
384 					error = ENOTDIR;
385 				} else {
386 					*dpp = dfp->f_vnode;
387 					vrefact(*dpp);
388 
389 					if ((dfp->f_flag & FSEARCH) != 0)
390 						cnp->cn_flags |= NOEXECCHECK;
391 				}
392 				fdrop(dfp, td);
393 			}
394 #ifdef CAPABILITIES
395 			/*
396 			 * If file descriptor doesn't have all rights,
397 			 * all lookups relative to it must also be
398 			 * strictly relative.
399 			 */
400 			CAP_ALL(&rights);
401 			if (!cap_rights_contains(&ndp->ni_filecaps.fc_rights,
402 			    &rights) ||
403 			    ndp->ni_filecaps.fc_fcntls != CAP_FCNTL_ALL ||
404 			    ndp->ni_filecaps.fc_nioctls != -1) {
405 				ndp->ni_lcf |= NI_LCF_STRICTRELATIVE;
406 				ndp->ni_resflags |= NIRES_STRICTREL;
407 			}
408 #endif
409 		}
410 		if (error == 0 && (*dpp)->v_type != VDIR)
411 			error = ENOTDIR;
412 	}
413 	if (error == 0 && (cnp->cn_flags & BENEATH) != 0) {
414 		if (ndp->ni_dirfd == AT_FDCWD) {
415 			ndp->ni_beneath_latch = pwd->pwd_cdir;
416 			vrefact(ndp->ni_beneath_latch);
417 		} else {
418 			rights = *ndp->ni_rightsneeded;
419 			cap_rights_set_one(&rights, CAP_LOOKUP);
420 			error = fgetvp_rights(td, ndp->ni_dirfd, &rights,
421 			    &dirfd_caps, &ndp->ni_beneath_latch);
422 			if (error == 0 && (*dpp)->v_type != VDIR) {
423 				vrele(ndp->ni_beneath_latch);
424 				error = ENOTDIR;
425 			}
426 		}
427 		if (error == 0)
428 			ndp->ni_lcf |= NI_LCF_LATCH;
429 	}
430 	if (error == 0 && (cnp->cn_flags & RBENEATH) != 0) {
431 		if (cnp->cn_pnbuf[0] == '/' ||
432 		    (ndp->ni_lcf & NI_LCF_BENEATH_ABS) != 0) {
433 			error = EINVAL;
434 		} else if ((ndp->ni_lcf & NI_LCF_STRICTRELATIVE) == 0) {
435 			ndp->ni_lcf |= NI_LCF_STRICTRELATIVE |
436 			    NI_LCF_CAP_DOTDOT;
437 		}
438 	}
439 
440 	/*
441 	 * If we are auditing the kernel pathname, save the user pathname.
442 	 */
443 	if (cnp->cn_flags & AUDITVNODE1)
444 		AUDIT_ARG_UPATH1_VP(td, ndp->ni_rootdir, *dpp, cnp->cn_pnbuf);
445 	if (cnp->cn_flags & AUDITVNODE2)
446 		AUDIT_ARG_UPATH2_VP(td, ndp->ni_rootdir, *dpp, cnp->cn_pnbuf);
447 	if (ndp->ni_startdir != NULL && !startdir_used)
448 		vrele(ndp->ni_startdir);
449 	if (error != 0) {
450 		if (*dpp != NULL)
451 			vrele(*dpp);
452 		pwd_drop(pwd);
453 		return (error);
454 	}
455 	MPASS((ndp->ni_lcf & (NI_LCF_BENEATH_ABS | NI_LCF_LATCH)) !=
456 	    NI_LCF_BENEATH_ABS);
457 	if (((ndp->ni_lcf & NI_LCF_STRICTRELATIVE) != 0 &&
458 	    lookup_cap_dotdot != 0) ||
459 	    ((ndp->ni_lcf & NI_LCF_STRICTRELATIVE) == 0 &&
460 	    (cnp->cn_flags & BENEATH) != 0))
461 		ndp->ni_lcf |= NI_LCF_CAP_DOTDOT;
462 	SDT_PROBE4(vfs, namei, lookup, entry, *dpp, cnp->cn_pnbuf,
463 	    cnp->cn_flags, false);
464 	*pwdp = pwd;
465 	return (0);
466 }
467 
468 static int
469 namei_getpath(struct nameidata *ndp)
470 {
471 	struct componentname *cnp;
472 	int error;
473 
474 	cnp = &ndp->ni_cnd;
475 
476 	/*
477 	 * Get a buffer for the name to be translated, and copy the
478 	 * name into the buffer.
479 	 */
480 	cnp->cn_pnbuf = uma_zalloc(namei_zone, M_WAITOK);
481 	if (ndp->ni_segflg == UIO_SYSSPACE) {
482 		error = copystr(ndp->ni_dirp, cnp->cn_pnbuf, MAXPATHLEN,
483 		    &ndp->ni_pathlen);
484 	} else {
485 		error = copyinstr(ndp->ni_dirp, cnp->cn_pnbuf, MAXPATHLEN,
486 		    &ndp->ni_pathlen);
487 	}
488 
489 	if (__predict_false(error != 0)) {
490 		namei_cleanup_cnp(cnp);
491 		return (error);
492 	}
493 
494 	/*
495 	 * Don't allow empty pathnames.
496 	 */
497 	if (__predict_false(*cnp->cn_pnbuf == '\0')) {
498 		namei_cleanup_cnp(cnp);
499 		return (ENOENT);
500 	}
501 
502 	cnp->cn_nameptr = cnp->cn_pnbuf;
503 	return (0);
504 }
505 
506 /*
507  * Convert a pathname into a pointer to a locked vnode.
508  *
509  * The FOLLOW flag is set when symbolic links are to be followed
510  * when they occur at the end of the name translation process.
511  * Symbolic links are always followed for all other pathname
512  * components other than the last.
513  *
514  * The segflg defines whether the name is to be copied from user
515  * space or kernel space.
516  *
517  * Overall outline of namei:
518  *
519  *	copy in name
520  *	get starting directory
521  *	while (!done && !error) {
522  *		call lookup to search path.
523  *		if symbolic link, massage name in buffer and continue
524  *	}
525  */
526 int
527 namei(struct nameidata *ndp)
528 {
529 	char *cp;		/* pointer into pathname argument */
530 	struct vnode *dp;	/* the directory we are searching */
531 	struct iovec aiov;		/* uio for reading symbolic links */
532 	struct componentname *cnp;
533 	struct thread *td;
534 	struct pwd *pwd;
535 	struct uio auio;
536 	int error, linklen;
537 	enum cache_fpl_status status;
538 
539 	cnp = &ndp->ni_cnd;
540 	td = cnp->cn_thread;
541 #ifdef INVARIANTS
542 	KASSERT((ndp->ni_debugflags & NAMEI_DBG_CALLED) == 0,
543 	    ("%s: repeated call to namei without NDREINIT", __func__));
544 	KASSERT(ndp->ni_debugflags == NAMEI_DBG_INITED,
545 	    ("%s: bad debugflags %d", __func__, ndp->ni_debugflags));
546 	ndp->ni_debugflags |= NAMEI_DBG_CALLED;
547 	if (ndp->ni_startdir != NULL)
548 		ndp->ni_debugflags |= NAMEI_DBG_HADSTARTDIR;
549 	if (cnp->cn_flags & FAILIFEXISTS) {
550 		KASSERT(cnp->cn_nameiop == CREATE,
551 		    ("%s: FAILIFEXISTS passed for op %d", __func__, cnp->cn_nameiop));
552 		/*
553 		 * The limitation below is to restrict hairy corner cases.
554 		 */
555 		KASSERT((cnp->cn_flags & (LOCKPARENT | LOCKLEAF)) == LOCKPARENT,
556 		    ("%s: FAILIFEXISTS must be passed with LOCKPARENT and without LOCKLEAF",
557 		    __func__));
558 	}
559 	/*
560 	 * For NDVALIDATE.
561 	 *
562 	 * While NDINIT may seem like a more natural place to do it, there are
563 	 * callers which directly modify flags past invoking init.
564 	 */
565 	cnp->cn_origflags = cnp->cn_flags;
566 #endif
567 	ndp->ni_cnd.cn_cred = ndp->ni_cnd.cn_thread->td_ucred;
568 	KASSERT(ndp->ni_resflags == 0, ("%s: garbage in ni_resflags: %x\n",
569 	    __func__, ndp->ni_resflags));
570 	KASSERT(cnp->cn_cred && td->td_proc, ("namei: bad cred/proc"));
571 	KASSERT((cnp->cn_flags & NAMEI_INTERNAL_FLAGS) == 0,
572 	    ("namei: unexpected flags: %" PRIx64 "\n",
573 	    cnp->cn_flags & NAMEI_INTERNAL_FLAGS));
574 	if (cnp->cn_flags & NOCACHE)
575 		KASSERT(cnp->cn_nameiop != LOOKUP,
576 		    ("%s: NOCACHE passed with LOOKUP", __func__));
577 	MPASS(ndp->ni_startdir == NULL || ndp->ni_startdir->v_type == VDIR ||
578 	    ndp->ni_startdir->v_type == VBAD);
579 
580 	ndp->ni_lcf = 0;
581 	ndp->ni_loopcnt = 0;
582 	ndp->ni_vp = NULL;
583 
584 	error = namei_getpath(ndp);
585 	if (__predict_false(error != 0)) {
586 		return (error);
587 	}
588 
589 #ifdef KTRACE
590 	if (KTRPOINT(td, KTR_NAMEI)) {
591 		KASSERT(cnp->cn_thread == curthread,
592 		    ("namei not using curthread"));
593 		ktrnamei(cnp->cn_pnbuf);
594 	}
595 #endif
596 
597 	/*
598 	 * First try looking up the target without locking any vnodes.
599 	 *
600 	 * We may need to start from scratch or pick up where it left off.
601 	 */
602 	error = cache_fplookup(ndp, &status, &pwd);
603 	switch (status) {
604 	case CACHE_FPL_STATUS_UNSET:
605 		__assert_unreachable();
606 		break;
607 	case CACHE_FPL_STATUS_HANDLED:
608 		if (error == 0)
609 			NDVALIDATE(ndp);
610 		return (error);
611 	case CACHE_FPL_STATUS_PARTIAL:
612 		TAILQ_INIT(&ndp->ni_cap_tracker);
613 		dp = ndp->ni_startdir;
614 		break;
615 	case CACHE_FPL_STATUS_DESTROYED:
616 		ndp->ni_loopcnt = 0;
617 		error = namei_getpath(ndp);
618 		if (__predict_false(error != 0)) {
619 			return (error);
620 		}
621 		/* FALLTHROUGH */
622 	case CACHE_FPL_STATUS_ABORTED:
623 		TAILQ_INIT(&ndp->ni_cap_tracker);
624 		MPASS(ndp->ni_lcf == 0);
625 		error = namei_setup(ndp, &dp, &pwd);
626 		if (error != 0) {
627 			namei_cleanup_cnp(cnp);
628 			return (error);
629 		}
630 		break;
631 	}
632 
633 	/*
634 	 * Locked lookup.
635 	 */
636 	for (;;) {
637 		ndp->ni_startdir = dp;
638 		error = lookup(ndp);
639 		if (error != 0) {
640 			/*
641 			 * Override an error to not allow user to use
642 			 * BENEATH as an oracle.
643 			 */
644 			if ((ndp->ni_lcf & (NI_LCF_LATCH |
645 			    NI_LCF_BENEATH_LATCHED)) == NI_LCF_LATCH)
646 				error = ENOTCAPABLE;
647 			goto out;
648 		}
649 
650 		/*
651 		 * If not a symbolic link, we're done.
652 		 */
653 		if ((cnp->cn_flags & ISSYMLINK) == 0) {
654 			SDT_PROBE4(vfs, namei, lookup, return, error,
655 			    (error == 0 ? ndp->ni_vp : NULL), false, ndp);
656 			if ((cnp->cn_flags & (SAVENAME | SAVESTART)) == 0) {
657 				namei_cleanup_cnp(cnp);
658 			} else
659 				cnp->cn_flags |= HASBUF;
660 			if ((ndp->ni_lcf & (NI_LCF_LATCH |
661 			    NI_LCF_BENEATH_LATCHED)) == NI_LCF_LATCH) {
662 				NDFREE(ndp, 0);
663 				error = ENOTCAPABLE;
664 			}
665 			nameicap_cleanup(ndp, true);
666 			pwd_drop(pwd);
667 			if (error == 0)
668 				NDVALIDATE(ndp);
669 			return (error);
670 		}
671 		if (ndp->ni_loopcnt++ >= MAXSYMLINKS) {
672 			error = ELOOP;
673 			break;
674 		}
675 #ifdef MAC
676 		if ((cnp->cn_flags & NOMACCHECK) == 0) {
677 			error = mac_vnode_check_readlink(td->td_ucred,
678 			    ndp->ni_vp);
679 			if (error != 0)
680 				break;
681 		}
682 #endif
683 		if (ndp->ni_pathlen > 1)
684 			cp = uma_zalloc(namei_zone, M_WAITOK);
685 		else
686 			cp = cnp->cn_pnbuf;
687 		aiov.iov_base = cp;
688 		aiov.iov_len = MAXPATHLEN;
689 		auio.uio_iov = &aiov;
690 		auio.uio_iovcnt = 1;
691 		auio.uio_offset = 0;
692 		auio.uio_rw = UIO_READ;
693 		auio.uio_segflg = UIO_SYSSPACE;
694 		auio.uio_td = td;
695 		auio.uio_resid = MAXPATHLEN;
696 		error = VOP_READLINK(ndp->ni_vp, &auio, cnp->cn_cred);
697 		if (error != 0) {
698 			if (ndp->ni_pathlen > 1)
699 				uma_zfree(namei_zone, cp);
700 			break;
701 		}
702 		linklen = MAXPATHLEN - auio.uio_resid;
703 		if (linklen == 0) {
704 			if (ndp->ni_pathlen > 1)
705 				uma_zfree(namei_zone, cp);
706 			error = ENOENT;
707 			break;
708 		}
709 		if (linklen + ndp->ni_pathlen > MAXPATHLEN) {
710 			if (ndp->ni_pathlen > 1)
711 				uma_zfree(namei_zone, cp);
712 			error = ENAMETOOLONG;
713 			break;
714 		}
715 		if (ndp->ni_pathlen > 1) {
716 			bcopy(ndp->ni_next, cp + linklen, ndp->ni_pathlen);
717 			uma_zfree(namei_zone, cnp->cn_pnbuf);
718 			cnp->cn_pnbuf = cp;
719 		} else
720 			cnp->cn_pnbuf[linklen] = '\0';
721 		ndp->ni_pathlen += linklen;
722 		vput(ndp->ni_vp);
723 		dp = ndp->ni_dvp;
724 		/*
725 		 * Check if root directory should replace current directory.
726 		 */
727 		cnp->cn_nameptr = cnp->cn_pnbuf;
728 		if (*(cnp->cn_nameptr) == '/') {
729 			vrele(dp);
730 			error = namei_handle_root(ndp, &dp);
731 			if (error != 0)
732 				goto out;
733 		}
734 	}
735 	vput(ndp->ni_vp);
736 	ndp->ni_vp = NULL;
737 	vrele(ndp->ni_dvp);
738 out:
739 	MPASS(error != 0);
740 	SDT_PROBE4(vfs, namei, lookup, return, error, NULL, false, ndp);
741 	namei_cleanup_cnp(cnp);
742 	nameicap_cleanup(ndp, true);
743 	pwd_drop(pwd);
744 	return (error);
745 }
746 
747 static int
748 compute_cn_lkflags(struct mount *mp, int lkflags, int cnflags)
749 {
750 
751 	if (mp == NULL || ((lkflags & LK_SHARED) &&
752 	    (!(mp->mnt_kern_flag & MNTK_LOOKUP_SHARED) ||
753 	    ((cnflags & ISDOTDOT) &&
754 	    (mp->mnt_kern_flag & MNTK_LOOKUP_EXCL_DOTDOT))))) {
755 		lkflags &= ~LK_SHARED;
756 		lkflags |= LK_EXCLUSIVE;
757 	}
758 	lkflags |= LK_NODDLKTREAT;
759 	return (lkflags);
760 }
761 
762 static __inline int
763 needs_exclusive_leaf(struct mount *mp, int flags)
764 {
765 
766 	/*
767 	 * Intermediate nodes can use shared locks, we only need to
768 	 * force an exclusive lock for leaf nodes.
769 	 */
770 	if ((flags & (ISLASTCN | LOCKLEAF)) != (ISLASTCN | LOCKLEAF))
771 		return (0);
772 
773 	/* Always use exclusive locks if LOCKSHARED isn't set. */
774 	if (!(flags & LOCKSHARED))
775 		return (1);
776 
777 	/*
778 	 * For lookups during open(), if the mount point supports
779 	 * extended shared operations, then use a shared lock for the
780 	 * leaf node, otherwise use an exclusive lock.
781 	 */
782 	if ((flags & ISOPEN) != 0)
783 		return (!MNT_EXTENDED_SHARED(mp));
784 
785 	/*
786 	 * Lookup requests outside of open() that specify LOCKSHARED
787 	 * only need a shared lock on the leaf vnode.
788 	 */
789 	return (0);
790 }
791 
792 /*
793  * Search a pathname.
794  * This is a very central and rather complicated routine.
795  *
796  * The pathname is pointed to by ni_ptr and is of length ni_pathlen.
797  * The starting directory is taken from ni_startdir. The pathname is
798  * descended until done, or a symbolic link is encountered. The variable
799  * ni_more is clear if the path is completed; it is set to one if a
800  * symbolic link needing interpretation is encountered.
801  *
802  * The flag argument is LOOKUP, CREATE, RENAME, or DELETE depending on
803  * whether the name is to be looked up, created, renamed, or deleted.
804  * When CREATE, RENAME, or DELETE is specified, information usable in
805  * creating, renaming, or deleting a directory entry may be calculated.
806  * If flag has LOCKPARENT or'ed into it, the parent directory is returned
807  * locked. If flag has WANTPARENT or'ed into it, the parent directory is
808  * returned unlocked. Otherwise the parent directory is not returned. If
809  * the target of the pathname exists and LOCKLEAF is or'ed into the flag
810  * the target is returned locked, otherwise it is returned unlocked.
811  * When creating or renaming and LOCKPARENT is specified, the target may not
812  * be ".".  When deleting and LOCKPARENT is specified, the target may be ".".
813  *
814  * Overall outline of lookup:
815  *
816  * dirloop:
817  *	identify next component of name at ndp->ni_ptr
818  *	handle degenerate case where name is null string
819  *	if .. and crossing mount points and on mounted filesys, find parent
820  *	call VOP_LOOKUP routine for next component name
821  *	    directory vnode returned in ni_dvp, unlocked unless LOCKPARENT set
822  *	    component vnode returned in ni_vp (if it exists), locked.
823  *	if result vnode is mounted on and crossing mount points,
824  *	    find mounted on vnode
825  *	if more components of name, do next level at dirloop
826  *	return the answer in ni_vp, locked if LOCKLEAF set
827  *	    if LOCKPARENT set, return locked parent in ni_dvp
828  *	    if WANTPARENT set, return unlocked parent in ni_dvp
829  */
830 int
831 lookup(struct nameidata *ndp)
832 {
833 	char *cp;			/* pointer into pathname argument */
834 	char *prev_ni_next;		/* saved ndp->ni_next */
835 	struct vnode *dp = NULL;	/* the directory we are searching */
836 	struct vnode *tdp;		/* saved dp */
837 	struct mount *mp;		/* mount table entry */
838 	struct prison *pr;
839 	size_t prev_ni_pathlen;		/* saved ndp->ni_pathlen */
840 	int docache;			/* == 0 do not cache last component */
841 	int wantparent;			/* 1 => wantparent or lockparent flag */
842 	int rdonly;			/* lookup read-only flag bit */
843 	int error = 0;
844 	int dpunlocked = 0;		/* dp has already been unlocked */
845 	int relookup = 0;		/* do not consume the path component */
846 	struct componentname *cnp = &ndp->ni_cnd;
847 	int lkflags_save;
848 	int ni_dvp_unlocked;
849 
850 	/*
851 	 * Setup: break out flag bits into variables.
852 	 */
853 	ni_dvp_unlocked = 0;
854 	wantparent = cnp->cn_flags & (LOCKPARENT | WANTPARENT);
855 	KASSERT(cnp->cn_nameiop == LOOKUP || wantparent,
856 	    ("CREATE, DELETE, RENAME require LOCKPARENT or WANTPARENT."));
857 	/*
858 	 * When set to zero, docache causes the last component of the
859 	 * pathname to be deleted from the cache and the full lookup
860 	 * of the name to be done (via VOP_CACHEDLOOKUP()). Often
861 	 * filesystems need some pre-computed values that are made
862 	 * during the full lookup, for instance UFS sets dp->i_offset.
863 	 *
864 	 * The docache variable is set to zero when requested by the
865 	 * NOCACHE flag and for all modifying operations except CREATE.
866 	 */
867 	docache = (cnp->cn_flags & NOCACHE) ^ NOCACHE;
868 	if (cnp->cn_nameiop == DELETE ||
869 	    (wantparent && cnp->cn_nameiop != CREATE &&
870 	     cnp->cn_nameiop != LOOKUP))
871 		docache = 0;
872 	rdonly = cnp->cn_flags & RDONLY;
873 	cnp->cn_flags &= ~ISSYMLINK;
874 	ndp->ni_dvp = NULL;
875 	/*
876 	 * We use shared locks until we hit the parent of the last cn then
877 	 * we adjust based on the requesting flags.
878 	 */
879 	cnp->cn_lkflags = LK_SHARED;
880 	dp = ndp->ni_startdir;
881 	ndp->ni_startdir = NULLVP;
882 	vn_lock(dp,
883 	    compute_cn_lkflags(dp->v_mount, cnp->cn_lkflags | LK_RETRY,
884 	    cnp->cn_flags));
885 
886 dirloop:
887 	/*
888 	 * Search a new directory.
889 	 *
890 	 * The last component of the filename is left accessible via
891 	 * cnp->cn_nameptr for callers that need the name. Callers needing
892 	 * the name set the SAVENAME flag. When done, they assume
893 	 * responsibility for freeing the pathname buffer.
894 	 */
895 	for (cp = cnp->cn_nameptr; *cp != 0 && *cp != '/'; cp++)
896 		continue;
897 	cnp->cn_namelen = cp - cnp->cn_nameptr;
898 	if (cnp->cn_namelen > NAME_MAX) {
899 		error = ENAMETOOLONG;
900 		goto bad;
901 	}
902 #ifdef NAMEI_DIAGNOSTIC
903 	{ char c = *cp;
904 	*cp = '\0';
905 	printf("{%s}: ", cnp->cn_nameptr);
906 	*cp = c; }
907 #endif
908 	prev_ni_pathlen = ndp->ni_pathlen;
909 	ndp->ni_pathlen -= cnp->cn_namelen;
910 	KASSERT(ndp->ni_pathlen <= PATH_MAX,
911 	    ("%s: ni_pathlen underflow to %zd\n", __func__, ndp->ni_pathlen));
912 	prev_ni_next = ndp->ni_next;
913 	ndp->ni_next = cp;
914 
915 	/*
916 	 * Replace multiple slashes by a single slash and trailing slashes
917 	 * by a null.  This must be done before VOP_LOOKUP() because some
918 	 * fs's don't know about trailing slashes.  Remember if there were
919 	 * trailing slashes to handle symlinks, existing non-directories
920 	 * and non-existing files that won't be directories specially later.
921 	 */
922 	while (*cp == '/' && (cp[1] == '/' || cp[1] == '\0')) {
923 		cp++;
924 		ndp->ni_pathlen--;
925 		if (*cp == '\0') {
926 			*ndp->ni_next = '\0';
927 			cnp->cn_flags |= TRAILINGSLASH;
928 		}
929 	}
930 	ndp->ni_next = cp;
931 
932 	cnp->cn_flags |= MAKEENTRY;
933 	if (*cp == '\0' && docache == 0)
934 		cnp->cn_flags &= ~MAKEENTRY;
935 	if (cnp->cn_namelen == 2 &&
936 	    cnp->cn_nameptr[1] == '.' && cnp->cn_nameptr[0] == '.')
937 		cnp->cn_flags |= ISDOTDOT;
938 	else
939 		cnp->cn_flags &= ~ISDOTDOT;
940 	if (*ndp->ni_next == 0)
941 		cnp->cn_flags |= ISLASTCN;
942 	else
943 		cnp->cn_flags &= ~ISLASTCN;
944 
945 	if ((cnp->cn_flags & ISLASTCN) != 0 &&
946 	    cnp->cn_namelen == 1 && cnp->cn_nameptr[0] == '.' &&
947 	    (cnp->cn_nameiop == DELETE || cnp->cn_nameiop == RENAME)) {
948 		error = EINVAL;
949 		goto bad;
950 	}
951 
952 	nameicap_tracker_add(ndp, dp);
953 
954 	/*
955 	 * Check for degenerate name (e.g. / or "")
956 	 * which is a way of talking about a directory,
957 	 * e.g. like "/." or ".".
958 	 */
959 	if (cnp->cn_nameptr[0] == '\0') {
960 		if (dp->v_type != VDIR) {
961 			error = ENOTDIR;
962 			goto bad;
963 		}
964 		if (cnp->cn_nameiop != LOOKUP) {
965 			error = EISDIR;
966 			goto bad;
967 		}
968 		if (wantparent) {
969 			ndp->ni_dvp = dp;
970 			VREF(dp);
971 		}
972 		ndp->ni_vp = dp;
973 
974 		if (cnp->cn_flags & AUDITVNODE1)
975 			AUDIT_ARG_VNODE1(dp);
976 		else if (cnp->cn_flags & AUDITVNODE2)
977 			AUDIT_ARG_VNODE2(dp);
978 
979 		if (!(cnp->cn_flags & (LOCKPARENT | LOCKLEAF)))
980 			VOP_UNLOCK(dp);
981 		/* XXX This should probably move to the top of function. */
982 		if (cnp->cn_flags & SAVESTART)
983 			panic("lookup: SAVESTART");
984 		goto success;
985 	}
986 
987 	/*
988 	 * Handle "..": five special cases.
989 	 * 0. If doing a capability lookup and lookup_cap_dotdot is
990 	 *    disabled, return ENOTCAPABLE.
991 	 * 1. Return an error if this is the last component of
992 	 *    the name and the operation is DELETE or RENAME.
993 	 * 2. If at root directory (e.g. after chroot)
994 	 *    or at absolute root directory
995 	 *    then ignore it so can't get out.
996 	 * 3. If this vnode is the root of a mounted
997 	 *    filesystem, then replace it with the
998 	 *    vnode which was mounted on so we take the
999 	 *    .. in the other filesystem.
1000 	 * 4. If the vnode is the top directory of
1001 	 *    the jail or chroot, don't let them out.
1002 	 * 5. If doing a capability lookup and lookup_cap_dotdot is
1003 	 *    enabled, return ENOTCAPABLE if the lookup would escape
1004 	 *    from the initial file descriptor directory.  Checks are
1005 	 *    done by ensuring that namei() already traversed the
1006 	 *    result of dotdot lookup.
1007 	 */
1008 	if (cnp->cn_flags & ISDOTDOT) {
1009 		if ((ndp->ni_lcf & (NI_LCF_STRICTRELATIVE | NI_LCF_CAP_DOTDOT))
1010 		    == NI_LCF_STRICTRELATIVE) {
1011 #ifdef KTRACE
1012 			if (KTRPOINT(curthread, KTR_CAPFAIL))
1013 				ktrcapfail(CAPFAIL_LOOKUP, NULL, NULL);
1014 #endif
1015 			error = ENOTCAPABLE;
1016 			goto bad;
1017 		}
1018 		if ((cnp->cn_flags & ISLASTCN) != 0 &&
1019 		    (cnp->cn_nameiop == DELETE || cnp->cn_nameiop == RENAME)) {
1020 			error = EINVAL;
1021 			goto bad;
1022 		}
1023 		for (;;) {
1024 			for (pr = cnp->cn_cred->cr_prison; pr != NULL;
1025 			     pr = pr->pr_parent)
1026 				if (dp == pr->pr_root)
1027 					break;
1028 			if (dp == ndp->ni_rootdir ||
1029 			    dp == ndp->ni_topdir ||
1030 			    dp == rootvnode ||
1031 			    pr != NULL ||
1032 			    ((dp->v_vflag & VV_ROOT) != 0 &&
1033 			     (cnp->cn_flags & NOCROSSMOUNT) != 0)) {
1034 				ndp->ni_dvp = dp;
1035 				ndp->ni_vp = dp;
1036 				VREF(dp);
1037 				goto nextname;
1038 			}
1039 			if ((dp->v_vflag & VV_ROOT) == 0)
1040 				break;
1041 			if (VN_IS_DOOMED(dp)) {	/* forced unmount */
1042 				error = ENOENT;
1043 				goto bad;
1044 			}
1045 			tdp = dp;
1046 			dp = dp->v_mount->mnt_vnodecovered;
1047 			VREF(dp);
1048 			vput(tdp);
1049 			vn_lock(dp,
1050 			    compute_cn_lkflags(dp->v_mount, cnp->cn_lkflags |
1051 			    LK_RETRY, ISDOTDOT));
1052 			error = nameicap_check_dotdot(ndp, dp);
1053 			if (error != 0) {
1054 #ifdef KTRACE
1055 				if (KTRPOINT(curthread, KTR_CAPFAIL))
1056 					ktrcapfail(CAPFAIL_LOOKUP, NULL, NULL);
1057 #endif
1058 				goto bad;
1059 			}
1060 		}
1061 	}
1062 
1063 	/*
1064 	 * We now have a segment name to search for, and a directory to search.
1065 	 */
1066 unionlookup:
1067 #ifdef MAC
1068 	error = mac_vnode_check_lookup(cnp->cn_thread->td_ucred, dp, cnp);
1069 	if (error)
1070 		goto bad;
1071 #endif
1072 	ndp->ni_dvp = dp;
1073 	ndp->ni_vp = NULL;
1074 	ASSERT_VOP_LOCKED(dp, "lookup");
1075 	/*
1076 	 * If we have a shared lock we may need to upgrade the lock for the
1077 	 * last operation.
1078 	 */
1079 	if ((cnp->cn_flags & LOCKPARENT) && (cnp->cn_flags & ISLASTCN) &&
1080 	    dp != vp_crossmp && VOP_ISLOCKED(dp) == LK_SHARED)
1081 		vn_lock(dp, LK_UPGRADE|LK_RETRY);
1082 	if (VN_IS_DOOMED(dp)) {
1083 		error = ENOENT;
1084 		goto bad;
1085 	}
1086 	/*
1087 	 * If we're looking up the last component and we need an exclusive
1088 	 * lock, adjust our lkflags.
1089 	 */
1090 	if (needs_exclusive_leaf(dp->v_mount, cnp->cn_flags))
1091 		cnp->cn_lkflags = LK_EXCLUSIVE;
1092 #ifdef NAMEI_DIAGNOSTIC
1093 	vn_printf(dp, "lookup in ");
1094 #endif
1095 	lkflags_save = cnp->cn_lkflags;
1096 	cnp->cn_lkflags = compute_cn_lkflags(dp->v_mount, cnp->cn_lkflags,
1097 	    cnp->cn_flags);
1098 	error = VOP_LOOKUP(dp, &ndp->ni_vp, cnp);
1099 	cnp->cn_lkflags = lkflags_save;
1100 	if (error != 0) {
1101 		KASSERT(ndp->ni_vp == NULL, ("leaf should be empty"));
1102 #ifdef NAMEI_DIAGNOSTIC
1103 		printf("not found\n");
1104 #endif
1105 		if ((error == ENOENT) &&
1106 		    (dp->v_vflag & VV_ROOT) && (dp->v_mount != NULL) &&
1107 		    (dp->v_mount->mnt_flag & MNT_UNION)) {
1108 			tdp = dp;
1109 			dp = dp->v_mount->mnt_vnodecovered;
1110 			VREF(dp);
1111 			vput(tdp);
1112 			vn_lock(dp,
1113 			    compute_cn_lkflags(dp->v_mount, cnp->cn_lkflags |
1114 			    LK_RETRY, cnp->cn_flags));
1115 			nameicap_tracker_add(ndp, dp);
1116 			goto unionlookup;
1117 		}
1118 
1119 		if (error == ERELOOKUP) {
1120 			vref(dp);
1121 			ndp->ni_vp = dp;
1122 			error = 0;
1123 			relookup = 1;
1124 			goto good;
1125 		}
1126 
1127 		if (error != EJUSTRETURN)
1128 			goto bad;
1129 		/*
1130 		 * At this point, we know we're at the end of the
1131 		 * pathname.  If creating / renaming, we can consider
1132 		 * allowing the file or directory to be created / renamed,
1133 		 * provided we're not on a read-only filesystem.
1134 		 */
1135 		if (rdonly) {
1136 			error = EROFS;
1137 			goto bad;
1138 		}
1139 		/* trailing slash only allowed for directories */
1140 		if ((cnp->cn_flags & TRAILINGSLASH) &&
1141 		    !(cnp->cn_flags & WILLBEDIR)) {
1142 			error = ENOENT;
1143 			goto bad;
1144 		}
1145 		if ((cnp->cn_flags & LOCKPARENT) == 0)
1146 			VOP_UNLOCK(dp);
1147 		/*
1148 		 * We return with ni_vp NULL to indicate that the entry
1149 		 * doesn't currently exist, leaving a pointer to the
1150 		 * (possibly locked) directory vnode in ndp->ni_dvp.
1151 		 */
1152 		if (cnp->cn_flags & SAVESTART) {
1153 			ndp->ni_startdir = ndp->ni_dvp;
1154 			VREF(ndp->ni_startdir);
1155 		}
1156 		goto success;
1157 	}
1158 
1159 good:
1160 #ifdef NAMEI_DIAGNOSTIC
1161 	printf("found\n");
1162 #endif
1163 	dp = ndp->ni_vp;
1164 
1165 	/*
1166 	 * Check to see if the vnode has been mounted on;
1167 	 * if so find the root of the mounted filesystem.
1168 	 */
1169 	while (dp->v_type == VDIR && (mp = dp->v_mountedhere) &&
1170 	       (cnp->cn_flags & NOCROSSMOUNT) == 0) {
1171 		if (vfs_busy(mp, 0))
1172 			continue;
1173 		vput(dp);
1174 		if (dp != ndp->ni_dvp)
1175 			vput(ndp->ni_dvp);
1176 		else
1177 			vrele(ndp->ni_dvp);
1178 		vrefact(vp_crossmp);
1179 		ndp->ni_dvp = vp_crossmp;
1180 		error = VFS_ROOT(mp, compute_cn_lkflags(mp, cnp->cn_lkflags,
1181 		    cnp->cn_flags), &tdp);
1182 		vfs_unbusy(mp);
1183 		if (vn_lock(vp_crossmp, LK_SHARED | LK_NOWAIT))
1184 			panic("vp_crossmp exclusively locked or reclaimed");
1185 		if (error) {
1186 			dpunlocked = 1;
1187 			goto bad2;
1188 		}
1189 		ndp->ni_vp = dp = tdp;
1190 	}
1191 
1192 	/*
1193 	 * Check for symbolic link
1194 	 */
1195 	if ((dp->v_type == VLNK) &&
1196 	    ((cnp->cn_flags & FOLLOW) || (cnp->cn_flags & TRAILINGSLASH) ||
1197 	     *ndp->ni_next == '/')) {
1198 		cnp->cn_flags |= ISSYMLINK;
1199 		if (VN_IS_DOOMED(dp)) {
1200 			/*
1201 			 * We can't know whether the directory was mounted with
1202 			 * NOSYMFOLLOW, so we can't follow safely.
1203 			 */
1204 			error = ENOENT;
1205 			goto bad2;
1206 		}
1207 		if (dp->v_mount->mnt_flag & MNT_NOSYMFOLLOW) {
1208 			error = EACCES;
1209 			goto bad2;
1210 		}
1211 		/*
1212 		 * Symlink code always expects an unlocked dvp.
1213 		 */
1214 		if (ndp->ni_dvp != ndp->ni_vp) {
1215 			VOP_UNLOCK(ndp->ni_dvp);
1216 			ni_dvp_unlocked = 1;
1217 		}
1218 		goto success;
1219 	}
1220 
1221 nextname:
1222 	/*
1223 	 * Not a symbolic link that we will follow.  Continue with the
1224 	 * next component if there is any; otherwise, we're done.
1225 	 */
1226 	KASSERT((cnp->cn_flags & ISLASTCN) || *ndp->ni_next == '/',
1227 	    ("lookup: invalid path state."));
1228 	if (relookup) {
1229 		relookup = 0;
1230 		ndp->ni_pathlen = prev_ni_pathlen;
1231 		ndp->ni_next = prev_ni_next;
1232 		if (ndp->ni_dvp != dp)
1233 			vput(ndp->ni_dvp);
1234 		else
1235 			vrele(ndp->ni_dvp);
1236 		goto dirloop;
1237 	}
1238 	if (cnp->cn_flags & ISDOTDOT) {
1239 		error = nameicap_check_dotdot(ndp, ndp->ni_vp);
1240 		if (error != 0) {
1241 #ifdef KTRACE
1242 			if (KTRPOINT(curthread, KTR_CAPFAIL))
1243 				ktrcapfail(CAPFAIL_LOOKUP, NULL, NULL);
1244 #endif
1245 			goto bad2;
1246 		}
1247 	}
1248 	if (*ndp->ni_next == '/') {
1249 		cnp->cn_nameptr = ndp->ni_next;
1250 		while (*cnp->cn_nameptr == '/') {
1251 			cnp->cn_nameptr++;
1252 			ndp->ni_pathlen--;
1253 		}
1254 		if (ndp->ni_dvp != dp)
1255 			vput(ndp->ni_dvp);
1256 		else
1257 			vrele(ndp->ni_dvp);
1258 		goto dirloop;
1259 	}
1260 	/*
1261 	 * If we're processing a path with a trailing slash,
1262 	 * check that the end result is a directory.
1263 	 */
1264 	if ((cnp->cn_flags & TRAILINGSLASH) && dp->v_type != VDIR) {
1265 		error = ENOTDIR;
1266 		goto bad2;
1267 	}
1268 	/*
1269 	 * Disallow directory write attempts on read-only filesystems.
1270 	 */
1271 	if (rdonly &&
1272 	    (cnp->cn_nameiop == DELETE || cnp->cn_nameiop == RENAME)) {
1273 		error = EROFS;
1274 		goto bad2;
1275 	}
1276 	if (cnp->cn_flags & SAVESTART) {
1277 		ndp->ni_startdir = ndp->ni_dvp;
1278 		VREF(ndp->ni_startdir);
1279 	}
1280 	if (!wantparent) {
1281 		ni_dvp_unlocked = 2;
1282 		if (ndp->ni_dvp != dp)
1283 			vput(ndp->ni_dvp);
1284 		else
1285 			vrele(ndp->ni_dvp);
1286 	} else if ((cnp->cn_flags & LOCKPARENT) == 0 && ndp->ni_dvp != dp) {
1287 		VOP_UNLOCK(ndp->ni_dvp);
1288 		ni_dvp_unlocked = 1;
1289 	}
1290 
1291 	if (cnp->cn_flags & AUDITVNODE1)
1292 		AUDIT_ARG_VNODE1(dp);
1293 	else if (cnp->cn_flags & AUDITVNODE2)
1294 		AUDIT_ARG_VNODE2(dp);
1295 
1296 	if ((cnp->cn_flags & LOCKLEAF) == 0)
1297 		VOP_UNLOCK(dp);
1298 success:
1299 	/*
1300 	 * FIXME: for lookups which only cross a mount point to fetch the
1301 	 * root vnode, ni_dvp will be set to vp_crossmp. This can be a problem
1302 	 * if either WANTPARENT or LOCKPARENT is set.
1303 	 */
1304 	/*
1305 	 * Because of shared lookup we may have the vnode shared locked, but
1306 	 * the caller may want it to be exclusively locked.
1307 	 */
1308 	if (needs_exclusive_leaf(dp->v_mount, cnp->cn_flags) &&
1309 	    VOP_ISLOCKED(dp) != LK_EXCLUSIVE) {
1310 		vn_lock(dp, LK_UPGRADE | LK_RETRY);
1311 		if (VN_IS_DOOMED(dp)) {
1312 			error = ENOENT;
1313 			goto bad2;
1314 		}
1315 	}
1316 	if (ndp->ni_vp != NULL) {
1317 		nameicap_tracker_add(ndp, ndp->ni_vp);
1318 		if ((cnp->cn_flags & (FAILIFEXISTS | ISSYMLINK)) == FAILIFEXISTS)
1319 			goto bad_eexist;
1320 	}
1321 	return (0);
1322 
1323 bad2:
1324 	if (ni_dvp_unlocked != 2) {
1325 		if (dp != ndp->ni_dvp && !ni_dvp_unlocked)
1326 			vput(ndp->ni_dvp);
1327 		else
1328 			vrele(ndp->ni_dvp);
1329 	}
1330 bad:
1331 	if (!dpunlocked)
1332 		vput(dp);
1333 	ndp->ni_vp = NULL;
1334 	return (error);
1335 bad_eexist:
1336 	/*
1337 	 * FAILIFEXISTS handling.
1338 	 *
1339 	 * XXX namei called with LOCKPARENT but not LOCKLEAF has the strange
1340 	 * behaviour of leaving the vnode unlocked if the target is the same
1341 	 * vnode as the parent.
1342 	 */
1343 	MPASS((cnp->cn_flags & ISSYMLINK) == 0);
1344 	if (ndp->ni_vp == ndp->ni_dvp)
1345 		vrele(ndp->ni_dvp);
1346 	else
1347 		vput(ndp->ni_dvp);
1348 	vrele(ndp->ni_vp);
1349 	ndp->ni_dvp = NULL;
1350 	ndp->ni_vp = NULL;
1351 	NDFREE(ndp, NDF_ONLY_PNBUF);
1352 	return (EEXIST);
1353 }
1354 
1355 /*
1356  * relookup - lookup a path name component
1357  *    Used by lookup to re-acquire things.
1358  */
1359 int
1360 relookup(struct vnode *dvp, struct vnode **vpp, struct componentname *cnp)
1361 {
1362 	struct vnode *dp = NULL;		/* the directory we are searching */
1363 	int rdonly;			/* lookup read-only flag bit */
1364 	int error = 0;
1365 
1366 	KASSERT(cnp->cn_flags & ISLASTCN,
1367 	    ("relookup: Not given last component."));
1368 	/*
1369 	 * Setup: break out flag bits into variables.
1370 	 */
1371 	KASSERT((cnp->cn_flags & (LOCKPARENT | WANTPARENT)) != 0,
1372 	    ("relookup: parent not wanted"));
1373 	rdonly = cnp->cn_flags & RDONLY;
1374 	cnp->cn_flags &= ~ISSYMLINK;
1375 	dp = dvp;
1376 	cnp->cn_lkflags = LK_EXCLUSIVE;
1377 	vn_lock(dp, LK_EXCLUSIVE | LK_RETRY);
1378 
1379 	/*
1380 	 * Search a new directory.
1381 	 *
1382 	 * The last component of the filename is left accessible via
1383 	 * cnp->cn_nameptr for callers that need the name. Callers needing
1384 	 * the name set the SAVENAME flag. When done, they assume
1385 	 * responsibility for freeing the pathname buffer.
1386 	 */
1387 #ifdef NAMEI_DIAGNOSTIC
1388 	printf("{%s}: ", cnp->cn_nameptr);
1389 #endif
1390 
1391 	/*
1392 	 * Check for "" which represents the root directory after slash
1393 	 * removal.
1394 	 */
1395 	if (cnp->cn_nameptr[0] == '\0') {
1396 		/*
1397 		 * Support only LOOKUP for "/" because lookup()
1398 		 * can't succeed for CREATE, DELETE and RENAME.
1399 		 */
1400 		KASSERT(cnp->cn_nameiop == LOOKUP, ("nameiop must be LOOKUP"));
1401 		KASSERT(dp->v_type == VDIR, ("dp is not a directory"));
1402 
1403 		if (!(cnp->cn_flags & LOCKLEAF))
1404 			VOP_UNLOCK(dp);
1405 		*vpp = dp;
1406 		/* XXX This should probably move to the top of function. */
1407 		if (cnp->cn_flags & SAVESTART)
1408 			panic("lookup: SAVESTART");
1409 		return (0);
1410 	}
1411 
1412 	if (cnp->cn_flags & ISDOTDOT)
1413 		panic ("relookup: lookup on dot-dot");
1414 
1415 	/*
1416 	 * We now have a segment name to search for, and a directory to search.
1417 	 */
1418 #ifdef NAMEI_DIAGNOSTIC
1419 	vn_printf(dp, "search in ");
1420 #endif
1421 	if ((error = VOP_LOOKUP(dp, vpp, cnp)) != 0) {
1422 		KASSERT(*vpp == NULL, ("leaf should be empty"));
1423 		if (error != EJUSTRETURN)
1424 			goto bad;
1425 		/*
1426 		 * If creating and at end of pathname, then can consider
1427 		 * allowing file to be created.
1428 		 */
1429 		if (rdonly) {
1430 			error = EROFS;
1431 			goto bad;
1432 		}
1433 		/* ASSERT(dvp == ndp->ni_startdir) */
1434 		if (cnp->cn_flags & SAVESTART)
1435 			VREF(dvp);
1436 		if ((cnp->cn_flags & LOCKPARENT) == 0)
1437 			VOP_UNLOCK(dp);
1438 		/*
1439 		 * We return with ni_vp NULL to indicate that the entry
1440 		 * doesn't currently exist, leaving a pointer to the
1441 		 * (possibly locked) directory vnode in ndp->ni_dvp.
1442 		 */
1443 		return (0);
1444 	}
1445 
1446 	dp = *vpp;
1447 
1448 	/*
1449 	 * Disallow directory write attempts on read-only filesystems.
1450 	 */
1451 	if (rdonly &&
1452 	    (cnp->cn_nameiop == DELETE || cnp->cn_nameiop == RENAME)) {
1453 		if (dvp == dp)
1454 			vrele(dvp);
1455 		else
1456 			vput(dvp);
1457 		error = EROFS;
1458 		goto bad;
1459 	}
1460 	/*
1461 	 * Set the parent lock/ref state to the requested state.
1462 	 */
1463 	if ((cnp->cn_flags & LOCKPARENT) == 0 && dvp != dp)
1464 		VOP_UNLOCK(dvp);
1465 	/*
1466 	 * Check for symbolic link
1467 	 */
1468 	KASSERT(dp->v_type != VLNK || !(cnp->cn_flags & FOLLOW),
1469 	    ("relookup: symlink found.\n"));
1470 
1471 	/* ASSERT(dvp == ndp->ni_startdir) */
1472 	if (cnp->cn_flags & SAVESTART)
1473 		VREF(dvp);
1474 
1475 	if ((cnp->cn_flags & LOCKLEAF) == 0)
1476 		VOP_UNLOCK(dp);
1477 	return (0);
1478 bad:
1479 	vput(dp);
1480 	*vpp = NULL;
1481 	return (error);
1482 }
1483 
1484 /*
1485  * Free data allocated by namei(); see namei(9) for details.
1486  */
1487 void
1488 NDFREE_PNBUF(struct nameidata *ndp)
1489 {
1490 
1491 	if ((ndp->ni_cnd.cn_flags & HASBUF) != 0) {
1492 		MPASS((ndp->ni_cnd.cn_flags & (SAVENAME | SAVESTART)) != 0);
1493 		uma_zfree(namei_zone, ndp->ni_cnd.cn_pnbuf);
1494 		ndp->ni_cnd.cn_flags &= ~HASBUF;
1495 	}
1496 }
1497 
1498 /*
1499  * NDFREE_PNBUF replacement for callers that know there is no buffer.
1500  *
1501  * This is a hack. Preferably the VFS layer would not produce anything more
1502  * than it was asked to do. Unfortunately several non-LOOKUP cases can add the
1503  * HASBUF flag to the result. Even then an interface could be implemented where
1504  * the caller specifies what they expected to see in the result and what they
1505  * are going to take care of.
1506  *
1507  * In the meantime provide this kludge as a trivial replacement for NDFREE_PNBUF
1508  * calls scattered throughout the kernel where we know for a fact the flag must not
1509  * be seen.
1510  */
1511 #ifdef INVARIANTS
1512 void
1513 NDFREE_NOTHING(struct nameidata *ndp)
1514 {
1515 	struct componentname *cnp;
1516 
1517 	cnp = &ndp->ni_cnd;
1518 	KASSERT(cnp->cn_nameiop == LOOKUP, ("%s: got non-LOOKUP op %d\n",
1519 	    __func__, cnp->cn_nameiop));
1520 	KASSERT((cnp->cn_flags & (SAVENAME | HASBUF)) == 0,
1521 	    ("%s: bad flags \%" PRIx64 "\n", __func__, cnp->cn_flags));
1522 }
1523 #endif
1524 
1525 void
1526 (NDFREE)(struct nameidata *ndp, const u_int flags)
1527 {
1528 	int unlock_dvp;
1529 	int unlock_vp;
1530 
1531 	unlock_dvp = 0;
1532 	unlock_vp = 0;
1533 
1534 	if (!(flags & NDF_NO_FREE_PNBUF)) {
1535 		NDFREE_PNBUF(ndp);
1536 	}
1537 	if (!(flags & NDF_NO_VP_UNLOCK) &&
1538 	    (ndp->ni_cnd.cn_flags & LOCKLEAF) && ndp->ni_vp)
1539 		unlock_vp = 1;
1540 	if (!(flags & NDF_NO_DVP_UNLOCK) &&
1541 	    (ndp->ni_cnd.cn_flags & LOCKPARENT) &&
1542 	    ndp->ni_dvp != ndp->ni_vp)
1543 		unlock_dvp = 1;
1544 	if (!(flags & NDF_NO_VP_RELE) && ndp->ni_vp) {
1545 		if (unlock_vp) {
1546 			vput(ndp->ni_vp);
1547 			unlock_vp = 0;
1548 		} else
1549 			vrele(ndp->ni_vp);
1550 		ndp->ni_vp = NULL;
1551 	}
1552 	if (unlock_vp)
1553 		VOP_UNLOCK(ndp->ni_vp);
1554 	if (!(flags & NDF_NO_DVP_RELE) &&
1555 	    (ndp->ni_cnd.cn_flags & (LOCKPARENT|WANTPARENT))) {
1556 		if (unlock_dvp) {
1557 			vput(ndp->ni_dvp);
1558 			unlock_dvp = 0;
1559 		} else
1560 			vrele(ndp->ni_dvp);
1561 		ndp->ni_dvp = NULL;
1562 	}
1563 	if (unlock_dvp)
1564 		VOP_UNLOCK(ndp->ni_dvp);
1565 	if (!(flags & NDF_NO_STARTDIR_RELE) &&
1566 	    (ndp->ni_cnd.cn_flags & SAVESTART)) {
1567 		vrele(ndp->ni_startdir);
1568 		ndp->ni_startdir = NULL;
1569 	}
1570 }
1571 
1572 #ifdef INVARIANTS
1573 /*
1574  * Validate the final state of ndp after the lookup.
1575  *
1576  * Historically filesystems were allowed to modify cn_flags. Most notably they
1577  * can add SAVENAME to the request, resulting in HASBUF and pushing subsequent
1578  * clean up to the consumer. In practice this seems to only concern != LOOKUP
1579  * operations.
1580  *
1581  * As a step towards stricter API contract this routine validates the state to
1582  * clean up. Note validation is a work in progress with the intent of becoming
1583  * stricter over time.
1584  */
1585 #define NDMODIFYINGFLAGS (LOCKLEAF | LOCKPARENT | WANTPARENT | SAVENAME | SAVESTART | HASBUF)
1586 void
1587 NDVALIDATE(struct nameidata *ndp)
1588 {
1589 	struct componentname *cnp;
1590 	u_int64_t used, orig;
1591 
1592 	cnp = &ndp->ni_cnd;
1593 	orig = cnp->cn_origflags;
1594 	used = cnp->cn_flags;
1595 	switch (cnp->cn_nameiop) {
1596 	case LOOKUP:
1597 		/*
1598 		 * For plain lookup we require strict conformance -- nothing
1599 		 * to clean up if it was not requested by the caller.
1600 		 */
1601 		orig &= NDMODIFYINGFLAGS;
1602 		used &= NDMODIFYINGFLAGS;
1603 		if ((orig & (SAVENAME | SAVESTART)) != 0)
1604 			orig |= HASBUF;
1605 		if (orig != used) {
1606 			goto out_mismatch;
1607 		}
1608 		break;
1609 	case CREATE:
1610 	case DELETE:
1611 	case RENAME:
1612 		/*
1613 		 * Some filesystems set SAVENAME to provoke HASBUF, accomodate
1614 		 * for it until it gets fixed.
1615 		 */
1616 		orig &= NDMODIFYINGFLAGS;
1617 		orig |= (SAVENAME | HASBUF);
1618 		used &= NDMODIFYINGFLAGS;
1619 		used |= (SAVENAME | HASBUF);
1620 		if (orig != used) {
1621 			goto out_mismatch;
1622 		}
1623 		break;
1624 	}
1625 	return;
1626 out_mismatch:
1627 	panic("%s: mismatched flags for op %d: added %" PRIx64 ", "
1628 	    "removed %" PRIx64" (%" PRIx64" != %" PRIx64"; stored %" PRIx64" != %" PRIx64")",
1629 	    __func__, cnp->cn_nameiop, used & ~orig, orig &~ used,
1630 	    orig, used, cnp->cn_origflags, cnp->cn_flags);
1631 }
1632 #endif
1633 
1634 /*
1635  * Determine if there is a suitable alternate filename under the specified
1636  * prefix for the specified path.  If the create flag is set, then the
1637  * alternate prefix will be used so long as the parent directory exists.
1638  * This is used by the various compatibility ABIs so that Linux binaries prefer
1639  * files under /compat/linux for example.  The chosen path (whether under
1640  * the prefix or under /) is returned in a kernel malloc'd buffer pointed
1641  * to by pathbuf.  The caller is responsible for free'ing the buffer from
1642  * the M_TEMP bucket if one is returned.
1643  */
1644 int
1645 kern_alternate_path(struct thread *td, const char *prefix, const char *path,
1646     enum uio_seg pathseg, char **pathbuf, int create, int dirfd)
1647 {
1648 	struct nameidata nd, ndroot;
1649 	char *ptr, *buf, *cp;
1650 	size_t len, sz;
1651 	int error;
1652 
1653 	buf = (char *) malloc(MAXPATHLEN, M_TEMP, M_WAITOK);
1654 	*pathbuf = buf;
1655 
1656 	/* Copy the prefix into the new pathname as a starting point. */
1657 	len = strlcpy(buf, prefix, MAXPATHLEN);
1658 	if (len >= MAXPATHLEN) {
1659 		*pathbuf = NULL;
1660 		free(buf, M_TEMP);
1661 		return (EINVAL);
1662 	}
1663 	sz = MAXPATHLEN - len;
1664 	ptr = buf + len;
1665 
1666 	/* Append the filename to the prefix. */
1667 	if (pathseg == UIO_SYSSPACE)
1668 		error = copystr(path, ptr, sz, &len);
1669 	else
1670 		error = copyinstr(path, ptr, sz, &len);
1671 
1672 	if (error) {
1673 		*pathbuf = NULL;
1674 		free(buf, M_TEMP);
1675 		return (error);
1676 	}
1677 
1678 	/* Only use a prefix with absolute pathnames. */
1679 	if (*ptr != '/') {
1680 		error = EINVAL;
1681 		goto keeporig;
1682 	}
1683 
1684 	if (dirfd != AT_FDCWD) {
1685 		/*
1686 		 * We want the original because the "prefix" is
1687 		 * included in the already opened dirfd.
1688 		 */
1689 		bcopy(ptr, buf, len);
1690 		return (0);
1691 	}
1692 
1693 	/*
1694 	 * We know that there is a / somewhere in this pathname.
1695 	 * Search backwards for it, to find the file's parent dir
1696 	 * to see if it exists in the alternate tree. If it does,
1697 	 * and we want to create a file (cflag is set). We don't
1698 	 * need to worry about the root comparison in this case.
1699 	 */
1700 
1701 	if (create) {
1702 		for (cp = &ptr[len] - 1; *cp != '/'; cp--);
1703 		*cp = '\0';
1704 
1705 		NDINIT(&nd, LOOKUP, NOFOLLOW, UIO_SYSSPACE, buf, td);
1706 		error = namei(&nd);
1707 		*cp = '/';
1708 		if (error != 0)
1709 			goto keeporig;
1710 	} else {
1711 		NDINIT(&nd, LOOKUP, NOFOLLOW, UIO_SYSSPACE, buf, td);
1712 
1713 		error = namei(&nd);
1714 		if (error != 0)
1715 			goto keeporig;
1716 
1717 		/*
1718 		 * We now compare the vnode of the prefix to the one
1719 		 * vnode asked. If they resolve to be the same, then we
1720 		 * ignore the match so that the real root gets used.
1721 		 * This avoids the problem of traversing "../.." to find the
1722 		 * root directory and never finding it, because "/" resolves
1723 		 * to the emulation root directory. This is expensive :-(
1724 		 */
1725 		NDINIT(&ndroot, LOOKUP, FOLLOW, UIO_SYSSPACE, prefix,
1726 		    td);
1727 
1728 		/* We shouldn't ever get an error from this namei(). */
1729 		error = namei(&ndroot);
1730 		if (error == 0) {
1731 			if (nd.ni_vp == ndroot.ni_vp)
1732 				error = ENOENT;
1733 
1734 			NDFREE(&ndroot, NDF_ONLY_PNBUF);
1735 			vrele(ndroot.ni_vp);
1736 		}
1737 	}
1738 
1739 	NDFREE(&nd, NDF_ONLY_PNBUF);
1740 	vrele(nd.ni_vp);
1741 
1742 keeporig:
1743 	/* If there was an error, use the original path name. */
1744 	if (error)
1745 		bcopy(ptr, buf, len);
1746 	return (error);
1747 }
1748