xref: /freebsd/sys/kern/vfs_lookup.c (revision 8ddb146abcdf061be9f2c0db7e391697dafad85c)
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/dirent.h>
48 #include <sys/kernel.h>
49 #include <sys/capsicum.h>
50 #include <sys/fcntl.h>
51 #include <sys/jail.h>
52 #include <sys/lock.h>
53 #include <sys/mutex.h>
54 #include <sys/namei.h>
55 #include <sys/vnode.h>
56 #include <sys/mount.h>
57 #include <sys/filedesc.h>
58 #include <sys/proc.h>
59 #include <sys/sdt.h>
60 #include <sys/syscallsubr.h>
61 #include <sys/sysctl.h>
62 #ifdef KTRACE
63 #include <sys/ktrace.h>
64 #endif
65 #ifdef INVARIANTS
66 #include <machine/_inttypes.h>
67 #endif
68 
69 #include <security/audit/audit.h>
70 #include <security/mac/mac_framework.h>
71 
72 #include <vm/uma.h>
73 
74 #define	NAMEI_DIAGNOSTIC 1
75 #undef NAMEI_DIAGNOSTIC
76 
77 SDT_PROVIDER_DEFINE(vfs);
78 SDT_PROBE_DEFINE4(vfs, namei, lookup, entry, "struct vnode *", "char *",
79     "unsigned long", "bool");
80 SDT_PROBE_DEFINE4(vfs, namei, lookup, return, "int", "struct vnode *", "bool",
81     "struct nameidata");
82 
83 /* Allocation zone for namei. */
84 uma_zone_t namei_zone;
85 
86 /* Placeholder vnode for mp traversal. */
87 static struct vnode *vp_crossmp;
88 
89 static int
90 crossmp_vop_islocked(struct vop_islocked_args *ap)
91 {
92 
93 	return (LK_SHARED);
94 }
95 
96 static int
97 crossmp_vop_lock1(struct vop_lock1_args *ap)
98 {
99 	struct vnode *vp;
100 	struct lock *lk __diagused;
101 	const char *file __witness_used;
102 	int flags, line __witness_used;
103 
104 	vp = ap->a_vp;
105 	lk = vp->v_vnlock;
106 	flags = ap->a_flags;
107 	file = ap->a_file;
108 	line = ap->a_line;
109 
110 	if ((flags & LK_SHARED) == 0)
111 		panic("invalid lock request for crossmp");
112 
113 	WITNESS_CHECKORDER(&lk->lock_object, LOP_NEWORDER, file, line,
114 	    flags & LK_INTERLOCK ? &VI_MTX(vp)->lock_object : NULL);
115 	WITNESS_LOCK(&lk->lock_object, 0, file, line);
116 	if ((flags & LK_INTERLOCK) != 0)
117 		VI_UNLOCK(vp);
118 	LOCK_LOG_LOCK("SLOCK", &lk->lock_object, 0, 0, ap->a_file, ap->a_line);
119 	return (0);
120 }
121 
122 static int
123 crossmp_vop_unlock(struct vop_unlock_args *ap)
124 {
125 	struct vnode *vp;
126 	struct lock *lk __diagused;
127 
128 	vp = ap->a_vp;
129 	lk = vp->v_vnlock;
130 
131 	WITNESS_UNLOCK(&lk->lock_object, 0, LOCK_FILE, LOCK_LINE);
132 	LOCK_LOG_LOCK("SUNLOCK", &lk->lock_object, 0, 0, LOCK_FILE,
133 	    LOCK_LINE);
134 	return (0);
135 }
136 
137 static struct vop_vector crossmp_vnodeops = {
138 	.vop_default =		&default_vnodeops,
139 	.vop_islocked =		crossmp_vop_islocked,
140 	.vop_lock1 =		crossmp_vop_lock1,
141 	.vop_unlock =		crossmp_vop_unlock,
142 };
143 /*
144  * VFS_VOP_VECTOR_REGISTER(crossmp_vnodeops) is not used here since the vnode
145  * gets allocated early. See nameiinit for the direct call below.
146  */
147 
148 struct nameicap_tracker {
149 	struct vnode *dp;
150 	TAILQ_ENTRY(nameicap_tracker) nm_link;
151 };
152 
153 /* Zone for cap mode tracker elements used for dotdot capability checks. */
154 MALLOC_DEFINE(M_NAMEITRACKER, "namei_tracker", "namei tracking for dotdot");
155 
156 static void
157 nameiinit(void *dummy __unused)
158 {
159 
160 	namei_zone = uma_zcreate("NAMEI", MAXPATHLEN, NULL, NULL, NULL, NULL,
161 	    UMA_ALIGN_PTR, 0);
162 	vfs_vector_op_register(&crossmp_vnodeops);
163 	getnewvnode("crossmp", NULL, &crossmp_vnodeops, &vp_crossmp);
164 }
165 SYSINIT(vfs, SI_SUB_VFS, SI_ORDER_SECOND, nameiinit, NULL);
166 
167 static int lookup_cap_dotdot = 1;
168 SYSCTL_INT(_vfs, OID_AUTO, lookup_cap_dotdot, CTLFLAG_RWTUN,
169     &lookup_cap_dotdot, 0,
170     "enables \"..\" components in path lookup in capability mode");
171 static int lookup_cap_dotdot_nonlocal = 1;
172 SYSCTL_INT(_vfs, OID_AUTO, lookup_cap_dotdot_nonlocal, CTLFLAG_RWTUN,
173     &lookup_cap_dotdot_nonlocal, 0,
174     "enables \"..\" components in path lookup in capability mode "
175     "on non-local mount");
176 
177 static void
178 nameicap_tracker_add(struct nameidata *ndp, struct vnode *dp)
179 {
180 	struct nameicap_tracker *nt;
181 
182 	if ((ndp->ni_lcf & NI_LCF_CAP_DOTDOT) == 0 || dp->v_type != VDIR)
183 		return;
184 	nt = TAILQ_LAST(&ndp->ni_cap_tracker, nameicap_tracker_head);
185 	if (nt != NULL && nt->dp == dp)
186 		return;
187 	nt = malloc(sizeof(*nt), M_NAMEITRACKER, M_WAITOK);
188 	vhold(dp);
189 	nt->dp = dp;
190 	TAILQ_INSERT_TAIL(&ndp->ni_cap_tracker, nt, nm_link);
191 }
192 
193 static void
194 nameicap_cleanup_from(struct nameidata *ndp, struct nameicap_tracker *first)
195 {
196 	struct nameicap_tracker *nt, *nt1;
197 
198 	nt = first;
199 	TAILQ_FOREACH_FROM_SAFE(nt, &ndp->ni_cap_tracker, nm_link, nt1) {
200 		TAILQ_REMOVE(&ndp->ni_cap_tracker, nt, nm_link);
201 		vdrop(nt->dp);
202 		free(nt, M_NAMEITRACKER);
203 	}
204 }
205 
206 static void
207 nameicap_cleanup(struct nameidata *ndp)
208 {
209 	KASSERT(TAILQ_EMPTY(&ndp->ni_cap_tracker) ||
210 	    (ndp->ni_lcf & NI_LCF_CAP_DOTDOT) != 0, ("not strictrelative"));
211 	nameicap_cleanup_from(ndp, NULL);
212 }
213 
214 /*
215  * For dotdot lookups in capability mode, only allow the component
216  * lookup to succeed if the resulting directory was already traversed
217  * during the operation.  This catches situations where already
218  * traversed directory is moved to different parent, and then we walk
219  * over it with dotdots.
220  *
221  * Also allow to force failure of dotdot lookups for non-local
222  * filesystems, where external agents might assist local lookups to
223  * escape the compartment.
224  */
225 static int
226 nameicap_check_dotdot(struct nameidata *ndp, struct vnode *dp)
227 {
228 	struct nameicap_tracker *nt;
229 	struct mount *mp;
230 
231 	if (dp == NULL || dp->v_type != VDIR || (ndp->ni_lcf &
232 	    NI_LCF_STRICTRELATIVE) == 0)
233 		return (0);
234 	if ((ndp->ni_lcf & NI_LCF_CAP_DOTDOT) == 0)
235 		return (ENOTCAPABLE);
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 (dp == nt->dp) {
243 			nt = TAILQ_NEXT(nt, nm_link);
244 			if (nt != NULL)
245 				nameicap_cleanup_from(ndp, nt);
246 			return (0);
247 		}
248 	}
249 	return (ENOTCAPABLE);
250 }
251 
252 static void
253 namei_cleanup_cnp(struct componentname *cnp)
254 {
255 
256 	uma_zfree(namei_zone, cnp->cn_pnbuf);
257 #ifdef DIAGNOSTIC
258 	cnp->cn_pnbuf = NULL;
259 	cnp->cn_nameptr = NULL;
260 #endif
261 }
262 
263 static int
264 namei_handle_root(struct nameidata *ndp, struct vnode **dpp)
265 {
266 	struct componentname *cnp;
267 
268 	cnp = &ndp->ni_cnd;
269 	if ((ndp->ni_lcf & NI_LCF_STRICTRELATIVE) != 0) {
270 #ifdef KTRACE
271 		if (KTRPOINT(curthread, KTR_CAPFAIL))
272 			ktrcapfail(CAPFAIL_LOOKUP, NULL, NULL);
273 #endif
274 		return (ENOTCAPABLE);
275 	}
276 	while (*(cnp->cn_nameptr) == '/') {
277 		cnp->cn_nameptr++;
278 		ndp->ni_pathlen--;
279 	}
280 	*dpp = ndp->ni_rootdir;
281 	vrefact(*dpp);
282 	return (0);
283 }
284 
285 static int
286 namei_setup(struct nameidata *ndp, struct vnode **dpp, struct pwd **pwdp)
287 {
288 	struct componentname *cnp;
289 	struct thread *td;
290 	struct pwd *pwd;
291 	int error;
292 	bool startdir_used;
293 
294 	cnp = &ndp->ni_cnd;
295 	td = curthread;
296 
297 	startdir_used = false;
298 	*pwdp = NULL;
299 	*dpp = NULL;
300 
301 #ifdef CAPABILITY_MODE
302 	/*
303 	 * In capability mode, lookups must be restricted to happen in
304 	 * the subtree with the root specified by the file descriptor:
305 	 * - The root must be real file descriptor, not the pseudo-descriptor
306 	 *   AT_FDCWD.
307 	 * - The passed path must be relative and not absolute.
308 	 * - If lookup_cap_dotdot is disabled, path must not contain the
309 	 *   '..' components.
310 	 * - If lookup_cap_dotdot is enabled, we verify that all '..'
311 	 *   components lookups result in the directories which were
312 	 *   previously walked by us, which prevents an escape from
313 	 *   the relative root.
314 	 */
315 	if (IN_CAPABILITY_MODE(td) && (cnp->cn_flags & NOCAPCHECK) == 0) {
316 		ndp->ni_lcf |= NI_LCF_STRICTRELATIVE;
317 		ndp->ni_resflags |= NIRES_STRICTREL;
318 		if (ndp->ni_dirfd == AT_FDCWD) {
319 #ifdef KTRACE
320 			if (KTRPOINT(td, KTR_CAPFAIL))
321 				ktrcapfail(CAPFAIL_LOOKUP, NULL, NULL);
322 #endif
323 			return (ECAPMODE);
324 		}
325 	}
326 #endif
327 	error = 0;
328 
329 	/*
330 	 * Get starting point for the translation.
331 	 */
332 	pwd = pwd_hold(td);
333 	/*
334 	 * The reference on ni_rootdir is acquired in the block below to avoid
335 	 * back-to-back atomics for absolute lookups.
336 	 */
337 	ndp->ni_rootdir = pwd->pwd_rdir;
338 	ndp->ni_topdir = pwd->pwd_jdir;
339 
340 	if (cnp->cn_pnbuf[0] == '/') {
341 		ndp->ni_resflags |= NIRES_ABS;
342 		error = namei_handle_root(ndp, dpp);
343 	} else {
344 		if (ndp->ni_startdir != NULL) {
345 			*dpp = ndp->ni_startdir;
346 			startdir_used = true;
347 		} else if (ndp->ni_dirfd == AT_FDCWD) {
348 			*dpp = pwd->pwd_cdir;
349 			vrefact(*dpp);
350 		} else {
351 			if (cnp->cn_flags & AUDITVNODE1)
352 				AUDIT_ARG_ATFD1(ndp->ni_dirfd);
353 			if (cnp->cn_flags & AUDITVNODE2)
354 				AUDIT_ARG_ATFD2(ndp->ni_dirfd);
355 
356 			error = fgetvp_lookup(ndp->ni_dirfd, ndp, dpp);
357 		}
358 		if (error == 0 && (*dpp)->v_type != VDIR &&
359 		    (cnp->cn_pnbuf[0] != '\0' ||
360 		    (cnp->cn_flags & EMPTYPATH) == 0))
361 			error = ENOTDIR;
362 	}
363 	if (error == 0 && (cnp->cn_flags & RBENEATH) != 0) {
364 		if (cnp->cn_pnbuf[0] == '/') {
365 			error = ENOTCAPABLE;
366 		} else if ((ndp->ni_lcf & NI_LCF_STRICTRELATIVE) == 0) {
367 			ndp->ni_lcf |= NI_LCF_STRICTRELATIVE |
368 			    NI_LCF_CAP_DOTDOT;
369 		}
370 	}
371 
372 	/*
373 	 * If we are auditing the kernel pathname, save the user pathname.
374 	 */
375 	if (cnp->cn_flags & AUDITVNODE1)
376 		AUDIT_ARG_UPATH1_VP(td, ndp->ni_rootdir, *dpp, cnp->cn_pnbuf);
377 	if (cnp->cn_flags & AUDITVNODE2)
378 		AUDIT_ARG_UPATH2_VP(td, ndp->ni_rootdir, *dpp, cnp->cn_pnbuf);
379 	if (ndp->ni_startdir != NULL && !startdir_used)
380 		vrele(ndp->ni_startdir);
381 	if (error != 0) {
382 		if (*dpp != NULL)
383 			vrele(*dpp);
384 		pwd_drop(pwd);
385 		return (error);
386 	}
387 	if ((ndp->ni_lcf & NI_LCF_STRICTRELATIVE) != 0 &&
388 	    lookup_cap_dotdot != 0)
389 		ndp->ni_lcf |= NI_LCF_CAP_DOTDOT;
390 	SDT_PROBE4(vfs, namei, lookup, entry, *dpp, cnp->cn_pnbuf,
391 	    cnp->cn_flags, false);
392 	*pwdp = pwd;
393 	return (0);
394 }
395 
396 static int
397 namei_getpath(struct nameidata *ndp)
398 {
399 	struct componentname *cnp;
400 	int error;
401 
402 	cnp = &ndp->ni_cnd;
403 
404 	/*
405 	 * Get a buffer for the name to be translated, and copy the
406 	 * name into the buffer.
407 	 */
408 	cnp->cn_pnbuf = uma_zalloc(namei_zone, M_WAITOK);
409 	if (ndp->ni_segflg == UIO_SYSSPACE) {
410 		error = copystr(ndp->ni_dirp, cnp->cn_pnbuf, MAXPATHLEN,
411 		    &ndp->ni_pathlen);
412 	} else {
413 		error = copyinstr(ndp->ni_dirp, cnp->cn_pnbuf, MAXPATHLEN,
414 		    &ndp->ni_pathlen);
415 	}
416 
417 	if (__predict_false(error != 0))
418 		return (error);
419 
420 	cnp->cn_nameptr = cnp->cn_pnbuf;
421 	return (0);
422 }
423 
424 static int
425 namei_emptypath(struct nameidata *ndp)
426 {
427 	struct componentname *cnp;
428 	struct pwd *pwd;
429 	struct vnode *dp;
430 	int error;
431 
432 	cnp = &ndp->ni_cnd;
433 	MPASS(*cnp->cn_pnbuf == '\0');
434 	MPASS((cnp->cn_flags & EMPTYPATH) != 0);
435 	MPASS((cnp->cn_flags & (LOCKPARENT | WANTPARENT)) == 0);
436 
437 	ndp->ni_resflags |= NIRES_EMPTYPATH;
438 	error = namei_setup(ndp, &dp, &pwd);
439 	if (error != 0) {
440 		namei_cleanup_cnp(cnp);
441 		goto errout;
442 	}
443 
444 	/*
445 	 * Usecount on dp already provided by namei_setup.
446 	 */
447 	ndp->ni_vp = dp;
448 	namei_cleanup_cnp(cnp);
449 	pwd_drop(pwd);
450 	NDVALIDATE(ndp);
451 	if ((cnp->cn_flags & LOCKLEAF) != 0) {
452 		VOP_LOCK(dp, (cnp->cn_flags & LOCKSHARED) != 0 ?
453 		    LK_SHARED : LK_EXCLUSIVE);
454 		if (VN_IS_DOOMED(dp)) {
455 			vput(dp);
456 			error = ENOENT;
457 			goto errout;
458 		}
459 	}
460 	SDT_PROBE4(vfs, namei, lookup, return, 0, ndp->ni_vp, false, ndp);
461 	return (0);
462 
463 errout:
464 	SDT_PROBE4(vfs, namei, lookup, return, error, NULL, false, ndp);
465 	return (error);
466 }
467 
468 static int __noinline
469 namei_follow_link(struct nameidata *ndp)
470 {
471 	char *cp;
472 	struct iovec aiov;
473 	struct uio auio;
474 	struct componentname *cnp;
475 	struct thread *td;
476 	int error, linklen;
477 
478 	error = 0;
479 	cnp = &ndp->ni_cnd;
480 	td = curthread;
481 
482 	if (ndp->ni_loopcnt++ >= MAXSYMLINKS) {
483 		error = ELOOP;
484 		goto out;
485 	}
486 #ifdef MAC
487 	if ((cnp->cn_flags & NOMACCHECK) == 0) {
488 		error = mac_vnode_check_readlink(td->td_ucred, ndp->ni_vp);
489 		if (error != 0)
490 			goto out;
491 	}
492 #endif
493 	if (ndp->ni_pathlen > 1)
494 		cp = uma_zalloc(namei_zone, M_WAITOK);
495 	else
496 		cp = cnp->cn_pnbuf;
497 	aiov.iov_base = cp;
498 	aiov.iov_len = MAXPATHLEN;
499 	auio.uio_iov = &aiov;
500 	auio.uio_iovcnt = 1;
501 	auio.uio_offset = 0;
502 	auio.uio_rw = UIO_READ;
503 	auio.uio_segflg = UIO_SYSSPACE;
504 	auio.uio_td = td;
505 	auio.uio_resid = MAXPATHLEN;
506 	error = VOP_READLINK(ndp->ni_vp, &auio, cnp->cn_cred);
507 	if (error != 0) {
508 		if (ndp->ni_pathlen > 1)
509 			uma_zfree(namei_zone, cp);
510 		goto out;
511 	}
512 	linklen = MAXPATHLEN - auio.uio_resid;
513 	if (linklen == 0) {
514 		if (ndp->ni_pathlen > 1)
515 			uma_zfree(namei_zone, cp);
516 		error = ENOENT;
517 		goto out;
518 	}
519 	if (linklen + ndp->ni_pathlen > MAXPATHLEN) {
520 		if (ndp->ni_pathlen > 1)
521 			uma_zfree(namei_zone, cp);
522 		error = ENAMETOOLONG;
523 		goto out;
524 	}
525 	if (ndp->ni_pathlen > 1) {
526 		bcopy(ndp->ni_next, cp + linklen, ndp->ni_pathlen);
527 		uma_zfree(namei_zone, cnp->cn_pnbuf);
528 		cnp->cn_pnbuf = cp;
529 	} else
530 		cnp->cn_pnbuf[linklen] = '\0';
531 	ndp->ni_pathlen += linklen;
532 out:
533 	return (error);
534 }
535 
536 /*
537  * Convert a pathname into a pointer to a locked vnode.
538  *
539  * The FOLLOW flag is set when symbolic links are to be followed
540  * when they occur at the end of the name translation process.
541  * Symbolic links are always followed for all other pathname
542  * components other than the last.
543  *
544  * The segflg defines whether the name is to be copied from user
545  * space or kernel space.
546  *
547  * Overall outline of namei:
548  *
549  *	copy in name
550  *	get starting directory
551  *	while (!done && !error) {
552  *		call lookup to search path.
553  *		if symbolic link, massage name in buffer and continue
554  *	}
555  */
556 int
557 namei(struct nameidata *ndp)
558 {
559 	struct vnode *dp;	/* the directory we are searching */
560 	struct componentname *cnp;
561 	struct thread *td;
562 	struct pwd *pwd;
563 	int error;
564 	enum cache_fpl_status status;
565 
566 	cnp = &ndp->ni_cnd;
567 	td = curthread;
568 #ifdef INVARIANTS
569 	KASSERT((ndp->ni_debugflags & NAMEI_DBG_CALLED) == 0,
570 	    ("%s: repeated call to namei without NDREINIT", __func__));
571 	KASSERT(ndp->ni_debugflags == NAMEI_DBG_INITED,
572 	    ("%s: bad debugflags %d", __func__, ndp->ni_debugflags));
573 	ndp->ni_debugflags |= NAMEI_DBG_CALLED;
574 	if (ndp->ni_startdir != NULL)
575 		ndp->ni_debugflags |= NAMEI_DBG_HADSTARTDIR;
576 	if (cnp->cn_flags & FAILIFEXISTS) {
577 		KASSERT(cnp->cn_nameiop == CREATE,
578 		    ("%s: FAILIFEXISTS passed for op %d", __func__, cnp->cn_nameiop));
579 		/*
580 		 * The limitation below is to restrict hairy corner cases.
581 		 */
582 		KASSERT((cnp->cn_flags & (LOCKPARENT | LOCKLEAF)) == LOCKPARENT,
583 		    ("%s: FAILIFEXISTS must be passed with LOCKPARENT and without LOCKLEAF",
584 		    __func__));
585 	}
586 	/*
587 	 * For NDVALIDATE.
588 	 *
589 	 * While NDINIT may seem like a more natural place to do it, there are
590 	 * callers which directly modify flags past invoking init.
591 	 */
592 	cnp->cn_origflags = cnp->cn_flags;
593 #endif
594 	ndp->ni_cnd.cn_cred = td->td_ucred;
595 	KASSERT(ndp->ni_resflags == 0, ("%s: garbage in ni_resflags: %x\n",
596 	    __func__, ndp->ni_resflags));
597 	KASSERT(cnp->cn_cred && td->td_proc, ("namei: bad cred/proc"));
598 	KASSERT((cnp->cn_flags & NAMEI_INTERNAL_FLAGS) == 0,
599 	    ("namei: unexpected flags: %" PRIx64 "\n",
600 	    cnp->cn_flags & NAMEI_INTERNAL_FLAGS));
601 	if (cnp->cn_flags & NOCACHE)
602 		KASSERT(cnp->cn_nameiop != LOOKUP,
603 		    ("%s: NOCACHE passed with LOOKUP", __func__));
604 	MPASS(ndp->ni_startdir == NULL || ndp->ni_startdir->v_type == VDIR ||
605 	    ndp->ni_startdir->v_type == VBAD);
606 
607 	ndp->ni_lcf = 0;
608 	ndp->ni_loopcnt = 0;
609 	ndp->ni_vp = NULL;
610 
611 	error = namei_getpath(ndp);
612 	if (__predict_false(error != 0)) {
613 		namei_cleanup_cnp(cnp);
614 		SDT_PROBE4(vfs, namei, lookup, return, error, NULL,
615 		    false, ndp);
616 		return (error);
617 	}
618 
619 #ifdef KTRACE
620 	if (KTRPOINT(td, KTR_NAMEI)) {
621 		ktrnamei(cnp->cn_pnbuf);
622 	}
623 #endif
624 	TSNAMEI(curthread->td_proc->p_pid, cnp->cn_pnbuf);
625 
626 	/*
627 	 * First try looking up the target without locking any vnodes.
628 	 *
629 	 * We may need to start from scratch or pick up where it left off.
630 	 */
631 	error = cache_fplookup(ndp, &status, &pwd);
632 	switch (status) {
633 	case CACHE_FPL_STATUS_UNSET:
634 		__assert_unreachable();
635 		break;
636 	case CACHE_FPL_STATUS_HANDLED:
637 		if (error == 0)
638 			NDVALIDATE(ndp);
639 		return (error);
640 	case CACHE_FPL_STATUS_PARTIAL:
641 		TAILQ_INIT(&ndp->ni_cap_tracker);
642 		dp = ndp->ni_startdir;
643 		break;
644 	case CACHE_FPL_STATUS_DESTROYED:
645 		ndp->ni_loopcnt = 0;
646 		error = namei_getpath(ndp);
647 		if (__predict_false(error != 0)) {
648 			namei_cleanup_cnp(cnp);
649 			return (error);
650 		}
651 		/* FALLTHROUGH */
652 	case CACHE_FPL_STATUS_ABORTED:
653 		TAILQ_INIT(&ndp->ni_cap_tracker);
654 		MPASS(ndp->ni_lcf == 0);
655 		if (*cnp->cn_pnbuf == '\0') {
656 			if ((cnp->cn_flags & EMPTYPATH) != 0) {
657 				return (namei_emptypath(ndp));
658 			}
659 			namei_cleanup_cnp(cnp);
660 			SDT_PROBE4(vfs, namei, lookup, return, ENOENT, NULL,
661 			    false, ndp);
662 			return (ENOENT);
663 		}
664 		error = namei_setup(ndp, &dp, &pwd);
665 		if (error != 0) {
666 			namei_cleanup_cnp(cnp);
667 			return (error);
668 		}
669 		break;
670 	}
671 
672 	/*
673 	 * Locked lookup.
674 	 */
675 	for (;;) {
676 		ndp->ni_startdir = dp;
677 		error = vfs_lookup(ndp);
678 		if (error != 0)
679 			goto out;
680 
681 		/*
682 		 * If not a symbolic link, we're done.
683 		 */
684 		if ((cnp->cn_flags & ISSYMLINK) == 0) {
685 			SDT_PROBE4(vfs, namei, lookup, return, error,
686 			    (error == 0 ? ndp->ni_vp : NULL), false, ndp);
687 			if ((cnp->cn_flags & (SAVENAME | SAVESTART)) == 0) {
688 				namei_cleanup_cnp(cnp);
689 			} else
690 				cnp->cn_flags |= HASBUF;
691 			nameicap_cleanup(ndp);
692 			pwd_drop(pwd);
693 			if (error == 0)
694 				NDVALIDATE(ndp);
695 			return (error);
696 		}
697 		error = namei_follow_link(ndp);
698 		if (error != 0)
699 			break;
700 		vput(ndp->ni_vp);
701 		dp = ndp->ni_dvp;
702 		/*
703 		 * Check if root directory should replace current directory.
704 		 */
705 		cnp->cn_nameptr = cnp->cn_pnbuf;
706 		if (*(cnp->cn_nameptr) == '/') {
707 			vrele(dp);
708 			error = namei_handle_root(ndp, &dp);
709 			if (error != 0)
710 				goto out;
711 		}
712 	}
713 	vput(ndp->ni_vp);
714 	ndp->ni_vp = NULL;
715 	vrele(ndp->ni_dvp);
716 out:
717 	MPASS(error != 0);
718 	SDT_PROBE4(vfs, namei, lookup, return, error, NULL, false, ndp);
719 	namei_cleanup_cnp(cnp);
720 	nameicap_cleanup(ndp);
721 	pwd_drop(pwd);
722 	return (error);
723 }
724 
725 static int
726 compute_cn_lkflags(struct mount *mp, int lkflags, int cnflags)
727 {
728 
729 	if (mp == NULL || ((lkflags & LK_SHARED) &&
730 	    !(mp->mnt_kern_flag & MNTK_LOOKUP_SHARED))) {
731 		lkflags &= ~LK_SHARED;
732 		lkflags |= LK_EXCLUSIVE;
733 	}
734 	lkflags |= LK_NODDLKTREAT;
735 	return (lkflags);
736 }
737 
738 static __inline int
739 needs_exclusive_leaf(struct mount *mp, int flags)
740 {
741 
742 	/*
743 	 * Intermediate nodes can use shared locks, we only need to
744 	 * force an exclusive lock for leaf nodes.
745 	 */
746 	if ((flags & (ISLASTCN | LOCKLEAF)) != (ISLASTCN | LOCKLEAF))
747 		return (0);
748 
749 	/* Always use exclusive locks if LOCKSHARED isn't set. */
750 	if (!(flags & LOCKSHARED))
751 		return (1);
752 
753 	/*
754 	 * For lookups during open(), if the mount point supports
755 	 * extended shared operations, then use a shared lock for the
756 	 * leaf node, otherwise use an exclusive lock.
757 	 */
758 	if ((flags & ISOPEN) != 0)
759 		return (!MNT_EXTENDED_SHARED(mp));
760 
761 	/*
762 	 * Lookup requests outside of open() that specify LOCKSHARED
763 	 * only need a shared lock on the leaf vnode.
764 	 */
765 	return (0);
766 }
767 
768 /*
769  * Various filesystems expect to be able to copy a name component with length
770  * bounded by NAME_MAX into a directory entry buffer of size MAXNAMLEN.  Make
771  * sure that these are the same size.
772  */
773 _Static_assert(MAXNAMLEN == NAME_MAX,
774     "MAXNAMLEN and NAME_MAX have different values");
775 
776 static int __noinline
777 vfs_lookup_degenerate(struct nameidata *ndp, struct vnode *dp, int wantparent)
778 {
779 	struct componentname *cnp;
780 	struct mount *mp;
781 	int error;
782 
783 	cnp = &ndp->ni_cnd;
784 
785 	cnp->cn_flags |= ISLASTCN;
786 
787 	mp = atomic_load_ptr(&dp->v_mount);
788 	if (needs_exclusive_leaf(mp, cnp->cn_flags)) {
789 		cnp->cn_lkflags &= ~LK_SHARED;
790 		cnp->cn_lkflags |= LK_EXCLUSIVE;
791 	}
792 
793 	vn_lock(dp,
794 	    compute_cn_lkflags(mp, cnp->cn_lkflags | LK_RETRY,
795 	    cnp->cn_flags));
796 
797 	if (dp->v_type != VDIR) {
798 		error = ENOTDIR;
799 		goto bad;
800 	}
801 	if (cnp->cn_nameiop != LOOKUP) {
802 		error = EISDIR;
803 		goto bad;
804 	}
805 	if (wantparent) {
806 		ndp->ni_dvp = dp;
807 		VREF(dp);
808 	}
809 	ndp->ni_vp = dp;
810 	cnp->cn_namelen = 0;
811 
812 	if (cnp->cn_flags & AUDITVNODE1)
813 		AUDIT_ARG_VNODE1(dp);
814 	else if (cnp->cn_flags & AUDITVNODE2)
815 		AUDIT_ARG_VNODE2(dp);
816 
817 	if (!(cnp->cn_flags & (LOCKPARENT | LOCKLEAF)))
818 		VOP_UNLOCK(dp);
819 	/* XXX This should probably move to the top of function. */
820 	if (cnp->cn_flags & SAVESTART)
821 		panic("lookup: SAVESTART");
822 	return (0);
823 bad:
824 	VOP_UNLOCK(dp);
825 	return (error);
826 }
827 
828 /*
829  * FAILIFEXISTS handling.
830  *
831  * XXX namei called with LOCKPARENT but not LOCKLEAF has the strange
832  * behaviour of leaving the vnode unlocked if the target is the same
833  * vnode as the parent.
834  */
835 static int __noinline
836 vfs_lookup_failifexists(struct nameidata *ndp)
837 {
838 	struct componentname *cnp __diagused;
839 
840 	cnp = &ndp->ni_cnd;
841 
842 	MPASS((cnp->cn_flags & ISSYMLINK) == 0);
843 	if (ndp->ni_vp == ndp->ni_dvp)
844 		vrele(ndp->ni_dvp);
845 	else
846 		vput(ndp->ni_dvp);
847 	vrele(ndp->ni_vp);
848 	ndp->ni_dvp = NULL;
849 	ndp->ni_vp = NULL;
850 	NDFREE_PNBUF(ndp);
851 	return (EEXIST);
852 }
853 
854 /*
855  * Search a pathname.
856  * This is a very central and rather complicated routine.
857  *
858  * The pathname is pointed to by ni_ptr and is of length ni_pathlen.
859  * The starting directory is taken from ni_startdir. The pathname is
860  * descended until done, or a symbolic link is encountered. The variable
861  * ni_more is clear if the path is completed; it is set to one if a
862  * symbolic link needing interpretation is encountered.
863  *
864  * The flag argument is LOOKUP, CREATE, RENAME, or DELETE depending on
865  * whether the name is to be looked up, created, renamed, or deleted.
866  * When CREATE, RENAME, or DELETE is specified, information usable in
867  * creating, renaming, or deleting a directory entry may be calculated.
868  * If flag has LOCKPARENT or'ed into it, the parent directory is returned
869  * locked. If flag has WANTPARENT or'ed into it, the parent directory is
870  * returned unlocked. Otherwise the parent directory is not returned. If
871  * the target of the pathname exists and LOCKLEAF is or'ed into the flag
872  * the target is returned locked, otherwise it is returned unlocked.
873  * When creating or renaming and LOCKPARENT is specified, the target may not
874  * be ".".  When deleting and LOCKPARENT is specified, the target may be ".".
875  *
876  * Overall outline of lookup:
877  *
878  * dirloop:
879  *	identify next component of name at ndp->ni_ptr
880  *	handle degenerate case where name is null string
881  *	if .. and crossing mount points and on mounted filesys, find parent
882  *	call VOP_LOOKUP routine for next component name
883  *	    directory vnode returned in ni_dvp, unlocked unless LOCKPARENT set
884  *	    component vnode returned in ni_vp (if it exists), locked.
885  *	if result vnode is mounted on and crossing mount points,
886  *	    find mounted on vnode
887  *	if more components of name, do next level at dirloop
888  *	return the answer in ni_vp, locked if LOCKLEAF set
889  *	    if LOCKPARENT set, return locked parent in ni_dvp
890  *	    if WANTPARENT set, return unlocked parent in ni_dvp
891  */
892 int
893 vfs_lookup(struct nameidata *ndp)
894 {
895 	char *cp;			/* pointer into pathname argument */
896 	char *prev_ni_next;		/* saved ndp->ni_next */
897 	char *nulchar;			/* location of '\0' in cn_pnbuf */
898 	char *lastchar;			/* location of the last character */
899 	struct vnode *dp = NULL;	/* the directory we are searching */
900 	struct vnode *tdp;		/* saved dp */
901 	struct mount *mp;		/* mount table entry */
902 	struct prison *pr;
903 	size_t prev_ni_pathlen;		/* saved ndp->ni_pathlen */
904 	int docache;			/* == 0 do not cache last component */
905 	int wantparent;			/* 1 => wantparent or lockparent flag */
906 	int rdonly;			/* lookup read-only flag bit */
907 	int error = 0;
908 	int dpunlocked = 0;		/* dp has already been unlocked */
909 	int relookup = 0;		/* do not consume the path component */
910 	struct componentname *cnp = &ndp->ni_cnd;
911 	int lkflags_save;
912 	int ni_dvp_unlocked;
913 
914 	/*
915 	 * Setup: break out flag bits into variables.
916 	 */
917 	ni_dvp_unlocked = 0;
918 	wantparent = cnp->cn_flags & (LOCKPARENT | WANTPARENT);
919 	KASSERT(cnp->cn_nameiop == LOOKUP || wantparent,
920 	    ("CREATE, DELETE, RENAME require LOCKPARENT or WANTPARENT."));
921 	/*
922 	 * When set to zero, docache causes the last component of the
923 	 * pathname to be deleted from the cache and the full lookup
924 	 * of the name to be done (via VOP_CACHEDLOOKUP()). Often
925 	 * filesystems need some pre-computed values that are made
926 	 * during the full lookup, for instance UFS sets dp->i_offset.
927 	 *
928 	 * The docache variable is set to zero when requested by the
929 	 * NOCACHE flag and for all modifying operations except CREATE.
930 	 */
931 	docache = (cnp->cn_flags & NOCACHE) ^ NOCACHE;
932 	if (cnp->cn_nameiop == DELETE ||
933 	    (wantparent && cnp->cn_nameiop != CREATE &&
934 	     cnp->cn_nameiop != LOOKUP))
935 		docache = 0;
936 	rdonly = cnp->cn_flags & RDONLY;
937 	cnp->cn_flags &= ~ISSYMLINK;
938 	ndp->ni_dvp = NULL;
939 
940 	cnp->cn_lkflags = LK_SHARED;
941 	dp = ndp->ni_startdir;
942 	ndp->ni_startdir = NULLVP;
943 
944 	/*
945 	 * Leading slashes, if any, are supposed to be skipped by the caller.
946 	 */
947 	MPASS(cnp->cn_nameptr[0] != '/');
948 
949 	/*
950 	 * Check for degenerate name (e.g. / or "") which is a way of talking
951 	 * about a directory, e.g. like "/." or ".".
952 	 */
953 	if (__predict_false(cnp->cn_nameptr[0] == '\0')) {
954 		error = vfs_lookup_degenerate(ndp, dp, wantparent);
955 		if (error == 0)
956 			goto success_right_lock;
957 		goto bad_unlocked;
958 	}
959 
960 	/*
961 	 * Nul-out trailing slashes (e.g., "foo///" -> "foo").
962 	 *
963 	 * This must be done before VOP_LOOKUP() because some fs's don't know
964 	 * about trailing slashes.  Remember if there were trailing slashes to
965 	 * handle symlinks, existing non-directories and non-existing files
966 	 * that won't be directories specially later.
967 	 */
968 	MPASS(ndp->ni_pathlen >= 2);
969 	lastchar = &cnp->cn_nameptr[ndp->ni_pathlen - 2];
970 	if (*lastchar == '/') {
971 		while (lastchar >= cnp->cn_pnbuf) {
972 			*lastchar = '\0';
973 			lastchar--;
974 			ndp->ni_pathlen--;
975 			if (*lastchar != '/') {
976 				break;
977 			}
978 		}
979 		cnp->cn_flags |= TRAILINGSLASH;
980 	}
981 
982 	/*
983 	 * We use shared locks until we hit the parent of the last cn then
984 	 * we adjust based on the requesting flags.
985 	 */
986 	vn_lock(dp,
987 	    compute_cn_lkflags(dp->v_mount, cnp->cn_lkflags | LK_RETRY,
988 	    cnp->cn_flags));
989 
990 dirloop:
991 	/*
992 	 * Search a new directory.
993 	 *
994 	 * The last component of the filename is left accessible via
995 	 * cnp->cn_nameptr for callers that need the name. Callers needing
996 	 * the name set the SAVENAME flag. When done, they assume
997 	 * responsibility for freeing the pathname buffer.
998 	 *
999 	 * Store / as a temporary sentinel so that we only have one character
1000 	 * to test for. Pathnames tend to be short so this should not be
1001 	 * resulting in cache misses.
1002 	 */
1003 	nulchar = &cnp->cn_nameptr[ndp->ni_pathlen - 1];
1004 	KASSERT(*nulchar == '\0',
1005 	    ("%s: expected nul at %p; string [%s]\n", __func__, nulchar,
1006 	    cnp->cn_pnbuf));
1007 	*nulchar = '/';
1008 	for (cp = cnp->cn_nameptr; *cp != '/'; cp++) {
1009 		KASSERT(*cp != '\0',
1010 		    ("%s: encountered unexpected nul; string [%s]\n", __func__,
1011 		    cnp->cn_nameptr));
1012 		continue;
1013 	}
1014 	*nulchar = '\0';
1015 	cnp->cn_namelen = cp - cnp->cn_nameptr;
1016 	if (__predict_false(cnp->cn_namelen > NAME_MAX)) {
1017 		error = ENAMETOOLONG;
1018 		goto bad;
1019 	}
1020 #ifdef NAMEI_DIAGNOSTIC
1021 	{ char c = *cp;
1022 	*cp = '\0';
1023 	printf("{%s}: ", cnp->cn_nameptr);
1024 	*cp = c; }
1025 #endif
1026 	prev_ni_pathlen = ndp->ni_pathlen;
1027 	ndp->ni_pathlen -= cnp->cn_namelen;
1028 	KASSERT(ndp->ni_pathlen <= PATH_MAX,
1029 	    ("%s: ni_pathlen underflow to %zd\n", __func__, ndp->ni_pathlen));
1030 	prev_ni_next = ndp->ni_next;
1031 	ndp->ni_next = cp;
1032 
1033 	/*
1034 	 * Something else should be clearing this.
1035 	 */
1036 	cnp->cn_flags &= ~(ISDOTDOT|ISLASTCN);
1037 
1038 	cnp->cn_flags |= MAKEENTRY;
1039 	if (*cp == '\0' && docache == 0)
1040 		cnp->cn_flags &= ~MAKEENTRY;
1041 	if (cnp->cn_namelen == 2 &&
1042 	    cnp->cn_nameptr[1] == '.' && cnp->cn_nameptr[0] == '.')
1043 		cnp->cn_flags |= ISDOTDOT;
1044 	if (*ndp->ni_next == 0) {
1045 		cnp->cn_flags |= ISLASTCN;
1046 
1047 		if (__predict_false(cnp->cn_namelen == 1 && cnp->cn_nameptr[0] == '.' &&
1048 		    (cnp->cn_nameiop == DELETE || cnp->cn_nameiop == RENAME))) {
1049 			error = EINVAL;
1050 			goto bad;
1051 		}
1052 	}
1053 
1054 	nameicap_tracker_add(ndp, dp);
1055 
1056 	/*
1057 	 * Make sure degenerate names don't get here, their handling was
1058 	 * previously found in this spot.
1059 	 */
1060 	MPASS(cnp->cn_nameptr[0] != '\0');
1061 
1062 	/*
1063 	 * Handle "..": five special cases.
1064 	 * 0. If doing a capability lookup and lookup_cap_dotdot is
1065 	 *    disabled, return ENOTCAPABLE.
1066 	 * 1. Return an error if this is the last component of
1067 	 *    the name and the operation is DELETE or RENAME.
1068 	 * 2. If at root directory (e.g. after chroot)
1069 	 *    or at absolute root directory
1070 	 *    then ignore it so can't get out.
1071 	 * 3. If this vnode is the root of a mounted
1072 	 *    filesystem, then replace it with the
1073 	 *    vnode which was mounted on so we take the
1074 	 *    .. in the other filesystem.
1075 	 * 4. If the vnode is the top directory of
1076 	 *    the jail or chroot, don't let them out.
1077 	 * 5. If doing a capability lookup and lookup_cap_dotdot is
1078 	 *    enabled, return ENOTCAPABLE if the lookup would escape
1079 	 *    from the initial file descriptor directory.  Checks are
1080 	 *    done by ensuring that namei() already traversed the
1081 	 *    result of dotdot lookup.
1082 	 */
1083 	if (cnp->cn_flags & ISDOTDOT) {
1084 		if ((ndp->ni_lcf & (NI_LCF_STRICTRELATIVE | NI_LCF_CAP_DOTDOT))
1085 		    == NI_LCF_STRICTRELATIVE) {
1086 #ifdef KTRACE
1087 			if (KTRPOINT(curthread, KTR_CAPFAIL))
1088 				ktrcapfail(CAPFAIL_LOOKUP, NULL, NULL);
1089 #endif
1090 			error = ENOTCAPABLE;
1091 			goto bad;
1092 		}
1093 		if ((cnp->cn_flags & ISLASTCN) != 0 &&
1094 		    (cnp->cn_nameiop == DELETE || cnp->cn_nameiop == RENAME)) {
1095 			error = EINVAL;
1096 			goto bad;
1097 		}
1098 		for (;;) {
1099 			for (pr = cnp->cn_cred->cr_prison; pr != NULL;
1100 			     pr = pr->pr_parent)
1101 				if (dp == pr->pr_root)
1102 					break;
1103 			if (dp == ndp->ni_rootdir ||
1104 			    dp == ndp->ni_topdir ||
1105 			    dp == rootvnode ||
1106 			    pr != NULL ||
1107 			    ((dp->v_vflag & VV_ROOT) != 0 &&
1108 			     (cnp->cn_flags & NOCROSSMOUNT) != 0)) {
1109 				ndp->ni_dvp = dp;
1110 				ndp->ni_vp = dp;
1111 				VREF(dp);
1112 				goto nextname;
1113 			}
1114 			if ((dp->v_vflag & VV_ROOT) == 0)
1115 				break;
1116 			if (VN_IS_DOOMED(dp)) {	/* forced unmount */
1117 				error = ENOENT;
1118 				goto bad;
1119 			}
1120 			tdp = dp;
1121 			dp = dp->v_mount->mnt_vnodecovered;
1122 			VREF(dp);
1123 			vput(tdp);
1124 			vn_lock(dp,
1125 			    compute_cn_lkflags(dp->v_mount, cnp->cn_lkflags |
1126 			    LK_RETRY, ISDOTDOT));
1127 			error = nameicap_check_dotdot(ndp, dp);
1128 			if (error != 0) {
1129 #ifdef KTRACE
1130 				if (KTRPOINT(curthread, KTR_CAPFAIL))
1131 					ktrcapfail(CAPFAIL_LOOKUP, NULL, NULL);
1132 #endif
1133 				goto bad;
1134 			}
1135 		}
1136 	}
1137 
1138 	/*
1139 	 * We now have a segment name to search for, and a directory to search.
1140 	 */
1141 unionlookup:
1142 #ifdef MAC
1143 	error = mac_vnode_check_lookup(cnp->cn_cred, dp, cnp);
1144 	if (__predict_false(error))
1145 		goto bad;
1146 #endif
1147 	ndp->ni_dvp = dp;
1148 	ndp->ni_vp = NULL;
1149 	ASSERT_VOP_LOCKED(dp, "lookup");
1150 	/*
1151 	 * If we have a shared lock we may need to upgrade the lock for the
1152 	 * last operation.
1153 	 */
1154 	if ((cnp->cn_flags & LOCKPARENT) && (cnp->cn_flags & ISLASTCN) &&
1155 	    dp != vp_crossmp && VOP_ISLOCKED(dp) == LK_SHARED)
1156 		vn_lock(dp, LK_UPGRADE|LK_RETRY);
1157 	if (VN_IS_DOOMED(dp)) {
1158 		error = ENOENT;
1159 		goto bad;
1160 	}
1161 	/*
1162 	 * If we're looking up the last component and we need an exclusive
1163 	 * lock, adjust our lkflags.
1164 	 */
1165 	if (needs_exclusive_leaf(dp->v_mount, cnp->cn_flags))
1166 		cnp->cn_lkflags = LK_EXCLUSIVE;
1167 #ifdef NAMEI_DIAGNOSTIC
1168 	vn_printf(dp, "lookup in ");
1169 #endif
1170 	lkflags_save = cnp->cn_lkflags;
1171 	cnp->cn_lkflags = compute_cn_lkflags(dp->v_mount, cnp->cn_lkflags,
1172 	    cnp->cn_flags);
1173 	error = VOP_LOOKUP(dp, &ndp->ni_vp, cnp);
1174 	cnp->cn_lkflags = lkflags_save;
1175 	if (error != 0) {
1176 		KASSERT(ndp->ni_vp == NULL, ("leaf should be empty"));
1177 #ifdef NAMEI_DIAGNOSTIC
1178 		printf("not found\n");
1179 #endif
1180 		if ((error == ENOENT) &&
1181 		    (dp->v_vflag & VV_ROOT) && (dp->v_mount != NULL) &&
1182 		    (dp->v_mount->mnt_flag & MNT_UNION)) {
1183 			tdp = dp;
1184 			dp = dp->v_mount->mnt_vnodecovered;
1185 			VREF(dp);
1186 			vput(tdp);
1187 			vn_lock(dp,
1188 			    compute_cn_lkflags(dp->v_mount, cnp->cn_lkflags |
1189 			    LK_RETRY, cnp->cn_flags));
1190 			nameicap_tracker_add(ndp, dp);
1191 			goto unionlookup;
1192 		}
1193 
1194 		if (error == ERELOOKUP) {
1195 			vref(dp);
1196 			ndp->ni_vp = dp;
1197 			error = 0;
1198 			relookup = 1;
1199 			goto good;
1200 		}
1201 
1202 		if (error != EJUSTRETURN)
1203 			goto bad;
1204 		/*
1205 		 * At this point, we know we're at the end of the
1206 		 * pathname.  If creating / renaming, we can consider
1207 		 * allowing the file or directory to be created / renamed,
1208 		 * provided we're not on a read-only filesystem.
1209 		 */
1210 		if (rdonly) {
1211 			error = EROFS;
1212 			goto bad;
1213 		}
1214 		/* trailing slash only allowed for directories */
1215 		if ((cnp->cn_flags & TRAILINGSLASH) &&
1216 		    !(cnp->cn_flags & WILLBEDIR)) {
1217 			error = ENOENT;
1218 			goto bad;
1219 		}
1220 		if ((cnp->cn_flags & LOCKPARENT) == 0)
1221 			VOP_UNLOCK(dp);
1222 		/*
1223 		 * We return with ni_vp NULL to indicate that the entry
1224 		 * doesn't currently exist, leaving a pointer to the
1225 		 * (possibly locked) directory vnode in ndp->ni_dvp.
1226 		 */
1227 		if (cnp->cn_flags & SAVESTART) {
1228 			ndp->ni_startdir = ndp->ni_dvp;
1229 			VREF(ndp->ni_startdir);
1230 		}
1231 		goto success;
1232 	}
1233 
1234 good:
1235 #ifdef NAMEI_DIAGNOSTIC
1236 	printf("found\n");
1237 #endif
1238 	dp = ndp->ni_vp;
1239 
1240 	/*
1241 	 * Check to see if the vnode has been mounted on;
1242 	 * if so find the root of the mounted filesystem.
1243 	 */
1244 	while (dp->v_type == VDIR && (mp = dp->v_mountedhere) &&
1245 	       (cnp->cn_flags & NOCROSSMOUNT) == 0) {
1246 		if (vfs_busy(mp, 0))
1247 			continue;
1248 		vput(dp);
1249 		if (dp != ndp->ni_dvp)
1250 			vput(ndp->ni_dvp);
1251 		else
1252 			vrele(ndp->ni_dvp);
1253 		vrefact(vp_crossmp);
1254 		ndp->ni_dvp = vp_crossmp;
1255 		error = VFS_ROOT(mp, compute_cn_lkflags(mp, cnp->cn_lkflags,
1256 		    cnp->cn_flags), &tdp);
1257 		vfs_unbusy(mp);
1258 		if (vn_lock(vp_crossmp, LK_SHARED | LK_NOWAIT))
1259 			panic("vp_crossmp exclusively locked or reclaimed");
1260 		if (error) {
1261 			dpunlocked = 1;
1262 			goto bad2;
1263 		}
1264 		ndp->ni_vp = dp = tdp;
1265 	}
1266 
1267 	/*
1268 	 * Check for symbolic link
1269 	 */
1270 	if ((dp->v_type == VLNK) &&
1271 	    ((cnp->cn_flags & FOLLOW) || (cnp->cn_flags & TRAILINGSLASH) ||
1272 	     *ndp->ni_next == '/')) {
1273 		cnp->cn_flags |= ISSYMLINK;
1274 		if (VN_IS_DOOMED(dp)) {
1275 			/*
1276 			 * We can't know whether the directory was mounted with
1277 			 * NOSYMFOLLOW, so we can't follow safely.
1278 			 */
1279 			error = ENOENT;
1280 			goto bad2;
1281 		}
1282 		if (dp->v_mount->mnt_flag & MNT_NOSYMFOLLOW) {
1283 			error = EACCES;
1284 			goto bad2;
1285 		}
1286 		/*
1287 		 * Symlink code always expects an unlocked dvp.
1288 		 */
1289 		if (ndp->ni_dvp != ndp->ni_vp) {
1290 			VOP_UNLOCK(ndp->ni_dvp);
1291 			ni_dvp_unlocked = 1;
1292 		}
1293 		goto success;
1294 	}
1295 
1296 nextname:
1297 	/*
1298 	 * Not a symbolic link that we will follow.  Continue with the
1299 	 * next component if there is any; otherwise, we're done.
1300 	 */
1301 	KASSERT((cnp->cn_flags & ISLASTCN) || *ndp->ni_next == '/',
1302 	    ("lookup: invalid path state."));
1303 	if (relookup) {
1304 		relookup = 0;
1305 		ndp->ni_pathlen = prev_ni_pathlen;
1306 		ndp->ni_next = prev_ni_next;
1307 		if (ndp->ni_dvp != dp)
1308 			vput(ndp->ni_dvp);
1309 		else
1310 			vrele(ndp->ni_dvp);
1311 		goto dirloop;
1312 	}
1313 	if (cnp->cn_flags & ISDOTDOT) {
1314 		error = nameicap_check_dotdot(ndp, ndp->ni_vp);
1315 		if (error != 0) {
1316 #ifdef KTRACE
1317 			if (KTRPOINT(curthread, KTR_CAPFAIL))
1318 				ktrcapfail(CAPFAIL_LOOKUP, NULL, NULL);
1319 #endif
1320 			goto bad2;
1321 		}
1322 	}
1323 	if (*ndp->ni_next == '/') {
1324 		cnp->cn_nameptr = ndp->ni_next;
1325 		while (*cnp->cn_nameptr == '/') {
1326 			cnp->cn_nameptr++;
1327 			ndp->ni_pathlen--;
1328 		}
1329 		if (ndp->ni_dvp != dp)
1330 			vput(ndp->ni_dvp);
1331 		else
1332 			vrele(ndp->ni_dvp);
1333 		goto dirloop;
1334 	}
1335 	/*
1336 	 * If we're processing a path with a trailing slash,
1337 	 * check that the end result is a directory.
1338 	 */
1339 	if ((cnp->cn_flags & TRAILINGSLASH) && dp->v_type != VDIR) {
1340 		error = ENOTDIR;
1341 		goto bad2;
1342 	}
1343 	/*
1344 	 * Disallow directory write attempts on read-only filesystems.
1345 	 */
1346 	if (rdonly &&
1347 	    (cnp->cn_nameiop == DELETE || cnp->cn_nameiop == RENAME)) {
1348 		error = EROFS;
1349 		goto bad2;
1350 	}
1351 	if (cnp->cn_flags & SAVESTART) {
1352 		ndp->ni_startdir = ndp->ni_dvp;
1353 		VREF(ndp->ni_startdir);
1354 	}
1355 	if (!wantparent) {
1356 		ni_dvp_unlocked = 2;
1357 		if (ndp->ni_dvp != dp)
1358 			vput(ndp->ni_dvp);
1359 		else
1360 			vrele(ndp->ni_dvp);
1361 	} else if ((cnp->cn_flags & LOCKPARENT) == 0 && ndp->ni_dvp != dp) {
1362 		VOP_UNLOCK(ndp->ni_dvp);
1363 		ni_dvp_unlocked = 1;
1364 	}
1365 
1366 	if (cnp->cn_flags & AUDITVNODE1)
1367 		AUDIT_ARG_VNODE1(dp);
1368 	else if (cnp->cn_flags & AUDITVNODE2)
1369 		AUDIT_ARG_VNODE2(dp);
1370 
1371 	if ((cnp->cn_flags & LOCKLEAF) == 0)
1372 		VOP_UNLOCK(dp);
1373 success:
1374 	/*
1375 	 * FIXME: for lookups which only cross a mount point to fetch the
1376 	 * root vnode, ni_dvp will be set to vp_crossmp. This can be a problem
1377 	 * if either WANTPARENT or LOCKPARENT is set.
1378 	 */
1379 	/*
1380 	 * Because of shared lookup we may have the vnode shared locked, but
1381 	 * the caller may want it to be exclusively locked.
1382 	 */
1383 	if (needs_exclusive_leaf(dp->v_mount, cnp->cn_flags) &&
1384 	    VOP_ISLOCKED(dp) != LK_EXCLUSIVE) {
1385 		vn_lock(dp, LK_UPGRADE | LK_RETRY);
1386 		if (VN_IS_DOOMED(dp)) {
1387 			error = ENOENT;
1388 			goto bad2;
1389 		}
1390 	}
1391 success_right_lock:
1392 	if (ndp->ni_vp != NULL) {
1393 		if ((cnp->cn_flags & ISDOTDOT) == 0)
1394 			nameicap_tracker_add(ndp, ndp->ni_vp);
1395 		if ((cnp->cn_flags & (FAILIFEXISTS | ISSYMLINK)) == FAILIFEXISTS)
1396 			return (vfs_lookup_failifexists(ndp));
1397 	}
1398 	return (0);
1399 
1400 bad2:
1401 	if (ni_dvp_unlocked != 2) {
1402 		if (dp != ndp->ni_dvp && !ni_dvp_unlocked)
1403 			vput(ndp->ni_dvp);
1404 		else
1405 			vrele(ndp->ni_dvp);
1406 	}
1407 bad:
1408 	if (!dpunlocked)
1409 		vput(dp);
1410 bad_unlocked:
1411 	ndp->ni_vp = NULL;
1412 	return (error);
1413 }
1414 
1415 /*
1416  * relookup - lookup a path name component
1417  *    Used by lookup to re-acquire things.
1418  */
1419 int
1420 vfs_relookup(struct vnode *dvp, struct vnode **vpp, struct componentname *cnp)
1421 {
1422 	struct vnode *dp = NULL;		/* the directory we are searching */
1423 	int rdonly;			/* lookup read-only flag bit */
1424 	int error = 0;
1425 
1426 	KASSERT(cnp->cn_flags & ISLASTCN,
1427 	    ("relookup: Not given last component."));
1428 	/*
1429 	 * Setup: break out flag bits into variables.
1430 	 */
1431 	KASSERT((cnp->cn_flags & (LOCKPARENT | WANTPARENT)) != 0,
1432 	    ("relookup: parent not wanted"));
1433 	rdonly = cnp->cn_flags & RDONLY;
1434 	cnp->cn_flags &= ~ISSYMLINK;
1435 	dp = dvp;
1436 	cnp->cn_lkflags = LK_EXCLUSIVE;
1437 	vn_lock(dp, LK_EXCLUSIVE | LK_RETRY);
1438 
1439 	/*
1440 	 * Search a new directory.
1441 	 *
1442 	 * The last component of the filename is left accessible via
1443 	 * cnp->cn_nameptr for callers that need the name. Callers needing
1444 	 * the name set the SAVENAME flag. When done, they assume
1445 	 * responsibility for freeing the pathname buffer.
1446 	 */
1447 #ifdef NAMEI_DIAGNOSTIC
1448 	printf("{%s}: ", cnp->cn_nameptr);
1449 #endif
1450 
1451 	/*
1452 	 * Check for "" which represents the root directory after slash
1453 	 * removal.
1454 	 */
1455 	if (cnp->cn_nameptr[0] == '\0') {
1456 		/*
1457 		 * Support only LOOKUP for "/" because lookup()
1458 		 * can't succeed for CREATE, DELETE and RENAME.
1459 		 */
1460 		KASSERT(cnp->cn_nameiop == LOOKUP, ("nameiop must be LOOKUP"));
1461 		KASSERT(dp->v_type == VDIR, ("dp is not a directory"));
1462 
1463 		if (!(cnp->cn_flags & LOCKLEAF))
1464 			VOP_UNLOCK(dp);
1465 		*vpp = dp;
1466 		/* XXX This should probably move to the top of function. */
1467 		if (cnp->cn_flags & SAVESTART)
1468 			panic("lookup: SAVESTART");
1469 		return (0);
1470 	}
1471 
1472 	if (cnp->cn_flags & ISDOTDOT)
1473 		panic ("relookup: lookup on dot-dot");
1474 
1475 	/*
1476 	 * We now have a segment name to search for, and a directory to search.
1477 	 */
1478 #ifdef NAMEI_DIAGNOSTIC
1479 	vn_printf(dp, "search in ");
1480 #endif
1481 	if ((error = VOP_LOOKUP(dp, vpp, cnp)) != 0) {
1482 		KASSERT(*vpp == NULL, ("leaf should be empty"));
1483 		if (error != EJUSTRETURN)
1484 			goto bad;
1485 		/*
1486 		 * If creating and at end of pathname, then can consider
1487 		 * allowing file to be created.
1488 		 */
1489 		if (rdonly) {
1490 			error = EROFS;
1491 			goto bad;
1492 		}
1493 		/* ASSERT(dvp == ndp->ni_startdir) */
1494 		if (cnp->cn_flags & SAVESTART)
1495 			VREF(dvp);
1496 		if ((cnp->cn_flags & LOCKPARENT) == 0)
1497 			VOP_UNLOCK(dp);
1498 		/*
1499 		 * We return with ni_vp NULL to indicate that the entry
1500 		 * doesn't currently exist, leaving a pointer to the
1501 		 * (possibly locked) directory vnode in ndp->ni_dvp.
1502 		 */
1503 		return (0);
1504 	}
1505 
1506 	dp = *vpp;
1507 
1508 	/*
1509 	 * Disallow directory write attempts on read-only filesystems.
1510 	 */
1511 	if (rdonly &&
1512 	    (cnp->cn_nameiop == DELETE || cnp->cn_nameiop == RENAME)) {
1513 		if (dvp == dp)
1514 			vrele(dvp);
1515 		else
1516 			vput(dvp);
1517 		error = EROFS;
1518 		goto bad;
1519 	}
1520 	/*
1521 	 * Set the parent lock/ref state to the requested state.
1522 	 */
1523 	if ((cnp->cn_flags & LOCKPARENT) == 0 && dvp != dp)
1524 		VOP_UNLOCK(dvp);
1525 	/*
1526 	 * Check for symbolic link
1527 	 */
1528 	KASSERT(dp->v_type != VLNK || !(cnp->cn_flags & FOLLOW),
1529 	    ("relookup: symlink found.\n"));
1530 
1531 	/* ASSERT(dvp == ndp->ni_startdir) */
1532 	if (cnp->cn_flags & SAVESTART)
1533 		VREF(dvp);
1534 
1535 	if ((cnp->cn_flags & LOCKLEAF) == 0)
1536 		VOP_UNLOCK(dp);
1537 	return (0);
1538 bad:
1539 	vput(dp);
1540 	*vpp = NULL;
1541 	return (error);
1542 }
1543 
1544 /*
1545  * Free data allocated by namei(); see namei(9) for details.
1546  */
1547 void
1548 NDFREE_PNBUF(struct nameidata *ndp)
1549 {
1550 
1551 	if ((ndp->ni_cnd.cn_flags & HASBUF) != 0) {
1552 		MPASS((ndp->ni_cnd.cn_flags & (SAVENAME | SAVESTART)) != 0);
1553 		uma_zfree(namei_zone, ndp->ni_cnd.cn_pnbuf);
1554 		ndp->ni_cnd.cn_flags &= ~HASBUF;
1555 	}
1556 }
1557 
1558 /*
1559  * NDFREE_PNBUF replacement for callers that know there is no buffer.
1560  *
1561  * This is a hack. Preferably the VFS layer would not produce anything more
1562  * than it was asked to do. Unfortunately several non-LOOKUP cases can add the
1563  * HASBUF flag to the result. Even then an interface could be implemented where
1564  * the caller specifies what they expected to see in the result and what they
1565  * are going to take care of.
1566  *
1567  * In the meantime provide this kludge as a trivial replacement for NDFREE_PNBUF
1568  * calls scattered throughout the kernel where we know for a fact the flag must not
1569  * be seen.
1570  */
1571 #ifdef INVARIANTS
1572 void
1573 NDFREE_NOTHING(struct nameidata *ndp)
1574 {
1575 	struct componentname *cnp;
1576 
1577 	cnp = &ndp->ni_cnd;
1578 	KASSERT(cnp->cn_nameiop == LOOKUP, ("%s: got non-LOOKUP op %d\n",
1579 	    __func__, cnp->cn_nameiop));
1580 	KASSERT((cnp->cn_flags & (SAVENAME | HASBUF)) == 0,
1581 	    ("%s: bad flags \%" PRIx64 "\n", __func__, cnp->cn_flags));
1582 }
1583 #endif
1584 
1585 void
1586 (NDFREE)(struct nameidata *ndp, const u_int flags)
1587 {
1588 	int unlock_dvp;
1589 	int unlock_vp;
1590 
1591 	unlock_dvp = 0;
1592 	unlock_vp = 0;
1593 
1594 	if (!(flags & NDF_NO_FREE_PNBUF)) {
1595 		NDFREE_PNBUF(ndp);
1596 	}
1597 	if (!(flags & NDF_NO_VP_UNLOCK) &&
1598 	    (ndp->ni_cnd.cn_flags & LOCKLEAF) && ndp->ni_vp)
1599 		unlock_vp = 1;
1600 	if (!(flags & NDF_NO_DVP_UNLOCK) &&
1601 	    (ndp->ni_cnd.cn_flags & LOCKPARENT) &&
1602 	    ndp->ni_dvp != ndp->ni_vp)
1603 		unlock_dvp = 1;
1604 	if (!(flags & NDF_NO_VP_RELE) && ndp->ni_vp) {
1605 		if (unlock_vp) {
1606 			vput(ndp->ni_vp);
1607 			unlock_vp = 0;
1608 		} else
1609 			vrele(ndp->ni_vp);
1610 		ndp->ni_vp = NULL;
1611 	}
1612 	if (unlock_vp)
1613 		VOP_UNLOCK(ndp->ni_vp);
1614 	if (!(flags & NDF_NO_DVP_RELE) &&
1615 	    (ndp->ni_cnd.cn_flags & (LOCKPARENT|WANTPARENT))) {
1616 		if (unlock_dvp) {
1617 			vput(ndp->ni_dvp);
1618 			unlock_dvp = 0;
1619 		} else
1620 			vrele(ndp->ni_dvp);
1621 		ndp->ni_dvp = NULL;
1622 	}
1623 	if (unlock_dvp)
1624 		VOP_UNLOCK(ndp->ni_dvp);
1625 	if (!(flags & NDF_NO_STARTDIR_RELE) &&
1626 	    (ndp->ni_cnd.cn_flags & SAVESTART)) {
1627 		vrele(ndp->ni_startdir);
1628 		ndp->ni_startdir = NULL;
1629 	}
1630 }
1631 
1632 #ifdef INVARIANTS
1633 /*
1634  * Validate the final state of ndp after the lookup.
1635  *
1636  * Historically filesystems were allowed to modify cn_flags. Most notably they
1637  * can add SAVENAME to the request, resulting in HASBUF and pushing subsequent
1638  * clean up to the consumer. In practice this seems to only concern != LOOKUP
1639  * operations.
1640  *
1641  * As a step towards stricter API contract this routine validates the state to
1642  * clean up. Note validation is a work in progress with the intent of becoming
1643  * stricter over time.
1644  */
1645 #define NDMODIFYINGFLAGS (LOCKLEAF | LOCKPARENT | WANTPARENT | SAVENAME | SAVESTART | HASBUF)
1646 void
1647 NDVALIDATE(struct nameidata *ndp)
1648 {
1649 	struct componentname *cnp;
1650 	uint64_t used, orig;
1651 
1652 	cnp = &ndp->ni_cnd;
1653 	orig = cnp->cn_origflags;
1654 	used = cnp->cn_flags;
1655 	switch (cnp->cn_nameiop) {
1656 	case LOOKUP:
1657 		/*
1658 		 * For plain lookup we require strict conformance -- nothing
1659 		 * to clean up if it was not requested by the caller.
1660 		 */
1661 		orig &= NDMODIFYINGFLAGS;
1662 		used &= NDMODIFYINGFLAGS;
1663 		if ((orig & (SAVENAME | SAVESTART)) != 0)
1664 			orig |= HASBUF;
1665 		if (orig != used) {
1666 			goto out_mismatch;
1667 		}
1668 		break;
1669 	case CREATE:
1670 	case DELETE:
1671 	case RENAME:
1672 		/*
1673 		 * Some filesystems set SAVENAME to provoke HASBUF, accommodate
1674 		 * for it until it gets fixed.
1675 		 */
1676 		orig &= NDMODIFYINGFLAGS;
1677 		orig |= (SAVENAME | HASBUF);
1678 		used &= NDMODIFYINGFLAGS;
1679 		used |= (SAVENAME | HASBUF);
1680 		if (orig != used) {
1681 			goto out_mismatch;
1682 		}
1683 		break;
1684 	}
1685 	return;
1686 out_mismatch:
1687 	panic("%s: mismatched flags for op %d: added %" PRIx64 ", "
1688 	    "removed %" PRIx64" (%" PRIx64" != %" PRIx64"; stored %" PRIx64" != %" PRIx64")",
1689 	    __func__, cnp->cn_nameiop, used & ~orig, orig &~ used,
1690 	    orig, used, cnp->cn_origflags, cnp->cn_flags);
1691 }
1692 #endif
1693 
1694 /*
1695  * Determine if there is a suitable alternate filename under the specified
1696  * prefix for the specified path.  If the create flag is set, then the
1697  * alternate prefix will be used so long as the parent directory exists.
1698  * This is used by the various compatibility ABIs so that Linux binaries prefer
1699  * files under /compat/linux for example.  The chosen path (whether under
1700  * the prefix or under /) is returned in a kernel malloc'd buffer pointed
1701  * to by pathbuf.  The caller is responsible for free'ing the buffer from
1702  * the M_TEMP bucket if one is returned.
1703  */
1704 int
1705 kern_alternate_path(const char *prefix, const char *path, enum uio_seg pathseg,
1706     char **pathbuf, int create, int dirfd)
1707 {
1708 	struct nameidata nd, ndroot;
1709 	char *ptr, *buf, *cp;
1710 	size_t len, sz;
1711 	int error;
1712 
1713 	buf = (char *) malloc(MAXPATHLEN, M_TEMP, M_WAITOK);
1714 	*pathbuf = buf;
1715 
1716 	/* Copy the prefix into the new pathname as a starting point. */
1717 	len = strlcpy(buf, prefix, MAXPATHLEN);
1718 	if (len >= MAXPATHLEN) {
1719 		*pathbuf = NULL;
1720 		free(buf, M_TEMP);
1721 		return (EINVAL);
1722 	}
1723 	sz = MAXPATHLEN - len;
1724 	ptr = buf + len;
1725 
1726 	/* Append the filename to the prefix. */
1727 	if (pathseg == UIO_SYSSPACE)
1728 		error = copystr(path, ptr, sz, &len);
1729 	else
1730 		error = copyinstr(path, ptr, sz, &len);
1731 
1732 	if (error) {
1733 		*pathbuf = NULL;
1734 		free(buf, M_TEMP);
1735 		return (error);
1736 	}
1737 
1738 	/* Only use a prefix with absolute pathnames. */
1739 	if (*ptr != '/') {
1740 		error = EINVAL;
1741 		goto keeporig;
1742 	}
1743 
1744 	if (dirfd != AT_FDCWD) {
1745 		/*
1746 		 * We want the original because the "prefix" is
1747 		 * included in the already opened dirfd.
1748 		 */
1749 		bcopy(ptr, buf, len);
1750 		return (0);
1751 	}
1752 
1753 	/*
1754 	 * We know that there is a / somewhere in this pathname.
1755 	 * Search backwards for it, to find the file's parent dir
1756 	 * to see if it exists in the alternate tree. If it does,
1757 	 * and we want to create a file (cflag is set). We don't
1758 	 * need to worry about the root comparison in this case.
1759 	 */
1760 
1761 	if (create) {
1762 		for (cp = &ptr[len] - 1; *cp != '/'; cp--);
1763 		*cp = '\0';
1764 
1765 		NDINIT(&nd, LOOKUP, NOFOLLOW, UIO_SYSSPACE, buf);
1766 		error = namei(&nd);
1767 		*cp = '/';
1768 		if (error != 0)
1769 			goto keeporig;
1770 	} else {
1771 		NDINIT(&nd, LOOKUP, NOFOLLOW, UIO_SYSSPACE, buf);
1772 
1773 		error = namei(&nd);
1774 		if (error != 0)
1775 			goto keeporig;
1776 
1777 		/*
1778 		 * We now compare the vnode of the prefix to the one
1779 		 * vnode asked. If they resolve to be the same, then we
1780 		 * ignore the match so that the real root gets used.
1781 		 * This avoids the problem of traversing "../.." to find the
1782 		 * root directory and never finding it, because "/" resolves
1783 		 * to the emulation root directory. This is expensive :-(
1784 		 */
1785 		NDINIT(&ndroot, LOOKUP, FOLLOW, UIO_SYSSPACE, prefix);
1786 
1787 		/* We shouldn't ever get an error from this namei(). */
1788 		error = namei(&ndroot);
1789 		if (error == 0) {
1790 			if (nd.ni_vp == ndroot.ni_vp)
1791 				error = ENOENT;
1792 
1793 			NDFREE_PNBUF(&ndroot);
1794 			vrele(ndroot.ni_vp);
1795 		}
1796 	}
1797 
1798 	NDFREE_PNBUF(&nd);
1799 	vrele(nd.ni_vp);
1800 
1801 keeporig:
1802 	/* If there was an error, use the original path name. */
1803 	if (error)
1804 		bcopy(ptr, buf, len);
1805 	return (error);
1806 }
1807