xref: /freebsd/sys/kern/vfs_lookup.c (revision 8a0edc914ffdda876987add5128da3ee236a6a12)
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_DEFINE3(vfs, namei, lookup, return, "int", "struct vnode *", "bool");
80 
81 /* Allocation zone for namei. */
82 uma_zone_t namei_zone;
83 
84 /* Placeholder vnode for mp traversal. */
85 static struct vnode *vp_crossmp;
86 
87 static int
88 crossmp_vop_islocked(struct vop_islocked_args *ap)
89 {
90 
91 	return (LK_SHARED);
92 }
93 
94 static int
95 crossmp_vop_lock1(struct vop_lock1_args *ap)
96 {
97 	struct vnode *vp;
98 	struct lock *lk __unused;
99 	const char *file __unused;
100 	int flags, line __unused;
101 
102 	vp = ap->a_vp;
103 	lk = vp->v_vnlock;
104 	flags = ap->a_flags;
105 	file = ap->a_file;
106 	line = ap->a_line;
107 
108 	if ((flags & LK_SHARED) == 0)
109 		panic("invalid lock request for crossmp");
110 
111 	WITNESS_CHECKORDER(&lk->lock_object, LOP_NEWORDER, file, line,
112 	    flags & LK_INTERLOCK ? &VI_MTX(vp)->lock_object : NULL);
113 	WITNESS_LOCK(&lk->lock_object, 0, file, line);
114 	if ((flags & LK_INTERLOCK) != 0)
115 		VI_UNLOCK(vp);
116 	LOCK_LOG_LOCK("SLOCK", &lk->lock_object, 0, 0, ap->a_file, line);
117 	return (0);
118 }
119 
120 static int
121 crossmp_vop_unlock(struct vop_unlock_args *ap)
122 {
123 	struct vnode *vp;
124 	struct lock *lk __unused;
125 
126 	vp = ap->a_vp;
127 	lk = vp->v_vnlock;
128 
129 	WITNESS_UNLOCK(&lk->lock_object, 0, LOCK_FILE, LOCK_LINE);
130 	LOCK_LOG_LOCK("SUNLOCK", &lk->lock_object, 0, 0, LOCK_FILE,
131 	    LOCK_LINE);
132 	return (0);
133 }
134 
135 static struct vop_vector crossmp_vnodeops = {
136 	.vop_default =		&default_vnodeops,
137 	.vop_islocked =		crossmp_vop_islocked,
138 	.vop_lock1 =		crossmp_vop_lock1,
139 	.vop_unlock =		crossmp_vop_unlock,
140 };
141 /*
142  * VFS_VOP_VECTOR_REGISTER(crossmp_vnodeops) is not used here since the vnode
143  * gets allocated early. See nameiinit for the direct call below.
144  */
145 
146 struct nameicap_tracker {
147 	struct vnode *dp;
148 	TAILQ_ENTRY(nameicap_tracker) nm_link;
149 };
150 
151 /* Zone for cap mode tracker elements used for dotdot capability checks. */
152 static uma_zone_t nt_zone;
153 
154 static void
155 nameiinit(void *dummy __unused)
156 {
157 
158 	namei_zone = uma_zcreate("NAMEI", MAXPATHLEN, NULL, NULL, NULL, NULL,
159 	    UMA_ALIGN_PTR, 0);
160 	nt_zone = uma_zcreate("rentr", sizeof(struct nameicap_tracker),
161 	    NULL, NULL, NULL, NULL, 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 	if ((ndp->ni_lcf & (NI_LCF_BENEATH_ABS | NI_LCF_BENEATH_LATCHED)) ==
185 	    NI_LCF_BENEATH_ABS) {
186 		MPASS((ndp->ni_lcf & NI_LCF_LATCH) != 0);
187 		if (dp != ndp->ni_beneath_latch)
188 			return;
189 		ndp->ni_lcf |= NI_LCF_BENEATH_LATCHED;
190 	}
191 	nt = uma_zalloc(nt_zone, M_WAITOK);
192 	vhold(dp);
193 	nt->dp = dp;
194 	TAILQ_INSERT_TAIL(&ndp->ni_cap_tracker, nt, nm_link);
195 }
196 
197 static void
198 nameicap_cleanup(struct nameidata *ndp, bool clean_latch)
199 {
200 	struct nameicap_tracker *nt, *nt1;
201 
202 	KASSERT(TAILQ_EMPTY(&ndp->ni_cap_tracker) ||
203 	    (ndp->ni_lcf & NI_LCF_CAP_DOTDOT) != 0, ("not strictrelative"));
204 	TAILQ_FOREACH_SAFE(nt, &ndp->ni_cap_tracker, nm_link, nt1) {
205 		TAILQ_REMOVE(&ndp->ni_cap_tracker, nt, nm_link);
206 		vdrop(nt->dp);
207 		uma_zfree(nt_zone, nt);
208 	}
209 	if (clean_latch && (ndp->ni_lcf & NI_LCF_LATCH) != 0) {
210 		ndp->ni_lcf &= ~NI_LCF_LATCH;
211 		vrele(ndp->ni_beneath_latch);
212 	}
213 }
214 
215 /*
216  * For dotdot lookups in capability mode, only allow the component
217  * lookup to succeed if the resulting directory was already traversed
218  * during the operation.  Also fail dotdot lookups for non-local
219  * filesystems, where external agents might assist local lookups to
220  * escape the compartment.
221  */
222 static int
223 nameicap_check_dotdot(struct nameidata *ndp, struct vnode *dp)
224 {
225 	struct nameicap_tracker *nt;
226 	struct mount *mp;
227 
228 	if ((ndp->ni_lcf & NI_LCF_CAP_DOTDOT) == 0 || dp == NULL ||
229 	    dp->v_type != VDIR)
230 		return (0);
231 	mp = dp->v_mount;
232 	if (lookup_cap_dotdot_nonlocal == 0 && mp != NULL &&
233 	    (mp->mnt_flag & MNT_LOCAL) == 0)
234 		return (ENOTCAPABLE);
235 	TAILQ_FOREACH_REVERSE(nt, &ndp->ni_cap_tracker, nameicap_tracker_head,
236 	    nm_link) {
237 		if (dp == nt->dp)
238 			return (0);
239 	}
240 	if ((ndp->ni_lcf & NI_LCF_BENEATH_ABS) != 0) {
241 		ndp->ni_lcf &= ~NI_LCF_BENEATH_LATCHED;
242 		nameicap_cleanup(ndp, false);
243 		return (0);
244 	}
245 	return (ENOTCAPABLE);
246 }
247 
248 static void
249 namei_cleanup_cnp(struct componentname *cnp)
250 {
251 
252 	uma_zfree(namei_zone, cnp->cn_pnbuf);
253 #ifdef DIAGNOSTIC
254 	cnp->cn_pnbuf = NULL;
255 	cnp->cn_nameptr = NULL;
256 #endif
257 }
258 
259 static int
260 namei_handle_root(struct nameidata *ndp, struct vnode **dpp)
261 {
262 	struct componentname *cnp;
263 
264 	cnp = &ndp->ni_cnd;
265 	if ((ndp->ni_lcf & NI_LCF_STRICTRELATIVE) != 0) {
266 #ifdef KTRACE
267 		if (KTRPOINT(curthread, KTR_CAPFAIL))
268 			ktrcapfail(CAPFAIL_LOOKUP, NULL, NULL);
269 #endif
270 		return (ENOTCAPABLE);
271 	}
272 	if ((cnp->cn_flags & BENEATH) != 0) {
273 		ndp->ni_lcf |= NI_LCF_BENEATH_ABS;
274 		ndp->ni_lcf &= ~NI_LCF_BENEATH_LATCHED;
275 		nameicap_cleanup(ndp, false);
276 	}
277 	while (*(cnp->cn_nameptr) == '/') {
278 		cnp->cn_nameptr++;
279 		ndp->ni_pathlen--;
280 	}
281 	*dpp = ndp->ni_rootdir;
282 	vrefact(*dpp);
283 	return (0);
284 }
285 
286 static int
287 namei_setup(struct nameidata *ndp, struct vnode **dpp, struct pwd **pwdp)
288 {
289 	struct componentname *cnp;
290 	struct file *dfp;
291 	struct thread *td;
292 	struct pwd *pwd;
293 	cap_rights_t rights;
294 	struct filecaps dirfd_caps;
295 	int error;
296 	bool startdir_used;
297 
298 	cnp = &ndp->ni_cnd;
299 	td = cnp->cn_thread;
300 
301 	startdir_used = false;
302 	*pwdp = NULL;
303 	*dpp = NULL;
304 
305 #ifdef CAPABILITY_MODE
306 	/*
307 	 * In capability mode, lookups must be restricted to happen in
308 	 * the subtree with the root specified by the file descriptor:
309 	 * - The root must be real file descriptor, not the pseudo-descriptor
310 	 *   AT_FDCWD.
311 	 * - The passed path must be relative and not absolute.
312 	 * - If lookup_cap_dotdot is disabled, path must not contain the
313 	 *   '..' components.
314 	 * - If lookup_cap_dotdot is enabled, we verify that all '..'
315 	 *   components lookups result in the directories which were
316 	 *   previously walked by us, which prevents an escape from
317 	 *   the relative root.
318 	 */
319 	if (IN_CAPABILITY_MODE(td) && (cnp->cn_flags & NOCAPCHECK) == 0) {
320 		ndp->ni_lcf |= NI_LCF_STRICTRELATIVE;
321 		if (ndp->ni_dirfd == AT_FDCWD) {
322 #ifdef KTRACE
323 			if (KTRPOINT(td, KTR_CAPFAIL))
324 				ktrcapfail(CAPFAIL_LOOKUP, NULL, NULL);
325 #endif
326 			return (ECAPMODE);
327 		}
328 	}
329 #endif
330 	error = 0;
331 
332 	/*
333 	 * Get starting point for the translation.
334 	 */
335 	pwd = pwd_hold(td);
336 	/*
337 	 * The reference on ni_rootdir is acquired in the block below to avoid
338 	 * back-to-back atomics for absolute lookups.
339 	 */
340 	ndp->ni_rootdir = pwd->pwd_rdir;
341 	ndp->ni_topdir = pwd->pwd_jdir;
342 
343 	if (cnp->cn_pnbuf[0] == '/') {
344 		ndp->ni_resflags |= NIRES_ABS;
345 		error = namei_handle_root(ndp, dpp);
346 	} else {
347 		if (ndp->ni_startdir != NULL) {
348 			*dpp = ndp->ni_startdir;
349 			startdir_used = true;
350 		} else if (ndp->ni_dirfd == AT_FDCWD) {
351 			*dpp = pwd->pwd_cdir;
352 			vrefact(*dpp);
353 		} else {
354 			rights = *ndp->ni_rightsneeded;
355 			cap_rights_set_one(&rights, CAP_LOOKUP);
356 
357 			if (cnp->cn_flags & AUDITVNODE1)
358 				AUDIT_ARG_ATFD1(ndp->ni_dirfd);
359 			if (cnp->cn_flags & AUDITVNODE2)
360 				AUDIT_ARG_ATFD2(ndp->ni_dirfd);
361 			/*
362 			 * Effectively inlined fgetvp_rights, because we need to
363 			 * inspect the file as well as grabbing the vnode.
364 			 */
365 			error = fget_cap(td, ndp->ni_dirfd, &rights,
366 			    &dfp, &ndp->ni_filecaps);
367 			if (error != 0) {
368 				/*
369 				 * Preserve the error; it should either be EBADF
370 				 * or capability-related, both of which can be
371 				 * safely returned to the caller.
372 				 */
373 			} else {
374 				if (dfp->f_ops == &badfileops) {
375 					error = EBADF;
376 				} else if (dfp->f_vnode == NULL) {
377 					error = ENOTDIR;
378 				} else {
379 					*dpp = dfp->f_vnode;
380 					vrefact(*dpp);
381 
382 					if ((dfp->f_flag & FSEARCH) != 0)
383 						cnp->cn_flags |= NOEXECCHECK;
384 				}
385 				fdrop(dfp, td);
386 			}
387 #ifdef CAPABILITIES
388 			/*
389 			 * If file descriptor doesn't have all rights,
390 			 * all lookups relative to it must also be
391 			 * strictly relative.
392 			 */
393 			CAP_ALL(&rights);
394 			if (!cap_rights_contains(&ndp->ni_filecaps.fc_rights,
395 			    &rights) ||
396 			    ndp->ni_filecaps.fc_fcntls != CAP_FCNTL_ALL ||
397 			    ndp->ni_filecaps.fc_nioctls != -1) {
398 				ndp->ni_lcf |= NI_LCF_STRICTRELATIVE;
399 			}
400 #endif
401 		}
402 		if (error == 0 && (*dpp)->v_type != VDIR)
403 			error = ENOTDIR;
404 	}
405 	if (error == 0 && (cnp->cn_flags & BENEATH) != 0) {
406 		if (ndp->ni_dirfd == AT_FDCWD) {
407 			ndp->ni_beneath_latch = pwd->pwd_cdir;
408 			vrefact(ndp->ni_beneath_latch);
409 		} else {
410 			rights = *ndp->ni_rightsneeded;
411 			cap_rights_set_one(&rights, CAP_LOOKUP);
412 			error = fgetvp_rights(td, ndp->ni_dirfd, &rights,
413 			    &dirfd_caps, &ndp->ni_beneath_latch);
414 			if (error == 0 && (*dpp)->v_type != VDIR) {
415 				vrele(ndp->ni_beneath_latch);
416 				error = ENOTDIR;
417 			}
418 		}
419 		if (error == 0)
420 			ndp->ni_lcf |= NI_LCF_LATCH;
421 	}
422 	/*
423 	 * If we are auditing the kernel pathname, save the user pathname.
424 	 */
425 	if (cnp->cn_flags & AUDITVNODE1)
426 		AUDIT_ARG_UPATH1_VP(td, ndp->ni_rootdir, *dpp, cnp->cn_pnbuf);
427 	if (cnp->cn_flags & AUDITVNODE2)
428 		AUDIT_ARG_UPATH2_VP(td, ndp->ni_rootdir, *dpp, cnp->cn_pnbuf);
429 	if (ndp->ni_startdir != NULL && !startdir_used)
430 		vrele(ndp->ni_startdir);
431 	if (error != 0) {
432 		if (*dpp != NULL)
433 			vrele(*dpp);
434 		pwd_drop(pwd);
435 		return (error);
436 	}
437 	MPASS((ndp->ni_lcf & (NI_LCF_BENEATH_ABS | NI_LCF_LATCH)) !=
438 	    NI_LCF_BENEATH_ABS);
439 	if (((ndp->ni_lcf & NI_LCF_STRICTRELATIVE) != 0 &&
440 	    lookup_cap_dotdot != 0) ||
441 	    ((ndp->ni_lcf & NI_LCF_STRICTRELATIVE) == 0 &&
442 	    (cnp->cn_flags & BENEATH) != 0))
443 		ndp->ni_lcf |= NI_LCF_CAP_DOTDOT;
444 	SDT_PROBE4(vfs, namei, lookup, entry, *dpp, cnp->cn_pnbuf,
445 	    cnp->cn_flags, false);
446 	*pwdp = pwd;
447 	return (0);
448 }
449 
450 /*
451  * Convert a pathname into a pointer to a locked vnode.
452  *
453  * The FOLLOW flag is set when symbolic links are to be followed
454  * when they occur at the end of the name translation process.
455  * Symbolic links are always followed for all other pathname
456  * components other than the last.
457  *
458  * The segflg defines whether the name is to be copied from user
459  * space or kernel space.
460  *
461  * Overall outline of namei:
462  *
463  *	copy in name
464  *	get starting directory
465  *	while (!done && !error) {
466  *		call lookup to search path.
467  *		if symbolic link, massage name in buffer and continue
468  *	}
469  */
470 int
471 namei(struct nameidata *ndp)
472 {
473 	char *cp;		/* pointer into pathname argument */
474 	struct vnode *dp;	/* the directory we are searching */
475 	struct iovec aiov;		/* uio for reading symbolic links */
476 	struct componentname *cnp;
477 	struct thread *td;
478 	struct pwd *pwd;
479 	struct uio auio;
480 	int error, linklen;
481 	enum cache_fpl_status status;
482 
483 	cnp = &ndp->ni_cnd;
484 	td = cnp->cn_thread;
485 	ndp->ni_cnd.cn_cred = ndp->ni_cnd.cn_thread->td_ucred;
486 	KASSERT(cnp->cn_cred && td->td_proc, ("namei: bad cred/proc"));
487 	KASSERT((cnp->cn_nameiop & (~OPMASK)) == 0,
488 	    ("namei: nameiop contaminated with flags"));
489 	KASSERT((cnp->cn_flags & OPMASK) == 0,
490 	    ("namei: flags contaminated with nameiops"));
491 	KASSERT((cnp->cn_flags & NAMEI_INTERNAL_FLAGS) == 0,
492 	    ("namei: unexpected flags: %" PRIx64 "\n",
493 	    cnp->cn_flags & NAMEI_INTERNAL_FLAGS));
494 	if (cnp->cn_flags & NOCACHE)
495 		KASSERT(cnp->cn_nameiop != LOOKUP,
496 		    ("%s: NOCACHE passed with LOOKUP", __func__));
497 	MPASS(ndp->ni_startdir == NULL || ndp->ni_startdir->v_type == VDIR ||
498 	    ndp->ni_startdir->v_type == VBAD);
499 
500 	ndp->ni_lcf = 0;
501 	ndp->ni_vp = NULL;
502 
503 	/*
504 	 * Get a buffer for the name to be translated, and copy the
505 	 * name into the buffer.
506 	 */
507 	if ((cnp->cn_flags & HASBUF) == 0)
508 		cnp->cn_pnbuf = uma_zalloc(namei_zone, M_WAITOK);
509 	if (ndp->ni_segflg == UIO_SYSSPACE)
510 		error = copystr(ndp->ni_dirp, cnp->cn_pnbuf, MAXPATHLEN,
511 		    &ndp->ni_pathlen);
512 	else
513 		error = copyinstr(ndp->ni_dirp, cnp->cn_pnbuf, MAXPATHLEN,
514 		    &ndp->ni_pathlen);
515 
516 	if (__predict_false(error != 0)) {
517 		namei_cleanup_cnp(cnp);
518 		return (error);
519 	}
520 
521 	/*
522 	 * Don't allow empty pathnames.
523 	 */
524 	if (__predict_false(*cnp->cn_pnbuf == '\0')) {
525 		namei_cleanup_cnp(cnp);
526 		return (ENOENT);
527 	}
528 
529 #ifdef KTRACE
530 	if (KTRPOINT(td, KTR_NAMEI)) {
531 		KASSERT(cnp->cn_thread == curthread,
532 		    ("namei not using curthread"));
533 		ktrnamei(cnp->cn_pnbuf);
534 	}
535 #endif
536 
537 	cnp->cn_nameptr = cnp->cn_pnbuf;
538 
539 	/*
540 	 * First try looking up the target without locking any vnodes.
541 	 *
542 	 * We may need to start from scratch or pick up where it left off.
543 	 */
544 	error = cache_fplookup(ndp, &status, &pwd);
545 	switch (status) {
546 	case CACHE_FPL_STATUS_UNSET:
547 		__assert_unreachable();
548 		break;
549 	case CACHE_FPL_STATUS_HANDLED:
550 		return (error);
551 	case CACHE_FPL_STATUS_PARTIAL:
552 		TAILQ_INIT(&ndp->ni_cap_tracker);
553 		dp = ndp->ni_startdir;
554 		break;
555 	case CACHE_FPL_STATUS_ABORTED:
556 		TAILQ_INIT(&ndp->ni_cap_tracker);
557 		error = namei_setup(ndp, &dp, &pwd);
558 		if (error != 0) {
559 			namei_cleanup_cnp(cnp);
560 			return (error);
561 		}
562 		break;
563 	}
564 
565 	ndp->ni_loopcnt = 0;
566 
567 	/*
568 	 * Locked lookup.
569 	 */
570 	for (;;) {
571 		ndp->ni_startdir = dp;
572 		error = lookup(ndp);
573 		if (error != 0)
574 			goto out;
575 		/*
576 		 * If not a symbolic link, we're done.
577 		 */
578 		if ((cnp->cn_flags & ISSYMLINK) == 0) {
579 			if ((cnp->cn_flags & (SAVENAME | SAVESTART)) == 0) {
580 				namei_cleanup_cnp(cnp);
581 			} else
582 				cnp->cn_flags |= HASBUF;
583 			if ((ndp->ni_lcf & (NI_LCF_BENEATH_ABS |
584 			    NI_LCF_BENEATH_LATCHED)) == NI_LCF_BENEATH_ABS) {
585 				NDFREE(ndp, 0);
586 				error = ENOTCAPABLE;
587 			}
588 			nameicap_cleanup(ndp, true);
589 			SDT_PROBE3(vfs, namei, lookup, return, error,
590 			    (error == 0 ? ndp->ni_vp : NULL), false);
591 			pwd_drop(pwd);
592 			return (error);
593 		}
594 		if (ndp->ni_loopcnt++ >= MAXSYMLINKS) {
595 			error = ELOOP;
596 			break;
597 		}
598 #ifdef MAC
599 		if ((cnp->cn_flags & NOMACCHECK) == 0) {
600 			error = mac_vnode_check_readlink(td->td_ucred,
601 			    ndp->ni_vp);
602 			if (error != 0)
603 				break;
604 		}
605 #endif
606 		if (ndp->ni_pathlen > 1)
607 			cp = uma_zalloc(namei_zone, M_WAITOK);
608 		else
609 			cp = cnp->cn_pnbuf;
610 		aiov.iov_base = cp;
611 		aiov.iov_len = MAXPATHLEN;
612 		auio.uio_iov = &aiov;
613 		auio.uio_iovcnt = 1;
614 		auio.uio_offset = 0;
615 		auio.uio_rw = UIO_READ;
616 		auio.uio_segflg = UIO_SYSSPACE;
617 		auio.uio_td = td;
618 		auio.uio_resid = MAXPATHLEN;
619 		error = VOP_READLINK(ndp->ni_vp, &auio, cnp->cn_cred);
620 		if (error != 0) {
621 			if (ndp->ni_pathlen > 1)
622 				uma_zfree(namei_zone, cp);
623 			break;
624 		}
625 		linklen = MAXPATHLEN - auio.uio_resid;
626 		if (linklen == 0) {
627 			if (ndp->ni_pathlen > 1)
628 				uma_zfree(namei_zone, cp);
629 			error = ENOENT;
630 			break;
631 		}
632 		if (linklen + ndp->ni_pathlen > MAXPATHLEN) {
633 			if (ndp->ni_pathlen > 1)
634 				uma_zfree(namei_zone, cp);
635 			error = ENAMETOOLONG;
636 			break;
637 		}
638 		if (ndp->ni_pathlen > 1) {
639 			bcopy(ndp->ni_next, cp + linklen, ndp->ni_pathlen);
640 			uma_zfree(namei_zone, cnp->cn_pnbuf);
641 			cnp->cn_pnbuf = cp;
642 		} else
643 			cnp->cn_pnbuf[linklen] = '\0';
644 		ndp->ni_pathlen += linklen;
645 		vput(ndp->ni_vp);
646 		dp = ndp->ni_dvp;
647 		/*
648 		 * Check if root directory should replace current directory.
649 		 */
650 		cnp->cn_nameptr = cnp->cn_pnbuf;
651 		if (*(cnp->cn_nameptr) == '/') {
652 			vrele(dp);
653 			error = namei_handle_root(ndp, &dp);
654 			if (error != 0)
655 				goto out;
656 		}
657 	}
658 	vput(ndp->ni_vp);
659 	ndp->ni_vp = NULL;
660 	vrele(ndp->ni_dvp);
661 out:
662 	MPASS(error != 0);
663 	namei_cleanup_cnp(cnp);
664 	nameicap_cleanup(ndp, true);
665 	SDT_PROBE3(vfs, namei, lookup, return, error, NULL, false);
666 	pwd_drop(pwd);
667 	return (error);
668 }
669 
670 static int
671 compute_cn_lkflags(struct mount *mp, int lkflags, int cnflags)
672 {
673 
674 	if (mp == NULL || ((lkflags & LK_SHARED) &&
675 	    (!(mp->mnt_kern_flag & MNTK_LOOKUP_SHARED) ||
676 	    ((cnflags & ISDOTDOT) &&
677 	    (mp->mnt_kern_flag & MNTK_LOOKUP_EXCL_DOTDOT))))) {
678 		lkflags &= ~LK_SHARED;
679 		lkflags |= LK_EXCLUSIVE;
680 	}
681 	lkflags |= LK_NODDLKTREAT;
682 	return (lkflags);
683 }
684 
685 static __inline int
686 needs_exclusive_leaf(struct mount *mp, int flags)
687 {
688 
689 	/*
690 	 * Intermediate nodes can use shared locks, we only need to
691 	 * force an exclusive lock for leaf nodes.
692 	 */
693 	if ((flags & (ISLASTCN | LOCKLEAF)) != (ISLASTCN | LOCKLEAF))
694 		return (0);
695 
696 	/* Always use exclusive locks if LOCKSHARED isn't set. */
697 	if (!(flags & LOCKSHARED))
698 		return (1);
699 
700 	/*
701 	 * For lookups during open(), if the mount point supports
702 	 * extended shared operations, then use a shared lock for the
703 	 * leaf node, otherwise use an exclusive lock.
704 	 */
705 	if ((flags & ISOPEN) != 0)
706 		return (!MNT_EXTENDED_SHARED(mp));
707 
708 	/*
709 	 * Lookup requests outside of open() that specify LOCKSHARED
710 	 * only need a shared lock on the leaf vnode.
711 	 */
712 	return (0);
713 }
714 
715 /*
716  * Search a pathname.
717  * This is a very central and rather complicated routine.
718  *
719  * The pathname is pointed to by ni_ptr and is of length ni_pathlen.
720  * The starting directory is taken from ni_startdir. The pathname is
721  * descended until done, or a symbolic link is encountered. The variable
722  * ni_more is clear if the path is completed; it is set to one if a
723  * symbolic link needing interpretation is encountered.
724  *
725  * The flag argument is LOOKUP, CREATE, RENAME, or DELETE depending on
726  * whether the name is to be looked up, created, renamed, or deleted.
727  * When CREATE, RENAME, or DELETE is specified, information usable in
728  * creating, renaming, or deleting a directory entry may be calculated.
729  * If flag has LOCKPARENT or'ed into it, the parent directory is returned
730  * locked. If flag has WANTPARENT or'ed into it, the parent directory is
731  * returned unlocked. Otherwise the parent directory is not returned. If
732  * the target of the pathname exists and LOCKLEAF is or'ed into the flag
733  * the target is returned locked, otherwise it is returned unlocked.
734  * When creating or renaming and LOCKPARENT is specified, the target may not
735  * be ".".  When deleting and LOCKPARENT is specified, the target may be ".".
736  *
737  * Overall outline of lookup:
738  *
739  * dirloop:
740  *	identify next component of name at ndp->ni_ptr
741  *	handle degenerate case where name is null string
742  *	if .. and crossing mount points and on mounted filesys, find parent
743  *	call VOP_LOOKUP routine for next component name
744  *	    directory vnode returned in ni_dvp, unlocked unless LOCKPARENT set
745  *	    component vnode returned in ni_vp (if it exists), locked.
746  *	if result vnode is mounted on and crossing mount points,
747  *	    find mounted on vnode
748  *	if more components of name, do next level at dirloop
749  *	return the answer in ni_vp, locked if LOCKLEAF set
750  *	    if LOCKPARENT set, return locked parent in ni_dvp
751  *	    if WANTPARENT set, return unlocked parent in ni_dvp
752  */
753 int
754 lookup(struct nameidata *ndp)
755 {
756 	char *cp;			/* pointer into pathname argument */
757 	char *prev_ni_next;		/* saved ndp->ni_next */
758 	struct vnode *dp = NULL;	/* the directory we are searching */
759 	struct vnode *tdp;		/* saved dp */
760 	struct mount *mp;		/* mount table entry */
761 	struct prison *pr;
762 	size_t prev_ni_pathlen;		/* saved ndp->ni_pathlen */
763 	int docache;			/* == 0 do not cache last component */
764 	int wantparent;			/* 1 => wantparent or lockparent flag */
765 	int rdonly;			/* lookup read-only flag bit */
766 	int error = 0;
767 	int dpunlocked = 0;		/* dp has already been unlocked */
768 	int relookup = 0;		/* do not consume the path component */
769 	struct componentname *cnp = &ndp->ni_cnd;
770 	int lkflags_save;
771 	int ni_dvp_unlocked;
772 
773 	/*
774 	 * Setup: break out flag bits into variables.
775 	 */
776 	ni_dvp_unlocked = 0;
777 	wantparent = cnp->cn_flags & (LOCKPARENT | WANTPARENT);
778 	KASSERT(cnp->cn_nameiop == LOOKUP || wantparent,
779 	    ("CREATE, DELETE, RENAME require LOCKPARENT or WANTPARENT."));
780 	docache = (cnp->cn_flags & NOCACHE) ^ NOCACHE;
781 	if (cnp->cn_nameiop == DELETE ||
782 	    (wantparent && cnp->cn_nameiop != CREATE &&
783 	     cnp->cn_nameiop != LOOKUP))
784 		docache = 0;
785 	rdonly = cnp->cn_flags & RDONLY;
786 	cnp->cn_flags &= ~ISSYMLINK;
787 	ndp->ni_dvp = NULL;
788 	/*
789 	 * We use shared locks until we hit the parent of the last cn then
790 	 * we adjust based on the requesting flags.
791 	 */
792 	cnp->cn_lkflags = LK_SHARED;
793 	dp = ndp->ni_startdir;
794 	ndp->ni_startdir = NULLVP;
795 	vn_lock(dp,
796 	    compute_cn_lkflags(dp->v_mount, cnp->cn_lkflags | LK_RETRY,
797 	    cnp->cn_flags));
798 
799 dirloop:
800 	/*
801 	 * Search a new directory.
802 	 *
803 	 * The last component of the filename is left accessible via
804 	 * cnp->cn_nameptr for callers that need the name. Callers needing
805 	 * the name set the SAVENAME flag. When done, they assume
806 	 * responsibility for freeing the pathname buffer.
807 	 */
808 	for (cp = cnp->cn_nameptr; *cp != 0 && *cp != '/'; cp++)
809 		continue;
810 	cnp->cn_namelen = cp - cnp->cn_nameptr;
811 	if (cnp->cn_namelen > NAME_MAX) {
812 		error = ENAMETOOLONG;
813 		goto bad;
814 	}
815 #ifdef NAMEI_DIAGNOSTIC
816 	{ char c = *cp;
817 	*cp = '\0';
818 	printf("{%s}: ", cnp->cn_nameptr);
819 	*cp = c; }
820 #endif
821 	prev_ni_pathlen = ndp->ni_pathlen;
822 	ndp->ni_pathlen -= cnp->cn_namelen;
823 	KASSERT(ndp->ni_pathlen <= PATH_MAX,
824 	    ("%s: ni_pathlen underflow to %zd\n", __func__, ndp->ni_pathlen));
825 	prev_ni_next = ndp->ni_next;
826 	ndp->ni_next = cp;
827 
828 	/*
829 	 * Replace multiple slashes by a single slash and trailing slashes
830 	 * by a null.  This must be done before VOP_LOOKUP() because some
831 	 * fs's don't know about trailing slashes.  Remember if there were
832 	 * trailing slashes to handle symlinks, existing non-directories
833 	 * and non-existing files that won't be directories specially later.
834 	 */
835 	while (*cp == '/' && (cp[1] == '/' || cp[1] == '\0')) {
836 		cp++;
837 		ndp->ni_pathlen--;
838 		if (*cp == '\0') {
839 			*ndp->ni_next = '\0';
840 			cnp->cn_flags |= TRAILINGSLASH;
841 		}
842 	}
843 	ndp->ni_next = cp;
844 
845 	cnp->cn_flags |= MAKEENTRY;
846 	if (*cp == '\0' && docache == 0)
847 		cnp->cn_flags &= ~MAKEENTRY;
848 	if (cnp->cn_namelen == 2 &&
849 	    cnp->cn_nameptr[1] == '.' && cnp->cn_nameptr[0] == '.')
850 		cnp->cn_flags |= ISDOTDOT;
851 	else
852 		cnp->cn_flags &= ~ISDOTDOT;
853 	if (*ndp->ni_next == 0)
854 		cnp->cn_flags |= ISLASTCN;
855 	else
856 		cnp->cn_flags &= ~ISLASTCN;
857 
858 	if ((cnp->cn_flags & ISLASTCN) != 0 &&
859 	    cnp->cn_namelen == 1 && cnp->cn_nameptr[0] == '.' &&
860 	    (cnp->cn_nameiop == DELETE || cnp->cn_nameiop == RENAME)) {
861 		error = EINVAL;
862 		goto bad;
863 	}
864 
865 	nameicap_tracker_add(ndp, dp);
866 
867 	/*
868 	 * Check for degenerate name (e.g. / or "")
869 	 * which is a way of talking about a directory,
870 	 * e.g. like "/." or ".".
871 	 */
872 	if (cnp->cn_nameptr[0] == '\0') {
873 		if (dp->v_type != VDIR) {
874 			error = ENOTDIR;
875 			goto bad;
876 		}
877 		if (cnp->cn_nameiop != LOOKUP) {
878 			error = EISDIR;
879 			goto bad;
880 		}
881 		if (wantparent) {
882 			ndp->ni_dvp = dp;
883 			VREF(dp);
884 		}
885 		ndp->ni_vp = dp;
886 
887 		if (cnp->cn_flags & AUDITVNODE1)
888 			AUDIT_ARG_VNODE1(dp);
889 		else if (cnp->cn_flags & AUDITVNODE2)
890 			AUDIT_ARG_VNODE2(dp);
891 
892 		if (!(cnp->cn_flags & (LOCKPARENT | LOCKLEAF)))
893 			VOP_UNLOCK(dp);
894 		/* XXX This should probably move to the top of function. */
895 		if (cnp->cn_flags & SAVESTART)
896 			panic("lookup: SAVESTART");
897 		goto success;
898 	}
899 
900 	/*
901 	 * Handle "..": five special cases.
902 	 * 0. If doing a capability lookup and lookup_cap_dotdot is
903 	 *    disabled, return ENOTCAPABLE.
904 	 * 1. Return an error if this is the last component of
905 	 *    the name and the operation is DELETE or RENAME.
906 	 * 2. If at root directory (e.g. after chroot)
907 	 *    or at absolute root directory
908 	 *    then ignore it so can't get out.
909 	 * 3. If this vnode is the root of a mounted
910 	 *    filesystem, then replace it with the
911 	 *    vnode which was mounted on so we take the
912 	 *    .. in the other filesystem.
913 	 * 4. If the vnode is the top directory of
914 	 *    the jail or chroot, don't let them out.
915 	 * 5. If doing a capability lookup and lookup_cap_dotdot is
916 	 *    enabled, return ENOTCAPABLE if the lookup would escape
917 	 *    from the initial file descriptor directory.  Checks are
918 	 *    done by ensuring that namei() already traversed the
919 	 *    result of dotdot lookup.
920 	 */
921 	if (cnp->cn_flags & ISDOTDOT) {
922 		if ((ndp->ni_lcf & (NI_LCF_STRICTRELATIVE | NI_LCF_CAP_DOTDOT))
923 		    == NI_LCF_STRICTRELATIVE) {
924 #ifdef KTRACE
925 			if (KTRPOINT(curthread, KTR_CAPFAIL))
926 				ktrcapfail(CAPFAIL_LOOKUP, NULL, NULL);
927 #endif
928 			error = ENOTCAPABLE;
929 			goto bad;
930 		}
931 		if ((cnp->cn_flags & ISLASTCN) != 0 &&
932 		    (cnp->cn_nameiop == DELETE || cnp->cn_nameiop == RENAME)) {
933 			error = EINVAL;
934 			goto bad;
935 		}
936 		for (;;) {
937 			for (pr = cnp->cn_cred->cr_prison; pr != NULL;
938 			     pr = pr->pr_parent)
939 				if (dp == pr->pr_root)
940 					break;
941 			if (dp == ndp->ni_rootdir ||
942 			    dp == ndp->ni_topdir ||
943 			    dp == rootvnode ||
944 			    pr != NULL ||
945 			    ((dp->v_vflag & VV_ROOT) != 0 &&
946 			     (cnp->cn_flags & NOCROSSMOUNT) != 0)) {
947 				ndp->ni_dvp = dp;
948 				ndp->ni_vp = dp;
949 				VREF(dp);
950 				goto nextname;
951 			}
952 			if ((dp->v_vflag & VV_ROOT) == 0)
953 				break;
954 			if (VN_IS_DOOMED(dp)) {	/* forced unmount */
955 				error = ENOENT;
956 				goto bad;
957 			}
958 			tdp = dp;
959 			dp = dp->v_mount->mnt_vnodecovered;
960 			VREF(dp);
961 			vput(tdp);
962 			vn_lock(dp,
963 			    compute_cn_lkflags(dp->v_mount, cnp->cn_lkflags |
964 			    LK_RETRY, ISDOTDOT));
965 			error = nameicap_check_dotdot(ndp, dp);
966 			if (error != 0) {
967 #ifdef KTRACE
968 				if (KTRPOINT(curthread, KTR_CAPFAIL))
969 					ktrcapfail(CAPFAIL_LOOKUP, NULL, NULL);
970 #endif
971 				goto bad;
972 			}
973 		}
974 	}
975 
976 	/*
977 	 * We now have a segment name to search for, and a directory to search.
978 	 */
979 unionlookup:
980 #ifdef MAC
981 	error = mac_vnode_check_lookup(cnp->cn_thread->td_ucred, dp, cnp);
982 	if (error)
983 		goto bad;
984 #endif
985 	ndp->ni_dvp = dp;
986 	ndp->ni_vp = NULL;
987 	ASSERT_VOP_LOCKED(dp, "lookup");
988 	/*
989 	 * If we have a shared lock we may need to upgrade the lock for the
990 	 * last operation.
991 	 */
992 	if ((cnp->cn_flags & LOCKPARENT) && (cnp->cn_flags & ISLASTCN) &&
993 	    dp != vp_crossmp && VOP_ISLOCKED(dp) == LK_SHARED)
994 		vn_lock(dp, LK_UPGRADE|LK_RETRY);
995 	if (VN_IS_DOOMED(dp)) {
996 		error = ENOENT;
997 		goto bad;
998 	}
999 	/*
1000 	 * If we're looking up the last component and we need an exclusive
1001 	 * lock, adjust our lkflags.
1002 	 */
1003 	if (needs_exclusive_leaf(dp->v_mount, cnp->cn_flags))
1004 		cnp->cn_lkflags = LK_EXCLUSIVE;
1005 #ifdef NAMEI_DIAGNOSTIC
1006 	vn_printf(dp, "lookup in ");
1007 #endif
1008 	lkflags_save = cnp->cn_lkflags;
1009 	cnp->cn_lkflags = compute_cn_lkflags(dp->v_mount, cnp->cn_lkflags,
1010 	    cnp->cn_flags);
1011 	error = VOP_LOOKUP(dp, &ndp->ni_vp, cnp);
1012 	cnp->cn_lkflags = lkflags_save;
1013 	if (error != 0) {
1014 		KASSERT(ndp->ni_vp == NULL, ("leaf should be empty"));
1015 #ifdef NAMEI_DIAGNOSTIC
1016 		printf("not found\n");
1017 #endif
1018 		if ((error == ENOENT) &&
1019 		    (dp->v_vflag & VV_ROOT) && (dp->v_mount != NULL) &&
1020 		    (dp->v_mount->mnt_flag & MNT_UNION)) {
1021 			tdp = dp;
1022 			dp = dp->v_mount->mnt_vnodecovered;
1023 			VREF(dp);
1024 			vput(tdp);
1025 			vn_lock(dp,
1026 			    compute_cn_lkflags(dp->v_mount, cnp->cn_lkflags |
1027 			    LK_RETRY, cnp->cn_flags));
1028 			nameicap_tracker_add(ndp, dp);
1029 			goto unionlookup;
1030 		}
1031 
1032 		if (error == ERELOOKUP) {
1033 			vref(dp);
1034 			ndp->ni_vp = dp;
1035 			error = 0;
1036 			relookup = 1;
1037 			goto good;
1038 		}
1039 
1040 		if (error != EJUSTRETURN)
1041 			goto bad;
1042 		/*
1043 		 * At this point, we know we're at the end of the
1044 		 * pathname.  If creating / renaming, we can consider
1045 		 * allowing the file or directory to be created / renamed,
1046 		 * provided we're not on a read-only filesystem.
1047 		 */
1048 		if (rdonly) {
1049 			error = EROFS;
1050 			goto bad;
1051 		}
1052 		/* trailing slash only allowed for directories */
1053 		if ((cnp->cn_flags & TRAILINGSLASH) &&
1054 		    !(cnp->cn_flags & WILLBEDIR)) {
1055 			error = ENOENT;
1056 			goto bad;
1057 		}
1058 		if ((cnp->cn_flags & LOCKPARENT) == 0)
1059 			VOP_UNLOCK(dp);
1060 		/*
1061 		 * We return with ni_vp NULL to indicate that the entry
1062 		 * doesn't currently exist, leaving a pointer to the
1063 		 * (possibly locked) directory vnode in ndp->ni_dvp.
1064 		 */
1065 		if (cnp->cn_flags & SAVESTART) {
1066 			ndp->ni_startdir = ndp->ni_dvp;
1067 			VREF(ndp->ni_startdir);
1068 		}
1069 		goto success;
1070 	}
1071 
1072 good:
1073 #ifdef NAMEI_DIAGNOSTIC
1074 	printf("found\n");
1075 #endif
1076 	dp = ndp->ni_vp;
1077 
1078 	/*
1079 	 * Check to see if the vnode has been mounted on;
1080 	 * if so find the root of the mounted filesystem.
1081 	 */
1082 	while (dp->v_type == VDIR && (mp = dp->v_mountedhere) &&
1083 	       (cnp->cn_flags & NOCROSSMOUNT) == 0) {
1084 		if (vfs_busy(mp, 0))
1085 			continue;
1086 		vput(dp);
1087 		if (dp != ndp->ni_dvp)
1088 			vput(ndp->ni_dvp);
1089 		else
1090 			vrele(ndp->ni_dvp);
1091 		vrefact(vp_crossmp);
1092 		ndp->ni_dvp = vp_crossmp;
1093 		error = VFS_ROOT(mp, compute_cn_lkflags(mp, cnp->cn_lkflags,
1094 		    cnp->cn_flags), &tdp);
1095 		vfs_unbusy(mp);
1096 		if (vn_lock(vp_crossmp, LK_SHARED | LK_NOWAIT))
1097 			panic("vp_crossmp exclusively locked or reclaimed");
1098 		if (error) {
1099 			dpunlocked = 1;
1100 			goto bad2;
1101 		}
1102 		ndp->ni_vp = dp = tdp;
1103 	}
1104 
1105 	/*
1106 	 * Check for symbolic link
1107 	 */
1108 	if ((dp->v_type == VLNK) &&
1109 	    ((cnp->cn_flags & FOLLOW) || (cnp->cn_flags & TRAILINGSLASH) ||
1110 	     *ndp->ni_next == '/')) {
1111 		cnp->cn_flags |= ISSYMLINK;
1112 		if (VN_IS_DOOMED(dp)) {
1113 			/*
1114 			 * We can't know whether the directory was mounted with
1115 			 * NOSYMFOLLOW, so we can't follow safely.
1116 			 */
1117 			error = ENOENT;
1118 			goto bad2;
1119 		}
1120 		if (dp->v_mount->mnt_flag & MNT_NOSYMFOLLOW) {
1121 			error = EACCES;
1122 			goto bad2;
1123 		}
1124 		/*
1125 		 * Symlink code always expects an unlocked dvp.
1126 		 */
1127 		if (ndp->ni_dvp != ndp->ni_vp) {
1128 			VOP_UNLOCK(ndp->ni_dvp);
1129 			ni_dvp_unlocked = 1;
1130 		}
1131 		goto success;
1132 	}
1133 
1134 nextname:
1135 	/*
1136 	 * Not a symbolic link that we will follow.  Continue with the
1137 	 * next component if there is any; otherwise, we're done.
1138 	 */
1139 	KASSERT((cnp->cn_flags & ISLASTCN) || *ndp->ni_next == '/',
1140 	    ("lookup: invalid path state."));
1141 	if (relookup) {
1142 		relookup = 0;
1143 		ndp->ni_pathlen = prev_ni_pathlen;
1144 		ndp->ni_next = prev_ni_next;
1145 		if (ndp->ni_dvp != dp)
1146 			vput(ndp->ni_dvp);
1147 		else
1148 			vrele(ndp->ni_dvp);
1149 		goto dirloop;
1150 	}
1151 	if (cnp->cn_flags & ISDOTDOT) {
1152 		error = nameicap_check_dotdot(ndp, ndp->ni_vp);
1153 		if (error != 0) {
1154 #ifdef KTRACE
1155 			if (KTRPOINT(curthread, KTR_CAPFAIL))
1156 				ktrcapfail(CAPFAIL_LOOKUP, NULL, NULL);
1157 #endif
1158 			goto bad2;
1159 		}
1160 	}
1161 	if (*ndp->ni_next == '/') {
1162 		cnp->cn_nameptr = ndp->ni_next;
1163 		while (*cnp->cn_nameptr == '/') {
1164 			cnp->cn_nameptr++;
1165 			ndp->ni_pathlen--;
1166 		}
1167 		if (ndp->ni_dvp != dp)
1168 			vput(ndp->ni_dvp);
1169 		else
1170 			vrele(ndp->ni_dvp);
1171 		goto dirloop;
1172 	}
1173 	/*
1174 	 * If we're processing a path with a trailing slash,
1175 	 * check that the end result is a directory.
1176 	 */
1177 	if ((cnp->cn_flags & TRAILINGSLASH) && dp->v_type != VDIR) {
1178 		error = ENOTDIR;
1179 		goto bad2;
1180 	}
1181 	/*
1182 	 * Disallow directory write attempts on read-only filesystems.
1183 	 */
1184 	if (rdonly &&
1185 	    (cnp->cn_nameiop == DELETE || cnp->cn_nameiop == RENAME)) {
1186 		error = EROFS;
1187 		goto bad2;
1188 	}
1189 	if (cnp->cn_flags & SAVESTART) {
1190 		ndp->ni_startdir = ndp->ni_dvp;
1191 		VREF(ndp->ni_startdir);
1192 	}
1193 	if (!wantparent) {
1194 		ni_dvp_unlocked = 2;
1195 		if (ndp->ni_dvp != dp)
1196 			vput(ndp->ni_dvp);
1197 		else
1198 			vrele(ndp->ni_dvp);
1199 	} else if ((cnp->cn_flags & LOCKPARENT) == 0 && ndp->ni_dvp != dp) {
1200 		VOP_UNLOCK(ndp->ni_dvp);
1201 		ni_dvp_unlocked = 1;
1202 	}
1203 
1204 	if (cnp->cn_flags & AUDITVNODE1)
1205 		AUDIT_ARG_VNODE1(dp);
1206 	else if (cnp->cn_flags & AUDITVNODE2)
1207 		AUDIT_ARG_VNODE2(dp);
1208 
1209 	if ((cnp->cn_flags & LOCKLEAF) == 0)
1210 		VOP_UNLOCK(dp);
1211 success:
1212 	/*
1213 	 * Because of shared lookup we may have the vnode shared locked, but
1214 	 * the caller may want it to be exclusively locked.
1215 	 */
1216 	if (needs_exclusive_leaf(dp->v_mount, cnp->cn_flags) &&
1217 	    VOP_ISLOCKED(dp) != LK_EXCLUSIVE) {
1218 		vn_lock(dp, LK_UPGRADE | LK_RETRY);
1219 		if (VN_IS_DOOMED(dp)) {
1220 			error = ENOENT;
1221 			goto bad2;
1222 		}
1223 	}
1224 	return (0);
1225 
1226 bad2:
1227 	if (ni_dvp_unlocked != 2) {
1228 		if (dp != ndp->ni_dvp && !ni_dvp_unlocked)
1229 			vput(ndp->ni_dvp);
1230 		else
1231 			vrele(ndp->ni_dvp);
1232 	}
1233 bad:
1234 	if (!dpunlocked)
1235 		vput(dp);
1236 	ndp->ni_vp = NULL;
1237 	return (error);
1238 }
1239 
1240 /*
1241  * relookup - lookup a path name component
1242  *    Used by lookup to re-acquire things.
1243  */
1244 int
1245 relookup(struct vnode *dvp, struct vnode **vpp, struct componentname *cnp)
1246 {
1247 	struct vnode *dp = NULL;		/* the directory we are searching */
1248 	int wantparent;			/* 1 => wantparent or lockparent flag */
1249 	int rdonly;			/* lookup read-only flag bit */
1250 	int error = 0;
1251 
1252 	KASSERT(cnp->cn_flags & ISLASTCN,
1253 	    ("relookup: Not given last component."));
1254 	/*
1255 	 * Setup: break out flag bits into variables.
1256 	 */
1257 	wantparent = cnp->cn_flags & (LOCKPARENT|WANTPARENT);
1258 	KASSERT(wantparent, ("relookup: parent not wanted."));
1259 	rdonly = cnp->cn_flags & RDONLY;
1260 	cnp->cn_flags &= ~ISSYMLINK;
1261 	dp = dvp;
1262 	cnp->cn_lkflags = LK_EXCLUSIVE;
1263 	vn_lock(dp, LK_EXCLUSIVE | LK_RETRY);
1264 
1265 	/*
1266 	 * Search a new directory.
1267 	 *
1268 	 * The last component of the filename is left accessible via
1269 	 * cnp->cn_nameptr for callers that need the name. Callers needing
1270 	 * the name set the SAVENAME flag. When done, they assume
1271 	 * responsibility for freeing the pathname buffer.
1272 	 */
1273 #ifdef NAMEI_DIAGNOSTIC
1274 	printf("{%s}: ", cnp->cn_nameptr);
1275 #endif
1276 
1277 	/*
1278 	 * Check for "" which represents the root directory after slash
1279 	 * removal.
1280 	 */
1281 	if (cnp->cn_nameptr[0] == '\0') {
1282 		/*
1283 		 * Support only LOOKUP for "/" because lookup()
1284 		 * can't succeed for CREATE, DELETE and RENAME.
1285 		 */
1286 		KASSERT(cnp->cn_nameiop == LOOKUP, ("nameiop must be LOOKUP"));
1287 		KASSERT(dp->v_type == VDIR, ("dp is not a directory"));
1288 
1289 		if (!(cnp->cn_flags & LOCKLEAF))
1290 			VOP_UNLOCK(dp);
1291 		*vpp = dp;
1292 		/* XXX This should probably move to the top of function. */
1293 		if (cnp->cn_flags & SAVESTART)
1294 			panic("lookup: SAVESTART");
1295 		return (0);
1296 	}
1297 
1298 	if (cnp->cn_flags & ISDOTDOT)
1299 		panic ("relookup: lookup on dot-dot");
1300 
1301 	/*
1302 	 * We now have a segment name to search for, and a directory to search.
1303 	 */
1304 #ifdef NAMEI_DIAGNOSTIC
1305 	vn_printf(dp, "search in ");
1306 #endif
1307 	if ((error = VOP_LOOKUP(dp, vpp, cnp)) != 0) {
1308 		KASSERT(*vpp == NULL, ("leaf should be empty"));
1309 		if (error != EJUSTRETURN)
1310 			goto bad;
1311 		/*
1312 		 * If creating and at end of pathname, then can consider
1313 		 * allowing file to be created.
1314 		 */
1315 		if (rdonly) {
1316 			error = EROFS;
1317 			goto bad;
1318 		}
1319 		/* ASSERT(dvp == ndp->ni_startdir) */
1320 		if (cnp->cn_flags & SAVESTART)
1321 			VREF(dvp);
1322 		if ((cnp->cn_flags & LOCKPARENT) == 0)
1323 			VOP_UNLOCK(dp);
1324 		/*
1325 		 * We return with ni_vp NULL to indicate that the entry
1326 		 * doesn't currently exist, leaving a pointer to the
1327 		 * (possibly locked) directory vnode in ndp->ni_dvp.
1328 		 */
1329 		return (0);
1330 	}
1331 
1332 	dp = *vpp;
1333 
1334 	/*
1335 	 * Disallow directory write attempts on read-only filesystems.
1336 	 */
1337 	if (rdonly &&
1338 	    (cnp->cn_nameiop == DELETE || cnp->cn_nameiop == RENAME)) {
1339 		if (dvp == dp)
1340 			vrele(dvp);
1341 		else
1342 			vput(dvp);
1343 		error = EROFS;
1344 		goto bad;
1345 	}
1346 	/*
1347 	 * Set the parent lock/ref state to the requested state.
1348 	 */
1349 	if ((cnp->cn_flags & LOCKPARENT) == 0 && dvp != dp) {
1350 		if (wantparent)
1351 			VOP_UNLOCK(dvp);
1352 		else
1353 			vput(dvp);
1354 	} else if (!wantparent)
1355 		vrele(dvp);
1356 	/*
1357 	 * Check for symbolic link
1358 	 */
1359 	KASSERT(dp->v_type != VLNK || !(cnp->cn_flags & FOLLOW),
1360 	    ("relookup: symlink found.\n"));
1361 
1362 	/* ASSERT(dvp == ndp->ni_startdir) */
1363 	if (cnp->cn_flags & SAVESTART)
1364 		VREF(dvp);
1365 
1366 	if ((cnp->cn_flags & LOCKLEAF) == 0)
1367 		VOP_UNLOCK(dp);
1368 	return (0);
1369 bad:
1370 	vput(dp);
1371 	*vpp = NULL;
1372 	return (error);
1373 }
1374 
1375 /*
1376  * Free data allocated by namei(); see namei(9) for details.
1377  */
1378 void
1379 NDFREE_PNBUF(struct nameidata *ndp)
1380 {
1381 
1382 	if ((ndp->ni_cnd.cn_flags & HASBUF) != 0) {
1383 		uma_zfree(namei_zone, ndp->ni_cnd.cn_pnbuf);
1384 		ndp->ni_cnd.cn_flags &= ~HASBUF;
1385 	}
1386 }
1387 
1388 void
1389 (NDFREE)(struct nameidata *ndp, const u_int flags)
1390 {
1391 	int unlock_dvp;
1392 	int unlock_vp;
1393 
1394 	unlock_dvp = 0;
1395 	unlock_vp = 0;
1396 
1397 	if (!(flags & NDF_NO_FREE_PNBUF)) {
1398 		NDFREE_PNBUF(ndp);
1399 	}
1400 	if (!(flags & NDF_NO_VP_UNLOCK) &&
1401 	    (ndp->ni_cnd.cn_flags & LOCKLEAF) && ndp->ni_vp)
1402 		unlock_vp = 1;
1403 	if (!(flags & NDF_NO_DVP_UNLOCK) &&
1404 	    (ndp->ni_cnd.cn_flags & LOCKPARENT) &&
1405 	    ndp->ni_dvp != ndp->ni_vp)
1406 		unlock_dvp = 1;
1407 	if (!(flags & NDF_NO_VP_RELE) && ndp->ni_vp) {
1408 		if (unlock_vp) {
1409 			vput(ndp->ni_vp);
1410 			unlock_vp = 0;
1411 		} else
1412 			vrele(ndp->ni_vp);
1413 		ndp->ni_vp = NULL;
1414 	}
1415 	if (unlock_vp)
1416 		VOP_UNLOCK(ndp->ni_vp);
1417 	if (!(flags & NDF_NO_DVP_RELE) &&
1418 	    (ndp->ni_cnd.cn_flags & (LOCKPARENT|WANTPARENT))) {
1419 		if (unlock_dvp) {
1420 			vput(ndp->ni_dvp);
1421 			unlock_dvp = 0;
1422 		} else
1423 			vrele(ndp->ni_dvp);
1424 		ndp->ni_dvp = NULL;
1425 	}
1426 	if (unlock_dvp)
1427 		VOP_UNLOCK(ndp->ni_dvp);
1428 	if (!(flags & NDF_NO_STARTDIR_RELE) &&
1429 	    (ndp->ni_cnd.cn_flags & SAVESTART)) {
1430 		vrele(ndp->ni_startdir);
1431 		ndp->ni_startdir = NULL;
1432 	}
1433 }
1434 
1435 /*
1436  * Determine if there is a suitable alternate filename under the specified
1437  * prefix for the specified path.  If the create flag is set, then the
1438  * alternate prefix will be used so long as the parent directory exists.
1439  * This is used by the various compatibility ABIs so that Linux binaries prefer
1440  * files under /compat/linux for example.  The chosen path (whether under
1441  * the prefix or under /) is returned in a kernel malloc'd buffer pointed
1442  * to by pathbuf.  The caller is responsible for free'ing the buffer from
1443  * the M_TEMP bucket if one is returned.
1444  */
1445 int
1446 kern_alternate_path(struct thread *td, const char *prefix, const char *path,
1447     enum uio_seg pathseg, char **pathbuf, int create, int dirfd)
1448 {
1449 	struct nameidata nd, ndroot;
1450 	char *ptr, *buf, *cp;
1451 	size_t len, sz;
1452 	int error;
1453 
1454 	buf = (char *) malloc(MAXPATHLEN, M_TEMP, M_WAITOK);
1455 	*pathbuf = buf;
1456 
1457 	/* Copy the prefix into the new pathname as a starting point. */
1458 	len = strlcpy(buf, prefix, MAXPATHLEN);
1459 	if (len >= MAXPATHLEN) {
1460 		*pathbuf = NULL;
1461 		free(buf, M_TEMP);
1462 		return (EINVAL);
1463 	}
1464 	sz = MAXPATHLEN - len;
1465 	ptr = buf + len;
1466 
1467 	/* Append the filename to the prefix. */
1468 	if (pathseg == UIO_SYSSPACE)
1469 		error = copystr(path, ptr, sz, &len);
1470 	else
1471 		error = copyinstr(path, ptr, sz, &len);
1472 
1473 	if (error) {
1474 		*pathbuf = NULL;
1475 		free(buf, M_TEMP);
1476 		return (error);
1477 	}
1478 
1479 	/* Only use a prefix with absolute pathnames. */
1480 	if (*ptr != '/') {
1481 		error = EINVAL;
1482 		goto keeporig;
1483 	}
1484 
1485 	if (dirfd != AT_FDCWD) {
1486 		/*
1487 		 * We want the original because the "prefix" is
1488 		 * included in the already opened dirfd.
1489 		 */
1490 		bcopy(ptr, buf, len);
1491 		return (0);
1492 	}
1493 
1494 	/*
1495 	 * We know that there is a / somewhere in this pathname.
1496 	 * Search backwards for it, to find the file's parent dir
1497 	 * to see if it exists in the alternate tree. If it does,
1498 	 * and we want to create a file (cflag is set). We don't
1499 	 * need to worry about the root comparison in this case.
1500 	 */
1501 
1502 	if (create) {
1503 		for (cp = &ptr[len] - 1; *cp != '/'; cp--);
1504 		*cp = '\0';
1505 
1506 		NDINIT(&nd, LOOKUP, NOFOLLOW, UIO_SYSSPACE, buf, td);
1507 		error = namei(&nd);
1508 		*cp = '/';
1509 		if (error != 0)
1510 			goto keeporig;
1511 	} else {
1512 		NDINIT(&nd, LOOKUP, NOFOLLOW, UIO_SYSSPACE, buf, td);
1513 
1514 		error = namei(&nd);
1515 		if (error != 0)
1516 			goto keeporig;
1517 
1518 		/*
1519 		 * We now compare the vnode of the prefix to the one
1520 		 * vnode asked. If they resolve to be the same, then we
1521 		 * ignore the match so that the real root gets used.
1522 		 * This avoids the problem of traversing "../.." to find the
1523 		 * root directory and never finding it, because "/" resolves
1524 		 * to the emulation root directory. This is expensive :-(
1525 		 */
1526 		NDINIT(&ndroot, LOOKUP, FOLLOW, UIO_SYSSPACE, prefix,
1527 		    td);
1528 
1529 		/* We shouldn't ever get an error from this namei(). */
1530 		error = namei(&ndroot);
1531 		if (error == 0) {
1532 			if (nd.ni_vp == ndroot.ni_vp)
1533 				error = ENOENT;
1534 
1535 			NDFREE(&ndroot, NDF_ONLY_PNBUF);
1536 			vrele(ndroot.ni_vp);
1537 		}
1538 	}
1539 
1540 	NDFREE(&nd, NDF_ONLY_PNBUF);
1541 	vrele(nd.ni_vp);
1542 
1543 keeporig:
1544 	/* If there was an error, use the original path name. */
1545 	if (error)
1546 		bcopy(ptr, buf, len);
1547 	return (error);
1548 }
1549