xref: /freebsd/sys/kern/vfs_lookup.c (revision d5e3895ea4fe4ef9db8823774e07b4368180a23e)
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 	cnp->cn_pnbuf = uma_zalloc(namei_zone, M_WAITOK);
508 	if (ndp->ni_segflg == UIO_SYSSPACE)
509 		error = copystr(ndp->ni_dirp, cnp->cn_pnbuf, MAXPATHLEN,
510 		    &ndp->ni_pathlen);
511 	else
512 		error = copyinstr(ndp->ni_dirp, cnp->cn_pnbuf, MAXPATHLEN,
513 		    &ndp->ni_pathlen);
514 
515 	if (__predict_false(error != 0)) {
516 		namei_cleanup_cnp(cnp);
517 		return (error);
518 	}
519 
520 	/*
521 	 * Don't allow empty pathnames.
522 	 */
523 	if (__predict_false(*cnp->cn_pnbuf == '\0')) {
524 		namei_cleanup_cnp(cnp);
525 		return (ENOENT);
526 	}
527 
528 #ifdef KTRACE
529 	if (KTRPOINT(td, KTR_NAMEI)) {
530 		KASSERT(cnp->cn_thread == curthread,
531 		    ("namei not using curthread"));
532 		ktrnamei(cnp->cn_pnbuf);
533 	}
534 #endif
535 
536 	cnp->cn_nameptr = cnp->cn_pnbuf;
537 
538 	/*
539 	 * First try looking up the target without locking any vnodes.
540 	 *
541 	 * We may need to start from scratch or pick up where it left off.
542 	 */
543 	error = cache_fplookup(ndp, &status, &pwd);
544 	switch (status) {
545 	case CACHE_FPL_STATUS_UNSET:
546 		__assert_unreachable();
547 		break;
548 	case CACHE_FPL_STATUS_HANDLED:
549 		return (error);
550 	case CACHE_FPL_STATUS_PARTIAL:
551 		TAILQ_INIT(&ndp->ni_cap_tracker);
552 		dp = ndp->ni_startdir;
553 		break;
554 	case CACHE_FPL_STATUS_ABORTED:
555 		TAILQ_INIT(&ndp->ni_cap_tracker);
556 		error = namei_setup(ndp, &dp, &pwd);
557 		if (error != 0) {
558 			namei_cleanup_cnp(cnp);
559 			return (error);
560 		}
561 		break;
562 	}
563 
564 	ndp->ni_loopcnt = 0;
565 
566 	/*
567 	 * Locked lookup.
568 	 */
569 	for (;;) {
570 		ndp->ni_startdir = dp;
571 		error = lookup(ndp);
572 		if (error != 0)
573 			goto out;
574 		/*
575 		 * If not a symbolic link, we're done.
576 		 */
577 		if ((cnp->cn_flags & ISSYMLINK) == 0) {
578 			if ((cnp->cn_flags & (SAVENAME | SAVESTART)) == 0) {
579 				namei_cleanup_cnp(cnp);
580 			} else
581 				cnp->cn_flags |= HASBUF;
582 			if ((ndp->ni_lcf & (NI_LCF_BENEATH_ABS |
583 			    NI_LCF_BENEATH_LATCHED)) == NI_LCF_BENEATH_ABS) {
584 				NDFREE(ndp, 0);
585 				error = ENOTCAPABLE;
586 			}
587 			nameicap_cleanup(ndp, true);
588 			SDT_PROBE3(vfs, namei, lookup, return, error,
589 			    (error == 0 ? ndp->ni_vp : NULL), false);
590 			pwd_drop(pwd);
591 			return (error);
592 		}
593 		if (ndp->ni_loopcnt++ >= MAXSYMLINKS) {
594 			error = ELOOP;
595 			break;
596 		}
597 #ifdef MAC
598 		if ((cnp->cn_flags & NOMACCHECK) == 0) {
599 			error = mac_vnode_check_readlink(td->td_ucred,
600 			    ndp->ni_vp);
601 			if (error != 0)
602 				break;
603 		}
604 #endif
605 		if (ndp->ni_pathlen > 1)
606 			cp = uma_zalloc(namei_zone, M_WAITOK);
607 		else
608 			cp = cnp->cn_pnbuf;
609 		aiov.iov_base = cp;
610 		aiov.iov_len = MAXPATHLEN;
611 		auio.uio_iov = &aiov;
612 		auio.uio_iovcnt = 1;
613 		auio.uio_offset = 0;
614 		auio.uio_rw = UIO_READ;
615 		auio.uio_segflg = UIO_SYSSPACE;
616 		auio.uio_td = td;
617 		auio.uio_resid = MAXPATHLEN;
618 		error = VOP_READLINK(ndp->ni_vp, &auio, cnp->cn_cred);
619 		if (error != 0) {
620 			if (ndp->ni_pathlen > 1)
621 				uma_zfree(namei_zone, cp);
622 			break;
623 		}
624 		linklen = MAXPATHLEN - auio.uio_resid;
625 		if (linklen == 0) {
626 			if (ndp->ni_pathlen > 1)
627 				uma_zfree(namei_zone, cp);
628 			error = ENOENT;
629 			break;
630 		}
631 		if (linklen + ndp->ni_pathlen > MAXPATHLEN) {
632 			if (ndp->ni_pathlen > 1)
633 				uma_zfree(namei_zone, cp);
634 			error = ENAMETOOLONG;
635 			break;
636 		}
637 		if (ndp->ni_pathlen > 1) {
638 			bcopy(ndp->ni_next, cp + linklen, ndp->ni_pathlen);
639 			uma_zfree(namei_zone, cnp->cn_pnbuf);
640 			cnp->cn_pnbuf = cp;
641 		} else
642 			cnp->cn_pnbuf[linklen] = '\0';
643 		ndp->ni_pathlen += linklen;
644 		vput(ndp->ni_vp);
645 		dp = ndp->ni_dvp;
646 		/*
647 		 * Check if root directory should replace current directory.
648 		 */
649 		cnp->cn_nameptr = cnp->cn_pnbuf;
650 		if (*(cnp->cn_nameptr) == '/') {
651 			vrele(dp);
652 			error = namei_handle_root(ndp, &dp);
653 			if (error != 0)
654 				goto out;
655 		}
656 	}
657 	vput(ndp->ni_vp);
658 	ndp->ni_vp = NULL;
659 	vrele(ndp->ni_dvp);
660 out:
661 	MPASS(error != 0);
662 	namei_cleanup_cnp(cnp);
663 	nameicap_cleanup(ndp, true);
664 	SDT_PROBE3(vfs, namei, lookup, return, error, NULL, false);
665 	pwd_drop(pwd);
666 	return (error);
667 }
668 
669 static int
670 compute_cn_lkflags(struct mount *mp, int lkflags, int cnflags)
671 {
672 
673 	if (mp == NULL || ((lkflags & LK_SHARED) &&
674 	    (!(mp->mnt_kern_flag & MNTK_LOOKUP_SHARED) ||
675 	    ((cnflags & ISDOTDOT) &&
676 	    (mp->mnt_kern_flag & MNTK_LOOKUP_EXCL_DOTDOT))))) {
677 		lkflags &= ~LK_SHARED;
678 		lkflags |= LK_EXCLUSIVE;
679 	}
680 	lkflags |= LK_NODDLKTREAT;
681 	return (lkflags);
682 }
683 
684 static __inline int
685 needs_exclusive_leaf(struct mount *mp, int flags)
686 {
687 
688 	/*
689 	 * Intermediate nodes can use shared locks, we only need to
690 	 * force an exclusive lock for leaf nodes.
691 	 */
692 	if ((flags & (ISLASTCN | LOCKLEAF)) != (ISLASTCN | LOCKLEAF))
693 		return (0);
694 
695 	/* Always use exclusive locks if LOCKSHARED isn't set. */
696 	if (!(flags & LOCKSHARED))
697 		return (1);
698 
699 	/*
700 	 * For lookups during open(), if the mount point supports
701 	 * extended shared operations, then use a shared lock for the
702 	 * leaf node, otherwise use an exclusive lock.
703 	 */
704 	if ((flags & ISOPEN) != 0)
705 		return (!MNT_EXTENDED_SHARED(mp));
706 
707 	/*
708 	 * Lookup requests outside of open() that specify LOCKSHARED
709 	 * only need a shared lock on the leaf vnode.
710 	 */
711 	return (0);
712 }
713 
714 /*
715  * Search a pathname.
716  * This is a very central and rather complicated routine.
717  *
718  * The pathname is pointed to by ni_ptr and is of length ni_pathlen.
719  * The starting directory is taken from ni_startdir. The pathname is
720  * descended until done, or a symbolic link is encountered. The variable
721  * ni_more is clear if the path is completed; it is set to one if a
722  * symbolic link needing interpretation is encountered.
723  *
724  * The flag argument is LOOKUP, CREATE, RENAME, or DELETE depending on
725  * whether the name is to be looked up, created, renamed, or deleted.
726  * When CREATE, RENAME, or DELETE is specified, information usable in
727  * creating, renaming, or deleting a directory entry may be calculated.
728  * If flag has LOCKPARENT or'ed into it, the parent directory is returned
729  * locked. If flag has WANTPARENT or'ed into it, the parent directory is
730  * returned unlocked. Otherwise the parent directory is not returned. If
731  * the target of the pathname exists and LOCKLEAF is or'ed into the flag
732  * the target is returned locked, otherwise it is returned unlocked.
733  * When creating or renaming and LOCKPARENT is specified, the target may not
734  * be ".".  When deleting and LOCKPARENT is specified, the target may be ".".
735  *
736  * Overall outline of lookup:
737  *
738  * dirloop:
739  *	identify next component of name at ndp->ni_ptr
740  *	handle degenerate case where name is null string
741  *	if .. and crossing mount points and on mounted filesys, find parent
742  *	call VOP_LOOKUP routine for next component name
743  *	    directory vnode returned in ni_dvp, unlocked unless LOCKPARENT set
744  *	    component vnode returned in ni_vp (if it exists), locked.
745  *	if result vnode is mounted on and crossing mount points,
746  *	    find mounted on vnode
747  *	if more components of name, do next level at dirloop
748  *	return the answer in ni_vp, locked if LOCKLEAF set
749  *	    if LOCKPARENT set, return locked parent in ni_dvp
750  *	    if WANTPARENT set, return unlocked parent in ni_dvp
751  */
752 int
753 lookup(struct nameidata *ndp)
754 {
755 	char *cp;			/* pointer into pathname argument */
756 	char *prev_ni_next;		/* saved ndp->ni_next */
757 	struct vnode *dp = NULL;	/* the directory we are searching */
758 	struct vnode *tdp;		/* saved dp */
759 	struct mount *mp;		/* mount table entry */
760 	struct prison *pr;
761 	size_t prev_ni_pathlen;		/* saved ndp->ni_pathlen */
762 	int docache;			/* == 0 do not cache last component */
763 	int wantparent;			/* 1 => wantparent or lockparent flag */
764 	int rdonly;			/* lookup read-only flag bit */
765 	int error = 0;
766 	int dpunlocked = 0;		/* dp has already been unlocked */
767 	int relookup = 0;		/* do not consume the path component */
768 	struct componentname *cnp = &ndp->ni_cnd;
769 	int lkflags_save;
770 	int ni_dvp_unlocked;
771 
772 	/*
773 	 * Setup: break out flag bits into variables.
774 	 */
775 	ni_dvp_unlocked = 0;
776 	wantparent = cnp->cn_flags & (LOCKPARENT | WANTPARENT);
777 	KASSERT(cnp->cn_nameiop == LOOKUP || wantparent,
778 	    ("CREATE, DELETE, RENAME require LOCKPARENT or WANTPARENT."));
779 	docache = (cnp->cn_flags & NOCACHE) ^ NOCACHE;
780 	if (cnp->cn_nameiop == DELETE ||
781 	    (wantparent && cnp->cn_nameiop != CREATE &&
782 	     cnp->cn_nameiop != LOOKUP))
783 		docache = 0;
784 	rdonly = cnp->cn_flags & RDONLY;
785 	cnp->cn_flags &= ~ISSYMLINK;
786 	ndp->ni_dvp = NULL;
787 	/*
788 	 * We use shared locks until we hit the parent of the last cn then
789 	 * we adjust based on the requesting flags.
790 	 */
791 	cnp->cn_lkflags = LK_SHARED;
792 	dp = ndp->ni_startdir;
793 	ndp->ni_startdir = NULLVP;
794 	vn_lock(dp,
795 	    compute_cn_lkflags(dp->v_mount, cnp->cn_lkflags | LK_RETRY,
796 	    cnp->cn_flags));
797 
798 dirloop:
799 	/*
800 	 * Search a new directory.
801 	 *
802 	 * The last component of the filename is left accessible via
803 	 * cnp->cn_nameptr for callers that need the name. Callers needing
804 	 * the name set the SAVENAME flag. When done, they assume
805 	 * responsibility for freeing the pathname buffer.
806 	 */
807 	for (cp = cnp->cn_nameptr; *cp != 0 && *cp != '/'; cp++)
808 		continue;
809 	cnp->cn_namelen = cp - cnp->cn_nameptr;
810 	if (cnp->cn_namelen > NAME_MAX) {
811 		error = ENAMETOOLONG;
812 		goto bad;
813 	}
814 #ifdef NAMEI_DIAGNOSTIC
815 	{ char c = *cp;
816 	*cp = '\0';
817 	printf("{%s}: ", cnp->cn_nameptr);
818 	*cp = c; }
819 #endif
820 	prev_ni_pathlen = ndp->ni_pathlen;
821 	ndp->ni_pathlen -= cnp->cn_namelen;
822 	KASSERT(ndp->ni_pathlen <= PATH_MAX,
823 	    ("%s: ni_pathlen underflow to %zd\n", __func__, ndp->ni_pathlen));
824 	prev_ni_next = ndp->ni_next;
825 	ndp->ni_next = cp;
826 
827 	/*
828 	 * Replace multiple slashes by a single slash and trailing slashes
829 	 * by a null.  This must be done before VOP_LOOKUP() because some
830 	 * fs's don't know about trailing slashes.  Remember if there were
831 	 * trailing slashes to handle symlinks, existing non-directories
832 	 * and non-existing files that won't be directories specially later.
833 	 */
834 	while (*cp == '/' && (cp[1] == '/' || cp[1] == '\0')) {
835 		cp++;
836 		ndp->ni_pathlen--;
837 		if (*cp == '\0') {
838 			*ndp->ni_next = '\0';
839 			cnp->cn_flags |= TRAILINGSLASH;
840 		}
841 	}
842 	ndp->ni_next = cp;
843 
844 	cnp->cn_flags |= MAKEENTRY;
845 	if (*cp == '\0' && docache == 0)
846 		cnp->cn_flags &= ~MAKEENTRY;
847 	if (cnp->cn_namelen == 2 &&
848 	    cnp->cn_nameptr[1] == '.' && cnp->cn_nameptr[0] == '.')
849 		cnp->cn_flags |= ISDOTDOT;
850 	else
851 		cnp->cn_flags &= ~ISDOTDOT;
852 	if (*ndp->ni_next == 0)
853 		cnp->cn_flags |= ISLASTCN;
854 	else
855 		cnp->cn_flags &= ~ISLASTCN;
856 
857 	if ((cnp->cn_flags & ISLASTCN) != 0 &&
858 	    cnp->cn_namelen == 1 && cnp->cn_nameptr[0] == '.' &&
859 	    (cnp->cn_nameiop == DELETE || cnp->cn_nameiop == RENAME)) {
860 		error = EINVAL;
861 		goto bad;
862 	}
863 
864 	nameicap_tracker_add(ndp, dp);
865 
866 	/*
867 	 * Check for degenerate name (e.g. / or "")
868 	 * which is a way of talking about a directory,
869 	 * e.g. like "/." or ".".
870 	 */
871 	if (cnp->cn_nameptr[0] == '\0') {
872 		if (dp->v_type != VDIR) {
873 			error = ENOTDIR;
874 			goto bad;
875 		}
876 		if (cnp->cn_nameiop != LOOKUP) {
877 			error = EISDIR;
878 			goto bad;
879 		}
880 		if (wantparent) {
881 			ndp->ni_dvp = dp;
882 			VREF(dp);
883 		}
884 		ndp->ni_vp = dp;
885 
886 		if (cnp->cn_flags & AUDITVNODE1)
887 			AUDIT_ARG_VNODE1(dp);
888 		else if (cnp->cn_flags & AUDITVNODE2)
889 			AUDIT_ARG_VNODE2(dp);
890 
891 		if (!(cnp->cn_flags & (LOCKPARENT | LOCKLEAF)))
892 			VOP_UNLOCK(dp);
893 		/* XXX This should probably move to the top of function. */
894 		if (cnp->cn_flags & SAVESTART)
895 			panic("lookup: SAVESTART");
896 		goto success;
897 	}
898 
899 	/*
900 	 * Handle "..": five special cases.
901 	 * 0. If doing a capability lookup and lookup_cap_dotdot is
902 	 *    disabled, return ENOTCAPABLE.
903 	 * 1. Return an error if this is the last component of
904 	 *    the name and the operation is DELETE or RENAME.
905 	 * 2. If at root directory (e.g. after chroot)
906 	 *    or at absolute root directory
907 	 *    then ignore it so can't get out.
908 	 * 3. If this vnode is the root of a mounted
909 	 *    filesystem, then replace it with the
910 	 *    vnode which was mounted on so we take the
911 	 *    .. in the other filesystem.
912 	 * 4. If the vnode is the top directory of
913 	 *    the jail or chroot, don't let them out.
914 	 * 5. If doing a capability lookup and lookup_cap_dotdot is
915 	 *    enabled, return ENOTCAPABLE if the lookup would escape
916 	 *    from the initial file descriptor directory.  Checks are
917 	 *    done by ensuring that namei() already traversed the
918 	 *    result of dotdot lookup.
919 	 */
920 	if (cnp->cn_flags & ISDOTDOT) {
921 		if ((ndp->ni_lcf & (NI_LCF_STRICTRELATIVE | NI_LCF_CAP_DOTDOT))
922 		    == NI_LCF_STRICTRELATIVE) {
923 #ifdef KTRACE
924 			if (KTRPOINT(curthread, KTR_CAPFAIL))
925 				ktrcapfail(CAPFAIL_LOOKUP, NULL, NULL);
926 #endif
927 			error = ENOTCAPABLE;
928 			goto bad;
929 		}
930 		if ((cnp->cn_flags & ISLASTCN) != 0 &&
931 		    (cnp->cn_nameiop == DELETE || cnp->cn_nameiop == RENAME)) {
932 			error = EINVAL;
933 			goto bad;
934 		}
935 		for (;;) {
936 			for (pr = cnp->cn_cred->cr_prison; pr != NULL;
937 			     pr = pr->pr_parent)
938 				if (dp == pr->pr_root)
939 					break;
940 			if (dp == ndp->ni_rootdir ||
941 			    dp == ndp->ni_topdir ||
942 			    dp == rootvnode ||
943 			    pr != NULL ||
944 			    ((dp->v_vflag & VV_ROOT) != 0 &&
945 			     (cnp->cn_flags & NOCROSSMOUNT) != 0)) {
946 				ndp->ni_dvp = dp;
947 				ndp->ni_vp = dp;
948 				VREF(dp);
949 				goto nextname;
950 			}
951 			if ((dp->v_vflag & VV_ROOT) == 0)
952 				break;
953 			if (VN_IS_DOOMED(dp)) {	/* forced unmount */
954 				error = ENOENT;
955 				goto bad;
956 			}
957 			tdp = dp;
958 			dp = dp->v_mount->mnt_vnodecovered;
959 			VREF(dp);
960 			vput(tdp);
961 			vn_lock(dp,
962 			    compute_cn_lkflags(dp->v_mount, cnp->cn_lkflags |
963 			    LK_RETRY, ISDOTDOT));
964 			error = nameicap_check_dotdot(ndp, dp);
965 			if (error != 0) {
966 #ifdef KTRACE
967 				if (KTRPOINT(curthread, KTR_CAPFAIL))
968 					ktrcapfail(CAPFAIL_LOOKUP, NULL, NULL);
969 #endif
970 				goto bad;
971 			}
972 		}
973 	}
974 
975 	/*
976 	 * We now have a segment name to search for, and a directory to search.
977 	 */
978 unionlookup:
979 #ifdef MAC
980 	error = mac_vnode_check_lookup(cnp->cn_thread->td_ucred, dp, cnp);
981 	if (error)
982 		goto bad;
983 #endif
984 	ndp->ni_dvp = dp;
985 	ndp->ni_vp = NULL;
986 	ASSERT_VOP_LOCKED(dp, "lookup");
987 	/*
988 	 * If we have a shared lock we may need to upgrade the lock for the
989 	 * last operation.
990 	 */
991 	if ((cnp->cn_flags & LOCKPARENT) && (cnp->cn_flags & ISLASTCN) &&
992 	    dp != vp_crossmp && VOP_ISLOCKED(dp) == LK_SHARED)
993 		vn_lock(dp, LK_UPGRADE|LK_RETRY);
994 	if (VN_IS_DOOMED(dp)) {
995 		error = ENOENT;
996 		goto bad;
997 	}
998 	/*
999 	 * If we're looking up the last component and we need an exclusive
1000 	 * lock, adjust our lkflags.
1001 	 */
1002 	if (needs_exclusive_leaf(dp->v_mount, cnp->cn_flags))
1003 		cnp->cn_lkflags = LK_EXCLUSIVE;
1004 #ifdef NAMEI_DIAGNOSTIC
1005 	vn_printf(dp, "lookup in ");
1006 #endif
1007 	lkflags_save = cnp->cn_lkflags;
1008 	cnp->cn_lkflags = compute_cn_lkflags(dp->v_mount, cnp->cn_lkflags,
1009 	    cnp->cn_flags);
1010 	error = VOP_LOOKUP(dp, &ndp->ni_vp, cnp);
1011 	cnp->cn_lkflags = lkflags_save;
1012 	if (error != 0) {
1013 		KASSERT(ndp->ni_vp == NULL, ("leaf should be empty"));
1014 #ifdef NAMEI_DIAGNOSTIC
1015 		printf("not found\n");
1016 #endif
1017 		if ((error == ENOENT) &&
1018 		    (dp->v_vflag & VV_ROOT) && (dp->v_mount != NULL) &&
1019 		    (dp->v_mount->mnt_flag & MNT_UNION)) {
1020 			tdp = dp;
1021 			dp = dp->v_mount->mnt_vnodecovered;
1022 			VREF(dp);
1023 			vput(tdp);
1024 			vn_lock(dp,
1025 			    compute_cn_lkflags(dp->v_mount, cnp->cn_lkflags |
1026 			    LK_RETRY, cnp->cn_flags));
1027 			nameicap_tracker_add(ndp, dp);
1028 			goto unionlookup;
1029 		}
1030 
1031 		if (error == ERELOOKUP) {
1032 			vref(dp);
1033 			ndp->ni_vp = dp;
1034 			error = 0;
1035 			relookup = 1;
1036 			goto good;
1037 		}
1038 
1039 		if (error != EJUSTRETURN)
1040 			goto bad;
1041 		/*
1042 		 * At this point, we know we're at the end of the
1043 		 * pathname.  If creating / renaming, we can consider
1044 		 * allowing the file or directory to be created / renamed,
1045 		 * provided we're not on a read-only filesystem.
1046 		 */
1047 		if (rdonly) {
1048 			error = EROFS;
1049 			goto bad;
1050 		}
1051 		/* trailing slash only allowed for directories */
1052 		if ((cnp->cn_flags & TRAILINGSLASH) &&
1053 		    !(cnp->cn_flags & WILLBEDIR)) {
1054 			error = ENOENT;
1055 			goto bad;
1056 		}
1057 		if ((cnp->cn_flags & LOCKPARENT) == 0)
1058 			VOP_UNLOCK(dp);
1059 		/*
1060 		 * We return with ni_vp NULL to indicate that the entry
1061 		 * doesn't currently exist, leaving a pointer to the
1062 		 * (possibly locked) directory vnode in ndp->ni_dvp.
1063 		 */
1064 		if (cnp->cn_flags & SAVESTART) {
1065 			ndp->ni_startdir = ndp->ni_dvp;
1066 			VREF(ndp->ni_startdir);
1067 		}
1068 		goto success;
1069 	}
1070 
1071 good:
1072 #ifdef NAMEI_DIAGNOSTIC
1073 	printf("found\n");
1074 #endif
1075 	dp = ndp->ni_vp;
1076 
1077 	/*
1078 	 * Check to see if the vnode has been mounted on;
1079 	 * if so find the root of the mounted filesystem.
1080 	 */
1081 	while (dp->v_type == VDIR && (mp = dp->v_mountedhere) &&
1082 	       (cnp->cn_flags & NOCROSSMOUNT) == 0) {
1083 		if (vfs_busy(mp, 0))
1084 			continue;
1085 		vput(dp);
1086 		if (dp != ndp->ni_dvp)
1087 			vput(ndp->ni_dvp);
1088 		else
1089 			vrele(ndp->ni_dvp);
1090 		vrefact(vp_crossmp);
1091 		ndp->ni_dvp = vp_crossmp;
1092 		error = VFS_ROOT(mp, compute_cn_lkflags(mp, cnp->cn_lkflags,
1093 		    cnp->cn_flags), &tdp);
1094 		vfs_unbusy(mp);
1095 		if (vn_lock(vp_crossmp, LK_SHARED | LK_NOWAIT))
1096 			panic("vp_crossmp exclusively locked or reclaimed");
1097 		if (error) {
1098 			dpunlocked = 1;
1099 			goto bad2;
1100 		}
1101 		ndp->ni_vp = dp = tdp;
1102 	}
1103 
1104 	/*
1105 	 * Check for symbolic link
1106 	 */
1107 	if ((dp->v_type == VLNK) &&
1108 	    ((cnp->cn_flags & FOLLOW) || (cnp->cn_flags & TRAILINGSLASH) ||
1109 	     *ndp->ni_next == '/')) {
1110 		cnp->cn_flags |= ISSYMLINK;
1111 		if (VN_IS_DOOMED(dp)) {
1112 			/*
1113 			 * We can't know whether the directory was mounted with
1114 			 * NOSYMFOLLOW, so we can't follow safely.
1115 			 */
1116 			error = ENOENT;
1117 			goto bad2;
1118 		}
1119 		if (dp->v_mount->mnt_flag & MNT_NOSYMFOLLOW) {
1120 			error = EACCES;
1121 			goto bad2;
1122 		}
1123 		/*
1124 		 * Symlink code always expects an unlocked dvp.
1125 		 */
1126 		if (ndp->ni_dvp != ndp->ni_vp) {
1127 			VOP_UNLOCK(ndp->ni_dvp);
1128 			ni_dvp_unlocked = 1;
1129 		}
1130 		goto success;
1131 	}
1132 
1133 nextname:
1134 	/*
1135 	 * Not a symbolic link that we will follow.  Continue with the
1136 	 * next component if there is any; otherwise, we're done.
1137 	 */
1138 	KASSERT((cnp->cn_flags & ISLASTCN) || *ndp->ni_next == '/',
1139 	    ("lookup: invalid path state."));
1140 	if (relookup) {
1141 		relookup = 0;
1142 		ndp->ni_pathlen = prev_ni_pathlen;
1143 		ndp->ni_next = prev_ni_next;
1144 		if (ndp->ni_dvp != dp)
1145 			vput(ndp->ni_dvp);
1146 		else
1147 			vrele(ndp->ni_dvp);
1148 		goto dirloop;
1149 	}
1150 	if (cnp->cn_flags & ISDOTDOT) {
1151 		error = nameicap_check_dotdot(ndp, ndp->ni_vp);
1152 		if (error != 0) {
1153 #ifdef KTRACE
1154 			if (KTRPOINT(curthread, KTR_CAPFAIL))
1155 				ktrcapfail(CAPFAIL_LOOKUP, NULL, NULL);
1156 #endif
1157 			goto bad2;
1158 		}
1159 	}
1160 	if (*ndp->ni_next == '/') {
1161 		cnp->cn_nameptr = ndp->ni_next;
1162 		while (*cnp->cn_nameptr == '/') {
1163 			cnp->cn_nameptr++;
1164 			ndp->ni_pathlen--;
1165 		}
1166 		if (ndp->ni_dvp != dp)
1167 			vput(ndp->ni_dvp);
1168 		else
1169 			vrele(ndp->ni_dvp);
1170 		goto dirloop;
1171 	}
1172 	/*
1173 	 * If we're processing a path with a trailing slash,
1174 	 * check that the end result is a directory.
1175 	 */
1176 	if ((cnp->cn_flags & TRAILINGSLASH) && dp->v_type != VDIR) {
1177 		error = ENOTDIR;
1178 		goto bad2;
1179 	}
1180 	/*
1181 	 * Disallow directory write attempts on read-only filesystems.
1182 	 */
1183 	if (rdonly &&
1184 	    (cnp->cn_nameiop == DELETE || cnp->cn_nameiop == RENAME)) {
1185 		error = EROFS;
1186 		goto bad2;
1187 	}
1188 	if (cnp->cn_flags & SAVESTART) {
1189 		ndp->ni_startdir = ndp->ni_dvp;
1190 		VREF(ndp->ni_startdir);
1191 	}
1192 	if (!wantparent) {
1193 		ni_dvp_unlocked = 2;
1194 		if (ndp->ni_dvp != dp)
1195 			vput(ndp->ni_dvp);
1196 		else
1197 			vrele(ndp->ni_dvp);
1198 	} else if ((cnp->cn_flags & LOCKPARENT) == 0 && ndp->ni_dvp != dp) {
1199 		VOP_UNLOCK(ndp->ni_dvp);
1200 		ni_dvp_unlocked = 1;
1201 	}
1202 
1203 	if (cnp->cn_flags & AUDITVNODE1)
1204 		AUDIT_ARG_VNODE1(dp);
1205 	else if (cnp->cn_flags & AUDITVNODE2)
1206 		AUDIT_ARG_VNODE2(dp);
1207 
1208 	if ((cnp->cn_flags & LOCKLEAF) == 0)
1209 		VOP_UNLOCK(dp);
1210 success:
1211 	/*
1212 	 * Because of shared lookup we may have the vnode shared locked, but
1213 	 * the caller may want it to be exclusively locked.
1214 	 */
1215 	if (needs_exclusive_leaf(dp->v_mount, cnp->cn_flags) &&
1216 	    VOP_ISLOCKED(dp) != LK_EXCLUSIVE) {
1217 		vn_lock(dp, LK_UPGRADE | LK_RETRY);
1218 		if (VN_IS_DOOMED(dp)) {
1219 			error = ENOENT;
1220 			goto bad2;
1221 		}
1222 	}
1223 	return (0);
1224 
1225 bad2:
1226 	if (ni_dvp_unlocked != 2) {
1227 		if (dp != ndp->ni_dvp && !ni_dvp_unlocked)
1228 			vput(ndp->ni_dvp);
1229 		else
1230 			vrele(ndp->ni_dvp);
1231 	}
1232 bad:
1233 	if (!dpunlocked)
1234 		vput(dp);
1235 	ndp->ni_vp = NULL;
1236 	return (error);
1237 }
1238 
1239 /*
1240  * relookup - lookup a path name component
1241  *    Used by lookup to re-acquire things.
1242  */
1243 int
1244 relookup(struct vnode *dvp, struct vnode **vpp, struct componentname *cnp)
1245 {
1246 	struct vnode *dp = NULL;		/* the directory we are searching */
1247 	int wantparent;			/* 1 => wantparent or lockparent flag */
1248 	int rdonly;			/* lookup read-only flag bit */
1249 	int error = 0;
1250 
1251 	KASSERT(cnp->cn_flags & ISLASTCN,
1252 	    ("relookup: Not given last component."));
1253 	/*
1254 	 * Setup: break out flag bits into variables.
1255 	 */
1256 	wantparent = cnp->cn_flags & (LOCKPARENT|WANTPARENT);
1257 	KASSERT(wantparent, ("relookup: parent not wanted."));
1258 	rdonly = cnp->cn_flags & RDONLY;
1259 	cnp->cn_flags &= ~ISSYMLINK;
1260 	dp = dvp;
1261 	cnp->cn_lkflags = LK_EXCLUSIVE;
1262 	vn_lock(dp, LK_EXCLUSIVE | LK_RETRY);
1263 
1264 	/*
1265 	 * Search a new directory.
1266 	 *
1267 	 * The last component of the filename is left accessible via
1268 	 * cnp->cn_nameptr for callers that need the name. Callers needing
1269 	 * the name set the SAVENAME flag. When done, they assume
1270 	 * responsibility for freeing the pathname buffer.
1271 	 */
1272 #ifdef NAMEI_DIAGNOSTIC
1273 	printf("{%s}: ", cnp->cn_nameptr);
1274 #endif
1275 
1276 	/*
1277 	 * Check for "" which represents the root directory after slash
1278 	 * removal.
1279 	 */
1280 	if (cnp->cn_nameptr[0] == '\0') {
1281 		/*
1282 		 * Support only LOOKUP for "/" because lookup()
1283 		 * can't succeed for CREATE, DELETE and RENAME.
1284 		 */
1285 		KASSERT(cnp->cn_nameiop == LOOKUP, ("nameiop must be LOOKUP"));
1286 		KASSERT(dp->v_type == VDIR, ("dp is not a directory"));
1287 
1288 		if (!(cnp->cn_flags & LOCKLEAF))
1289 			VOP_UNLOCK(dp);
1290 		*vpp = dp;
1291 		/* XXX This should probably move to the top of function. */
1292 		if (cnp->cn_flags & SAVESTART)
1293 			panic("lookup: SAVESTART");
1294 		return (0);
1295 	}
1296 
1297 	if (cnp->cn_flags & ISDOTDOT)
1298 		panic ("relookup: lookup on dot-dot");
1299 
1300 	/*
1301 	 * We now have a segment name to search for, and a directory to search.
1302 	 */
1303 #ifdef NAMEI_DIAGNOSTIC
1304 	vn_printf(dp, "search in ");
1305 #endif
1306 	if ((error = VOP_LOOKUP(dp, vpp, cnp)) != 0) {
1307 		KASSERT(*vpp == NULL, ("leaf should be empty"));
1308 		if (error != EJUSTRETURN)
1309 			goto bad;
1310 		/*
1311 		 * If creating and at end of pathname, then can consider
1312 		 * allowing file to be created.
1313 		 */
1314 		if (rdonly) {
1315 			error = EROFS;
1316 			goto bad;
1317 		}
1318 		/* ASSERT(dvp == ndp->ni_startdir) */
1319 		if (cnp->cn_flags & SAVESTART)
1320 			VREF(dvp);
1321 		if ((cnp->cn_flags & LOCKPARENT) == 0)
1322 			VOP_UNLOCK(dp);
1323 		/*
1324 		 * We return with ni_vp NULL to indicate that the entry
1325 		 * doesn't currently exist, leaving a pointer to the
1326 		 * (possibly locked) directory vnode in ndp->ni_dvp.
1327 		 */
1328 		return (0);
1329 	}
1330 
1331 	dp = *vpp;
1332 
1333 	/*
1334 	 * Disallow directory write attempts on read-only filesystems.
1335 	 */
1336 	if (rdonly &&
1337 	    (cnp->cn_nameiop == DELETE || cnp->cn_nameiop == RENAME)) {
1338 		if (dvp == dp)
1339 			vrele(dvp);
1340 		else
1341 			vput(dvp);
1342 		error = EROFS;
1343 		goto bad;
1344 	}
1345 	/*
1346 	 * Set the parent lock/ref state to the requested state.
1347 	 */
1348 	if ((cnp->cn_flags & LOCKPARENT) == 0 && dvp != dp) {
1349 		if (wantparent)
1350 			VOP_UNLOCK(dvp);
1351 		else
1352 			vput(dvp);
1353 	} else if (!wantparent)
1354 		vrele(dvp);
1355 	/*
1356 	 * Check for symbolic link
1357 	 */
1358 	KASSERT(dp->v_type != VLNK || !(cnp->cn_flags & FOLLOW),
1359 	    ("relookup: symlink found.\n"));
1360 
1361 	/* ASSERT(dvp == ndp->ni_startdir) */
1362 	if (cnp->cn_flags & SAVESTART)
1363 		VREF(dvp);
1364 
1365 	if ((cnp->cn_flags & LOCKLEAF) == 0)
1366 		VOP_UNLOCK(dp);
1367 	return (0);
1368 bad:
1369 	vput(dp);
1370 	*vpp = NULL;
1371 	return (error);
1372 }
1373 
1374 /*
1375  * Free data allocated by namei(); see namei(9) for details.
1376  */
1377 void
1378 NDFREE_PNBUF(struct nameidata *ndp)
1379 {
1380 
1381 	if ((ndp->ni_cnd.cn_flags & HASBUF) != 0) {
1382 		uma_zfree(namei_zone, ndp->ni_cnd.cn_pnbuf);
1383 		ndp->ni_cnd.cn_flags &= ~HASBUF;
1384 	}
1385 }
1386 
1387 void
1388 (NDFREE)(struct nameidata *ndp, const u_int flags)
1389 {
1390 	int unlock_dvp;
1391 	int unlock_vp;
1392 
1393 	unlock_dvp = 0;
1394 	unlock_vp = 0;
1395 
1396 	if (!(flags & NDF_NO_FREE_PNBUF)) {
1397 		NDFREE_PNBUF(ndp);
1398 	}
1399 	if (!(flags & NDF_NO_VP_UNLOCK) &&
1400 	    (ndp->ni_cnd.cn_flags & LOCKLEAF) && ndp->ni_vp)
1401 		unlock_vp = 1;
1402 	if (!(flags & NDF_NO_DVP_UNLOCK) &&
1403 	    (ndp->ni_cnd.cn_flags & LOCKPARENT) &&
1404 	    ndp->ni_dvp != ndp->ni_vp)
1405 		unlock_dvp = 1;
1406 	if (!(flags & NDF_NO_VP_RELE) && ndp->ni_vp) {
1407 		if (unlock_vp) {
1408 			vput(ndp->ni_vp);
1409 			unlock_vp = 0;
1410 		} else
1411 			vrele(ndp->ni_vp);
1412 		ndp->ni_vp = NULL;
1413 	}
1414 	if (unlock_vp)
1415 		VOP_UNLOCK(ndp->ni_vp);
1416 	if (!(flags & NDF_NO_DVP_RELE) &&
1417 	    (ndp->ni_cnd.cn_flags & (LOCKPARENT|WANTPARENT))) {
1418 		if (unlock_dvp) {
1419 			vput(ndp->ni_dvp);
1420 			unlock_dvp = 0;
1421 		} else
1422 			vrele(ndp->ni_dvp);
1423 		ndp->ni_dvp = NULL;
1424 	}
1425 	if (unlock_dvp)
1426 		VOP_UNLOCK(ndp->ni_dvp);
1427 	if (!(flags & NDF_NO_STARTDIR_RELE) &&
1428 	    (ndp->ni_cnd.cn_flags & SAVESTART)) {
1429 		vrele(ndp->ni_startdir);
1430 		ndp->ni_startdir = NULL;
1431 	}
1432 }
1433 
1434 /*
1435  * Determine if there is a suitable alternate filename under the specified
1436  * prefix for the specified path.  If the create flag is set, then the
1437  * alternate prefix will be used so long as the parent directory exists.
1438  * This is used by the various compatibility ABIs so that Linux binaries prefer
1439  * files under /compat/linux for example.  The chosen path (whether under
1440  * the prefix or under /) is returned in a kernel malloc'd buffer pointed
1441  * to by pathbuf.  The caller is responsible for free'ing the buffer from
1442  * the M_TEMP bucket if one is returned.
1443  */
1444 int
1445 kern_alternate_path(struct thread *td, const char *prefix, const char *path,
1446     enum uio_seg pathseg, char **pathbuf, int create, int dirfd)
1447 {
1448 	struct nameidata nd, ndroot;
1449 	char *ptr, *buf, *cp;
1450 	size_t len, sz;
1451 	int error;
1452 
1453 	buf = (char *) malloc(MAXPATHLEN, M_TEMP, M_WAITOK);
1454 	*pathbuf = buf;
1455 
1456 	/* Copy the prefix into the new pathname as a starting point. */
1457 	len = strlcpy(buf, prefix, MAXPATHLEN);
1458 	if (len >= MAXPATHLEN) {
1459 		*pathbuf = NULL;
1460 		free(buf, M_TEMP);
1461 		return (EINVAL);
1462 	}
1463 	sz = MAXPATHLEN - len;
1464 	ptr = buf + len;
1465 
1466 	/* Append the filename to the prefix. */
1467 	if (pathseg == UIO_SYSSPACE)
1468 		error = copystr(path, ptr, sz, &len);
1469 	else
1470 		error = copyinstr(path, ptr, sz, &len);
1471 
1472 	if (error) {
1473 		*pathbuf = NULL;
1474 		free(buf, M_TEMP);
1475 		return (error);
1476 	}
1477 
1478 	/* Only use a prefix with absolute pathnames. */
1479 	if (*ptr != '/') {
1480 		error = EINVAL;
1481 		goto keeporig;
1482 	}
1483 
1484 	if (dirfd != AT_FDCWD) {
1485 		/*
1486 		 * We want the original because the "prefix" is
1487 		 * included in the already opened dirfd.
1488 		 */
1489 		bcopy(ptr, buf, len);
1490 		return (0);
1491 	}
1492 
1493 	/*
1494 	 * We know that there is a / somewhere in this pathname.
1495 	 * Search backwards for it, to find the file's parent dir
1496 	 * to see if it exists in the alternate tree. If it does,
1497 	 * and we want to create a file (cflag is set). We don't
1498 	 * need to worry about the root comparison in this case.
1499 	 */
1500 
1501 	if (create) {
1502 		for (cp = &ptr[len] - 1; *cp != '/'; cp--);
1503 		*cp = '\0';
1504 
1505 		NDINIT(&nd, LOOKUP, NOFOLLOW, UIO_SYSSPACE, buf, td);
1506 		error = namei(&nd);
1507 		*cp = '/';
1508 		if (error != 0)
1509 			goto keeporig;
1510 	} else {
1511 		NDINIT(&nd, LOOKUP, NOFOLLOW, UIO_SYSSPACE, buf, td);
1512 
1513 		error = namei(&nd);
1514 		if (error != 0)
1515 			goto keeporig;
1516 
1517 		/*
1518 		 * We now compare the vnode of the prefix to the one
1519 		 * vnode asked. If they resolve to be the same, then we
1520 		 * ignore the match so that the real root gets used.
1521 		 * This avoids the problem of traversing "../.." to find the
1522 		 * root directory and never finding it, because "/" resolves
1523 		 * to the emulation root directory. This is expensive :-(
1524 		 */
1525 		NDINIT(&ndroot, LOOKUP, FOLLOW, UIO_SYSSPACE, prefix,
1526 		    td);
1527 
1528 		/* We shouldn't ever get an error from this namei(). */
1529 		error = namei(&ndroot);
1530 		if (error == 0) {
1531 			if (nd.ni_vp == ndroot.ni_vp)
1532 				error = ENOENT;
1533 
1534 			NDFREE(&ndroot, NDF_ONLY_PNBUF);
1535 			vrele(ndroot.ni_vp);
1536 		}
1537 	}
1538 
1539 	NDFREE(&nd, NDF_ONLY_PNBUF);
1540 	vrele(nd.ni_vp);
1541 
1542 keeporig:
1543 	/* If there was an error, use the original path name. */
1544 	if (error)
1545 		bcopy(ptr, buf, len);
1546 	return (error);
1547 }
1548