xref: /freebsd/sys/kern/vfs_lookup.c (revision 28f4385e45a2681c14bd04b83fe1796eaefe8265)
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 
65 #include <security/audit/audit.h>
66 #include <security/mac/mac_framework.h>
67 
68 #include <vm/uma.h>
69 
70 #define	NAMEI_DIAGNOSTIC 1
71 #undef NAMEI_DIAGNOSTIC
72 
73 SDT_PROVIDER_DECLARE(vfs);
74 SDT_PROBE_DEFINE3(vfs, namei, lookup, entry, "struct vnode *", "char *",
75     "unsigned long");
76 SDT_PROBE_DEFINE2(vfs, namei, lookup, return, "int", "struct vnode *");
77 
78 /* Allocation zone for namei. */
79 uma_zone_t namei_zone;
80 
81 /* Placeholder vnode for mp traversal. */
82 static struct vnode *vp_crossmp;
83 
84 static int
85 crossmp_vop_islocked(struct vop_islocked_args *ap)
86 {
87 
88 	return (LK_SHARED);
89 }
90 
91 static int
92 crossmp_vop_lock1(struct vop_lock1_args *ap)
93 {
94 	struct vnode *vp;
95 	struct lock *lk __unused;
96 	const char *file __unused;
97 	int flags, line __unused;
98 
99 	vp = ap->a_vp;
100 	lk = vp->v_vnlock;
101 	flags = ap->a_flags;
102 	file = ap->a_file;
103 	line = ap->a_line;
104 
105 	if ((flags & LK_SHARED) == 0)
106 		panic("invalid lock request for crossmp");
107 
108 	WITNESS_CHECKORDER(&lk->lock_object, LOP_NEWORDER, file, line,
109 	    flags & LK_INTERLOCK ? &VI_MTX(vp)->lock_object : NULL);
110 	WITNESS_LOCK(&lk->lock_object, 0, file, line);
111 	if ((flags & LK_INTERLOCK) != 0)
112 		VI_UNLOCK(vp);
113 	LOCK_LOG_LOCK("SLOCK", &lk->lock_object, 0, 0, ap->a_file, line);
114 	return (0);
115 }
116 
117 static int
118 crossmp_vop_unlock(struct vop_unlock_args *ap)
119 {
120 	struct vnode *vp;
121 	struct lock *lk __unused;
122 	int flags;
123 
124 	vp = ap->a_vp;
125 	lk = vp->v_vnlock;
126 	flags = ap->a_flags;
127 
128 	if ((flags & LK_INTERLOCK) != 0)
129 		VI_UNLOCK(vp);
130 	WITNESS_UNLOCK(&lk->lock_object, 0, LOCK_FILE, LOCK_LINE);
131 	LOCK_LOG_LOCK("SUNLOCK", &lk->lock_object, 0, 0, LOCK_FILE,
132 	    LOCK_LINE);
133 	return (0);
134 }
135 
136 static struct vop_vector crossmp_vnodeops = {
137 	.vop_default =		&default_vnodeops,
138 	.vop_islocked =		crossmp_vop_islocked,
139 	.vop_lock1 =		crossmp_vop_lock1,
140 	.vop_unlock =		crossmp_vop_unlock,
141 };
142 
143 struct nameicap_tracker {
144 	struct vnode *dp;
145 	TAILQ_ENTRY(nameicap_tracker) nm_link;
146 };
147 
148 /* Zone for cap mode tracker elements used for dotdot capability checks. */
149 static uma_zone_t nt_zone;
150 
151 static void
152 nameiinit(void *dummy __unused)
153 {
154 
155 	namei_zone = uma_zcreate("NAMEI", MAXPATHLEN, NULL, NULL, NULL, NULL,
156 	    UMA_ALIGN_PTR, 0);
157 	nt_zone = uma_zcreate("rentr", sizeof(struct nameicap_tracker),
158 	    NULL, NULL, NULL, NULL, UMA_ALIGN_PTR, 0);
159 	getnewvnode("crossmp", NULL, &crossmp_vnodeops, &vp_crossmp);
160 }
161 SYSINIT(vfs, SI_SUB_VFS, SI_ORDER_SECOND, nameiinit, NULL);
162 
163 static int lookup_cap_dotdot = 1;
164 SYSCTL_INT(_vfs, OID_AUTO, lookup_cap_dotdot, CTLFLAG_RWTUN,
165     &lookup_cap_dotdot, 0,
166     "enables \"..\" components in path lookup in capability mode");
167 static int lookup_cap_dotdot_nonlocal = 1;
168 SYSCTL_INT(_vfs, OID_AUTO, lookup_cap_dotdot_nonlocal, CTLFLAG_RWTUN,
169     &lookup_cap_dotdot_nonlocal, 0,
170     "enables \"..\" components in path lookup in capability mode "
171     "on non-local mount");
172 
173 static void
174 nameicap_tracker_add(struct nameidata *ndp, struct vnode *dp)
175 {
176 	struct nameicap_tracker *nt;
177 
178 	if ((ndp->ni_lcf & NI_LCF_CAP_DOTDOT) == 0 || dp->v_type != VDIR)
179 		return;
180 	if ((ndp->ni_lcf & (NI_LCF_BENEATH_ABS | NI_LCF_BENEATH_LATCHED)) ==
181 	    NI_LCF_BENEATH_ABS) {
182 		MPASS((ndp->ni_lcf & NI_LCF_LATCH) != 0);
183 		if (dp != ndp->ni_beneath_latch)
184 			return;
185 		ndp->ni_lcf |= NI_LCF_BENEATH_LATCHED;
186 	}
187 	nt = uma_zalloc(nt_zone, M_WAITOK);
188 	vhold(dp);
189 	nt->dp = dp;
190 	TAILQ_INSERT_TAIL(&ndp->ni_cap_tracker, nt, nm_link);
191 }
192 
193 static void
194 nameicap_cleanup(struct nameidata *ndp, bool clean_latch)
195 {
196 	struct nameicap_tracker *nt, *nt1;
197 
198 	KASSERT(TAILQ_EMPTY(&ndp->ni_cap_tracker) ||
199 	    (ndp->ni_lcf & NI_LCF_CAP_DOTDOT) != 0, ("not strictrelative"));
200 	TAILQ_FOREACH_SAFE(nt, &ndp->ni_cap_tracker, nm_link, nt1) {
201 		TAILQ_REMOVE(&ndp->ni_cap_tracker, nt, nm_link);
202 		vdrop(nt->dp);
203 		uma_zfree(nt_zone, nt);
204 	}
205 	if (clean_latch && (ndp->ni_lcf & NI_LCF_LATCH) != 0) {
206 		ndp->ni_lcf &= ~NI_LCF_LATCH;
207 		vrele(ndp->ni_beneath_latch);
208 	}
209 }
210 
211 /*
212  * For dotdot lookups in capability mode, only allow the component
213  * lookup to succeed if the resulting directory was already traversed
214  * during the operation.  Also fail dotdot lookups for non-local
215  * filesystems, where external agents might assist local lookups to
216  * escape the compartment.
217  */
218 static int
219 nameicap_check_dotdot(struct nameidata *ndp, struct vnode *dp)
220 {
221 	struct nameicap_tracker *nt;
222 	struct mount *mp;
223 
224 	if ((ndp->ni_lcf & NI_LCF_CAP_DOTDOT) == 0 || dp == NULL ||
225 	    dp->v_type != VDIR)
226 		return (0);
227 	mp = dp->v_mount;
228 	if (lookup_cap_dotdot_nonlocal == 0 && mp != NULL &&
229 	    (mp->mnt_flag & MNT_LOCAL) == 0)
230 		return (ENOTCAPABLE);
231 	TAILQ_FOREACH_REVERSE(nt, &ndp->ni_cap_tracker, nameicap_tracker_head,
232 	    nm_link) {
233 		if (dp == nt->dp)
234 			return (0);
235 	}
236 	if ((ndp->ni_lcf & NI_LCF_BENEATH_ABS) != 0) {
237 		ndp->ni_lcf &= ~NI_LCF_BENEATH_LATCHED;
238 		nameicap_cleanup(ndp, false);
239 		return (0);
240 	}
241 	return (ENOTCAPABLE);
242 }
243 
244 static void
245 namei_cleanup_cnp(struct componentname *cnp)
246 {
247 
248 	uma_zfree(namei_zone, cnp->cn_pnbuf);
249 #ifdef DIAGNOSTIC
250 	cnp->cn_pnbuf = NULL;
251 	cnp->cn_nameptr = NULL;
252 #endif
253 }
254 
255 static int
256 namei_handle_root(struct nameidata *ndp, struct vnode **dpp)
257 {
258 	struct componentname *cnp;
259 
260 	cnp = &ndp->ni_cnd;
261 	if ((ndp->ni_lcf & NI_LCF_STRICTRELATIVE) != 0) {
262 #ifdef KTRACE
263 		if (KTRPOINT(curthread, KTR_CAPFAIL))
264 			ktrcapfail(CAPFAIL_LOOKUP, NULL, NULL);
265 #endif
266 		return (ENOTCAPABLE);
267 	}
268 	if ((cnp->cn_flags & BENEATH) != 0) {
269 		ndp->ni_lcf |= NI_LCF_BENEATH_ABS;
270 		ndp->ni_lcf &= ~NI_LCF_BENEATH_LATCHED;
271 		nameicap_cleanup(ndp, false);
272 	}
273 	while (*(cnp->cn_nameptr) == '/') {
274 		cnp->cn_nameptr++;
275 		ndp->ni_pathlen--;
276 	}
277 	*dpp = ndp->ni_rootdir;
278 	vrefact(*dpp);
279 	return (0);
280 }
281 
282 /*
283  * Convert a pathname into a pointer to a locked vnode.
284  *
285  * The FOLLOW flag is set when symbolic links are to be followed
286  * when they occur at the end of the name translation process.
287  * Symbolic links are always followed for all other pathname
288  * components other than the last.
289  *
290  * The segflg defines whether the name is to be copied from user
291  * space or kernel space.
292  *
293  * Overall outline of namei:
294  *
295  *	copy in name
296  *	get starting directory
297  *	while (!done && !error) {
298  *		call lookup to search path.
299  *		if symbolic link, massage name in buffer and continue
300  *	}
301  */
302 int
303 namei(struct nameidata *ndp)
304 {
305 	struct filedesc *fdp;	/* pointer to file descriptor state */
306 	char *cp;		/* pointer into pathname argument */
307 	struct vnode *dp;	/* the directory we are searching */
308 	struct iovec aiov;		/* uio for reading symbolic links */
309 	struct componentname *cnp;
310 	struct thread *td;
311 	struct proc *p;
312 	cap_rights_t rights;
313 	struct filecaps dirfd_caps;
314 	struct uio auio;
315 	int error, linklen, startdir_used;
316 
317 	cnp = &ndp->ni_cnd;
318 	td = cnp->cn_thread;
319 	p = td->td_proc;
320 	ndp->ni_cnd.cn_cred = ndp->ni_cnd.cn_thread->td_ucred;
321 	KASSERT(cnp->cn_cred && p, ("namei: bad cred/proc"));
322 	KASSERT((cnp->cn_nameiop & (~OPMASK)) == 0,
323 	    ("namei: nameiop contaminated with flags"));
324 	KASSERT((cnp->cn_flags & OPMASK) == 0,
325 	    ("namei: flags contaminated with nameiops"));
326 	MPASS(ndp->ni_startdir == NULL || ndp->ni_startdir->v_type == VDIR ||
327 	    ndp->ni_startdir->v_type == VBAD);
328 	fdp = p->p_fd;
329 	TAILQ_INIT(&ndp->ni_cap_tracker);
330 	ndp->ni_lcf = 0;
331 
332 	/* We will set this ourselves if we need it. */
333 	cnp->cn_flags &= ~TRAILINGSLASH;
334 
335 	/*
336 	 * Get a buffer for the name to be translated, and copy the
337 	 * name into the buffer.
338 	 */
339 	if ((cnp->cn_flags & HASBUF) == 0)
340 		cnp->cn_pnbuf = uma_zalloc(namei_zone, M_WAITOK);
341 	if (ndp->ni_segflg == UIO_SYSSPACE)
342 		error = copystr(ndp->ni_dirp, cnp->cn_pnbuf, MAXPATHLEN,
343 		    &ndp->ni_pathlen);
344 	else
345 		error = copyinstr(ndp->ni_dirp, cnp->cn_pnbuf, MAXPATHLEN,
346 		    &ndp->ni_pathlen);
347 
348 	/*
349 	 * Don't allow empty pathnames.
350 	 */
351 	if (error == 0 && *cnp->cn_pnbuf == '\0')
352 		error = ENOENT;
353 
354 #ifdef CAPABILITY_MODE
355 	/*
356 	 * In capability mode, lookups must be restricted to happen in
357 	 * the subtree with the root specified by the file descriptor:
358 	 * - The root must be real file descriptor, not the pseudo-descriptor
359 	 *   AT_FDCWD.
360 	 * - The passed path must be relative and not absolute.
361 	 * - If lookup_cap_dotdot is disabled, path must not contain the
362 	 *   '..' components.
363 	 * - If lookup_cap_dotdot is enabled, we verify that all '..'
364 	 *   components lookups result in the directories which were
365 	 *   previously walked by us, which prevents an escape from
366 	 *   the relative root.
367 	 */
368 	if (error == 0 && IN_CAPABILITY_MODE(td) &&
369 	    (cnp->cn_flags & NOCAPCHECK) == 0) {
370 		ndp->ni_lcf |= NI_LCF_STRICTRELATIVE;
371 		if (ndp->ni_dirfd == AT_FDCWD) {
372 #ifdef KTRACE
373 			if (KTRPOINT(td, KTR_CAPFAIL))
374 				ktrcapfail(CAPFAIL_LOOKUP, NULL, NULL);
375 #endif
376 			error = ECAPMODE;
377 		}
378 	}
379 #endif
380 	if (error != 0) {
381 		namei_cleanup_cnp(cnp);
382 		ndp->ni_vp = NULL;
383 		return (error);
384 	}
385 	ndp->ni_loopcnt = 0;
386 #ifdef KTRACE
387 	if (KTRPOINT(td, KTR_NAMEI)) {
388 		KASSERT(cnp->cn_thread == curthread,
389 		    ("namei not using curthread"));
390 		ktrnamei(cnp->cn_pnbuf);
391 	}
392 #endif
393 	/*
394 	 * Get starting point for the translation.
395 	 */
396 	FILEDESC_SLOCK(fdp);
397 	ndp->ni_rootdir = fdp->fd_rdir;
398 	vrefact(ndp->ni_rootdir);
399 	ndp->ni_topdir = fdp->fd_jdir;
400 
401 	/*
402 	 * If we are auditing the kernel pathname, save the user pathname.
403 	 */
404 	if (cnp->cn_flags & AUDITVNODE1)
405 		AUDIT_ARG_UPATH1(td, ndp->ni_dirfd, cnp->cn_pnbuf);
406 	if (cnp->cn_flags & AUDITVNODE2)
407 		AUDIT_ARG_UPATH2(td, ndp->ni_dirfd, cnp->cn_pnbuf);
408 
409 	startdir_used = 0;
410 	dp = NULL;
411 	cnp->cn_nameptr = cnp->cn_pnbuf;
412 	if (cnp->cn_pnbuf[0] == '/') {
413 		error = namei_handle_root(ndp, &dp);
414 	} else {
415 		if (ndp->ni_startdir != NULL) {
416 			dp = ndp->ni_startdir;
417 			startdir_used = 1;
418 		} else if (ndp->ni_dirfd == AT_FDCWD) {
419 			dp = fdp->fd_cdir;
420 			vrefact(dp);
421 		} else {
422 			rights = ndp->ni_rightsneeded;
423 			cap_rights_set(&rights, CAP_LOOKUP);
424 
425 			if (cnp->cn_flags & AUDITVNODE1)
426 				AUDIT_ARG_ATFD1(ndp->ni_dirfd);
427 			if (cnp->cn_flags & AUDITVNODE2)
428 				AUDIT_ARG_ATFD2(ndp->ni_dirfd);
429 			error = fgetvp_rights(td, ndp->ni_dirfd,
430 			    &rights, &ndp->ni_filecaps, &dp);
431 			if (error == EINVAL)
432 				error = ENOTDIR;
433 #ifdef CAPABILITIES
434 			/*
435 			 * If file descriptor doesn't have all rights,
436 			 * all lookups relative to it must also be
437 			 * strictly relative.
438 			 */
439 			CAP_ALL(&rights);
440 			if (!cap_rights_contains(&ndp->ni_filecaps.fc_rights,
441 			    &rights) ||
442 			    ndp->ni_filecaps.fc_fcntls != CAP_FCNTL_ALL ||
443 			    ndp->ni_filecaps.fc_nioctls != -1) {
444 				ndp->ni_lcf |= NI_LCF_STRICTRELATIVE;
445 			}
446 #endif
447 		}
448 		if (error == 0 && dp->v_type != VDIR)
449 			error = ENOTDIR;
450 	}
451 	if (error == 0 && (cnp->cn_flags & BENEATH) != 0) {
452 		if (ndp->ni_dirfd == AT_FDCWD) {
453 			ndp->ni_beneath_latch = fdp->fd_cdir;
454 			vrefact(ndp->ni_beneath_latch);
455 		} else {
456 			rights = ndp->ni_rightsneeded;
457 			cap_rights_set(&rights, CAP_LOOKUP);
458 			error = fgetvp_rights(td, ndp->ni_dirfd, &rights,
459 			    &dirfd_caps, &ndp->ni_beneath_latch);
460 			if (error == 0 && dp->v_type != VDIR) {
461 				vrele(ndp->ni_beneath_latch);
462 				error = ENOTDIR;
463 			}
464 		}
465 		if (error == 0)
466 			ndp->ni_lcf |= NI_LCF_LATCH;
467 	}
468 	FILEDESC_SUNLOCK(fdp);
469 	if (ndp->ni_startdir != NULL && !startdir_used)
470 		vrele(ndp->ni_startdir);
471 	if (error != 0) {
472 		if (dp != NULL)
473 			vrele(dp);
474 		goto out;
475 	}
476 	MPASS((ndp->ni_lcf & (NI_LCF_BENEATH_ABS | NI_LCF_LATCH)) !=
477 	    NI_LCF_BENEATH_ABS);
478 	if (((ndp->ni_lcf & NI_LCF_STRICTRELATIVE) != 0 &&
479 	    lookup_cap_dotdot != 0) ||
480 	    ((ndp->ni_lcf & NI_LCF_STRICTRELATIVE) == 0 &&
481 	    (cnp->cn_flags & BENEATH) != 0))
482 		ndp->ni_lcf |= NI_LCF_CAP_DOTDOT;
483 	SDT_PROBE3(vfs, namei, lookup, entry, dp, cnp->cn_pnbuf,
484 	    cnp->cn_flags);
485 	for (;;) {
486 		ndp->ni_startdir = dp;
487 		error = lookup(ndp);
488 		if (error != 0)
489 			goto out;
490 		/*
491 		 * If not a symbolic link, we're done.
492 		 */
493 		if ((cnp->cn_flags & ISSYMLINK) == 0) {
494 			vrele(ndp->ni_rootdir);
495 			if ((cnp->cn_flags & (SAVENAME | SAVESTART)) == 0) {
496 				namei_cleanup_cnp(cnp);
497 			} else
498 				cnp->cn_flags |= HASBUF;
499 			if ((ndp->ni_lcf & (NI_LCF_BENEATH_ABS |
500 			    NI_LCF_BENEATH_LATCHED)) == NI_LCF_BENEATH_ABS) {
501 				NDFREE(ndp, 0);
502 				error = ENOTCAPABLE;
503 			}
504 			nameicap_cleanup(ndp, true);
505 			SDT_PROBE2(vfs, namei, lookup, return, error,
506 			    (error == 0 ? ndp->ni_vp : NULL));
507 			return (error);
508 		}
509 		if (ndp->ni_loopcnt++ >= MAXSYMLINKS) {
510 			error = ELOOP;
511 			break;
512 		}
513 #ifdef MAC
514 		if ((cnp->cn_flags & NOMACCHECK) == 0) {
515 			error = mac_vnode_check_readlink(td->td_ucred,
516 			    ndp->ni_vp);
517 			if (error != 0)
518 				break;
519 		}
520 #endif
521 		if (ndp->ni_pathlen > 1)
522 			cp = uma_zalloc(namei_zone, M_WAITOK);
523 		else
524 			cp = cnp->cn_pnbuf;
525 		aiov.iov_base = cp;
526 		aiov.iov_len = MAXPATHLEN;
527 		auio.uio_iov = &aiov;
528 		auio.uio_iovcnt = 1;
529 		auio.uio_offset = 0;
530 		auio.uio_rw = UIO_READ;
531 		auio.uio_segflg = UIO_SYSSPACE;
532 		auio.uio_td = td;
533 		auio.uio_resid = MAXPATHLEN;
534 		error = VOP_READLINK(ndp->ni_vp, &auio, cnp->cn_cred);
535 		if (error != 0) {
536 			if (ndp->ni_pathlen > 1)
537 				uma_zfree(namei_zone, cp);
538 			break;
539 		}
540 		linklen = MAXPATHLEN - auio.uio_resid;
541 		if (linklen == 0) {
542 			if (ndp->ni_pathlen > 1)
543 				uma_zfree(namei_zone, cp);
544 			error = ENOENT;
545 			break;
546 		}
547 		if (linklen + ndp->ni_pathlen > MAXPATHLEN) {
548 			if (ndp->ni_pathlen > 1)
549 				uma_zfree(namei_zone, cp);
550 			error = ENAMETOOLONG;
551 			break;
552 		}
553 		if (ndp->ni_pathlen > 1) {
554 			bcopy(ndp->ni_next, cp + linklen, ndp->ni_pathlen);
555 			uma_zfree(namei_zone, cnp->cn_pnbuf);
556 			cnp->cn_pnbuf = cp;
557 		} else
558 			cnp->cn_pnbuf[linklen] = '\0';
559 		ndp->ni_pathlen += linklen;
560 		vput(ndp->ni_vp);
561 		dp = ndp->ni_dvp;
562 		/*
563 		 * Check if root directory should replace current directory.
564 		 */
565 		cnp->cn_nameptr = cnp->cn_pnbuf;
566 		if (*(cnp->cn_nameptr) == '/') {
567 			vrele(dp);
568 			error = namei_handle_root(ndp, &dp);
569 			if (error != 0)
570 				goto out;
571 		}
572 	}
573 	vput(ndp->ni_vp);
574 	ndp->ni_vp = NULL;
575 	vrele(ndp->ni_dvp);
576 out:
577 	vrele(ndp->ni_rootdir);
578 	MPASS(error != 0);
579 	namei_cleanup_cnp(cnp);
580 	nameicap_cleanup(ndp, true);
581 	SDT_PROBE2(vfs, namei, lookup, return, error, NULL);
582 	return (error);
583 }
584 
585 static int
586 compute_cn_lkflags(struct mount *mp, int lkflags, int cnflags)
587 {
588 
589 	if (mp == NULL || ((lkflags & LK_SHARED) &&
590 	    (!(mp->mnt_kern_flag & MNTK_LOOKUP_SHARED) ||
591 	    ((cnflags & ISDOTDOT) &&
592 	    (mp->mnt_kern_flag & MNTK_LOOKUP_EXCL_DOTDOT))))) {
593 		lkflags &= ~LK_SHARED;
594 		lkflags |= LK_EXCLUSIVE;
595 	}
596 	lkflags |= LK_NODDLKTREAT;
597 	return (lkflags);
598 }
599 
600 static __inline int
601 needs_exclusive_leaf(struct mount *mp, int flags)
602 {
603 
604 	/*
605 	 * Intermediate nodes can use shared locks, we only need to
606 	 * force an exclusive lock for leaf nodes.
607 	 */
608 	if ((flags & (ISLASTCN | LOCKLEAF)) != (ISLASTCN | LOCKLEAF))
609 		return (0);
610 
611 	/* Always use exclusive locks if LOCKSHARED isn't set. */
612 	if (!(flags & LOCKSHARED))
613 		return (1);
614 
615 	/*
616 	 * For lookups during open(), if the mount point supports
617 	 * extended shared operations, then use a shared lock for the
618 	 * leaf node, otherwise use an exclusive lock.
619 	 */
620 	if ((flags & ISOPEN) != 0)
621 		return (!MNT_EXTENDED_SHARED(mp));
622 
623 	/*
624 	 * Lookup requests outside of open() that specify LOCKSHARED
625 	 * only need a shared lock on the leaf vnode.
626 	 */
627 	return (0);
628 }
629 
630 /*
631  * Search a pathname.
632  * This is a very central and rather complicated routine.
633  *
634  * The pathname is pointed to by ni_ptr and is of length ni_pathlen.
635  * The starting directory is taken from ni_startdir. The pathname is
636  * descended until done, or a symbolic link is encountered. The variable
637  * ni_more is clear if the path is completed; it is set to one if a
638  * symbolic link needing interpretation is encountered.
639  *
640  * The flag argument is LOOKUP, CREATE, RENAME, or DELETE depending on
641  * whether the name is to be looked up, created, renamed, or deleted.
642  * When CREATE, RENAME, or DELETE is specified, information usable in
643  * creating, renaming, or deleting a directory entry may be calculated.
644  * If flag has LOCKPARENT or'ed into it, the parent directory is returned
645  * locked. If flag has WANTPARENT or'ed into it, the parent directory is
646  * returned unlocked. Otherwise the parent directory is not returned. If
647  * the target of the pathname exists and LOCKLEAF is or'ed into the flag
648  * the target is returned locked, otherwise it is returned unlocked.
649  * When creating or renaming and LOCKPARENT is specified, the target may not
650  * be ".".  When deleting and LOCKPARENT is specified, the target may be ".".
651  *
652  * Overall outline of lookup:
653  *
654  * dirloop:
655  *	identify next component of name at ndp->ni_ptr
656  *	handle degenerate case where name is null string
657  *	if .. and crossing mount points and on mounted filesys, find parent
658  *	call VOP_LOOKUP routine for next component name
659  *	    directory vnode returned in ni_dvp, unlocked unless LOCKPARENT set
660  *	    component vnode returned in ni_vp (if it exists), locked.
661  *	if result vnode is mounted on and crossing mount points,
662  *	    find mounted on vnode
663  *	if more components of name, do next level at dirloop
664  *	return the answer in ni_vp, locked if LOCKLEAF set
665  *	    if LOCKPARENT set, return locked parent in ni_dvp
666  *	    if WANTPARENT set, return unlocked parent in ni_dvp
667  */
668 int
669 lookup(struct nameidata *ndp)
670 {
671 	char *cp;			/* pointer into pathname argument */
672 	char *prev_ni_next;		/* saved ndp->ni_next */
673 	struct vnode *dp = NULL;	/* the directory we are searching */
674 	struct vnode *tdp;		/* saved dp */
675 	struct mount *mp;		/* mount table entry */
676 	struct prison *pr;
677 	size_t prev_ni_pathlen;		/* saved ndp->ni_pathlen */
678 	int docache;			/* == 0 do not cache last component */
679 	int wantparent;			/* 1 => wantparent or lockparent flag */
680 	int rdonly;			/* lookup read-only flag bit */
681 	int error = 0;
682 	int dpunlocked = 0;		/* dp has already been unlocked */
683 	int relookup = 0;		/* do not consume the path component */
684 	struct componentname *cnp = &ndp->ni_cnd;
685 	int lkflags_save;
686 	int ni_dvp_unlocked;
687 
688 	/*
689 	 * Setup: break out flag bits into variables.
690 	 */
691 	ni_dvp_unlocked = 0;
692 	wantparent = cnp->cn_flags & (LOCKPARENT | WANTPARENT);
693 	KASSERT(cnp->cn_nameiop == LOOKUP || wantparent,
694 	    ("CREATE, DELETE, RENAME require LOCKPARENT or WANTPARENT."));
695 	docache = (cnp->cn_flags & NOCACHE) ^ NOCACHE;
696 	if (cnp->cn_nameiop == DELETE ||
697 	    (wantparent && cnp->cn_nameiop != CREATE &&
698 	     cnp->cn_nameiop != LOOKUP))
699 		docache = 0;
700 	rdonly = cnp->cn_flags & RDONLY;
701 	cnp->cn_flags &= ~ISSYMLINK;
702 	ndp->ni_dvp = NULL;
703 	/*
704 	 * We use shared locks until we hit the parent of the last cn then
705 	 * we adjust based on the requesting flags.
706 	 */
707 	cnp->cn_lkflags = LK_SHARED;
708 	dp = ndp->ni_startdir;
709 	ndp->ni_startdir = NULLVP;
710 	vn_lock(dp,
711 	    compute_cn_lkflags(dp->v_mount, cnp->cn_lkflags | LK_RETRY,
712 	    cnp->cn_flags));
713 
714 dirloop:
715 	/*
716 	 * Search a new directory.
717 	 *
718 	 * The last component of the filename is left accessible via
719 	 * cnp->cn_nameptr for callers that need the name. Callers needing
720 	 * the name set the SAVENAME flag. When done, they assume
721 	 * responsibility for freeing the pathname buffer.
722 	 */
723 	for (cp = cnp->cn_nameptr; *cp != 0 && *cp != '/'; cp++)
724 		continue;
725 	cnp->cn_namelen = cp - cnp->cn_nameptr;
726 	if (cnp->cn_namelen > NAME_MAX) {
727 		error = ENAMETOOLONG;
728 		goto bad;
729 	}
730 #ifdef NAMEI_DIAGNOSTIC
731 	{ char c = *cp;
732 	*cp = '\0';
733 	printf("{%s}: ", cnp->cn_nameptr);
734 	*cp = c; }
735 #endif
736 	prev_ni_pathlen = ndp->ni_pathlen;
737 	ndp->ni_pathlen -= cnp->cn_namelen;
738 	KASSERT(ndp->ni_pathlen <= PATH_MAX,
739 	    ("%s: ni_pathlen underflow to %zd\n", __func__, ndp->ni_pathlen));
740 	prev_ni_next = ndp->ni_next;
741 	ndp->ni_next = cp;
742 
743 	/*
744 	 * Replace multiple slashes by a single slash and trailing slashes
745 	 * by a null.  This must be done before VOP_LOOKUP() because some
746 	 * fs's don't know about trailing slashes.  Remember if there were
747 	 * trailing slashes to handle symlinks, existing non-directories
748 	 * and non-existing files that won't be directories specially later.
749 	 */
750 	while (*cp == '/' && (cp[1] == '/' || cp[1] == '\0')) {
751 		cp++;
752 		ndp->ni_pathlen--;
753 		if (*cp == '\0') {
754 			*ndp->ni_next = '\0';
755 			cnp->cn_flags |= TRAILINGSLASH;
756 		}
757 	}
758 	ndp->ni_next = cp;
759 
760 	cnp->cn_flags |= MAKEENTRY;
761 	if (*cp == '\0' && docache == 0)
762 		cnp->cn_flags &= ~MAKEENTRY;
763 	if (cnp->cn_namelen == 2 &&
764 	    cnp->cn_nameptr[1] == '.' && cnp->cn_nameptr[0] == '.')
765 		cnp->cn_flags |= ISDOTDOT;
766 	else
767 		cnp->cn_flags &= ~ISDOTDOT;
768 	if (*ndp->ni_next == 0)
769 		cnp->cn_flags |= ISLASTCN;
770 	else
771 		cnp->cn_flags &= ~ISLASTCN;
772 
773 	if ((cnp->cn_flags & ISLASTCN) != 0 &&
774 	    cnp->cn_namelen == 1 && cnp->cn_nameptr[0] == '.' &&
775 	    (cnp->cn_nameiop == DELETE || cnp->cn_nameiop == RENAME)) {
776 		error = EINVAL;
777 		goto bad;
778 	}
779 
780 	nameicap_tracker_add(ndp, dp);
781 
782 	/*
783 	 * Check for degenerate name (e.g. / or "")
784 	 * which is a way of talking about a directory,
785 	 * e.g. like "/." or ".".
786 	 */
787 	if (cnp->cn_nameptr[0] == '\0') {
788 		if (dp->v_type != VDIR) {
789 			error = ENOTDIR;
790 			goto bad;
791 		}
792 		if (cnp->cn_nameiop != LOOKUP) {
793 			error = EISDIR;
794 			goto bad;
795 		}
796 		if (wantparent) {
797 			ndp->ni_dvp = dp;
798 			VREF(dp);
799 		}
800 		ndp->ni_vp = dp;
801 
802 		if (cnp->cn_flags & AUDITVNODE1)
803 			AUDIT_ARG_VNODE1(dp);
804 		else if (cnp->cn_flags & AUDITVNODE2)
805 			AUDIT_ARG_VNODE2(dp);
806 
807 		if (!(cnp->cn_flags & (LOCKPARENT | LOCKLEAF)))
808 			VOP_UNLOCK(dp, 0);
809 		/* XXX This should probably move to the top of function. */
810 		if (cnp->cn_flags & SAVESTART)
811 			panic("lookup: SAVESTART");
812 		goto success;
813 	}
814 
815 	/*
816 	 * Handle "..": five special cases.
817 	 * 0. If doing a capability lookup and lookup_cap_dotdot is
818 	 *    disabled, return ENOTCAPABLE.
819 	 * 1. Return an error if this is the last component of
820 	 *    the name and the operation is DELETE or RENAME.
821 	 * 2. If at root directory (e.g. after chroot)
822 	 *    or at absolute root directory
823 	 *    then ignore it so can't get out.
824 	 * 3. If this vnode is the root of a mounted
825 	 *    filesystem, then replace it with the
826 	 *    vnode which was mounted on so we take the
827 	 *    .. in the other filesystem.
828 	 * 4. If the vnode is the top directory of
829 	 *    the jail or chroot, don't let them out.
830 	 * 5. If doing a capability lookup and lookup_cap_dotdot is
831 	 *    enabled, return ENOTCAPABLE if the lookup would escape
832 	 *    from the initial file descriptor directory.  Checks are
833 	 *    done by ensuring that namei() already traversed the
834 	 *    result of dotdot lookup.
835 	 */
836 	if (cnp->cn_flags & ISDOTDOT) {
837 		if ((ndp->ni_lcf & (NI_LCF_STRICTRELATIVE | NI_LCF_CAP_DOTDOT))
838 		    == NI_LCF_STRICTRELATIVE) {
839 #ifdef KTRACE
840 			if (KTRPOINT(curthread, KTR_CAPFAIL))
841 				ktrcapfail(CAPFAIL_LOOKUP, NULL, NULL);
842 #endif
843 			error = ENOTCAPABLE;
844 			goto bad;
845 		}
846 		if ((cnp->cn_flags & ISLASTCN) != 0 &&
847 		    (cnp->cn_nameiop == DELETE || cnp->cn_nameiop == RENAME)) {
848 			error = EINVAL;
849 			goto bad;
850 		}
851 		for (;;) {
852 			for (pr = cnp->cn_cred->cr_prison; pr != NULL;
853 			     pr = pr->pr_parent)
854 				if (dp == pr->pr_root)
855 					break;
856 			if (dp == ndp->ni_rootdir ||
857 			    dp == ndp->ni_topdir ||
858 			    dp == rootvnode ||
859 			    pr != NULL ||
860 			    ((dp->v_vflag & VV_ROOT) != 0 &&
861 			     (cnp->cn_flags & NOCROSSMOUNT) != 0)) {
862 				ndp->ni_dvp = dp;
863 				ndp->ni_vp = dp;
864 				VREF(dp);
865 				goto nextname;
866 			}
867 			if ((dp->v_vflag & VV_ROOT) == 0)
868 				break;
869 			if (dp->v_iflag & VI_DOOMED) {	/* forced unmount */
870 				error = ENOENT;
871 				goto bad;
872 			}
873 			tdp = dp;
874 			dp = dp->v_mount->mnt_vnodecovered;
875 			VREF(dp);
876 			vput(tdp);
877 			vn_lock(dp,
878 			    compute_cn_lkflags(dp->v_mount, cnp->cn_lkflags |
879 			    LK_RETRY, ISDOTDOT));
880 			error = nameicap_check_dotdot(ndp, dp);
881 			if (error != 0) {
882 #ifdef KTRACE
883 				if (KTRPOINT(curthread, KTR_CAPFAIL))
884 					ktrcapfail(CAPFAIL_LOOKUP, NULL, NULL);
885 #endif
886 				goto bad;
887 			}
888 		}
889 	}
890 
891 	/*
892 	 * We now have a segment name to search for, and a directory to search.
893 	 */
894 unionlookup:
895 #ifdef MAC
896 	if ((cnp->cn_flags & NOMACCHECK) == 0) {
897 		error = mac_vnode_check_lookup(cnp->cn_thread->td_ucred, dp,
898 		    cnp);
899 		if (error)
900 			goto bad;
901 	}
902 #endif
903 	ndp->ni_dvp = dp;
904 	ndp->ni_vp = NULL;
905 	ASSERT_VOP_LOCKED(dp, "lookup");
906 	/*
907 	 * If we have a shared lock we may need to upgrade the lock for the
908 	 * last operation.
909 	 */
910 	if ((cnp->cn_flags & LOCKPARENT) && (cnp->cn_flags & ISLASTCN) &&
911 	    dp != vp_crossmp && VOP_ISLOCKED(dp) == LK_SHARED)
912 		vn_lock(dp, LK_UPGRADE|LK_RETRY);
913 	if ((dp->v_iflag & VI_DOOMED) != 0) {
914 		error = ENOENT;
915 		goto bad;
916 	}
917 	/*
918 	 * If we're looking up the last component and we need an exclusive
919 	 * lock, adjust our lkflags.
920 	 */
921 	if (needs_exclusive_leaf(dp->v_mount, cnp->cn_flags))
922 		cnp->cn_lkflags = LK_EXCLUSIVE;
923 #ifdef NAMEI_DIAGNOSTIC
924 	vn_printf(dp, "lookup in ");
925 #endif
926 	lkflags_save = cnp->cn_lkflags;
927 	cnp->cn_lkflags = compute_cn_lkflags(dp->v_mount, cnp->cn_lkflags,
928 	    cnp->cn_flags);
929 	error = VOP_LOOKUP(dp, &ndp->ni_vp, cnp);
930 	cnp->cn_lkflags = lkflags_save;
931 	if (error != 0) {
932 		KASSERT(ndp->ni_vp == NULL, ("leaf should be empty"));
933 #ifdef NAMEI_DIAGNOSTIC
934 		printf("not found\n");
935 #endif
936 		if ((error == ENOENT) &&
937 		    (dp->v_vflag & VV_ROOT) && (dp->v_mount != NULL) &&
938 		    (dp->v_mount->mnt_flag & MNT_UNION)) {
939 			tdp = dp;
940 			dp = dp->v_mount->mnt_vnodecovered;
941 			VREF(dp);
942 			vput(tdp);
943 			vn_lock(dp,
944 			    compute_cn_lkflags(dp->v_mount, cnp->cn_lkflags |
945 			    LK_RETRY, cnp->cn_flags));
946 			nameicap_tracker_add(ndp, dp);
947 			goto unionlookup;
948 		}
949 
950 		if (error == ERELOOKUP) {
951 			vref(dp);
952 			ndp->ni_vp = dp;
953 			error = 0;
954 			relookup = 1;
955 			goto good;
956 		}
957 
958 		if (error != EJUSTRETURN)
959 			goto bad;
960 		/*
961 		 * At this point, we know we're at the end of the
962 		 * pathname.  If creating / renaming, we can consider
963 		 * allowing the file or directory to be created / renamed,
964 		 * provided we're not on a read-only filesystem.
965 		 */
966 		if (rdonly) {
967 			error = EROFS;
968 			goto bad;
969 		}
970 		/* trailing slash only allowed for directories */
971 		if ((cnp->cn_flags & TRAILINGSLASH) &&
972 		    !(cnp->cn_flags & WILLBEDIR)) {
973 			error = ENOENT;
974 			goto bad;
975 		}
976 		if ((cnp->cn_flags & LOCKPARENT) == 0)
977 			VOP_UNLOCK(dp, 0);
978 		/*
979 		 * We return with ni_vp NULL to indicate that the entry
980 		 * doesn't currently exist, leaving a pointer to the
981 		 * (possibly locked) directory vnode in ndp->ni_dvp.
982 		 */
983 		if (cnp->cn_flags & SAVESTART) {
984 			ndp->ni_startdir = ndp->ni_dvp;
985 			VREF(ndp->ni_startdir);
986 		}
987 		goto success;
988 	}
989 
990 good:
991 #ifdef NAMEI_DIAGNOSTIC
992 	printf("found\n");
993 #endif
994 	dp = ndp->ni_vp;
995 
996 	/*
997 	 * Check to see if the vnode has been mounted on;
998 	 * if so find the root of the mounted filesystem.
999 	 */
1000 	while (dp->v_type == VDIR && (mp = dp->v_mountedhere) &&
1001 	       (cnp->cn_flags & NOCROSSMOUNT) == 0) {
1002 		if (vfs_busy(mp, 0))
1003 			continue;
1004 		vput(dp);
1005 		if (dp != ndp->ni_dvp)
1006 			vput(ndp->ni_dvp);
1007 		else
1008 			vrele(ndp->ni_dvp);
1009 		vrefact(vp_crossmp);
1010 		ndp->ni_dvp = vp_crossmp;
1011 		error = VFS_ROOT(mp, compute_cn_lkflags(mp, cnp->cn_lkflags,
1012 		    cnp->cn_flags), &tdp);
1013 		vfs_unbusy(mp);
1014 		if (vn_lock(vp_crossmp, LK_SHARED | LK_NOWAIT))
1015 			panic("vp_crossmp exclusively locked or reclaimed");
1016 		if (error) {
1017 			dpunlocked = 1;
1018 			goto bad2;
1019 		}
1020 		ndp->ni_vp = dp = tdp;
1021 	}
1022 
1023 	/*
1024 	 * Check for symbolic link
1025 	 */
1026 	if ((dp->v_type == VLNK) &&
1027 	    ((cnp->cn_flags & FOLLOW) || (cnp->cn_flags & TRAILINGSLASH) ||
1028 	     *ndp->ni_next == '/')) {
1029 		cnp->cn_flags |= ISSYMLINK;
1030 		if (dp->v_iflag & VI_DOOMED) {
1031 			/*
1032 			 * We can't know whether the directory was mounted with
1033 			 * NOSYMFOLLOW, so we can't follow safely.
1034 			 */
1035 			error = ENOENT;
1036 			goto bad2;
1037 		}
1038 		if (dp->v_mount->mnt_flag & MNT_NOSYMFOLLOW) {
1039 			error = EACCES;
1040 			goto bad2;
1041 		}
1042 		/*
1043 		 * Symlink code always expects an unlocked dvp.
1044 		 */
1045 		if (ndp->ni_dvp != ndp->ni_vp) {
1046 			VOP_UNLOCK(ndp->ni_dvp, 0);
1047 			ni_dvp_unlocked = 1;
1048 		}
1049 		goto success;
1050 	}
1051 
1052 nextname:
1053 	/*
1054 	 * Not a symbolic link that we will follow.  Continue with the
1055 	 * next component if there is any; otherwise, we're done.
1056 	 */
1057 	KASSERT((cnp->cn_flags & ISLASTCN) || *ndp->ni_next == '/',
1058 	    ("lookup: invalid path state."));
1059 	if (relookup) {
1060 		relookup = 0;
1061 		ndp->ni_pathlen = prev_ni_pathlen;
1062 		ndp->ni_next = prev_ni_next;
1063 		if (ndp->ni_dvp != dp)
1064 			vput(ndp->ni_dvp);
1065 		else
1066 			vrele(ndp->ni_dvp);
1067 		goto dirloop;
1068 	}
1069 	if (cnp->cn_flags & ISDOTDOT) {
1070 		error = nameicap_check_dotdot(ndp, ndp->ni_vp);
1071 		if (error != 0) {
1072 #ifdef KTRACE
1073 			if (KTRPOINT(curthread, KTR_CAPFAIL))
1074 				ktrcapfail(CAPFAIL_LOOKUP, NULL, NULL);
1075 #endif
1076 			goto bad2;
1077 		}
1078 	}
1079 	if (*ndp->ni_next == '/') {
1080 		cnp->cn_nameptr = ndp->ni_next;
1081 		while (*cnp->cn_nameptr == '/') {
1082 			cnp->cn_nameptr++;
1083 			ndp->ni_pathlen--;
1084 		}
1085 		if (ndp->ni_dvp != dp)
1086 			vput(ndp->ni_dvp);
1087 		else
1088 			vrele(ndp->ni_dvp);
1089 		goto dirloop;
1090 	}
1091 	/*
1092 	 * If we're processing a path with a trailing slash,
1093 	 * check that the end result is a directory.
1094 	 */
1095 	if ((cnp->cn_flags & TRAILINGSLASH) && dp->v_type != VDIR) {
1096 		error = ENOTDIR;
1097 		goto bad2;
1098 	}
1099 	/*
1100 	 * Disallow directory write attempts on read-only filesystems.
1101 	 */
1102 	if (rdonly &&
1103 	    (cnp->cn_nameiop == DELETE || cnp->cn_nameiop == RENAME)) {
1104 		error = EROFS;
1105 		goto bad2;
1106 	}
1107 	if (cnp->cn_flags & SAVESTART) {
1108 		ndp->ni_startdir = ndp->ni_dvp;
1109 		VREF(ndp->ni_startdir);
1110 	}
1111 	if (!wantparent) {
1112 		ni_dvp_unlocked = 2;
1113 		if (ndp->ni_dvp != dp)
1114 			vput(ndp->ni_dvp);
1115 		else
1116 			vrele(ndp->ni_dvp);
1117 	} else if ((cnp->cn_flags & LOCKPARENT) == 0 && ndp->ni_dvp != dp) {
1118 		VOP_UNLOCK(ndp->ni_dvp, 0);
1119 		ni_dvp_unlocked = 1;
1120 	}
1121 
1122 	if (cnp->cn_flags & AUDITVNODE1)
1123 		AUDIT_ARG_VNODE1(dp);
1124 	else if (cnp->cn_flags & AUDITVNODE2)
1125 		AUDIT_ARG_VNODE2(dp);
1126 
1127 	if ((cnp->cn_flags & LOCKLEAF) == 0)
1128 		VOP_UNLOCK(dp, 0);
1129 success:
1130 	/*
1131 	 * Because of shared lookup we may have the vnode shared locked, but
1132 	 * the caller may want it to be exclusively locked.
1133 	 */
1134 	if (needs_exclusive_leaf(dp->v_mount, cnp->cn_flags) &&
1135 	    VOP_ISLOCKED(dp) != LK_EXCLUSIVE) {
1136 		vn_lock(dp, LK_UPGRADE | LK_RETRY);
1137 		if (dp->v_iflag & VI_DOOMED) {
1138 			error = ENOENT;
1139 			goto bad2;
1140 		}
1141 	}
1142 	return (0);
1143 
1144 bad2:
1145 	if (ni_dvp_unlocked != 2) {
1146 		if (dp != ndp->ni_dvp && !ni_dvp_unlocked)
1147 			vput(ndp->ni_dvp);
1148 		else
1149 			vrele(ndp->ni_dvp);
1150 	}
1151 bad:
1152 	if (!dpunlocked)
1153 		vput(dp);
1154 	ndp->ni_vp = NULL;
1155 	return (error);
1156 }
1157 
1158 /*
1159  * relookup - lookup a path name component
1160  *    Used by lookup to re-acquire things.
1161  */
1162 int
1163 relookup(struct vnode *dvp, struct vnode **vpp, struct componentname *cnp)
1164 {
1165 	struct vnode *dp = NULL;		/* the directory we are searching */
1166 	int wantparent;			/* 1 => wantparent or lockparent flag */
1167 	int rdonly;			/* lookup read-only flag bit */
1168 	int error = 0;
1169 
1170 	KASSERT(cnp->cn_flags & ISLASTCN,
1171 	    ("relookup: Not given last component."));
1172 	/*
1173 	 * Setup: break out flag bits into variables.
1174 	 */
1175 	wantparent = cnp->cn_flags & (LOCKPARENT|WANTPARENT);
1176 	KASSERT(wantparent, ("relookup: parent not wanted."));
1177 	rdonly = cnp->cn_flags & RDONLY;
1178 	cnp->cn_flags &= ~ISSYMLINK;
1179 	dp = dvp;
1180 	cnp->cn_lkflags = LK_EXCLUSIVE;
1181 	vn_lock(dp, LK_EXCLUSIVE | LK_RETRY);
1182 
1183 	/*
1184 	 * Search a new directory.
1185 	 *
1186 	 * The last component of the filename is left accessible via
1187 	 * cnp->cn_nameptr for callers that need the name. Callers needing
1188 	 * the name set the SAVENAME flag. When done, they assume
1189 	 * responsibility for freeing the pathname buffer.
1190 	 */
1191 #ifdef NAMEI_DIAGNOSTIC
1192 	printf("{%s}: ", cnp->cn_nameptr);
1193 #endif
1194 
1195 	/*
1196 	 * Check for "" which represents the root directory after slash
1197 	 * removal.
1198 	 */
1199 	if (cnp->cn_nameptr[0] == '\0') {
1200 		/*
1201 		 * Support only LOOKUP for "/" because lookup()
1202 		 * can't succeed for CREATE, DELETE and RENAME.
1203 		 */
1204 		KASSERT(cnp->cn_nameiop == LOOKUP, ("nameiop must be LOOKUP"));
1205 		KASSERT(dp->v_type == VDIR, ("dp is not a directory"));
1206 
1207 		if (!(cnp->cn_flags & LOCKLEAF))
1208 			VOP_UNLOCK(dp, 0);
1209 		*vpp = dp;
1210 		/* XXX This should probably move to the top of function. */
1211 		if (cnp->cn_flags & SAVESTART)
1212 			panic("lookup: SAVESTART");
1213 		return (0);
1214 	}
1215 
1216 	if (cnp->cn_flags & ISDOTDOT)
1217 		panic ("relookup: lookup on dot-dot");
1218 
1219 	/*
1220 	 * We now have a segment name to search for, and a directory to search.
1221 	 */
1222 #ifdef NAMEI_DIAGNOSTIC
1223 	vn_printf(dp, "search in ");
1224 #endif
1225 	if ((error = VOP_LOOKUP(dp, vpp, cnp)) != 0) {
1226 		KASSERT(*vpp == NULL, ("leaf should be empty"));
1227 		if (error != EJUSTRETURN)
1228 			goto bad;
1229 		/*
1230 		 * If creating and at end of pathname, then can consider
1231 		 * allowing file to be created.
1232 		 */
1233 		if (rdonly) {
1234 			error = EROFS;
1235 			goto bad;
1236 		}
1237 		/* ASSERT(dvp == ndp->ni_startdir) */
1238 		if (cnp->cn_flags & SAVESTART)
1239 			VREF(dvp);
1240 		if ((cnp->cn_flags & LOCKPARENT) == 0)
1241 			VOP_UNLOCK(dp, 0);
1242 		/*
1243 		 * We return with ni_vp NULL to indicate that the entry
1244 		 * doesn't currently exist, leaving a pointer to the
1245 		 * (possibly locked) directory vnode in ndp->ni_dvp.
1246 		 */
1247 		return (0);
1248 	}
1249 
1250 	dp = *vpp;
1251 
1252 	/*
1253 	 * Disallow directory write attempts on read-only filesystems.
1254 	 */
1255 	if (rdonly &&
1256 	    (cnp->cn_nameiop == DELETE || cnp->cn_nameiop == RENAME)) {
1257 		if (dvp == dp)
1258 			vrele(dvp);
1259 		else
1260 			vput(dvp);
1261 		error = EROFS;
1262 		goto bad;
1263 	}
1264 	/*
1265 	 * Set the parent lock/ref state to the requested state.
1266 	 */
1267 	if ((cnp->cn_flags & LOCKPARENT) == 0 && dvp != dp) {
1268 		if (wantparent)
1269 			VOP_UNLOCK(dvp, 0);
1270 		else
1271 			vput(dvp);
1272 	} else if (!wantparent)
1273 		vrele(dvp);
1274 	/*
1275 	 * Check for symbolic link
1276 	 */
1277 	KASSERT(dp->v_type != VLNK || !(cnp->cn_flags & FOLLOW),
1278 	    ("relookup: symlink found.\n"));
1279 
1280 	/* ASSERT(dvp == ndp->ni_startdir) */
1281 	if (cnp->cn_flags & SAVESTART)
1282 		VREF(dvp);
1283 
1284 	if ((cnp->cn_flags & LOCKLEAF) == 0)
1285 		VOP_UNLOCK(dp, 0);
1286 	return (0);
1287 bad:
1288 	vput(dp);
1289 	*vpp = NULL;
1290 	return (error);
1291 }
1292 
1293 void
1294 NDINIT_ALL(struct nameidata *ndp, u_long op, u_long flags, enum uio_seg segflg,
1295     const char *namep, int dirfd, struct vnode *startdir, cap_rights_t *rightsp,
1296     struct thread *td)
1297 {
1298 
1299 	ndp->ni_cnd.cn_nameiop = op;
1300 	ndp->ni_cnd.cn_flags = flags;
1301 	ndp->ni_segflg = segflg;
1302 	ndp->ni_dirp = namep;
1303 	ndp->ni_dirfd = dirfd;
1304 	ndp->ni_startdir = startdir;
1305 	filecaps_init(&ndp->ni_filecaps);
1306 	ndp->ni_cnd.cn_thread = td;
1307 	if (rightsp != NULL)
1308 		ndp->ni_rightsneeded = *rightsp;
1309 	else
1310 		cap_rights_init(&ndp->ni_rightsneeded);
1311 }
1312 
1313 /*
1314  * Free data allocated by namei(); see namei(9) for details.
1315  */
1316 void
1317 NDFREE(struct nameidata *ndp, const u_int flags)
1318 {
1319 	int unlock_dvp;
1320 	int unlock_vp;
1321 
1322 	unlock_dvp = 0;
1323 	unlock_vp = 0;
1324 
1325 	if (!(flags & NDF_NO_FREE_PNBUF) &&
1326 	    (ndp->ni_cnd.cn_flags & HASBUF)) {
1327 		uma_zfree(namei_zone, ndp->ni_cnd.cn_pnbuf);
1328 		ndp->ni_cnd.cn_flags &= ~HASBUF;
1329 	}
1330 	if (!(flags & NDF_NO_VP_UNLOCK) &&
1331 	    (ndp->ni_cnd.cn_flags & LOCKLEAF) && ndp->ni_vp)
1332 		unlock_vp = 1;
1333 	if (!(flags & NDF_NO_VP_RELE) && ndp->ni_vp) {
1334 		if (unlock_vp) {
1335 			vput(ndp->ni_vp);
1336 			unlock_vp = 0;
1337 		} else
1338 			vrele(ndp->ni_vp);
1339 		ndp->ni_vp = NULL;
1340 	}
1341 	if (unlock_vp)
1342 		VOP_UNLOCK(ndp->ni_vp, 0);
1343 	if (!(flags & NDF_NO_DVP_UNLOCK) &&
1344 	    (ndp->ni_cnd.cn_flags & LOCKPARENT) &&
1345 	    ndp->ni_dvp != ndp->ni_vp)
1346 		unlock_dvp = 1;
1347 	if (!(flags & NDF_NO_DVP_RELE) &&
1348 	    (ndp->ni_cnd.cn_flags & (LOCKPARENT|WANTPARENT))) {
1349 		if (unlock_dvp) {
1350 			vput(ndp->ni_dvp);
1351 			unlock_dvp = 0;
1352 		} else
1353 			vrele(ndp->ni_dvp);
1354 		ndp->ni_dvp = NULL;
1355 	}
1356 	if (unlock_dvp)
1357 		VOP_UNLOCK(ndp->ni_dvp, 0);
1358 	if (!(flags & NDF_NO_STARTDIR_RELE) &&
1359 	    (ndp->ni_cnd.cn_flags & SAVESTART)) {
1360 		vrele(ndp->ni_startdir);
1361 		ndp->ni_startdir = NULL;
1362 	}
1363 }
1364 
1365 /*
1366  * Determine if there is a suitable alternate filename under the specified
1367  * prefix for the specified path.  If the create flag is set, then the
1368  * alternate prefix will be used so long as the parent directory exists.
1369  * This is used by the various compatibility ABIs so that Linux binaries prefer
1370  * files under /compat/linux for example.  The chosen path (whether under
1371  * the prefix or under /) is returned in a kernel malloc'd buffer pointed
1372  * to by pathbuf.  The caller is responsible for free'ing the buffer from
1373  * the M_TEMP bucket if one is returned.
1374  */
1375 int
1376 kern_alternate_path(struct thread *td, const char *prefix, const char *path,
1377     enum uio_seg pathseg, char **pathbuf, int create, int dirfd)
1378 {
1379 	struct nameidata nd, ndroot;
1380 	char *ptr, *buf, *cp;
1381 	size_t len, sz;
1382 	int error;
1383 
1384 	buf = (char *) malloc(MAXPATHLEN, M_TEMP, M_WAITOK);
1385 	*pathbuf = buf;
1386 
1387 	/* Copy the prefix into the new pathname as a starting point. */
1388 	len = strlcpy(buf, prefix, MAXPATHLEN);
1389 	if (len >= MAXPATHLEN) {
1390 		*pathbuf = NULL;
1391 		free(buf, M_TEMP);
1392 		return (EINVAL);
1393 	}
1394 	sz = MAXPATHLEN - len;
1395 	ptr = buf + len;
1396 
1397 	/* Append the filename to the prefix. */
1398 	if (pathseg == UIO_SYSSPACE)
1399 		error = copystr(path, ptr, sz, &len);
1400 	else
1401 		error = copyinstr(path, ptr, sz, &len);
1402 
1403 	if (error) {
1404 		*pathbuf = NULL;
1405 		free(buf, M_TEMP);
1406 		return (error);
1407 	}
1408 
1409 	/* Only use a prefix with absolute pathnames. */
1410 	if (*ptr != '/') {
1411 		error = EINVAL;
1412 		goto keeporig;
1413 	}
1414 
1415 	if (dirfd != AT_FDCWD) {
1416 		/*
1417 		 * We want the original because the "prefix" is
1418 		 * included in the already opened dirfd.
1419 		 */
1420 		bcopy(ptr, buf, len);
1421 		return (0);
1422 	}
1423 
1424 	/*
1425 	 * We know that there is a / somewhere in this pathname.
1426 	 * Search backwards for it, to find the file's parent dir
1427 	 * to see if it exists in the alternate tree. If it does,
1428 	 * and we want to create a file (cflag is set). We don't
1429 	 * need to worry about the root comparison in this case.
1430 	 */
1431 
1432 	if (create) {
1433 		for (cp = &ptr[len] - 1; *cp != '/'; cp--);
1434 		*cp = '\0';
1435 
1436 		NDINIT(&nd, LOOKUP, NOFOLLOW, UIO_SYSSPACE, buf, td);
1437 		error = namei(&nd);
1438 		*cp = '/';
1439 		if (error != 0)
1440 			goto keeporig;
1441 	} else {
1442 		NDINIT(&nd, LOOKUP, NOFOLLOW, UIO_SYSSPACE, buf, td);
1443 
1444 		error = namei(&nd);
1445 		if (error != 0)
1446 			goto keeporig;
1447 
1448 		/*
1449 		 * We now compare the vnode of the prefix to the one
1450 		 * vnode asked. If they resolve to be the same, then we
1451 		 * ignore the match so that the real root gets used.
1452 		 * This avoids the problem of traversing "../.." to find the
1453 		 * root directory and never finding it, because "/" resolves
1454 		 * to the emulation root directory. This is expensive :-(
1455 		 */
1456 		NDINIT(&ndroot, LOOKUP, FOLLOW, UIO_SYSSPACE, prefix,
1457 		    td);
1458 
1459 		/* We shouldn't ever get an error from this namei(). */
1460 		error = namei(&ndroot);
1461 		if (error == 0) {
1462 			if (nd.ni_vp == ndroot.ni_vp)
1463 				error = ENOENT;
1464 
1465 			NDFREE(&ndroot, NDF_ONLY_PNBUF);
1466 			vrele(ndroot.ni_vp);
1467 		}
1468 	}
1469 
1470 	NDFREE(&nd, NDF_ONLY_PNBUF);
1471 	vrele(nd.ni_vp);
1472 
1473 keeporig:
1474 	/* If there was an error, use the original path name. */
1475 	if (error)
1476 		bcopy(ptr, buf, len);
1477 	return (error);
1478 }
1479