xref: /freebsd/sys/kern/vfs_lookup.c (revision 4bfebc8d2c5d0e813dfdcbe7038b36ffc2bb9f1b)
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 	struct componentname *cnp;
182 
183 	if ((ndp->ni_lcf & NI_LCF_CAP_DOTDOT) == 0 || dp->v_type != VDIR)
184 		return;
185 	cnp = &ndp->ni_cnd;
186 	if ((cnp->cn_flags & BENEATH) != 0 &&
187 	    (ndp->ni_lcf & NI_LCF_BENEATH_LATCHED) == 0) {
188 		MPASS((ndp->ni_lcf & NI_LCF_LATCH) != 0);
189 		if (dp != ndp->ni_beneath_latch)
190 			return;
191 		ndp->ni_lcf |= NI_LCF_BENEATH_LATCHED;
192 	}
193 	nt = uma_zalloc(nt_zone, M_WAITOK);
194 	vhold(dp);
195 	nt->dp = dp;
196 	TAILQ_INSERT_TAIL(&ndp->ni_cap_tracker, nt, nm_link);
197 }
198 
199 static void
200 nameicap_cleanup(struct nameidata *ndp, bool clean_latch)
201 {
202 	struct nameicap_tracker *nt, *nt1;
203 
204 	KASSERT(TAILQ_EMPTY(&ndp->ni_cap_tracker) ||
205 	    (ndp->ni_lcf & NI_LCF_CAP_DOTDOT) != 0, ("not strictrelative"));
206 	TAILQ_FOREACH_SAFE(nt, &ndp->ni_cap_tracker, nm_link, nt1) {
207 		TAILQ_REMOVE(&ndp->ni_cap_tracker, nt, nm_link);
208 		vdrop(nt->dp);
209 		uma_zfree(nt_zone, nt);
210 	}
211 	if (clean_latch && (ndp->ni_lcf & NI_LCF_LATCH) != 0) {
212 		ndp->ni_lcf &= ~NI_LCF_LATCH;
213 		vrele(ndp->ni_beneath_latch);
214 	}
215 }
216 
217 /*
218  * For dotdot lookups in capability mode, only allow the component
219  * lookup to succeed if the resulting directory was already traversed
220  * during the operation.  This catches situations where already
221  * traversed directory is moved to different parent, and then we walk
222  * over it with dotdots.
223  *
224  * Also allow to force failure of dotdot lookups for non-local
225  * filesystems, where external agents might assist local lookups to
226  * escape the compartment.
227  */
228 static int
229 nameicap_check_dotdot(struct nameidata *ndp, struct vnode *dp)
230 {
231 	struct nameicap_tracker *nt;
232 	struct mount *mp;
233 
234 	if ((ndp->ni_lcf & NI_LCF_CAP_DOTDOT) == 0 || dp == NULL ||
235 	    dp->v_type != VDIR)
236 		return (0);
237 	mp = dp->v_mount;
238 	if (lookup_cap_dotdot_nonlocal == 0 && mp != NULL &&
239 	    (mp->mnt_flag & MNT_LOCAL) == 0)
240 		return (ENOTCAPABLE);
241 	TAILQ_FOREACH_REVERSE(nt, &ndp->ni_cap_tracker, nameicap_tracker_head,
242 	    nm_link) {
243 		if ((ndp->ni_lcf & NI_LCF_LATCH) != 0 &&
244 		    ndp->ni_beneath_latch == nt->dp) {
245 			ndp->ni_lcf &= ~NI_LCF_BENEATH_LATCHED;
246 			nameicap_cleanup(ndp, false);
247 			return (0);
248 		}
249 		if (dp == nt->dp)
250 			return (0);
251 	}
252 	return (ENOTCAPABLE);
253 }
254 
255 static void
256 namei_cleanup_cnp(struct componentname *cnp)
257 {
258 
259 	uma_zfree(namei_zone, cnp->cn_pnbuf);
260 #ifdef DIAGNOSTIC
261 	cnp->cn_pnbuf = NULL;
262 	cnp->cn_nameptr = NULL;
263 #endif
264 }
265 
266 static int
267 namei_handle_root(struct nameidata *ndp, struct vnode **dpp)
268 {
269 	struct componentname *cnp;
270 
271 	cnp = &ndp->ni_cnd;
272 	if ((ndp->ni_lcf & NI_LCF_STRICTRELATIVE) != 0) {
273 #ifdef KTRACE
274 		if (KTRPOINT(curthread, KTR_CAPFAIL))
275 			ktrcapfail(CAPFAIL_LOOKUP, NULL, NULL);
276 #endif
277 		return (ENOTCAPABLE);
278 	}
279 	if ((cnp->cn_flags & BENEATH) != 0) {
280 		ndp->ni_lcf |= NI_LCF_BENEATH_ABS;
281 		ndp->ni_lcf &= ~NI_LCF_BENEATH_LATCHED;
282 		nameicap_cleanup(ndp, false);
283 	}
284 	while (*(cnp->cn_nameptr) == '/') {
285 		cnp->cn_nameptr++;
286 		ndp->ni_pathlen--;
287 	}
288 	*dpp = ndp->ni_rootdir;
289 	vrefact(*dpp);
290 	return (0);
291 }
292 
293 static int
294 namei_setup(struct nameidata *ndp, struct vnode **dpp, struct pwd **pwdp)
295 {
296 	struct componentname *cnp;
297 	struct file *dfp;
298 	struct thread *td;
299 	struct pwd *pwd;
300 	cap_rights_t rights;
301 	struct filecaps dirfd_caps;
302 	int error;
303 	bool startdir_used;
304 
305 	cnp = &ndp->ni_cnd;
306 	td = cnp->cn_thread;
307 
308 	startdir_used = false;
309 	*pwdp = NULL;
310 	*dpp = NULL;
311 
312 #ifdef CAPABILITY_MODE
313 	/*
314 	 * In capability mode, lookups must be restricted to happen in
315 	 * the subtree with the root specified by the file descriptor:
316 	 * - The root must be real file descriptor, not the pseudo-descriptor
317 	 *   AT_FDCWD.
318 	 * - The passed path must be relative and not absolute.
319 	 * - If lookup_cap_dotdot is disabled, path must not contain the
320 	 *   '..' components.
321 	 * - If lookup_cap_dotdot is enabled, we verify that all '..'
322 	 *   components lookups result in the directories which were
323 	 *   previously walked by us, which prevents an escape from
324 	 *   the relative root.
325 	 */
326 	if (IN_CAPABILITY_MODE(td) && (cnp->cn_flags & NOCAPCHECK) == 0) {
327 		ndp->ni_lcf |= NI_LCF_STRICTRELATIVE;
328 		ndp->ni_resflags |= NIRES_STRICTREL;
329 		if (ndp->ni_dirfd == AT_FDCWD) {
330 #ifdef KTRACE
331 			if (KTRPOINT(td, KTR_CAPFAIL))
332 				ktrcapfail(CAPFAIL_LOOKUP, NULL, NULL);
333 #endif
334 			return (ECAPMODE);
335 		}
336 	}
337 #endif
338 	error = 0;
339 
340 	/*
341 	 * Get starting point for the translation.
342 	 */
343 	pwd = pwd_hold(td);
344 	/*
345 	 * The reference on ni_rootdir is acquired in the block below to avoid
346 	 * back-to-back atomics for absolute lookups.
347 	 */
348 	ndp->ni_rootdir = pwd->pwd_rdir;
349 	ndp->ni_topdir = pwd->pwd_jdir;
350 
351 	if (cnp->cn_pnbuf[0] == '/') {
352 		ndp->ni_resflags |= NIRES_ABS;
353 		error = namei_handle_root(ndp, dpp);
354 	} else {
355 		if (ndp->ni_startdir != NULL) {
356 			*dpp = ndp->ni_startdir;
357 			startdir_used = true;
358 		} else if (ndp->ni_dirfd == AT_FDCWD) {
359 			*dpp = pwd->pwd_cdir;
360 			vrefact(*dpp);
361 		} else {
362 			rights = *ndp->ni_rightsneeded;
363 			cap_rights_set_one(&rights, CAP_LOOKUP);
364 
365 			if (cnp->cn_flags & AUDITVNODE1)
366 				AUDIT_ARG_ATFD1(ndp->ni_dirfd);
367 			if (cnp->cn_flags & AUDITVNODE2)
368 				AUDIT_ARG_ATFD2(ndp->ni_dirfd);
369 			/*
370 			 * Effectively inlined fgetvp_rights, because we need to
371 			 * inspect the file as well as grabbing the vnode.
372 			 */
373 			error = fget_cap(td, ndp->ni_dirfd, &rights,
374 			    &dfp, &ndp->ni_filecaps);
375 			if (error != 0) {
376 				/*
377 				 * Preserve the error; it should either be EBADF
378 				 * or capability-related, both of which can be
379 				 * safely returned to the caller.
380 				 */
381 			} else {
382 				if (dfp->f_ops == &badfileops) {
383 					error = EBADF;
384 				} else if (dfp->f_vnode == NULL) {
385 					error = ENOTDIR;
386 				} else {
387 					*dpp = dfp->f_vnode;
388 					vrefact(*dpp);
389 
390 					if ((dfp->f_flag & FSEARCH) != 0)
391 						cnp->cn_flags |= NOEXECCHECK;
392 				}
393 				fdrop(dfp, td);
394 			}
395 #ifdef CAPABILITIES
396 			/*
397 			 * If file descriptor doesn't have all rights,
398 			 * all lookups relative to it must also be
399 			 * strictly relative.
400 			 */
401 			CAP_ALL(&rights);
402 			if (!cap_rights_contains(&ndp->ni_filecaps.fc_rights,
403 			    &rights) ||
404 			    ndp->ni_filecaps.fc_fcntls != CAP_FCNTL_ALL ||
405 			    ndp->ni_filecaps.fc_nioctls != -1) {
406 				ndp->ni_lcf |= NI_LCF_STRICTRELATIVE;
407 				ndp->ni_resflags |= NIRES_STRICTREL;
408 			}
409 #endif
410 		}
411 		if (error == 0 && (*dpp)->v_type != VDIR)
412 			error = ENOTDIR;
413 	}
414 	if (error == 0 && (cnp->cn_flags & BENEATH) != 0) {
415 		if (ndp->ni_dirfd == AT_FDCWD) {
416 			ndp->ni_beneath_latch = pwd->pwd_cdir;
417 			vrefact(ndp->ni_beneath_latch);
418 		} else {
419 			rights = *ndp->ni_rightsneeded;
420 			cap_rights_set_one(&rights, CAP_LOOKUP);
421 			error = fgetvp_rights(td, ndp->ni_dirfd, &rights,
422 			    &dirfd_caps, &ndp->ni_beneath_latch);
423 			if (error == 0 && (*dpp)->v_type != VDIR) {
424 				vrele(ndp->ni_beneath_latch);
425 				error = ENOTDIR;
426 			}
427 		}
428 		if (error == 0)
429 			ndp->ni_lcf |= NI_LCF_LATCH;
430 	}
431 	if (error == 0 && (cnp->cn_flags & RBENEATH) != 0) {
432 		if (cnp->cn_pnbuf[0] == '/' ||
433 		    (ndp->ni_lcf & NI_LCF_BENEATH_ABS) != 0) {
434 			error = EINVAL;
435 		} else if ((ndp->ni_lcf & NI_LCF_STRICTRELATIVE) == 0) {
436 			ndp->ni_lcf |= NI_LCF_STRICTRELATIVE |
437 			    NI_LCF_CAP_DOTDOT;
438 		}
439 	}
440 
441 	/*
442 	 * If we are auditing the kernel pathname, save the user pathname.
443 	 */
444 	if (cnp->cn_flags & AUDITVNODE1)
445 		AUDIT_ARG_UPATH1_VP(td, ndp->ni_rootdir, *dpp, cnp->cn_pnbuf);
446 	if (cnp->cn_flags & AUDITVNODE2)
447 		AUDIT_ARG_UPATH2_VP(td, ndp->ni_rootdir, *dpp, cnp->cn_pnbuf);
448 	if (ndp->ni_startdir != NULL && !startdir_used)
449 		vrele(ndp->ni_startdir);
450 	if (error != 0) {
451 		if (*dpp != NULL)
452 			vrele(*dpp);
453 		pwd_drop(pwd);
454 		return (error);
455 	}
456 	MPASS((ndp->ni_lcf & (NI_LCF_BENEATH_ABS | NI_LCF_LATCH)) !=
457 	    NI_LCF_BENEATH_ABS);
458 	if (((ndp->ni_lcf & NI_LCF_STRICTRELATIVE) != 0 &&
459 	    lookup_cap_dotdot != 0) ||
460 	    ((ndp->ni_lcf & NI_LCF_STRICTRELATIVE) == 0 &&
461 	    (cnp->cn_flags & BENEATH) != 0))
462 		ndp->ni_lcf |= NI_LCF_CAP_DOTDOT;
463 	SDT_PROBE4(vfs, namei, lookup, entry, *dpp, cnp->cn_pnbuf,
464 	    cnp->cn_flags, false);
465 	*pwdp = pwd;
466 	return (0);
467 }
468 
469 /*
470  * Convert a pathname into a pointer to a locked vnode.
471  *
472  * The FOLLOW flag is set when symbolic links are to be followed
473  * when they occur at the end of the name translation process.
474  * Symbolic links are always followed for all other pathname
475  * components other than the last.
476  *
477  * The segflg defines whether the name is to be copied from user
478  * space or kernel space.
479  *
480  * Overall outline of namei:
481  *
482  *	copy in name
483  *	get starting directory
484  *	while (!done && !error) {
485  *		call lookup to search path.
486  *		if symbolic link, massage name in buffer and continue
487  *	}
488  */
489 int
490 namei(struct nameidata *ndp)
491 {
492 	char *cp;		/* pointer into pathname argument */
493 	struct vnode *dp;	/* the directory we are searching */
494 	struct iovec aiov;		/* uio for reading symbolic links */
495 	struct componentname *cnp;
496 	struct thread *td;
497 	struct pwd *pwd;
498 	struct uio auio;
499 	int error, linklen;
500 	enum cache_fpl_status status;
501 
502 	cnp = &ndp->ni_cnd;
503 	td = cnp->cn_thread;
504 #ifdef INVARIANTS
505 	KASSERT((ndp->ni_debugflags & NAMEI_DBG_CALLED) == 0,
506 	    ("%s: repeated call to namei without NDREINIT", __func__));
507 	KASSERT(ndp->ni_debugflags == NAMEI_DBG_INITED,
508 	    ("%s: bad debugflags %d", __func__, ndp->ni_debugflags));
509 	ndp->ni_debugflags |= NAMEI_DBG_CALLED;
510 	if (ndp->ni_startdir != NULL)
511 		ndp->ni_debugflags |= NAMEI_DBG_HADSTARTDIR;
512 	/*
513 	 * For NDVALIDATE.
514 	 *
515 	 * While NDINIT may seem like a more natural place to do it, there are
516 	 * callers which directly modify flags past invoking init.
517 	 */
518 	cnp->cn_origflags = cnp->cn_flags;
519 #endif
520 	ndp->ni_cnd.cn_cred = ndp->ni_cnd.cn_thread->td_ucred;
521 	KASSERT(ndp->ni_resflags == 0, ("%s: garbage in ni_resflags: %x\n",
522 	    __func__, ndp->ni_resflags));
523 	KASSERT(cnp->cn_cred && td->td_proc, ("namei: bad cred/proc"));
524 	KASSERT((cnp->cn_flags & NAMEI_INTERNAL_FLAGS) == 0,
525 	    ("namei: unexpected flags: %" PRIx64 "\n",
526 	    cnp->cn_flags & NAMEI_INTERNAL_FLAGS));
527 	if (cnp->cn_flags & NOCACHE)
528 		KASSERT(cnp->cn_nameiop != LOOKUP,
529 		    ("%s: NOCACHE passed with LOOKUP", __func__));
530 	MPASS(ndp->ni_startdir == NULL || ndp->ni_startdir->v_type == VDIR ||
531 	    ndp->ni_startdir->v_type == VBAD);
532 
533 	ndp->ni_lcf = 0;
534 	ndp->ni_vp = NULL;
535 
536 	/*
537 	 * Get a buffer for the name to be translated, and copy the
538 	 * name into the buffer.
539 	 */
540 	cnp->cn_pnbuf = uma_zalloc(namei_zone, M_WAITOK);
541 	if (ndp->ni_segflg == UIO_SYSSPACE)
542 		error = copystr(ndp->ni_dirp, cnp->cn_pnbuf, MAXPATHLEN,
543 		    &ndp->ni_pathlen);
544 	else
545 		error = copyinstr(ndp->ni_dirp, cnp->cn_pnbuf, MAXPATHLEN,
546 		    &ndp->ni_pathlen);
547 
548 	if (__predict_false(error != 0)) {
549 		namei_cleanup_cnp(cnp);
550 		return (error);
551 	}
552 
553 	/*
554 	 * Don't allow empty pathnames.
555 	 */
556 	if (__predict_false(*cnp->cn_pnbuf == '\0')) {
557 		namei_cleanup_cnp(cnp);
558 		return (ENOENT);
559 	}
560 
561 #ifdef KTRACE
562 	if (KTRPOINT(td, KTR_NAMEI)) {
563 		KASSERT(cnp->cn_thread == curthread,
564 		    ("namei not using curthread"));
565 		ktrnamei(cnp->cn_pnbuf);
566 	}
567 #endif
568 
569 	cnp->cn_nameptr = cnp->cn_pnbuf;
570 
571 	/*
572 	 * First try looking up the target without locking any vnodes.
573 	 *
574 	 * We may need to start from scratch or pick up where it left off.
575 	 */
576 	error = cache_fplookup(ndp, &status, &pwd);
577 	switch (status) {
578 	case CACHE_FPL_STATUS_UNSET:
579 		__assert_unreachable();
580 		break;
581 	case CACHE_FPL_STATUS_HANDLED:
582 		if (error == 0)
583 			NDVALIDATE(ndp);
584 		return (error);
585 	case CACHE_FPL_STATUS_PARTIAL:
586 		TAILQ_INIT(&ndp->ni_cap_tracker);
587 		dp = ndp->ni_startdir;
588 		break;
589 	case CACHE_FPL_STATUS_ABORTED:
590 		TAILQ_INIT(&ndp->ni_cap_tracker);
591 		error = namei_setup(ndp, &dp, &pwd);
592 		if (error != 0) {
593 			namei_cleanup_cnp(cnp);
594 			return (error);
595 		}
596 		break;
597 	}
598 
599 	ndp->ni_loopcnt = 0;
600 
601 	/*
602 	 * Locked lookup.
603 	 */
604 	for (;;) {
605 		ndp->ni_startdir = dp;
606 		error = lookup(ndp);
607 		if (error != 0) {
608 			/*
609 			 * Override an error to not allow user to use
610 			 * BENEATH as an oracle.
611 			 */
612 			if ((ndp->ni_lcf & (NI_LCF_LATCH |
613 			    NI_LCF_BENEATH_LATCHED)) == NI_LCF_LATCH)
614 				error = ENOTCAPABLE;
615 			goto out;
616 		}
617 
618 		/*
619 		 * If not a symbolic link, we're done.
620 		 */
621 		if ((cnp->cn_flags & ISSYMLINK) == 0) {
622 			if ((cnp->cn_flags & (SAVENAME | SAVESTART)) == 0) {
623 				namei_cleanup_cnp(cnp);
624 			} else
625 				cnp->cn_flags |= HASBUF;
626 			if ((ndp->ni_lcf & (NI_LCF_LATCH |
627 			    NI_LCF_BENEATH_LATCHED)) == NI_LCF_LATCH) {
628 				NDFREE(ndp, 0);
629 				error = ENOTCAPABLE;
630 			}
631 			nameicap_cleanup(ndp, true);
632 			SDT_PROBE3(vfs, namei, lookup, return, error,
633 			    (error == 0 ? ndp->ni_vp : NULL), false);
634 			pwd_drop(pwd);
635 			if (error == 0)
636 				NDVALIDATE(ndp);
637 			return (error);
638 		}
639 		if (ndp->ni_loopcnt++ >= MAXSYMLINKS) {
640 			error = ELOOP;
641 			break;
642 		}
643 #ifdef MAC
644 		if ((cnp->cn_flags & NOMACCHECK) == 0) {
645 			error = mac_vnode_check_readlink(td->td_ucred,
646 			    ndp->ni_vp);
647 			if (error != 0)
648 				break;
649 		}
650 #endif
651 		if (ndp->ni_pathlen > 1)
652 			cp = uma_zalloc(namei_zone, M_WAITOK);
653 		else
654 			cp = cnp->cn_pnbuf;
655 		aiov.iov_base = cp;
656 		aiov.iov_len = MAXPATHLEN;
657 		auio.uio_iov = &aiov;
658 		auio.uio_iovcnt = 1;
659 		auio.uio_offset = 0;
660 		auio.uio_rw = UIO_READ;
661 		auio.uio_segflg = UIO_SYSSPACE;
662 		auio.uio_td = td;
663 		auio.uio_resid = MAXPATHLEN;
664 		error = VOP_READLINK(ndp->ni_vp, &auio, cnp->cn_cred);
665 		if (error != 0) {
666 			if (ndp->ni_pathlen > 1)
667 				uma_zfree(namei_zone, cp);
668 			break;
669 		}
670 		linklen = MAXPATHLEN - auio.uio_resid;
671 		if (linklen == 0) {
672 			if (ndp->ni_pathlen > 1)
673 				uma_zfree(namei_zone, cp);
674 			error = ENOENT;
675 			break;
676 		}
677 		if (linklen + ndp->ni_pathlen > MAXPATHLEN) {
678 			if (ndp->ni_pathlen > 1)
679 				uma_zfree(namei_zone, cp);
680 			error = ENAMETOOLONG;
681 			break;
682 		}
683 		if (ndp->ni_pathlen > 1) {
684 			bcopy(ndp->ni_next, cp + linklen, ndp->ni_pathlen);
685 			uma_zfree(namei_zone, cnp->cn_pnbuf);
686 			cnp->cn_pnbuf = cp;
687 		} else
688 			cnp->cn_pnbuf[linklen] = '\0';
689 		ndp->ni_pathlen += linklen;
690 		vput(ndp->ni_vp);
691 		dp = ndp->ni_dvp;
692 		/*
693 		 * Check if root directory should replace current directory.
694 		 */
695 		cnp->cn_nameptr = cnp->cn_pnbuf;
696 		if (*(cnp->cn_nameptr) == '/') {
697 			vrele(dp);
698 			error = namei_handle_root(ndp, &dp);
699 			if (error != 0)
700 				goto out;
701 		}
702 	}
703 	vput(ndp->ni_vp);
704 	ndp->ni_vp = NULL;
705 	vrele(ndp->ni_dvp);
706 out:
707 	MPASS(error != 0);
708 	namei_cleanup_cnp(cnp);
709 	nameicap_cleanup(ndp, true);
710 	SDT_PROBE3(vfs, namei, lookup, return, error, NULL, false);
711 	pwd_drop(pwd);
712 	return (error);
713 }
714 
715 static int
716 compute_cn_lkflags(struct mount *mp, int lkflags, int cnflags)
717 {
718 
719 	if (mp == NULL || ((lkflags & LK_SHARED) &&
720 	    (!(mp->mnt_kern_flag & MNTK_LOOKUP_SHARED) ||
721 	    ((cnflags & ISDOTDOT) &&
722 	    (mp->mnt_kern_flag & MNTK_LOOKUP_EXCL_DOTDOT))))) {
723 		lkflags &= ~LK_SHARED;
724 		lkflags |= LK_EXCLUSIVE;
725 	}
726 	lkflags |= LK_NODDLKTREAT;
727 	return (lkflags);
728 }
729 
730 static __inline int
731 needs_exclusive_leaf(struct mount *mp, int flags)
732 {
733 
734 	/*
735 	 * Intermediate nodes can use shared locks, we only need to
736 	 * force an exclusive lock for leaf nodes.
737 	 */
738 	if ((flags & (ISLASTCN | LOCKLEAF)) != (ISLASTCN | LOCKLEAF))
739 		return (0);
740 
741 	/* Always use exclusive locks if LOCKSHARED isn't set. */
742 	if (!(flags & LOCKSHARED))
743 		return (1);
744 
745 	/*
746 	 * For lookups during open(), if the mount point supports
747 	 * extended shared operations, then use a shared lock for the
748 	 * leaf node, otherwise use an exclusive lock.
749 	 */
750 	if ((flags & ISOPEN) != 0)
751 		return (!MNT_EXTENDED_SHARED(mp));
752 
753 	/*
754 	 * Lookup requests outside of open() that specify LOCKSHARED
755 	 * only need a shared lock on the leaf vnode.
756 	 */
757 	return (0);
758 }
759 
760 /*
761  * Search a pathname.
762  * This is a very central and rather complicated routine.
763  *
764  * The pathname is pointed to by ni_ptr and is of length ni_pathlen.
765  * The starting directory is taken from ni_startdir. The pathname is
766  * descended until done, or a symbolic link is encountered. The variable
767  * ni_more is clear if the path is completed; it is set to one if a
768  * symbolic link needing interpretation is encountered.
769  *
770  * The flag argument is LOOKUP, CREATE, RENAME, or DELETE depending on
771  * whether the name is to be looked up, created, renamed, or deleted.
772  * When CREATE, RENAME, or DELETE is specified, information usable in
773  * creating, renaming, or deleting a directory entry may be calculated.
774  * If flag has LOCKPARENT or'ed into it, the parent directory is returned
775  * locked. If flag has WANTPARENT or'ed into it, the parent directory is
776  * returned unlocked. Otherwise the parent directory is not returned. If
777  * the target of the pathname exists and LOCKLEAF is or'ed into the flag
778  * the target is returned locked, otherwise it is returned unlocked.
779  * When creating or renaming and LOCKPARENT is specified, the target may not
780  * be ".".  When deleting and LOCKPARENT is specified, the target may be ".".
781  *
782  * Overall outline of lookup:
783  *
784  * dirloop:
785  *	identify next component of name at ndp->ni_ptr
786  *	handle degenerate case where name is null string
787  *	if .. and crossing mount points and on mounted filesys, find parent
788  *	call VOP_LOOKUP routine for next component name
789  *	    directory vnode returned in ni_dvp, unlocked unless LOCKPARENT set
790  *	    component vnode returned in ni_vp (if it exists), locked.
791  *	if result vnode is mounted on and crossing mount points,
792  *	    find mounted on vnode
793  *	if more components of name, do next level at dirloop
794  *	return the answer in ni_vp, locked if LOCKLEAF set
795  *	    if LOCKPARENT set, return locked parent in ni_dvp
796  *	    if WANTPARENT set, return unlocked parent in ni_dvp
797  */
798 int
799 lookup(struct nameidata *ndp)
800 {
801 	char *cp;			/* pointer into pathname argument */
802 	char *prev_ni_next;		/* saved ndp->ni_next */
803 	struct vnode *dp = NULL;	/* the directory we are searching */
804 	struct vnode *tdp;		/* saved dp */
805 	struct mount *mp;		/* mount table entry */
806 	struct prison *pr;
807 	size_t prev_ni_pathlen;		/* saved ndp->ni_pathlen */
808 	int docache;			/* == 0 do not cache last component */
809 	int wantparent;			/* 1 => wantparent or lockparent flag */
810 	int rdonly;			/* lookup read-only flag bit */
811 	int error = 0;
812 	int dpunlocked = 0;		/* dp has already been unlocked */
813 	int relookup = 0;		/* do not consume the path component */
814 	struct componentname *cnp = &ndp->ni_cnd;
815 	int lkflags_save;
816 	int ni_dvp_unlocked;
817 
818 	/*
819 	 * Setup: break out flag bits into variables.
820 	 */
821 	ni_dvp_unlocked = 0;
822 	wantparent = cnp->cn_flags & (LOCKPARENT | WANTPARENT);
823 	KASSERT(cnp->cn_nameiop == LOOKUP || wantparent,
824 	    ("CREATE, DELETE, RENAME require LOCKPARENT or WANTPARENT."));
825 	/*
826 	 * When set to zero, docache causes the last component of the
827 	 * pathname to be deleted from the cache and the full lookup
828 	 * of the name to be done (via VOP_CACHEDLOOKUP()). Often
829 	 * filesystems need some pre-computed values that are made
830 	 * during the full lookup, for instance UFS sets dp->i_offset.
831 	 *
832 	 * The docache variable is set to zero when requested by the
833 	 * NOCACHE flag and for all modifying operations except CREATE.
834 	 */
835 	docache = (cnp->cn_flags & NOCACHE) ^ NOCACHE;
836 	if (cnp->cn_nameiop == DELETE ||
837 	    (wantparent && cnp->cn_nameiop != CREATE &&
838 	     cnp->cn_nameiop != LOOKUP))
839 		docache = 0;
840 	rdonly = cnp->cn_flags & RDONLY;
841 	cnp->cn_flags &= ~ISSYMLINK;
842 	ndp->ni_dvp = NULL;
843 	/*
844 	 * We use shared locks until we hit the parent of the last cn then
845 	 * we adjust based on the requesting flags.
846 	 */
847 	cnp->cn_lkflags = LK_SHARED;
848 	dp = ndp->ni_startdir;
849 	ndp->ni_startdir = NULLVP;
850 	vn_lock(dp,
851 	    compute_cn_lkflags(dp->v_mount, cnp->cn_lkflags | LK_RETRY,
852 	    cnp->cn_flags));
853 
854 dirloop:
855 	/*
856 	 * Search a new directory.
857 	 *
858 	 * The last component of the filename is left accessible via
859 	 * cnp->cn_nameptr for callers that need the name. Callers needing
860 	 * the name set the SAVENAME flag. When done, they assume
861 	 * responsibility for freeing the pathname buffer.
862 	 */
863 	for (cp = cnp->cn_nameptr; *cp != 0 && *cp != '/'; cp++)
864 		continue;
865 	cnp->cn_namelen = cp - cnp->cn_nameptr;
866 	if (cnp->cn_namelen > NAME_MAX) {
867 		error = ENAMETOOLONG;
868 		goto bad;
869 	}
870 #ifdef NAMEI_DIAGNOSTIC
871 	{ char c = *cp;
872 	*cp = '\0';
873 	printf("{%s}: ", cnp->cn_nameptr);
874 	*cp = c; }
875 #endif
876 	prev_ni_pathlen = ndp->ni_pathlen;
877 	ndp->ni_pathlen -= cnp->cn_namelen;
878 	KASSERT(ndp->ni_pathlen <= PATH_MAX,
879 	    ("%s: ni_pathlen underflow to %zd\n", __func__, ndp->ni_pathlen));
880 	prev_ni_next = ndp->ni_next;
881 	ndp->ni_next = cp;
882 
883 	/*
884 	 * Replace multiple slashes by a single slash and trailing slashes
885 	 * by a null.  This must be done before VOP_LOOKUP() because some
886 	 * fs's don't know about trailing slashes.  Remember if there were
887 	 * trailing slashes to handle symlinks, existing non-directories
888 	 * and non-existing files that won't be directories specially later.
889 	 */
890 	while (*cp == '/' && (cp[1] == '/' || cp[1] == '\0')) {
891 		cp++;
892 		ndp->ni_pathlen--;
893 		if (*cp == '\0') {
894 			*ndp->ni_next = '\0';
895 			cnp->cn_flags |= TRAILINGSLASH;
896 		}
897 	}
898 	ndp->ni_next = cp;
899 
900 	cnp->cn_flags |= MAKEENTRY;
901 	if (*cp == '\0' && docache == 0)
902 		cnp->cn_flags &= ~MAKEENTRY;
903 	if (cnp->cn_namelen == 2 &&
904 	    cnp->cn_nameptr[1] == '.' && cnp->cn_nameptr[0] == '.')
905 		cnp->cn_flags |= ISDOTDOT;
906 	else
907 		cnp->cn_flags &= ~ISDOTDOT;
908 	if (*ndp->ni_next == 0)
909 		cnp->cn_flags |= ISLASTCN;
910 	else
911 		cnp->cn_flags &= ~ISLASTCN;
912 
913 	if ((cnp->cn_flags & ISLASTCN) != 0 &&
914 	    cnp->cn_namelen == 1 && cnp->cn_nameptr[0] == '.' &&
915 	    (cnp->cn_nameiop == DELETE || cnp->cn_nameiop == RENAME)) {
916 		error = EINVAL;
917 		goto bad;
918 	}
919 
920 	nameicap_tracker_add(ndp, dp);
921 
922 	/*
923 	 * Check for degenerate name (e.g. / or "")
924 	 * which is a way of talking about a directory,
925 	 * e.g. like "/." or ".".
926 	 */
927 	if (cnp->cn_nameptr[0] == '\0') {
928 		if (dp->v_type != VDIR) {
929 			error = ENOTDIR;
930 			goto bad;
931 		}
932 		if (cnp->cn_nameiop != LOOKUP) {
933 			error = EISDIR;
934 			goto bad;
935 		}
936 		if (wantparent) {
937 			ndp->ni_dvp = dp;
938 			VREF(dp);
939 		}
940 		ndp->ni_vp = dp;
941 
942 		if (cnp->cn_flags & AUDITVNODE1)
943 			AUDIT_ARG_VNODE1(dp);
944 		else if (cnp->cn_flags & AUDITVNODE2)
945 			AUDIT_ARG_VNODE2(dp);
946 
947 		if (!(cnp->cn_flags & (LOCKPARENT | LOCKLEAF)))
948 			VOP_UNLOCK(dp);
949 		/* XXX This should probably move to the top of function. */
950 		if (cnp->cn_flags & SAVESTART)
951 			panic("lookup: SAVESTART");
952 		goto success;
953 	}
954 
955 	/*
956 	 * Handle "..": five special cases.
957 	 * 0. If doing a capability lookup and lookup_cap_dotdot is
958 	 *    disabled, return ENOTCAPABLE.
959 	 * 1. Return an error if this is the last component of
960 	 *    the name and the operation is DELETE or RENAME.
961 	 * 2. If at root directory (e.g. after chroot)
962 	 *    or at absolute root directory
963 	 *    then ignore it so can't get out.
964 	 * 3. If this vnode is the root of a mounted
965 	 *    filesystem, then replace it with the
966 	 *    vnode which was mounted on so we take the
967 	 *    .. in the other filesystem.
968 	 * 4. If the vnode is the top directory of
969 	 *    the jail or chroot, don't let them out.
970 	 * 5. If doing a capability lookup and lookup_cap_dotdot is
971 	 *    enabled, return ENOTCAPABLE if the lookup would escape
972 	 *    from the initial file descriptor directory.  Checks are
973 	 *    done by ensuring that namei() already traversed the
974 	 *    result of dotdot lookup.
975 	 */
976 	if (cnp->cn_flags & ISDOTDOT) {
977 		if ((ndp->ni_lcf & (NI_LCF_STRICTRELATIVE | NI_LCF_CAP_DOTDOT))
978 		    == NI_LCF_STRICTRELATIVE) {
979 #ifdef KTRACE
980 			if (KTRPOINT(curthread, KTR_CAPFAIL))
981 				ktrcapfail(CAPFAIL_LOOKUP, NULL, NULL);
982 #endif
983 			error = ENOTCAPABLE;
984 			goto bad;
985 		}
986 		if ((cnp->cn_flags & ISLASTCN) != 0 &&
987 		    (cnp->cn_nameiop == DELETE || cnp->cn_nameiop == RENAME)) {
988 			error = EINVAL;
989 			goto bad;
990 		}
991 		for (;;) {
992 			for (pr = cnp->cn_cred->cr_prison; pr != NULL;
993 			     pr = pr->pr_parent)
994 				if (dp == pr->pr_root)
995 					break;
996 			if (dp == ndp->ni_rootdir ||
997 			    dp == ndp->ni_topdir ||
998 			    dp == rootvnode ||
999 			    pr != NULL ||
1000 			    ((dp->v_vflag & VV_ROOT) != 0 &&
1001 			     (cnp->cn_flags & NOCROSSMOUNT) != 0)) {
1002 				ndp->ni_dvp = dp;
1003 				ndp->ni_vp = dp;
1004 				VREF(dp);
1005 				goto nextname;
1006 			}
1007 			if ((dp->v_vflag & VV_ROOT) == 0)
1008 				break;
1009 			if (VN_IS_DOOMED(dp)) {	/* forced unmount */
1010 				error = ENOENT;
1011 				goto bad;
1012 			}
1013 			tdp = dp;
1014 			dp = dp->v_mount->mnt_vnodecovered;
1015 			VREF(dp);
1016 			vput(tdp);
1017 			vn_lock(dp,
1018 			    compute_cn_lkflags(dp->v_mount, cnp->cn_lkflags |
1019 			    LK_RETRY, ISDOTDOT));
1020 			error = nameicap_check_dotdot(ndp, dp);
1021 			if (error != 0) {
1022 #ifdef KTRACE
1023 				if (KTRPOINT(curthread, KTR_CAPFAIL))
1024 					ktrcapfail(CAPFAIL_LOOKUP, NULL, NULL);
1025 #endif
1026 				goto bad;
1027 			}
1028 		}
1029 	}
1030 
1031 	/*
1032 	 * We now have a segment name to search for, and a directory to search.
1033 	 */
1034 unionlookup:
1035 #ifdef MAC
1036 	error = mac_vnode_check_lookup(cnp->cn_thread->td_ucred, dp, cnp);
1037 	if (error)
1038 		goto bad;
1039 #endif
1040 	ndp->ni_dvp = dp;
1041 	ndp->ni_vp = NULL;
1042 	ASSERT_VOP_LOCKED(dp, "lookup");
1043 	/*
1044 	 * If we have a shared lock we may need to upgrade the lock for the
1045 	 * last operation.
1046 	 */
1047 	if ((cnp->cn_flags & LOCKPARENT) && (cnp->cn_flags & ISLASTCN) &&
1048 	    dp != vp_crossmp && VOP_ISLOCKED(dp) == LK_SHARED)
1049 		vn_lock(dp, LK_UPGRADE|LK_RETRY);
1050 	if (VN_IS_DOOMED(dp)) {
1051 		error = ENOENT;
1052 		goto bad;
1053 	}
1054 	/*
1055 	 * If we're looking up the last component and we need an exclusive
1056 	 * lock, adjust our lkflags.
1057 	 */
1058 	if (needs_exclusive_leaf(dp->v_mount, cnp->cn_flags))
1059 		cnp->cn_lkflags = LK_EXCLUSIVE;
1060 #ifdef NAMEI_DIAGNOSTIC
1061 	vn_printf(dp, "lookup in ");
1062 #endif
1063 	lkflags_save = cnp->cn_lkflags;
1064 	cnp->cn_lkflags = compute_cn_lkflags(dp->v_mount, cnp->cn_lkflags,
1065 	    cnp->cn_flags);
1066 	error = VOP_LOOKUP(dp, &ndp->ni_vp, cnp);
1067 	cnp->cn_lkflags = lkflags_save;
1068 	if (error != 0) {
1069 		KASSERT(ndp->ni_vp == NULL, ("leaf should be empty"));
1070 #ifdef NAMEI_DIAGNOSTIC
1071 		printf("not found\n");
1072 #endif
1073 		if ((error == ENOENT) &&
1074 		    (dp->v_vflag & VV_ROOT) && (dp->v_mount != NULL) &&
1075 		    (dp->v_mount->mnt_flag & MNT_UNION)) {
1076 			tdp = dp;
1077 			dp = dp->v_mount->mnt_vnodecovered;
1078 			VREF(dp);
1079 			vput(tdp);
1080 			vn_lock(dp,
1081 			    compute_cn_lkflags(dp->v_mount, cnp->cn_lkflags |
1082 			    LK_RETRY, cnp->cn_flags));
1083 			nameicap_tracker_add(ndp, dp);
1084 			goto unionlookup;
1085 		}
1086 
1087 		if (error == ERELOOKUP) {
1088 			vref(dp);
1089 			ndp->ni_vp = dp;
1090 			error = 0;
1091 			relookup = 1;
1092 			goto good;
1093 		}
1094 
1095 		if (error != EJUSTRETURN)
1096 			goto bad;
1097 		/*
1098 		 * At this point, we know we're at the end of the
1099 		 * pathname.  If creating / renaming, we can consider
1100 		 * allowing the file or directory to be created / renamed,
1101 		 * provided we're not on a read-only filesystem.
1102 		 */
1103 		if (rdonly) {
1104 			error = EROFS;
1105 			goto bad;
1106 		}
1107 		/* trailing slash only allowed for directories */
1108 		if ((cnp->cn_flags & TRAILINGSLASH) &&
1109 		    !(cnp->cn_flags & WILLBEDIR)) {
1110 			error = ENOENT;
1111 			goto bad;
1112 		}
1113 		if ((cnp->cn_flags & LOCKPARENT) == 0)
1114 			VOP_UNLOCK(dp);
1115 		/*
1116 		 * We return with ni_vp NULL to indicate that the entry
1117 		 * doesn't currently exist, leaving a pointer to the
1118 		 * (possibly locked) directory vnode in ndp->ni_dvp.
1119 		 */
1120 		if (cnp->cn_flags & SAVESTART) {
1121 			ndp->ni_startdir = ndp->ni_dvp;
1122 			VREF(ndp->ni_startdir);
1123 		}
1124 		goto success;
1125 	}
1126 
1127 good:
1128 #ifdef NAMEI_DIAGNOSTIC
1129 	printf("found\n");
1130 #endif
1131 	dp = ndp->ni_vp;
1132 
1133 	/*
1134 	 * Check to see if the vnode has been mounted on;
1135 	 * if so find the root of the mounted filesystem.
1136 	 */
1137 	while (dp->v_type == VDIR && (mp = dp->v_mountedhere) &&
1138 	       (cnp->cn_flags & NOCROSSMOUNT) == 0) {
1139 		if (vfs_busy(mp, 0))
1140 			continue;
1141 		vput(dp);
1142 		if (dp != ndp->ni_dvp)
1143 			vput(ndp->ni_dvp);
1144 		else
1145 			vrele(ndp->ni_dvp);
1146 		vrefact(vp_crossmp);
1147 		ndp->ni_dvp = vp_crossmp;
1148 		error = VFS_ROOT(mp, compute_cn_lkflags(mp, cnp->cn_lkflags,
1149 		    cnp->cn_flags), &tdp);
1150 		vfs_unbusy(mp);
1151 		if (vn_lock(vp_crossmp, LK_SHARED | LK_NOWAIT))
1152 			panic("vp_crossmp exclusively locked or reclaimed");
1153 		if (error) {
1154 			dpunlocked = 1;
1155 			goto bad2;
1156 		}
1157 		ndp->ni_vp = dp = tdp;
1158 	}
1159 
1160 	/*
1161 	 * Check for symbolic link
1162 	 */
1163 	if ((dp->v_type == VLNK) &&
1164 	    ((cnp->cn_flags & FOLLOW) || (cnp->cn_flags & TRAILINGSLASH) ||
1165 	     *ndp->ni_next == '/')) {
1166 		cnp->cn_flags |= ISSYMLINK;
1167 		if (VN_IS_DOOMED(dp)) {
1168 			/*
1169 			 * We can't know whether the directory was mounted with
1170 			 * NOSYMFOLLOW, so we can't follow safely.
1171 			 */
1172 			error = ENOENT;
1173 			goto bad2;
1174 		}
1175 		if (dp->v_mount->mnt_flag & MNT_NOSYMFOLLOW) {
1176 			error = EACCES;
1177 			goto bad2;
1178 		}
1179 		/*
1180 		 * Symlink code always expects an unlocked dvp.
1181 		 */
1182 		if (ndp->ni_dvp != ndp->ni_vp) {
1183 			VOP_UNLOCK(ndp->ni_dvp);
1184 			ni_dvp_unlocked = 1;
1185 		}
1186 		goto success;
1187 	}
1188 
1189 nextname:
1190 	/*
1191 	 * Not a symbolic link that we will follow.  Continue with the
1192 	 * next component if there is any; otherwise, we're done.
1193 	 */
1194 	KASSERT((cnp->cn_flags & ISLASTCN) || *ndp->ni_next == '/',
1195 	    ("lookup: invalid path state."));
1196 	if (relookup) {
1197 		relookup = 0;
1198 		ndp->ni_pathlen = prev_ni_pathlen;
1199 		ndp->ni_next = prev_ni_next;
1200 		if (ndp->ni_dvp != dp)
1201 			vput(ndp->ni_dvp);
1202 		else
1203 			vrele(ndp->ni_dvp);
1204 		goto dirloop;
1205 	}
1206 	if (cnp->cn_flags & ISDOTDOT) {
1207 		error = nameicap_check_dotdot(ndp, ndp->ni_vp);
1208 		if (error != 0) {
1209 #ifdef KTRACE
1210 			if (KTRPOINT(curthread, KTR_CAPFAIL))
1211 				ktrcapfail(CAPFAIL_LOOKUP, NULL, NULL);
1212 #endif
1213 			goto bad2;
1214 		}
1215 	}
1216 	if (*ndp->ni_next == '/') {
1217 		cnp->cn_nameptr = ndp->ni_next;
1218 		while (*cnp->cn_nameptr == '/') {
1219 			cnp->cn_nameptr++;
1220 			ndp->ni_pathlen--;
1221 		}
1222 		if (ndp->ni_dvp != dp)
1223 			vput(ndp->ni_dvp);
1224 		else
1225 			vrele(ndp->ni_dvp);
1226 		goto dirloop;
1227 	}
1228 	/*
1229 	 * If we're processing a path with a trailing slash,
1230 	 * check that the end result is a directory.
1231 	 */
1232 	if ((cnp->cn_flags & TRAILINGSLASH) && dp->v_type != VDIR) {
1233 		error = ENOTDIR;
1234 		goto bad2;
1235 	}
1236 	/*
1237 	 * Disallow directory write attempts on read-only filesystems.
1238 	 */
1239 	if (rdonly &&
1240 	    (cnp->cn_nameiop == DELETE || cnp->cn_nameiop == RENAME)) {
1241 		error = EROFS;
1242 		goto bad2;
1243 	}
1244 	if (cnp->cn_flags & SAVESTART) {
1245 		ndp->ni_startdir = ndp->ni_dvp;
1246 		VREF(ndp->ni_startdir);
1247 	}
1248 	if (!wantparent) {
1249 		ni_dvp_unlocked = 2;
1250 		if (ndp->ni_dvp != dp)
1251 			vput(ndp->ni_dvp);
1252 		else
1253 			vrele(ndp->ni_dvp);
1254 	} else if ((cnp->cn_flags & LOCKPARENT) == 0 && ndp->ni_dvp != dp) {
1255 		VOP_UNLOCK(ndp->ni_dvp);
1256 		ni_dvp_unlocked = 1;
1257 	}
1258 
1259 	if (cnp->cn_flags & AUDITVNODE1)
1260 		AUDIT_ARG_VNODE1(dp);
1261 	else if (cnp->cn_flags & AUDITVNODE2)
1262 		AUDIT_ARG_VNODE2(dp);
1263 
1264 	if ((cnp->cn_flags & LOCKLEAF) == 0)
1265 		VOP_UNLOCK(dp);
1266 success:
1267 	/*
1268 	 * FIXME: for lookups which only cross a mount point to fetch the
1269 	 * root vnode, ni_dvp will be set to vp_crossmp. This can be a problem
1270 	 * if either WANTPARENT or LOCKPARENT is set.
1271 	 */
1272 	/*
1273 	 * Because of shared lookup we may have the vnode shared locked, but
1274 	 * the caller may want it to be exclusively locked.
1275 	 */
1276 	if (needs_exclusive_leaf(dp->v_mount, cnp->cn_flags) &&
1277 	    VOP_ISLOCKED(dp) != LK_EXCLUSIVE) {
1278 		vn_lock(dp, LK_UPGRADE | LK_RETRY);
1279 		if (VN_IS_DOOMED(dp)) {
1280 			error = ENOENT;
1281 			goto bad2;
1282 		}
1283 	}
1284 	if (ndp->ni_vp != NULL && ndp->ni_vp->v_type == VDIR)
1285 		nameicap_tracker_add(ndp, ndp->ni_vp);
1286 	return (0);
1287 
1288 bad2:
1289 	if (ni_dvp_unlocked != 2) {
1290 		if (dp != ndp->ni_dvp && !ni_dvp_unlocked)
1291 			vput(ndp->ni_dvp);
1292 		else
1293 			vrele(ndp->ni_dvp);
1294 	}
1295 bad:
1296 	if (!dpunlocked)
1297 		vput(dp);
1298 	ndp->ni_vp = NULL;
1299 	return (error);
1300 }
1301 
1302 /*
1303  * relookup - lookup a path name component
1304  *    Used by lookup to re-acquire things.
1305  */
1306 int
1307 relookup(struct vnode *dvp, struct vnode **vpp, struct componentname *cnp)
1308 {
1309 	struct vnode *dp = NULL;		/* the directory we are searching */
1310 	int wantparent;			/* 1 => wantparent or lockparent flag */
1311 	int rdonly;			/* lookup read-only flag bit */
1312 	int error = 0;
1313 
1314 	KASSERT(cnp->cn_flags & ISLASTCN,
1315 	    ("relookup: Not given last component."));
1316 	/*
1317 	 * Setup: break out flag bits into variables.
1318 	 */
1319 	wantparent = cnp->cn_flags & (LOCKPARENT|WANTPARENT);
1320 	KASSERT(wantparent, ("relookup: parent not wanted."));
1321 	rdonly = cnp->cn_flags & RDONLY;
1322 	cnp->cn_flags &= ~ISSYMLINK;
1323 	dp = dvp;
1324 	cnp->cn_lkflags = LK_EXCLUSIVE;
1325 	vn_lock(dp, LK_EXCLUSIVE | LK_RETRY);
1326 
1327 	/*
1328 	 * Search a new directory.
1329 	 *
1330 	 * The last component of the filename is left accessible via
1331 	 * cnp->cn_nameptr for callers that need the name. Callers needing
1332 	 * the name set the SAVENAME flag. When done, they assume
1333 	 * responsibility for freeing the pathname buffer.
1334 	 */
1335 #ifdef NAMEI_DIAGNOSTIC
1336 	printf("{%s}: ", cnp->cn_nameptr);
1337 #endif
1338 
1339 	/*
1340 	 * Check for "" which represents the root directory after slash
1341 	 * removal.
1342 	 */
1343 	if (cnp->cn_nameptr[0] == '\0') {
1344 		/*
1345 		 * Support only LOOKUP for "/" because lookup()
1346 		 * can't succeed for CREATE, DELETE and RENAME.
1347 		 */
1348 		KASSERT(cnp->cn_nameiop == LOOKUP, ("nameiop must be LOOKUP"));
1349 		KASSERT(dp->v_type == VDIR, ("dp is not a directory"));
1350 
1351 		if (!(cnp->cn_flags & LOCKLEAF))
1352 			VOP_UNLOCK(dp);
1353 		*vpp = dp;
1354 		/* XXX This should probably move to the top of function. */
1355 		if (cnp->cn_flags & SAVESTART)
1356 			panic("lookup: SAVESTART");
1357 		return (0);
1358 	}
1359 
1360 	if (cnp->cn_flags & ISDOTDOT)
1361 		panic ("relookup: lookup on dot-dot");
1362 
1363 	/*
1364 	 * We now have a segment name to search for, and a directory to search.
1365 	 */
1366 #ifdef NAMEI_DIAGNOSTIC
1367 	vn_printf(dp, "search in ");
1368 #endif
1369 	if ((error = VOP_LOOKUP(dp, vpp, cnp)) != 0) {
1370 		KASSERT(*vpp == NULL, ("leaf should be empty"));
1371 		if (error != EJUSTRETURN)
1372 			goto bad;
1373 		/*
1374 		 * If creating and at end of pathname, then can consider
1375 		 * allowing file to be created.
1376 		 */
1377 		if (rdonly) {
1378 			error = EROFS;
1379 			goto bad;
1380 		}
1381 		/* ASSERT(dvp == ndp->ni_startdir) */
1382 		if (cnp->cn_flags & SAVESTART)
1383 			VREF(dvp);
1384 		if ((cnp->cn_flags & LOCKPARENT) == 0)
1385 			VOP_UNLOCK(dp);
1386 		/*
1387 		 * We return with ni_vp NULL to indicate that the entry
1388 		 * doesn't currently exist, leaving a pointer to the
1389 		 * (possibly locked) directory vnode in ndp->ni_dvp.
1390 		 */
1391 		return (0);
1392 	}
1393 
1394 	dp = *vpp;
1395 
1396 	/*
1397 	 * Disallow directory write attempts on read-only filesystems.
1398 	 */
1399 	if (rdonly &&
1400 	    (cnp->cn_nameiop == DELETE || cnp->cn_nameiop == RENAME)) {
1401 		if (dvp == dp)
1402 			vrele(dvp);
1403 		else
1404 			vput(dvp);
1405 		error = EROFS;
1406 		goto bad;
1407 	}
1408 	/*
1409 	 * Set the parent lock/ref state to the requested state.
1410 	 */
1411 	if ((cnp->cn_flags & LOCKPARENT) == 0 && dvp != dp) {
1412 		if (wantparent)
1413 			VOP_UNLOCK(dvp);
1414 		else
1415 			vput(dvp);
1416 	} else if (!wantparent)
1417 		vrele(dvp);
1418 	/*
1419 	 * Check for symbolic link
1420 	 */
1421 	KASSERT(dp->v_type != VLNK || !(cnp->cn_flags & FOLLOW),
1422 	    ("relookup: symlink found.\n"));
1423 
1424 	/* ASSERT(dvp == ndp->ni_startdir) */
1425 	if (cnp->cn_flags & SAVESTART)
1426 		VREF(dvp);
1427 
1428 	if ((cnp->cn_flags & LOCKLEAF) == 0)
1429 		VOP_UNLOCK(dp);
1430 	return (0);
1431 bad:
1432 	vput(dp);
1433 	*vpp = NULL;
1434 	return (error);
1435 }
1436 
1437 /*
1438  * Free data allocated by namei(); see namei(9) for details.
1439  */
1440 void
1441 NDFREE_PNBUF(struct nameidata *ndp)
1442 {
1443 
1444 	if ((ndp->ni_cnd.cn_flags & HASBUF) != 0) {
1445 		MPASS((ndp->ni_cnd.cn_flags & (SAVENAME | SAVESTART)) != 0);
1446 		uma_zfree(namei_zone, ndp->ni_cnd.cn_pnbuf);
1447 		ndp->ni_cnd.cn_flags &= ~HASBUF;
1448 	}
1449 }
1450 
1451 void
1452 (NDFREE)(struct nameidata *ndp, const u_int flags)
1453 {
1454 	int unlock_dvp;
1455 	int unlock_vp;
1456 
1457 	unlock_dvp = 0;
1458 	unlock_vp = 0;
1459 
1460 	if (!(flags & NDF_NO_FREE_PNBUF)) {
1461 		NDFREE_PNBUF(ndp);
1462 	}
1463 	if (!(flags & NDF_NO_VP_UNLOCK) &&
1464 	    (ndp->ni_cnd.cn_flags & LOCKLEAF) && ndp->ni_vp)
1465 		unlock_vp = 1;
1466 	if (!(flags & NDF_NO_DVP_UNLOCK) &&
1467 	    (ndp->ni_cnd.cn_flags & LOCKPARENT) &&
1468 	    ndp->ni_dvp != ndp->ni_vp)
1469 		unlock_dvp = 1;
1470 	if (!(flags & NDF_NO_VP_RELE) && ndp->ni_vp) {
1471 		if (unlock_vp) {
1472 			vput(ndp->ni_vp);
1473 			unlock_vp = 0;
1474 		} else
1475 			vrele(ndp->ni_vp);
1476 		ndp->ni_vp = NULL;
1477 	}
1478 	if (unlock_vp)
1479 		VOP_UNLOCK(ndp->ni_vp);
1480 	if (!(flags & NDF_NO_DVP_RELE) &&
1481 	    (ndp->ni_cnd.cn_flags & (LOCKPARENT|WANTPARENT))) {
1482 		if (unlock_dvp) {
1483 			vput(ndp->ni_dvp);
1484 			unlock_dvp = 0;
1485 		} else
1486 			vrele(ndp->ni_dvp);
1487 		ndp->ni_dvp = NULL;
1488 	}
1489 	if (unlock_dvp)
1490 		VOP_UNLOCK(ndp->ni_dvp);
1491 	if (!(flags & NDF_NO_STARTDIR_RELE) &&
1492 	    (ndp->ni_cnd.cn_flags & SAVESTART)) {
1493 		vrele(ndp->ni_startdir);
1494 		ndp->ni_startdir = NULL;
1495 	}
1496 }
1497 
1498 #ifdef INVARIANTS
1499 /*
1500  * Validate the final state of ndp after the lookup.
1501  *
1502  * Historically filesystems were allowed to modify cn_flags. Most notably they
1503  * can add SAVENAME to the request, resulting in HASBUF and pushing subsequent
1504  * clean up to the consumer. In practice this seems to only concern != LOOKUP
1505  * operations.
1506  *
1507  * As a step towards stricter API contract this routine validates the state to
1508  * clean up. Note validation is a work in progress with the intent of becoming
1509  * stricter over time.
1510  */
1511 #define NDMODIFYINGFLAGS (LOCKLEAF | LOCKPARENT | WANTPARENT | SAVENAME | SAVESTART | HASBUF)
1512 void
1513 NDVALIDATE(struct nameidata *ndp)
1514 {
1515 	struct componentname *cnp;
1516 	u_int64_t used, orig;
1517 
1518 	cnp = &ndp->ni_cnd;
1519 	orig = cnp->cn_origflags;
1520 	used = cnp->cn_flags;
1521 	switch (cnp->cn_nameiop) {
1522 	case LOOKUP:
1523 		/*
1524 		 * For plain lookup we require strict conformance -- nothing
1525 		 * to clean up if it was not requested by the caller.
1526 		 */
1527 		orig &= NDMODIFYINGFLAGS;
1528 		used &= NDMODIFYINGFLAGS;
1529 		if ((orig & (SAVENAME | SAVESTART)) != 0)
1530 			orig |= HASBUF;
1531 		if (orig != used) {
1532 			goto out_mismatch;
1533 		}
1534 		break;
1535 	case CREATE:
1536 	case DELETE:
1537 	case RENAME:
1538 		/*
1539 		 * Some filesystems set SAVENAME to provoke HASBUF, accomodate
1540 		 * for it until it gets fixed.
1541 		 */
1542 		orig &= NDMODIFYINGFLAGS;
1543 		orig |= (SAVENAME | HASBUF);
1544 		used &= NDMODIFYINGFLAGS;
1545 		used |= (SAVENAME | HASBUF);
1546 		if (orig != used) {
1547 			goto out_mismatch;
1548 		}
1549 		break;
1550 	}
1551 	return;
1552 out_mismatch:
1553 	panic("%s: mismatched flags for op %d: added %" PRIx64 ", "
1554 	    "removed %" PRIx64" (%" PRIx64" != %" PRIx64"; stored %" PRIx64" != %" PRIx64")",
1555 	    __func__, cnp->cn_nameiop, used & ~orig, orig &~ used,
1556 	    orig, used, cnp->cn_origflags, cnp->cn_flags);
1557 }
1558 #endif
1559 
1560 /*
1561  * Determine if there is a suitable alternate filename under the specified
1562  * prefix for the specified path.  If the create flag is set, then the
1563  * alternate prefix will be used so long as the parent directory exists.
1564  * This is used by the various compatibility ABIs so that Linux binaries prefer
1565  * files under /compat/linux for example.  The chosen path (whether under
1566  * the prefix or under /) is returned in a kernel malloc'd buffer pointed
1567  * to by pathbuf.  The caller is responsible for free'ing the buffer from
1568  * the M_TEMP bucket if one is returned.
1569  */
1570 int
1571 kern_alternate_path(struct thread *td, const char *prefix, const char *path,
1572     enum uio_seg pathseg, char **pathbuf, int create, int dirfd)
1573 {
1574 	struct nameidata nd, ndroot;
1575 	char *ptr, *buf, *cp;
1576 	size_t len, sz;
1577 	int error;
1578 
1579 	buf = (char *) malloc(MAXPATHLEN, M_TEMP, M_WAITOK);
1580 	*pathbuf = buf;
1581 
1582 	/* Copy the prefix into the new pathname as a starting point. */
1583 	len = strlcpy(buf, prefix, MAXPATHLEN);
1584 	if (len >= MAXPATHLEN) {
1585 		*pathbuf = NULL;
1586 		free(buf, M_TEMP);
1587 		return (EINVAL);
1588 	}
1589 	sz = MAXPATHLEN - len;
1590 	ptr = buf + len;
1591 
1592 	/* Append the filename to the prefix. */
1593 	if (pathseg == UIO_SYSSPACE)
1594 		error = copystr(path, ptr, sz, &len);
1595 	else
1596 		error = copyinstr(path, ptr, sz, &len);
1597 
1598 	if (error) {
1599 		*pathbuf = NULL;
1600 		free(buf, M_TEMP);
1601 		return (error);
1602 	}
1603 
1604 	/* Only use a prefix with absolute pathnames. */
1605 	if (*ptr != '/') {
1606 		error = EINVAL;
1607 		goto keeporig;
1608 	}
1609 
1610 	if (dirfd != AT_FDCWD) {
1611 		/*
1612 		 * We want the original because the "prefix" is
1613 		 * included in the already opened dirfd.
1614 		 */
1615 		bcopy(ptr, buf, len);
1616 		return (0);
1617 	}
1618 
1619 	/*
1620 	 * We know that there is a / somewhere in this pathname.
1621 	 * Search backwards for it, to find the file's parent dir
1622 	 * to see if it exists in the alternate tree. If it does,
1623 	 * and we want to create a file (cflag is set). We don't
1624 	 * need to worry about the root comparison in this case.
1625 	 */
1626 
1627 	if (create) {
1628 		for (cp = &ptr[len] - 1; *cp != '/'; cp--);
1629 		*cp = '\0';
1630 
1631 		NDINIT(&nd, LOOKUP, NOFOLLOW, UIO_SYSSPACE, buf, td);
1632 		error = namei(&nd);
1633 		*cp = '/';
1634 		if (error != 0)
1635 			goto keeporig;
1636 	} else {
1637 		NDINIT(&nd, LOOKUP, NOFOLLOW, UIO_SYSSPACE, buf, td);
1638 
1639 		error = namei(&nd);
1640 		if (error != 0)
1641 			goto keeporig;
1642 
1643 		/*
1644 		 * We now compare the vnode of the prefix to the one
1645 		 * vnode asked. If they resolve to be the same, then we
1646 		 * ignore the match so that the real root gets used.
1647 		 * This avoids the problem of traversing "../.." to find the
1648 		 * root directory and never finding it, because "/" resolves
1649 		 * to the emulation root directory. This is expensive :-(
1650 		 */
1651 		NDINIT(&ndroot, LOOKUP, FOLLOW, UIO_SYSSPACE, prefix,
1652 		    td);
1653 
1654 		/* We shouldn't ever get an error from this namei(). */
1655 		error = namei(&ndroot);
1656 		if (error == 0) {
1657 			if (nd.ni_vp == ndroot.ni_vp)
1658 				error = ENOENT;
1659 
1660 			NDFREE(&ndroot, NDF_ONLY_PNBUF);
1661 			vrele(ndroot.ni_vp);
1662 		}
1663 	}
1664 
1665 	NDFREE(&nd, NDF_ONLY_PNBUF);
1666 	vrele(nd.ni_vp);
1667 
1668 keeporig:
1669 	/* If there was an error, use the original path name. */
1670 	if (error)
1671 		bcopy(ptr, buf, len);
1672 	return (error);
1673 }
1674