xref: /freebsd/sys/kern/vfs_lookup.c (revision ff19fd624233a938b6a09ac75a87a2c69d65df08)
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 MALLOC_DEFINE(M_NAMEITRACKER, "namei_tracker", "namei tracking for dotdot");
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 	vfs_vector_op_register(&crossmp_vnodeops);
161 	getnewvnode("crossmp", NULL, &crossmp_vnodeops, &vp_crossmp);
162 }
163 SYSINIT(vfs, SI_SUB_VFS, SI_ORDER_SECOND, nameiinit, NULL);
164 
165 static int lookup_cap_dotdot = 1;
166 SYSCTL_INT(_vfs, OID_AUTO, lookup_cap_dotdot, CTLFLAG_RWTUN,
167     &lookup_cap_dotdot, 0,
168     "enables \"..\" components in path lookup in capability mode");
169 static int lookup_cap_dotdot_nonlocal = 1;
170 SYSCTL_INT(_vfs, OID_AUTO, lookup_cap_dotdot_nonlocal, CTLFLAG_RWTUN,
171     &lookup_cap_dotdot_nonlocal, 0,
172     "enables \"..\" components in path lookup in capability mode "
173     "on non-local mount");
174 
175 static void
176 nameicap_tracker_add(struct nameidata *ndp, struct vnode *dp)
177 {
178 	struct nameicap_tracker *nt;
179 	struct componentname *cnp;
180 
181 	if ((ndp->ni_lcf & NI_LCF_CAP_DOTDOT) == 0 || dp->v_type != VDIR)
182 		return;
183 	cnp = &ndp->ni_cnd;
184 	if ((cnp->cn_flags & BENEATH) != 0 &&
185 	    (ndp->ni_lcf & NI_LCF_BENEATH_LATCHED) == 0) {
186 		MPASS((ndp->ni_lcf & NI_LCF_LATCH) != 0);
187 		if (dp != ndp->ni_beneath_latch)
188 			return;
189 		ndp->ni_lcf |= NI_LCF_BENEATH_LATCHED;
190 	}
191 	nt = malloc(sizeof(*nt), M_NAMEITRACKER, M_WAITOK);
192 	vhold(dp);
193 	nt->dp = dp;
194 	TAILQ_INSERT_TAIL(&ndp->ni_cap_tracker, nt, nm_link);
195 }
196 
197 static void
198 nameicap_cleanup(struct nameidata *ndp, bool clean_latch)
199 {
200 	struct nameicap_tracker *nt, *nt1;
201 
202 	KASSERT(TAILQ_EMPTY(&ndp->ni_cap_tracker) ||
203 	    (ndp->ni_lcf & NI_LCF_CAP_DOTDOT) != 0, ("not strictrelative"));
204 	TAILQ_FOREACH_SAFE(nt, &ndp->ni_cap_tracker, nm_link, nt1) {
205 		TAILQ_REMOVE(&ndp->ni_cap_tracker, nt, nm_link);
206 		vdrop(nt->dp);
207 		free(nt, M_NAMEITRACKER);
208 	}
209 	if (clean_latch && (ndp->ni_lcf & NI_LCF_LATCH) != 0) {
210 		ndp->ni_lcf &= ~NI_LCF_LATCH;
211 		vrele(ndp->ni_beneath_latch);
212 	}
213 }
214 
215 /*
216  * For dotdot lookups in capability mode, only allow the component
217  * lookup to succeed if the resulting directory was already traversed
218  * during the operation.  This catches situations where already
219  * traversed directory is moved to different parent, and then we walk
220  * over it with dotdots.
221  *
222  * Also allow to force failure of dotdot lookups for non-local
223  * filesystems, where external agents might assist local lookups to
224  * escape the compartment.
225  */
226 static int
227 nameicap_check_dotdot(struct nameidata *ndp, struct vnode *dp)
228 {
229 	struct nameicap_tracker *nt;
230 	struct mount *mp;
231 
232 	if ((ndp->ni_lcf & NI_LCF_CAP_DOTDOT) == 0 || dp == NULL ||
233 	    dp->v_type != VDIR)
234 		return (0);
235 	mp = dp->v_mount;
236 	if (lookup_cap_dotdot_nonlocal == 0 && mp != NULL &&
237 	    (mp->mnt_flag & MNT_LOCAL) == 0)
238 		return (ENOTCAPABLE);
239 	TAILQ_FOREACH_REVERSE(nt, &ndp->ni_cap_tracker, nameicap_tracker_head,
240 	    nm_link) {
241 		if ((ndp->ni_lcf & NI_LCF_LATCH) != 0 &&
242 		    ndp->ni_beneath_latch == nt->dp) {
243 			ndp->ni_lcf &= ~NI_LCF_BENEATH_LATCHED;
244 			nameicap_cleanup(ndp, false);
245 			return (0);
246 		}
247 		if (dp == nt->dp)
248 			return (0);
249 	}
250 	return (ENOTCAPABLE);
251 }
252 
253 static void
254 namei_cleanup_cnp(struct componentname *cnp)
255 {
256 
257 	uma_zfree(namei_zone, cnp->cn_pnbuf);
258 #ifdef DIAGNOSTIC
259 	cnp->cn_pnbuf = NULL;
260 	cnp->cn_nameptr = NULL;
261 #endif
262 }
263 
264 static int
265 namei_handle_root(struct nameidata *ndp, struct vnode **dpp)
266 {
267 	struct componentname *cnp;
268 
269 	cnp = &ndp->ni_cnd;
270 	if ((ndp->ni_lcf & NI_LCF_STRICTRELATIVE) != 0) {
271 #ifdef KTRACE
272 		if (KTRPOINT(curthread, KTR_CAPFAIL))
273 			ktrcapfail(CAPFAIL_LOOKUP, NULL, NULL);
274 #endif
275 		return (ENOTCAPABLE);
276 	}
277 	if ((cnp->cn_flags & BENEATH) != 0) {
278 		ndp->ni_lcf |= NI_LCF_BENEATH_ABS;
279 		ndp->ni_lcf &= ~NI_LCF_BENEATH_LATCHED;
280 		nameicap_cleanup(ndp, false);
281 	}
282 	while (*(cnp->cn_nameptr) == '/') {
283 		cnp->cn_nameptr++;
284 		ndp->ni_pathlen--;
285 	}
286 	*dpp = ndp->ni_rootdir;
287 	vrefact(*dpp);
288 	return (0);
289 }
290 
291 static int
292 namei_setup(struct nameidata *ndp, struct vnode **dpp, struct pwd **pwdp)
293 {
294 	struct componentname *cnp;
295 	struct file *dfp;
296 	struct thread *td;
297 	struct pwd *pwd;
298 	cap_rights_t rights;
299 	struct filecaps dirfd_caps;
300 	int error;
301 	bool startdir_used;
302 
303 	cnp = &ndp->ni_cnd;
304 	td = cnp->cn_thread;
305 
306 	startdir_used = false;
307 	*pwdp = NULL;
308 	*dpp = NULL;
309 
310 #ifdef CAPABILITY_MODE
311 	/*
312 	 * In capability mode, lookups must be restricted to happen in
313 	 * the subtree with the root specified by the file descriptor:
314 	 * - The root must be real file descriptor, not the pseudo-descriptor
315 	 *   AT_FDCWD.
316 	 * - The passed path must be relative and not absolute.
317 	 * - If lookup_cap_dotdot is disabled, path must not contain the
318 	 *   '..' components.
319 	 * - If lookup_cap_dotdot is enabled, we verify that all '..'
320 	 *   components lookups result in the directories which were
321 	 *   previously walked by us, which prevents an escape from
322 	 *   the relative root.
323 	 */
324 	if (IN_CAPABILITY_MODE(td) && (cnp->cn_flags & NOCAPCHECK) == 0) {
325 		ndp->ni_lcf |= NI_LCF_STRICTRELATIVE;
326 		ndp->ni_resflags |= NIRES_STRICTREL;
327 		if (ndp->ni_dirfd == AT_FDCWD) {
328 #ifdef KTRACE
329 			if (KTRPOINT(td, KTR_CAPFAIL))
330 				ktrcapfail(CAPFAIL_LOOKUP, NULL, NULL);
331 #endif
332 			return (ECAPMODE);
333 		}
334 	}
335 #endif
336 	error = 0;
337 
338 	/*
339 	 * Get starting point for the translation.
340 	 */
341 	pwd = pwd_hold(td);
342 	/*
343 	 * The reference on ni_rootdir is acquired in the block below to avoid
344 	 * back-to-back atomics for absolute lookups.
345 	 */
346 	ndp->ni_rootdir = pwd->pwd_rdir;
347 	ndp->ni_topdir = pwd->pwd_jdir;
348 
349 	if (cnp->cn_pnbuf[0] == '/') {
350 		ndp->ni_resflags |= NIRES_ABS;
351 		error = namei_handle_root(ndp, dpp);
352 	} else {
353 		if (ndp->ni_startdir != NULL) {
354 			*dpp = ndp->ni_startdir;
355 			startdir_used = true;
356 		} else if (ndp->ni_dirfd == AT_FDCWD) {
357 			*dpp = pwd->pwd_cdir;
358 			vrefact(*dpp);
359 		} else {
360 			rights = *ndp->ni_rightsneeded;
361 			cap_rights_set_one(&rights, CAP_LOOKUP);
362 
363 			if (cnp->cn_flags & AUDITVNODE1)
364 				AUDIT_ARG_ATFD1(ndp->ni_dirfd);
365 			if (cnp->cn_flags & AUDITVNODE2)
366 				AUDIT_ARG_ATFD2(ndp->ni_dirfd);
367 			/*
368 			 * Effectively inlined fgetvp_rights, because we need to
369 			 * inspect the file as well as grabbing the vnode.
370 			 */
371 			error = fget_cap(td, ndp->ni_dirfd, &rights,
372 			    &dfp, &ndp->ni_filecaps);
373 			if (error != 0) {
374 				/*
375 				 * Preserve the error; it should either be EBADF
376 				 * or capability-related, both of which can be
377 				 * safely returned to the caller.
378 				 */
379 			} else {
380 				if (dfp->f_ops == &badfileops) {
381 					error = EBADF;
382 				} else if (dfp->f_vnode == NULL) {
383 					error = ENOTDIR;
384 				} else {
385 					*dpp = dfp->f_vnode;
386 					vrefact(*dpp);
387 
388 					if ((dfp->f_flag & FSEARCH) != 0)
389 						cnp->cn_flags |= NOEXECCHECK;
390 				}
391 				fdrop(dfp, td);
392 			}
393 #ifdef CAPABILITIES
394 			/*
395 			 * If file descriptor doesn't have all rights,
396 			 * all lookups relative to it must also be
397 			 * strictly relative.
398 			 */
399 			CAP_ALL(&rights);
400 			if (!cap_rights_contains(&ndp->ni_filecaps.fc_rights,
401 			    &rights) ||
402 			    ndp->ni_filecaps.fc_fcntls != CAP_FCNTL_ALL ||
403 			    ndp->ni_filecaps.fc_nioctls != -1) {
404 				ndp->ni_lcf |= NI_LCF_STRICTRELATIVE;
405 				ndp->ni_resflags |= NIRES_STRICTREL;
406 			}
407 #endif
408 		}
409 		if (error == 0 && (*dpp)->v_type != VDIR)
410 			error = ENOTDIR;
411 	}
412 	if (error == 0 && (cnp->cn_flags & BENEATH) != 0) {
413 		if (ndp->ni_dirfd == AT_FDCWD) {
414 			ndp->ni_beneath_latch = pwd->pwd_cdir;
415 			vrefact(ndp->ni_beneath_latch);
416 		} else {
417 			rights = *ndp->ni_rightsneeded;
418 			cap_rights_set_one(&rights, CAP_LOOKUP);
419 			error = fgetvp_rights(td, ndp->ni_dirfd, &rights,
420 			    &dirfd_caps, &ndp->ni_beneath_latch);
421 			if (error == 0 && (*dpp)->v_type != VDIR) {
422 				vrele(ndp->ni_beneath_latch);
423 				error = ENOTDIR;
424 			}
425 		}
426 		if (error == 0)
427 			ndp->ni_lcf |= NI_LCF_LATCH;
428 	}
429 	if (error == 0 && (cnp->cn_flags & RBENEATH) != 0) {
430 		if (cnp->cn_pnbuf[0] == '/' ||
431 		    (ndp->ni_lcf & NI_LCF_BENEATH_ABS) != 0) {
432 			error = EINVAL;
433 		} else if ((ndp->ni_lcf & NI_LCF_STRICTRELATIVE) == 0) {
434 			ndp->ni_lcf |= NI_LCF_STRICTRELATIVE |
435 			    NI_LCF_CAP_DOTDOT;
436 		}
437 	}
438 
439 	/*
440 	 * If we are auditing the kernel pathname, save the user pathname.
441 	 */
442 	if (cnp->cn_flags & AUDITVNODE1)
443 		AUDIT_ARG_UPATH1_VP(td, ndp->ni_rootdir, *dpp, cnp->cn_pnbuf);
444 	if (cnp->cn_flags & AUDITVNODE2)
445 		AUDIT_ARG_UPATH2_VP(td, ndp->ni_rootdir, *dpp, cnp->cn_pnbuf);
446 	if (ndp->ni_startdir != NULL && !startdir_used)
447 		vrele(ndp->ni_startdir);
448 	if (error != 0) {
449 		if (*dpp != NULL)
450 			vrele(*dpp);
451 		pwd_drop(pwd);
452 		return (error);
453 	}
454 	MPASS((ndp->ni_lcf & (NI_LCF_BENEATH_ABS | NI_LCF_LATCH)) !=
455 	    NI_LCF_BENEATH_ABS);
456 	if (((ndp->ni_lcf & NI_LCF_STRICTRELATIVE) != 0 &&
457 	    lookup_cap_dotdot != 0) ||
458 	    ((ndp->ni_lcf & NI_LCF_STRICTRELATIVE) == 0 &&
459 	    (cnp->cn_flags & BENEATH) != 0))
460 		ndp->ni_lcf |= NI_LCF_CAP_DOTDOT;
461 	SDT_PROBE4(vfs, namei, lookup, entry, *dpp, cnp->cn_pnbuf,
462 	    cnp->cn_flags, false);
463 	*pwdp = pwd;
464 	return (0);
465 }
466 
467 /*
468  * Convert a pathname into a pointer to a locked vnode.
469  *
470  * The FOLLOW flag is set when symbolic links are to be followed
471  * when they occur at the end of the name translation process.
472  * Symbolic links are always followed for all other pathname
473  * components other than the last.
474  *
475  * The segflg defines whether the name is to be copied from user
476  * space or kernel space.
477  *
478  * Overall outline of namei:
479  *
480  *	copy in name
481  *	get starting directory
482  *	while (!done && !error) {
483  *		call lookup to search path.
484  *		if symbolic link, massage name in buffer and continue
485  *	}
486  */
487 int
488 namei(struct nameidata *ndp)
489 {
490 	char *cp;		/* pointer into pathname argument */
491 	struct vnode *dp;	/* the directory we are searching */
492 	struct iovec aiov;		/* uio for reading symbolic links */
493 	struct componentname *cnp;
494 	struct thread *td;
495 	struct pwd *pwd;
496 	struct uio auio;
497 	int error, linklen;
498 	enum cache_fpl_status status;
499 
500 	cnp = &ndp->ni_cnd;
501 	td = cnp->cn_thread;
502 #ifdef INVARIANTS
503 	KASSERT((ndp->ni_debugflags & NAMEI_DBG_CALLED) == 0,
504 	    ("%s: repeated call to namei without NDREINIT", __func__));
505 	KASSERT(ndp->ni_debugflags == NAMEI_DBG_INITED,
506 	    ("%s: bad debugflags %d", __func__, ndp->ni_debugflags));
507 	ndp->ni_debugflags |= NAMEI_DBG_CALLED;
508 	if (ndp->ni_startdir != NULL)
509 		ndp->ni_debugflags |= NAMEI_DBG_HADSTARTDIR;
510 	/*
511 	 * For NDVALIDATE.
512 	 *
513 	 * While NDINIT may seem like a more natural place to do it, there are
514 	 * callers which directly modify flags past invoking init.
515 	 */
516 	cnp->cn_origflags = cnp->cn_flags;
517 #endif
518 	ndp->ni_cnd.cn_cred = ndp->ni_cnd.cn_thread->td_ucred;
519 	KASSERT(ndp->ni_resflags == 0, ("%s: garbage in ni_resflags: %x\n",
520 	    __func__, ndp->ni_resflags));
521 	KASSERT(cnp->cn_cred && td->td_proc, ("namei: bad cred/proc"));
522 	KASSERT((cnp->cn_flags & NAMEI_INTERNAL_FLAGS) == 0,
523 	    ("namei: unexpected flags: %" PRIx64 "\n",
524 	    cnp->cn_flags & NAMEI_INTERNAL_FLAGS));
525 	if (cnp->cn_flags & NOCACHE)
526 		KASSERT(cnp->cn_nameiop != LOOKUP,
527 		    ("%s: NOCACHE passed with LOOKUP", __func__));
528 	MPASS(ndp->ni_startdir == NULL || ndp->ni_startdir->v_type == VDIR ||
529 	    ndp->ni_startdir->v_type == VBAD);
530 
531 	ndp->ni_lcf = 0;
532 	ndp->ni_vp = NULL;
533 
534 	/*
535 	 * Get a buffer for the name to be translated, and copy the
536 	 * name into the buffer.
537 	 */
538 	cnp->cn_pnbuf = uma_zalloc(namei_zone, M_WAITOK);
539 	if (ndp->ni_segflg == UIO_SYSSPACE)
540 		error = copystr(ndp->ni_dirp, cnp->cn_pnbuf, MAXPATHLEN,
541 		    &ndp->ni_pathlen);
542 	else
543 		error = copyinstr(ndp->ni_dirp, cnp->cn_pnbuf, MAXPATHLEN,
544 		    &ndp->ni_pathlen);
545 
546 	if (__predict_false(error != 0)) {
547 		namei_cleanup_cnp(cnp);
548 		return (error);
549 	}
550 
551 	/*
552 	 * Don't allow empty pathnames.
553 	 */
554 	if (__predict_false(*cnp->cn_pnbuf == '\0')) {
555 		namei_cleanup_cnp(cnp);
556 		return (ENOENT);
557 	}
558 
559 #ifdef KTRACE
560 	if (KTRPOINT(td, KTR_NAMEI)) {
561 		KASSERT(cnp->cn_thread == curthread,
562 		    ("namei not using curthread"));
563 		ktrnamei(cnp->cn_pnbuf);
564 	}
565 #endif
566 
567 	cnp->cn_nameptr = cnp->cn_pnbuf;
568 
569 	/*
570 	 * First try looking up the target without locking any vnodes.
571 	 *
572 	 * We may need to start from scratch or pick up where it left off.
573 	 */
574 	error = cache_fplookup(ndp, &status, &pwd);
575 	switch (status) {
576 	case CACHE_FPL_STATUS_UNSET:
577 		__assert_unreachable();
578 		break;
579 	case CACHE_FPL_STATUS_HANDLED:
580 		if (error == 0)
581 			NDVALIDATE(ndp);
582 		return (error);
583 	case CACHE_FPL_STATUS_PARTIAL:
584 		TAILQ_INIT(&ndp->ni_cap_tracker);
585 		dp = ndp->ni_startdir;
586 		break;
587 	case CACHE_FPL_STATUS_ABORTED:
588 		TAILQ_INIT(&ndp->ni_cap_tracker);
589 		error = namei_setup(ndp, &dp, &pwd);
590 		if (error != 0) {
591 			namei_cleanup_cnp(cnp);
592 			return (error);
593 		}
594 		break;
595 	}
596 
597 	ndp->ni_loopcnt = 0;
598 
599 	/*
600 	 * Locked lookup.
601 	 */
602 	for (;;) {
603 		ndp->ni_startdir = dp;
604 		error = lookup(ndp);
605 		if (error != 0) {
606 			/*
607 			 * Override an error to not allow user to use
608 			 * BENEATH as an oracle.
609 			 */
610 			if ((ndp->ni_lcf & (NI_LCF_LATCH |
611 			    NI_LCF_BENEATH_LATCHED)) == NI_LCF_LATCH)
612 				error = ENOTCAPABLE;
613 			goto out;
614 		}
615 
616 		/*
617 		 * If not a symbolic link, we're done.
618 		 */
619 		if ((cnp->cn_flags & ISSYMLINK) == 0) {
620 			if ((cnp->cn_flags & (SAVENAME | SAVESTART)) == 0) {
621 				namei_cleanup_cnp(cnp);
622 			} else
623 				cnp->cn_flags |= HASBUF;
624 			if ((ndp->ni_lcf & (NI_LCF_LATCH |
625 			    NI_LCF_BENEATH_LATCHED)) == NI_LCF_LATCH) {
626 				NDFREE(ndp, 0);
627 				error = ENOTCAPABLE;
628 			}
629 			nameicap_cleanup(ndp, true);
630 			SDT_PROBE3(vfs, namei, lookup, return, error,
631 			    (error == 0 ? ndp->ni_vp : NULL), false);
632 			pwd_drop(pwd);
633 			if (error == 0)
634 				NDVALIDATE(ndp);
635 			return (error);
636 		}
637 		if (ndp->ni_loopcnt++ >= MAXSYMLINKS) {
638 			error = ELOOP;
639 			break;
640 		}
641 #ifdef MAC
642 		if ((cnp->cn_flags & NOMACCHECK) == 0) {
643 			error = mac_vnode_check_readlink(td->td_ucred,
644 			    ndp->ni_vp);
645 			if (error != 0)
646 				break;
647 		}
648 #endif
649 		if (ndp->ni_pathlen > 1)
650 			cp = uma_zalloc(namei_zone, M_WAITOK);
651 		else
652 			cp = cnp->cn_pnbuf;
653 		aiov.iov_base = cp;
654 		aiov.iov_len = MAXPATHLEN;
655 		auio.uio_iov = &aiov;
656 		auio.uio_iovcnt = 1;
657 		auio.uio_offset = 0;
658 		auio.uio_rw = UIO_READ;
659 		auio.uio_segflg = UIO_SYSSPACE;
660 		auio.uio_td = td;
661 		auio.uio_resid = MAXPATHLEN;
662 		error = VOP_READLINK(ndp->ni_vp, &auio, cnp->cn_cred);
663 		if (error != 0) {
664 			if (ndp->ni_pathlen > 1)
665 				uma_zfree(namei_zone, cp);
666 			break;
667 		}
668 		linklen = MAXPATHLEN - auio.uio_resid;
669 		if (linklen == 0) {
670 			if (ndp->ni_pathlen > 1)
671 				uma_zfree(namei_zone, cp);
672 			error = ENOENT;
673 			break;
674 		}
675 		if (linklen + ndp->ni_pathlen > MAXPATHLEN) {
676 			if (ndp->ni_pathlen > 1)
677 				uma_zfree(namei_zone, cp);
678 			error = ENAMETOOLONG;
679 			break;
680 		}
681 		if (ndp->ni_pathlen > 1) {
682 			bcopy(ndp->ni_next, cp + linklen, ndp->ni_pathlen);
683 			uma_zfree(namei_zone, cnp->cn_pnbuf);
684 			cnp->cn_pnbuf = cp;
685 		} else
686 			cnp->cn_pnbuf[linklen] = '\0';
687 		ndp->ni_pathlen += linklen;
688 		vput(ndp->ni_vp);
689 		dp = ndp->ni_dvp;
690 		/*
691 		 * Check if root directory should replace current directory.
692 		 */
693 		cnp->cn_nameptr = cnp->cn_pnbuf;
694 		if (*(cnp->cn_nameptr) == '/') {
695 			vrele(dp);
696 			error = namei_handle_root(ndp, &dp);
697 			if (error != 0)
698 				goto out;
699 		}
700 	}
701 	vput(ndp->ni_vp);
702 	ndp->ni_vp = NULL;
703 	vrele(ndp->ni_dvp);
704 out:
705 	MPASS(error != 0);
706 	namei_cleanup_cnp(cnp);
707 	nameicap_cleanup(ndp, true);
708 	SDT_PROBE3(vfs, namei, lookup, return, error, NULL, false);
709 	pwd_drop(pwd);
710 	return (error);
711 }
712 
713 static int
714 compute_cn_lkflags(struct mount *mp, int lkflags, int cnflags)
715 {
716 
717 	if (mp == NULL || ((lkflags & LK_SHARED) &&
718 	    (!(mp->mnt_kern_flag & MNTK_LOOKUP_SHARED) ||
719 	    ((cnflags & ISDOTDOT) &&
720 	    (mp->mnt_kern_flag & MNTK_LOOKUP_EXCL_DOTDOT))))) {
721 		lkflags &= ~LK_SHARED;
722 		lkflags |= LK_EXCLUSIVE;
723 	}
724 	lkflags |= LK_NODDLKTREAT;
725 	return (lkflags);
726 }
727 
728 static __inline int
729 needs_exclusive_leaf(struct mount *mp, int flags)
730 {
731 
732 	/*
733 	 * Intermediate nodes can use shared locks, we only need to
734 	 * force an exclusive lock for leaf nodes.
735 	 */
736 	if ((flags & (ISLASTCN | LOCKLEAF)) != (ISLASTCN | LOCKLEAF))
737 		return (0);
738 
739 	/* Always use exclusive locks if LOCKSHARED isn't set. */
740 	if (!(flags & LOCKSHARED))
741 		return (1);
742 
743 	/*
744 	 * For lookups during open(), if the mount point supports
745 	 * extended shared operations, then use a shared lock for the
746 	 * leaf node, otherwise use an exclusive lock.
747 	 */
748 	if ((flags & ISOPEN) != 0)
749 		return (!MNT_EXTENDED_SHARED(mp));
750 
751 	/*
752 	 * Lookup requests outside of open() that specify LOCKSHARED
753 	 * only need a shared lock on the leaf vnode.
754 	 */
755 	return (0);
756 }
757 
758 /*
759  * Search a pathname.
760  * This is a very central and rather complicated routine.
761  *
762  * The pathname is pointed to by ni_ptr and is of length ni_pathlen.
763  * The starting directory is taken from ni_startdir. The pathname is
764  * descended until done, or a symbolic link is encountered. The variable
765  * ni_more is clear if the path is completed; it is set to one if a
766  * symbolic link needing interpretation is encountered.
767  *
768  * The flag argument is LOOKUP, CREATE, RENAME, or DELETE depending on
769  * whether the name is to be looked up, created, renamed, or deleted.
770  * When CREATE, RENAME, or DELETE is specified, information usable in
771  * creating, renaming, or deleting a directory entry may be calculated.
772  * If flag has LOCKPARENT or'ed into it, the parent directory is returned
773  * locked. If flag has WANTPARENT or'ed into it, the parent directory is
774  * returned unlocked. Otherwise the parent directory is not returned. If
775  * the target of the pathname exists and LOCKLEAF is or'ed into the flag
776  * the target is returned locked, otherwise it is returned unlocked.
777  * When creating or renaming and LOCKPARENT is specified, the target may not
778  * be ".".  When deleting and LOCKPARENT is specified, the target may be ".".
779  *
780  * Overall outline of lookup:
781  *
782  * dirloop:
783  *	identify next component of name at ndp->ni_ptr
784  *	handle degenerate case where name is null string
785  *	if .. and crossing mount points and on mounted filesys, find parent
786  *	call VOP_LOOKUP routine for next component name
787  *	    directory vnode returned in ni_dvp, unlocked unless LOCKPARENT set
788  *	    component vnode returned in ni_vp (if it exists), locked.
789  *	if result vnode is mounted on and crossing mount points,
790  *	    find mounted on vnode
791  *	if more components of name, do next level at dirloop
792  *	return the answer in ni_vp, locked if LOCKLEAF set
793  *	    if LOCKPARENT set, return locked parent in ni_dvp
794  *	    if WANTPARENT set, return unlocked parent in ni_dvp
795  */
796 int
797 lookup(struct nameidata *ndp)
798 {
799 	char *cp;			/* pointer into pathname argument */
800 	char *prev_ni_next;		/* saved ndp->ni_next */
801 	struct vnode *dp = NULL;	/* the directory we are searching */
802 	struct vnode *tdp;		/* saved dp */
803 	struct mount *mp;		/* mount table entry */
804 	struct prison *pr;
805 	size_t prev_ni_pathlen;		/* saved ndp->ni_pathlen */
806 	int docache;			/* == 0 do not cache last component */
807 	int wantparent;			/* 1 => wantparent or lockparent flag */
808 	int rdonly;			/* lookup read-only flag bit */
809 	int error = 0;
810 	int dpunlocked = 0;		/* dp has already been unlocked */
811 	int relookup = 0;		/* do not consume the path component */
812 	struct componentname *cnp = &ndp->ni_cnd;
813 	int lkflags_save;
814 	int ni_dvp_unlocked;
815 
816 	/*
817 	 * Setup: break out flag bits into variables.
818 	 */
819 	ni_dvp_unlocked = 0;
820 	wantparent = cnp->cn_flags & (LOCKPARENT | WANTPARENT);
821 	KASSERT(cnp->cn_nameiop == LOOKUP || wantparent,
822 	    ("CREATE, DELETE, RENAME require LOCKPARENT or WANTPARENT."));
823 	/*
824 	 * When set to zero, docache causes the last component of the
825 	 * pathname to be deleted from the cache and the full lookup
826 	 * of the name to be done (via VOP_CACHEDLOOKUP()). Often
827 	 * filesystems need some pre-computed values that are made
828 	 * during the full lookup, for instance UFS sets dp->i_offset.
829 	 *
830 	 * The docache variable is set to zero when requested by the
831 	 * NOCACHE flag and for all modifying operations except CREATE.
832 	 */
833 	docache = (cnp->cn_flags & NOCACHE) ^ NOCACHE;
834 	if (cnp->cn_nameiop == DELETE ||
835 	    (wantparent && cnp->cn_nameiop != CREATE &&
836 	     cnp->cn_nameiop != LOOKUP))
837 		docache = 0;
838 	rdonly = cnp->cn_flags & RDONLY;
839 	cnp->cn_flags &= ~ISSYMLINK;
840 	ndp->ni_dvp = NULL;
841 	/*
842 	 * We use shared locks until we hit the parent of the last cn then
843 	 * we adjust based on the requesting flags.
844 	 */
845 	cnp->cn_lkflags = LK_SHARED;
846 	dp = ndp->ni_startdir;
847 	ndp->ni_startdir = NULLVP;
848 	vn_lock(dp,
849 	    compute_cn_lkflags(dp->v_mount, cnp->cn_lkflags | LK_RETRY,
850 	    cnp->cn_flags));
851 
852 dirloop:
853 	/*
854 	 * Search a new directory.
855 	 *
856 	 * The last component of the filename is left accessible via
857 	 * cnp->cn_nameptr for callers that need the name. Callers needing
858 	 * the name set the SAVENAME flag. When done, they assume
859 	 * responsibility for freeing the pathname buffer.
860 	 */
861 	for (cp = cnp->cn_nameptr; *cp != 0 && *cp != '/'; cp++)
862 		continue;
863 	cnp->cn_namelen = cp - cnp->cn_nameptr;
864 	if (cnp->cn_namelen > NAME_MAX) {
865 		error = ENAMETOOLONG;
866 		goto bad;
867 	}
868 #ifdef NAMEI_DIAGNOSTIC
869 	{ char c = *cp;
870 	*cp = '\0';
871 	printf("{%s}: ", cnp->cn_nameptr);
872 	*cp = c; }
873 #endif
874 	prev_ni_pathlen = ndp->ni_pathlen;
875 	ndp->ni_pathlen -= cnp->cn_namelen;
876 	KASSERT(ndp->ni_pathlen <= PATH_MAX,
877 	    ("%s: ni_pathlen underflow to %zd\n", __func__, ndp->ni_pathlen));
878 	prev_ni_next = ndp->ni_next;
879 	ndp->ni_next = cp;
880 
881 	/*
882 	 * Replace multiple slashes by a single slash and trailing slashes
883 	 * by a null.  This must be done before VOP_LOOKUP() because some
884 	 * fs's don't know about trailing slashes.  Remember if there were
885 	 * trailing slashes to handle symlinks, existing non-directories
886 	 * and non-existing files that won't be directories specially later.
887 	 */
888 	while (*cp == '/' && (cp[1] == '/' || cp[1] == '\0')) {
889 		cp++;
890 		ndp->ni_pathlen--;
891 		if (*cp == '\0') {
892 			*ndp->ni_next = '\0';
893 			cnp->cn_flags |= TRAILINGSLASH;
894 		}
895 	}
896 	ndp->ni_next = cp;
897 
898 	cnp->cn_flags |= MAKEENTRY;
899 	if (*cp == '\0' && docache == 0)
900 		cnp->cn_flags &= ~MAKEENTRY;
901 	if (cnp->cn_namelen == 2 &&
902 	    cnp->cn_nameptr[1] == '.' && cnp->cn_nameptr[0] == '.')
903 		cnp->cn_flags |= ISDOTDOT;
904 	else
905 		cnp->cn_flags &= ~ISDOTDOT;
906 	if (*ndp->ni_next == 0)
907 		cnp->cn_flags |= ISLASTCN;
908 	else
909 		cnp->cn_flags &= ~ISLASTCN;
910 
911 	if ((cnp->cn_flags & ISLASTCN) != 0 &&
912 	    cnp->cn_namelen == 1 && cnp->cn_nameptr[0] == '.' &&
913 	    (cnp->cn_nameiop == DELETE || cnp->cn_nameiop == RENAME)) {
914 		error = EINVAL;
915 		goto bad;
916 	}
917 
918 	nameicap_tracker_add(ndp, dp);
919 
920 	/*
921 	 * Check for degenerate name (e.g. / or "")
922 	 * which is a way of talking about a directory,
923 	 * e.g. like "/." or ".".
924 	 */
925 	if (cnp->cn_nameptr[0] == '\0') {
926 		if (dp->v_type != VDIR) {
927 			error = ENOTDIR;
928 			goto bad;
929 		}
930 		if (cnp->cn_nameiop != LOOKUP) {
931 			error = EISDIR;
932 			goto bad;
933 		}
934 		if (wantparent) {
935 			ndp->ni_dvp = dp;
936 			VREF(dp);
937 		}
938 		ndp->ni_vp = dp;
939 
940 		if (cnp->cn_flags & AUDITVNODE1)
941 			AUDIT_ARG_VNODE1(dp);
942 		else if (cnp->cn_flags & AUDITVNODE2)
943 			AUDIT_ARG_VNODE2(dp);
944 
945 		if (!(cnp->cn_flags & (LOCKPARENT | LOCKLEAF)))
946 			VOP_UNLOCK(dp);
947 		/* XXX This should probably move to the top of function. */
948 		if (cnp->cn_flags & SAVESTART)
949 			panic("lookup: SAVESTART");
950 		goto success;
951 	}
952 
953 	/*
954 	 * Handle "..": five special cases.
955 	 * 0. If doing a capability lookup and lookup_cap_dotdot is
956 	 *    disabled, return ENOTCAPABLE.
957 	 * 1. Return an error if this is the last component of
958 	 *    the name and the operation is DELETE or RENAME.
959 	 * 2. If at root directory (e.g. after chroot)
960 	 *    or at absolute root directory
961 	 *    then ignore it so can't get out.
962 	 * 3. If this vnode is the root of a mounted
963 	 *    filesystem, then replace it with the
964 	 *    vnode which was mounted on so we take the
965 	 *    .. in the other filesystem.
966 	 * 4. If the vnode is the top directory of
967 	 *    the jail or chroot, don't let them out.
968 	 * 5. If doing a capability lookup and lookup_cap_dotdot is
969 	 *    enabled, return ENOTCAPABLE if the lookup would escape
970 	 *    from the initial file descriptor directory.  Checks are
971 	 *    done by ensuring that namei() already traversed the
972 	 *    result of dotdot lookup.
973 	 */
974 	if (cnp->cn_flags & ISDOTDOT) {
975 		if ((ndp->ni_lcf & (NI_LCF_STRICTRELATIVE | NI_LCF_CAP_DOTDOT))
976 		    == NI_LCF_STRICTRELATIVE) {
977 #ifdef KTRACE
978 			if (KTRPOINT(curthread, KTR_CAPFAIL))
979 				ktrcapfail(CAPFAIL_LOOKUP, NULL, NULL);
980 #endif
981 			error = ENOTCAPABLE;
982 			goto bad;
983 		}
984 		if ((cnp->cn_flags & ISLASTCN) != 0 &&
985 		    (cnp->cn_nameiop == DELETE || cnp->cn_nameiop == RENAME)) {
986 			error = EINVAL;
987 			goto bad;
988 		}
989 		for (;;) {
990 			for (pr = cnp->cn_cred->cr_prison; pr != NULL;
991 			     pr = pr->pr_parent)
992 				if (dp == pr->pr_root)
993 					break;
994 			if (dp == ndp->ni_rootdir ||
995 			    dp == ndp->ni_topdir ||
996 			    dp == rootvnode ||
997 			    pr != NULL ||
998 			    ((dp->v_vflag & VV_ROOT) != 0 &&
999 			     (cnp->cn_flags & NOCROSSMOUNT) != 0)) {
1000 				ndp->ni_dvp = dp;
1001 				ndp->ni_vp = dp;
1002 				VREF(dp);
1003 				goto nextname;
1004 			}
1005 			if ((dp->v_vflag & VV_ROOT) == 0)
1006 				break;
1007 			if (VN_IS_DOOMED(dp)) {	/* forced unmount */
1008 				error = ENOENT;
1009 				goto bad;
1010 			}
1011 			tdp = dp;
1012 			dp = dp->v_mount->mnt_vnodecovered;
1013 			VREF(dp);
1014 			vput(tdp);
1015 			vn_lock(dp,
1016 			    compute_cn_lkflags(dp->v_mount, cnp->cn_lkflags |
1017 			    LK_RETRY, ISDOTDOT));
1018 			error = nameicap_check_dotdot(ndp, dp);
1019 			if (error != 0) {
1020 #ifdef KTRACE
1021 				if (KTRPOINT(curthread, KTR_CAPFAIL))
1022 					ktrcapfail(CAPFAIL_LOOKUP, NULL, NULL);
1023 #endif
1024 				goto bad;
1025 			}
1026 		}
1027 	}
1028 
1029 	/*
1030 	 * We now have a segment name to search for, and a directory to search.
1031 	 */
1032 unionlookup:
1033 #ifdef MAC
1034 	error = mac_vnode_check_lookup(cnp->cn_thread->td_ucred, dp, cnp);
1035 	if (error)
1036 		goto bad;
1037 #endif
1038 	ndp->ni_dvp = dp;
1039 	ndp->ni_vp = NULL;
1040 	ASSERT_VOP_LOCKED(dp, "lookup");
1041 	/*
1042 	 * If we have a shared lock we may need to upgrade the lock for the
1043 	 * last operation.
1044 	 */
1045 	if ((cnp->cn_flags & LOCKPARENT) && (cnp->cn_flags & ISLASTCN) &&
1046 	    dp != vp_crossmp && VOP_ISLOCKED(dp) == LK_SHARED)
1047 		vn_lock(dp, LK_UPGRADE|LK_RETRY);
1048 	if (VN_IS_DOOMED(dp)) {
1049 		error = ENOENT;
1050 		goto bad;
1051 	}
1052 	/*
1053 	 * If we're looking up the last component and we need an exclusive
1054 	 * lock, adjust our lkflags.
1055 	 */
1056 	if (needs_exclusive_leaf(dp->v_mount, cnp->cn_flags))
1057 		cnp->cn_lkflags = LK_EXCLUSIVE;
1058 #ifdef NAMEI_DIAGNOSTIC
1059 	vn_printf(dp, "lookup in ");
1060 #endif
1061 	lkflags_save = cnp->cn_lkflags;
1062 	cnp->cn_lkflags = compute_cn_lkflags(dp->v_mount, cnp->cn_lkflags,
1063 	    cnp->cn_flags);
1064 	error = VOP_LOOKUP(dp, &ndp->ni_vp, cnp);
1065 	cnp->cn_lkflags = lkflags_save;
1066 	if (error != 0) {
1067 		KASSERT(ndp->ni_vp == NULL, ("leaf should be empty"));
1068 #ifdef NAMEI_DIAGNOSTIC
1069 		printf("not found\n");
1070 #endif
1071 		if ((error == ENOENT) &&
1072 		    (dp->v_vflag & VV_ROOT) && (dp->v_mount != NULL) &&
1073 		    (dp->v_mount->mnt_flag & MNT_UNION)) {
1074 			tdp = dp;
1075 			dp = dp->v_mount->mnt_vnodecovered;
1076 			VREF(dp);
1077 			vput(tdp);
1078 			vn_lock(dp,
1079 			    compute_cn_lkflags(dp->v_mount, cnp->cn_lkflags |
1080 			    LK_RETRY, cnp->cn_flags));
1081 			nameicap_tracker_add(ndp, dp);
1082 			goto unionlookup;
1083 		}
1084 
1085 		if (error == ERELOOKUP) {
1086 			vref(dp);
1087 			ndp->ni_vp = dp;
1088 			error = 0;
1089 			relookup = 1;
1090 			goto good;
1091 		}
1092 
1093 		if (error != EJUSTRETURN)
1094 			goto bad;
1095 		/*
1096 		 * At this point, we know we're at the end of the
1097 		 * pathname.  If creating / renaming, we can consider
1098 		 * allowing the file or directory to be created / renamed,
1099 		 * provided we're not on a read-only filesystem.
1100 		 */
1101 		if (rdonly) {
1102 			error = EROFS;
1103 			goto bad;
1104 		}
1105 		/* trailing slash only allowed for directories */
1106 		if ((cnp->cn_flags & TRAILINGSLASH) &&
1107 		    !(cnp->cn_flags & WILLBEDIR)) {
1108 			error = ENOENT;
1109 			goto bad;
1110 		}
1111 		if ((cnp->cn_flags & LOCKPARENT) == 0)
1112 			VOP_UNLOCK(dp);
1113 		/*
1114 		 * We return with ni_vp NULL to indicate that the entry
1115 		 * doesn't currently exist, leaving a pointer to the
1116 		 * (possibly locked) directory vnode in ndp->ni_dvp.
1117 		 */
1118 		if (cnp->cn_flags & SAVESTART) {
1119 			ndp->ni_startdir = ndp->ni_dvp;
1120 			VREF(ndp->ni_startdir);
1121 		}
1122 		goto success;
1123 	}
1124 
1125 good:
1126 #ifdef NAMEI_DIAGNOSTIC
1127 	printf("found\n");
1128 #endif
1129 	dp = ndp->ni_vp;
1130 
1131 	/*
1132 	 * Check to see if the vnode has been mounted on;
1133 	 * if so find the root of the mounted filesystem.
1134 	 */
1135 	while (dp->v_type == VDIR && (mp = dp->v_mountedhere) &&
1136 	       (cnp->cn_flags & NOCROSSMOUNT) == 0) {
1137 		if (vfs_busy(mp, 0))
1138 			continue;
1139 		vput(dp);
1140 		if (dp != ndp->ni_dvp)
1141 			vput(ndp->ni_dvp);
1142 		else
1143 			vrele(ndp->ni_dvp);
1144 		vrefact(vp_crossmp);
1145 		ndp->ni_dvp = vp_crossmp;
1146 		error = VFS_ROOT(mp, compute_cn_lkflags(mp, cnp->cn_lkflags,
1147 		    cnp->cn_flags), &tdp);
1148 		vfs_unbusy(mp);
1149 		if (vn_lock(vp_crossmp, LK_SHARED | LK_NOWAIT))
1150 			panic("vp_crossmp exclusively locked or reclaimed");
1151 		if (error) {
1152 			dpunlocked = 1;
1153 			goto bad2;
1154 		}
1155 		ndp->ni_vp = dp = tdp;
1156 	}
1157 
1158 	/*
1159 	 * Check for symbolic link
1160 	 */
1161 	if ((dp->v_type == VLNK) &&
1162 	    ((cnp->cn_flags & FOLLOW) || (cnp->cn_flags & TRAILINGSLASH) ||
1163 	     *ndp->ni_next == '/')) {
1164 		cnp->cn_flags |= ISSYMLINK;
1165 		if (VN_IS_DOOMED(dp)) {
1166 			/*
1167 			 * We can't know whether the directory was mounted with
1168 			 * NOSYMFOLLOW, so we can't follow safely.
1169 			 */
1170 			error = ENOENT;
1171 			goto bad2;
1172 		}
1173 		if (dp->v_mount->mnt_flag & MNT_NOSYMFOLLOW) {
1174 			error = EACCES;
1175 			goto bad2;
1176 		}
1177 		/*
1178 		 * Symlink code always expects an unlocked dvp.
1179 		 */
1180 		if (ndp->ni_dvp != ndp->ni_vp) {
1181 			VOP_UNLOCK(ndp->ni_dvp);
1182 			ni_dvp_unlocked = 1;
1183 		}
1184 		goto success;
1185 	}
1186 
1187 nextname:
1188 	/*
1189 	 * Not a symbolic link that we will follow.  Continue with the
1190 	 * next component if there is any; otherwise, we're done.
1191 	 */
1192 	KASSERT((cnp->cn_flags & ISLASTCN) || *ndp->ni_next == '/',
1193 	    ("lookup: invalid path state."));
1194 	if (relookup) {
1195 		relookup = 0;
1196 		ndp->ni_pathlen = prev_ni_pathlen;
1197 		ndp->ni_next = prev_ni_next;
1198 		if (ndp->ni_dvp != dp)
1199 			vput(ndp->ni_dvp);
1200 		else
1201 			vrele(ndp->ni_dvp);
1202 		goto dirloop;
1203 	}
1204 	if (cnp->cn_flags & ISDOTDOT) {
1205 		error = nameicap_check_dotdot(ndp, ndp->ni_vp);
1206 		if (error != 0) {
1207 #ifdef KTRACE
1208 			if (KTRPOINT(curthread, KTR_CAPFAIL))
1209 				ktrcapfail(CAPFAIL_LOOKUP, NULL, NULL);
1210 #endif
1211 			goto bad2;
1212 		}
1213 	}
1214 	if (*ndp->ni_next == '/') {
1215 		cnp->cn_nameptr = ndp->ni_next;
1216 		while (*cnp->cn_nameptr == '/') {
1217 			cnp->cn_nameptr++;
1218 			ndp->ni_pathlen--;
1219 		}
1220 		if (ndp->ni_dvp != dp)
1221 			vput(ndp->ni_dvp);
1222 		else
1223 			vrele(ndp->ni_dvp);
1224 		goto dirloop;
1225 	}
1226 	/*
1227 	 * If we're processing a path with a trailing slash,
1228 	 * check that the end result is a directory.
1229 	 */
1230 	if ((cnp->cn_flags & TRAILINGSLASH) && dp->v_type != VDIR) {
1231 		error = ENOTDIR;
1232 		goto bad2;
1233 	}
1234 	/*
1235 	 * Disallow directory write attempts on read-only filesystems.
1236 	 */
1237 	if (rdonly &&
1238 	    (cnp->cn_nameiop == DELETE || cnp->cn_nameiop == RENAME)) {
1239 		error = EROFS;
1240 		goto bad2;
1241 	}
1242 	if (cnp->cn_flags & SAVESTART) {
1243 		ndp->ni_startdir = ndp->ni_dvp;
1244 		VREF(ndp->ni_startdir);
1245 	}
1246 	if (!wantparent) {
1247 		ni_dvp_unlocked = 2;
1248 		if (ndp->ni_dvp != dp)
1249 			vput(ndp->ni_dvp);
1250 		else
1251 			vrele(ndp->ni_dvp);
1252 	} else if ((cnp->cn_flags & LOCKPARENT) == 0 && ndp->ni_dvp != dp) {
1253 		VOP_UNLOCK(ndp->ni_dvp);
1254 		ni_dvp_unlocked = 1;
1255 	}
1256 
1257 	if (cnp->cn_flags & AUDITVNODE1)
1258 		AUDIT_ARG_VNODE1(dp);
1259 	else if (cnp->cn_flags & AUDITVNODE2)
1260 		AUDIT_ARG_VNODE2(dp);
1261 
1262 	if ((cnp->cn_flags & LOCKLEAF) == 0)
1263 		VOP_UNLOCK(dp);
1264 success:
1265 	/*
1266 	 * FIXME: for lookups which only cross a mount point to fetch the
1267 	 * root vnode, ni_dvp will be set to vp_crossmp. This can be a problem
1268 	 * if either WANTPARENT or LOCKPARENT is set.
1269 	 */
1270 	/*
1271 	 * Because of shared lookup we may have the vnode shared locked, but
1272 	 * the caller may want it to be exclusively locked.
1273 	 */
1274 	if (needs_exclusive_leaf(dp->v_mount, cnp->cn_flags) &&
1275 	    VOP_ISLOCKED(dp) != LK_EXCLUSIVE) {
1276 		vn_lock(dp, LK_UPGRADE | LK_RETRY);
1277 		if (VN_IS_DOOMED(dp)) {
1278 			error = ENOENT;
1279 			goto bad2;
1280 		}
1281 	}
1282 	if (ndp->ni_vp != NULL && ndp->ni_vp->v_type == VDIR)
1283 		nameicap_tracker_add(ndp, ndp->ni_vp);
1284 	return (0);
1285 
1286 bad2:
1287 	if (ni_dvp_unlocked != 2) {
1288 		if (dp != ndp->ni_dvp && !ni_dvp_unlocked)
1289 			vput(ndp->ni_dvp);
1290 		else
1291 			vrele(ndp->ni_dvp);
1292 	}
1293 bad:
1294 	if (!dpunlocked)
1295 		vput(dp);
1296 	ndp->ni_vp = NULL;
1297 	return (error);
1298 }
1299 
1300 /*
1301  * relookup - lookup a path name component
1302  *    Used by lookup to re-acquire things.
1303  */
1304 int
1305 relookup(struct vnode *dvp, struct vnode **vpp, struct componentname *cnp)
1306 {
1307 	struct vnode *dp = NULL;		/* the directory we are searching */
1308 	int wantparent;			/* 1 => wantparent or lockparent flag */
1309 	int rdonly;			/* lookup read-only flag bit */
1310 	int error = 0;
1311 
1312 	KASSERT(cnp->cn_flags & ISLASTCN,
1313 	    ("relookup: Not given last component."));
1314 	/*
1315 	 * Setup: break out flag bits into variables.
1316 	 */
1317 	wantparent = cnp->cn_flags & (LOCKPARENT|WANTPARENT);
1318 	KASSERT(wantparent, ("relookup: parent not wanted."));
1319 	rdonly = cnp->cn_flags & RDONLY;
1320 	cnp->cn_flags &= ~ISSYMLINK;
1321 	dp = dvp;
1322 	cnp->cn_lkflags = LK_EXCLUSIVE;
1323 	vn_lock(dp, LK_EXCLUSIVE | LK_RETRY);
1324 
1325 	/*
1326 	 * Search a new directory.
1327 	 *
1328 	 * The last component of the filename is left accessible via
1329 	 * cnp->cn_nameptr for callers that need the name. Callers needing
1330 	 * the name set the SAVENAME flag. When done, they assume
1331 	 * responsibility for freeing the pathname buffer.
1332 	 */
1333 #ifdef NAMEI_DIAGNOSTIC
1334 	printf("{%s}: ", cnp->cn_nameptr);
1335 #endif
1336 
1337 	/*
1338 	 * Check for "" which represents the root directory after slash
1339 	 * removal.
1340 	 */
1341 	if (cnp->cn_nameptr[0] == '\0') {
1342 		/*
1343 		 * Support only LOOKUP for "/" because lookup()
1344 		 * can't succeed for CREATE, DELETE and RENAME.
1345 		 */
1346 		KASSERT(cnp->cn_nameiop == LOOKUP, ("nameiop must be LOOKUP"));
1347 		KASSERT(dp->v_type == VDIR, ("dp is not a directory"));
1348 
1349 		if (!(cnp->cn_flags & LOCKLEAF))
1350 			VOP_UNLOCK(dp);
1351 		*vpp = dp;
1352 		/* XXX This should probably move to the top of function. */
1353 		if (cnp->cn_flags & SAVESTART)
1354 			panic("lookup: SAVESTART");
1355 		return (0);
1356 	}
1357 
1358 	if (cnp->cn_flags & ISDOTDOT)
1359 		panic ("relookup: lookup on dot-dot");
1360 
1361 	/*
1362 	 * We now have a segment name to search for, and a directory to search.
1363 	 */
1364 #ifdef NAMEI_DIAGNOSTIC
1365 	vn_printf(dp, "search in ");
1366 #endif
1367 	if ((error = VOP_LOOKUP(dp, vpp, cnp)) != 0) {
1368 		KASSERT(*vpp == NULL, ("leaf should be empty"));
1369 		if (error != EJUSTRETURN)
1370 			goto bad;
1371 		/*
1372 		 * If creating and at end of pathname, then can consider
1373 		 * allowing file to be created.
1374 		 */
1375 		if (rdonly) {
1376 			error = EROFS;
1377 			goto bad;
1378 		}
1379 		/* ASSERT(dvp == ndp->ni_startdir) */
1380 		if (cnp->cn_flags & SAVESTART)
1381 			VREF(dvp);
1382 		if ((cnp->cn_flags & LOCKPARENT) == 0)
1383 			VOP_UNLOCK(dp);
1384 		/*
1385 		 * We return with ni_vp NULL to indicate that the entry
1386 		 * doesn't currently exist, leaving a pointer to the
1387 		 * (possibly locked) directory vnode in ndp->ni_dvp.
1388 		 */
1389 		return (0);
1390 	}
1391 
1392 	dp = *vpp;
1393 
1394 	/*
1395 	 * Disallow directory write attempts on read-only filesystems.
1396 	 */
1397 	if (rdonly &&
1398 	    (cnp->cn_nameiop == DELETE || cnp->cn_nameiop == RENAME)) {
1399 		if (dvp == dp)
1400 			vrele(dvp);
1401 		else
1402 			vput(dvp);
1403 		error = EROFS;
1404 		goto bad;
1405 	}
1406 	/*
1407 	 * Set the parent lock/ref state to the requested state.
1408 	 */
1409 	if ((cnp->cn_flags & LOCKPARENT) == 0 && dvp != dp) {
1410 		if (wantparent)
1411 			VOP_UNLOCK(dvp);
1412 		else
1413 			vput(dvp);
1414 	} else if (!wantparent)
1415 		vrele(dvp);
1416 	/*
1417 	 * Check for symbolic link
1418 	 */
1419 	KASSERT(dp->v_type != VLNK || !(cnp->cn_flags & FOLLOW),
1420 	    ("relookup: symlink found.\n"));
1421 
1422 	/* ASSERT(dvp == ndp->ni_startdir) */
1423 	if (cnp->cn_flags & SAVESTART)
1424 		VREF(dvp);
1425 
1426 	if ((cnp->cn_flags & LOCKLEAF) == 0)
1427 		VOP_UNLOCK(dp);
1428 	return (0);
1429 bad:
1430 	vput(dp);
1431 	*vpp = NULL;
1432 	return (error);
1433 }
1434 
1435 /*
1436  * Free data allocated by namei(); see namei(9) for details.
1437  */
1438 void
1439 NDFREE_PNBUF(struct nameidata *ndp)
1440 {
1441 
1442 	if ((ndp->ni_cnd.cn_flags & HASBUF) != 0) {
1443 		MPASS((ndp->ni_cnd.cn_flags & (SAVENAME | SAVESTART)) != 0);
1444 		uma_zfree(namei_zone, ndp->ni_cnd.cn_pnbuf);
1445 		ndp->ni_cnd.cn_flags &= ~HASBUF;
1446 	}
1447 }
1448 
1449 void
1450 (NDFREE)(struct nameidata *ndp, const u_int flags)
1451 {
1452 	int unlock_dvp;
1453 	int unlock_vp;
1454 
1455 	unlock_dvp = 0;
1456 	unlock_vp = 0;
1457 
1458 	if (!(flags & NDF_NO_FREE_PNBUF)) {
1459 		NDFREE_PNBUF(ndp);
1460 	}
1461 	if (!(flags & NDF_NO_VP_UNLOCK) &&
1462 	    (ndp->ni_cnd.cn_flags & LOCKLEAF) && ndp->ni_vp)
1463 		unlock_vp = 1;
1464 	if (!(flags & NDF_NO_DVP_UNLOCK) &&
1465 	    (ndp->ni_cnd.cn_flags & LOCKPARENT) &&
1466 	    ndp->ni_dvp != ndp->ni_vp)
1467 		unlock_dvp = 1;
1468 	if (!(flags & NDF_NO_VP_RELE) && ndp->ni_vp) {
1469 		if (unlock_vp) {
1470 			vput(ndp->ni_vp);
1471 			unlock_vp = 0;
1472 		} else
1473 			vrele(ndp->ni_vp);
1474 		ndp->ni_vp = NULL;
1475 	}
1476 	if (unlock_vp)
1477 		VOP_UNLOCK(ndp->ni_vp);
1478 	if (!(flags & NDF_NO_DVP_RELE) &&
1479 	    (ndp->ni_cnd.cn_flags & (LOCKPARENT|WANTPARENT))) {
1480 		if (unlock_dvp) {
1481 			vput(ndp->ni_dvp);
1482 			unlock_dvp = 0;
1483 		} else
1484 			vrele(ndp->ni_dvp);
1485 		ndp->ni_dvp = NULL;
1486 	}
1487 	if (unlock_dvp)
1488 		VOP_UNLOCK(ndp->ni_dvp);
1489 	if (!(flags & NDF_NO_STARTDIR_RELE) &&
1490 	    (ndp->ni_cnd.cn_flags & SAVESTART)) {
1491 		vrele(ndp->ni_startdir);
1492 		ndp->ni_startdir = NULL;
1493 	}
1494 }
1495 
1496 #ifdef INVARIANTS
1497 /*
1498  * Validate the final state of ndp after the lookup.
1499  *
1500  * Historically filesystems were allowed to modify cn_flags. Most notably they
1501  * can add SAVENAME to the request, resulting in HASBUF and pushing subsequent
1502  * clean up to the consumer. In practice this seems to only concern != LOOKUP
1503  * operations.
1504  *
1505  * As a step towards stricter API contract this routine validates the state to
1506  * clean up. Note validation is a work in progress with the intent of becoming
1507  * stricter over time.
1508  */
1509 #define NDMODIFYINGFLAGS (LOCKLEAF | LOCKPARENT | WANTPARENT | SAVENAME | SAVESTART | HASBUF)
1510 void
1511 NDVALIDATE(struct nameidata *ndp)
1512 {
1513 	struct componentname *cnp;
1514 	u_int64_t used, orig;
1515 
1516 	cnp = &ndp->ni_cnd;
1517 	orig = cnp->cn_origflags;
1518 	used = cnp->cn_flags;
1519 	switch (cnp->cn_nameiop) {
1520 	case LOOKUP:
1521 		/*
1522 		 * For plain lookup we require strict conformance -- nothing
1523 		 * to clean up if it was not requested by the caller.
1524 		 */
1525 		orig &= NDMODIFYINGFLAGS;
1526 		used &= NDMODIFYINGFLAGS;
1527 		if ((orig & (SAVENAME | SAVESTART)) != 0)
1528 			orig |= HASBUF;
1529 		if (orig != used) {
1530 			goto out_mismatch;
1531 		}
1532 		break;
1533 	case CREATE:
1534 	case DELETE:
1535 	case RENAME:
1536 		/*
1537 		 * Some filesystems set SAVENAME to provoke HASBUF, accomodate
1538 		 * for it until it gets fixed.
1539 		 */
1540 		orig &= NDMODIFYINGFLAGS;
1541 		orig |= (SAVENAME | HASBUF);
1542 		used &= NDMODIFYINGFLAGS;
1543 		used |= (SAVENAME | HASBUF);
1544 		if (orig != used) {
1545 			goto out_mismatch;
1546 		}
1547 		break;
1548 	}
1549 	return;
1550 out_mismatch:
1551 	panic("%s: mismatched flags for op %d: added %" PRIx64 ", "
1552 	    "removed %" PRIx64" (%" PRIx64" != %" PRIx64"; stored %" PRIx64" != %" PRIx64")",
1553 	    __func__, cnp->cn_nameiop, used & ~orig, orig &~ used,
1554 	    orig, used, cnp->cn_origflags, cnp->cn_flags);
1555 }
1556 #endif
1557 
1558 /*
1559  * Determine if there is a suitable alternate filename under the specified
1560  * prefix for the specified path.  If the create flag is set, then the
1561  * alternate prefix will be used so long as the parent directory exists.
1562  * This is used by the various compatibility ABIs so that Linux binaries prefer
1563  * files under /compat/linux for example.  The chosen path (whether under
1564  * the prefix or under /) is returned in a kernel malloc'd buffer pointed
1565  * to by pathbuf.  The caller is responsible for free'ing the buffer from
1566  * the M_TEMP bucket if one is returned.
1567  */
1568 int
1569 kern_alternate_path(struct thread *td, const char *prefix, const char *path,
1570     enum uio_seg pathseg, char **pathbuf, int create, int dirfd)
1571 {
1572 	struct nameidata nd, ndroot;
1573 	char *ptr, *buf, *cp;
1574 	size_t len, sz;
1575 	int error;
1576 
1577 	buf = (char *) malloc(MAXPATHLEN, M_TEMP, M_WAITOK);
1578 	*pathbuf = buf;
1579 
1580 	/* Copy the prefix into the new pathname as a starting point. */
1581 	len = strlcpy(buf, prefix, MAXPATHLEN);
1582 	if (len >= MAXPATHLEN) {
1583 		*pathbuf = NULL;
1584 		free(buf, M_TEMP);
1585 		return (EINVAL);
1586 	}
1587 	sz = MAXPATHLEN - len;
1588 	ptr = buf + len;
1589 
1590 	/* Append the filename to the prefix. */
1591 	if (pathseg == UIO_SYSSPACE)
1592 		error = copystr(path, ptr, sz, &len);
1593 	else
1594 		error = copyinstr(path, ptr, sz, &len);
1595 
1596 	if (error) {
1597 		*pathbuf = NULL;
1598 		free(buf, M_TEMP);
1599 		return (error);
1600 	}
1601 
1602 	/* Only use a prefix with absolute pathnames. */
1603 	if (*ptr != '/') {
1604 		error = EINVAL;
1605 		goto keeporig;
1606 	}
1607 
1608 	if (dirfd != AT_FDCWD) {
1609 		/*
1610 		 * We want the original because the "prefix" is
1611 		 * included in the already opened dirfd.
1612 		 */
1613 		bcopy(ptr, buf, len);
1614 		return (0);
1615 	}
1616 
1617 	/*
1618 	 * We know that there is a / somewhere in this pathname.
1619 	 * Search backwards for it, to find the file's parent dir
1620 	 * to see if it exists in the alternate tree. If it does,
1621 	 * and we want to create a file (cflag is set). We don't
1622 	 * need to worry about the root comparison in this case.
1623 	 */
1624 
1625 	if (create) {
1626 		for (cp = &ptr[len] - 1; *cp != '/'; cp--);
1627 		*cp = '\0';
1628 
1629 		NDINIT(&nd, LOOKUP, NOFOLLOW, UIO_SYSSPACE, buf, td);
1630 		error = namei(&nd);
1631 		*cp = '/';
1632 		if (error != 0)
1633 			goto keeporig;
1634 	} else {
1635 		NDINIT(&nd, LOOKUP, NOFOLLOW, UIO_SYSSPACE, buf, td);
1636 
1637 		error = namei(&nd);
1638 		if (error != 0)
1639 			goto keeporig;
1640 
1641 		/*
1642 		 * We now compare the vnode of the prefix to the one
1643 		 * vnode asked. If they resolve to be the same, then we
1644 		 * ignore the match so that the real root gets used.
1645 		 * This avoids the problem of traversing "../.." to find the
1646 		 * root directory and never finding it, because "/" resolves
1647 		 * to the emulation root directory. This is expensive :-(
1648 		 */
1649 		NDINIT(&ndroot, LOOKUP, FOLLOW, UIO_SYSSPACE, prefix,
1650 		    td);
1651 
1652 		/* We shouldn't ever get an error from this namei(). */
1653 		error = namei(&ndroot);
1654 		if (error == 0) {
1655 			if (nd.ni_vp == ndroot.ni_vp)
1656 				error = ENOENT;
1657 
1658 			NDFREE(&ndroot, NDF_ONLY_PNBUF);
1659 			vrele(ndroot.ni_vp);
1660 		}
1661 	}
1662 
1663 	NDFREE(&nd, NDF_ONLY_PNBUF);
1664 	vrele(nd.ni_vp);
1665 
1666 keeporig:
1667 	/* If there was an error, use the original path name. */
1668 	if (error)
1669 		bcopy(ptr, buf, len);
1670 	return (error);
1671 }
1672