xref: /freebsd/sys/kern/vfs_lookup.c (revision eb24e1491f9900e922c78e53af588f22a3e9535f)
1 /*-
2  * SPDX-License-Identifier: BSD-3-Clause
3  *
4  * Copyright (c) 1982, 1986, 1989, 1993
5  *	The Regents of the University of California.  All rights reserved.
6  * (c) UNIX System Laboratories, Inc.
7  * All or some portions of this file are derived from material licensed
8  * to the University of California by American Telephone and Telegraph
9  * Co. or Unix System Laboratories, Inc. and are reproduced herein with
10  * the permission of UNIX System Laboratories, Inc.
11  *
12  * Redistribution and use in source and binary forms, with or without
13  * modification, are permitted provided that the following conditions
14  * are met:
15  * 1. Redistributions of source code must retain the above copyright
16  *    notice, this list of conditions and the following disclaimer.
17  * 2. Redistributions in binary form must reproduce the above copyright
18  *    notice, this list of conditions and the following disclaimer in the
19  *    documentation and/or other materials provided with the distribution.
20  * 3. Neither the name of the University nor the names of its contributors
21  *    may be used to endorse or promote products derived from this software
22  *    without specific prior written permission.
23  *
24  * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
25  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
26  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
27  * ARE DISCLAIMED.  IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
28  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
29  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
30  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
31  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
32  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
33  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
34  * SUCH DAMAGE.
35  *
36  *	@(#)vfs_lookup.c	8.4 (Berkeley) 2/16/94
37  */
38 
39 #include <sys/cdefs.h>
40 __FBSDID("$FreeBSD$");
41 
42 #include "opt_capsicum.h"
43 #include "opt_ktrace.h"
44 
45 #include <sys/param.h>
46 #include <sys/systm.h>
47 #include <sys/kernel.h>
48 #include <sys/capsicum.h>
49 #include <sys/fcntl.h>
50 #include <sys/jail.h>
51 #include <sys/lock.h>
52 #include <sys/mutex.h>
53 #include <sys/namei.h>
54 #include <sys/vnode.h>
55 #include <sys/mount.h>
56 #include <sys/filedesc.h>
57 #include <sys/proc.h>
58 #include <sys/sdt.h>
59 #include <sys/syscallsubr.h>
60 #include <sys/sysctl.h>
61 #ifdef KTRACE
62 #include <sys/ktrace.h>
63 #endif
64 
65 #include <security/audit/audit.h>
66 #include <security/mac/mac_framework.h>
67 
68 #include <vm/uma.h>
69 
70 #define	NAMEI_DIAGNOSTIC 1
71 #undef NAMEI_DIAGNOSTIC
72 
73 SDT_PROVIDER_DECLARE(vfs);
74 SDT_PROBE_DEFINE3(vfs, namei, lookup, entry, "struct vnode *", "char *",
75     "unsigned long");
76 SDT_PROBE_DEFINE2(vfs, namei, lookup, return, "int", "struct vnode *");
77 
78 /* Allocation zone for namei. */
79 uma_zone_t namei_zone;
80 
81 /* Placeholder vnode for mp traversal. */
82 static struct vnode *vp_crossmp;
83 
84 static int
85 crossmp_vop_islocked(struct vop_islocked_args *ap)
86 {
87 
88 	return (LK_SHARED);
89 }
90 
91 static int
92 crossmp_vop_lock1(struct vop_lock1_args *ap)
93 {
94 	struct vnode *vp;
95 	struct lock *lk __unused;
96 	const char *file __unused;
97 	int flags, line __unused;
98 
99 	vp = ap->a_vp;
100 	lk = vp->v_vnlock;
101 	flags = ap->a_flags;
102 	file = ap->a_file;
103 	line = ap->a_line;
104 
105 	if ((flags & LK_SHARED) == 0)
106 		panic("invalid lock request for crossmp");
107 
108 	WITNESS_CHECKORDER(&lk->lock_object, LOP_NEWORDER, file, line,
109 	    flags & LK_INTERLOCK ? &VI_MTX(vp)->lock_object : NULL);
110 	WITNESS_LOCK(&lk->lock_object, 0, file, line);
111 	if ((flags & LK_INTERLOCK) != 0)
112 		VI_UNLOCK(vp);
113 	LOCK_LOG_LOCK("SLOCK", &lk->lock_object, 0, 0, ap->a_file, line);
114 	return (0);
115 }
116 
117 static int
118 crossmp_vop_unlock(struct vop_unlock_args *ap)
119 {
120 	struct vnode *vp;
121 	struct lock *lk __unused;
122 
123 	vp = ap->a_vp;
124 	lk = vp->v_vnlock;
125 
126 	WITNESS_UNLOCK(&lk->lock_object, 0, LOCK_FILE, LOCK_LINE);
127 	LOCK_LOG_LOCK("SUNLOCK", &lk->lock_object, 0, 0, LOCK_FILE,
128 	    LOCK_LINE);
129 	return (0);
130 }
131 
132 static struct vop_vector crossmp_vnodeops = {
133 	.vop_default =		&default_vnodeops,
134 	.vop_islocked =		crossmp_vop_islocked,
135 	.vop_lock1 =		crossmp_vop_lock1,
136 	.vop_unlock =		crossmp_vop_unlock,
137 };
138 /*
139  * VFS_VOP_VECTOR_REGISTER(crossmp_vnodeops) is not used here since the vnode
140  * gets allocated early. See nameiinit for the direct call below.
141  */
142 
143 struct nameicap_tracker {
144 	struct vnode *dp;
145 	TAILQ_ENTRY(nameicap_tracker) nm_link;
146 };
147 
148 /* Zone for cap mode tracker elements used for dotdot capability checks. */
149 static uma_zone_t nt_zone;
150 
151 static void
152 nameiinit(void *dummy __unused)
153 {
154 
155 	namei_zone = uma_zcreate("NAMEI", MAXPATHLEN, NULL, NULL, NULL, NULL,
156 	    UMA_ALIGN_PTR, 0);
157 	nt_zone = uma_zcreate("rentr", sizeof(struct nameicap_tracker),
158 	    NULL, NULL, NULL, NULL, UMA_ALIGN_PTR, 0);
159 	vfs_vector_op_register(&crossmp_vnodeops);
160 	getnewvnode("crossmp", NULL, &crossmp_vnodeops, &vp_crossmp);
161 }
162 SYSINIT(vfs, SI_SUB_VFS, SI_ORDER_SECOND, nameiinit, NULL);
163 
164 static int lookup_cap_dotdot = 1;
165 SYSCTL_INT(_vfs, OID_AUTO, lookup_cap_dotdot, CTLFLAG_RWTUN,
166     &lookup_cap_dotdot, 0,
167     "enables \"..\" components in path lookup in capability mode");
168 static int lookup_cap_dotdot_nonlocal = 1;
169 SYSCTL_INT(_vfs, OID_AUTO, lookup_cap_dotdot_nonlocal, CTLFLAG_RWTUN,
170     &lookup_cap_dotdot_nonlocal, 0,
171     "enables \"..\" components in path lookup in capability mode "
172     "on non-local mount");
173 
174 static void
175 nameicap_tracker_add(struct nameidata *ndp, struct vnode *dp)
176 {
177 	struct nameicap_tracker *nt;
178 
179 	if ((ndp->ni_lcf & NI_LCF_CAP_DOTDOT) == 0 || dp->v_type != VDIR)
180 		return;
181 	if ((ndp->ni_lcf & (NI_LCF_BENEATH_ABS | NI_LCF_BENEATH_LATCHED)) ==
182 	    NI_LCF_BENEATH_ABS) {
183 		MPASS((ndp->ni_lcf & NI_LCF_LATCH) != 0);
184 		if (dp != ndp->ni_beneath_latch)
185 			return;
186 		ndp->ni_lcf |= NI_LCF_BENEATH_LATCHED;
187 	}
188 	nt = uma_zalloc(nt_zone, M_WAITOK);
189 	vhold(dp);
190 	nt->dp = dp;
191 	TAILQ_INSERT_TAIL(&ndp->ni_cap_tracker, nt, nm_link);
192 }
193 
194 static void
195 nameicap_cleanup(struct nameidata *ndp, bool clean_latch)
196 {
197 	struct nameicap_tracker *nt, *nt1;
198 
199 	KASSERT(TAILQ_EMPTY(&ndp->ni_cap_tracker) ||
200 	    (ndp->ni_lcf & NI_LCF_CAP_DOTDOT) != 0, ("not strictrelative"));
201 	TAILQ_FOREACH_SAFE(nt, &ndp->ni_cap_tracker, nm_link, nt1) {
202 		TAILQ_REMOVE(&ndp->ni_cap_tracker, nt, nm_link);
203 		vdrop(nt->dp);
204 		uma_zfree(nt_zone, nt);
205 	}
206 	if (clean_latch && (ndp->ni_lcf & NI_LCF_LATCH) != 0) {
207 		ndp->ni_lcf &= ~NI_LCF_LATCH;
208 		vrele(ndp->ni_beneath_latch);
209 	}
210 }
211 
212 /*
213  * For dotdot lookups in capability mode, only allow the component
214  * lookup to succeed if the resulting directory was already traversed
215  * during the operation.  Also fail dotdot lookups for non-local
216  * filesystems, where external agents might assist local lookups to
217  * escape the compartment.
218  */
219 static int
220 nameicap_check_dotdot(struct nameidata *ndp, struct vnode *dp)
221 {
222 	struct nameicap_tracker *nt;
223 	struct mount *mp;
224 
225 	if ((ndp->ni_lcf & NI_LCF_CAP_DOTDOT) == 0 || dp == NULL ||
226 	    dp->v_type != VDIR)
227 		return (0);
228 	mp = dp->v_mount;
229 	if (lookup_cap_dotdot_nonlocal == 0 && mp != NULL &&
230 	    (mp->mnt_flag & MNT_LOCAL) == 0)
231 		return (ENOTCAPABLE);
232 	TAILQ_FOREACH_REVERSE(nt, &ndp->ni_cap_tracker, nameicap_tracker_head,
233 	    nm_link) {
234 		if (dp == nt->dp)
235 			return (0);
236 	}
237 	if ((ndp->ni_lcf & NI_LCF_BENEATH_ABS) != 0) {
238 		ndp->ni_lcf &= ~NI_LCF_BENEATH_LATCHED;
239 		nameicap_cleanup(ndp, false);
240 		return (0);
241 	}
242 	return (ENOTCAPABLE);
243 }
244 
245 static void
246 namei_cleanup_cnp(struct componentname *cnp)
247 {
248 
249 	uma_zfree(namei_zone, cnp->cn_pnbuf);
250 #ifdef DIAGNOSTIC
251 	cnp->cn_pnbuf = NULL;
252 	cnp->cn_nameptr = NULL;
253 #endif
254 }
255 
256 static int
257 namei_handle_root(struct nameidata *ndp, struct vnode **dpp, u_int n)
258 {
259 	struct componentname *cnp;
260 
261 	cnp = &ndp->ni_cnd;
262 	if ((ndp->ni_lcf & NI_LCF_STRICTRELATIVE) != 0) {
263 #ifdef KTRACE
264 		if (KTRPOINT(curthread, KTR_CAPFAIL))
265 			ktrcapfail(CAPFAIL_LOOKUP, NULL, NULL);
266 #endif
267 		return (ENOTCAPABLE);
268 	}
269 	if ((cnp->cn_flags & BENEATH) != 0) {
270 		ndp->ni_lcf |= NI_LCF_BENEATH_ABS;
271 		ndp->ni_lcf &= ~NI_LCF_BENEATH_LATCHED;
272 		nameicap_cleanup(ndp, false);
273 	}
274 	while (*(cnp->cn_nameptr) == '/') {
275 		cnp->cn_nameptr++;
276 		ndp->ni_pathlen--;
277 	}
278 	*dpp = ndp->ni_rootdir;
279 	vrefactn(*dpp, n);
280 	return (0);
281 }
282 
283 /*
284  * Convert a pathname into a pointer to a locked vnode.
285  *
286  * The FOLLOW flag is set when symbolic links are to be followed
287  * when they occur at the end of the name translation process.
288  * Symbolic links are always followed for all other pathname
289  * components other than the last.
290  *
291  * The segflg defines whether the name is to be copied from user
292  * space or kernel space.
293  *
294  * Overall outline of namei:
295  *
296  *	copy in name
297  *	get starting directory
298  *	while (!done && !error) {
299  *		call lookup to search path.
300  *		if symbolic link, massage name in buffer and continue
301  *	}
302  */
303 int
304 namei(struct nameidata *ndp)
305 {
306 	struct filedesc *fdp;	/* pointer to file descriptor state */
307 	char *cp;		/* pointer into pathname argument */
308 	struct vnode *dp;	/* the directory we are searching */
309 	struct iovec aiov;		/* uio for reading symbolic links */
310 	struct componentname *cnp;
311 	struct file *dfp;
312 	struct thread *td;
313 	struct proc *p;
314 	cap_rights_t rights;
315 	struct filecaps dirfd_caps;
316 	struct uio auio;
317 	int error, linklen, startdir_used;
318 
319 	cnp = &ndp->ni_cnd;
320 	td = cnp->cn_thread;
321 	p = td->td_proc;
322 	ndp->ni_cnd.cn_cred = ndp->ni_cnd.cn_thread->td_ucred;
323 	KASSERT(cnp->cn_cred && p, ("namei: bad cred/proc"));
324 	KASSERT((cnp->cn_nameiop & (~OPMASK)) == 0,
325 	    ("namei: nameiop contaminated with flags"));
326 	KASSERT((cnp->cn_flags & OPMASK) == 0,
327 	    ("namei: flags contaminated with nameiops"));
328 	MPASS(ndp->ni_startdir == NULL || ndp->ni_startdir->v_type == VDIR ||
329 	    ndp->ni_startdir->v_type == VBAD);
330 	fdp = p->p_fd;
331 	TAILQ_INIT(&ndp->ni_cap_tracker);
332 	ndp->ni_lcf = 0;
333 
334 	/* We will set this ourselves if we need it. */
335 	cnp->cn_flags &= ~TRAILINGSLASH;
336 
337 	/*
338 	 * Get a buffer for the name to be translated, and copy the
339 	 * name into the buffer.
340 	 */
341 	if ((cnp->cn_flags & HASBUF) == 0)
342 		cnp->cn_pnbuf = uma_zalloc(namei_zone, M_WAITOK);
343 	if (ndp->ni_segflg == UIO_SYSSPACE)
344 		error = copystr(ndp->ni_dirp, cnp->cn_pnbuf, MAXPATHLEN,
345 		    &ndp->ni_pathlen);
346 	else
347 		error = copyinstr(ndp->ni_dirp, cnp->cn_pnbuf, MAXPATHLEN,
348 		    &ndp->ni_pathlen);
349 
350 	/*
351 	 * Don't allow empty pathnames.
352 	 */
353 	if (error == 0 && *cnp->cn_pnbuf == '\0')
354 		error = ENOENT;
355 
356 #ifdef CAPABILITY_MODE
357 	/*
358 	 * In capability mode, lookups must be restricted to happen in
359 	 * the subtree with the root specified by the file descriptor:
360 	 * - The root must be real file descriptor, not the pseudo-descriptor
361 	 *   AT_FDCWD.
362 	 * - The passed path must be relative and not absolute.
363 	 * - If lookup_cap_dotdot is disabled, path must not contain the
364 	 *   '..' components.
365 	 * - If lookup_cap_dotdot is enabled, we verify that all '..'
366 	 *   components lookups result in the directories which were
367 	 *   previously walked by us, which prevents an escape from
368 	 *   the relative root.
369 	 */
370 	if (error == 0 && IN_CAPABILITY_MODE(td) &&
371 	    (cnp->cn_flags & NOCAPCHECK) == 0) {
372 		ndp->ni_lcf |= NI_LCF_STRICTRELATIVE;
373 		if (ndp->ni_dirfd == AT_FDCWD) {
374 #ifdef KTRACE
375 			if (KTRPOINT(td, KTR_CAPFAIL))
376 				ktrcapfail(CAPFAIL_LOOKUP, NULL, NULL);
377 #endif
378 			error = ECAPMODE;
379 		}
380 	}
381 #endif
382 	if (error != 0) {
383 		namei_cleanup_cnp(cnp);
384 		ndp->ni_vp = NULL;
385 		return (error);
386 	}
387 	ndp->ni_loopcnt = 0;
388 #ifdef KTRACE
389 	if (KTRPOINT(td, KTR_NAMEI)) {
390 		KASSERT(cnp->cn_thread == curthread,
391 		    ("namei not using curthread"));
392 		ktrnamei(cnp->cn_pnbuf);
393 	}
394 #endif
395 	/*
396 	 * Get starting point for the translation.
397 	 */
398 	FILEDESC_SLOCK(fdp);
399 	/*
400 	 * The reference on ni_rootdir is acquired in the block below to avoid
401 	 * back-to-back atomics for absolute lookups.
402 	 */
403 	ndp->ni_rootdir = fdp->fd_rdir;
404 	ndp->ni_topdir = fdp->fd_jdir;
405 
406 	/*
407 	 * If we are auditing the kernel pathname, save the user pathname.
408 	 */
409 	if (cnp->cn_flags & AUDITVNODE1)
410 		AUDIT_ARG_UPATH1(td, ndp->ni_dirfd, cnp->cn_pnbuf);
411 	if (cnp->cn_flags & AUDITVNODE2)
412 		AUDIT_ARG_UPATH2(td, ndp->ni_dirfd, cnp->cn_pnbuf);
413 
414 	startdir_used = 0;
415 	dp = NULL;
416 	cnp->cn_nameptr = cnp->cn_pnbuf;
417 	if (cnp->cn_pnbuf[0] == '/') {
418 		ndp->ni_resflags |= NIRES_ABS;
419 		error = namei_handle_root(ndp, &dp, 2);
420 		if (error != 0) {
421 			/*
422 			 * Simplify error handling, we should almost never be
423 			 * here.
424 			 */
425 			vrefact(ndp->ni_rootdir);
426 		}
427 	} else {
428 		if (ndp->ni_startdir != NULL) {
429 			vrefact(ndp->ni_rootdir);
430 			dp = ndp->ni_startdir;
431 			startdir_used = 1;
432 		} else if (ndp->ni_dirfd == AT_FDCWD) {
433 			dp = fdp->fd_cdir;
434 			if (dp == ndp->ni_rootdir) {
435 				vrefactn(dp, 2);
436 			} else {
437 				vrefact(ndp->ni_rootdir);
438 				vrefact(dp);
439 			}
440 		} else {
441 			vrefact(ndp->ni_rootdir);
442 			rights = ndp->ni_rightsneeded;
443 			cap_rights_set(&rights, CAP_LOOKUP);
444 
445 			if (cnp->cn_flags & AUDITVNODE1)
446 				AUDIT_ARG_ATFD1(ndp->ni_dirfd);
447 			if (cnp->cn_flags & AUDITVNODE2)
448 				AUDIT_ARG_ATFD2(ndp->ni_dirfd);
449 			/*
450 			 * Effectively inlined fgetvp_rights, because we need to
451 			 * inspect the file as well as grabbing the vnode.
452 			 */
453 			error = fget_cap_locked(fdp, ndp->ni_dirfd, &rights,
454 			    &dfp, &ndp->ni_filecaps);
455 			if (error != 0 || dfp->f_ops == &badfileops ||
456 			    dfp->f_vnode == NULL) {
457 				error = ENOTDIR;
458 			} else {
459 				dp = dfp->f_vnode;
460 				vrefact(dp);
461 
462 				if ((dfp->f_flag & FSEARCH) != 0)
463 					cnp->cn_flags |= NOEXECCHECK;
464 			}
465 #ifdef CAPABILITIES
466 			/*
467 			 * If file descriptor doesn't have all rights,
468 			 * all lookups relative to it must also be
469 			 * strictly relative.
470 			 */
471 			CAP_ALL(&rights);
472 			if (!cap_rights_contains(&ndp->ni_filecaps.fc_rights,
473 			    &rights) ||
474 			    ndp->ni_filecaps.fc_fcntls != CAP_FCNTL_ALL ||
475 			    ndp->ni_filecaps.fc_nioctls != -1) {
476 				ndp->ni_lcf |= NI_LCF_STRICTRELATIVE;
477 			}
478 #endif
479 		}
480 		if (error == 0 && dp->v_type != VDIR)
481 			error = ENOTDIR;
482 	}
483 	if (error == 0 && (cnp->cn_flags & BENEATH) != 0) {
484 		if (ndp->ni_dirfd == AT_FDCWD) {
485 			ndp->ni_beneath_latch = fdp->fd_cdir;
486 			vrefact(ndp->ni_beneath_latch);
487 		} else {
488 			rights = ndp->ni_rightsneeded;
489 			cap_rights_set(&rights, CAP_LOOKUP);
490 			error = fgetvp_rights(td, ndp->ni_dirfd, &rights,
491 			    &dirfd_caps, &ndp->ni_beneath_latch);
492 			if (error == 0 && dp->v_type != VDIR) {
493 				vrele(ndp->ni_beneath_latch);
494 				error = ENOTDIR;
495 			}
496 		}
497 		if (error == 0)
498 			ndp->ni_lcf |= NI_LCF_LATCH;
499 	}
500 	FILEDESC_SUNLOCK(fdp);
501 	if (ndp->ni_startdir != NULL && !startdir_used)
502 		vrele(ndp->ni_startdir);
503 	if (error != 0) {
504 		if (dp != NULL)
505 			vrele(dp);
506 		goto out;
507 	}
508 	MPASS((ndp->ni_lcf & (NI_LCF_BENEATH_ABS | NI_LCF_LATCH)) !=
509 	    NI_LCF_BENEATH_ABS);
510 	if (((ndp->ni_lcf & NI_LCF_STRICTRELATIVE) != 0 &&
511 	    lookup_cap_dotdot != 0) ||
512 	    ((ndp->ni_lcf & NI_LCF_STRICTRELATIVE) == 0 &&
513 	    (cnp->cn_flags & BENEATH) != 0))
514 		ndp->ni_lcf |= NI_LCF_CAP_DOTDOT;
515 	SDT_PROBE3(vfs, namei, lookup, entry, dp, cnp->cn_pnbuf,
516 	    cnp->cn_flags);
517 	for (;;) {
518 		ndp->ni_startdir = dp;
519 		error = lookup(ndp);
520 		if (error != 0)
521 			goto out;
522 		/*
523 		 * If not a symbolic link, we're done.
524 		 */
525 		if ((cnp->cn_flags & ISSYMLINK) == 0) {
526 			vrele(ndp->ni_rootdir);
527 			if ((cnp->cn_flags & (SAVENAME | SAVESTART)) == 0) {
528 				namei_cleanup_cnp(cnp);
529 			} else
530 				cnp->cn_flags |= HASBUF;
531 			if ((ndp->ni_lcf & (NI_LCF_BENEATH_ABS |
532 			    NI_LCF_BENEATH_LATCHED)) == NI_LCF_BENEATH_ABS) {
533 				NDFREE(ndp, 0);
534 				error = ENOTCAPABLE;
535 			}
536 			nameicap_cleanup(ndp, true);
537 			SDT_PROBE2(vfs, namei, lookup, return, error,
538 			    (error == 0 ? ndp->ni_vp : NULL));
539 			return (error);
540 		}
541 		if (ndp->ni_loopcnt++ >= MAXSYMLINKS) {
542 			error = ELOOP;
543 			break;
544 		}
545 #ifdef MAC
546 		if ((cnp->cn_flags & NOMACCHECK) == 0) {
547 			error = mac_vnode_check_readlink(td->td_ucred,
548 			    ndp->ni_vp);
549 			if (error != 0)
550 				break;
551 		}
552 #endif
553 		if (ndp->ni_pathlen > 1)
554 			cp = uma_zalloc(namei_zone, M_WAITOK);
555 		else
556 			cp = cnp->cn_pnbuf;
557 		aiov.iov_base = cp;
558 		aiov.iov_len = MAXPATHLEN;
559 		auio.uio_iov = &aiov;
560 		auio.uio_iovcnt = 1;
561 		auio.uio_offset = 0;
562 		auio.uio_rw = UIO_READ;
563 		auio.uio_segflg = UIO_SYSSPACE;
564 		auio.uio_td = td;
565 		auio.uio_resid = MAXPATHLEN;
566 		error = VOP_READLINK(ndp->ni_vp, &auio, cnp->cn_cred);
567 		if (error != 0) {
568 			if (ndp->ni_pathlen > 1)
569 				uma_zfree(namei_zone, cp);
570 			break;
571 		}
572 		linklen = MAXPATHLEN - auio.uio_resid;
573 		if (linklen == 0) {
574 			if (ndp->ni_pathlen > 1)
575 				uma_zfree(namei_zone, cp);
576 			error = ENOENT;
577 			break;
578 		}
579 		if (linklen + ndp->ni_pathlen > MAXPATHLEN) {
580 			if (ndp->ni_pathlen > 1)
581 				uma_zfree(namei_zone, cp);
582 			error = ENAMETOOLONG;
583 			break;
584 		}
585 		if (ndp->ni_pathlen > 1) {
586 			bcopy(ndp->ni_next, cp + linklen, ndp->ni_pathlen);
587 			uma_zfree(namei_zone, cnp->cn_pnbuf);
588 			cnp->cn_pnbuf = cp;
589 		} else
590 			cnp->cn_pnbuf[linklen] = '\0';
591 		ndp->ni_pathlen += linklen;
592 		vput(ndp->ni_vp);
593 		dp = ndp->ni_dvp;
594 		/*
595 		 * Check if root directory should replace current directory.
596 		 */
597 		cnp->cn_nameptr = cnp->cn_pnbuf;
598 		if (*(cnp->cn_nameptr) == '/') {
599 			vrele(dp);
600 			error = namei_handle_root(ndp, &dp, 1);
601 			if (error != 0)
602 				goto out;
603 		}
604 	}
605 	vput(ndp->ni_vp);
606 	ndp->ni_vp = NULL;
607 	vrele(ndp->ni_dvp);
608 out:
609 	vrele(ndp->ni_rootdir);
610 	MPASS(error != 0);
611 	namei_cleanup_cnp(cnp);
612 	nameicap_cleanup(ndp, true);
613 	SDT_PROBE2(vfs, namei, lookup, return, error, NULL);
614 	return (error);
615 }
616 
617 static int
618 compute_cn_lkflags(struct mount *mp, int lkflags, int cnflags)
619 {
620 
621 	if (mp == NULL || ((lkflags & LK_SHARED) &&
622 	    (!(mp->mnt_kern_flag & MNTK_LOOKUP_SHARED) ||
623 	    ((cnflags & ISDOTDOT) &&
624 	    (mp->mnt_kern_flag & MNTK_LOOKUP_EXCL_DOTDOT))))) {
625 		lkflags &= ~LK_SHARED;
626 		lkflags |= LK_EXCLUSIVE;
627 	}
628 	lkflags |= LK_NODDLKTREAT;
629 	return (lkflags);
630 }
631 
632 static __inline int
633 needs_exclusive_leaf(struct mount *mp, int flags)
634 {
635 
636 	/*
637 	 * Intermediate nodes can use shared locks, we only need to
638 	 * force an exclusive lock for leaf nodes.
639 	 */
640 	if ((flags & (ISLASTCN | LOCKLEAF)) != (ISLASTCN | LOCKLEAF))
641 		return (0);
642 
643 	/* Always use exclusive locks if LOCKSHARED isn't set. */
644 	if (!(flags & LOCKSHARED))
645 		return (1);
646 
647 	/*
648 	 * For lookups during open(), if the mount point supports
649 	 * extended shared operations, then use a shared lock for the
650 	 * leaf node, otherwise use an exclusive lock.
651 	 */
652 	if ((flags & ISOPEN) != 0)
653 		return (!MNT_EXTENDED_SHARED(mp));
654 
655 	/*
656 	 * Lookup requests outside of open() that specify LOCKSHARED
657 	 * only need a shared lock on the leaf vnode.
658 	 */
659 	return (0);
660 }
661 
662 /*
663  * Search a pathname.
664  * This is a very central and rather complicated routine.
665  *
666  * The pathname is pointed to by ni_ptr and is of length ni_pathlen.
667  * The starting directory is taken from ni_startdir. The pathname is
668  * descended until done, or a symbolic link is encountered. The variable
669  * ni_more is clear if the path is completed; it is set to one if a
670  * symbolic link needing interpretation is encountered.
671  *
672  * The flag argument is LOOKUP, CREATE, RENAME, or DELETE depending on
673  * whether the name is to be looked up, created, renamed, or deleted.
674  * When CREATE, RENAME, or DELETE is specified, information usable in
675  * creating, renaming, or deleting a directory entry may be calculated.
676  * If flag has LOCKPARENT or'ed into it, the parent directory is returned
677  * locked. If flag has WANTPARENT or'ed into it, the parent directory is
678  * returned unlocked. Otherwise the parent directory is not returned. If
679  * the target of the pathname exists and LOCKLEAF is or'ed into the flag
680  * the target is returned locked, otherwise it is returned unlocked.
681  * When creating or renaming and LOCKPARENT is specified, the target may not
682  * be ".".  When deleting and LOCKPARENT is specified, the target may be ".".
683  *
684  * Overall outline of lookup:
685  *
686  * dirloop:
687  *	identify next component of name at ndp->ni_ptr
688  *	handle degenerate case where name is null string
689  *	if .. and crossing mount points and on mounted filesys, find parent
690  *	call VOP_LOOKUP routine for next component name
691  *	    directory vnode returned in ni_dvp, unlocked unless LOCKPARENT set
692  *	    component vnode returned in ni_vp (if it exists), locked.
693  *	if result vnode is mounted on and crossing mount points,
694  *	    find mounted on vnode
695  *	if more components of name, do next level at dirloop
696  *	return the answer in ni_vp, locked if LOCKLEAF set
697  *	    if LOCKPARENT set, return locked parent in ni_dvp
698  *	    if WANTPARENT set, return unlocked parent in ni_dvp
699  */
700 int
701 lookup(struct nameidata *ndp)
702 {
703 	char *cp;			/* pointer into pathname argument */
704 	char *prev_ni_next;		/* saved ndp->ni_next */
705 	struct vnode *dp = NULL;	/* the directory we are searching */
706 	struct vnode *tdp;		/* saved dp */
707 	struct mount *mp;		/* mount table entry */
708 	struct prison *pr;
709 	size_t prev_ni_pathlen;		/* saved ndp->ni_pathlen */
710 	int docache;			/* == 0 do not cache last component */
711 	int wantparent;			/* 1 => wantparent or lockparent flag */
712 	int rdonly;			/* lookup read-only flag bit */
713 	int error = 0;
714 	int dpunlocked = 0;		/* dp has already been unlocked */
715 	int relookup = 0;		/* do not consume the path component */
716 	struct componentname *cnp = &ndp->ni_cnd;
717 	int lkflags_save;
718 	int ni_dvp_unlocked;
719 
720 	/*
721 	 * Setup: break out flag bits into variables.
722 	 */
723 	ni_dvp_unlocked = 0;
724 	wantparent = cnp->cn_flags & (LOCKPARENT | WANTPARENT);
725 	KASSERT(cnp->cn_nameiop == LOOKUP || wantparent,
726 	    ("CREATE, DELETE, RENAME require LOCKPARENT or WANTPARENT."));
727 	docache = (cnp->cn_flags & NOCACHE) ^ NOCACHE;
728 	if (cnp->cn_nameiop == DELETE ||
729 	    (wantparent && cnp->cn_nameiop != CREATE &&
730 	     cnp->cn_nameiop != LOOKUP))
731 		docache = 0;
732 	rdonly = cnp->cn_flags & RDONLY;
733 	cnp->cn_flags &= ~ISSYMLINK;
734 	ndp->ni_dvp = NULL;
735 	/*
736 	 * We use shared locks until we hit the parent of the last cn then
737 	 * we adjust based on the requesting flags.
738 	 */
739 	cnp->cn_lkflags = LK_SHARED;
740 	dp = ndp->ni_startdir;
741 	ndp->ni_startdir = NULLVP;
742 	vn_lock(dp,
743 	    compute_cn_lkflags(dp->v_mount, cnp->cn_lkflags | LK_RETRY,
744 	    cnp->cn_flags));
745 
746 dirloop:
747 	/*
748 	 * Search a new directory.
749 	 *
750 	 * The last component of the filename is left accessible via
751 	 * cnp->cn_nameptr for callers that need the name. Callers needing
752 	 * the name set the SAVENAME flag. When done, they assume
753 	 * responsibility for freeing the pathname buffer.
754 	 */
755 	for (cp = cnp->cn_nameptr; *cp != 0 && *cp != '/'; cp++)
756 		continue;
757 	cnp->cn_namelen = cp - cnp->cn_nameptr;
758 	if (cnp->cn_namelen > NAME_MAX) {
759 		error = ENAMETOOLONG;
760 		goto bad;
761 	}
762 #ifdef NAMEI_DIAGNOSTIC
763 	{ char c = *cp;
764 	*cp = '\0';
765 	printf("{%s}: ", cnp->cn_nameptr);
766 	*cp = c; }
767 #endif
768 	prev_ni_pathlen = ndp->ni_pathlen;
769 	ndp->ni_pathlen -= cnp->cn_namelen;
770 	KASSERT(ndp->ni_pathlen <= PATH_MAX,
771 	    ("%s: ni_pathlen underflow to %zd\n", __func__, ndp->ni_pathlen));
772 	prev_ni_next = ndp->ni_next;
773 	ndp->ni_next = cp;
774 
775 	/*
776 	 * Replace multiple slashes by a single slash and trailing slashes
777 	 * by a null.  This must be done before VOP_LOOKUP() because some
778 	 * fs's don't know about trailing slashes.  Remember if there were
779 	 * trailing slashes to handle symlinks, existing non-directories
780 	 * and non-existing files that won't be directories specially later.
781 	 */
782 	while (*cp == '/' && (cp[1] == '/' || cp[1] == '\0')) {
783 		cp++;
784 		ndp->ni_pathlen--;
785 		if (*cp == '\0') {
786 			*ndp->ni_next = '\0';
787 			cnp->cn_flags |= TRAILINGSLASH;
788 		}
789 	}
790 	ndp->ni_next = cp;
791 
792 	cnp->cn_flags |= MAKEENTRY;
793 	if (*cp == '\0' && docache == 0)
794 		cnp->cn_flags &= ~MAKEENTRY;
795 	if (cnp->cn_namelen == 2 &&
796 	    cnp->cn_nameptr[1] == '.' && cnp->cn_nameptr[0] == '.')
797 		cnp->cn_flags |= ISDOTDOT;
798 	else
799 		cnp->cn_flags &= ~ISDOTDOT;
800 	if (*ndp->ni_next == 0)
801 		cnp->cn_flags |= ISLASTCN;
802 	else
803 		cnp->cn_flags &= ~ISLASTCN;
804 
805 	if ((cnp->cn_flags & ISLASTCN) != 0 &&
806 	    cnp->cn_namelen == 1 && cnp->cn_nameptr[0] == '.' &&
807 	    (cnp->cn_nameiop == DELETE || cnp->cn_nameiop == RENAME)) {
808 		error = EINVAL;
809 		goto bad;
810 	}
811 
812 	nameicap_tracker_add(ndp, dp);
813 
814 	/*
815 	 * Check for degenerate name (e.g. / or "")
816 	 * which is a way of talking about a directory,
817 	 * e.g. like "/." or ".".
818 	 */
819 	if (cnp->cn_nameptr[0] == '\0') {
820 		if (dp->v_type != VDIR) {
821 			error = ENOTDIR;
822 			goto bad;
823 		}
824 		if (cnp->cn_nameiop != LOOKUP) {
825 			error = EISDIR;
826 			goto bad;
827 		}
828 		if (wantparent) {
829 			ndp->ni_dvp = dp;
830 			VREF(dp);
831 		}
832 		ndp->ni_vp = dp;
833 
834 		if (cnp->cn_flags & AUDITVNODE1)
835 			AUDIT_ARG_VNODE1(dp);
836 		else if (cnp->cn_flags & AUDITVNODE2)
837 			AUDIT_ARG_VNODE2(dp);
838 
839 		if (!(cnp->cn_flags & (LOCKPARENT | LOCKLEAF)))
840 			VOP_UNLOCK(dp);
841 		/* XXX This should probably move to the top of function. */
842 		if (cnp->cn_flags & SAVESTART)
843 			panic("lookup: SAVESTART");
844 		goto success;
845 	}
846 
847 	/*
848 	 * Handle "..": five special cases.
849 	 * 0. If doing a capability lookup and lookup_cap_dotdot is
850 	 *    disabled, return ENOTCAPABLE.
851 	 * 1. Return an error if this is the last component of
852 	 *    the name and the operation is DELETE or RENAME.
853 	 * 2. If at root directory (e.g. after chroot)
854 	 *    or at absolute root directory
855 	 *    then ignore it so can't get out.
856 	 * 3. If this vnode is the root of a mounted
857 	 *    filesystem, then replace it with the
858 	 *    vnode which was mounted on so we take the
859 	 *    .. in the other filesystem.
860 	 * 4. If the vnode is the top directory of
861 	 *    the jail or chroot, don't let them out.
862 	 * 5. If doing a capability lookup and lookup_cap_dotdot is
863 	 *    enabled, return ENOTCAPABLE if the lookup would escape
864 	 *    from the initial file descriptor directory.  Checks are
865 	 *    done by ensuring that namei() already traversed the
866 	 *    result of dotdot lookup.
867 	 */
868 	if (cnp->cn_flags & ISDOTDOT) {
869 		if ((ndp->ni_lcf & (NI_LCF_STRICTRELATIVE | NI_LCF_CAP_DOTDOT))
870 		    == NI_LCF_STRICTRELATIVE) {
871 #ifdef KTRACE
872 			if (KTRPOINT(curthread, KTR_CAPFAIL))
873 				ktrcapfail(CAPFAIL_LOOKUP, NULL, NULL);
874 #endif
875 			error = ENOTCAPABLE;
876 			goto bad;
877 		}
878 		if ((cnp->cn_flags & ISLASTCN) != 0 &&
879 		    (cnp->cn_nameiop == DELETE || cnp->cn_nameiop == RENAME)) {
880 			error = EINVAL;
881 			goto bad;
882 		}
883 		for (;;) {
884 			for (pr = cnp->cn_cred->cr_prison; pr != NULL;
885 			     pr = pr->pr_parent)
886 				if (dp == pr->pr_root)
887 					break;
888 			if (dp == ndp->ni_rootdir ||
889 			    dp == ndp->ni_topdir ||
890 			    dp == rootvnode ||
891 			    pr != NULL ||
892 			    ((dp->v_vflag & VV_ROOT) != 0 &&
893 			     (cnp->cn_flags & NOCROSSMOUNT) != 0)) {
894 				ndp->ni_dvp = dp;
895 				ndp->ni_vp = dp;
896 				VREF(dp);
897 				goto nextname;
898 			}
899 			if ((dp->v_vflag & VV_ROOT) == 0)
900 				break;
901 			if (VN_IS_DOOMED(dp)) {	/* forced unmount */
902 				error = ENOENT;
903 				goto bad;
904 			}
905 			tdp = dp;
906 			dp = dp->v_mount->mnt_vnodecovered;
907 			VREF(dp);
908 			vput(tdp);
909 			vn_lock(dp,
910 			    compute_cn_lkflags(dp->v_mount, cnp->cn_lkflags |
911 			    LK_RETRY, ISDOTDOT));
912 			error = nameicap_check_dotdot(ndp, dp);
913 			if (error != 0) {
914 #ifdef KTRACE
915 				if (KTRPOINT(curthread, KTR_CAPFAIL))
916 					ktrcapfail(CAPFAIL_LOOKUP, NULL, NULL);
917 #endif
918 				goto bad;
919 			}
920 		}
921 	}
922 
923 	/*
924 	 * We now have a segment name to search for, and a directory to search.
925 	 */
926 unionlookup:
927 #ifdef MAC
928 	if ((cnp->cn_flags & NOMACCHECK) == 0) {
929 		error = mac_vnode_check_lookup(cnp->cn_thread->td_ucred, dp,
930 		    cnp);
931 		if (error)
932 			goto bad;
933 	}
934 #endif
935 	ndp->ni_dvp = dp;
936 	ndp->ni_vp = NULL;
937 	ASSERT_VOP_LOCKED(dp, "lookup");
938 	/*
939 	 * If we have a shared lock we may need to upgrade the lock for the
940 	 * last operation.
941 	 */
942 	if ((cnp->cn_flags & LOCKPARENT) && (cnp->cn_flags & ISLASTCN) &&
943 	    dp != vp_crossmp && VOP_ISLOCKED(dp) == LK_SHARED)
944 		vn_lock(dp, LK_UPGRADE|LK_RETRY);
945 	if (VN_IS_DOOMED(dp)) {
946 		error = ENOENT;
947 		goto bad;
948 	}
949 	/*
950 	 * If we're looking up the last component and we need an exclusive
951 	 * lock, adjust our lkflags.
952 	 */
953 	if (needs_exclusive_leaf(dp->v_mount, cnp->cn_flags))
954 		cnp->cn_lkflags = LK_EXCLUSIVE;
955 #ifdef NAMEI_DIAGNOSTIC
956 	vn_printf(dp, "lookup in ");
957 #endif
958 	lkflags_save = cnp->cn_lkflags;
959 	cnp->cn_lkflags = compute_cn_lkflags(dp->v_mount, cnp->cn_lkflags,
960 	    cnp->cn_flags);
961 	error = VOP_LOOKUP(dp, &ndp->ni_vp, cnp);
962 	cnp->cn_lkflags = lkflags_save;
963 	if (error != 0) {
964 		KASSERT(ndp->ni_vp == NULL, ("leaf should be empty"));
965 #ifdef NAMEI_DIAGNOSTIC
966 		printf("not found\n");
967 #endif
968 		if ((error == ENOENT) &&
969 		    (dp->v_vflag & VV_ROOT) && (dp->v_mount != NULL) &&
970 		    (dp->v_mount->mnt_flag & MNT_UNION)) {
971 			tdp = dp;
972 			dp = dp->v_mount->mnt_vnodecovered;
973 			VREF(dp);
974 			vput(tdp);
975 			vn_lock(dp,
976 			    compute_cn_lkflags(dp->v_mount, cnp->cn_lkflags |
977 			    LK_RETRY, cnp->cn_flags));
978 			nameicap_tracker_add(ndp, dp);
979 			goto unionlookup;
980 		}
981 
982 		if (error == ERELOOKUP) {
983 			vref(dp);
984 			ndp->ni_vp = dp;
985 			error = 0;
986 			relookup = 1;
987 			goto good;
988 		}
989 
990 		if (error != EJUSTRETURN)
991 			goto bad;
992 		/*
993 		 * At this point, we know we're at the end of the
994 		 * pathname.  If creating / renaming, we can consider
995 		 * allowing the file or directory to be created / renamed,
996 		 * provided we're not on a read-only filesystem.
997 		 */
998 		if (rdonly) {
999 			error = EROFS;
1000 			goto bad;
1001 		}
1002 		/* trailing slash only allowed for directories */
1003 		if ((cnp->cn_flags & TRAILINGSLASH) &&
1004 		    !(cnp->cn_flags & WILLBEDIR)) {
1005 			error = ENOENT;
1006 			goto bad;
1007 		}
1008 		if ((cnp->cn_flags & LOCKPARENT) == 0)
1009 			VOP_UNLOCK(dp);
1010 		/*
1011 		 * We return with ni_vp NULL to indicate that the entry
1012 		 * doesn't currently exist, leaving a pointer to the
1013 		 * (possibly locked) directory vnode in ndp->ni_dvp.
1014 		 */
1015 		if (cnp->cn_flags & SAVESTART) {
1016 			ndp->ni_startdir = ndp->ni_dvp;
1017 			VREF(ndp->ni_startdir);
1018 		}
1019 		goto success;
1020 	}
1021 
1022 good:
1023 #ifdef NAMEI_DIAGNOSTIC
1024 	printf("found\n");
1025 #endif
1026 	dp = ndp->ni_vp;
1027 
1028 	/*
1029 	 * Check to see if the vnode has been mounted on;
1030 	 * if so find the root of the mounted filesystem.
1031 	 */
1032 	while (dp->v_type == VDIR && (mp = dp->v_mountedhere) &&
1033 	       (cnp->cn_flags & NOCROSSMOUNT) == 0) {
1034 		if (vfs_busy(mp, 0))
1035 			continue;
1036 		vput(dp);
1037 		if (dp != ndp->ni_dvp)
1038 			vput(ndp->ni_dvp);
1039 		else
1040 			vrele(ndp->ni_dvp);
1041 		vrefact(vp_crossmp);
1042 		ndp->ni_dvp = vp_crossmp;
1043 		error = VFS_ROOT(mp, compute_cn_lkflags(mp, cnp->cn_lkflags,
1044 		    cnp->cn_flags), &tdp);
1045 		vfs_unbusy(mp);
1046 		if (vn_lock(vp_crossmp, LK_SHARED | LK_NOWAIT))
1047 			panic("vp_crossmp exclusively locked or reclaimed");
1048 		if (error) {
1049 			dpunlocked = 1;
1050 			goto bad2;
1051 		}
1052 		ndp->ni_vp = dp = tdp;
1053 	}
1054 
1055 	/*
1056 	 * Check for symbolic link
1057 	 */
1058 	if ((dp->v_type == VLNK) &&
1059 	    ((cnp->cn_flags & FOLLOW) || (cnp->cn_flags & TRAILINGSLASH) ||
1060 	     *ndp->ni_next == '/')) {
1061 		cnp->cn_flags |= ISSYMLINK;
1062 		if (VN_IS_DOOMED(dp)) {
1063 			/*
1064 			 * We can't know whether the directory was mounted with
1065 			 * NOSYMFOLLOW, so we can't follow safely.
1066 			 */
1067 			error = ENOENT;
1068 			goto bad2;
1069 		}
1070 		if (dp->v_mount->mnt_flag & MNT_NOSYMFOLLOW) {
1071 			error = EACCES;
1072 			goto bad2;
1073 		}
1074 		/*
1075 		 * Symlink code always expects an unlocked dvp.
1076 		 */
1077 		if (ndp->ni_dvp != ndp->ni_vp) {
1078 			VOP_UNLOCK(ndp->ni_dvp);
1079 			ni_dvp_unlocked = 1;
1080 		}
1081 		goto success;
1082 	}
1083 
1084 nextname:
1085 	/*
1086 	 * Not a symbolic link that we will follow.  Continue with the
1087 	 * next component if there is any; otherwise, we're done.
1088 	 */
1089 	KASSERT((cnp->cn_flags & ISLASTCN) || *ndp->ni_next == '/',
1090 	    ("lookup: invalid path state."));
1091 	if (relookup) {
1092 		relookup = 0;
1093 		ndp->ni_pathlen = prev_ni_pathlen;
1094 		ndp->ni_next = prev_ni_next;
1095 		if (ndp->ni_dvp != dp)
1096 			vput(ndp->ni_dvp);
1097 		else
1098 			vrele(ndp->ni_dvp);
1099 		goto dirloop;
1100 	}
1101 	if (cnp->cn_flags & ISDOTDOT) {
1102 		error = nameicap_check_dotdot(ndp, ndp->ni_vp);
1103 		if (error != 0) {
1104 #ifdef KTRACE
1105 			if (KTRPOINT(curthread, KTR_CAPFAIL))
1106 				ktrcapfail(CAPFAIL_LOOKUP, NULL, NULL);
1107 #endif
1108 			goto bad2;
1109 		}
1110 	}
1111 	if (*ndp->ni_next == '/') {
1112 		cnp->cn_nameptr = ndp->ni_next;
1113 		while (*cnp->cn_nameptr == '/') {
1114 			cnp->cn_nameptr++;
1115 			ndp->ni_pathlen--;
1116 		}
1117 		if (ndp->ni_dvp != dp)
1118 			vput(ndp->ni_dvp);
1119 		else
1120 			vrele(ndp->ni_dvp);
1121 		goto dirloop;
1122 	}
1123 	/*
1124 	 * If we're processing a path with a trailing slash,
1125 	 * check that the end result is a directory.
1126 	 */
1127 	if ((cnp->cn_flags & TRAILINGSLASH) && dp->v_type != VDIR) {
1128 		error = ENOTDIR;
1129 		goto bad2;
1130 	}
1131 	/*
1132 	 * Disallow directory write attempts on read-only filesystems.
1133 	 */
1134 	if (rdonly &&
1135 	    (cnp->cn_nameiop == DELETE || cnp->cn_nameiop == RENAME)) {
1136 		error = EROFS;
1137 		goto bad2;
1138 	}
1139 	if (cnp->cn_flags & SAVESTART) {
1140 		ndp->ni_startdir = ndp->ni_dvp;
1141 		VREF(ndp->ni_startdir);
1142 	}
1143 	if (!wantparent) {
1144 		ni_dvp_unlocked = 2;
1145 		if (ndp->ni_dvp != dp)
1146 			vput(ndp->ni_dvp);
1147 		else
1148 			vrele(ndp->ni_dvp);
1149 	} else if ((cnp->cn_flags & LOCKPARENT) == 0 && ndp->ni_dvp != dp) {
1150 		VOP_UNLOCK(ndp->ni_dvp);
1151 		ni_dvp_unlocked = 1;
1152 	}
1153 
1154 	if (cnp->cn_flags & AUDITVNODE1)
1155 		AUDIT_ARG_VNODE1(dp);
1156 	else if (cnp->cn_flags & AUDITVNODE2)
1157 		AUDIT_ARG_VNODE2(dp);
1158 
1159 	if ((cnp->cn_flags & LOCKLEAF) == 0)
1160 		VOP_UNLOCK(dp);
1161 success:
1162 	/*
1163 	 * Because of shared lookup we may have the vnode shared locked, but
1164 	 * the caller may want it to be exclusively locked.
1165 	 */
1166 	if (needs_exclusive_leaf(dp->v_mount, cnp->cn_flags) &&
1167 	    VOP_ISLOCKED(dp) != LK_EXCLUSIVE) {
1168 		vn_lock(dp, LK_UPGRADE | LK_RETRY);
1169 		if (VN_IS_DOOMED(dp)) {
1170 			error = ENOENT;
1171 			goto bad2;
1172 		}
1173 	}
1174 	return (0);
1175 
1176 bad2:
1177 	if (ni_dvp_unlocked != 2) {
1178 		if (dp != ndp->ni_dvp && !ni_dvp_unlocked)
1179 			vput(ndp->ni_dvp);
1180 		else
1181 			vrele(ndp->ni_dvp);
1182 	}
1183 bad:
1184 	if (!dpunlocked)
1185 		vput(dp);
1186 	ndp->ni_vp = NULL;
1187 	return (error);
1188 }
1189 
1190 /*
1191  * relookup - lookup a path name component
1192  *    Used by lookup to re-acquire things.
1193  */
1194 int
1195 relookup(struct vnode *dvp, struct vnode **vpp, struct componentname *cnp)
1196 {
1197 	struct vnode *dp = NULL;		/* the directory we are searching */
1198 	int wantparent;			/* 1 => wantparent or lockparent flag */
1199 	int rdonly;			/* lookup read-only flag bit */
1200 	int error = 0;
1201 
1202 	KASSERT(cnp->cn_flags & ISLASTCN,
1203 	    ("relookup: Not given last component."));
1204 	/*
1205 	 * Setup: break out flag bits into variables.
1206 	 */
1207 	wantparent = cnp->cn_flags & (LOCKPARENT|WANTPARENT);
1208 	KASSERT(wantparent, ("relookup: parent not wanted."));
1209 	rdonly = cnp->cn_flags & RDONLY;
1210 	cnp->cn_flags &= ~ISSYMLINK;
1211 	dp = dvp;
1212 	cnp->cn_lkflags = LK_EXCLUSIVE;
1213 	vn_lock(dp, LK_EXCLUSIVE | LK_RETRY);
1214 
1215 	/*
1216 	 * Search a new directory.
1217 	 *
1218 	 * The last component of the filename is left accessible via
1219 	 * cnp->cn_nameptr for callers that need the name. Callers needing
1220 	 * the name set the SAVENAME flag. When done, they assume
1221 	 * responsibility for freeing the pathname buffer.
1222 	 */
1223 #ifdef NAMEI_DIAGNOSTIC
1224 	printf("{%s}: ", cnp->cn_nameptr);
1225 #endif
1226 
1227 	/*
1228 	 * Check for "" which represents the root directory after slash
1229 	 * removal.
1230 	 */
1231 	if (cnp->cn_nameptr[0] == '\0') {
1232 		/*
1233 		 * Support only LOOKUP for "/" because lookup()
1234 		 * can't succeed for CREATE, DELETE and RENAME.
1235 		 */
1236 		KASSERT(cnp->cn_nameiop == LOOKUP, ("nameiop must be LOOKUP"));
1237 		KASSERT(dp->v_type == VDIR, ("dp is not a directory"));
1238 
1239 		if (!(cnp->cn_flags & LOCKLEAF))
1240 			VOP_UNLOCK(dp);
1241 		*vpp = dp;
1242 		/* XXX This should probably move to the top of function. */
1243 		if (cnp->cn_flags & SAVESTART)
1244 			panic("lookup: SAVESTART");
1245 		return (0);
1246 	}
1247 
1248 	if (cnp->cn_flags & ISDOTDOT)
1249 		panic ("relookup: lookup on dot-dot");
1250 
1251 	/*
1252 	 * We now have a segment name to search for, and a directory to search.
1253 	 */
1254 #ifdef NAMEI_DIAGNOSTIC
1255 	vn_printf(dp, "search in ");
1256 #endif
1257 	if ((error = VOP_LOOKUP(dp, vpp, cnp)) != 0) {
1258 		KASSERT(*vpp == NULL, ("leaf should be empty"));
1259 		if (error != EJUSTRETURN)
1260 			goto bad;
1261 		/*
1262 		 * If creating and at end of pathname, then can consider
1263 		 * allowing file to be created.
1264 		 */
1265 		if (rdonly) {
1266 			error = EROFS;
1267 			goto bad;
1268 		}
1269 		/* ASSERT(dvp == ndp->ni_startdir) */
1270 		if (cnp->cn_flags & SAVESTART)
1271 			VREF(dvp);
1272 		if ((cnp->cn_flags & LOCKPARENT) == 0)
1273 			VOP_UNLOCK(dp);
1274 		/*
1275 		 * We return with ni_vp NULL to indicate that the entry
1276 		 * doesn't currently exist, leaving a pointer to the
1277 		 * (possibly locked) directory vnode in ndp->ni_dvp.
1278 		 */
1279 		return (0);
1280 	}
1281 
1282 	dp = *vpp;
1283 
1284 	/*
1285 	 * Disallow directory write attempts on read-only filesystems.
1286 	 */
1287 	if (rdonly &&
1288 	    (cnp->cn_nameiop == DELETE || cnp->cn_nameiop == RENAME)) {
1289 		if (dvp == dp)
1290 			vrele(dvp);
1291 		else
1292 			vput(dvp);
1293 		error = EROFS;
1294 		goto bad;
1295 	}
1296 	/*
1297 	 * Set the parent lock/ref state to the requested state.
1298 	 */
1299 	if ((cnp->cn_flags & LOCKPARENT) == 0 && dvp != dp) {
1300 		if (wantparent)
1301 			VOP_UNLOCK(dvp);
1302 		else
1303 			vput(dvp);
1304 	} else if (!wantparent)
1305 		vrele(dvp);
1306 	/*
1307 	 * Check for symbolic link
1308 	 */
1309 	KASSERT(dp->v_type != VLNK || !(cnp->cn_flags & FOLLOW),
1310 	    ("relookup: symlink found.\n"));
1311 
1312 	/* ASSERT(dvp == ndp->ni_startdir) */
1313 	if (cnp->cn_flags & SAVESTART)
1314 		VREF(dvp);
1315 
1316 	if ((cnp->cn_flags & LOCKLEAF) == 0)
1317 		VOP_UNLOCK(dp);
1318 	return (0);
1319 bad:
1320 	vput(dp);
1321 	*vpp = NULL;
1322 	return (error);
1323 }
1324 
1325 void
1326 NDINIT_ALL(struct nameidata *ndp, u_long op, u_long flags, enum uio_seg segflg,
1327     const char *namep, int dirfd, struct vnode *startdir, cap_rights_t *rightsp,
1328     struct thread *td)
1329 {
1330 
1331 	ndp->ni_cnd.cn_nameiop = op;
1332 	ndp->ni_cnd.cn_flags = flags;
1333 	ndp->ni_segflg = segflg;
1334 	ndp->ni_dirp = namep;
1335 	ndp->ni_dirfd = dirfd;
1336 	ndp->ni_startdir = startdir;
1337 	ndp->ni_resflags = 0;
1338 	filecaps_init(&ndp->ni_filecaps);
1339 	ndp->ni_cnd.cn_thread = td;
1340 	if (rightsp != NULL)
1341 		ndp->ni_rightsneeded = *rightsp;
1342 	else
1343 		cap_rights_init(&ndp->ni_rightsneeded);
1344 }
1345 
1346 /*
1347  * Free data allocated by namei(); see namei(9) for details.
1348  */
1349 void
1350 NDFREE(struct nameidata *ndp, const u_int flags)
1351 {
1352 	int unlock_dvp;
1353 	int unlock_vp;
1354 
1355 	unlock_dvp = 0;
1356 	unlock_vp = 0;
1357 
1358 	if (!(flags & NDF_NO_FREE_PNBUF) &&
1359 	    (ndp->ni_cnd.cn_flags & HASBUF)) {
1360 		uma_zfree(namei_zone, ndp->ni_cnd.cn_pnbuf);
1361 		ndp->ni_cnd.cn_flags &= ~HASBUF;
1362 	}
1363 	if (!(flags & NDF_NO_VP_UNLOCK) &&
1364 	    (ndp->ni_cnd.cn_flags & LOCKLEAF) && ndp->ni_vp)
1365 		unlock_vp = 1;
1366 	if (!(flags & NDF_NO_DVP_UNLOCK) &&
1367 	    (ndp->ni_cnd.cn_flags & LOCKPARENT) &&
1368 	    ndp->ni_dvp != ndp->ni_vp)
1369 		unlock_dvp = 1;
1370 	if (!(flags & NDF_NO_VP_RELE) && ndp->ni_vp) {
1371 		if (unlock_vp) {
1372 			vput(ndp->ni_vp);
1373 			unlock_vp = 0;
1374 		} else
1375 			vrele(ndp->ni_vp);
1376 		ndp->ni_vp = NULL;
1377 	}
1378 	if (unlock_vp)
1379 		VOP_UNLOCK(ndp->ni_vp);
1380 	if (!(flags & NDF_NO_DVP_RELE) &&
1381 	    (ndp->ni_cnd.cn_flags & (LOCKPARENT|WANTPARENT))) {
1382 		if (unlock_dvp) {
1383 			vput(ndp->ni_dvp);
1384 			unlock_dvp = 0;
1385 		} else
1386 			vrele(ndp->ni_dvp);
1387 		ndp->ni_dvp = NULL;
1388 	}
1389 	if (unlock_dvp)
1390 		VOP_UNLOCK(ndp->ni_dvp);
1391 	if (!(flags & NDF_NO_STARTDIR_RELE) &&
1392 	    (ndp->ni_cnd.cn_flags & SAVESTART)) {
1393 		vrele(ndp->ni_startdir);
1394 		ndp->ni_startdir = NULL;
1395 	}
1396 }
1397 
1398 /*
1399  * Determine if there is a suitable alternate filename under the specified
1400  * prefix for the specified path.  If the create flag is set, then the
1401  * alternate prefix will be used so long as the parent directory exists.
1402  * This is used by the various compatibility ABIs so that Linux binaries prefer
1403  * files under /compat/linux for example.  The chosen path (whether under
1404  * the prefix or under /) is returned in a kernel malloc'd buffer pointed
1405  * to by pathbuf.  The caller is responsible for free'ing the buffer from
1406  * the M_TEMP bucket if one is returned.
1407  */
1408 int
1409 kern_alternate_path(struct thread *td, const char *prefix, const char *path,
1410     enum uio_seg pathseg, char **pathbuf, int create, int dirfd)
1411 {
1412 	struct nameidata nd, ndroot;
1413 	char *ptr, *buf, *cp;
1414 	size_t len, sz;
1415 	int error;
1416 
1417 	buf = (char *) malloc(MAXPATHLEN, M_TEMP, M_WAITOK);
1418 	*pathbuf = buf;
1419 
1420 	/* Copy the prefix into the new pathname as a starting point. */
1421 	len = strlcpy(buf, prefix, MAXPATHLEN);
1422 	if (len >= MAXPATHLEN) {
1423 		*pathbuf = NULL;
1424 		free(buf, M_TEMP);
1425 		return (EINVAL);
1426 	}
1427 	sz = MAXPATHLEN - len;
1428 	ptr = buf + len;
1429 
1430 	/* Append the filename to the prefix. */
1431 	if (pathseg == UIO_SYSSPACE)
1432 		error = copystr(path, ptr, sz, &len);
1433 	else
1434 		error = copyinstr(path, ptr, sz, &len);
1435 
1436 	if (error) {
1437 		*pathbuf = NULL;
1438 		free(buf, M_TEMP);
1439 		return (error);
1440 	}
1441 
1442 	/* Only use a prefix with absolute pathnames. */
1443 	if (*ptr != '/') {
1444 		error = EINVAL;
1445 		goto keeporig;
1446 	}
1447 
1448 	if (dirfd != AT_FDCWD) {
1449 		/*
1450 		 * We want the original because the "prefix" is
1451 		 * included in the already opened dirfd.
1452 		 */
1453 		bcopy(ptr, buf, len);
1454 		return (0);
1455 	}
1456 
1457 	/*
1458 	 * We know that there is a / somewhere in this pathname.
1459 	 * Search backwards for it, to find the file's parent dir
1460 	 * to see if it exists in the alternate tree. If it does,
1461 	 * and we want to create a file (cflag is set). We don't
1462 	 * need to worry about the root comparison in this case.
1463 	 */
1464 
1465 	if (create) {
1466 		for (cp = &ptr[len] - 1; *cp != '/'; cp--);
1467 		*cp = '\0';
1468 
1469 		NDINIT(&nd, LOOKUP, NOFOLLOW, UIO_SYSSPACE, buf, td);
1470 		error = namei(&nd);
1471 		*cp = '/';
1472 		if (error != 0)
1473 			goto keeporig;
1474 	} else {
1475 		NDINIT(&nd, LOOKUP, NOFOLLOW, UIO_SYSSPACE, buf, td);
1476 
1477 		error = namei(&nd);
1478 		if (error != 0)
1479 			goto keeporig;
1480 
1481 		/*
1482 		 * We now compare the vnode of the prefix to the one
1483 		 * vnode asked. If they resolve to be the same, then we
1484 		 * ignore the match so that the real root gets used.
1485 		 * This avoids the problem of traversing "../.." to find the
1486 		 * root directory and never finding it, because "/" resolves
1487 		 * to the emulation root directory. This is expensive :-(
1488 		 */
1489 		NDINIT(&ndroot, LOOKUP, FOLLOW, UIO_SYSSPACE, prefix,
1490 		    td);
1491 
1492 		/* We shouldn't ever get an error from this namei(). */
1493 		error = namei(&ndroot);
1494 		if (error == 0) {
1495 			if (nd.ni_vp == ndroot.ni_vp)
1496 				error = ENOENT;
1497 
1498 			NDFREE(&ndroot, NDF_ONLY_PNBUF);
1499 			vrele(ndroot.ni_vp);
1500 		}
1501 	}
1502 
1503 	NDFREE(&nd, NDF_ONLY_PNBUF);
1504 	vrele(nd.ni_vp);
1505 
1506 keeporig:
1507 	/* If there was an error, use the original path name. */
1508 	if (error)
1509 		bcopy(ptr, buf, len);
1510 	return (error);
1511 }
1512