xref: /freebsd/sys/kern/vfs_lookup.c (revision 82397d791966b09d344251bc709cd9db2b3a1902)
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 static int
468 namei_getpath(struct nameidata *ndp)
469 {
470 	struct componentname *cnp;
471 	int error;
472 
473 	cnp = &ndp->ni_cnd;
474 
475 	/*
476 	 * Get a buffer for the name to be translated, and copy the
477 	 * name into the buffer.
478 	 */
479 	cnp->cn_pnbuf = uma_zalloc(namei_zone, M_WAITOK);
480 	if (ndp->ni_segflg == UIO_SYSSPACE) {
481 		error = copystr(ndp->ni_dirp, cnp->cn_pnbuf, MAXPATHLEN,
482 		    &ndp->ni_pathlen);
483 	} else {
484 		error = copyinstr(ndp->ni_dirp, cnp->cn_pnbuf, MAXPATHLEN,
485 		    &ndp->ni_pathlen);
486 	}
487 
488 	if (__predict_false(error != 0)) {
489 		namei_cleanup_cnp(cnp);
490 		return (error);
491 	}
492 
493 	/*
494 	 * Don't allow empty pathnames.
495 	 */
496 	if (__predict_false(*cnp->cn_pnbuf == '\0')) {
497 		namei_cleanup_cnp(cnp);
498 		return (ENOENT);
499 	}
500 
501 	cnp->cn_nameptr = cnp->cn_pnbuf;
502 	return (0);
503 }
504 
505 /*
506  * Convert a pathname into a pointer to a locked vnode.
507  *
508  * The FOLLOW flag is set when symbolic links are to be followed
509  * when they occur at the end of the name translation process.
510  * Symbolic links are always followed for all other pathname
511  * components other than the last.
512  *
513  * The segflg defines whether the name is to be copied from user
514  * space or kernel space.
515  *
516  * Overall outline of namei:
517  *
518  *	copy in name
519  *	get starting directory
520  *	while (!done && !error) {
521  *		call lookup to search path.
522  *		if symbolic link, massage name in buffer and continue
523  *	}
524  */
525 int
526 namei(struct nameidata *ndp)
527 {
528 	char *cp;		/* pointer into pathname argument */
529 	struct vnode *dp;	/* the directory we are searching */
530 	struct iovec aiov;		/* uio for reading symbolic links */
531 	struct componentname *cnp;
532 	struct thread *td;
533 	struct pwd *pwd;
534 	struct uio auio;
535 	int error, linklen;
536 	enum cache_fpl_status status;
537 
538 	cnp = &ndp->ni_cnd;
539 	td = cnp->cn_thread;
540 #ifdef INVARIANTS
541 	KASSERT((ndp->ni_debugflags & NAMEI_DBG_CALLED) == 0,
542 	    ("%s: repeated call to namei without NDREINIT", __func__));
543 	KASSERT(ndp->ni_debugflags == NAMEI_DBG_INITED,
544 	    ("%s: bad debugflags %d", __func__, ndp->ni_debugflags));
545 	ndp->ni_debugflags |= NAMEI_DBG_CALLED;
546 	if (ndp->ni_startdir != NULL)
547 		ndp->ni_debugflags |= NAMEI_DBG_HADSTARTDIR;
548 	if (cnp->cn_flags & FAILIFEXISTS) {
549 		KASSERT(cnp->cn_nameiop == CREATE,
550 		    ("%s: FAILIFEXISTS passed for op %d", __func__, cnp->cn_nameiop));
551 		/*
552 		 * The limitation below is to restrict hairy corner cases.
553 		 */
554 		KASSERT((cnp->cn_flags & (LOCKPARENT | LOCKLEAF)) == LOCKPARENT,
555 		    ("%s: FAILIFEXISTS must be passed with LOCKPARENT and without LOCKLEAF",
556 		    __func__));
557 	}
558 	/*
559 	 * For NDVALIDATE.
560 	 *
561 	 * While NDINIT may seem like a more natural place to do it, there are
562 	 * callers which directly modify flags past invoking init.
563 	 */
564 	cnp->cn_origflags = cnp->cn_flags;
565 #endif
566 	ndp->ni_cnd.cn_cred = ndp->ni_cnd.cn_thread->td_ucred;
567 	KASSERT(ndp->ni_resflags == 0, ("%s: garbage in ni_resflags: %x\n",
568 	    __func__, ndp->ni_resflags));
569 	KASSERT(cnp->cn_cred && td->td_proc, ("namei: bad cred/proc"));
570 	KASSERT((cnp->cn_flags & NAMEI_INTERNAL_FLAGS) == 0,
571 	    ("namei: unexpected flags: %" PRIx64 "\n",
572 	    cnp->cn_flags & NAMEI_INTERNAL_FLAGS));
573 	if (cnp->cn_flags & NOCACHE)
574 		KASSERT(cnp->cn_nameiop != LOOKUP,
575 		    ("%s: NOCACHE passed with LOOKUP", __func__));
576 	MPASS(ndp->ni_startdir == NULL || ndp->ni_startdir->v_type == VDIR ||
577 	    ndp->ni_startdir->v_type == VBAD);
578 
579 	ndp->ni_lcf = 0;
580 	ndp->ni_vp = NULL;
581 
582 	error = namei_getpath(ndp);
583 	if (__predict_false(error != 0)) {
584 		return (error);
585 	}
586 
587 #ifdef KTRACE
588 	if (KTRPOINT(td, KTR_NAMEI)) {
589 		KASSERT(cnp->cn_thread == curthread,
590 		    ("namei not using curthread"));
591 		ktrnamei(cnp->cn_pnbuf);
592 	}
593 #endif
594 
595 	/*
596 	 * First try looking up the target without locking any vnodes.
597 	 *
598 	 * We may need to start from scratch or pick up where it left off.
599 	 */
600 	error = cache_fplookup(ndp, &status, &pwd);
601 	switch (status) {
602 	case CACHE_FPL_STATUS_UNSET:
603 		__assert_unreachable();
604 		break;
605 	case CACHE_FPL_STATUS_HANDLED:
606 		if (error == 0)
607 			NDVALIDATE(ndp);
608 		return (error);
609 	case CACHE_FPL_STATUS_PARTIAL:
610 		TAILQ_INIT(&ndp->ni_cap_tracker);
611 		dp = ndp->ni_startdir;
612 		break;
613 	case CACHE_FPL_STATUS_ABORTED:
614 		TAILQ_INIT(&ndp->ni_cap_tracker);
615 		error = namei_setup(ndp, &dp, &pwd);
616 		if (error != 0) {
617 			namei_cleanup_cnp(cnp);
618 			return (error);
619 		}
620 		break;
621 	}
622 
623 	ndp->ni_loopcnt = 0;
624 
625 	/*
626 	 * Locked lookup.
627 	 */
628 	for (;;) {
629 		ndp->ni_startdir = dp;
630 		error = lookup(ndp);
631 		if (error != 0) {
632 			/*
633 			 * Override an error to not allow user to use
634 			 * BENEATH as an oracle.
635 			 */
636 			if ((ndp->ni_lcf & (NI_LCF_LATCH |
637 			    NI_LCF_BENEATH_LATCHED)) == NI_LCF_LATCH)
638 				error = ENOTCAPABLE;
639 			goto out;
640 		}
641 
642 		/*
643 		 * If not a symbolic link, we're done.
644 		 */
645 		if ((cnp->cn_flags & ISSYMLINK) == 0) {
646 			if ((cnp->cn_flags & (SAVENAME | SAVESTART)) == 0) {
647 				namei_cleanup_cnp(cnp);
648 			} else
649 				cnp->cn_flags |= HASBUF;
650 			if ((ndp->ni_lcf & (NI_LCF_LATCH |
651 			    NI_LCF_BENEATH_LATCHED)) == NI_LCF_LATCH) {
652 				NDFREE(ndp, 0);
653 				error = ENOTCAPABLE;
654 			}
655 			nameicap_cleanup(ndp, true);
656 			SDT_PROBE3(vfs, namei, lookup, return, error,
657 			    (error == 0 ? ndp->ni_vp : NULL), false);
658 			pwd_drop(pwd);
659 			if (error == 0)
660 				NDVALIDATE(ndp);
661 			return (error);
662 		}
663 		if (ndp->ni_loopcnt++ >= MAXSYMLINKS) {
664 			error = ELOOP;
665 			break;
666 		}
667 #ifdef MAC
668 		if ((cnp->cn_flags & NOMACCHECK) == 0) {
669 			error = mac_vnode_check_readlink(td->td_ucred,
670 			    ndp->ni_vp);
671 			if (error != 0)
672 				break;
673 		}
674 #endif
675 		if (ndp->ni_pathlen > 1)
676 			cp = uma_zalloc(namei_zone, M_WAITOK);
677 		else
678 			cp = cnp->cn_pnbuf;
679 		aiov.iov_base = cp;
680 		aiov.iov_len = MAXPATHLEN;
681 		auio.uio_iov = &aiov;
682 		auio.uio_iovcnt = 1;
683 		auio.uio_offset = 0;
684 		auio.uio_rw = UIO_READ;
685 		auio.uio_segflg = UIO_SYSSPACE;
686 		auio.uio_td = td;
687 		auio.uio_resid = MAXPATHLEN;
688 		error = VOP_READLINK(ndp->ni_vp, &auio, cnp->cn_cred);
689 		if (error != 0) {
690 			if (ndp->ni_pathlen > 1)
691 				uma_zfree(namei_zone, cp);
692 			break;
693 		}
694 		linklen = MAXPATHLEN - auio.uio_resid;
695 		if (linklen == 0) {
696 			if (ndp->ni_pathlen > 1)
697 				uma_zfree(namei_zone, cp);
698 			error = ENOENT;
699 			break;
700 		}
701 		if (linklen + ndp->ni_pathlen > MAXPATHLEN) {
702 			if (ndp->ni_pathlen > 1)
703 				uma_zfree(namei_zone, cp);
704 			error = ENAMETOOLONG;
705 			break;
706 		}
707 		if (ndp->ni_pathlen > 1) {
708 			bcopy(ndp->ni_next, cp + linklen, ndp->ni_pathlen);
709 			uma_zfree(namei_zone, cnp->cn_pnbuf);
710 			cnp->cn_pnbuf = cp;
711 		} else
712 			cnp->cn_pnbuf[linklen] = '\0';
713 		ndp->ni_pathlen += linklen;
714 		vput(ndp->ni_vp);
715 		dp = ndp->ni_dvp;
716 		/*
717 		 * Check if root directory should replace current directory.
718 		 */
719 		cnp->cn_nameptr = cnp->cn_pnbuf;
720 		if (*(cnp->cn_nameptr) == '/') {
721 			vrele(dp);
722 			error = namei_handle_root(ndp, &dp);
723 			if (error != 0)
724 				goto out;
725 		}
726 	}
727 	vput(ndp->ni_vp);
728 	ndp->ni_vp = NULL;
729 	vrele(ndp->ni_dvp);
730 out:
731 	MPASS(error != 0);
732 	namei_cleanup_cnp(cnp);
733 	nameicap_cleanup(ndp, true);
734 	SDT_PROBE3(vfs, namei, lookup, return, error, NULL, false);
735 	pwd_drop(pwd);
736 	return (error);
737 }
738 
739 static int
740 compute_cn_lkflags(struct mount *mp, int lkflags, int cnflags)
741 {
742 
743 	if (mp == NULL || ((lkflags & LK_SHARED) &&
744 	    (!(mp->mnt_kern_flag & MNTK_LOOKUP_SHARED) ||
745 	    ((cnflags & ISDOTDOT) &&
746 	    (mp->mnt_kern_flag & MNTK_LOOKUP_EXCL_DOTDOT))))) {
747 		lkflags &= ~LK_SHARED;
748 		lkflags |= LK_EXCLUSIVE;
749 	}
750 	lkflags |= LK_NODDLKTREAT;
751 	return (lkflags);
752 }
753 
754 static __inline int
755 needs_exclusive_leaf(struct mount *mp, int flags)
756 {
757 
758 	/*
759 	 * Intermediate nodes can use shared locks, we only need to
760 	 * force an exclusive lock for leaf nodes.
761 	 */
762 	if ((flags & (ISLASTCN | LOCKLEAF)) != (ISLASTCN | LOCKLEAF))
763 		return (0);
764 
765 	/* Always use exclusive locks if LOCKSHARED isn't set. */
766 	if (!(flags & LOCKSHARED))
767 		return (1);
768 
769 	/*
770 	 * For lookups during open(), if the mount point supports
771 	 * extended shared operations, then use a shared lock for the
772 	 * leaf node, otherwise use an exclusive lock.
773 	 */
774 	if ((flags & ISOPEN) != 0)
775 		return (!MNT_EXTENDED_SHARED(mp));
776 
777 	/*
778 	 * Lookup requests outside of open() that specify LOCKSHARED
779 	 * only need a shared lock on the leaf vnode.
780 	 */
781 	return (0);
782 }
783 
784 /*
785  * Search a pathname.
786  * This is a very central and rather complicated routine.
787  *
788  * The pathname is pointed to by ni_ptr and is of length ni_pathlen.
789  * The starting directory is taken from ni_startdir. The pathname is
790  * descended until done, or a symbolic link is encountered. The variable
791  * ni_more is clear if the path is completed; it is set to one if a
792  * symbolic link needing interpretation is encountered.
793  *
794  * The flag argument is LOOKUP, CREATE, RENAME, or DELETE depending on
795  * whether the name is to be looked up, created, renamed, or deleted.
796  * When CREATE, RENAME, or DELETE is specified, information usable in
797  * creating, renaming, or deleting a directory entry may be calculated.
798  * If flag has LOCKPARENT or'ed into it, the parent directory is returned
799  * locked. If flag has WANTPARENT or'ed into it, the parent directory is
800  * returned unlocked. Otherwise the parent directory is not returned. If
801  * the target of the pathname exists and LOCKLEAF is or'ed into the flag
802  * the target is returned locked, otherwise it is returned unlocked.
803  * When creating or renaming and LOCKPARENT is specified, the target may not
804  * be ".".  When deleting and LOCKPARENT is specified, the target may be ".".
805  *
806  * Overall outline of lookup:
807  *
808  * dirloop:
809  *	identify next component of name at ndp->ni_ptr
810  *	handle degenerate case where name is null string
811  *	if .. and crossing mount points and on mounted filesys, find parent
812  *	call VOP_LOOKUP routine for next component name
813  *	    directory vnode returned in ni_dvp, unlocked unless LOCKPARENT set
814  *	    component vnode returned in ni_vp (if it exists), locked.
815  *	if result vnode is mounted on and crossing mount points,
816  *	    find mounted on vnode
817  *	if more components of name, do next level at dirloop
818  *	return the answer in ni_vp, locked if LOCKLEAF set
819  *	    if LOCKPARENT set, return locked parent in ni_dvp
820  *	    if WANTPARENT set, return unlocked parent in ni_dvp
821  */
822 int
823 lookup(struct nameidata *ndp)
824 {
825 	char *cp;			/* pointer into pathname argument */
826 	char *prev_ni_next;		/* saved ndp->ni_next */
827 	struct vnode *dp = NULL;	/* the directory we are searching */
828 	struct vnode *tdp;		/* saved dp */
829 	struct mount *mp;		/* mount table entry */
830 	struct prison *pr;
831 	size_t prev_ni_pathlen;		/* saved ndp->ni_pathlen */
832 	int docache;			/* == 0 do not cache last component */
833 	int wantparent;			/* 1 => wantparent or lockparent flag */
834 	int rdonly;			/* lookup read-only flag bit */
835 	int error = 0;
836 	int dpunlocked = 0;		/* dp has already been unlocked */
837 	int relookup = 0;		/* do not consume the path component */
838 	struct componentname *cnp = &ndp->ni_cnd;
839 	int lkflags_save;
840 	int ni_dvp_unlocked;
841 
842 	/*
843 	 * Setup: break out flag bits into variables.
844 	 */
845 	ni_dvp_unlocked = 0;
846 	wantparent = cnp->cn_flags & (LOCKPARENT | WANTPARENT);
847 	KASSERT(cnp->cn_nameiop == LOOKUP || wantparent,
848 	    ("CREATE, DELETE, RENAME require LOCKPARENT or WANTPARENT."));
849 	/*
850 	 * When set to zero, docache causes the last component of the
851 	 * pathname to be deleted from the cache and the full lookup
852 	 * of the name to be done (via VOP_CACHEDLOOKUP()). Often
853 	 * filesystems need some pre-computed values that are made
854 	 * during the full lookup, for instance UFS sets dp->i_offset.
855 	 *
856 	 * The docache variable is set to zero when requested by the
857 	 * NOCACHE flag and for all modifying operations except CREATE.
858 	 */
859 	docache = (cnp->cn_flags & NOCACHE) ^ NOCACHE;
860 	if (cnp->cn_nameiop == DELETE ||
861 	    (wantparent && cnp->cn_nameiop != CREATE &&
862 	     cnp->cn_nameiop != LOOKUP))
863 		docache = 0;
864 	rdonly = cnp->cn_flags & RDONLY;
865 	cnp->cn_flags &= ~ISSYMLINK;
866 	ndp->ni_dvp = NULL;
867 	/*
868 	 * We use shared locks until we hit the parent of the last cn then
869 	 * we adjust based on the requesting flags.
870 	 */
871 	cnp->cn_lkflags = LK_SHARED;
872 	dp = ndp->ni_startdir;
873 	ndp->ni_startdir = NULLVP;
874 	vn_lock(dp,
875 	    compute_cn_lkflags(dp->v_mount, cnp->cn_lkflags | LK_RETRY,
876 	    cnp->cn_flags));
877 
878 dirloop:
879 	/*
880 	 * Search a new directory.
881 	 *
882 	 * The last component of the filename is left accessible via
883 	 * cnp->cn_nameptr for callers that need the name. Callers needing
884 	 * the name set the SAVENAME flag. When done, they assume
885 	 * responsibility for freeing the pathname buffer.
886 	 */
887 	for (cp = cnp->cn_nameptr; *cp != 0 && *cp != '/'; cp++)
888 		continue;
889 	cnp->cn_namelen = cp - cnp->cn_nameptr;
890 	if (cnp->cn_namelen > NAME_MAX) {
891 		error = ENAMETOOLONG;
892 		goto bad;
893 	}
894 #ifdef NAMEI_DIAGNOSTIC
895 	{ char c = *cp;
896 	*cp = '\0';
897 	printf("{%s}: ", cnp->cn_nameptr);
898 	*cp = c; }
899 #endif
900 	prev_ni_pathlen = ndp->ni_pathlen;
901 	ndp->ni_pathlen -= cnp->cn_namelen;
902 	KASSERT(ndp->ni_pathlen <= PATH_MAX,
903 	    ("%s: ni_pathlen underflow to %zd\n", __func__, ndp->ni_pathlen));
904 	prev_ni_next = ndp->ni_next;
905 	ndp->ni_next = cp;
906 
907 	/*
908 	 * Replace multiple slashes by a single slash and trailing slashes
909 	 * by a null.  This must be done before VOP_LOOKUP() because some
910 	 * fs's don't know about trailing slashes.  Remember if there were
911 	 * trailing slashes to handle symlinks, existing non-directories
912 	 * and non-existing files that won't be directories specially later.
913 	 */
914 	while (*cp == '/' && (cp[1] == '/' || cp[1] == '\0')) {
915 		cp++;
916 		ndp->ni_pathlen--;
917 		if (*cp == '\0') {
918 			*ndp->ni_next = '\0';
919 			cnp->cn_flags |= TRAILINGSLASH;
920 		}
921 	}
922 	ndp->ni_next = cp;
923 
924 	cnp->cn_flags |= MAKEENTRY;
925 	if (*cp == '\0' && docache == 0)
926 		cnp->cn_flags &= ~MAKEENTRY;
927 	if (cnp->cn_namelen == 2 &&
928 	    cnp->cn_nameptr[1] == '.' && cnp->cn_nameptr[0] == '.')
929 		cnp->cn_flags |= ISDOTDOT;
930 	else
931 		cnp->cn_flags &= ~ISDOTDOT;
932 	if (*ndp->ni_next == 0)
933 		cnp->cn_flags |= ISLASTCN;
934 	else
935 		cnp->cn_flags &= ~ISLASTCN;
936 
937 	if ((cnp->cn_flags & ISLASTCN) != 0 &&
938 	    cnp->cn_namelen == 1 && cnp->cn_nameptr[0] == '.' &&
939 	    (cnp->cn_nameiop == DELETE || cnp->cn_nameiop == RENAME)) {
940 		error = EINVAL;
941 		goto bad;
942 	}
943 
944 	nameicap_tracker_add(ndp, dp);
945 
946 	/*
947 	 * Check for degenerate name (e.g. / or "")
948 	 * which is a way of talking about a directory,
949 	 * e.g. like "/." or ".".
950 	 */
951 	if (cnp->cn_nameptr[0] == '\0') {
952 		if (dp->v_type != VDIR) {
953 			error = ENOTDIR;
954 			goto bad;
955 		}
956 		if (cnp->cn_nameiop != LOOKUP) {
957 			error = EISDIR;
958 			goto bad;
959 		}
960 		if (wantparent) {
961 			ndp->ni_dvp = dp;
962 			VREF(dp);
963 		}
964 		ndp->ni_vp = dp;
965 
966 		if (cnp->cn_flags & AUDITVNODE1)
967 			AUDIT_ARG_VNODE1(dp);
968 		else if (cnp->cn_flags & AUDITVNODE2)
969 			AUDIT_ARG_VNODE2(dp);
970 
971 		if (!(cnp->cn_flags & (LOCKPARENT | LOCKLEAF)))
972 			VOP_UNLOCK(dp);
973 		/* XXX This should probably move to the top of function. */
974 		if (cnp->cn_flags & SAVESTART)
975 			panic("lookup: SAVESTART");
976 		goto success;
977 	}
978 
979 	/*
980 	 * Handle "..": five special cases.
981 	 * 0. If doing a capability lookup and lookup_cap_dotdot is
982 	 *    disabled, return ENOTCAPABLE.
983 	 * 1. Return an error if this is the last component of
984 	 *    the name and the operation is DELETE or RENAME.
985 	 * 2. If at root directory (e.g. after chroot)
986 	 *    or at absolute root directory
987 	 *    then ignore it so can't get out.
988 	 * 3. If this vnode is the root of a mounted
989 	 *    filesystem, then replace it with the
990 	 *    vnode which was mounted on so we take the
991 	 *    .. in the other filesystem.
992 	 * 4. If the vnode is the top directory of
993 	 *    the jail or chroot, don't let them out.
994 	 * 5. If doing a capability lookup and lookup_cap_dotdot is
995 	 *    enabled, return ENOTCAPABLE if the lookup would escape
996 	 *    from the initial file descriptor directory.  Checks are
997 	 *    done by ensuring that namei() already traversed the
998 	 *    result of dotdot lookup.
999 	 */
1000 	if (cnp->cn_flags & ISDOTDOT) {
1001 		if ((ndp->ni_lcf & (NI_LCF_STRICTRELATIVE | NI_LCF_CAP_DOTDOT))
1002 		    == NI_LCF_STRICTRELATIVE) {
1003 #ifdef KTRACE
1004 			if (KTRPOINT(curthread, KTR_CAPFAIL))
1005 				ktrcapfail(CAPFAIL_LOOKUP, NULL, NULL);
1006 #endif
1007 			error = ENOTCAPABLE;
1008 			goto bad;
1009 		}
1010 		if ((cnp->cn_flags & ISLASTCN) != 0 &&
1011 		    (cnp->cn_nameiop == DELETE || cnp->cn_nameiop == RENAME)) {
1012 			error = EINVAL;
1013 			goto bad;
1014 		}
1015 		for (;;) {
1016 			for (pr = cnp->cn_cred->cr_prison; pr != NULL;
1017 			     pr = pr->pr_parent)
1018 				if (dp == pr->pr_root)
1019 					break;
1020 			if (dp == ndp->ni_rootdir ||
1021 			    dp == ndp->ni_topdir ||
1022 			    dp == rootvnode ||
1023 			    pr != NULL ||
1024 			    ((dp->v_vflag & VV_ROOT) != 0 &&
1025 			     (cnp->cn_flags & NOCROSSMOUNT) != 0)) {
1026 				ndp->ni_dvp = dp;
1027 				ndp->ni_vp = dp;
1028 				VREF(dp);
1029 				goto nextname;
1030 			}
1031 			if ((dp->v_vflag & VV_ROOT) == 0)
1032 				break;
1033 			if (VN_IS_DOOMED(dp)) {	/* forced unmount */
1034 				error = ENOENT;
1035 				goto bad;
1036 			}
1037 			tdp = dp;
1038 			dp = dp->v_mount->mnt_vnodecovered;
1039 			VREF(dp);
1040 			vput(tdp);
1041 			vn_lock(dp,
1042 			    compute_cn_lkflags(dp->v_mount, cnp->cn_lkflags |
1043 			    LK_RETRY, ISDOTDOT));
1044 			error = nameicap_check_dotdot(ndp, dp);
1045 			if (error != 0) {
1046 #ifdef KTRACE
1047 				if (KTRPOINT(curthread, KTR_CAPFAIL))
1048 					ktrcapfail(CAPFAIL_LOOKUP, NULL, NULL);
1049 #endif
1050 				goto bad;
1051 			}
1052 		}
1053 	}
1054 
1055 	/*
1056 	 * We now have a segment name to search for, and a directory to search.
1057 	 */
1058 unionlookup:
1059 #ifdef MAC
1060 	error = mac_vnode_check_lookup(cnp->cn_thread->td_ucred, dp, cnp);
1061 	if (error)
1062 		goto bad;
1063 #endif
1064 	ndp->ni_dvp = dp;
1065 	ndp->ni_vp = NULL;
1066 	ASSERT_VOP_LOCKED(dp, "lookup");
1067 	/*
1068 	 * If we have a shared lock we may need to upgrade the lock for the
1069 	 * last operation.
1070 	 */
1071 	if ((cnp->cn_flags & LOCKPARENT) && (cnp->cn_flags & ISLASTCN) &&
1072 	    dp != vp_crossmp && VOP_ISLOCKED(dp) == LK_SHARED)
1073 		vn_lock(dp, LK_UPGRADE|LK_RETRY);
1074 	if (VN_IS_DOOMED(dp)) {
1075 		error = ENOENT;
1076 		goto bad;
1077 	}
1078 	/*
1079 	 * If we're looking up the last component and we need an exclusive
1080 	 * lock, adjust our lkflags.
1081 	 */
1082 	if (needs_exclusive_leaf(dp->v_mount, cnp->cn_flags))
1083 		cnp->cn_lkflags = LK_EXCLUSIVE;
1084 #ifdef NAMEI_DIAGNOSTIC
1085 	vn_printf(dp, "lookup in ");
1086 #endif
1087 	lkflags_save = cnp->cn_lkflags;
1088 	cnp->cn_lkflags = compute_cn_lkflags(dp->v_mount, cnp->cn_lkflags,
1089 	    cnp->cn_flags);
1090 	error = VOP_LOOKUP(dp, &ndp->ni_vp, cnp);
1091 	cnp->cn_lkflags = lkflags_save;
1092 	if (error != 0) {
1093 		KASSERT(ndp->ni_vp == NULL, ("leaf should be empty"));
1094 #ifdef NAMEI_DIAGNOSTIC
1095 		printf("not found\n");
1096 #endif
1097 		if ((error == ENOENT) &&
1098 		    (dp->v_vflag & VV_ROOT) && (dp->v_mount != NULL) &&
1099 		    (dp->v_mount->mnt_flag & MNT_UNION)) {
1100 			tdp = dp;
1101 			dp = dp->v_mount->mnt_vnodecovered;
1102 			VREF(dp);
1103 			vput(tdp);
1104 			vn_lock(dp,
1105 			    compute_cn_lkflags(dp->v_mount, cnp->cn_lkflags |
1106 			    LK_RETRY, cnp->cn_flags));
1107 			nameicap_tracker_add(ndp, dp);
1108 			goto unionlookup;
1109 		}
1110 
1111 		if (error == ERELOOKUP) {
1112 			vref(dp);
1113 			ndp->ni_vp = dp;
1114 			error = 0;
1115 			relookup = 1;
1116 			goto good;
1117 		}
1118 
1119 		if (error != EJUSTRETURN)
1120 			goto bad;
1121 		/*
1122 		 * At this point, we know we're at the end of the
1123 		 * pathname.  If creating / renaming, we can consider
1124 		 * allowing the file or directory to be created / renamed,
1125 		 * provided we're not on a read-only filesystem.
1126 		 */
1127 		if (rdonly) {
1128 			error = EROFS;
1129 			goto bad;
1130 		}
1131 		/* trailing slash only allowed for directories */
1132 		if ((cnp->cn_flags & TRAILINGSLASH) &&
1133 		    !(cnp->cn_flags & WILLBEDIR)) {
1134 			error = ENOENT;
1135 			goto bad;
1136 		}
1137 		if ((cnp->cn_flags & LOCKPARENT) == 0)
1138 			VOP_UNLOCK(dp);
1139 		/*
1140 		 * We return with ni_vp NULL to indicate that the entry
1141 		 * doesn't currently exist, leaving a pointer to the
1142 		 * (possibly locked) directory vnode in ndp->ni_dvp.
1143 		 */
1144 		if (cnp->cn_flags & SAVESTART) {
1145 			ndp->ni_startdir = ndp->ni_dvp;
1146 			VREF(ndp->ni_startdir);
1147 		}
1148 		goto success;
1149 	}
1150 
1151 good:
1152 #ifdef NAMEI_DIAGNOSTIC
1153 	printf("found\n");
1154 #endif
1155 	dp = ndp->ni_vp;
1156 
1157 	/*
1158 	 * Check to see if the vnode has been mounted on;
1159 	 * if so find the root of the mounted filesystem.
1160 	 */
1161 	while (dp->v_type == VDIR && (mp = dp->v_mountedhere) &&
1162 	       (cnp->cn_flags & NOCROSSMOUNT) == 0) {
1163 		if (vfs_busy(mp, 0))
1164 			continue;
1165 		vput(dp);
1166 		if (dp != ndp->ni_dvp)
1167 			vput(ndp->ni_dvp);
1168 		else
1169 			vrele(ndp->ni_dvp);
1170 		vrefact(vp_crossmp);
1171 		ndp->ni_dvp = vp_crossmp;
1172 		error = VFS_ROOT(mp, compute_cn_lkflags(mp, cnp->cn_lkflags,
1173 		    cnp->cn_flags), &tdp);
1174 		vfs_unbusy(mp);
1175 		if (vn_lock(vp_crossmp, LK_SHARED | LK_NOWAIT))
1176 			panic("vp_crossmp exclusively locked or reclaimed");
1177 		if (error) {
1178 			dpunlocked = 1;
1179 			goto bad2;
1180 		}
1181 		ndp->ni_vp = dp = tdp;
1182 	}
1183 
1184 	/*
1185 	 * Check for symbolic link
1186 	 */
1187 	if ((dp->v_type == VLNK) &&
1188 	    ((cnp->cn_flags & FOLLOW) || (cnp->cn_flags & TRAILINGSLASH) ||
1189 	     *ndp->ni_next == '/')) {
1190 		cnp->cn_flags |= ISSYMLINK;
1191 		if (VN_IS_DOOMED(dp)) {
1192 			/*
1193 			 * We can't know whether the directory was mounted with
1194 			 * NOSYMFOLLOW, so we can't follow safely.
1195 			 */
1196 			error = ENOENT;
1197 			goto bad2;
1198 		}
1199 		if (dp->v_mount->mnt_flag & MNT_NOSYMFOLLOW) {
1200 			error = EACCES;
1201 			goto bad2;
1202 		}
1203 		/*
1204 		 * Symlink code always expects an unlocked dvp.
1205 		 */
1206 		if (ndp->ni_dvp != ndp->ni_vp) {
1207 			VOP_UNLOCK(ndp->ni_dvp);
1208 			ni_dvp_unlocked = 1;
1209 		}
1210 		goto success;
1211 	}
1212 
1213 nextname:
1214 	/*
1215 	 * Not a symbolic link that we will follow.  Continue with the
1216 	 * next component if there is any; otherwise, we're done.
1217 	 */
1218 	KASSERT((cnp->cn_flags & ISLASTCN) || *ndp->ni_next == '/',
1219 	    ("lookup: invalid path state."));
1220 	if (relookup) {
1221 		relookup = 0;
1222 		ndp->ni_pathlen = prev_ni_pathlen;
1223 		ndp->ni_next = prev_ni_next;
1224 		if (ndp->ni_dvp != dp)
1225 			vput(ndp->ni_dvp);
1226 		else
1227 			vrele(ndp->ni_dvp);
1228 		goto dirloop;
1229 	}
1230 	if (cnp->cn_flags & ISDOTDOT) {
1231 		error = nameicap_check_dotdot(ndp, ndp->ni_vp);
1232 		if (error != 0) {
1233 #ifdef KTRACE
1234 			if (KTRPOINT(curthread, KTR_CAPFAIL))
1235 				ktrcapfail(CAPFAIL_LOOKUP, NULL, NULL);
1236 #endif
1237 			goto bad2;
1238 		}
1239 	}
1240 	if (*ndp->ni_next == '/') {
1241 		cnp->cn_nameptr = ndp->ni_next;
1242 		while (*cnp->cn_nameptr == '/') {
1243 			cnp->cn_nameptr++;
1244 			ndp->ni_pathlen--;
1245 		}
1246 		if (ndp->ni_dvp != dp)
1247 			vput(ndp->ni_dvp);
1248 		else
1249 			vrele(ndp->ni_dvp);
1250 		goto dirloop;
1251 	}
1252 	/*
1253 	 * If we're processing a path with a trailing slash,
1254 	 * check that the end result is a directory.
1255 	 */
1256 	if ((cnp->cn_flags & TRAILINGSLASH) && dp->v_type != VDIR) {
1257 		error = ENOTDIR;
1258 		goto bad2;
1259 	}
1260 	/*
1261 	 * Disallow directory write attempts on read-only filesystems.
1262 	 */
1263 	if (rdonly &&
1264 	    (cnp->cn_nameiop == DELETE || cnp->cn_nameiop == RENAME)) {
1265 		error = EROFS;
1266 		goto bad2;
1267 	}
1268 	if (cnp->cn_flags & SAVESTART) {
1269 		ndp->ni_startdir = ndp->ni_dvp;
1270 		VREF(ndp->ni_startdir);
1271 	}
1272 	if (!wantparent) {
1273 		ni_dvp_unlocked = 2;
1274 		if (ndp->ni_dvp != dp)
1275 			vput(ndp->ni_dvp);
1276 		else
1277 			vrele(ndp->ni_dvp);
1278 	} else if ((cnp->cn_flags & LOCKPARENT) == 0 && ndp->ni_dvp != dp) {
1279 		VOP_UNLOCK(ndp->ni_dvp);
1280 		ni_dvp_unlocked = 1;
1281 	}
1282 
1283 	if (cnp->cn_flags & AUDITVNODE1)
1284 		AUDIT_ARG_VNODE1(dp);
1285 	else if (cnp->cn_flags & AUDITVNODE2)
1286 		AUDIT_ARG_VNODE2(dp);
1287 
1288 	if ((cnp->cn_flags & LOCKLEAF) == 0)
1289 		VOP_UNLOCK(dp);
1290 success:
1291 	/*
1292 	 * FIXME: for lookups which only cross a mount point to fetch the
1293 	 * root vnode, ni_dvp will be set to vp_crossmp. This can be a problem
1294 	 * if either WANTPARENT or LOCKPARENT is set.
1295 	 */
1296 	/*
1297 	 * Because of shared lookup we may have the vnode shared locked, but
1298 	 * the caller may want it to be exclusively locked.
1299 	 */
1300 	if (needs_exclusive_leaf(dp->v_mount, cnp->cn_flags) &&
1301 	    VOP_ISLOCKED(dp) != LK_EXCLUSIVE) {
1302 		vn_lock(dp, LK_UPGRADE | LK_RETRY);
1303 		if (VN_IS_DOOMED(dp)) {
1304 			error = ENOENT;
1305 			goto bad2;
1306 		}
1307 	}
1308 	if (ndp->ni_vp != NULL) {
1309 		nameicap_tracker_add(ndp, ndp->ni_vp);
1310 		if ((cnp->cn_flags & (FAILIFEXISTS | ISSYMLINK)) == FAILIFEXISTS)
1311 			goto bad_eexist;
1312 	}
1313 	return (0);
1314 
1315 bad2:
1316 	if (ni_dvp_unlocked != 2) {
1317 		if (dp != ndp->ni_dvp && !ni_dvp_unlocked)
1318 			vput(ndp->ni_dvp);
1319 		else
1320 			vrele(ndp->ni_dvp);
1321 	}
1322 bad:
1323 	if (!dpunlocked)
1324 		vput(dp);
1325 	ndp->ni_vp = NULL;
1326 	return (error);
1327 bad_eexist:
1328 	/*
1329 	 * FAILIFEXISTS handling.
1330 	 *
1331 	 * XXX namei called with LOCKPARENT but not LOCKLEAF has the strange
1332 	 * behaviour of leaving the vnode unlocked if the target is the same
1333 	 * vnode as the parent.
1334 	 */
1335 	MPASS((cnp->cn_flags & ISSYMLINK) == 0);
1336 	if (ndp->ni_vp == ndp->ni_dvp)
1337 		vrele(ndp->ni_dvp);
1338 	else
1339 		vput(ndp->ni_dvp);
1340 	vrele(ndp->ni_vp);
1341 	ndp->ni_dvp = NULL;
1342 	ndp->ni_vp = NULL;
1343 	NDFREE(ndp, NDF_ONLY_PNBUF);
1344 	return (EEXIST);
1345 }
1346 
1347 /*
1348  * relookup - lookup a path name component
1349  *    Used by lookup to re-acquire things.
1350  */
1351 int
1352 relookup(struct vnode *dvp, struct vnode **vpp, struct componentname *cnp)
1353 {
1354 	struct vnode *dp = NULL;		/* the directory we are searching */
1355 	int rdonly;			/* lookup read-only flag bit */
1356 	int error = 0;
1357 
1358 	KASSERT(cnp->cn_flags & ISLASTCN,
1359 	    ("relookup: Not given last component."));
1360 	/*
1361 	 * Setup: break out flag bits into variables.
1362 	 */
1363 	KASSERT((cnp->cn_flags & (LOCKPARENT | WANTPARENT)) != 0,
1364 	    ("relookup: parent not wanted"));
1365 	rdonly = cnp->cn_flags & RDONLY;
1366 	cnp->cn_flags &= ~ISSYMLINK;
1367 	dp = dvp;
1368 	cnp->cn_lkflags = LK_EXCLUSIVE;
1369 	vn_lock(dp, LK_EXCLUSIVE | LK_RETRY);
1370 
1371 	/*
1372 	 * Search a new directory.
1373 	 *
1374 	 * The last component of the filename is left accessible via
1375 	 * cnp->cn_nameptr for callers that need the name. Callers needing
1376 	 * the name set the SAVENAME flag. When done, they assume
1377 	 * responsibility for freeing the pathname buffer.
1378 	 */
1379 #ifdef NAMEI_DIAGNOSTIC
1380 	printf("{%s}: ", cnp->cn_nameptr);
1381 #endif
1382 
1383 	/*
1384 	 * Check for "" which represents the root directory after slash
1385 	 * removal.
1386 	 */
1387 	if (cnp->cn_nameptr[0] == '\0') {
1388 		/*
1389 		 * Support only LOOKUP for "/" because lookup()
1390 		 * can't succeed for CREATE, DELETE and RENAME.
1391 		 */
1392 		KASSERT(cnp->cn_nameiop == LOOKUP, ("nameiop must be LOOKUP"));
1393 		KASSERT(dp->v_type == VDIR, ("dp is not a directory"));
1394 
1395 		if (!(cnp->cn_flags & LOCKLEAF))
1396 			VOP_UNLOCK(dp);
1397 		*vpp = dp;
1398 		/* XXX This should probably move to the top of function. */
1399 		if (cnp->cn_flags & SAVESTART)
1400 			panic("lookup: SAVESTART");
1401 		return (0);
1402 	}
1403 
1404 	if (cnp->cn_flags & ISDOTDOT)
1405 		panic ("relookup: lookup on dot-dot");
1406 
1407 	/*
1408 	 * We now have a segment name to search for, and a directory to search.
1409 	 */
1410 #ifdef NAMEI_DIAGNOSTIC
1411 	vn_printf(dp, "search in ");
1412 #endif
1413 	if ((error = VOP_LOOKUP(dp, vpp, cnp)) != 0) {
1414 		KASSERT(*vpp == NULL, ("leaf should be empty"));
1415 		if (error != EJUSTRETURN)
1416 			goto bad;
1417 		/*
1418 		 * If creating and at end of pathname, then can consider
1419 		 * allowing file to be created.
1420 		 */
1421 		if (rdonly) {
1422 			error = EROFS;
1423 			goto bad;
1424 		}
1425 		/* ASSERT(dvp == ndp->ni_startdir) */
1426 		if (cnp->cn_flags & SAVESTART)
1427 			VREF(dvp);
1428 		if ((cnp->cn_flags & LOCKPARENT) == 0)
1429 			VOP_UNLOCK(dp);
1430 		/*
1431 		 * We return with ni_vp NULL to indicate that the entry
1432 		 * doesn't currently exist, leaving a pointer to the
1433 		 * (possibly locked) directory vnode in ndp->ni_dvp.
1434 		 */
1435 		return (0);
1436 	}
1437 
1438 	dp = *vpp;
1439 
1440 	/*
1441 	 * Disallow directory write attempts on read-only filesystems.
1442 	 */
1443 	if (rdonly &&
1444 	    (cnp->cn_nameiop == DELETE || cnp->cn_nameiop == RENAME)) {
1445 		if (dvp == dp)
1446 			vrele(dvp);
1447 		else
1448 			vput(dvp);
1449 		error = EROFS;
1450 		goto bad;
1451 	}
1452 	/*
1453 	 * Set the parent lock/ref state to the requested state.
1454 	 */
1455 	if ((cnp->cn_flags & LOCKPARENT) == 0 && dvp != dp)
1456 		VOP_UNLOCK(dvp);
1457 	/*
1458 	 * Check for symbolic link
1459 	 */
1460 	KASSERT(dp->v_type != VLNK || !(cnp->cn_flags & FOLLOW),
1461 	    ("relookup: symlink found.\n"));
1462 
1463 	/* ASSERT(dvp == ndp->ni_startdir) */
1464 	if (cnp->cn_flags & SAVESTART)
1465 		VREF(dvp);
1466 
1467 	if ((cnp->cn_flags & LOCKLEAF) == 0)
1468 		VOP_UNLOCK(dp);
1469 	return (0);
1470 bad:
1471 	vput(dp);
1472 	*vpp = NULL;
1473 	return (error);
1474 }
1475 
1476 /*
1477  * Free data allocated by namei(); see namei(9) for details.
1478  */
1479 void
1480 NDFREE_PNBUF(struct nameidata *ndp)
1481 {
1482 
1483 	if ((ndp->ni_cnd.cn_flags & HASBUF) != 0) {
1484 		MPASS((ndp->ni_cnd.cn_flags & (SAVENAME | SAVESTART)) != 0);
1485 		uma_zfree(namei_zone, ndp->ni_cnd.cn_pnbuf);
1486 		ndp->ni_cnd.cn_flags &= ~HASBUF;
1487 	}
1488 }
1489 
1490 void
1491 (NDFREE)(struct nameidata *ndp, const u_int flags)
1492 {
1493 	int unlock_dvp;
1494 	int unlock_vp;
1495 
1496 	unlock_dvp = 0;
1497 	unlock_vp = 0;
1498 
1499 	if (!(flags & NDF_NO_FREE_PNBUF)) {
1500 		NDFREE_PNBUF(ndp);
1501 	}
1502 	if (!(flags & NDF_NO_VP_UNLOCK) &&
1503 	    (ndp->ni_cnd.cn_flags & LOCKLEAF) && ndp->ni_vp)
1504 		unlock_vp = 1;
1505 	if (!(flags & NDF_NO_DVP_UNLOCK) &&
1506 	    (ndp->ni_cnd.cn_flags & LOCKPARENT) &&
1507 	    ndp->ni_dvp != ndp->ni_vp)
1508 		unlock_dvp = 1;
1509 	if (!(flags & NDF_NO_VP_RELE) && ndp->ni_vp) {
1510 		if (unlock_vp) {
1511 			vput(ndp->ni_vp);
1512 			unlock_vp = 0;
1513 		} else
1514 			vrele(ndp->ni_vp);
1515 		ndp->ni_vp = NULL;
1516 	}
1517 	if (unlock_vp)
1518 		VOP_UNLOCK(ndp->ni_vp);
1519 	if (!(flags & NDF_NO_DVP_RELE) &&
1520 	    (ndp->ni_cnd.cn_flags & (LOCKPARENT|WANTPARENT))) {
1521 		if (unlock_dvp) {
1522 			vput(ndp->ni_dvp);
1523 			unlock_dvp = 0;
1524 		} else
1525 			vrele(ndp->ni_dvp);
1526 		ndp->ni_dvp = NULL;
1527 	}
1528 	if (unlock_dvp)
1529 		VOP_UNLOCK(ndp->ni_dvp);
1530 	if (!(flags & NDF_NO_STARTDIR_RELE) &&
1531 	    (ndp->ni_cnd.cn_flags & SAVESTART)) {
1532 		vrele(ndp->ni_startdir);
1533 		ndp->ni_startdir = NULL;
1534 	}
1535 }
1536 
1537 #ifdef INVARIANTS
1538 /*
1539  * Validate the final state of ndp after the lookup.
1540  *
1541  * Historically filesystems were allowed to modify cn_flags. Most notably they
1542  * can add SAVENAME to the request, resulting in HASBUF and pushing subsequent
1543  * clean up to the consumer. In practice this seems to only concern != LOOKUP
1544  * operations.
1545  *
1546  * As a step towards stricter API contract this routine validates the state to
1547  * clean up. Note validation is a work in progress with the intent of becoming
1548  * stricter over time.
1549  */
1550 #define NDMODIFYINGFLAGS (LOCKLEAF | LOCKPARENT | WANTPARENT | SAVENAME | SAVESTART | HASBUF)
1551 void
1552 NDVALIDATE(struct nameidata *ndp)
1553 {
1554 	struct componentname *cnp;
1555 	u_int64_t used, orig;
1556 
1557 	cnp = &ndp->ni_cnd;
1558 	orig = cnp->cn_origflags;
1559 	used = cnp->cn_flags;
1560 	switch (cnp->cn_nameiop) {
1561 	case LOOKUP:
1562 		/*
1563 		 * For plain lookup we require strict conformance -- nothing
1564 		 * to clean up if it was not requested by the caller.
1565 		 */
1566 		orig &= NDMODIFYINGFLAGS;
1567 		used &= NDMODIFYINGFLAGS;
1568 		if ((orig & (SAVENAME | SAVESTART)) != 0)
1569 			orig |= HASBUF;
1570 		if (orig != used) {
1571 			goto out_mismatch;
1572 		}
1573 		break;
1574 	case CREATE:
1575 	case DELETE:
1576 	case RENAME:
1577 		/*
1578 		 * Some filesystems set SAVENAME to provoke HASBUF, accomodate
1579 		 * for it until it gets fixed.
1580 		 */
1581 		orig &= NDMODIFYINGFLAGS;
1582 		orig |= (SAVENAME | HASBUF);
1583 		used &= NDMODIFYINGFLAGS;
1584 		used |= (SAVENAME | HASBUF);
1585 		if (orig != used) {
1586 			goto out_mismatch;
1587 		}
1588 		break;
1589 	}
1590 	return;
1591 out_mismatch:
1592 	panic("%s: mismatched flags for op %d: added %" PRIx64 ", "
1593 	    "removed %" PRIx64" (%" PRIx64" != %" PRIx64"; stored %" PRIx64" != %" PRIx64")",
1594 	    __func__, cnp->cn_nameiop, used & ~orig, orig &~ used,
1595 	    orig, used, cnp->cn_origflags, cnp->cn_flags);
1596 }
1597 #endif
1598 
1599 /*
1600  * Determine if there is a suitable alternate filename under the specified
1601  * prefix for the specified path.  If the create flag is set, then the
1602  * alternate prefix will be used so long as the parent directory exists.
1603  * This is used by the various compatibility ABIs so that Linux binaries prefer
1604  * files under /compat/linux for example.  The chosen path (whether under
1605  * the prefix or under /) is returned in a kernel malloc'd buffer pointed
1606  * to by pathbuf.  The caller is responsible for free'ing the buffer from
1607  * the M_TEMP bucket if one is returned.
1608  */
1609 int
1610 kern_alternate_path(struct thread *td, const char *prefix, const char *path,
1611     enum uio_seg pathseg, char **pathbuf, int create, int dirfd)
1612 {
1613 	struct nameidata nd, ndroot;
1614 	char *ptr, *buf, *cp;
1615 	size_t len, sz;
1616 	int error;
1617 
1618 	buf = (char *) malloc(MAXPATHLEN, M_TEMP, M_WAITOK);
1619 	*pathbuf = buf;
1620 
1621 	/* Copy the prefix into the new pathname as a starting point. */
1622 	len = strlcpy(buf, prefix, MAXPATHLEN);
1623 	if (len >= MAXPATHLEN) {
1624 		*pathbuf = NULL;
1625 		free(buf, M_TEMP);
1626 		return (EINVAL);
1627 	}
1628 	sz = MAXPATHLEN - len;
1629 	ptr = buf + len;
1630 
1631 	/* Append the filename to the prefix. */
1632 	if (pathseg == UIO_SYSSPACE)
1633 		error = copystr(path, ptr, sz, &len);
1634 	else
1635 		error = copyinstr(path, ptr, sz, &len);
1636 
1637 	if (error) {
1638 		*pathbuf = NULL;
1639 		free(buf, M_TEMP);
1640 		return (error);
1641 	}
1642 
1643 	/* Only use a prefix with absolute pathnames. */
1644 	if (*ptr != '/') {
1645 		error = EINVAL;
1646 		goto keeporig;
1647 	}
1648 
1649 	if (dirfd != AT_FDCWD) {
1650 		/*
1651 		 * We want the original because the "prefix" is
1652 		 * included in the already opened dirfd.
1653 		 */
1654 		bcopy(ptr, buf, len);
1655 		return (0);
1656 	}
1657 
1658 	/*
1659 	 * We know that there is a / somewhere in this pathname.
1660 	 * Search backwards for it, to find the file's parent dir
1661 	 * to see if it exists in the alternate tree. If it does,
1662 	 * and we want to create a file (cflag is set). We don't
1663 	 * need to worry about the root comparison in this case.
1664 	 */
1665 
1666 	if (create) {
1667 		for (cp = &ptr[len] - 1; *cp != '/'; cp--);
1668 		*cp = '\0';
1669 
1670 		NDINIT(&nd, LOOKUP, NOFOLLOW, UIO_SYSSPACE, buf, td);
1671 		error = namei(&nd);
1672 		*cp = '/';
1673 		if (error != 0)
1674 			goto keeporig;
1675 	} else {
1676 		NDINIT(&nd, LOOKUP, NOFOLLOW, UIO_SYSSPACE, buf, td);
1677 
1678 		error = namei(&nd);
1679 		if (error != 0)
1680 			goto keeporig;
1681 
1682 		/*
1683 		 * We now compare the vnode of the prefix to the one
1684 		 * vnode asked. If they resolve to be the same, then we
1685 		 * ignore the match so that the real root gets used.
1686 		 * This avoids the problem of traversing "../.." to find the
1687 		 * root directory and never finding it, because "/" resolves
1688 		 * to the emulation root directory. This is expensive :-(
1689 		 */
1690 		NDINIT(&ndroot, LOOKUP, FOLLOW, UIO_SYSSPACE, prefix,
1691 		    td);
1692 
1693 		/* We shouldn't ever get an error from this namei(). */
1694 		error = namei(&ndroot);
1695 		if (error == 0) {
1696 			if (nd.ni_vp == ndroot.ni_vp)
1697 				error = ENOENT;
1698 
1699 			NDFREE(&ndroot, NDF_ONLY_PNBUF);
1700 			vrele(ndroot.ni_vp);
1701 		}
1702 	}
1703 
1704 	NDFREE(&nd, NDF_ONLY_PNBUF);
1705 	vrele(nd.ni_vp);
1706 
1707 keeporig:
1708 	/* If there was an error, use the original path name. */
1709 	if (error)
1710 		bcopy(ptr, buf, len);
1711 	return (error);
1712 }
1713