xref: /freebsd/sys/kern/vfs_lookup.c (revision c6bfc8968df9eee0452db8f70472e4bbf68e7b90)
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 		return (error);
490 	}
491 
492 	/*
493 	 * Don't allow empty pathnames.
494 	 */
495 	if (__predict_false(*cnp->cn_pnbuf == '\0')) {
496 		namei_cleanup_cnp(cnp);
497 		return (ENOENT);
498 	}
499 
500 	cnp->cn_nameptr = cnp->cn_pnbuf;
501 	return (0);
502 }
503 
504 /*
505  * Convert a pathname into a pointer to a locked vnode.
506  *
507  * The FOLLOW flag is set when symbolic links are to be followed
508  * when they occur at the end of the name translation process.
509  * Symbolic links are always followed for all other pathname
510  * components other than the last.
511  *
512  * The segflg defines whether the name is to be copied from user
513  * space or kernel space.
514  *
515  * Overall outline of namei:
516  *
517  *	copy in name
518  *	get starting directory
519  *	while (!done && !error) {
520  *		call lookup to search path.
521  *		if symbolic link, massage name in buffer and continue
522  *	}
523  */
524 int
525 namei(struct nameidata *ndp)
526 {
527 	char *cp;		/* pointer into pathname argument */
528 	struct vnode *dp;	/* the directory we are searching */
529 	struct iovec aiov;		/* uio for reading symbolic links */
530 	struct componentname *cnp;
531 	struct thread *td;
532 	struct pwd *pwd;
533 	struct uio auio;
534 	int error, linklen;
535 	enum cache_fpl_status status;
536 
537 	cnp = &ndp->ni_cnd;
538 	td = cnp->cn_thread;
539 #ifdef INVARIANTS
540 	KASSERT((ndp->ni_debugflags & NAMEI_DBG_CALLED) == 0,
541 	    ("%s: repeated call to namei without NDREINIT", __func__));
542 	KASSERT(ndp->ni_debugflags == NAMEI_DBG_INITED,
543 	    ("%s: bad debugflags %d", __func__, ndp->ni_debugflags));
544 	ndp->ni_debugflags |= NAMEI_DBG_CALLED;
545 	if (ndp->ni_startdir != NULL)
546 		ndp->ni_debugflags |= NAMEI_DBG_HADSTARTDIR;
547 	/*
548 	 * For NDVALIDATE.
549 	 *
550 	 * While NDINIT may seem like a more natural place to do it, there are
551 	 * callers which directly modify flags past invoking init.
552 	 */
553 	cnp->cn_origflags = cnp->cn_flags;
554 #endif
555 	ndp->ni_cnd.cn_cred = ndp->ni_cnd.cn_thread->td_ucred;
556 	KASSERT(ndp->ni_resflags == 0, ("%s: garbage in ni_resflags: %x\n",
557 	    __func__, ndp->ni_resflags));
558 	KASSERT(cnp->cn_cred && td->td_proc, ("namei: bad cred/proc"));
559 	KASSERT((cnp->cn_flags & NAMEI_INTERNAL_FLAGS) == 0,
560 	    ("namei: unexpected flags: %" PRIx64 "\n",
561 	    cnp->cn_flags & NAMEI_INTERNAL_FLAGS));
562 	if (cnp->cn_flags & NOCACHE)
563 		KASSERT(cnp->cn_nameiop != LOOKUP,
564 		    ("%s: NOCACHE passed with LOOKUP", __func__));
565 	MPASS(ndp->ni_startdir == NULL || ndp->ni_startdir->v_type == VDIR ||
566 	    ndp->ni_startdir->v_type == VBAD);
567 
568 	ndp->ni_lcf = 0;
569 	ndp->ni_vp = NULL;
570 
571 	error = namei_getpath(ndp);
572 	if (__predict_false(error != 0)) {
573 		return (error);
574 	}
575 
576 #ifdef KTRACE
577 	if (KTRPOINT(td, KTR_NAMEI)) {
578 		KASSERT(cnp->cn_thread == curthread,
579 		    ("namei not using curthread"));
580 		ktrnamei(cnp->cn_pnbuf);
581 	}
582 #endif
583 
584 	/*
585 	 * First try looking up the target without locking any vnodes.
586 	 *
587 	 * We may need to start from scratch or pick up where it left off.
588 	 */
589 	error = cache_fplookup(ndp, &status, &pwd);
590 	switch (status) {
591 	case CACHE_FPL_STATUS_UNSET:
592 		__assert_unreachable();
593 		break;
594 	case CACHE_FPL_STATUS_HANDLED:
595 		if (error == 0)
596 			NDVALIDATE(ndp);
597 		return (error);
598 	case CACHE_FPL_STATUS_PARTIAL:
599 		TAILQ_INIT(&ndp->ni_cap_tracker);
600 		dp = ndp->ni_startdir;
601 		break;
602 	case CACHE_FPL_STATUS_ABORTED:
603 		TAILQ_INIT(&ndp->ni_cap_tracker);
604 		error = namei_setup(ndp, &dp, &pwd);
605 		if (error != 0) {
606 			namei_cleanup_cnp(cnp);
607 			return (error);
608 		}
609 		break;
610 	}
611 
612 	ndp->ni_loopcnt = 0;
613 
614 	/*
615 	 * Locked lookup.
616 	 */
617 	for (;;) {
618 		ndp->ni_startdir = dp;
619 		error = lookup(ndp);
620 		if (error != 0) {
621 			/*
622 			 * Override an error to not allow user to use
623 			 * BENEATH as an oracle.
624 			 */
625 			if ((ndp->ni_lcf & (NI_LCF_LATCH |
626 			    NI_LCF_BENEATH_LATCHED)) == NI_LCF_LATCH)
627 				error = ENOTCAPABLE;
628 			goto out;
629 		}
630 
631 		/*
632 		 * If not a symbolic link, we're done.
633 		 */
634 		if ((cnp->cn_flags & ISSYMLINK) == 0) {
635 			if ((cnp->cn_flags & (SAVENAME | SAVESTART)) == 0) {
636 				namei_cleanup_cnp(cnp);
637 			} else
638 				cnp->cn_flags |= HASBUF;
639 			if ((ndp->ni_lcf & (NI_LCF_LATCH |
640 			    NI_LCF_BENEATH_LATCHED)) == NI_LCF_LATCH) {
641 				NDFREE(ndp, 0);
642 				error = ENOTCAPABLE;
643 			}
644 			nameicap_cleanup(ndp, true);
645 			SDT_PROBE3(vfs, namei, lookup, return, error,
646 			    (error == 0 ? ndp->ni_vp : NULL), false);
647 			pwd_drop(pwd);
648 			if (error == 0)
649 				NDVALIDATE(ndp);
650 			return (error);
651 		}
652 		if (ndp->ni_loopcnt++ >= MAXSYMLINKS) {
653 			error = ELOOP;
654 			break;
655 		}
656 #ifdef MAC
657 		if ((cnp->cn_flags & NOMACCHECK) == 0) {
658 			error = mac_vnode_check_readlink(td->td_ucred,
659 			    ndp->ni_vp);
660 			if (error != 0)
661 				break;
662 		}
663 #endif
664 		if (ndp->ni_pathlen > 1)
665 			cp = uma_zalloc(namei_zone, M_WAITOK);
666 		else
667 			cp = cnp->cn_pnbuf;
668 		aiov.iov_base = cp;
669 		aiov.iov_len = MAXPATHLEN;
670 		auio.uio_iov = &aiov;
671 		auio.uio_iovcnt = 1;
672 		auio.uio_offset = 0;
673 		auio.uio_rw = UIO_READ;
674 		auio.uio_segflg = UIO_SYSSPACE;
675 		auio.uio_td = td;
676 		auio.uio_resid = MAXPATHLEN;
677 		error = VOP_READLINK(ndp->ni_vp, &auio, cnp->cn_cred);
678 		if (error != 0) {
679 			if (ndp->ni_pathlen > 1)
680 				uma_zfree(namei_zone, cp);
681 			break;
682 		}
683 		linklen = MAXPATHLEN - auio.uio_resid;
684 		if (linklen == 0) {
685 			if (ndp->ni_pathlen > 1)
686 				uma_zfree(namei_zone, cp);
687 			error = ENOENT;
688 			break;
689 		}
690 		if (linklen + ndp->ni_pathlen > MAXPATHLEN) {
691 			if (ndp->ni_pathlen > 1)
692 				uma_zfree(namei_zone, cp);
693 			error = ENAMETOOLONG;
694 			break;
695 		}
696 		if (ndp->ni_pathlen > 1) {
697 			bcopy(ndp->ni_next, cp + linklen, ndp->ni_pathlen);
698 			uma_zfree(namei_zone, cnp->cn_pnbuf);
699 			cnp->cn_pnbuf = cp;
700 		} else
701 			cnp->cn_pnbuf[linklen] = '\0';
702 		ndp->ni_pathlen += linklen;
703 		vput(ndp->ni_vp);
704 		dp = ndp->ni_dvp;
705 		/*
706 		 * Check if root directory should replace current directory.
707 		 */
708 		cnp->cn_nameptr = cnp->cn_pnbuf;
709 		if (*(cnp->cn_nameptr) == '/') {
710 			vrele(dp);
711 			error = namei_handle_root(ndp, &dp);
712 			if (error != 0)
713 				goto out;
714 		}
715 	}
716 	vput(ndp->ni_vp);
717 	ndp->ni_vp = NULL;
718 	vrele(ndp->ni_dvp);
719 out:
720 	MPASS(error != 0);
721 	namei_cleanup_cnp(cnp);
722 	nameicap_cleanup(ndp, true);
723 	SDT_PROBE3(vfs, namei, lookup, return, error, NULL, false);
724 	pwd_drop(pwd);
725 	return (error);
726 }
727 
728 static int
729 compute_cn_lkflags(struct mount *mp, int lkflags, int cnflags)
730 {
731 
732 	if (mp == NULL || ((lkflags & LK_SHARED) &&
733 	    (!(mp->mnt_kern_flag & MNTK_LOOKUP_SHARED) ||
734 	    ((cnflags & ISDOTDOT) &&
735 	    (mp->mnt_kern_flag & MNTK_LOOKUP_EXCL_DOTDOT))))) {
736 		lkflags &= ~LK_SHARED;
737 		lkflags |= LK_EXCLUSIVE;
738 	}
739 	lkflags |= LK_NODDLKTREAT;
740 	return (lkflags);
741 }
742 
743 static __inline int
744 needs_exclusive_leaf(struct mount *mp, int flags)
745 {
746 
747 	/*
748 	 * Intermediate nodes can use shared locks, we only need to
749 	 * force an exclusive lock for leaf nodes.
750 	 */
751 	if ((flags & (ISLASTCN | LOCKLEAF)) != (ISLASTCN | LOCKLEAF))
752 		return (0);
753 
754 	/* Always use exclusive locks if LOCKSHARED isn't set. */
755 	if (!(flags & LOCKSHARED))
756 		return (1);
757 
758 	/*
759 	 * For lookups during open(), if the mount point supports
760 	 * extended shared operations, then use a shared lock for the
761 	 * leaf node, otherwise use an exclusive lock.
762 	 */
763 	if ((flags & ISOPEN) != 0)
764 		return (!MNT_EXTENDED_SHARED(mp));
765 
766 	/*
767 	 * Lookup requests outside of open() that specify LOCKSHARED
768 	 * only need a shared lock on the leaf vnode.
769 	 */
770 	return (0);
771 }
772 
773 /*
774  * Search a pathname.
775  * This is a very central and rather complicated routine.
776  *
777  * The pathname is pointed to by ni_ptr and is of length ni_pathlen.
778  * The starting directory is taken from ni_startdir. The pathname is
779  * descended until done, or a symbolic link is encountered. The variable
780  * ni_more is clear if the path is completed; it is set to one if a
781  * symbolic link needing interpretation is encountered.
782  *
783  * The flag argument is LOOKUP, CREATE, RENAME, or DELETE depending on
784  * whether the name is to be looked up, created, renamed, or deleted.
785  * When CREATE, RENAME, or DELETE is specified, information usable in
786  * creating, renaming, or deleting a directory entry may be calculated.
787  * If flag has LOCKPARENT or'ed into it, the parent directory is returned
788  * locked. If flag has WANTPARENT or'ed into it, the parent directory is
789  * returned unlocked. Otherwise the parent directory is not returned. If
790  * the target of the pathname exists and LOCKLEAF is or'ed into the flag
791  * the target is returned locked, otherwise it is returned unlocked.
792  * When creating or renaming and LOCKPARENT is specified, the target may not
793  * be ".".  When deleting and LOCKPARENT is specified, the target may be ".".
794  *
795  * Overall outline of lookup:
796  *
797  * dirloop:
798  *	identify next component of name at ndp->ni_ptr
799  *	handle degenerate case where name is null string
800  *	if .. and crossing mount points and on mounted filesys, find parent
801  *	call VOP_LOOKUP routine for next component name
802  *	    directory vnode returned in ni_dvp, unlocked unless LOCKPARENT set
803  *	    component vnode returned in ni_vp (if it exists), locked.
804  *	if result vnode is mounted on and crossing mount points,
805  *	    find mounted on vnode
806  *	if more components of name, do next level at dirloop
807  *	return the answer in ni_vp, locked if LOCKLEAF set
808  *	    if LOCKPARENT set, return locked parent in ni_dvp
809  *	    if WANTPARENT set, return unlocked parent in ni_dvp
810  */
811 int
812 lookup(struct nameidata *ndp)
813 {
814 	char *cp;			/* pointer into pathname argument */
815 	char *prev_ni_next;		/* saved ndp->ni_next */
816 	struct vnode *dp = NULL;	/* the directory we are searching */
817 	struct vnode *tdp;		/* saved dp */
818 	struct mount *mp;		/* mount table entry */
819 	struct prison *pr;
820 	size_t prev_ni_pathlen;		/* saved ndp->ni_pathlen */
821 	int docache;			/* == 0 do not cache last component */
822 	int wantparent;			/* 1 => wantparent or lockparent flag */
823 	int rdonly;			/* lookup read-only flag bit */
824 	int error = 0;
825 	int dpunlocked = 0;		/* dp has already been unlocked */
826 	int relookup = 0;		/* do not consume the path component */
827 	struct componentname *cnp = &ndp->ni_cnd;
828 	int lkflags_save;
829 	int ni_dvp_unlocked;
830 
831 	/*
832 	 * Setup: break out flag bits into variables.
833 	 */
834 	ni_dvp_unlocked = 0;
835 	wantparent = cnp->cn_flags & (LOCKPARENT | WANTPARENT);
836 	KASSERT(cnp->cn_nameiop == LOOKUP || wantparent,
837 	    ("CREATE, DELETE, RENAME require LOCKPARENT or WANTPARENT."));
838 	/*
839 	 * When set to zero, docache causes the last component of the
840 	 * pathname to be deleted from the cache and the full lookup
841 	 * of the name to be done (via VOP_CACHEDLOOKUP()). Often
842 	 * filesystems need some pre-computed values that are made
843 	 * during the full lookup, for instance UFS sets dp->i_offset.
844 	 *
845 	 * The docache variable is set to zero when requested by the
846 	 * NOCACHE flag and for all modifying operations except CREATE.
847 	 */
848 	docache = (cnp->cn_flags & NOCACHE) ^ NOCACHE;
849 	if (cnp->cn_nameiop == DELETE ||
850 	    (wantparent && cnp->cn_nameiop != CREATE &&
851 	     cnp->cn_nameiop != LOOKUP))
852 		docache = 0;
853 	rdonly = cnp->cn_flags & RDONLY;
854 	cnp->cn_flags &= ~ISSYMLINK;
855 	ndp->ni_dvp = NULL;
856 	/*
857 	 * We use shared locks until we hit the parent of the last cn then
858 	 * we adjust based on the requesting flags.
859 	 */
860 	cnp->cn_lkflags = LK_SHARED;
861 	dp = ndp->ni_startdir;
862 	ndp->ni_startdir = NULLVP;
863 	vn_lock(dp,
864 	    compute_cn_lkflags(dp->v_mount, cnp->cn_lkflags | LK_RETRY,
865 	    cnp->cn_flags));
866 
867 dirloop:
868 	/*
869 	 * Search a new directory.
870 	 *
871 	 * The last component of the filename is left accessible via
872 	 * cnp->cn_nameptr for callers that need the name. Callers needing
873 	 * the name set the SAVENAME flag. When done, they assume
874 	 * responsibility for freeing the pathname buffer.
875 	 */
876 	for (cp = cnp->cn_nameptr; *cp != 0 && *cp != '/'; cp++)
877 		continue;
878 	cnp->cn_namelen = cp - cnp->cn_nameptr;
879 	if (cnp->cn_namelen > NAME_MAX) {
880 		error = ENAMETOOLONG;
881 		goto bad;
882 	}
883 #ifdef NAMEI_DIAGNOSTIC
884 	{ char c = *cp;
885 	*cp = '\0';
886 	printf("{%s}: ", cnp->cn_nameptr);
887 	*cp = c; }
888 #endif
889 	prev_ni_pathlen = ndp->ni_pathlen;
890 	ndp->ni_pathlen -= cnp->cn_namelen;
891 	KASSERT(ndp->ni_pathlen <= PATH_MAX,
892 	    ("%s: ni_pathlen underflow to %zd\n", __func__, ndp->ni_pathlen));
893 	prev_ni_next = ndp->ni_next;
894 	ndp->ni_next = cp;
895 
896 	/*
897 	 * Replace multiple slashes by a single slash and trailing slashes
898 	 * by a null.  This must be done before VOP_LOOKUP() because some
899 	 * fs's don't know about trailing slashes.  Remember if there were
900 	 * trailing slashes to handle symlinks, existing non-directories
901 	 * and non-existing files that won't be directories specially later.
902 	 */
903 	while (*cp == '/' && (cp[1] == '/' || cp[1] == '\0')) {
904 		cp++;
905 		ndp->ni_pathlen--;
906 		if (*cp == '\0') {
907 			*ndp->ni_next = '\0';
908 			cnp->cn_flags |= TRAILINGSLASH;
909 		}
910 	}
911 	ndp->ni_next = cp;
912 
913 	cnp->cn_flags |= MAKEENTRY;
914 	if (*cp == '\0' && docache == 0)
915 		cnp->cn_flags &= ~MAKEENTRY;
916 	if (cnp->cn_namelen == 2 &&
917 	    cnp->cn_nameptr[1] == '.' && cnp->cn_nameptr[0] == '.')
918 		cnp->cn_flags |= ISDOTDOT;
919 	else
920 		cnp->cn_flags &= ~ISDOTDOT;
921 	if (*ndp->ni_next == 0)
922 		cnp->cn_flags |= ISLASTCN;
923 	else
924 		cnp->cn_flags &= ~ISLASTCN;
925 
926 	if ((cnp->cn_flags & ISLASTCN) != 0 &&
927 	    cnp->cn_namelen == 1 && cnp->cn_nameptr[0] == '.' &&
928 	    (cnp->cn_nameiop == DELETE || cnp->cn_nameiop == RENAME)) {
929 		error = EINVAL;
930 		goto bad;
931 	}
932 
933 	nameicap_tracker_add(ndp, dp);
934 
935 	/*
936 	 * Check for degenerate name (e.g. / or "")
937 	 * which is a way of talking about a directory,
938 	 * e.g. like "/." or ".".
939 	 */
940 	if (cnp->cn_nameptr[0] == '\0') {
941 		if (dp->v_type != VDIR) {
942 			error = ENOTDIR;
943 			goto bad;
944 		}
945 		if (cnp->cn_nameiop != LOOKUP) {
946 			error = EISDIR;
947 			goto bad;
948 		}
949 		if (wantparent) {
950 			ndp->ni_dvp = dp;
951 			VREF(dp);
952 		}
953 		ndp->ni_vp = dp;
954 
955 		if (cnp->cn_flags & AUDITVNODE1)
956 			AUDIT_ARG_VNODE1(dp);
957 		else if (cnp->cn_flags & AUDITVNODE2)
958 			AUDIT_ARG_VNODE2(dp);
959 
960 		if (!(cnp->cn_flags & (LOCKPARENT | LOCKLEAF)))
961 			VOP_UNLOCK(dp);
962 		/* XXX This should probably move to the top of function. */
963 		if (cnp->cn_flags & SAVESTART)
964 			panic("lookup: SAVESTART");
965 		goto success;
966 	}
967 
968 	/*
969 	 * Handle "..": five special cases.
970 	 * 0. If doing a capability lookup and lookup_cap_dotdot is
971 	 *    disabled, return ENOTCAPABLE.
972 	 * 1. Return an error if this is the last component of
973 	 *    the name and the operation is DELETE or RENAME.
974 	 * 2. If at root directory (e.g. after chroot)
975 	 *    or at absolute root directory
976 	 *    then ignore it so can't get out.
977 	 * 3. If this vnode is the root of a mounted
978 	 *    filesystem, then replace it with the
979 	 *    vnode which was mounted on so we take the
980 	 *    .. in the other filesystem.
981 	 * 4. If the vnode is the top directory of
982 	 *    the jail or chroot, don't let them out.
983 	 * 5. If doing a capability lookup and lookup_cap_dotdot is
984 	 *    enabled, return ENOTCAPABLE if the lookup would escape
985 	 *    from the initial file descriptor directory.  Checks are
986 	 *    done by ensuring that namei() already traversed the
987 	 *    result of dotdot lookup.
988 	 */
989 	if (cnp->cn_flags & ISDOTDOT) {
990 		if ((ndp->ni_lcf & (NI_LCF_STRICTRELATIVE | NI_LCF_CAP_DOTDOT))
991 		    == NI_LCF_STRICTRELATIVE) {
992 #ifdef KTRACE
993 			if (KTRPOINT(curthread, KTR_CAPFAIL))
994 				ktrcapfail(CAPFAIL_LOOKUP, NULL, NULL);
995 #endif
996 			error = ENOTCAPABLE;
997 			goto bad;
998 		}
999 		if ((cnp->cn_flags & ISLASTCN) != 0 &&
1000 		    (cnp->cn_nameiop == DELETE || cnp->cn_nameiop == RENAME)) {
1001 			error = EINVAL;
1002 			goto bad;
1003 		}
1004 		for (;;) {
1005 			for (pr = cnp->cn_cred->cr_prison; pr != NULL;
1006 			     pr = pr->pr_parent)
1007 				if (dp == pr->pr_root)
1008 					break;
1009 			if (dp == ndp->ni_rootdir ||
1010 			    dp == ndp->ni_topdir ||
1011 			    dp == rootvnode ||
1012 			    pr != NULL ||
1013 			    ((dp->v_vflag & VV_ROOT) != 0 &&
1014 			     (cnp->cn_flags & NOCROSSMOUNT) != 0)) {
1015 				ndp->ni_dvp = dp;
1016 				ndp->ni_vp = dp;
1017 				VREF(dp);
1018 				goto nextname;
1019 			}
1020 			if ((dp->v_vflag & VV_ROOT) == 0)
1021 				break;
1022 			if (VN_IS_DOOMED(dp)) {	/* forced unmount */
1023 				error = ENOENT;
1024 				goto bad;
1025 			}
1026 			tdp = dp;
1027 			dp = dp->v_mount->mnt_vnodecovered;
1028 			VREF(dp);
1029 			vput(tdp);
1030 			vn_lock(dp,
1031 			    compute_cn_lkflags(dp->v_mount, cnp->cn_lkflags |
1032 			    LK_RETRY, ISDOTDOT));
1033 			error = nameicap_check_dotdot(ndp, dp);
1034 			if (error != 0) {
1035 #ifdef KTRACE
1036 				if (KTRPOINT(curthread, KTR_CAPFAIL))
1037 					ktrcapfail(CAPFAIL_LOOKUP, NULL, NULL);
1038 #endif
1039 				goto bad;
1040 			}
1041 		}
1042 	}
1043 
1044 	/*
1045 	 * We now have a segment name to search for, and a directory to search.
1046 	 */
1047 unionlookup:
1048 #ifdef MAC
1049 	error = mac_vnode_check_lookup(cnp->cn_thread->td_ucred, dp, cnp);
1050 	if (error)
1051 		goto bad;
1052 #endif
1053 	ndp->ni_dvp = dp;
1054 	ndp->ni_vp = NULL;
1055 	ASSERT_VOP_LOCKED(dp, "lookup");
1056 	/*
1057 	 * If we have a shared lock we may need to upgrade the lock for the
1058 	 * last operation.
1059 	 */
1060 	if ((cnp->cn_flags & LOCKPARENT) && (cnp->cn_flags & ISLASTCN) &&
1061 	    dp != vp_crossmp && VOP_ISLOCKED(dp) == LK_SHARED)
1062 		vn_lock(dp, LK_UPGRADE|LK_RETRY);
1063 	if (VN_IS_DOOMED(dp)) {
1064 		error = ENOENT;
1065 		goto bad;
1066 	}
1067 	/*
1068 	 * If we're looking up the last component and we need an exclusive
1069 	 * lock, adjust our lkflags.
1070 	 */
1071 	if (needs_exclusive_leaf(dp->v_mount, cnp->cn_flags))
1072 		cnp->cn_lkflags = LK_EXCLUSIVE;
1073 #ifdef NAMEI_DIAGNOSTIC
1074 	vn_printf(dp, "lookup in ");
1075 #endif
1076 	lkflags_save = cnp->cn_lkflags;
1077 	cnp->cn_lkflags = compute_cn_lkflags(dp->v_mount, cnp->cn_lkflags,
1078 	    cnp->cn_flags);
1079 	error = VOP_LOOKUP(dp, &ndp->ni_vp, cnp);
1080 	cnp->cn_lkflags = lkflags_save;
1081 	if (error != 0) {
1082 		KASSERT(ndp->ni_vp == NULL, ("leaf should be empty"));
1083 #ifdef NAMEI_DIAGNOSTIC
1084 		printf("not found\n");
1085 #endif
1086 		if ((error == ENOENT) &&
1087 		    (dp->v_vflag & VV_ROOT) && (dp->v_mount != NULL) &&
1088 		    (dp->v_mount->mnt_flag & MNT_UNION)) {
1089 			tdp = dp;
1090 			dp = dp->v_mount->mnt_vnodecovered;
1091 			VREF(dp);
1092 			vput(tdp);
1093 			vn_lock(dp,
1094 			    compute_cn_lkflags(dp->v_mount, cnp->cn_lkflags |
1095 			    LK_RETRY, cnp->cn_flags));
1096 			nameicap_tracker_add(ndp, dp);
1097 			goto unionlookup;
1098 		}
1099 
1100 		if (error == ERELOOKUP) {
1101 			vref(dp);
1102 			ndp->ni_vp = dp;
1103 			error = 0;
1104 			relookup = 1;
1105 			goto good;
1106 		}
1107 
1108 		if (error != EJUSTRETURN)
1109 			goto bad;
1110 		/*
1111 		 * At this point, we know we're at the end of the
1112 		 * pathname.  If creating / renaming, we can consider
1113 		 * allowing the file or directory to be created / renamed,
1114 		 * provided we're not on a read-only filesystem.
1115 		 */
1116 		if (rdonly) {
1117 			error = EROFS;
1118 			goto bad;
1119 		}
1120 		/* trailing slash only allowed for directories */
1121 		if ((cnp->cn_flags & TRAILINGSLASH) &&
1122 		    !(cnp->cn_flags & WILLBEDIR)) {
1123 			error = ENOENT;
1124 			goto bad;
1125 		}
1126 		if ((cnp->cn_flags & LOCKPARENT) == 0)
1127 			VOP_UNLOCK(dp);
1128 		/*
1129 		 * We return with ni_vp NULL to indicate that the entry
1130 		 * doesn't currently exist, leaving a pointer to the
1131 		 * (possibly locked) directory vnode in ndp->ni_dvp.
1132 		 */
1133 		if (cnp->cn_flags & SAVESTART) {
1134 			ndp->ni_startdir = ndp->ni_dvp;
1135 			VREF(ndp->ni_startdir);
1136 		}
1137 		goto success;
1138 	}
1139 
1140 good:
1141 #ifdef NAMEI_DIAGNOSTIC
1142 	printf("found\n");
1143 #endif
1144 	dp = ndp->ni_vp;
1145 
1146 	/*
1147 	 * Check to see if the vnode has been mounted on;
1148 	 * if so find the root of the mounted filesystem.
1149 	 */
1150 	while (dp->v_type == VDIR && (mp = dp->v_mountedhere) &&
1151 	       (cnp->cn_flags & NOCROSSMOUNT) == 0) {
1152 		if (vfs_busy(mp, 0))
1153 			continue;
1154 		vput(dp);
1155 		if (dp != ndp->ni_dvp)
1156 			vput(ndp->ni_dvp);
1157 		else
1158 			vrele(ndp->ni_dvp);
1159 		vrefact(vp_crossmp);
1160 		ndp->ni_dvp = vp_crossmp;
1161 		error = VFS_ROOT(mp, compute_cn_lkflags(mp, cnp->cn_lkflags,
1162 		    cnp->cn_flags), &tdp);
1163 		vfs_unbusy(mp);
1164 		if (vn_lock(vp_crossmp, LK_SHARED | LK_NOWAIT))
1165 			panic("vp_crossmp exclusively locked or reclaimed");
1166 		if (error) {
1167 			dpunlocked = 1;
1168 			goto bad2;
1169 		}
1170 		ndp->ni_vp = dp = tdp;
1171 	}
1172 
1173 	/*
1174 	 * Check for symbolic link
1175 	 */
1176 	if ((dp->v_type == VLNK) &&
1177 	    ((cnp->cn_flags & FOLLOW) || (cnp->cn_flags & TRAILINGSLASH) ||
1178 	     *ndp->ni_next == '/')) {
1179 		cnp->cn_flags |= ISSYMLINK;
1180 		if (VN_IS_DOOMED(dp)) {
1181 			/*
1182 			 * We can't know whether the directory was mounted with
1183 			 * NOSYMFOLLOW, so we can't follow safely.
1184 			 */
1185 			error = ENOENT;
1186 			goto bad2;
1187 		}
1188 		if (dp->v_mount->mnt_flag & MNT_NOSYMFOLLOW) {
1189 			error = EACCES;
1190 			goto bad2;
1191 		}
1192 		/*
1193 		 * Symlink code always expects an unlocked dvp.
1194 		 */
1195 		if (ndp->ni_dvp != ndp->ni_vp) {
1196 			VOP_UNLOCK(ndp->ni_dvp);
1197 			ni_dvp_unlocked = 1;
1198 		}
1199 		goto success;
1200 	}
1201 
1202 nextname:
1203 	/*
1204 	 * Not a symbolic link that we will follow.  Continue with the
1205 	 * next component if there is any; otherwise, we're done.
1206 	 */
1207 	KASSERT((cnp->cn_flags & ISLASTCN) || *ndp->ni_next == '/',
1208 	    ("lookup: invalid path state."));
1209 	if (relookup) {
1210 		relookup = 0;
1211 		ndp->ni_pathlen = prev_ni_pathlen;
1212 		ndp->ni_next = prev_ni_next;
1213 		if (ndp->ni_dvp != dp)
1214 			vput(ndp->ni_dvp);
1215 		else
1216 			vrele(ndp->ni_dvp);
1217 		goto dirloop;
1218 	}
1219 	if (cnp->cn_flags & ISDOTDOT) {
1220 		error = nameicap_check_dotdot(ndp, ndp->ni_vp);
1221 		if (error != 0) {
1222 #ifdef KTRACE
1223 			if (KTRPOINT(curthread, KTR_CAPFAIL))
1224 				ktrcapfail(CAPFAIL_LOOKUP, NULL, NULL);
1225 #endif
1226 			goto bad2;
1227 		}
1228 	}
1229 	if (*ndp->ni_next == '/') {
1230 		cnp->cn_nameptr = ndp->ni_next;
1231 		while (*cnp->cn_nameptr == '/') {
1232 			cnp->cn_nameptr++;
1233 			ndp->ni_pathlen--;
1234 		}
1235 		if (ndp->ni_dvp != dp)
1236 			vput(ndp->ni_dvp);
1237 		else
1238 			vrele(ndp->ni_dvp);
1239 		goto dirloop;
1240 	}
1241 	/*
1242 	 * If we're processing a path with a trailing slash,
1243 	 * check that the end result is a directory.
1244 	 */
1245 	if ((cnp->cn_flags & TRAILINGSLASH) && dp->v_type != VDIR) {
1246 		error = ENOTDIR;
1247 		goto bad2;
1248 	}
1249 	/*
1250 	 * Disallow directory write attempts on read-only filesystems.
1251 	 */
1252 	if (rdonly &&
1253 	    (cnp->cn_nameiop == DELETE || cnp->cn_nameiop == RENAME)) {
1254 		error = EROFS;
1255 		goto bad2;
1256 	}
1257 	if (cnp->cn_flags & SAVESTART) {
1258 		ndp->ni_startdir = ndp->ni_dvp;
1259 		VREF(ndp->ni_startdir);
1260 	}
1261 	if (!wantparent) {
1262 		ni_dvp_unlocked = 2;
1263 		if (ndp->ni_dvp != dp)
1264 			vput(ndp->ni_dvp);
1265 		else
1266 			vrele(ndp->ni_dvp);
1267 	} else if ((cnp->cn_flags & LOCKPARENT) == 0 && ndp->ni_dvp != dp) {
1268 		VOP_UNLOCK(ndp->ni_dvp);
1269 		ni_dvp_unlocked = 1;
1270 	}
1271 
1272 	if (cnp->cn_flags & AUDITVNODE1)
1273 		AUDIT_ARG_VNODE1(dp);
1274 	else if (cnp->cn_flags & AUDITVNODE2)
1275 		AUDIT_ARG_VNODE2(dp);
1276 
1277 	if ((cnp->cn_flags & LOCKLEAF) == 0)
1278 		VOP_UNLOCK(dp);
1279 success:
1280 	/*
1281 	 * FIXME: for lookups which only cross a mount point to fetch the
1282 	 * root vnode, ni_dvp will be set to vp_crossmp. This can be a problem
1283 	 * if either WANTPARENT or LOCKPARENT is set.
1284 	 */
1285 	/*
1286 	 * Because of shared lookup we may have the vnode shared locked, but
1287 	 * the caller may want it to be exclusively locked.
1288 	 */
1289 	if (needs_exclusive_leaf(dp->v_mount, cnp->cn_flags) &&
1290 	    VOP_ISLOCKED(dp) != LK_EXCLUSIVE) {
1291 		vn_lock(dp, LK_UPGRADE | LK_RETRY);
1292 		if (VN_IS_DOOMED(dp)) {
1293 			error = ENOENT;
1294 			goto bad2;
1295 		}
1296 	}
1297 	if (ndp->ni_vp != NULL && ndp->ni_vp->v_type == VDIR)
1298 		nameicap_tracker_add(ndp, ndp->ni_vp);
1299 	return (0);
1300 
1301 bad2:
1302 	if (ni_dvp_unlocked != 2) {
1303 		if (dp != ndp->ni_dvp && !ni_dvp_unlocked)
1304 			vput(ndp->ni_dvp);
1305 		else
1306 			vrele(ndp->ni_dvp);
1307 	}
1308 bad:
1309 	if (!dpunlocked)
1310 		vput(dp);
1311 	ndp->ni_vp = NULL;
1312 	return (error);
1313 }
1314 
1315 /*
1316  * relookup - lookup a path name component
1317  *    Used by lookup to re-acquire things.
1318  */
1319 int
1320 relookup(struct vnode *dvp, struct vnode **vpp, struct componentname *cnp)
1321 {
1322 	struct vnode *dp = NULL;		/* the directory we are searching */
1323 	int rdonly;			/* lookup read-only flag bit */
1324 	int error = 0;
1325 
1326 	KASSERT(cnp->cn_flags & ISLASTCN,
1327 	    ("relookup: Not given last component."));
1328 	/*
1329 	 * Setup: break out flag bits into variables.
1330 	 */
1331 	KASSERT((cnp->cn_flags & (LOCKPARENT | WANTPARENT)) != 0,
1332 	    ("relookup: parent not wanted"));
1333 	rdonly = cnp->cn_flags & RDONLY;
1334 	cnp->cn_flags &= ~ISSYMLINK;
1335 	dp = dvp;
1336 	cnp->cn_lkflags = LK_EXCLUSIVE;
1337 	vn_lock(dp, LK_EXCLUSIVE | LK_RETRY);
1338 
1339 	/*
1340 	 * Search a new directory.
1341 	 *
1342 	 * The last component of the filename is left accessible via
1343 	 * cnp->cn_nameptr for callers that need the name. Callers needing
1344 	 * the name set the SAVENAME flag. When done, they assume
1345 	 * responsibility for freeing the pathname buffer.
1346 	 */
1347 #ifdef NAMEI_DIAGNOSTIC
1348 	printf("{%s}: ", cnp->cn_nameptr);
1349 #endif
1350 
1351 	/*
1352 	 * Check for "" which represents the root directory after slash
1353 	 * removal.
1354 	 */
1355 	if (cnp->cn_nameptr[0] == '\0') {
1356 		/*
1357 		 * Support only LOOKUP for "/" because lookup()
1358 		 * can't succeed for CREATE, DELETE and RENAME.
1359 		 */
1360 		KASSERT(cnp->cn_nameiop == LOOKUP, ("nameiop must be LOOKUP"));
1361 		KASSERT(dp->v_type == VDIR, ("dp is not a directory"));
1362 
1363 		if (!(cnp->cn_flags & LOCKLEAF))
1364 			VOP_UNLOCK(dp);
1365 		*vpp = dp;
1366 		/* XXX This should probably move to the top of function. */
1367 		if (cnp->cn_flags & SAVESTART)
1368 			panic("lookup: SAVESTART");
1369 		return (0);
1370 	}
1371 
1372 	if (cnp->cn_flags & ISDOTDOT)
1373 		panic ("relookup: lookup on dot-dot");
1374 
1375 	/*
1376 	 * We now have a segment name to search for, and a directory to search.
1377 	 */
1378 #ifdef NAMEI_DIAGNOSTIC
1379 	vn_printf(dp, "search in ");
1380 #endif
1381 	if ((error = VOP_LOOKUP(dp, vpp, cnp)) != 0) {
1382 		KASSERT(*vpp == NULL, ("leaf should be empty"));
1383 		if (error != EJUSTRETURN)
1384 			goto bad;
1385 		/*
1386 		 * If creating and at end of pathname, then can consider
1387 		 * allowing file to be created.
1388 		 */
1389 		if (rdonly) {
1390 			error = EROFS;
1391 			goto bad;
1392 		}
1393 		/* ASSERT(dvp == ndp->ni_startdir) */
1394 		if (cnp->cn_flags & SAVESTART)
1395 			VREF(dvp);
1396 		if ((cnp->cn_flags & LOCKPARENT) == 0)
1397 			VOP_UNLOCK(dp);
1398 		/*
1399 		 * We return with ni_vp NULL to indicate that the entry
1400 		 * doesn't currently exist, leaving a pointer to the
1401 		 * (possibly locked) directory vnode in ndp->ni_dvp.
1402 		 */
1403 		return (0);
1404 	}
1405 
1406 	dp = *vpp;
1407 
1408 	/*
1409 	 * Disallow directory write attempts on read-only filesystems.
1410 	 */
1411 	if (rdonly &&
1412 	    (cnp->cn_nameiop == DELETE || cnp->cn_nameiop == RENAME)) {
1413 		if (dvp == dp)
1414 			vrele(dvp);
1415 		else
1416 			vput(dvp);
1417 		error = EROFS;
1418 		goto bad;
1419 	}
1420 	/*
1421 	 * Set the parent lock/ref state to the requested state.
1422 	 */
1423 	if ((cnp->cn_flags & LOCKPARENT) == 0 && dvp != dp)
1424 		VOP_UNLOCK(dvp);
1425 	/*
1426 	 * Check for symbolic link
1427 	 */
1428 	KASSERT(dp->v_type != VLNK || !(cnp->cn_flags & FOLLOW),
1429 	    ("relookup: symlink found.\n"));
1430 
1431 	/* ASSERT(dvp == ndp->ni_startdir) */
1432 	if (cnp->cn_flags & SAVESTART)
1433 		VREF(dvp);
1434 
1435 	if ((cnp->cn_flags & LOCKLEAF) == 0)
1436 		VOP_UNLOCK(dp);
1437 	return (0);
1438 bad:
1439 	vput(dp);
1440 	*vpp = NULL;
1441 	return (error);
1442 }
1443 
1444 /*
1445  * Free data allocated by namei(); see namei(9) for details.
1446  */
1447 void
1448 NDFREE_PNBUF(struct nameidata *ndp)
1449 {
1450 
1451 	if ((ndp->ni_cnd.cn_flags & HASBUF) != 0) {
1452 		MPASS((ndp->ni_cnd.cn_flags & (SAVENAME | SAVESTART)) != 0);
1453 		uma_zfree(namei_zone, ndp->ni_cnd.cn_pnbuf);
1454 		ndp->ni_cnd.cn_flags &= ~HASBUF;
1455 	}
1456 }
1457 
1458 void
1459 (NDFREE)(struct nameidata *ndp, const u_int flags)
1460 {
1461 	int unlock_dvp;
1462 	int unlock_vp;
1463 
1464 	unlock_dvp = 0;
1465 	unlock_vp = 0;
1466 
1467 	if (!(flags & NDF_NO_FREE_PNBUF)) {
1468 		NDFREE_PNBUF(ndp);
1469 	}
1470 	if (!(flags & NDF_NO_VP_UNLOCK) &&
1471 	    (ndp->ni_cnd.cn_flags & LOCKLEAF) && ndp->ni_vp)
1472 		unlock_vp = 1;
1473 	if (!(flags & NDF_NO_DVP_UNLOCK) &&
1474 	    (ndp->ni_cnd.cn_flags & LOCKPARENT) &&
1475 	    ndp->ni_dvp != ndp->ni_vp)
1476 		unlock_dvp = 1;
1477 	if (!(flags & NDF_NO_VP_RELE) && ndp->ni_vp) {
1478 		if (unlock_vp) {
1479 			vput(ndp->ni_vp);
1480 			unlock_vp = 0;
1481 		} else
1482 			vrele(ndp->ni_vp);
1483 		ndp->ni_vp = NULL;
1484 	}
1485 	if (unlock_vp)
1486 		VOP_UNLOCK(ndp->ni_vp);
1487 	if (!(flags & NDF_NO_DVP_RELE) &&
1488 	    (ndp->ni_cnd.cn_flags & (LOCKPARENT|WANTPARENT))) {
1489 		if (unlock_dvp) {
1490 			vput(ndp->ni_dvp);
1491 			unlock_dvp = 0;
1492 		} else
1493 			vrele(ndp->ni_dvp);
1494 		ndp->ni_dvp = NULL;
1495 	}
1496 	if (unlock_dvp)
1497 		VOP_UNLOCK(ndp->ni_dvp);
1498 	if (!(flags & NDF_NO_STARTDIR_RELE) &&
1499 	    (ndp->ni_cnd.cn_flags & SAVESTART)) {
1500 		vrele(ndp->ni_startdir);
1501 		ndp->ni_startdir = NULL;
1502 	}
1503 }
1504 
1505 #ifdef INVARIANTS
1506 /*
1507  * Validate the final state of ndp after the lookup.
1508  *
1509  * Historically filesystems were allowed to modify cn_flags. Most notably they
1510  * can add SAVENAME to the request, resulting in HASBUF and pushing subsequent
1511  * clean up to the consumer. In practice this seems to only concern != LOOKUP
1512  * operations.
1513  *
1514  * As a step towards stricter API contract this routine validates the state to
1515  * clean up. Note validation is a work in progress with the intent of becoming
1516  * stricter over time.
1517  */
1518 #define NDMODIFYINGFLAGS (LOCKLEAF | LOCKPARENT | WANTPARENT | SAVENAME | SAVESTART | HASBUF)
1519 void
1520 NDVALIDATE(struct nameidata *ndp)
1521 {
1522 	struct componentname *cnp;
1523 	u_int64_t used, orig;
1524 
1525 	cnp = &ndp->ni_cnd;
1526 	orig = cnp->cn_origflags;
1527 	used = cnp->cn_flags;
1528 	switch (cnp->cn_nameiop) {
1529 	case LOOKUP:
1530 		/*
1531 		 * For plain lookup we require strict conformance -- nothing
1532 		 * to clean up if it was not requested by the caller.
1533 		 */
1534 		orig &= NDMODIFYINGFLAGS;
1535 		used &= NDMODIFYINGFLAGS;
1536 		if ((orig & (SAVENAME | SAVESTART)) != 0)
1537 			orig |= HASBUF;
1538 		if (orig != used) {
1539 			goto out_mismatch;
1540 		}
1541 		break;
1542 	case CREATE:
1543 	case DELETE:
1544 	case RENAME:
1545 		/*
1546 		 * Some filesystems set SAVENAME to provoke HASBUF, accomodate
1547 		 * for it until it gets fixed.
1548 		 */
1549 		orig &= NDMODIFYINGFLAGS;
1550 		orig |= (SAVENAME | HASBUF);
1551 		used &= NDMODIFYINGFLAGS;
1552 		used |= (SAVENAME | HASBUF);
1553 		if (orig != used) {
1554 			goto out_mismatch;
1555 		}
1556 		break;
1557 	}
1558 	return;
1559 out_mismatch:
1560 	panic("%s: mismatched flags for op %d: added %" PRIx64 ", "
1561 	    "removed %" PRIx64" (%" PRIx64" != %" PRIx64"; stored %" PRIx64" != %" PRIx64")",
1562 	    __func__, cnp->cn_nameiop, used & ~orig, orig &~ used,
1563 	    orig, used, cnp->cn_origflags, cnp->cn_flags);
1564 }
1565 #endif
1566 
1567 /*
1568  * Determine if there is a suitable alternate filename under the specified
1569  * prefix for the specified path.  If the create flag is set, then the
1570  * alternate prefix will be used so long as the parent directory exists.
1571  * This is used by the various compatibility ABIs so that Linux binaries prefer
1572  * files under /compat/linux for example.  The chosen path (whether under
1573  * the prefix or under /) is returned in a kernel malloc'd buffer pointed
1574  * to by pathbuf.  The caller is responsible for free'ing the buffer from
1575  * the M_TEMP bucket if one is returned.
1576  */
1577 int
1578 kern_alternate_path(struct thread *td, const char *prefix, const char *path,
1579     enum uio_seg pathseg, char **pathbuf, int create, int dirfd)
1580 {
1581 	struct nameidata nd, ndroot;
1582 	char *ptr, *buf, *cp;
1583 	size_t len, sz;
1584 	int error;
1585 
1586 	buf = (char *) malloc(MAXPATHLEN, M_TEMP, M_WAITOK);
1587 	*pathbuf = buf;
1588 
1589 	/* Copy the prefix into the new pathname as a starting point. */
1590 	len = strlcpy(buf, prefix, MAXPATHLEN);
1591 	if (len >= MAXPATHLEN) {
1592 		*pathbuf = NULL;
1593 		free(buf, M_TEMP);
1594 		return (EINVAL);
1595 	}
1596 	sz = MAXPATHLEN - len;
1597 	ptr = buf + len;
1598 
1599 	/* Append the filename to the prefix. */
1600 	if (pathseg == UIO_SYSSPACE)
1601 		error = copystr(path, ptr, sz, &len);
1602 	else
1603 		error = copyinstr(path, ptr, sz, &len);
1604 
1605 	if (error) {
1606 		*pathbuf = NULL;
1607 		free(buf, M_TEMP);
1608 		return (error);
1609 	}
1610 
1611 	/* Only use a prefix with absolute pathnames. */
1612 	if (*ptr != '/') {
1613 		error = EINVAL;
1614 		goto keeporig;
1615 	}
1616 
1617 	if (dirfd != AT_FDCWD) {
1618 		/*
1619 		 * We want the original because the "prefix" is
1620 		 * included in the already opened dirfd.
1621 		 */
1622 		bcopy(ptr, buf, len);
1623 		return (0);
1624 	}
1625 
1626 	/*
1627 	 * We know that there is a / somewhere in this pathname.
1628 	 * Search backwards for it, to find the file's parent dir
1629 	 * to see if it exists in the alternate tree. If it does,
1630 	 * and we want to create a file (cflag is set). We don't
1631 	 * need to worry about the root comparison in this case.
1632 	 */
1633 
1634 	if (create) {
1635 		for (cp = &ptr[len] - 1; *cp != '/'; cp--);
1636 		*cp = '\0';
1637 
1638 		NDINIT(&nd, LOOKUP, NOFOLLOW, UIO_SYSSPACE, buf, td);
1639 		error = namei(&nd);
1640 		*cp = '/';
1641 		if (error != 0)
1642 			goto keeporig;
1643 	} else {
1644 		NDINIT(&nd, LOOKUP, NOFOLLOW, UIO_SYSSPACE, buf, td);
1645 
1646 		error = namei(&nd);
1647 		if (error != 0)
1648 			goto keeporig;
1649 
1650 		/*
1651 		 * We now compare the vnode of the prefix to the one
1652 		 * vnode asked. If they resolve to be the same, then we
1653 		 * ignore the match so that the real root gets used.
1654 		 * This avoids the problem of traversing "../.." to find the
1655 		 * root directory and never finding it, because "/" resolves
1656 		 * to the emulation root directory. This is expensive :-(
1657 		 */
1658 		NDINIT(&ndroot, LOOKUP, FOLLOW, UIO_SYSSPACE, prefix,
1659 		    td);
1660 
1661 		/* We shouldn't ever get an error from this namei(). */
1662 		error = namei(&ndroot);
1663 		if (error == 0) {
1664 			if (nd.ni_vp == ndroot.ni_vp)
1665 				error = ENOENT;
1666 
1667 			NDFREE(&ndroot, NDF_ONLY_PNBUF);
1668 			vrele(ndroot.ni_vp);
1669 		}
1670 	}
1671 
1672 	NDFREE(&nd, NDF_ONLY_PNBUF);
1673 	vrele(nd.ni_vp);
1674 
1675 keeporig:
1676 	/* If there was an error, use the original path name. */
1677 	if (error)
1678 		bcopy(ptr, buf, len);
1679 	return (error);
1680 }
1681