xref: /freebsd/sys/kern/vfs_lookup.c (revision 7587f6d4840f8d363e457cddc14c184cf1fe7cc1)
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 
37 #include <sys/cdefs.h>
38 #include "opt_capsicum.h"
39 #include "opt_ktrace.h"
40 
41 #include <sys/param.h>
42 #include <sys/systm.h>
43 #include <sys/dirent.h>
44 #include <sys/kernel.h>
45 #include <sys/capsicum.h>
46 #include <sys/fcntl.h>
47 #include <sys/jail.h>
48 #include <sys/lock.h>
49 #include <sys/mutex.h>
50 #include <sys/namei.h>
51 #include <sys/vnode.h>
52 #include <sys/mount.h>
53 #include <sys/filedesc.h>
54 #include <sys/proc.h>
55 #include <sys/sdt.h>
56 #include <sys/syscallsubr.h>
57 #include <sys/sysctl.h>
58 #ifdef KTRACE
59 #include <sys/ktrace.h>
60 #endif
61 #ifdef INVARIANTS
62 #include <machine/_inttypes.h>
63 #endif
64 
65 #include <security/audit/audit.h>
66 #include <security/mac/mac_framework.h>
67 
68 #include <vm/uma.h>
69 
70 #ifdef INVARIANTS
71 static void NDVALIDATE_impl(struct nameidata *, int);
72 #define NDVALIDATE(ndp) NDVALIDATE_impl(ndp, __LINE__)
73 #else
74 #define NDVALIDATE(ndp)
75 #endif
76 
77 /*
78  * Prepare namei() to restart. Reset components to its original state and set
79  * ISRESTARTED flag which signals the underlying lookup code to change the root
80  * from ABI root to actual root and prevents a further restarts.
81  */
82 #define	NDRESTART(ndp) do {						\
83 	NDREINIT_DBG(ndp);						\
84 	ndp->ni_resflags = 0;						\
85 	ndp->ni_cnd.cn_flags &= ~NAMEI_INTERNAL_FLAGS;			\
86 	ndp->ni_cnd.cn_flags |= ISRESTARTED;				\
87 } while (0)
88 
89 #ifdef KTRACE
90 #define	NIKTRCAPFAIL(path)	ktrcapfail(CAPFAIL_NAMEI, (path))
91 #else
92 #define	NIKTRCAPFAIL(path)
93 #endif
94 
95 #define	NI_CAP_VIOLATION(ndp, path)	do {			\
96 	NIKTRCAPFAIL(path);					\
97 	(ndp)->ni_lcf &= ~NI_LCF_KTR_FLAGS;			\
98 } while (0)
99 
100 SDT_PROVIDER_DEFINE(vfs);
101 SDT_PROBE_DEFINE4(vfs, namei, lookup, entry, "struct vnode *", "char *",
102     "unsigned long", "bool");
103 SDT_PROBE_DEFINE4(vfs, namei, lookup, return, "int", "struct vnode *", "bool",
104     "struct nameidata");
105 
106 /* Allocation zone for namei. */
107 uma_zone_t namei_zone;
108 
109 /* Placeholder vnode for mp traversal. */
110 static struct vnode *vp_crossmp;
111 
112 static int
crossmp_vop_islocked(struct vop_islocked_args * ap)113 crossmp_vop_islocked(struct vop_islocked_args *ap)
114 {
115 
116 	return (LK_SHARED);
117 }
118 
119 static int
crossmp_vop_lock1(struct vop_lock1_args * ap)120 crossmp_vop_lock1(struct vop_lock1_args *ap)
121 {
122 	struct vnode *vp;
123 	struct lock *lk __diagused;
124 	int flags;
125 
126 	vp = ap->a_vp;
127 	lk = vp->v_vnlock;
128 	flags = ap->a_flags;
129 
130 	KASSERT((flags & (LK_SHARED | LK_NOWAIT)) == (LK_SHARED | LK_NOWAIT),
131 	    ("%s: invalid lock request 0x%x for crossmp", __func__, flags));
132 
133 	if ((flags & LK_INTERLOCK) != 0)
134 		VI_UNLOCK(vp);
135 	LOCK_LOG_LOCK("SLOCK", &lk->lock_object, 0, 0, ap->a_file, ap->a_line);
136 	return (0);
137 }
138 
139 static int
crossmp_vop_unlock(struct vop_unlock_args * ap)140 crossmp_vop_unlock(struct vop_unlock_args *ap)
141 {
142 	struct vnode *vp;
143 	struct lock *lk __diagused;
144 
145 	vp = ap->a_vp;
146 	lk = vp->v_vnlock;
147 
148 	LOCK_LOG_LOCK("SUNLOCK", &lk->lock_object, 0, 0, LOCK_FILE,
149 	    LOCK_LINE);
150 	return (0);
151 }
152 
153 static struct vop_vector crossmp_vnodeops = {
154 	.vop_default =		&default_vnodeops,
155 	.vop_islocked =		crossmp_vop_islocked,
156 	.vop_lock1 =		crossmp_vop_lock1,
157 	.vop_unlock =		crossmp_vop_unlock,
158 };
159 /*
160  * VFS_VOP_VECTOR_REGISTER(crossmp_vnodeops) is not used here since the vnode
161  * gets allocated early. See nameiinit for the direct call below.
162  */
163 
164 struct nameicap_tracker {
165 	struct vnode *dp;
166 	TAILQ_ENTRY(nameicap_tracker) nm_link;
167 };
168 
169 /* Zone for cap mode tracker elements used for dotdot capability checks. */
170 MALLOC_DEFINE(M_NAMEITRACKER, "namei_tracker", "namei tracking for dotdot");
171 
172 static void
nameiinit(void * dummy __unused)173 nameiinit(void *dummy __unused)
174 {
175 
176 	namei_zone = uma_zcreate("NAMEI", MAXPATHLEN, NULL, NULL, NULL, NULL,
177 	    UMA_ALIGN_PTR, 0);
178 	vfs_vector_op_register(&crossmp_vnodeops);
179 	getnewvnode("crossmp", NULL, &crossmp_vnodeops, &vp_crossmp);
180 	vp_crossmp->v_state = VSTATE_CONSTRUCTED;
181 	vp_crossmp->v_irflag |= VIRF_CROSSMP;
182 }
183 SYSINIT(vfs, SI_SUB_VFS, SI_ORDER_SECOND, nameiinit, NULL);
184 
185 static int lookup_cap_dotdot = 1;
186 SYSCTL_INT(_vfs, OID_AUTO, lookup_cap_dotdot, CTLFLAG_RWTUN,
187     &lookup_cap_dotdot, 0,
188     "enables \"..\" components in path lookup in capability mode");
189 static int lookup_cap_dotdot_nonlocal = 1;
190 SYSCTL_INT(_vfs, OID_AUTO, lookup_cap_dotdot_nonlocal, CTLFLAG_RWTUN,
191     &lookup_cap_dotdot_nonlocal, 0,
192     "enables \"..\" components in path lookup in capability mode "
193     "on non-local mount");
194 
195 static void
nameicap_tracker_add(struct nameidata * ndp,struct vnode * dp)196 nameicap_tracker_add(struct nameidata *ndp, struct vnode *dp)
197 {
198 	struct nameicap_tracker *nt;
199 
200 	if ((ndp->ni_lcf & NI_LCF_CAP_DOTDOT) == 0 || dp->v_type != VDIR)
201 		return;
202 	nt = TAILQ_LAST(&ndp->ni_cap_tracker, nameicap_tracker_head);
203 	if (nt != NULL && nt->dp == dp)
204 		return;
205 	nt = malloc(sizeof(*nt), M_NAMEITRACKER, M_WAITOK);
206 	vhold(dp);
207 	nt->dp = dp;
208 	TAILQ_INSERT_TAIL(&ndp->ni_cap_tracker, nt, nm_link);
209 }
210 
211 static void
nameicap_cleanup_from(struct nameidata * ndp,struct nameicap_tracker * first)212 nameicap_cleanup_from(struct nameidata *ndp, struct nameicap_tracker *first)
213 {
214 	struct nameicap_tracker *nt, *nt1;
215 
216 	nt = first;
217 	TAILQ_FOREACH_FROM_SAFE(nt, &ndp->ni_cap_tracker, nm_link, nt1) {
218 		TAILQ_REMOVE(&ndp->ni_cap_tracker, nt, nm_link);
219 		vdrop(nt->dp);
220 		free(nt, M_NAMEITRACKER);
221 	}
222 }
223 
224 static void
nameicap_cleanup(struct nameidata * ndp)225 nameicap_cleanup(struct nameidata *ndp)
226 {
227 	KASSERT(TAILQ_EMPTY(&ndp->ni_cap_tracker) ||
228 	    (ndp->ni_lcf & NI_LCF_CAP_DOTDOT) != 0, ("not strictrelative"));
229 	nameicap_cleanup_from(ndp, NULL);
230 }
231 
232 /*
233  * For dotdot lookups in capability mode, only allow the component
234  * lookup to succeed if the resulting directory was already traversed
235  * during the operation.  This catches situations where already
236  * traversed directory is moved to different parent, and then we walk
237  * over it with dotdots.
238  *
239  * Also allow to force failure of dotdot lookups for non-local
240  * filesystems, where external agents might assist local lookups to
241  * escape the compartment.
242  */
243 static int
nameicap_check_dotdot(struct nameidata * ndp,struct vnode * dp)244 nameicap_check_dotdot(struct nameidata *ndp, struct vnode *dp)
245 {
246 	struct nameicap_tracker *nt;
247 	struct mount *mp;
248 
249 	if (dp == NULL || dp->v_type != VDIR || (ndp->ni_lcf &
250 	    NI_LCF_STRICTREL) == 0)
251 		return (0);
252 	if (__predict_false((ndp->ni_lcf & (NI_LCF_STRICTREL_KTR |
253 	    NI_LCF_CAP_DOTDOT_KTR)) == NI_LCF_STRICTREL_KTR))
254 		NI_CAP_VIOLATION(ndp, ndp->ni_cnd.cn_pnbuf);
255 	if ((ndp->ni_lcf & NI_LCF_CAP_DOTDOT) == 0)
256 		return (ENOTCAPABLE);
257 	mp = dp->v_mount;
258 	if (lookup_cap_dotdot_nonlocal == 0 && mp != NULL &&
259 	    (mp->mnt_flag & MNT_LOCAL) == 0)
260 		goto capfail;
261 	TAILQ_FOREACH_REVERSE(nt, &ndp->ni_cap_tracker, nameicap_tracker_head,
262 	    nm_link) {
263 		if (dp == nt->dp) {
264 			nt = TAILQ_NEXT(nt, nm_link);
265 			if (nt != NULL)
266 				nameicap_cleanup_from(ndp, nt);
267 			return (0);
268 		}
269 	}
270 
271 capfail:
272 	if (__predict_false((ndp->ni_lcf & NI_LCF_STRICTREL_KTR) != 0))
273 		NI_CAP_VIOLATION(ndp, ndp->ni_cnd.cn_pnbuf);
274 	return (ENOTCAPABLE);
275 }
276 
277 static void
namei_cleanup_cnp(struct componentname * cnp)278 namei_cleanup_cnp(struct componentname *cnp)
279 {
280 
281 	uma_zfree(namei_zone, cnp->cn_pnbuf);
282 	cnp->cn_pnbuf = NULL;
283 	cnp->cn_nameptr = NULL;
284 }
285 
286 static int
namei_handle_root(struct nameidata * ndp,struct vnode ** dpp)287 namei_handle_root(struct nameidata *ndp, struct vnode **dpp)
288 {
289 	struct componentname *cnp;
290 
291 	cnp = &ndp->ni_cnd;
292 	if (__predict_false((ndp->ni_lcf & (NI_LCF_STRICTREL |
293 	    NI_LCF_STRICTREL_KTR)) != 0)) {
294 		if ((ndp->ni_lcf & NI_LCF_STRICTREL_KTR) != 0)
295 			NI_CAP_VIOLATION(ndp, cnp->cn_pnbuf);
296 		if ((ndp->ni_lcf & NI_LCF_STRICTREL) != 0)
297 			return (ENOTCAPABLE);
298 	}
299 	while (*(cnp->cn_nameptr) == '/') {
300 		cnp->cn_nameptr++;
301 		ndp->ni_pathlen--;
302 	}
303 	*dpp = ndp->ni_rootdir;
304 	vrefact(*dpp);
305 	return (0);
306 }
307 
308 static int
namei_setup(struct nameidata * ndp,struct vnode ** dpp,struct pwd ** pwdp)309 namei_setup(struct nameidata *ndp, struct vnode **dpp, struct pwd **pwdp)
310 {
311 	struct componentname *cnp;
312 	struct thread *td;
313 	struct pwd *pwd;
314 	int error;
315 	bool startdir_used;
316 
317 	cnp = &ndp->ni_cnd;
318 	td = curthread;
319 
320 	startdir_used = false;
321 	*pwdp = NULL;
322 	*dpp = NULL;
323 
324 #ifdef CAPABILITY_MODE
325 	/*
326 	 * In capability mode, lookups must be restricted to happen in
327 	 * the subtree with the root specified by the file descriptor:
328 	 * - The root must be real file descriptor, not the pseudo-descriptor
329 	 *   AT_FDCWD.
330 	 * - The passed path must be relative and not absolute.
331 	 * - If lookup_cap_dotdot is disabled, path must not contain the
332 	 *   '..' components.
333 	 * - If lookup_cap_dotdot is enabled, we verify that all '..'
334 	 *   components lookups result in the directories which were
335 	 *   previously walked by us, which prevents an escape from
336 	 *   the relative root.
337 	 */
338 	if ((cnp->cn_flags & NOCAPCHECK) == 0) {
339 		if (CAP_TRACING(td)) {
340 			ndp->ni_lcf |= NI_LCF_STRICTREL_KTR;
341 			if (ndp->ni_dirfd == AT_FDCWD)
342 				NI_CAP_VIOLATION(ndp, "AT_FDCWD");
343 		}
344 		if (IN_CAPABILITY_MODE(td)) {
345 			ndp->ni_lcf |= NI_LCF_STRICTREL;
346 			ndp->ni_resflags |= NIRES_STRICTREL;
347 			if (ndp->ni_dirfd == AT_FDCWD)
348 				return (ECAPMODE);
349 		}
350 	}
351 #endif
352 	error = 0;
353 
354 	/*
355 	 * Get starting point for the translation.
356 	 */
357 	pwd = pwd_hold(td);
358 	/*
359 	 * The reference on ni_rootdir is acquired in the block below to avoid
360 	 * back-to-back atomics for absolute lookups.
361 	 */
362 	namei_setup_rootdir(ndp, cnp, pwd);
363 	ndp->ni_topdir = pwd->pwd_jdir;
364 
365 	if (cnp->cn_pnbuf[0] == '/') {
366 		ndp->ni_resflags |= NIRES_ABS;
367 		error = namei_handle_root(ndp, dpp);
368 	} else {
369 		if (ndp->ni_startdir != NULL) {
370 			*dpp = ndp->ni_startdir;
371 			startdir_used = true;
372 		} else if (ndp->ni_dirfd == AT_FDCWD) {
373 			*dpp = pwd->pwd_cdir;
374 			vrefact(*dpp);
375 		} else {
376 			if (cnp->cn_flags & AUDITVNODE1)
377 				AUDIT_ARG_ATFD1(ndp->ni_dirfd);
378 			if (cnp->cn_flags & AUDITVNODE2)
379 				AUDIT_ARG_ATFD2(ndp->ni_dirfd);
380 
381 			error = fgetvp_lookup(ndp, dpp);
382 		}
383 		if (error == 0 && (*dpp)->v_type != VDIR &&
384 		    (cnp->cn_flags & OPENNAMED) == 0 &&
385 		    (cnp->cn_pnbuf[0] != '\0' ||
386 		    (cnp->cn_flags & EMPTYPATH) == 0))
387 			error = ENOTDIR;
388 	}
389 	if (error == 0 && (cnp->cn_flags & RBENEATH) != 0) {
390 		if (cnp->cn_pnbuf[0] == '/') {
391 			error = ENOTCAPABLE;
392 		} else if ((ndp->ni_lcf & NI_LCF_STRICTREL) == 0) {
393 			ndp->ni_lcf |= NI_LCF_STRICTREL |
394 			    NI_LCF_CAP_DOTDOT;
395 		}
396 	}
397 
398 	/*
399 	 * If we are auditing the kernel pathname, save the user pathname.
400 	 */
401 	if (AUDITING_TD(td)) {
402 		if (cnp->cn_flags & AUDITVNODE1)
403 			AUDIT_ARG_UPATH1_VP(td, ndp->ni_rootdir, *dpp, cnp->cn_pnbuf);
404 		if (cnp->cn_flags & AUDITVNODE2)
405 			AUDIT_ARG_UPATH2_VP(td, ndp->ni_rootdir, *dpp, cnp->cn_pnbuf);
406 	}
407 	if (ndp->ni_startdir != NULL && !startdir_used)
408 		vrele(ndp->ni_startdir);
409 	if (error != 0) {
410 		if (*dpp != NULL)
411 			vrele(*dpp);
412 		pwd_drop(pwd);
413 		return (error);
414 	}
415 	if (lookup_cap_dotdot != 0) {
416 		if ((ndp->ni_lcf & NI_LCF_STRICTREL_KTR) != 0)
417 			ndp->ni_lcf |= NI_LCF_CAP_DOTDOT_KTR;
418 		if ((ndp->ni_lcf & NI_LCF_STRICTREL) != 0)
419 			ndp->ni_lcf |= NI_LCF_CAP_DOTDOT;
420 	}
421 	SDT_PROBE4(vfs, namei, lookup, entry, *dpp, cnp->cn_pnbuf,
422 	    cnp->cn_flags, false);
423 	*pwdp = pwd;
424 	return (0);
425 }
426 
427 static int
namei_getpath(struct nameidata * ndp)428 namei_getpath(struct nameidata *ndp)
429 {
430 	struct componentname *cnp;
431 	int error;
432 
433 	cnp = &ndp->ni_cnd;
434 
435 	/*
436 	 * Get a buffer for the name to be translated, and copy the
437 	 * name into the buffer.
438 	 */
439 	cnp->cn_pnbuf = uma_zalloc(namei_zone, M_WAITOK);
440 	if (ndp->ni_segflg == UIO_SYSSPACE) {
441 		error = copystr(ndp->ni_dirp, cnp->cn_pnbuf, MAXPATHLEN,
442 		    &ndp->ni_pathlen);
443 	} else {
444 		error = copyinstr(ndp->ni_dirp, cnp->cn_pnbuf, MAXPATHLEN,
445 		    &ndp->ni_pathlen);
446 	}
447 
448 	return (error);
449 }
450 
451 static int
namei_emptypath(struct nameidata * ndp)452 namei_emptypath(struct nameidata *ndp)
453 {
454 	struct componentname *cnp;
455 	struct pwd *pwd;
456 	struct vnode *dp;
457 	int error;
458 
459 	cnp = &ndp->ni_cnd;
460 	MPASS(*cnp->cn_pnbuf == '\0');
461 	MPASS((cnp->cn_flags & EMPTYPATH) != 0);
462 	MPASS((cnp->cn_flags & (LOCKPARENT | WANTPARENT)) == 0);
463 
464 	ndp->ni_resflags |= NIRES_EMPTYPATH;
465 	error = namei_setup(ndp, &dp, &pwd);
466 	if (error != 0) {
467 		goto errout;
468 	}
469 
470 	/*
471 	 * Usecount on dp already provided by namei_setup.
472 	 */
473 	ndp->ni_vp = dp;
474 	pwd_drop(pwd);
475 	NDVALIDATE(ndp);
476 	if ((cnp->cn_flags & LOCKLEAF) != 0) {
477 		VOP_LOCK(dp, (cnp->cn_flags & LOCKSHARED) != 0 ?
478 		    LK_SHARED : LK_EXCLUSIVE);
479 		if (VN_IS_DOOMED(dp)) {
480 			vput(dp);
481 			error = ENOENT;
482 			goto errout;
483 		}
484 	}
485 	SDT_PROBE4(vfs, namei, lookup, return, 0, ndp->ni_vp, false, ndp);
486 	return (0);
487 
488 errout:
489 	SDT_PROBE4(vfs, namei, lookup, return, error, NULL, false, ndp);
490 	namei_cleanup_cnp(cnp);
491 	return (error);
492 }
493 
494 static int __noinline
namei_follow_link(struct nameidata * ndp)495 namei_follow_link(struct nameidata *ndp)
496 {
497 	char *cp;
498 	struct iovec aiov;
499 	struct uio auio;
500 	struct componentname *cnp;
501 	struct thread *td;
502 	int error, linklen;
503 
504 	error = 0;
505 	cnp = &ndp->ni_cnd;
506 	td = curthread;
507 
508 	if (ndp->ni_loopcnt++ >= MAXSYMLINKS) {
509 		error = ELOOP;
510 		goto out;
511 	}
512 #ifdef MAC
513 	if ((cnp->cn_flags & NOMACCHECK) == 0) {
514 		error = mac_vnode_check_readlink(td->td_ucred, ndp->ni_vp);
515 		if (error != 0)
516 			goto out;
517 	}
518 #endif
519 	if (ndp->ni_pathlen > 1)
520 		cp = uma_zalloc(namei_zone, M_WAITOK);
521 	else
522 		cp = cnp->cn_pnbuf;
523 	aiov.iov_base = cp;
524 	aiov.iov_len = MAXPATHLEN;
525 	auio.uio_iov = &aiov;
526 	auio.uio_iovcnt = 1;
527 	auio.uio_offset = 0;
528 	auio.uio_rw = UIO_READ;
529 	auio.uio_segflg = UIO_SYSSPACE;
530 	auio.uio_td = td;
531 	auio.uio_resid = MAXPATHLEN;
532 	error = VOP_READLINK(ndp->ni_vp, &auio, cnp->cn_cred);
533 	if (error != 0) {
534 		if (ndp->ni_pathlen > 1)
535 			uma_zfree(namei_zone, cp);
536 		goto out;
537 	}
538 	linklen = MAXPATHLEN - auio.uio_resid;
539 	if (linklen == 0) {
540 		if (ndp->ni_pathlen > 1)
541 			uma_zfree(namei_zone, cp);
542 		error = ENOENT;
543 		goto out;
544 	}
545 	if (linklen + ndp->ni_pathlen > MAXPATHLEN) {
546 		if (ndp->ni_pathlen > 1)
547 			uma_zfree(namei_zone, cp);
548 		error = ENAMETOOLONG;
549 		goto out;
550 	}
551 	if (ndp->ni_pathlen > 1) {
552 		bcopy(ndp->ni_next, cp + linklen, ndp->ni_pathlen);
553 		uma_zfree(namei_zone, cnp->cn_pnbuf);
554 		cnp->cn_pnbuf = cp;
555 	} else
556 		cnp->cn_pnbuf[linklen] = '\0';
557 	ndp->ni_pathlen += linklen;
558 out:
559 	return (error);
560 }
561 
562 /*
563  * Convert a pathname into a pointer to a locked vnode.
564  *
565  * The FOLLOW flag is set when symbolic links are to be followed
566  * when they occur at the end of the name translation process.
567  * Symbolic links are always followed for all other pathname
568  * components other than the last.
569  *
570  * The segflg defines whether the name is to be copied from user
571  * space or kernel space.
572  *
573  * Overall outline of namei:
574  *
575  *	copy in name
576  *	get starting directory
577  *	while (!done && !error) {
578  *		call lookup to search path.
579  *		if symbolic link, massage name in buffer and continue
580  *	}
581  */
582 int
namei(struct nameidata * ndp)583 namei(struct nameidata *ndp)
584 {
585 	struct vnode *dp;	/* the directory we are searching */
586 	struct componentname *cnp;
587 	struct thread *td;
588 	struct pwd *pwd;
589 	int error;
590 	enum cache_fpl_status status;
591 
592 	cnp = &ndp->ni_cnd;
593 	td = curthread;
594 #ifdef INVARIANTS
595 	KASSERT((ndp->ni_debugflags & NAMEI_DBG_CALLED) == 0,
596 	    ("%s: repeated call to namei without NDREINIT", __func__));
597 	KASSERT(ndp->ni_debugflags == NAMEI_DBG_INITED,
598 	    ("%s: bad debugflags %d", __func__, ndp->ni_debugflags));
599 	ndp->ni_debugflags |= NAMEI_DBG_CALLED;
600 	if (ndp->ni_startdir != NULL)
601 		ndp->ni_debugflags |= NAMEI_DBG_HADSTARTDIR;
602 	if (cnp->cn_flags & FAILIFEXISTS) {
603 		KASSERT(cnp->cn_nameiop == CREATE,
604 		    ("%s: FAILIFEXISTS passed for op %d", __func__, cnp->cn_nameiop));
605 		/*
606 		 * The limitation below is to restrict hairy corner cases.
607 		 */
608 		KASSERT((cnp->cn_flags & (LOCKPARENT | LOCKLEAF)) == LOCKPARENT,
609 		    ("%s: FAILIFEXISTS must be passed with LOCKPARENT and without LOCKLEAF",
610 		    __func__));
611 	}
612 #endif
613 	ndp->ni_cnd.cn_cred = td->td_ucred;
614 	KASSERT(ndp->ni_resflags == 0, ("%s: garbage in ni_resflags: %x",
615 	    __func__, ndp->ni_resflags));
616 	KASSERT(cnp->cn_cred && td->td_proc, ("namei: bad cred/proc"));
617 	KASSERT((cnp->cn_flags & NAMEI_INTERNAL_FLAGS) == 0,
618 	    ("namei: unexpected flags: %#jx",
619 	    (uintmax_t)(cnp->cn_flags & NAMEI_INTERNAL_FLAGS)));
620 	if (cnp->cn_flags & NOCACHE)
621 		KASSERT(cnp->cn_nameiop != LOOKUP,
622 		    ("%s: NOCACHE passed with LOOKUP", __func__));
623 	MPASS(ndp->ni_startdir == NULL || ndp->ni_startdir->v_type == VDIR ||
624 	    ndp->ni_startdir->v_type == VBAD);
625 
626 restart:
627 	ndp->ni_lcf = 0;
628 	ndp->ni_loopcnt = 0;
629 	ndp->ni_vp = NULL;
630 
631 	error = namei_getpath(ndp);
632 	if (__predict_false(error != 0)) {
633 		namei_cleanup_cnp(cnp);
634 		SDT_PROBE4(vfs, namei, lookup, return, error, NULL,
635 		    false, ndp);
636 		return (error);
637 	}
638 
639 	cnp->cn_nameptr = cnp->cn_pnbuf;
640 
641 #ifdef KTRACE
642 	if (KTRPOINT(td, KTR_NAMEI)) {
643 		ktrnamei(cnp->cn_pnbuf);
644 	}
645 #endif
646 	TSNAMEI(curthread->td_proc->p_pid, cnp->cn_pnbuf);
647 
648 	/*
649 	 * First try looking up the target without locking any vnodes.
650 	 *
651 	 * We may need to start from scratch or pick up where it left off.
652 	 */
653 	error = cache_fplookup(ndp, &status, &pwd);
654 	switch (status) {
655 	case CACHE_FPL_STATUS_UNSET:
656 		__assert_unreachable();
657 		break;
658 	case CACHE_FPL_STATUS_HANDLED:
659 		if (error == 0)
660 			NDVALIDATE(ndp);
661 		else if (__predict_false(pwd->pwd_adir != pwd->pwd_rdir &&
662 		    (cnp->cn_flags & ISRESTARTED) == 0)) {
663 			namei_cleanup_cnp(cnp);
664 			NDRESTART(ndp);
665 			goto restart;
666 		}
667 		return (error);
668 	case CACHE_FPL_STATUS_PARTIAL:
669 		TAILQ_INIT(&ndp->ni_cap_tracker);
670 		dp = ndp->ni_startdir;
671 		break;
672 	case CACHE_FPL_STATUS_DESTROYED:
673 		ndp->ni_loopcnt = 0;
674 		error = namei_getpath(ndp);
675 		if (__predict_false(error != 0)) {
676 			namei_cleanup_cnp(cnp);
677 			return (error);
678 		}
679 		cnp->cn_nameptr = cnp->cn_pnbuf;
680 		/* FALLTHROUGH */
681 	case CACHE_FPL_STATUS_ABORTED:
682 		TAILQ_INIT(&ndp->ni_cap_tracker);
683 		MPASS(ndp->ni_lcf == 0);
684 		if (*cnp->cn_pnbuf == '\0') {
685 			if ((cnp->cn_flags & EMPTYPATH) != 0) {
686 				return (namei_emptypath(ndp));
687 			}
688 			namei_cleanup_cnp(cnp);
689 			SDT_PROBE4(vfs, namei, lookup, return, ENOENT, NULL,
690 			    false, ndp);
691 			return (ENOENT);
692 		}
693 		error = namei_setup(ndp, &dp, &pwd);
694 		if (error != 0) {
695 			namei_cleanup_cnp(cnp);
696 			return (error);
697 		}
698 		break;
699 	}
700 
701 	/*
702 	 * Locked lookup.
703 	 */
704 	for (;;) {
705 		ndp->ni_startdir = dp;
706 		error = vfs_lookup(ndp);
707 		if (error != 0) {
708 			if (__predict_false(pwd->pwd_adir != pwd->pwd_rdir &&
709 			    error == ENOENT &&
710 			    (cnp->cn_flags & ISRESTARTED) == 0)) {
711 				nameicap_cleanup(ndp);
712 				pwd_drop(pwd);
713 				namei_cleanup_cnp(cnp);
714 				NDRESTART(ndp);
715 				goto restart;
716 			} else
717 				goto out;
718 		}
719 
720 		/*
721 		 * If not a symbolic link, we're done.
722 		 */
723 		if ((cnp->cn_flags & ISSYMLINK) == 0) {
724 			SDT_PROBE4(vfs, namei, lookup, return, error,
725 			    ndp->ni_vp, false, ndp);
726 			nameicap_cleanup(ndp);
727 			pwd_drop(pwd);
728 			NDVALIDATE(ndp);
729 			return (0);
730 		}
731 		error = namei_follow_link(ndp);
732 		if (error != 0)
733 			break;
734 		vput(ndp->ni_vp);
735 		dp = ndp->ni_dvp;
736 		/*
737 		 * Check if root directory should replace current directory.
738 		 */
739 		cnp->cn_nameptr = cnp->cn_pnbuf;
740 		if (*(cnp->cn_nameptr) == '/') {
741 			/*
742 			 * Reset the lookup to start from the real root without
743 			 * origin path name reloading.
744 			 */
745 			if (__predict_false(ndp->ni_rootdir != pwd->pwd_rdir)) {
746 				cnp->cn_flags |= ISRESTARTED;
747 				ndp->ni_rootdir = pwd->pwd_rdir;
748 			}
749 			vrele(dp);
750 			error = namei_handle_root(ndp, &dp);
751 			if (error != 0)
752 				goto out;
753 		}
754 	}
755 	vput(ndp->ni_vp);
756 	ndp->ni_vp = NULL;
757 	vrele(ndp->ni_dvp);
758 out:
759 	MPASS(error != 0);
760 	SDT_PROBE4(vfs, namei, lookup, return, error, NULL, false, ndp);
761 	namei_cleanup_cnp(cnp);
762 	nameicap_cleanup(ndp);
763 	pwd_drop(pwd);
764 	return (error);
765 }
766 
767 static int
enforce_lkflags(struct mount * mp,int lkflags)768 enforce_lkflags(struct mount *mp, int lkflags)
769 {
770 
771 	if (mp == NULL || ((lkflags & LK_SHARED) &&
772 	    !(mp->mnt_kern_flag & MNTK_LOOKUP_SHARED))) {
773 		lkflags &= ~LK_SHARED;
774 		lkflags |= LK_EXCLUSIVE;
775 	}
776 	lkflags |= LK_NODDLKTREAT;
777 	return (lkflags);
778 }
779 
780 static __inline int
needs_exclusive_leaf(struct mount * mp,int flags)781 needs_exclusive_leaf(struct mount *mp, int flags)
782 {
783 
784 	/*
785 	 * Intermediate nodes can use shared locks, we only need to
786 	 * force an exclusive lock for leaf nodes.
787 	 */
788 	if ((flags & (ISLASTCN | LOCKLEAF)) != (ISLASTCN | LOCKLEAF))
789 		return (0);
790 
791 	/* Always use exclusive locks if LOCKSHARED isn't set. */
792 	if (!(flags & LOCKSHARED))
793 		return (1);
794 
795 	/*
796 	 * For lookups during open(), if the mount point supports
797 	 * extended shared operations, then use a shared lock for the
798 	 * leaf node, otherwise use an exclusive lock.
799 	 */
800 	if ((flags & ISOPEN) != 0)
801 		return (!MNT_EXTENDED_SHARED(mp));
802 
803 	/*
804 	 * Lookup requests outside of open() that specify LOCKSHARED
805 	 * only need a shared lock on the leaf vnode.
806 	 */
807 	return (0);
808 }
809 
810 /*
811  * Various filesystems expect to be able to copy a name component with length
812  * bounded by NAME_MAX into a directory entry buffer of size MAXNAMLEN.  Make
813  * sure that these are the same size.
814  */
815 _Static_assert(MAXNAMLEN == NAME_MAX,
816     "MAXNAMLEN and NAME_MAX have different values");
817 
818 static int __noinline
vfs_lookup_degenerate(struct nameidata * ndp,struct vnode * dp,int wantparent)819 vfs_lookup_degenerate(struct nameidata *ndp, struct vnode *dp, int wantparent)
820 {
821 	struct componentname *cnp;
822 	struct mount *mp;
823 	int error;
824 
825 	cnp = &ndp->ni_cnd;
826 
827 	cnp->cn_flags |= ISLASTCN;
828 
829 	mp = atomic_load_ptr(&dp->v_mount);
830 	if (needs_exclusive_leaf(mp, cnp->cn_flags)) {
831 		cnp->cn_lkflags &= ~LK_SHARED;
832 		cnp->cn_lkflags |= LK_EXCLUSIVE;
833 	}
834 
835 	vn_lock(dp, enforce_lkflags(mp, cnp->cn_lkflags | LK_RETRY));
836 
837 	if (dp->v_type != VDIR) {
838 		error = ENOTDIR;
839 		goto bad;
840 	}
841 	if (cnp->cn_nameiop != LOOKUP) {
842 		error = EISDIR;
843 		goto bad;
844 	}
845 	if (wantparent) {
846 		ndp->ni_dvp = dp;
847 		VREF(dp);
848 	}
849 	ndp->ni_vp = dp;
850 	cnp->cn_namelen = 0;
851 
852 	if (cnp->cn_flags & AUDITVNODE1)
853 		AUDIT_ARG_VNODE1(dp);
854 	else if (cnp->cn_flags & AUDITVNODE2)
855 		AUDIT_ARG_VNODE2(dp);
856 
857 	if (!(cnp->cn_flags & (LOCKPARENT | LOCKLEAF)))
858 		VOP_UNLOCK(dp);
859 	return (0);
860 bad:
861 	VOP_UNLOCK(dp);
862 	return (error);
863 }
864 
865 struct nameidata *
vfs_lookup_nameidata(struct componentname * cnp)866 vfs_lookup_nameidata(struct componentname *cnp)
867 {
868 	if ((cnp->cn_flags & NAMEILOOKUP) == 0)
869 		return (NULL);
870 	return (__containerof(cnp, struct nameidata, ni_cnd));
871 }
872 
873 /*
874  * Would a dotdot lookup relative to dvp cause this lookup to cross a jail or
875  * chroot boundary?
876  */
877 bool
vfs_lookup_isroot(struct nameidata * ndp,struct vnode * dvp)878 vfs_lookup_isroot(struct nameidata *ndp, struct vnode *dvp)
879 {
880 	for (struct prison *pr = ndp->ni_cnd.cn_cred->cr_prison; pr != NULL;
881 	    pr = pr->pr_parent) {
882 		if (dvp == pr->pr_root)
883 			return (true);
884 	}
885 	return (dvp == ndp->ni_rootdir || dvp == ndp->ni_topdir ||
886 	    dvp == rootvnode);
887 }
888 
889 /*
890  * FAILIFEXISTS handling.
891  *
892  * XXX namei called with LOCKPARENT but not LOCKLEAF has the strange
893  * behaviour of leaving the vnode unlocked if the target is the same
894  * vnode as the parent.
895  */
896 static int __noinline
vfs_lookup_failifexists(struct nameidata * ndp)897 vfs_lookup_failifexists(struct nameidata *ndp)
898 {
899 	struct componentname *cnp __diagused;
900 
901 	cnp = &ndp->ni_cnd;
902 
903 	MPASS((cnp->cn_flags & ISSYMLINK) == 0);
904 	if (ndp->ni_vp == ndp->ni_dvp)
905 		vrele(ndp->ni_dvp);
906 	else
907 		vput(ndp->ni_dvp);
908 	vrele(ndp->ni_vp);
909 	ndp->ni_dvp = NULL;
910 	ndp->ni_vp = NULL;
911 	NDFREE_PNBUF(ndp);
912 	return (EEXIST);
913 }
914 
915 static int __noinline
vfs_lookup_cross_mount(struct nameidata * ndp)916 vfs_lookup_cross_mount(struct nameidata *ndp)
917 {
918 	struct componentname *cnp;
919 	struct mount *mp;
920 	struct vnode *dp, *tdp;
921 	int error, crosslkflags;
922 	bool crosslock;
923 
924 	cnp = &ndp->ni_cnd;
925 	dp = ndp->ni_vp;
926 
927 	/*
928 	 * The vnode has been mounted on, find the root of the mounted
929 	 * filesystem.
930 	 */
931 	do {
932 		mp = dp->v_mountedhere;
933 		ASSERT_VOP_LOCKED(dp, __func__);
934 		VNPASS((vn_irflag_read(dp) & VIRF_MOUNTPOINT) != 0 && mp != NULL, dp);
935 
936 		crosslock = (dp->v_vflag & VV_CROSSLOCK) != 0;
937 		crosslkflags = enforce_lkflags(mp, cnp->cn_lkflags);
938 		if (__predict_false(crosslock)) {
939 			/*
940 			 * We are going to be holding the vnode lock, which
941 			 * in this case is shared by the root vnode of the
942 			 * filesystem mounted at mp, across the call to
943 			 * VFS_ROOT().  Make the situation clear to the
944 			 * filesystem by passing LK_CANRECURSE if the
945 			 * lock is held exclusive, or by clearinng
946 			 * LK_NODDLKTREAT to allow recursion on the shared
947 			 * lock in the presence of an exclusive waiter.
948 			 */
949 			if (VOP_ISLOCKED(dp) == LK_EXCLUSIVE) {
950 				crosslkflags &= ~LK_SHARED;
951 				crosslkflags |= LK_EXCLUSIVE | LK_CANRECURSE;
952 			} else if ((crosslkflags & LK_EXCLUSIVE) != 0) {
953 				error = vn_lock(dp, LK_UPGRADE);
954 				if (error != 0) {
955 					MPASS(error == ENOENT);
956 					vrele(dp);
957 					if (dp != ndp->ni_dvp)
958 						vput(ndp->ni_dvp);
959 					else
960 						vrele(ndp->ni_dvp);
961 					break;
962 				}
963 				if (dp->v_mountedhere != mp) {
964 					/*
965 					 * Note that we rely on the
966 					 * VIRF_MOUNTPOINT loop condition to
967 					 * ensure we stop iterating if dp is
968 					 * no longer a mountpoint at all.
969 					 */
970 					continue;
971 				}
972 			} else
973 				crosslkflags &= ~LK_NODDLKTREAT;
974 		}
975 		if (vfs_busy(mp, 0) != 0)
976 			continue;
977 		if (__predict_true(!crosslock))
978 			vput(dp);
979 		if (dp != ndp->ni_dvp)
980 			vput(ndp->ni_dvp);
981 		else
982 			vrele(ndp->ni_dvp);
983 		vrefact(vp_crossmp);
984 		ndp->ni_dvp = vp_crossmp;
985 		error = VFS_ROOT(mp, crosslkflags, &tdp);
986 		vfs_unbusy(mp);
987 		if (__predict_false(crosslock))
988 			vput(dp);
989 		if (vn_lock(vp_crossmp, LK_SHARED | LK_NOWAIT))
990 			panic("vp_crossmp exclusively locked or reclaimed");
991 		if (error != 0)
992 			break;
993 		ndp->ni_vp = dp = tdp;
994 	} while ((vn_irflag_read(dp) & VIRF_MOUNTPOINT) != 0);
995 
996 	return (error);
997 }
998 
999 /*
1000  * Search a pathname.
1001  * This is a very central and rather complicated routine.
1002  *
1003  * The pathname is pointed to by cn_nameptr and is of length ni_pathlen.
1004  * The starting directory is taken from ni_startdir. The pathname is
1005  * descended until done, or a symbolic link is encountered. The cn_flags
1006  * has ISLASTCN or'ed if the path is completed or ISSYMLINK or'ed if a
1007  * symbolic link needing interpretation is encountered.
1008  *
1009  * The cn_nameiop is LOOKUP, CREATE, RENAME, or DELETE depending on
1010  * whether the name is to be looked up, created, renamed, or deleted.
1011  * When CREATE, RENAME, or DELETE is specified, information usable in
1012  * creating, renaming, or deleting a directory entry may be calculated.
1013  * If cn_flags has LOCKPARENT or'ed into it, the parent directory is returned
1014  * locked. If it has WANTPARENT or'ed into it, the parent directory is
1015  * returned unlocked. Otherwise the parent directory is not returned. If
1016  * the target of the pathname exists and LOCKLEAF is or'ed into the cn_flags
1017  * the target is returned locked, otherwise it is returned unlocked.
1018  *
1019  * Overall outline of lookup:
1020  *
1021  *	handle degenerate case where name is null string
1022  *
1023  * dirloop:
1024  *	identify next component of name at ndp->ni_cnd.cn_nameptr
1025  *	handle .. special cases related to capabilities, chroot, jail
1026  *	if .. and crossing mount points and on mounted filesys, find parent
1027  *	call VOP_LOOKUP routine for next component name
1028  *	    directory vnode returned in ni_dvp, unlocked unless LOCKPARENT set
1029  *	    component vnode returned in ni_vp (if it exists), locked.
1030  *	if result vnode is mounted on and crossing mount points,
1031  *	    find mounted on vnode
1032  *	if more components of name, do next level at dirloop
1033  *	if VOP_LOOKUP returns ERELOOKUP, repeat the same level at dirloop
1034  *	return the answer in ni_vp, locked if LOCKLEAF set
1035  *	    if LOCKPARENT set, return locked parent in ni_dvp
1036  *	    if WANTPARENT set, return unlocked parent in ni_dvp
1037  */
1038 int
vfs_lookup(struct nameidata * ndp)1039 vfs_lookup(struct nameidata *ndp)
1040 {
1041 	char *cp;			/* pointer into pathname argument */
1042 	char *prev_ni_next;		/* saved ndp->ni_next */
1043 	char *nulchar;			/* location of '\0' in cn_pnbuf */
1044 	char *lastchar;			/* location of the last character */
1045 	struct vnode *dp = NULL;	/* the directory we are searching */
1046 	struct vnode *tdp;		/* saved dp */
1047 	size_t prev_ni_pathlen;		/* saved ndp->ni_pathlen */
1048 	int docache;			/* == 0 do not cache last component */
1049 	int wantparent;			/* 1 => wantparent or lockparent flag */
1050 	int rdonly;			/* lookup read-only flag bit */
1051 	int error = 0;
1052 	int relookup = 0;		/* do not consume the path component */
1053 	struct componentname *cnp = &ndp->ni_cnd;
1054 	int lkflags_save;
1055 	int ni_dvp_unlocked;
1056 
1057 	/*
1058 	 * Setup: break out flag bits into variables.
1059 	 */
1060 	ni_dvp_unlocked = 0;
1061 	wantparent = cnp->cn_flags & (LOCKPARENT | WANTPARENT);
1062 	KASSERT(cnp->cn_nameiop == LOOKUP || wantparent,
1063 	    ("CREATE, DELETE, RENAME require LOCKPARENT or WANTPARENT."));
1064 	/*
1065 	 * When set to zero, docache causes the last component of the
1066 	 * pathname to be deleted from the cache and the full lookup
1067 	 * of the name to be done (via VOP_CACHEDLOOKUP()). Often
1068 	 * filesystems need some pre-computed values that are made
1069 	 * during the full lookup, for instance UFS sets dp->i_offset.
1070 	 *
1071 	 * The docache variable is set to zero when requested by the
1072 	 * NOCACHE flag and for all modifying operations except CREATE.
1073 	 */
1074 	docache = (cnp->cn_flags & NOCACHE) ^ NOCACHE;
1075 	if (cnp->cn_nameiop == DELETE ||
1076 	    (wantparent && cnp->cn_nameiop != CREATE &&
1077 	     cnp->cn_nameiop != LOOKUP))
1078 		docache = 0;
1079 	rdonly = cnp->cn_flags & RDONLY;
1080 	cnp->cn_flags &= ~ISSYMLINK;
1081 	ndp->ni_dvp = NULL;
1082 
1083 	cnp->cn_lkflags = LK_SHARED;
1084 	dp = ndp->ni_startdir;
1085 	ndp->ni_startdir = NULLVP;
1086 
1087 	/*
1088 	 * Leading slashes, if any, are supposed to be skipped by the caller.
1089 	 */
1090 	MPASS(cnp->cn_nameptr[0] != '/');
1091 
1092 	/*
1093 	 * Check for degenerate name (e.g. / or "") which is a way of talking
1094 	 * about a directory, e.g. like "/." or ".".
1095 	 */
1096 	if (__predict_false(cnp->cn_nameptr[0] == '\0')) {
1097 		error = vfs_lookup_degenerate(ndp, dp, wantparent);
1098 		if (error == 0)
1099 			goto success_right_lock;
1100 		goto bad_unlocked;
1101 	}
1102 
1103 	/*
1104 	 * Nul-out trailing slashes (e.g., "foo///" -> "foo").
1105 	 *
1106 	 * This must be done before VOP_LOOKUP() because some fs's don't know
1107 	 * about trailing slashes.  Remember if there were trailing slashes to
1108 	 * handle symlinks, existing non-directories and non-existing files
1109 	 * that won't be directories specially later.
1110 	 */
1111 	MPASS(ndp->ni_pathlen >= 2);
1112 	lastchar = &cnp->cn_nameptr[ndp->ni_pathlen - 2];
1113 	if (*lastchar == '/') {
1114 		while (lastchar >= cnp->cn_pnbuf) {
1115 			*lastchar = '\0';
1116 			lastchar--;
1117 			ndp->ni_pathlen--;
1118 			if (*lastchar != '/') {
1119 				break;
1120 			}
1121 		}
1122 		cnp->cn_flags |= TRAILINGSLASH;
1123 	}
1124 
1125 	/*
1126 	 * We use shared locks until we hit the parent of the last cn then
1127 	 * we adjust based on the requesting flags.
1128 	 */
1129 	vn_lock(dp,
1130 	    enforce_lkflags(dp->v_mount, cnp->cn_lkflags | LK_RETRY));
1131 
1132 dirloop:
1133 	/*
1134 	 * Search a new directory.
1135 	 *
1136 	 * The last component of the filename is left accessible via
1137 	 * cnp->cn_nameptr. It has to be freed with a call to NDFREE*.
1138 	 *
1139 	 * Store / as a temporary sentinel so that we only have one character
1140 	 * to test for. Pathnames tend to be short so this should not be
1141 	 * resulting in cache misses.
1142 	 */
1143 	nulchar = &cnp->cn_nameptr[ndp->ni_pathlen - 1];
1144 	KASSERT(*nulchar == '\0',
1145 	    ("%s: expected nul at %p; string [%s]\n", __func__, nulchar,
1146 	    cnp->cn_pnbuf));
1147 	*nulchar = '/';
1148 	for (cp = cnp->cn_nameptr; *cp != '/'; cp++) {
1149 		KASSERT(*cp != '\0',
1150 		    ("%s: encountered unexpected nul; string [%s]\n", __func__,
1151 		    cnp->cn_nameptr));
1152 		continue;
1153 	}
1154 	*nulchar = '\0';
1155 	cnp->cn_namelen = cp - cnp->cn_nameptr;
1156 	if (__predict_false(cnp->cn_namelen > NAME_MAX)) {
1157 		error = ENAMETOOLONG;
1158 		goto bad;
1159 	}
1160 	prev_ni_pathlen = ndp->ni_pathlen;
1161 	ndp->ni_pathlen -= cnp->cn_namelen;
1162 	KASSERT(ndp->ni_pathlen <= PATH_MAX,
1163 	    ("%s: ni_pathlen underflow to %zd\n", __func__, ndp->ni_pathlen));
1164 	prev_ni_next = ndp->ni_next;
1165 	ndp->ni_next = cp;
1166 
1167 	/*
1168 	 * Something else should be clearing this.
1169 	 */
1170 	cnp->cn_flags &= ~(ISDOTDOT|ISLASTCN);
1171 
1172 	cnp->cn_flags |= MAKEENTRY;
1173 	if (*cp == '\0' && docache == 0)
1174 		cnp->cn_flags &= ~MAKEENTRY;
1175 	if (cnp->cn_namelen == 2 &&
1176 	    cnp->cn_nameptr[1] == '.' && cnp->cn_nameptr[0] == '.')
1177 		cnp->cn_flags |= ISDOTDOT;
1178 	if (*ndp->ni_next == 0) {
1179 		cnp->cn_flags |= ISLASTCN;
1180 
1181 		if (__predict_false(cnp->cn_namelen == 1 && cnp->cn_nameptr[0] == '.' &&
1182 		    (cnp->cn_nameiop == DELETE || cnp->cn_nameiop == RENAME))) {
1183 			error = EINVAL;
1184 			goto bad;
1185 		}
1186 	}
1187 
1188 	nameicap_tracker_add(ndp, dp);
1189 
1190 	/*
1191 	 * Make sure degenerate names don't get here, their handling was
1192 	 * previously found in this spot.
1193 	 */
1194 	MPASS(cnp->cn_nameptr[0] != '\0');
1195 
1196 	/*
1197 	 * Handle "..": five special cases.
1198 	 * 0. If doing a capability lookup and lookup_cap_dotdot is
1199 	 *    disabled, return ENOTCAPABLE.
1200 	 * 1. Return an error if this is the last component of
1201 	 *    the name and the operation is DELETE or RENAME.
1202 	 * 2. If at root directory (e.g. after chroot)
1203 	 *    or at absolute root directory
1204 	 *    then ignore it so can't get out.
1205 	 * 3. If this vnode is the root of a mounted
1206 	 *    filesystem, then replace it with the
1207 	 *    vnode which was mounted on so we take the
1208 	 *    .. in the other filesystem.
1209 	 * 4. If the vnode is the top directory of
1210 	 *    the jail or chroot, don't let them out.
1211 	 * 5. If doing a capability lookup and lookup_cap_dotdot is
1212 	 *    enabled, return ENOTCAPABLE if the lookup would escape
1213 	 *    from the initial file descriptor directory.  Checks are
1214 	 *    done by ensuring that namei() already traversed the
1215 	 *    result of dotdot lookup.
1216 	 */
1217 	if (cnp->cn_flags & ISDOTDOT) {
1218 		if (__predict_false((ndp->ni_lcf & (NI_LCF_STRICTREL_KTR |
1219 		    NI_LCF_CAP_DOTDOT_KTR)) == NI_LCF_STRICTREL_KTR))
1220 			NI_CAP_VIOLATION(ndp, cnp->cn_pnbuf);
1221 		if (__predict_false((ndp->ni_lcf & (NI_LCF_STRICTREL |
1222 		    NI_LCF_CAP_DOTDOT)) == NI_LCF_STRICTREL)) {
1223 			error = ENOTCAPABLE;
1224 			goto bad;
1225 		}
1226 		if ((cnp->cn_flags & ISLASTCN) != 0 &&
1227 		    (cnp->cn_nameiop == DELETE || cnp->cn_nameiop == RENAME)) {
1228 			error = EINVAL;
1229 			goto bad;
1230 		}
1231 		for (;;) {
1232 			bool isroot;
1233 
1234 			isroot = vfs_lookup_isroot(ndp, dp);
1235 			if (__predict_false(isroot && (ndp->ni_lcf &
1236 			    (NI_LCF_STRICTREL | NI_LCF_STRICTREL_KTR)) != 0)) {
1237 				if ((ndp->ni_lcf & NI_LCF_STRICTREL_KTR) != 0)
1238 					NI_CAP_VIOLATION(ndp, cnp->cn_pnbuf);
1239 				if ((ndp->ni_lcf & NI_LCF_STRICTREL) != 0) {
1240 					error = ENOTCAPABLE;
1241 					goto capdotdot;
1242 				}
1243 			}
1244 			if (isroot || ((dp->v_vflag & VV_ROOT) != 0 &&
1245 			    (cnp->cn_flags & NOCROSSMOUNT) != 0)) {
1246 				ndp->ni_dvp = dp;
1247 				ndp->ni_vp = dp;
1248 				VREF(dp);
1249 				goto nextname;
1250 			}
1251 			if ((dp->v_vflag & VV_ROOT) == 0)
1252 				break;
1253 			if (VN_IS_DOOMED(dp)) {	/* forced unmount */
1254 				error = ENOENT;
1255 				goto bad;
1256 			}
1257 			tdp = dp;
1258 			dp = dp->v_mount->mnt_vnodecovered;
1259 			VREF(dp);
1260 			vput(tdp);
1261 			vn_lock(dp,
1262 			    enforce_lkflags(dp->v_mount, cnp->cn_lkflags |
1263 			    LK_RETRY));
1264 			error = nameicap_check_dotdot(ndp, dp);
1265 			if (error != 0) {
1266 capdotdot:
1267 				goto bad;
1268 			}
1269 		}
1270 	}
1271 
1272 	/*
1273 	 * We now have a segment name to search for, and a directory to search.
1274 	 */
1275 unionlookup:
1276 #ifdef MAC
1277 	error = mac_vnode_check_lookup(cnp->cn_cred, dp, cnp);
1278 	if (__predict_false(error))
1279 		goto bad;
1280 #endif
1281 	ndp->ni_dvp = dp;
1282 	ndp->ni_vp = NULL;
1283 	ASSERT_VOP_LOCKED(dp, "lookup");
1284 	/*
1285 	 * If we have a shared lock we may need to upgrade the lock for the
1286 	 * last operation.
1287 	 */
1288 	if ((cnp->cn_flags & LOCKPARENT) && (cnp->cn_flags & ISLASTCN) &&
1289 	    dp != vp_crossmp && VOP_ISLOCKED(dp) == LK_SHARED)
1290 		vn_lock(dp, LK_UPGRADE|LK_RETRY);
1291 	if (VN_IS_DOOMED(dp)) {
1292 		error = ENOENT;
1293 		goto bad;
1294 	}
1295 	/*
1296 	 * If we're looking up the last component and we need an exclusive
1297 	 * lock, adjust our lkflags.
1298 	 */
1299 	if (needs_exclusive_leaf(dp->v_mount, cnp->cn_flags))
1300 		cnp->cn_lkflags = LK_EXCLUSIVE;
1301 	lkflags_save = cnp->cn_lkflags;
1302 	cnp->cn_lkflags = enforce_lkflags(dp->v_mount, cnp->cn_lkflags);
1303 	error = VOP_LOOKUP(dp, &ndp->ni_vp, cnp);
1304 	cnp->cn_lkflags = lkflags_save;
1305 	if (error != 0) {
1306 		KASSERT(ndp->ni_vp == NULL, ("leaf should be empty"));
1307 		if ((error == ENOENT) &&
1308 		    (dp->v_vflag & VV_ROOT) && (dp->v_mount != NULL) &&
1309 		    (dp->v_mount->mnt_flag & MNT_UNION)) {
1310 			tdp = dp;
1311 			dp = dp->v_mount->mnt_vnodecovered;
1312 			VREF(dp);
1313 			vput(tdp);
1314 			vn_lock(dp,
1315 			    enforce_lkflags(dp->v_mount, cnp->cn_lkflags |
1316 			    LK_RETRY));
1317 			nameicap_tracker_add(ndp, dp);
1318 			goto unionlookup;
1319 		}
1320 
1321 		if (error == ERELOOKUP) {
1322 			vref(dp);
1323 			ndp->ni_vp = dp;
1324 			error = 0;
1325 			relookup = 1;
1326 			goto good;
1327 		}
1328 
1329 		if (error != EJUSTRETURN)
1330 			goto bad;
1331 		/*
1332 		 * At this point, we know we're at the end of the
1333 		 * pathname.  If creating / renaming, we can consider
1334 		 * allowing the file or directory to be created / renamed,
1335 		 * provided we're not on a read-only filesystem.
1336 		 */
1337 		if (rdonly) {
1338 			error = EROFS;
1339 			goto bad;
1340 		}
1341 		/* trailing slash only allowed for directories */
1342 		if ((cnp->cn_flags & TRAILINGSLASH) &&
1343 		    !(cnp->cn_flags & WILLBEDIR)) {
1344 			error = ENOENT;
1345 			goto bad;
1346 		}
1347 		if ((cnp->cn_flags & LOCKPARENT) == 0)
1348 			VOP_UNLOCK(dp);
1349 		/*
1350 		 * We return with ni_vp NULL to indicate that the entry
1351 		 * doesn't currently exist, leaving a pointer to the
1352 		 * (possibly locked) directory vnode in ndp->ni_dvp.
1353 		 */
1354 		goto success;
1355 	}
1356 
1357 good:
1358 	dp = ndp->ni_vp;
1359 
1360 	/*
1361 	 * Check for symbolic link
1362 	 */
1363 	if ((dp->v_type == VLNK) &&
1364 	    ((cnp->cn_flags & FOLLOW) || (cnp->cn_flags & TRAILINGSLASH) ||
1365 	     *ndp->ni_next == '/')) {
1366 		cnp->cn_flags |= ISSYMLINK;
1367 		if (VN_IS_DOOMED(dp)) {
1368 			/*
1369 			 * We can't know whether the directory was mounted with
1370 			 * NOSYMFOLLOW, so we can't follow safely.
1371 			 */
1372 			error = ENOENT;
1373 			goto bad2;
1374 		}
1375 		if (dp->v_mount->mnt_flag & MNT_NOSYMFOLLOW) {
1376 			error = EACCES;
1377 			goto bad2;
1378 		}
1379 		/*
1380 		 * Symlink code always expects an unlocked dvp.
1381 		 */
1382 		if (ndp->ni_dvp != ndp->ni_vp) {
1383 			VOP_UNLOCK(ndp->ni_dvp);
1384 			ni_dvp_unlocked = 1;
1385 		}
1386 		goto success;
1387 	}
1388 
1389 	if ((vn_irflag_read(dp) & VIRF_MOUNTPOINT) != 0 &&
1390 	    (cnp->cn_flags & NOCROSSMOUNT) == 0) {
1391 		error = vfs_lookup_cross_mount(ndp);
1392 		if (error != 0)
1393 			goto bad_unlocked;
1394 		/*
1395 		 * FALLTHROUGH to nextname
1396 		 */
1397 		dp = ndp->ni_vp;
1398 	}
1399 
1400 nextname:
1401 	/*
1402 	 * Not a symbolic link that we will follow.  Continue with the
1403 	 * next component if there is any; otherwise, we're done.
1404 	 */
1405 	KASSERT((cnp->cn_flags & ISLASTCN) || *ndp->ni_next == '/',
1406 	    ("lookup: invalid path state."));
1407 	if (relookup) {
1408 		relookup = 0;
1409 		ndp->ni_pathlen = prev_ni_pathlen;
1410 		ndp->ni_next = prev_ni_next;
1411 		if (ndp->ni_dvp != dp)
1412 			vput(ndp->ni_dvp);
1413 		else
1414 			vrele(ndp->ni_dvp);
1415 		goto dirloop;
1416 	}
1417 	if (cnp->cn_flags & ISDOTDOT) {
1418 		error = nameicap_check_dotdot(ndp, ndp->ni_vp);
1419 		if (error != 0)
1420 			goto bad2;
1421 	}
1422 	if (*ndp->ni_next == '/') {
1423 		cnp->cn_nameptr = ndp->ni_next;
1424 		while (*cnp->cn_nameptr == '/') {
1425 			cnp->cn_nameptr++;
1426 			ndp->ni_pathlen--;
1427 		}
1428 		if (ndp->ni_dvp != dp)
1429 			vput(ndp->ni_dvp);
1430 		else
1431 			vrele(ndp->ni_dvp);
1432 		goto dirloop;
1433 	}
1434 	/*
1435 	 * If we're processing a path with a trailing slash,
1436 	 * check that the end result is a directory.
1437 	 */
1438 	if ((cnp->cn_flags & TRAILINGSLASH) && dp->v_type != VDIR) {
1439 		error = ENOTDIR;
1440 		goto bad2;
1441 	}
1442 	/*
1443 	 * Disallow directory write attempts on read-only filesystems.
1444 	 */
1445 	if (rdonly &&
1446 	    (cnp->cn_nameiop == DELETE || cnp->cn_nameiop == RENAME)) {
1447 		error = EROFS;
1448 		goto bad2;
1449 	}
1450 	if (!wantparent) {
1451 		ni_dvp_unlocked = 2;
1452 		if (ndp->ni_dvp != dp)
1453 			vput(ndp->ni_dvp);
1454 		else
1455 			vrele(ndp->ni_dvp);
1456 	} else if ((cnp->cn_flags & LOCKPARENT) == 0 && ndp->ni_dvp != dp) {
1457 		VOP_UNLOCK(ndp->ni_dvp);
1458 		ni_dvp_unlocked = 1;
1459 	}
1460 
1461 	if (cnp->cn_flags & AUDITVNODE1)
1462 		AUDIT_ARG_VNODE1(dp);
1463 	else if (cnp->cn_flags & AUDITVNODE2)
1464 		AUDIT_ARG_VNODE2(dp);
1465 
1466 	if ((cnp->cn_flags & LOCKLEAF) == 0)
1467 		VOP_UNLOCK(dp);
1468 success:
1469 	/*
1470 	 * FIXME: for lookups which only cross a mount point to fetch the
1471 	 * root vnode, ni_dvp will be set to vp_crossmp. This can be a problem
1472 	 * if either WANTPARENT or LOCKPARENT is set.
1473 	 */
1474 	/*
1475 	 * Because of shared lookup we may have the vnode shared locked, but
1476 	 * the caller may want it to be exclusively locked.
1477 	 */
1478 	if (needs_exclusive_leaf(dp->v_mount, cnp->cn_flags) &&
1479 	    VOP_ISLOCKED(dp) != LK_EXCLUSIVE) {
1480 		vn_lock(dp, LK_UPGRADE | LK_RETRY);
1481 		if (VN_IS_DOOMED(dp)) {
1482 			error = ENOENT;
1483 			goto bad2;
1484 		}
1485 	}
1486 success_right_lock:
1487 	if (ndp->ni_vp != NULL) {
1488 		if ((cnp->cn_flags & ISDOTDOT) == 0)
1489 			nameicap_tracker_add(ndp, ndp->ni_vp);
1490 		if ((cnp->cn_flags & (FAILIFEXISTS | ISSYMLINK)) == FAILIFEXISTS)
1491 			return (vfs_lookup_failifexists(ndp));
1492 	}
1493 	return (0);
1494 
1495 bad2:
1496 	if (ni_dvp_unlocked != 2) {
1497 		if (dp != ndp->ni_dvp && !ni_dvp_unlocked)
1498 			vput(ndp->ni_dvp);
1499 		else
1500 			vrele(ndp->ni_dvp);
1501 	}
1502 bad:
1503 	vput(dp);
1504 bad_unlocked:
1505 	ndp->ni_vp = NULL;
1506 	return (error);
1507 }
1508 
1509 /*
1510  * relookup - lookup a path name component
1511  *    Used by lookup to re-acquire things.
1512  */
1513 int
vfs_relookup(struct vnode * dvp,struct vnode ** vpp,struct componentname * cnp,bool refstart)1514 vfs_relookup(struct vnode *dvp, struct vnode **vpp, struct componentname *cnp,
1515     bool refstart)
1516 {
1517 	struct vnode *dp = NULL;		/* the directory we are searching */
1518 	int rdonly;			/* lookup read-only flag bit */
1519 	int error = 0;
1520 
1521 	KASSERT(cnp->cn_flags & ISLASTCN,
1522 	    ("relookup: Not given last component."));
1523 	/*
1524 	 * Setup: break out flag bits into variables.
1525 	 */
1526 	KASSERT((cnp->cn_flags & (LOCKPARENT | WANTPARENT)) != 0,
1527 	    ("relookup: parent not wanted"));
1528 	rdonly = cnp->cn_flags & RDONLY;
1529 	cnp->cn_flags &= ~ISSYMLINK;
1530 	dp = dvp;
1531 	cnp->cn_lkflags = LK_EXCLUSIVE;
1532 	vn_lock(dp, LK_EXCLUSIVE | LK_RETRY);
1533 
1534 	/*
1535 	 * Search a new directory.
1536 	 *
1537 	 * See a comment in vfs_lookup for cnp->cn_nameptr.
1538 	 *
1539 	 * Check for "" which represents the root directory after slash
1540 	 * removal.
1541 	 */
1542 	if (cnp->cn_nameptr[0] == '\0') {
1543 		/*
1544 		 * Support only LOOKUP for "/" because lookup()
1545 		 * can't succeed for CREATE, DELETE and RENAME.
1546 		 */
1547 		KASSERT(cnp->cn_nameiop == LOOKUP, ("nameiop must be LOOKUP"));
1548 		KASSERT(dp->v_type == VDIR, ("dp is not a directory"));
1549 
1550 		if (!(cnp->cn_flags & LOCKLEAF))
1551 			VOP_UNLOCK(dp);
1552 		*vpp = dp;
1553 		/* XXX This should probably move to the top of function. */
1554 		if (refstart)
1555 			panic("lookup: SAVESTART");
1556 		return (0);
1557 	}
1558 
1559 	if (cnp->cn_flags & ISDOTDOT)
1560 		panic ("relookup: lookup on dot-dot");
1561 
1562 	/*
1563 	 * We now have a segment name to search for, and a directory to search.
1564 	 */
1565 	if ((error = VOP_LOOKUP(dp, vpp, cnp)) != 0) {
1566 		KASSERT(*vpp == NULL, ("leaf should be empty"));
1567 		if (error != EJUSTRETURN)
1568 			goto bad;
1569 		/*
1570 		 * If creating and at end of pathname, then can consider
1571 		 * allowing file to be created.
1572 		 */
1573 		if (rdonly) {
1574 			error = EROFS;
1575 			goto bad;
1576 		}
1577 		/* ASSERT(dvp == ndp->ni_startdir) */
1578 		if (refstart)
1579 			VREF(dvp);
1580 		if ((cnp->cn_flags & LOCKPARENT) == 0)
1581 			VOP_UNLOCK(dp);
1582 		/*
1583 		 * We return with ni_vp NULL to indicate that the entry
1584 		 * doesn't currently exist, leaving a pointer to the
1585 		 * (possibly locked) directory vnode in ndp->ni_dvp.
1586 		 */
1587 		return (0);
1588 	}
1589 
1590 	dp = *vpp;
1591 
1592 	/*
1593 	 * Disallow directory write attempts on read-only filesystems.
1594 	 */
1595 	if (rdonly &&
1596 	    (cnp->cn_nameiop == DELETE || cnp->cn_nameiop == RENAME)) {
1597 		if (dvp == dp)
1598 			vrele(dvp);
1599 		else
1600 			vput(dvp);
1601 		error = EROFS;
1602 		goto bad;
1603 	}
1604 	/*
1605 	 * Set the parent lock/ref state to the requested state.
1606 	 */
1607 	if ((cnp->cn_flags & LOCKPARENT) == 0 && dvp != dp)
1608 		VOP_UNLOCK(dvp);
1609 	/*
1610 	 * Check for symbolic link
1611 	 */
1612 	KASSERT(dp->v_type != VLNK || !(cnp->cn_flags & FOLLOW),
1613 	    ("relookup: symlink found.\n"));
1614 
1615 	/* ASSERT(dvp == ndp->ni_startdir) */
1616 	if (refstart)
1617 		VREF(dvp);
1618 
1619 	if ((cnp->cn_flags & LOCKLEAF) == 0)
1620 		VOP_UNLOCK(dp);
1621 	return (0);
1622 bad:
1623 	vput(dp);
1624 	*vpp = NULL;
1625 	return (error);
1626 }
1627 
1628 #ifdef INVARIANTS
1629 /*
1630  * Validate the final state of ndp after the lookup.
1631  */
1632 static void
NDVALIDATE_impl(struct nameidata * ndp,int line)1633 NDVALIDATE_impl(struct nameidata *ndp, int line)
1634 {
1635 	struct componentname *cnp;
1636 
1637 	cnp = &ndp->ni_cnd;
1638 	if (cnp->cn_pnbuf == NULL)
1639 		panic("%s: got no buf! called from %d", __func__, line);
1640 }
1641 
1642 #endif
1643