2 * SPDX-License-Identifier: BSD-3-Clause
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.
12 * Redistribution and use in source and binary forms, with or without
13 * modification, are permitted provided that the following conditions
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.
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
36 * @(#)vfs_lookup.c 8.4 (Berkeley) 2/16/94
39 #include <sys/cdefs.h>
40 __FBSDID("$FreeBSD$");
42 #include "opt_capsicum.h"
43 #include "opt_ktrace.h"
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>
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>
59 #include <sys/syscallsubr.h>
60 #include <sys/sysctl.h>
62 #include <sys/ktrace.h>
65 #include <machine/_inttypes.h>
68 #include <security/audit/audit.h>
69 #include <security/mac/mac_framework.h>
73 #define NAMEI_DIAGNOSTIC 1
74 #undef NAMEI_DIAGNOSTIC
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");
81 /* Allocation zone for namei. */
82 uma_zone_t namei_zone;
84 /* Placeholder vnode for mp traversal. */
85 static struct vnode *vp_crossmp;
88 crossmp_vop_islocked(struct vop_islocked_args *ap)
95 crossmp_vop_lock1(struct vop_lock1_args *ap)
98 struct lock *lk __unused;
99 const char *file __unused;
100 int flags, line __unused;
108 if ((flags & LK_SHARED) == 0)
109 panic("invalid lock request for crossmp");
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)
116 LOCK_LOG_LOCK("SLOCK", &lk->lock_object, 0, 0, ap->a_file, line);
121 crossmp_vop_unlock(struct vop_unlock_args *ap)
124 struct lock *lk __unused;
129 WITNESS_UNLOCK(&lk->lock_object, 0, LOCK_FILE, LOCK_LINE);
130 LOCK_LOG_LOCK("SUNLOCK", &lk->lock_object, 0, 0, LOCK_FILE,
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,
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.
146 struct nameicap_tracker {
148 TAILQ_ENTRY(nameicap_tracker) nm_link;
151 /* Zone for cap mode tracker elements used for dotdot capability checks. */
152 static uma_zone_t nt_zone;
155 nameiinit(void *dummy __unused)
158 namei_zone = uma_zcreate("NAMEI", MAXPATHLEN, NULL, NULL, NULL, NULL,
160 nt_zone = uma_zcreate("rentr", sizeof(struct nameicap_tracker),
161 NULL, NULL, NULL, NULL, UMA_ALIGN_PTR, 0);
162 vfs_vector_op_register(&crossmp_vnodeops);
163 getnewvnode("crossmp", NULL, &crossmp_vnodeops, &vp_crossmp);
165 SYSINIT(vfs, SI_SUB_VFS, SI_ORDER_SECOND, nameiinit, NULL);
167 static int lookup_cap_dotdot = 1;
168 SYSCTL_INT(_vfs, OID_AUTO, lookup_cap_dotdot, CTLFLAG_RWTUN,
169 &lookup_cap_dotdot, 0,
170 "enables \"..\" components in path lookup in capability mode");
171 static int lookup_cap_dotdot_nonlocal = 1;
172 SYSCTL_INT(_vfs, OID_AUTO, lookup_cap_dotdot_nonlocal, CTLFLAG_RWTUN,
173 &lookup_cap_dotdot_nonlocal, 0,
174 "enables \"..\" components in path lookup in capability mode "
175 "on non-local mount");
178 nameicap_tracker_add(struct nameidata *ndp, struct vnode *dp)
180 struct nameicap_tracker *nt;
182 if ((ndp->ni_lcf & NI_LCF_CAP_DOTDOT) == 0 || dp->v_type != VDIR)
184 if ((ndp->ni_lcf & (NI_LCF_BENEATH_ABS | NI_LCF_BENEATH_LATCHED)) ==
185 NI_LCF_BENEATH_ABS) {
186 MPASS((ndp->ni_lcf & NI_LCF_LATCH) != 0);
187 if (dp != ndp->ni_beneath_latch)
189 ndp->ni_lcf |= NI_LCF_BENEATH_LATCHED;
191 nt = uma_zalloc(nt_zone, M_WAITOK);
194 TAILQ_INSERT_TAIL(&ndp->ni_cap_tracker, nt, nm_link);
198 nameicap_cleanup(struct nameidata *ndp, bool clean_latch)
200 struct nameicap_tracker *nt, *nt1;
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);
207 uma_zfree(nt_zone, nt);
209 if (clean_latch && (ndp->ni_lcf & NI_LCF_LATCH) != 0) {
210 ndp->ni_lcf &= ~NI_LCF_LATCH;
211 vrele(ndp->ni_beneath_latch);
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. Also fail dotdot lookups for non-local
219 * filesystems, where external agents might assist local lookups to
220 * escape the compartment.
223 nameicap_check_dotdot(struct nameidata *ndp, struct vnode *dp)
225 struct nameicap_tracker *nt;
228 if ((ndp->ni_lcf & NI_LCF_CAP_DOTDOT) == 0 || dp == NULL ||
232 if (lookup_cap_dotdot_nonlocal == 0 && mp != NULL &&
233 (mp->mnt_flag & MNT_LOCAL) == 0)
234 return (ENOTCAPABLE);
235 TAILQ_FOREACH_REVERSE(nt, &ndp->ni_cap_tracker, nameicap_tracker_head,
240 if ((ndp->ni_lcf & NI_LCF_BENEATH_ABS) != 0) {
241 ndp->ni_lcf &= ~NI_LCF_BENEATH_LATCHED;
242 nameicap_cleanup(ndp, false);
245 return (ENOTCAPABLE);
249 namei_cleanup_cnp(struct componentname *cnp)
252 uma_zfree(namei_zone, cnp->cn_pnbuf);
254 cnp->cn_pnbuf = NULL;
255 cnp->cn_nameptr = NULL;
260 namei_handle_root(struct nameidata *ndp, struct vnode **dpp)
262 struct componentname *cnp;
265 if ((ndp->ni_lcf & NI_LCF_STRICTRELATIVE) != 0) {
267 if (KTRPOINT(curthread, KTR_CAPFAIL))
268 ktrcapfail(CAPFAIL_LOOKUP, NULL, NULL);
270 return (ENOTCAPABLE);
272 if ((cnp->cn_flags & BENEATH) != 0) {
273 ndp->ni_lcf |= NI_LCF_BENEATH_ABS;
274 ndp->ni_lcf &= ~NI_LCF_BENEATH_LATCHED;
275 nameicap_cleanup(ndp, false);
277 while (*(cnp->cn_nameptr) == '/') {
281 *dpp = ndp->ni_rootdir;
287 namei_setup(struct nameidata *ndp, struct vnode **dpp, struct pwd **pwdp)
289 struct componentname *cnp;
294 struct filecaps dirfd_caps;
301 startdir_used = false;
305 #ifdef CAPABILITY_MODE
307 * In capability mode, lookups must be restricted to happen in
308 * the subtree with the root specified by the file descriptor:
309 * - The root must be real file descriptor, not the pseudo-descriptor
311 * - The passed path must be relative and not absolute.
312 * - If lookup_cap_dotdot is disabled, path must not contain the
314 * - If lookup_cap_dotdot is enabled, we verify that all '..'
315 * components lookups result in the directories which were
316 * previously walked by us, which prevents an escape from
319 if (IN_CAPABILITY_MODE(td) && (cnp->cn_flags & NOCAPCHECK) == 0) {
320 ndp->ni_lcf |= NI_LCF_STRICTRELATIVE;
321 if (ndp->ni_dirfd == AT_FDCWD) {
323 if (KTRPOINT(td, KTR_CAPFAIL))
324 ktrcapfail(CAPFAIL_LOOKUP, NULL, NULL);
333 * Get starting point for the translation.
337 * The reference on ni_rootdir is acquired in the block below to avoid
338 * back-to-back atomics for absolute lookups.
340 ndp->ni_rootdir = pwd->pwd_rdir;
341 ndp->ni_topdir = pwd->pwd_jdir;
343 if (cnp->cn_pnbuf[0] == '/') {
344 ndp->ni_resflags |= NIRES_ABS;
345 error = namei_handle_root(ndp, dpp);
347 if (ndp->ni_startdir != NULL) {
348 *dpp = ndp->ni_startdir;
349 startdir_used = true;
350 } else if (ndp->ni_dirfd == AT_FDCWD) {
351 *dpp = pwd->pwd_cdir;
354 rights = *ndp->ni_rightsneeded;
355 cap_rights_set_one(&rights, CAP_LOOKUP);
357 if (cnp->cn_flags & AUDITVNODE1)
358 AUDIT_ARG_ATFD1(ndp->ni_dirfd);
359 if (cnp->cn_flags & AUDITVNODE2)
360 AUDIT_ARG_ATFD2(ndp->ni_dirfd);
362 * Effectively inlined fgetvp_rights, because we need to
363 * inspect the file as well as grabbing the vnode.
365 error = fget_cap(td, ndp->ni_dirfd, &rights,
366 &dfp, &ndp->ni_filecaps);
369 * Preserve the error; it should either be EBADF
370 * or capability-related, both of which can be
371 * safely returned to the caller.
374 if (dfp->f_ops == &badfileops) {
376 } else if (dfp->f_vnode == NULL) {
382 if ((dfp->f_flag & FSEARCH) != 0)
383 cnp->cn_flags |= NOEXECCHECK;
389 * If file descriptor doesn't have all rights,
390 * all lookups relative to it must also be
394 if (!cap_rights_contains(&ndp->ni_filecaps.fc_rights,
396 ndp->ni_filecaps.fc_fcntls != CAP_FCNTL_ALL ||
397 ndp->ni_filecaps.fc_nioctls != -1) {
398 ndp->ni_lcf |= NI_LCF_STRICTRELATIVE;
402 if (error == 0 && (*dpp)->v_type != VDIR)
405 if (error == 0 && (cnp->cn_flags & BENEATH) != 0) {
406 if (ndp->ni_dirfd == AT_FDCWD) {
407 ndp->ni_beneath_latch = pwd->pwd_cdir;
408 vrefact(ndp->ni_beneath_latch);
410 rights = *ndp->ni_rightsneeded;
411 cap_rights_set_one(&rights, CAP_LOOKUP);
412 error = fgetvp_rights(td, ndp->ni_dirfd, &rights,
413 &dirfd_caps, &ndp->ni_beneath_latch);
414 if (error == 0 && (*dpp)->v_type != VDIR) {
415 vrele(ndp->ni_beneath_latch);
420 ndp->ni_lcf |= NI_LCF_LATCH;
423 * If we are auditing the kernel pathname, save the user pathname.
425 if (cnp->cn_flags & AUDITVNODE1)
426 AUDIT_ARG_UPATH1_VP(td, ndp->ni_rootdir, *dpp, cnp->cn_pnbuf);
427 if (cnp->cn_flags & AUDITVNODE2)
428 AUDIT_ARG_UPATH2_VP(td, ndp->ni_rootdir, *dpp, cnp->cn_pnbuf);
429 if (ndp->ni_startdir != NULL && !startdir_used)
430 vrele(ndp->ni_startdir);
437 MPASS((ndp->ni_lcf & (NI_LCF_BENEATH_ABS | NI_LCF_LATCH)) !=
439 if (((ndp->ni_lcf & NI_LCF_STRICTRELATIVE) != 0 &&
440 lookup_cap_dotdot != 0) ||
441 ((ndp->ni_lcf & NI_LCF_STRICTRELATIVE) == 0 &&
442 (cnp->cn_flags & BENEATH) != 0))
443 ndp->ni_lcf |= NI_LCF_CAP_DOTDOT;
444 SDT_PROBE4(vfs, namei, lookup, entry, *dpp, cnp->cn_pnbuf,
445 cnp->cn_flags, false);
451 * Convert a pathname into a pointer to a locked vnode.
453 * The FOLLOW flag is set when symbolic links are to be followed
454 * when they occur at the end of the name translation process.
455 * Symbolic links are always followed for all other pathname
456 * components other than the last.
458 * The segflg defines whether the name is to be copied from user
459 * space or kernel space.
461 * Overall outline of namei:
464 * get starting directory
465 * while (!done && !error) {
466 * call lookup to search path.
467 * if symbolic link, massage name in buffer and continue
471 namei(struct nameidata *ndp)
473 char *cp; /* pointer into pathname argument */
474 struct vnode *dp; /* the directory we are searching */
475 struct iovec aiov; /* uio for reading symbolic links */
476 struct componentname *cnp;
481 enum cache_fpl_status status;
489 * While NDINIT may seem like a more natural place to do it, there are
490 * callers which directly modify flags past invoking init.
492 cnp->cn_origflags = cnp->cn_flags;
494 ndp->ni_cnd.cn_cred = ndp->ni_cnd.cn_thread->td_ucred;
495 KASSERT(cnp->cn_cred && td->td_proc, ("namei: bad cred/proc"));
496 KASSERT((cnp->cn_flags & NAMEI_INTERNAL_FLAGS) == 0,
497 ("namei: unexpected flags: %" PRIx64 "\n",
498 cnp->cn_flags & NAMEI_INTERNAL_FLAGS));
499 if (cnp->cn_flags & NOCACHE)
500 KASSERT(cnp->cn_nameiop != LOOKUP,
501 ("%s: NOCACHE passed with LOOKUP", __func__));
502 MPASS(ndp->ni_startdir == NULL || ndp->ni_startdir->v_type == VDIR ||
503 ndp->ni_startdir->v_type == VBAD);
509 * Get a buffer for the name to be translated, and copy the
510 * name into the buffer.
512 cnp->cn_pnbuf = uma_zalloc(namei_zone, M_WAITOK);
513 if (ndp->ni_segflg == UIO_SYSSPACE)
514 error = copystr(ndp->ni_dirp, cnp->cn_pnbuf, MAXPATHLEN,
517 error = copyinstr(ndp->ni_dirp, cnp->cn_pnbuf, MAXPATHLEN,
520 if (__predict_false(error != 0)) {
521 namei_cleanup_cnp(cnp);
526 * Don't allow empty pathnames.
528 if (__predict_false(*cnp->cn_pnbuf == '\0')) {
529 namei_cleanup_cnp(cnp);
534 if (KTRPOINT(td, KTR_NAMEI)) {
535 KASSERT(cnp->cn_thread == curthread,
536 ("namei not using curthread"));
537 ktrnamei(cnp->cn_pnbuf);
541 cnp->cn_nameptr = cnp->cn_pnbuf;
544 * First try looking up the target without locking any vnodes.
546 * We may need to start from scratch or pick up where it left off.
548 error = cache_fplookup(ndp, &status, &pwd);
550 case CACHE_FPL_STATUS_UNSET:
551 __assert_unreachable();
553 case CACHE_FPL_STATUS_HANDLED:
557 case CACHE_FPL_STATUS_PARTIAL:
558 TAILQ_INIT(&ndp->ni_cap_tracker);
559 dp = ndp->ni_startdir;
561 case CACHE_FPL_STATUS_ABORTED:
562 TAILQ_INIT(&ndp->ni_cap_tracker);
563 error = namei_setup(ndp, &dp, &pwd);
565 namei_cleanup_cnp(cnp);
577 ndp->ni_startdir = dp;
582 * If not a symbolic link, we're done.
584 if ((cnp->cn_flags & ISSYMLINK) == 0) {
585 if ((cnp->cn_flags & (SAVENAME | SAVESTART)) == 0) {
586 namei_cleanup_cnp(cnp);
588 cnp->cn_flags |= HASBUF;
589 if ((ndp->ni_lcf & (NI_LCF_BENEATH_ABS |
590 NI_LCF_BENEATH_LATCHED)) == NI_LCF_BENEATH_ABS) {
594 nameicap_cleanup(ndp, true);
595 SDT_PROBE3(vfs, namei, lookup, return, error,
596 (error == 0 ? ndp->ni_vp : NULL), false);
602 if (ndp->ni_loopcnt++ >= MAXSYMLINKS) {
607 if ((cnp->cn_flags & NOMACCHECK) == 0) {
608 error = mac_vnode_check_readlink(td->td_ucred,
614 if (ndp->ni_pathlen > 1)
615 cp = uma_zalloc(namei_zone, M_WAITOK);
619 aiov.iov_len = MAXPATHLEN;
620 auio.uio_iov = &aiov;
623 auio.uio_rw = UIO_READ;
624 auio.uio_segflg = UIO_SYSSPACE;
626 auio.uio_resid = MAXPATHLEN;
627 error = VOP_READLINK(ndp->ni_vp, &auio, cnp->cn_cred);
629 if (ndp->ni_pathlen > 1)
630 uma_zfree(namei_zone, cp);
633 linklen = MAXPATHLEN - auio.uio_resid;
635 if (ndp->ni_pathlen > 1)
636 uma_zfree(namei_zone, cp);
640 if (linklen + ndp->ni_pathlen > MAXPATHLEN) {
641 if (ndp->ni_pathlen > 1)
642 uma_zfree(namei_zone, cp);
643 error = ENAMETOOLONG;
646 if (ndp->ni_pathlen > 1) {
647 bcopy(ndp->ni_next, cp + linklen, ndp->ni_pathlen);
648 uma_zfree(namei_zone, cnp->cn_pnbuf);
651 cnp->cn_pnbuf[linklen] = '\0';
652 ndp->ni_pathlen += linklen;
656 * Check if root directory should replace current directory.
658 cnp->cn_nameptr = cnp->cn_pnbuf;
659 if (*(cnp->cn_nameptr) == '/') {
661 error = namei_handle_root(ndp, &dp);
671 namei_cleanup_cnp(cnp);
672 nameicap_cleanup(ndp, true);
673 SDT_PROBE3(vfs, namei, lookup, return, error, NULL, false);
679 compute_cn_lkflags(struct mount *mp, int lkflags, int cnflags)
682 if (mp == NULL || ((lkflags & LK_SHARED) &&
683 (!(mp->mnt_kern_flag & MNTK_LOOKUP_SHARED) ||
684 ((cnflags & ISDOTDOT) &&
685 (mp->mnt_kern_flag & MNTK_LOOKUP_EXCL_DOTDOT))))) {
686 lkflags &= ~LK_SHARED;
687 lkflags |= LK_EXCLUSIVE;
689 lkflags |= LK_NODDLKTREAT;
694 needs_exclusive_leaf(struct mount *mp, int flags)
698 * Intermediate nodes can use shared locks, we only need to
699 * force an exclusive lock for leaf nodes.
701 if ((flags & (ISLASTCN | LOCKLEAF)) != (ISLASTCN | LOCKLEAF))
704 /* Always use exclusive locks if LOCKSHARED isn't set. */
705 if (!(flags & LOCKSHARED))
709 * For lookups during open(), if the mount point supports
710 * extended shared operations, then use a shared lock for the
711 * leaf node, otherwise use an exclusive lock.
713 if ((flags & ISOPEN) != 0)
714 return (!MNT_EXTENDED_SHARED(mp));
717 * Lookup requests outside of open() that specify LOCKSHARED
718 * only need a shared lock on the leaf vnode.
725 * This is a very central and rather complicated routine.
727 * The pathname is pointed to by ni_ptr and is of length ni_pathlen.
728 * The starting directory is taken from ni_startdir. The pathname is
729 * descended until done, or a symbolic link is encountered. The variable
730 * ni_more is clear if the path is completed; it is set to one if a
731 * symbolic link needing interpretation is encountered.
733 * The flag argument is LOOKUP, CREATE, RENAME, or DELETE depending on
734 * whether the name is to be looked up, created, renamed, or deleted.
735 * When CREATE, RENAME, or DELETE is specified, information usable in
736 * creating, renaming, or deleting a directory entry may be calculated.
737 * If flag has LOCKPARENT or'ed into it, the parent directory is returned
738 * locked. If flag has WANTPARENT or'ed into it, the parent directory is
739 * returned unlocked. Otherwise the parent directory is not returned. If
740 * the target of the pathname exists and LOCKLEAF is or'ed into the flag
741 * the target is returned locked, otherwise it is returned unlocked.
742 * When creating or renaming and LOCKPARENT is specified, the target may not
743 * be ".". When deleting and LOCKPARENT is specified, the target may be ".".
745 * Overall outline of lookup:
748 * identify next component of name at ndp->ni_ptr
749 * handle degenerate case where name is null string
750 * if .. and crossing mount points and on mounted filesys, find parent
751 * call VOP_LOOKUP routine for next component name
752 * directory vnode returned in ni_dvp, unlocked unless LOCKPARENT set
753 * component vnode returned in ni_vp (if it exists), locked.
754 * if result vnode is mounted on and crossing mount points,
755 * find mounted on vnode
756 * if more components of name, do next level at dirloop
757 * return the answer in ni_vp, locked if LOCKLEAF set
758 * if LOCKPARENT set, return locked parent in ni_dvp
759 * if WANTPARENT set, return unlocked parent in ni_dvp
762 lookup(struct nameidata *ndp)
764 char *cp; /* pointer into pathname argument */
765 char *prev_ni_next; /* saved ndp->ni_next */
766 struct vnode *dp = NULL; /* the directory we are searching */
767 struct vnode *tdp; /* saved dp */
768 struct mount *mp; /* mount table entry */
770 size_t prev_ni_pathlen; /* saved ndp->ni_pathlen */
771 int docache; /* == 0 do not cache last component */
772 int wantparent; /* 1 => wantparent or lockparent flag */
773 int rdonly; /* lookup read-only flag bit */
775 int dpunlocked = 0; /* dp has already been unlocked */
776 int relookup = 0; /* do not consume the path component */
777 struct componentname *cnp = &ndp->ni_cnd;
782 * Setup: break out flag bits into variables.
785 wantparent = cnp->cn_flags & (LOCKPARENT | WANTPARENT);
786 KASSERT(cnp->cn_nameiop == LOOKUP || wantparent,
787 ("CREATE, DELETE, RENAME require LOCKPARENT or WANTPARENT."));
789 * When set to zero, docache causes the last component of the
790 * pathname to be deleted from the cache and the full lookup
791 * of the name to be done (via VOP_CACHEDLOOKUP()). Often
792 * filesystems need some pre-computed values that are made
793 * during the full lookup, for instance UFS sets dp->i_offset.
795 * The docache variable is set to zero when requested by the
796 * NOCACHE flag and for all modifying operations except CREATE.
798 docache = (cnp->cn_flags & NOCACHE) ^ NOCACHE;
799 if (cnp->cn_nameiop == DELETE ||
800 (wantparent && cnp->cn_nameiop != CREATE &&
801 cnp->cn_nameiop != LOOKUP))
803 rdonly = cnp->cn_flags & RDONLY;
804 cnp->cn_flags &= ~ISSYMLINK;
807 * We use shared locks until we hit the parent of the last cn then
808 * we adjust based on the requesting flags.
810 cnp->cn_lkflags = LK_SHARED;
811 dp = ndp->ni_startdir;
812 ndp->ni_startdir = NULLVP;
814 compute_cn_lkflags(dp->v_mount, cnp->cn_lkflags | LK_RETRY,
819 * Search a new directory.
821 * The last component of the filename is left accessible via
822 * cnp->cn_nameptr for callers that need the name. Callers needing
823 * the name set the SAVENAME flag. When done, they assume
824 * responsibility for freeing the pathname buffer.
826 for (cp = cnp->cn_nameptr; *cp != 0 && *cp != '/'; cp++)
828 cnp->cn_namelen = cp - cnp->cn_nameptr;
829 if (cnp->cn_namelen > NAME_MAX) {
830 error = ENAMETOOLONG;
833 #ifdef NAMEI_DIAGNOSTIC
836 printf("{%s}: ", cnp->cn_nameptr);
839 prev_ni_pathlen = ndp->ni_pathlen;
840 ndp->ni_pathlen -= cnp->cn_namelen;
841 KASSERT(ndp->ni_pathlen <= PATH_MAX,
842 ("%s: ni_pathlen underflow to %zd\n", __func__, ndp->ni_pathlen));
843 prev_ni_next = ndp->ni_next;
847 * Replace multiple slashes by a single slash and trailing slashes
848 * by a null. This must be done before VOP_LOOKUP() because some
849 * fs's don't know about trailing slashes. Remember if there were
850 * trailing slashes to handle symlinks, existing non-directories
851 * and non-existing files that won't be directories specially later.
853 while (*cp == '/' && (cp[1] == '/' || cp[1] == '\0')) {
857 *ndp->ni_next = '\0';
858 cnp->cn_flags |= TRAILINGSLASH;
863 cnp->cn_flags |= MAKEENTRY;
864 if (*cp == '\0' && docache == 0)
865 cnp->cn_flags &= ~MAKEENTRY;
866 if (cnp->cn_namelen == 2 &&
867 cnp->cn_nameptr[1] == '.' && cnp->cn_nameptr[0] == '.')
868 cnp->cn_flags |= ISDOTDOT;
870 cnp->cn_flags &= ~ISDOTDOT;
871 if (*ndp->ni_next == 0)
872 cnp->cn_flags |= ISLASTCN;
874 cnp->cn_flags &= ~ISLASTCN;
876 if ((cnp->cn_flags & ISLASTCN) != 0 &&
877 cnp->cn_namelen == 1 && cnp->cn_nameptr[0] == '.' &&
878 (cnp->cn_nameiop == DELETE || cnp->cn_nameiop == RENAME)) {
883 nameicap_tracker_add(ndp, dp);
886 * Check for degenerate name (e.g. / or "")
887 * which is a way of talking about a directory,
888 * e.g. like "/." or ".".
890 if (cnp->cn_nameptr[0] == '\0') {
891 if (dp->v_type != VDIR) {
895 if (cnp->cn_nameiop != LOOKUP) {
905 if (cnp->cn_flags & AUDITVNODE1)
906 AUDIT_ARG_VNODE1(dp);
907 else if (cnp->cn_flags & AUDITVNODE2)
908 AUDIT_ARG_VNODE2(dp);
910 if (!(cnp->cn_flags & (LOCKPARENT | LOCKLEAF)))
912 /* XXX This should probably move to the top of function. */
913 if (cnp->cn_flags & SAVESTART)
914 panic("lookup: SAVESTART");
919 * Handle "..": five special cases.
920 * 0. If doing a capability lookup and lookup_cap_dotdot is
921 * disabled, return ENOTCAPABLE.
922 * 1. Return an error if this is the last component of
923 * the name and the operation is DELETE or RENAME.
924 * 2. If at root directory (e.g. after chroot)
925 * or at absolute root directory
926 * then ignore it so can't get out.
927 * 3. If this vnode is the root of a mounted
928 * filesystem, then replace it with the
929 * vnode which was mounted on so we take the
930 * .. in the other filesystem.
931 * 4. If the vnode is the top directory of
932 * the jail or chroot, don't let them out.
933 * 5. If doing a capability lookup and lookup_cap_dotdot is
934 * enabled, return ENOTCAPABLE if the lookup would escape
935 * from the initial file descriptor directory. Checks are
936 * done by ensuring that namei() already traversed the
937 * result of dotdot lookup.
939 if (cnp->cn_flags & ISDOTDOT) {
940 if ((ndp->ni_lcf & (NI_LCF_STRICTRELATIVE | NI_LCF_CAP_DOTDOT))
941 == NI_LCF_STRICTRELATIVE) {
943 if (KTRPOINT(curthread, KTR_CAPFAIL))
944 ktrcapfail(CAPFAIL_LOOKUP, NULL, NULL);
949 if ((cnp->cn_flags & ISLASTCN) != 0 &&
950 (cnp->cn_nameiop == DELETE || cnp->cn_nameiop == RENAME)) {
955 for (pr = cnp->cn_cred->cr_prison; pr != NULL;
957 if (dp == pr->pr_root)
959 if (dp == ndp->ni_rootdir ||
960 dp == ndp->ni_topdir ||
963 ((dp->v_vflag & VV_ROOT) != 0 &&
964 (cnp->cn_flags & NOCROSSMOUNT) != 0)) {
970 if ((dp->v_vflag & VV_ROOT) == 0)
972 if (VN_IS_DOOMED(dp)) { /* forced unmount */
977 dp = dp->v_mount->mnt_vnodecovered;
981 compute_cn_lkflags(dp->v_mount, cnp->cn_lkflags |
982 LK_RETRY, ISDOTDOT));
983 error = nameicap_check_dotdot(ndp, dp);
986 if (KTRPOINT(curthread, KTR_CAPFAIL))
987 ktrcapfail(CAPFAIL_LOOKUP, NULL, NULL);
995 * We now have a segment name to search for, and a directory to search.
999 error = mac_vnode_check_lookup(cnp->cn_thread->td_ucred, dp, cnp);
1005 ASSERT_VOP_LOCKED(dp, "lookup");
1007 * If we have a shared lock we may need to upgrade the lock for the
1010 if ((cnp->cn_flags & LOCKPARENT) && (cnp->cn_flags & ISLASTCN) &&
1011 dp != vp_crossmp && VOP_ISLOCKED(dp) == LK_SHARED)
1012 vn_lock(dp, LK_UPGRADE|LK_RETRY);
1013 if (VN_IS_DOOMED(dp)) {
1018 * If we're looking up the last component and we need an exclusive
1019 * lock, adjust our lkflags.
1021 if (needs_exclusive_leaf(dp->v_mount, cnp->cn_flags))
1022 cnp->cn_lkflags = LK_EXCLUSIVE;
1023 #ifdef NAMEI_DIAGNOSTIC
1024 vn_printf(dp, "lookup in ");
1026 lkflags_save = cnp->cn_lkflags;
1027 cnp->cn_lkflags = compute_cn_lkflags(dp->v_mount, cnp->cn_lkflags,
1029 error = VOP_LOOKUP(dp, &ndp->ni_vp, cnp);
1030 cnp->cn_lkflags = lkflags_save;
1032 KASSERT(ndp->ni_vp == NULL, ("leaf should be empty"));
1033 #ifdef NAMEI_DIAGNOSTIC
1034 printf("not found\n");
1036 if ((error == ENOENT) &&
1037 (dp->v_vflag & VV_ROOT) && (dp->v_mount != NULL) &&
1038 (dp->v_mount->mnt_flag & MNT_UNION)) {
1040 dp = dp->v_mount->mnt_vnodecovered;
1044 compute_cn_lkflags(dp->v_mount, cnp->cn_lkflags |
1045 LK_RETRY, cnp->cn_flags));
1046 nameicap_tracker_add(ndp, dp);
1050 if (error == ERELOOKUP) {
1058 if (error != EJUSTRETURN)
1061 * At this point, we know we're at the end of the
1062 * pathname. If creating / renaming, we can consider
1063 * allowing the file or directory to be created / renamed,
1064 * provided we're not on a read-only filesystem.
1070 /* trailing slash only allowed for directories */
1071 if ((cnp->cn_flags & TRAILINGSLASH) &&
1072 !(cnp->cn_flags & WILLBEDIR)) {
1076 if ((cnp->cn_flags & LOCKPARENT) == 0)
1079 * We return with ni_vp NULL to indicate that the entry
1080 * doesn't currently exist, leaving a pointer to the
1081 * (possibly locked) directory vnode in ndp->ni_dvp.
1083 if (cnp->cn_flags & SAVESTART) {
1084 ndp->ni_startdir = ndp->ni_dvp;
1085 VREF(ndp->ni_startdir);
1091 #ifdef NAMEI_DIAGNOSTIC
1097 * Check to see if the vnode has been mounted on;
1098 * if so find the root of the mounted filesystem.
1100 while (dp->v_type == VDIR && (mp = dp->v_mountedhere) &&
1101 (cnp->cn_flags & NOCROSSMOUNT) == 0) {
1102 if (vfs_busy(mp, 0))
1105 if (dp != ndp->ni_dvp)
1109 vrefact(vp_crossmp);
1110 ndp->ni_dvp = vp_crossmp;
1111 error = VFS_ROOT(mp, compute_cn_lkflags(mp, cnp->cn_lkflags,
1112 cnp->cn_flags), &tdp);
1114 if (vn_lock(vp_crossmp, LK_SHARED | LK_NOWAIT))
1115 panic("vp_crossmp exclusively locked or reclaimed");
1120 ndp->ni_vp = dp = tdp;
1124 * Check for symbolic link
1126 if ((dp->v_type == VLNK) &&
1127 ((cnp->cn_flags & FOLLOW) || (cnp->cn_flags & TRAILINGSLASH) ||
1128 *ndp->ni_next == '/')) {
1129 cnp->cn_flags |= ISSYMLINK;
1130 if (VN_IS_DOOMED(dp)) {
1132 * We can't know whether the directory was mounted with
1133 * NOSYMFOLLOW, so we can't follow safely.
1138 if (dp->v_mount->mnt_flag & MNT_NOSYMFOLLOW) {
1143 * Symlink code always expects an unlocked dvp.
1145 if (ndp->ni_dvp != ndp->ni_vp) {
1146 VOP_UNLOCK(ndp->ni_dvp);
1147 ni_dvp_unlocked = 1;
1154 * Not a symbolic link that we will follow. Continue with the
1155 * next component if there is any; otherwise, we're done.
1157 KASSERT((cnp->cn_flags & ISLASTCN) || *ndp->ni_next == '/',
1158 ("lookup: invalid path state."));
1161 ndp->ni_pathlen = prev_ni_pathlen;
1162 ndp->ni_next = prev_ni_next;
1163 if (ndp->ni_dvp != dp)
1169 if (cnp->cn_flags & ISDOTDOT) {
1170 error = nameicap_check_dotdot(ndp, ndp->ni_vp);
1173 if (KTRPOINT(curthread, KTR_CAPFAIL))
1174 ktrcapfail(CAPFAIL_LOOKUP, NULL, NULL);
1179 if (*ndp->ni_next == '/') {
1180 cnp->cn_nameptr = ndp->ni_next;
1181 while (*cnp->cn_nameptr == '/') {
1185 if (ndp->ni_dvp != dp)
1192 * If we're processing a path with a trailing slash,
1193 * check that the end result is a directory.
1195 if ((cnp->cn_flags & TRAILINGSLASH) && dp->v_type != VDIR) {
1200 * Disallow directory write attempts on read-only filesystems.
1203 (cnp->cn_nameiop == DELETE || cnp->cn_nameiop == RENAME)) {
1207 if (cnp->cn_flags & SAVESTART) {
1208 ndp->ni_startdir = ndp->ni_dvp;
1209 VREF(ndp->ni_startdir);
1212 ni_dvp_unlocked = 2;
1213 if (ndp->ni_dvp != dp)
1217 } else if ((cnp->cn_flags & LOCKPARENT) == 0 && ndp->ni_dvp != dp) {
1218 VOP_UNLOCK(ndp->ni_dvp);
1219 ni_dvp_unlocked = 1;
1222 if (cnp->cn_flags & AUDITVNODE1)
1223 AUDIT_ARG_VNODE1(dp);
1224 else if (cnp->cn_flags & AUDITVNODE2)
1225 AUDIT_ARG_VNODE2(dp);
1227 if ((cnp->cn_flags & LOCKLEAF) == 0)
1231 * FIXME: for lookups which only cross a mount point to fetch the
1232 * root vnode, ni_dvp will be set to vp_crossmp. This can be a problem
1233 * if either WANTPARENT or LOCKPARENT is set.
1236 * Because of shared lookup we may have the vnode shared locked, but
1237 * the caller may want it to be exclusively locked.
1239 if (needs_exclusive_leaf(dp->v_mount, cnp->cn_flags) &&
1240 VOP_ISLOCKED(dp) != LK_EXCLUSIVE) {
1241 vn_lock(dp, LK_UPGRADE | LK_RETRY);
1242 if (VN_IS_DOOMED(dp)) {
1250 if (ni_dvp_unlocked != 2) {
1251 if (dp != ndp->ni_dvp && !ni_dvp_unlocked)
1264 * relookup - lookup a path name component
1265 * Used by lookup to re-acquire things.
1268 relookup(struct vnode *dvp, struct vnode **vpp, struct componentname *cnp)
1270 struct vnode *dp = NULL; /* the directory we are searching */
1271 int wantparent; /* 1 => wantparent or lockparent flag */
1272 int rdonly; /* lookup read-only flag bit */
1275 KASSERT(cnp->cn_flags & ISLASTCN,
1276 ("relookup: Not given last component."));
1278 * Setup: break out flag bits into variables.
1280 wantparent = cnp->cn_flags & (LOCKPARENT|WANTPARENT);
1281 KASSERT(wantparent, ("relookup: parent not wanted."));
1282 rdonly = cnp->cn_flags & RDONLY;
1283 cnp->cn_flags &= ~ISSYMLINK;
1285 cnp->cn_lkflags = LK_EXCLUSIVE;
1286 vn_lock(dp, LK_EXCLUSIVE | LK_RETRY);
1289 * Search a new directory.
1291 * The last component of the filename is left accessible via
1292 * cnp->cn_nameptr for callers that need the name. Callers needing
1293 * the name set the SAVENAME flag. When done, they assume
1294 * responsibility for freeing the pathname buffer.
1296 #ifdef NAMEI_DIAGNOSTIC
1297 printf("{%s}: ", cnp->cn_nameptr);
1301 * Check for "" which represents the root directory after slash
1304 if (cnp->cn_nameptr[0] == '\0') {
1306 * Support only LOOKUP for "/" because lookup()
1307 * can't succeed for CREATE, DELETE and RENAME.
1309 KASSERT(cnp->cn_nameiop == LOOKUP, ("nameiop must be LOOKUP"));
1310 KASSERT(dp->v_type == VDIR, ("dp is not a directory"));
1312 if (!(cnp->cn_flags & LOCKLEAF))
1315 /* XXX This should probably move to the top of function. */
1316 if (cnp->cn_flags & SAVESTART)
1317 panic("lookup: SAVESTART");
1321 if (cnp->cn_flags & ISDOTDOT)
1322 panic ("relookup: lookup on dot-dot");
1325 * We now have a segment name to search for, and a directory to search.
1327 #ifdef NAMEI_DIAGNOSTIC
1328 vn_printf(dp, "search in ");
1330 if ((error = VOP_LOOKUP(dp, vpp, cnp)) != 0) {
1331 KASSERT(*vpp == NULL, ("leaf should be empty"));
1332 if (error != EJUSTRETURN)
1335 * If creating and at end of pathname, then can consider
1336 * allowing file to be created.
1342 /* ASSERT(dvp == ndp->ni_startdir) */
1343 if (cnp->cn_flags & SAVESTART)
1345 if ((cnp->cn_flags & LOCKPARENT) == 0)
1348 * We return with ni_vp NULL to indicate that the entry
1349 * doesn't currently exist, leaving a pointer to the
1350 * (possibly locked) directory vnode in ndp->ni_dvp.
1358 * Disallow directory write attempts on read-only filesystems.
1361 (cnp->cn_nameiop == DELETE || cnp->cn_nameiop == RENAME)) {
1370 * Set the parent lock/ref state to the requested state.
1372 if ((cnp->cn_flags & LOCKPARENT) == 0 && dvp != dp) {
1377 } else if (!wantparent)
1380 * Check for symbolic link
1382 KASSERT(dp->v_type != VLNK || !(cnp->cn_flags & FOLLOW),
1383 ("relookup: symlink found.\n"));
1385 /* ASSERT(dvp == ndp->ni_startdir) */
1386 if (cnp->cn_flags & SAVESTART)
1389 if ((cnp->cn_flags & LOCKLEAF) == 0)
1399 * Free data allocated by namei(); see namei(9) for details.
1402 NDFREE_PNBUF(struct nameidata *ndp)
1405 if ((ndp->ni_cnd.cn_flags & HASBUF) != 0) {
1406 MPASS((ndp->ni_cnd.cn_flags & (SAVENAME | SAVESTART)) != 0);
1407 uma_zfree(namei_zone, ndp->ni_cnd.cn_pnbuf);
1408 ndp->ni_cnd.cn_flags &= ~HASBUF;
1413 (NDFREE)(struct nameidata *ndp, const u_int flags)
1421 if (!(flags & NDF_NO_FREE_PNBUF)) {
1424 if (!(flags & NDF_NO_VP_UNLOCK) &&
1425 (ndp->ni_cnd.cn_flags & LOCKLEAF) && ndp->ni_vp)
1427 if (!(flags & NDF_NO_DVP_UNLOCK) &&
1428 (ndp->ni_cnd.cn_flags & LOCKPARENT) &&
1429 ndp->ni_dvp != ndp->ni_vp)
1431 if (!(flags & NDF_NO_VP_RELE) && ndp->ni_vp) {
1440 VOP_UNLOCK(ndp->ni_vp);
1441 if (!(flags & NDF_NO_DVP_RELE) &&
1442 (ndp->ni_cnd.cn_flags & (LOCKPARENT|WANTPARENT))) {
1451 VOP_UNLOCK(ndp->ni_dvp);
1452 if (!(flags & NDF_NO_STARTDIR_RELE) &&
1453 (ndp->ni_cnd.cn_flags & SAVESTART)) {
1454 vrele(ndp->ni_startdir);
1455 ndp->ni_startdir = NULL;
1461 * Validate the final state of ndp after the lookup.
1463 * Historically filesystems were allowed to modify cn_flags. Most notably they
1464 * can add SAVENAME to the request, resulting in HASBUF and pushing subsequent
1465 * clean up to the consumer. In practice this seems to only concern != LOOKUP
1468 * As a step towards stricter API contract this routine validates the state to
1469 * clean up. Note validation is a work in progress with the intent of becoming
1470 * stricter over time.
1472 #define NDMODIFYINGFLAGS (LOCKLEAF | LOCKPARENT | WANTPARENT | SAVENAME | SAVESTART | HASBUF)
1474 NDVALIDATE(struct nameidata *ndp)
1476 struct componentname *cnp;
1477 u_int64_t used, orig;
1480 orig = cnp->cn_origflags;
1481 used = cnp->cn_flags;
1482 switch (cnp->cn_nameiop) {
1485 * For plain lookup we require strict conformance -- nothing
1486 * to clean up if it was not requested by the caller.
1488 orig &= NDMODIFYINGFLAGS;
1489 used &= NDMODIFYINGFLAGS;
1490 if ((orig & (SAVENAME | SAVESTART)) != 0)
1500 * Some filesystems set SAVENAME to provoke HASBUF, accomodate
1501 * for it until it gets fixed.
1503 orig &= NDMODIFYINGFLAGS;
1504 orig |= (SAVENAME | HASBUF);
1505 used &= NDMODIFYINGFLAGS;
1506 used |= (SAVENAME | HASBUF);
1514 panic("%s: mismatched flags for op %d: added %" PRIx64 ", "
1515 "removed %" PRIx64" (%" PRIx64" != %" PRIx64"; stored %" PRIx64" != %" PRIx64")",
1516 __func__, cnp->cn_nameiop, used & ~orig, orig &~ used,
1517 orig, used, cnp->cn_origflags, cnp->cn_flags);
1522 * Determine if there is a suitable alternate filename under the specified
1523 * prefix for the specified path. If the create flag is set, then the
1524 * alternate prefix will be used so long as the parent directory exists.
1525 * This is used by the various compatibility ABIs so that Linux binaries prefer
1526 * files under /compat/linux for example. The chosen path (whether under
1527 * the prefix or under /) is returned in a kernel malloc'd buffer pointed
1528 * to by pathbuf. The caller is responsible for free'ing the buffer from
1529 * the M_TEMP bucket if one is returned.
1532 kern_alternate_path(struct thread *td, const char *prefix, const char *path,
1533 enum uio_seg pathseg, char **pathbuf, int create, int dirfd)
1535 struct nameidata nd, ndroot;
1536 char *ptr, *buf, *cp;
1540 buf = (char *) malloc(MAXPATHLEN, M_TEMP, M_WAITOK);
1543 /* Copy the prefix into the new pathname as a starting point. */
1544 len = strlcpy(buf, prefix, MAXPATHLEN);
1545 if (len >= MAXPATHLEN) {
1550 sz = MAXPATHLEN - len;
1553 /* Append the filename to the prefix. */
1554 if (pathseg == UIO_SYSSPACE)
1555 error = copystr(path, ptr, sz, &len);
1557 error = copyinstr(path, ptr, sz, &len);
1565 /* Only use a prefix with absolute pathnames. */
1571 if (dirfd != AT_FDCWD) {
1573 * We want the original because the "prefix" is
1574 * included in the already opened dirfd.
1576 bcopy(ptr, buf, len);
1581 * We know that there is a / somewhere in this pathname.
1582 * Search backwards for it, to find the file's parent dir
1583 * to see if it exists in the alternate tree. If it does,
1584 * and we want to create a file (cflag is set). We don't
1585 * need to worry about the root comparison in this case.
1589 for (cp = &ptr[len] - 1; *cp != '/'; cp--);
1592 NDINIT(&nd, LOOKUP, NOFOLLOW, UIO_SYSSPACE, buf, td);
1598 NDINIT(&nd, LOOKUP, NOFOLLOW, UIO_SYSSPACE, buf, td);
1605 * We now compare the vnode of the prefix to the one
1606 * vnode asked. If they resolve to be the same, then we
1607 * ignore the match so that the real root gets used.
1608 * This avoids the problem of traversing "../.." to find the
1609 * root directory and never finding it, because "/" resolves
1610 * to the emulation root directory. This is expensive :-(
1612 NDINIT(&ndroot, LOOKUP, FOLLOW, UIO_SYSSPACE, prefix,
1615 /* We shouldn't ever get an error from this namei(). */
1616 error = namei(&ndroot);
1618 if (nd.ni_vp == ndroot.ni_vp)
1621 NDFREE(&ndroot, NDF_ONLY_PNBUF);
1622 vrele(ndroot.ni_vp);
1626 NDFREE(&nd, NDF_ONLY_PNBUF);
1630 /* If there was an error, use the original path name. */
1632 bcopy(ptr, buf, len);