4 * The contents of this file are subject to the terms of the
5 * Common Development and Distribution License (the "License").
6 * You may not use this file except in compliance with the License.
8 * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE
9 * or http://www.opensolaris.org/os/licensing.
10 * See the License for the specific language governing permissions
11 * and limitations under the License.
13 * When distributing Covered Code, include this CDDL HEADER in each
14 * file and include the License file at usr/src/OPENSOLARIS.LICENSE.
15 * If applicable, add the following below this CDDL HEADER, with the
16 * fields enclosed by brackets "[]" replaced with your own identifying
17 * information: Portions Copyright [yyyy] [name of copyright owner]
23 * Copyright 2019 Nexenta by DDN, Inc. All rights reserved.
24 * Copyright (c) 1988, 2010, Oracle and/or its affiliates. All rights reserved.
25 * Copyright 2016 Joyent, Inc.
26 * Copyright 2023 RackTop Systems, Inc.
29 /* Copyright (c) 1983, 1984, 1985, 1986, 1987, 1988, 1989 AT&T */
30 /* All Rights Reserved */
33 * University Copyright- Copyright (c) 1982, 1986, 1988
34 * The Regents of the University of California
37 * University Acknowledgment- Portions of this document are derived from
38 * software developed by the University of California, Berkeley, and its
42 #include <sys/types.h>
43 #include <sys/param.h>
44 #include <sys/systm.h>
45 #include <sys/cpuvar.h>
46 #include <sys/errno.h>
51 #include <sys/vnode.h>
52 #include <sys/pathname.h>
54 #include <sys/vtrace.h>
55 #include <sys/sysmacros.h>
56 #include <sys/debug.h>
57 #include <sys/dirent.h>
61 #include <sys/fs/snode.h>
63 /* Controls whether paths are stored with vnodes. */
64 int vfs_vnode_path
= 1;
74 return (lookupnameatcred(fnamep
, seg
, followlink
, dirvpp
, compvpp
, NULL
,
79 * Lookup the user file name,
80 * Handle allocation and freeing of pathname buffer, return error.
84 const char *fnamep
, /* user pathname */
85 enum uio_seg seg
, /* addr space that name is in */
86 int followlink
, /* follow sym links */
87 vnode_t
**dirvpp
, /* ret for ptr to parent dir vnode */
88 vnode_t
**compvpp
, /* ret for ptr to component vnode */
89 vnode_t
*startvp
, /* start path search from vp */
90 cred_t
*cr
) /* credential */
92 char namebuf
[TYPICALMAXPATHLEN
];
93 struct pathname lookpn
;
96 error
= pn_get_buf(fnamep
, seg
, &lookpn
, namebuf
, sizeof (namebuf
));
98 error
= lookuppnatcred(&lookpn
, NULL
, followlink
,
99 dirvpp
, compvpp
, startvp
, cr
);
101 if (error
== ENAMETOOLONG
) {
103 * This thread used a pathname > TYPICALMAXPATHLEN bytes long.
105 if (error
= pn_get(fnamep
, seg
, &lookpn
))
107 error
= lookuppnatcred(&lookpn
, NULL
, followlink
,
108 dirvpp
, compvpp
, startvp
, cr
);
116 lookupnameat(const char *fnamep
, enum uio_seg seg
, int followlink
,
117 vnode_t
**dirvpp
, vnode_t
**compvpp
, vnode_t
*startvp
)
119 return (lookupnameatcred(fnamep
, seg
, followlink
, dirvpp
, compvpp
,
125 struct pathname
*pnp
,
126 struct pathname
*rpnp
,
131 return (lookuppnatcred(pnp
, rpnp
, followlink
, dirvpp
, compvpp
, NULL
,
136 * Lookup the user file name from a given vp, using a specific credential.
140 struct pathname
*pnp
, /* pathname to lookup */
141 struct pathname
*rpnp
, /* if non-NULL, return resolved path */
142 int followlink
, /* (don't) follow sym links */
143 vnode_t
**dirvpp
, /* ptr for parent vnode */
144 vnode_t
**compvpp
, /* ptr for entry vnode */
145 vnode_t
*startvp
, /* start search from this vp */
146 cred_t
*cr
) /* user credential */
148 vnode_t
*vp
; /* current directory vp */
152 if (pnp
->pn_pathlen
== 0)
155 mutex_enter(&p
->p_lock
); /* for u_rdir and u_cdir */
156 if ((rootvp
= PTOU(p
)->u_rdir
) == NULL
)
158 else if (rootvp
!= rootdir
) /* no need to VN_HOLD rootdir */
161 if (pnp
->pn_path
[0] == '/') {
164 vp
= (startvp
== NULL
) ? PTOU(p
)->u_cdir
: startvp
;
167 mutex_exit(&p
->p_lock
);
170 * Skip over leading slashes
172 if (pnp
->pn_path
[0] == '/') {
176 } while (pnp
->pn_path
[0] == '/');
179 return (lookuppnvp(pnp
, rpnp
, followlink
, dirvpp
,
180 compvpp
, rootvp
, vp
, cr
));
184 lookuppnat(struct pathname
*pnp
, struct pathname
*rpnp
,
185 int followlink
, vnode_t
**dirvpp
, vnode_t
**compvpp
,
188 return (lookuppnatcred(pnp
, rpnp
, followlink
, dirvpp
, compvpp
, startvp
,
192 /* Private flag to do our getcwd() dirty work */
193 #define LOOKUP_CHECKREAD 0x10
194 #define LOOKUP_MASK (~LOOKUP_CHECKREAD)
197 * Starting at current directory, translate pathname pnp to end.
198 * Leave pathname of final component in pnp, return the vnode
199 * for the final component in *compvpp, and return the vnode
200 * for the parent of the final component in dirvpp.
202 * This is the central routine in pathname translation and handles
203 * multiple components in pathnames, separating them at /'s. It also
204 * implements mounted file systems and processes symbolic links.
206 * vp is the vnode where the directory search should start.
208 * Reference counts: vp must be held prior to calling this function. rootvp
209 * should only be held if rootvp != rootdir.
213 struct pathname
*pnp
, /* pathname to lookup */
214 struct pathname
*rpnp
, /* if non-NULL, return resolved path */
215 int flags
, /* follow symlinks */
216 vnode_t
**dirvpp
, /* ptr for parent vnode */
217 vnode_t
**compvpp
, /* ptr for entry vnode */
218 vnode_t
*rootvp
, /* rootvp */
219 vnode_t
*vp
, /* directory to start search at */
220 cred_t
*cr
) /* user's credential */
222 vnode_t
*cvp
; /* current component vp */
223 char component
[MAXNAMELEN
]; /* buffer for component (incl null) */
227 struct pathname presrvd
; /* case preserved name */
228 struct pathname
*pp
= NULL
;
230 vnode_t
*zonevp
= curproc
->p_zone
->zone_rootvp
; /* zone root */
231 int must_be_directory
= 0;
232 boolean_t retry_with_kcred
;
233 uint32_t auditing
= AU_AUDITING();
235 CPU_STATS_ADDQ(CPU
, sys
, namei
, 1);
239 rpnp
->pn_pathlen
= 0;
241 lookup_flags
= dirvpp
? LOOKUP_DIR
: 0;
242 if (flags
& FIGNORECASE
) {
243 lookup_flags
|= FIGNORECASE
;
247 if ((flags
& LOOKUP_NOACLCHECK
) != 0) {
248 lookup_flags
|= LOOKUP_NOACLCHECK
;
249 flags
&= ~LOOKUP_NOACLCHECK
;
253 audit_anchorpath(pnp
, vp
== rootvp
);
256 * Eliminate any trailing slashes in the pathname.
257 * If there are any, we must follow all symlinks.
258 * Also, we must guarantee that the last component is a directory.
260 if (pn_fixslash(pnp
)) {
262 must_be_directory
= 1;
267 retry_with_kcred
= B_FALSE
;
270 * Make sure we have a directory.
272 if (vp
->v_type
!= VDIR
) {
277 if (rpnp
&& VN_CMP(vp
, rootvp
))
278 (void) pn_set(rpnp
, "/");
281 * Process the next component of the pathname.
283 if (error
= pn_getcomponent(pnp
, component
)) {
288 * Handle "..": two special cases.
289 * 1. If we're at the root directory (e.g. after chroot or
290 * zone_enter) then change ".." to "." so we can't get
291 * out of this subtree.
292 * 2. If this vnode is the root of a mounted file system,
293 * then replace it with the vnode that was mounted on
294 * so that we take the ".." in the other file system.
296 if (component
[0] == '.' && component
[1] == '.' && component
[2] == 0) {
298 if (VN_CMP(vp
, rootvp
) || VN_CMP(vp
, zonevp
)) {
300 } else if (vp
->v_flag
& VROOT
) {
305 * While we deal with the vfs pointer from the vnode
306 * the filesystem could have been forcefully unmounted
307 * and the vnode's v_vfsp could have been invalidated
308 * by VFS_UNMOUNT. Hence, we cache v_vfsp and use it
309 * with vfs_rlock_wait/vfs_unlock.
310 * It is safe to use the v_vfsp even it is freed by
311 * VFS_UNMOUNT because vfs_rlock_wait/vfs_unlock
312 * do not dereference v_vfsp. It is just used as a
314 * One more corner case here is the memory getting
315 * reused for another vfs structure. In this case
316 * lookuppnvp's vfs_rlock_wait will succeed, domount's
317 * vfs_lock will fail and domount will bail out with an
323 * This lock is used to synchronize
324 * mounts/unmounts and lookups.
325 * Threads doing mounts/unmounts hold the
326 * writers version vfs_lock_wait().
329 vfs_rlock_wait(vfsp
);
332 * If this vnode is on a file system that
333 * has been forcibly unmounted,
334 * we can't proceed. Cancel this operation
337 * vfs_vnodecovered is NULL if unmounted.
338 * Currently, nfs uses VFS_UNMOUNTED to
339 * check if it's a forced-umount. Keep the
340 * same checking here as well even though it
343 if (((vp
= cvp
->v_vfsp
->vfs_vnodecovered
) == NULL
) ||
344 (cvp
->v_vfsp
->vfs_flag
& VFS_UNMOUNTED
)) {
356 * Crossing mount points. For eg: We are doing
357 * a lookup of ".." for file systems root vnode
358 * mounted here, and VOP_LOOKUP() (with covered vnode)
359 * will be on underlying file systems mount point
360 * vnode. Set retry_with_kcred flag as we might end
361 * up doing VOP_LOOKUP() with kcred if required.
363 retry_with_kcred
= B_TRUE
;
369 * LOOKUP_CHECKREAD is a private flag used by vnodetopath() to indicate
370 * that we need to have read permission on every directory in the entire
371 * path. This is used to ensure that a forward-lookup of a cached value
372 * has the same effect as a reverse-lookup when the cached value cannot
375 if ((flags
& LOOKUP_CHECKREAD
) &&
376 (error
= VOP_ACCESS(vp
, VREAD
, 0, cr
, NULL
)) != 0)
380 * Perform a lookup in the current directory.
382 error
= VOP_LOOKUP(vp
, component
, &cvp
, pnp
, lookup_flags
,
383 rootvp
, cr
, NULL
, NULL
, pp
);
386 * Retry with kcred - If crossing mount points & error is EACCES.
388 * If we are crossing mount points here and doing ".." lookup,
389 * VOP_LOOKUP() might fail if the underlying file systems
390 * mount point has no execute permission. In cases like these,
391 * we retry VOP_LOOKUP() by giving as much privilage as possible
392 * by passing kcred credentials.
394 * In case of hierarchical file systems, passing kcred still may
396 * For eg: UFS FS --> Mount NFS FS --> Again mount UFS on some
397 * directory inside NFS FS.
399 if ((error
== EACCES
) && retry_with_kcred
)
400 error
= VOP_LOOKUP(vp
, component
, &cvp
, pnp
, lookup_flags
,
401 rootvp
, zone_kcred(), NULL
, NULL
, pp
);
406 * On error, return hard error if
407 * (a) we're not at the end of the pathname yet, or
408 * (b) the caller didn't want the parent directory, or
409 * (c) we failed for some reason other than a missing entry.
411 if (pn_pathleft(pnp
) || dirvpp
== NULL
|| error
!= ENOENT
)
413 if (auditing
) { /* directory access */
414 if (error
= audit_savepath(pnp
, vp
, vp
, error
, cr
))
420 * We inform the caller that the desired entry must be
421 * a directory by adding a '/' to the component name.
423 if (must_be_directory
&& (error
= pn_addslash(pnp
)) != 0)
428 if (rootvp
!= rootdir
)
436 * Traverse mount points.
437 * XXX why don't we need to hold a read lock here (call vn_vfsrlock)?
438 * What prevents a concurrent update to v_vfsmountedhere?
439 * Possible answer: if mounting, we might not see the mount
440 * if it is concurrently coming into existence, but that's
441 * really not much different from the thread running a bit slower.
442 * If unmounting, we may get into traverse() when we shouldn't,
443 * but traverse() will catch this case for us.
444 * (For this to work, fetching v_vfsmountedhere had better
447 if (vn_mountedvfs(cvp
) != NULL
) {
448 if ((error
= traverse(&cvp
)) != 0)
453 * If we hit a symbolic link and there is more path to be
454 * translated or this operation does not wish to apply
455 * to a link, then place the contents of the link at the
456 * front of the remaining pathname.
458 if (cvp
->v_type
== VLNK
&& ((flags
& FOLLOW
) || pn_pathleft(pnp
))) {
459 struct pathname linkpath
;
461 if (++nlink
> MAXSYMLINKS
) {
466 if (error
= pn_getsymlink(cvp
, &linkpath
, cr
)) {
472 audit_symlink(pnp
, &linkpath
);
474 if (pn_pathleft(&linkpath
) == 0)
475 (void) pn_set(&linkpath
, ".");
476 error
= pn_insert(pnp
, &linkpath
, strlen(component
));
482 if (pnp
->pn_pathlen
== 0) {
486 if (pnp
->pn_path
[0] == '/') {
490 } while (pnp
->pn_path
[0] == '/');
496 audit_anchorpath(pnp
, vp
== rootvp
);
497 if (pn_fixslash(pnp
)) {
499 must_be_directory
= 1;
505 * If rpnp is non-NULL, remember the resolved path name therein.
506 * Do not include "." components. Collapse occurrences of
507 * "previous/..", so long as "previous" is not itself "..".
508 * Exhausting rpnp results in error ENAMETOOLONG.
510 if (rpnp
&& strcmp(component
, ".") != 0) {
513 if (strcmp(component
, "..") == 0 &&
514 rpnp
->pn_pathlen
!= 0 &&
515 !((rpnp
->pn_pathlen
> 2 &&
516 strncmp(rpnp
->pn_path
+rpnp
->pn_pathlen
-3, "/..", 3) == 0) ||
517 (rpnp
->pn_pathlen
== 2 &&
518 strncmp(rpnp
->pn_path
, "..", 2) == 0))) {
519 while (rpnp
->pn_pathlen
&&
520 rpnp
->pn_path
[rpnp
->pn_pathlen
-1] != '/')
522 if (rpnp
->pn_pathlen
> 1)
524 rpnp
->pn_path
[rpnp
->pn_pathlen
] = '\0';
526 if (rpnp
->pn_pathlen
!= 0 &&
527 rpnp
->pn_path
[rpnp
->pn_pathlen
-1] != '/')
528 rpnp
->pn_path
[rpnp
->pn_pathlen
++] = '/';
529 if (flags
& FIGNORECASE
) {
531 * Return the case-preserved name
532 * within the resolved path.
534 error
= copystr(pp
->pn_buf
,
535 rpnp
->pn_path
+ rpnp
->pn_pathlen
,
536 rpnp
->pn_bufsize
- rpnp
->pn_pathlen
, &len
);
538 error
= copystr(component
,
539 rpnp
->pn_path
+ rpnp
->pn_pathlen
,
540 rpnp
->pn_bufsize
- rpnp
->pn_pathlen
, &len
);
542 if (error
) /* copystr() returns ENAMETOOLONG */
544 rpnp
->pn_pathlen
+= (len
- 1);
545 ASSERT(rpnp
->pn_bufsize
> rpnp
->pn_pathlen
);
550 * If no more components, return last directory (if wanted) and
551 * last component (if wanted).
553 if (pn_pathleft(pnp
) == 0) {
555 * If there was a trailing slash in the pathname,
556 * make sure the last component is a directory.
558 if (must_be_directory
&& cvp
->v_type
!= VDIR
) {
562 if (dirvpp
!= NULL
) {
564 * Check that we have the real parent and not
565 * an alias of the last component.
567 if (vn_compare(vp
, cvp
)) {
569 (void) audit_savepath(pnp
, cvp
, vp
,
574 if (rootvp
!= rootdir
)
584 (void) audit_savepath(pnp
, cvp
, vp
, 0, cr
);
585 if (pnp
->pn_path
== pnp
->pn_buf
)
586 (void) pn_set(pnp
, ".");
590 if (VN_CMP(cvp
, rootvp
))
591 (void) pn_set(rpnp
, "/");
592 else if (rpnp
->pn_pathlen
== 0)
593 (void) pn_set(rpnp
, ".");
600 if (rootvp
!= rootdir
)
608 * Skip over slashes from end of last component.
610 while (pnp
->pn_path
[0] == '/') {
616 * Searched through another level of directory:
617 * release previous directory handle and save new (result
618 * of lookup) as current directory.
626 if (auditing
) /* reached end of path */
627 (void) audit_savepath(pnp
, cvp
, vp
, error
, cr
);
630 * Error. Release vnodes and return.
635 * If the error was ESTALE and the current directory to look in
636 * was the root for this lookup, the root for a mounted file
637 * system, or the starting directory for lookups, then
638 * return ENOENT instead of ESTALE. In this case, no recovery
639 * is possible by the higher level. If ESTALE was returned for
640 * some intermediate directory along the path, then recovery
641 * is potentially possible and retrying from the higher level
642 * will either correct the situation by purging stale cache
643 * entries or eventually get back to the point where no recovery
646 if (error
== ESTALE
&&
647 (VN_CMP(vp
, rootvp
) || (vp
->v_flag
& VROOT
) || vp
== startvp
))
650 if (rootvp
!= rootdir
)
658 * Traverse a mount point. Routine accepts a vnode pointer as a reference
659 * parameter and performs the indirection, releasing the original vnode.
662 traverse(vnode_t
**cvpp
)
672 * If this vnode is mounted on, then we transparently indirect
673 * to the vnode which is the root of the mounted file system.
674 * Before we do this we must check that an unmount is not in
675 * progress on this vnode.
680 * Try to read lock the vnode. If this fails because
681 * the vnode is already write locked, then check to
682 * see whether it is the current thread which locked
683 * the vnode. If it is not, then read lock the vnode
684 * by waiting to acquire the lock.
686 * The code path in domount() is an example of support
687 * which needs to look up two pathnames and locks one
688 * of them in between the two lookups.
690 error
= vn_vfsrlock(cvp
);
692 if (!vn_vfswlock_held(cvp
))
693 error
= vn_vfsrlock_wait(cvp
);
696 * lookuppn() expects a held vnode to be
697 * returned because it promptly calls
698 * VN_RELE after the error return
706 * Reached the end of the mount chain?
708 vfsp
= vn_mountedvfs(cvp
);
715 * The read lock must be held across the call to VFS_ROOT() to
716 * prevent a concurrent unmount from destroying the vfs.
718 error
= VFS_ROOT(vfsp
, &tvp
);
734 * Return the lowermost vnode if this is a mountpoint.
737 vn_under(vnode_t
*vp
)
742 while (vp
->v_flag
& VROOT
) {
745 vfs_rlock_wait(vfsp
);
746 if ((uvp
= vfsp
->vfs_vnodecovered
) == NULL
||
747 (vfsp
->vfs_flag
& VFS_UNMOUNTED
)) {
761 vnode_match(vnode_t
*v1
, vnode_t
*v2
, cred_t
*cr
)
763 vattr_t v1attr
, v2attr
;
766 * If we have a device file, check to see if is a cloned open of the
767 * same device. For self-cloning devices, the major numbers will match.
768 * For devices cloned through the 'clone' driver, the minor number of
769 * the source device will be the same as the major number of the cloned
772 if ((v1
->v_type
== VCHR
|| v1
->v_type
== VBLK
) &&
773 v1
->v_type
== v2
->v_type
) {
774 if ((spec_is_selfclone(v1
) || spec_is_selfclone(v2
)) &&
775 getmajor(v1
->v_rdev
) == getmajor(v2
->v_rdev
))
778 if (spec_is_clone(v1
) &&
779 getmajor(v1
->v_rdev
) == getminor(v2
->v_rdev
))
782 if (spec_is_clone(v2
) &&
783 getmajor(v2
->v_rdev
) == getminor(v1
->v_rdev
))
787 v1attr
.va_mask
= v2attr
.va_mask
= AT_TYPE
;
790 * This check for symbolic links handles the pseudo-symlinks in procfs.
791 * These particular links have v_type of VDIR, but the attributes have a
792 * type of VLNK. We need to avoid these links because otherwise if we
793 * are currently in '/proc/self/fd', then '/proc/self/cwd' will compare
796 if (VOP_GETATTR(v1
, &v1attr
, 0, cr
, NULL
) != 0 ||
797 VOP_GETATTR(v2
, &v2attr
, 0, cr
, NULL
) != 0 ||
798 v1attr
.va_type
== VLNK
|| v2attr
.va_type
== VLNK
)
801 v1attr
.va_mask
= v2attr
.va_mask
= AT_TYPE
| AT_FSID
| AT_NODEID
;
803 if (VOP_GETATTR(v1
, &v1attr
, ATTR_REAL
, cr
, NULL
) != 0 ||
804 VOP_GETATTR(v2
, &v2attr
, ATTR_REAL
, cr
, NULL
) != 0)
807 return (v1attr
.va_fsid
== v2attr
.va_fsid
&&
808 v1attr
.va_nodeid
== v2attr
.va_nodeid
);
813 * Find the entry in the directory corresponding to the target vnode.
816 dirfindvp(vnode_t
*vrootp
, vnode_t
*dvp
, vnode_t
*tvp
, cred_t
*cr
, char *dbuf
,
817 size_t dlen
, dirent64_t
**rdp
)
828 ASSERT(dvp
->v_type
== VDIR
);
831 * This is necessary because of the strange semantics of VOP_LOOKUP().
833 bzero(&pnp
, sizeof (pnp
));
837 uio
.uio_segflg
= UIO_SYSSPACE
;
839 uio
.uio_extflg
= UIO_COPY_CACHED
;
842 if ((error
= VOP_ACCESS(dvp
, VREAD
, 0, cr
, NULL
)) != 0)
849 uio
.uio_resid
= dlen
;
853 (void) VOP_RWLOCK(dvp
, V_WRITELOCK_FALSE
, NULL
);
854 error
= VOP_READDIR(dvp
, &uio
, cr
, &eof
, NULL
, 0);
855 VOP_RWUNLOCK(dvp
, V_WRITELOCK_FALSE
, NULL
);
857 dbuflen
= dlen
- uio
.uio_resid
;
859 if (error
|| dbuflen
== 0)
862 dp
= (dirent64_t
*)dbuf
;
863 while ((intptr_t)dp
< (intptr_t)dbuf
+ dbuflen
) {
865 * Ignore '.' and '..' entries
867 if (strcmp(dp
->d_name
, ".") == 0 ||
868 strcmp(dp
->d_name
, "..") == 0) {
869 dp
= (dirent64_t
*)((intptr_t)dp
+
874 error
= VOP_LOOKUP(dvp
, dp
->d_name
, &cmpvp
, &pnp
, 0,
875 vrootp
, cr
, NULL
, NULL
, NULL
);
878 * We only want to bail out if there was an error other
879 * than ENOENT. Otherwise, it could be that someone
880 * just removed an entry since the readdir() call, and
881 * the entry we want is further on in the directory.
884 if (vnode_match(tvp
, cmpvp
, cr
)) {
891 } else if (error
!= ENOENT
) {
895 dp
= (dirent64_t
*)((intptr_t)dp
+ dp
->d_reclen
);
900 * Something strange has happened, this directory does not contain the
901 * specified vnode. This should never happen in the normal case, since
902 * we ensured that dvp is the parent of vp. This is possible in some
903 * rare conditions (races and the special .zfs directory).
906 error
= VOP_LOOKUP(dvp
, ".zfs", &cmpvp
, &pnp
, 0, vrootp
, cr
,
909 if (vnode_match(tvp
, cmpvp
, cr
)) {
910 (void) strcpy(dp
->d_name
, ".zfs");
911 dp
->d_reclen
= strlen(".zfs");
926 * Given a global path (from rootdir), and a vnode that is the current root,
927 * return the portion of the path that is beneath the current root or NULL on
928 * failure. The path MUST be a resolved path (no '..' entries or symlinks),
929 * otherwise this function will fail.
932 localpath(char *path
, struct vnode
*vrootp
, cred_t
*cr
)
936 char component
[MAXNAMELEN
];
941 * We use vn_compare() instead of VN_CMP() in order to detect lofs
942 * mounts and stacked vnodes.
944 if (vn_compare(vrootp
, rootdir
))
947 if (pn_get(path
, UIO_SYSSPACE
, &pn
) != 0)
953 if (vn_ismntpt(vp
) && traverse(&vp
) != 0) {
959 while (pn_pathleft(&pn
)) {
962 if (pn_getcomponent(&pn
, component
) != 0)
965 if (VOP_LOOKUP(vp
, component
, &cvp
, &pn
, 0, rootdir
, cr
,
966 NULL
, NULL
, NULL
) != 0)
971 if (vn_ismntpt(vp
) && traverse(&vp
) != 0)
974 if (vn_compare(vp
, vrootp
)) {
975 ret
= path
+ (pn
.pn_path
- pn
.pn_buf
);
987 * Clean a stale v_path from a vnode. This is only performed if the v_path has
988 * not been altered since it was found to be stale
991 vnode_clear_vpath(vnode_t
*vp
, char *vpath_old
)
993 mutex_enter(&vp
->v_lock
);
994 if (vp
->v_path
!= vn_vpath_empty
&& vp
->v_path
== vpath_old
) {
995 vp
->v_path
= vn_vpath_empty
;
996 mutex_exit(&vp
->v_lock
);
997 kmem_free(vpath_old
, strlen(vpath_old
) + 1);
999 mutex_exit(&vp
->v_lock
);
1004 * Validate that a pathname refers to a given vnode.
1007 vnode_valid_pn(vnode_t
*vp
, vnode_t
*vrootp
, pathname_t
*pn
, pathname_t
*rpn
,
1008 int flags
, cred_t
*cr
)
1012 * If we are in a zone or a chroot environment, then we have to
1013 * take additional steps, since the path to the root might not
1014 * be readable with the current credentials, even though the
1015 * process can legitmately access the file. In this case, we
1018 * lookuppnvp() with all privileges to get the resolved path.
1019 * call localpath() to get the local portion of the path, and
1020 * continue as normal.
1022 * If the the conversion to a local path fails, then we continue
1023 * as normal. This is a heuristic to make process object file
1024 * paths available from within a zone. Because lofs doesn't
1025 * support page operations, the vnode stored in the seg_t is
1026 * actually the underlying real vnode, not the lofs node itself.
1027 * Most of the time, the lofs path is the same as the underlying
1028 * vnode (for example, /usr/lib/libc.so.1).
1030 if (vrootp
!= rootdir
) {
1034 if (lookuppnvp(pn
, rpn
, FOLLOW
, NULL
, &compvp
, rootdir
,
1035 rootdir
, kcred
) == 0) {
1036 local
= localpath(rpn
->pn_path
, vrootp
, kcred
);
1041 * The original pn was changed through lookuppnvp().
1042 * Set it to local for next validation attempt.
1045 (void) pn_set(pn
, local
);
1052 * We should have a local path at this point, so start the search from
1053 * the root of the current process.
1056 if (vrootp
!= rootdir
)
1060 * The FOLLOW flag only determines, if the final path component
1061 * is a symlink, whether lookuppnvp will return the symlink, or its
1064 * If the vp is a VLNK, then passing the FOLLOW flag will cause
1065 * lookuppnvp to return the vnode of its target, instead of itself, and
1066 * so vn_compare will fail. Therefore, we do not pass FOLLOW when our vp
1069 * If the vp is not a VLNK, then we pass FOLLOW on the off-chance that
1070 * the stored v_path ends at a symlink, instead of the symlink's target.
1072 if (vp
->v_type
!= VLNK
)
1077 if (lookuppnvp(pn
, rpn
, flags
, NULL
, &compvp
, vrootp
, vrootp
,
1080 * Check to see if the returned vnode is the same as the one we
1083 if (vn_compare(vp
, compvp
) ||
1084 vnode_match(vp
, compvp
, cr
)) {
1096 * Struct for tracking vnodes with invalidated v_path entries during a
1097 * dirtopath reverse lookup. By keeping adequate state, those vnodes can be
1098 * revisted to populate v_path.
1100 struct dirpath_walk
{
1101 struct dirpath_walk
*dw_next
;
1109 * Given a directory, return the full, resolved path. This looks up "..",
1110 * searches for the given vnode in the parent, appends the component, etc. It
1111 * is used to implement vnodetopath() and getcwd() when the cached path fails.
1114 dirtopath(vnode_t
*vrootp
, vnode_t
*vp
, char *buf
, size_t buflen
, int flags
,
1117 pathname_t pn
, rpn
, emptypn
;
1118 vnode_t
*pvp
= NULL
, *startvp
= vp
;
1122 char *bufloc
, *dbuf
;
1123 const size_t dlen
= DIRENT64_RECLEN(MAXPATHLEN
);
1124 struct dirpath_walk
*dw_chain
= NULL
, *dw_entry
;
1126 /* Operation only allowed on directories */
1127 ASSERT(vp
->v_type
== VDIR
);
1129 /* We must have at least enough space for "/" */
1131 return (ENAMETOOLONG
);
1133 /* Start at end of string with terminating null */
1134 bufloc
= &buf
[buflen
- 1];
1139 dbuf
= kmem_alloc(dlen
, KM_SLEEP
);
1140 bzero(&emptypn
, sizeof (emptypn
));
1143 * Begin with an additional reference on vp. This will be decremented
1150 hrtime_t cached_stamp
;
1153 * Return if we've reached the root. If the buffer is empty,
1154 * return '/'. We explicitly don't use vn_compare(), since it
1155 * compares the real vnodes. A lofs mount of '/' would produce
1156 * incorrect results otherwise.
1158 if (VN_CMP(vrootp
, vp
)) {
1159 if (*bufloc
== '\0')
1165 * If we've reached the VFS root, something has gone wrong. We
1166 * should have reached the root in the above check. The only
1167 * explantation is that 'vp' is not contained withing the given
1168 * root, in which case we return EPERM.
1170 if (VN_CMP(rootdir
, vp
)) {
1176 * Shortcut: see if this vnode has correct v_path. If so,
1177 * we have the work done.
1179 mutex_enter(&vp
->v_lock
);
1180 if (vp
->v_path
!= vn_vpath_empty
&&
1181 pn_set(&pn
, vp
->v_path
) == 0) {
1182 cached_stamp
= vp
->v_path_stamp
;
1183 mutex_exit(&vp
->v_lock
);
1184 rpn
.pn_path
= rpn
.pn_buf
;
1186 /* Ensure the v_path pointing to correct vnode */
1187 if (vnode_valid_pn(vp
, vrootp
, &pn
, &rpn
, flags
,
1189 complen
= strlen(rpn
.pn_path
);
1195 bcopy(rpn
.pn_path
, bufloc
, complen
);
1199 * Immediately nuke cached v_path entries known
1202 vn_clearpath(vp
, cached_stamp
);
1205 mutex_exit(&vp
->v_lock
);
1209 * Shortcuts failed, search for this vnode in its parent. If
1210 * this is a mountpoint, then get the vnode underneath.
1212 if (vp
->v_flag
& VROOT
)
1214 if ((err
= VOP_LOOKUP(vp
, "..", &pvp
, &emptypn
, 0, vrootp
, cr
,
1215 NULL
, NULL
, NULL
)) != 0)
1219 * With extended attributes, it's possible for a directory to
1220 * have a parent that is a regular file. Check for that here.
1222 if (pvp
->v_type
!= VDIR
) {
1228 * If this is true, something strange has happened. This is
1229 * only true if we are the root of a filesystem, which should
1230 * have been caught by the check above.
1232 if (VN_CMP(pvp
, vp
)) {
1238 * Check if we have read and search privilege so, that
1239 * we can lookup the path in the directory
1241 vprivs
= (flags
& LOOKUP_CHECKREAD
) ? VREAD
| VEXEC
: VEXEC
;
1242 if ((err
= VOP_ACCESS(pvp
, vprivs
, 0, cr
, NULL
)) != 0) {
1247 * Search the parent directory for the entry corresponding to
1250 if ((err
= dirfindvp(vrootp
, pvp
, vp
, cr
, dbuf
, dlen
, &dp
))
1253 complen
= strlen(dp
->d_name
);
1255 if (bufloc
<= buf
) {
1259 bcopy(dp
->d_name
, bufloc
, complen
);
1261 /* Prepend a slash to the current path. */
1265 * Record the name and directory for later reconstruction and
1266 * link it up with the others.
1268 dw_entry
= kmem_alloc(sizeof (*dw_entry
), KM_SLEEP
);
1269 dw_entry
->dw_name
= kmem_alloc(complen
+ 1, KM_SLEEP
);
1270 VN_HOLD(dw_entry
->dw_vnode
= vp
);
1271 VN_HOLD(dw_entry
->dw_pvnode
= pvp
);
1272 bcopy(dp
->d_name
, dw_entry
->dw_name
, complen
+ 1);
1273 dw_entry
->dw_len
= complen
;
1274 dw_entry
->dw_next
= dw_chain
;
1275 dw_chain
= dw_entry
;
1277 /* And continue with the next component */
1284 * Place the path at the beginning of the buffer.
1287 ovbcopy(bufloc
, buf
, buflen
- (bufloc
- buf
));
1291 * Walk over encountered directory entries which were afflicted with a
1292 * stale or absent v_path. If the dirtopath was successful, we should
1293 * possess the necessary information to populate all of them with a
1296 * While processing this list, it is safe to call vn_setpath despite
1297 * the fact that racing vnode actions may have altered v_path entries
1298 * while the above loopwas still executing. Any updated entries will
1299 * have a newer v_path_stamp value which prevents an invalid overwrite.
1301 * If an error was encountered during the search, freeing the chain is
1304 dw_entry
= dw_chain
;
1305 while (dw_entry
!= NULL
) {
1306 struct dirpath_walk
*next
= dw_entry
->dw_next
;
1309 vn_setpath(NULL
, dw_entry
->dw_pvnode
,
1310 dw_entry
->dw_vnode
, dw_entry
->dw_name
,
1314 VN_RELE(dw_entry
->dw_vnode
);
1315 VN_RELE(dw_entry
->dw_pvnode
);
1316 kmem_free(dw_entry
->dw_name
, dw_entry
->dw_len
+ 1);
1317 kmem_free(dw_entry
, sizeof (*dw_entry
));
1322 * If the error was ESTALE and the current directory to look in
1323 * was the root for this lookup, the root for a mounted file
1324 * system, or the starting directory for lookups, then
1325 * return ENOENT instead of ESTALE. In this case, no recovery
1326 * is possible by the higher level. If ESTALE was returned for
1327 * some intermediate directory along the path, then recovery
1328 * is potentially possible and retrying from the higher level
1329 * will either correct the situation by purging stale cache
1330 * entries or eventually get back to the point where no recovery
1333 if (err
== ESTALE
&&
1334 (VN_CMP(vp
, vrootp
) || (vp
->v_flag
& VROOT
) || vp
== startvp
))
1337 kmem_free(dbuf
, dlen
);
1348 * The additional flag, LOOKUP_CHECKREAD, is used to enforce artificial
1349 * constraints in order to be standards compliant. For example, if we have
1350 * the cached path of '/foo/bar', and '/foo' has permissions 100 (execute
1351 * only), then we can legitimately look up the path to the current working
1352 * directory without needing read permission. Existing standards tests,
1353 * however, assume that we are determining the path by repeatedly looking up
1354 * "..". We need to keep this behavior in order to maintain backwards
1358 vnodetopath_common(vnode_t
*vrootp
, vnode_t
*vp
, char *buf
, size_t buflen
,
1359 cred_t
*cr
, int flags
)
1364 boolean_t doclose
= B_FALSE
;
1367 * If vrootp is NULL, get the root for curproc. Callers with any other
1368 * requirements should pass in a different vrootp.
1370 if (vrootp
== NULL
) {
1371 proc_t
*p
= curproc
;
1373 mutex_enter(&p
->p_lock
);
1374 if ((vrootp
= PTOU(p
)->u_rdir
) == NULL
)
1377 mutex_exit(&p
->p_lock
);
1383 * This is to get around an annoying artifact of the /proc filesystem,
1384 * which is the behavior of {cwd/root}. Trying to resolve this path
1385 * will result in /proc/pid/cwd instead of whatever the real working
1386 * directory is. We can't rely on VOP_REALVP(), since that will break
1387 * lofs. The only difference between procfs and lofs is that opening
1388 * the file will return the underling vnode in the case of procfs.
1390 if (vp
->v_type
== VDIR
&& VOP_REALVP(vp
, &realvp
, NULL
) == 0 &&
1393 if (VOP_OPEN(&vp
, FREAD
, cr
, NULL
) == 0)
1400 * Check to see if we have a valid cached path in the vnode.
1403 mutex_enter(&vp
->v_lock
);
1404 if (vp
->v_path
!= vn_vpath_empty
) {
1405 hrtime_t cached_stamp
;
1408 cached_stamp
= vp
->v_path_stamp
;
1409 (void) pn_set(&pn
, vp
->v_path
);
1410 mutex_exit(&vp
->v_lock
);
1412 /* We should only cache absolute paths */
1413 ASSERT(pn
.pn_buf
[0] == '/');
1416 if (vnode_valid_pn(vp
, vrootp
, &pn
, &rpn
, flags
, cr
) == 0) {
1417 /* Return the result, if we're able. */
1418 if (buflen
> rpn
.pn_pathlen
) {
1419 bcopy(rpn
.pn_path
, buf
, rpn
.pn_pathlen
+ 1);
1428 vn_clearpath(vp
, cached_stamp
);
1430 mutex_exit(&vp
->v_lock
);
1434 if (vp
->v_type
!= VDIR
) {
1436 * The reverse lookup tricks used by dirtopath aren't possible
1437 * for non-directory entries. The best which can be done is
1438 * clearing any stale v_path so later lookups can potentially
1439 * repopulate it with a valid path.
1443 ret
= dirtopath(vrootp
, vp
, buf
, buflen
, flags
, cr
);
1449 (void) VOP_CLOSE(vp
, FREAD
, 1, 0, cr
, NULL
);
1457 vnodetopath(vnode_t
*vrootp
, vnode_t
*vp
, char *buf
, size_t buflen
, cred_t
*cr
)
1459 return (vnodetopath_common(vrootp
, vp
, buf
, buflen
, cr
, 0));
1463 dogetcwd(char *buf
, size_t buflen
)
1468 refstr_t
*cwd
, *oldcwd
;
1470 pathname_t rpnp
, pnp
;
1471 proc_t
*p
= curproc
;
1474 * Check to see if there is a cached version of the cwd. If so, lookup
1475 * the cached value and make sure it is the same vnode.
1477 mutex_enter(&p
->p_lock
);
1478 if ((cwd
= PTOU(p
)->u_cwd
) != NULL
)
1480 vp
= PTOU(p
)->u_cdir
;
1482 mutex_exit(&p
->p_lock
);
1485 * Make sure we have permission to access the current directory.
1487 if ((ret
= VOP_ACCESS(vp
, VEXEC
, 0, CRED(), NULL
)) != 0) {
1495 value
= refstr_value(cwd
);
1496 if ((ret
= pn_get((char *)value
, UIO_SYSSPACE
, &pnp
)) != 0) {
1504 if (lookuppn(&pnp
, &rpnp
, NO_FOLLOW
, NULL
, &compvp
) == 0) {
1506 if (VN_CMP(vp
, compvp
) &&
1507 strcmp(value
, rpnp
.pn_path
) == 0) {
1512 if (strlen(value
) + 1 > buflen
) {
1514 return (ENAMETOOLONG
);
1516 bcopy(value
, buf
, strlen(value
) + 1);
1530 ret
= vnodetopath_common(NULL
, vp
, buf
, buflen
, CRED(),
1536 * Store the new cwd and replace the existing cached copy.
1539 cwd
= refstr_alloc(buf
);
1543 mutex_enter(&p
->p_lock
);
1544 oldcwd
= PTOU(p
)->u_cwd
;
1545 PTOU(p
)->u_cwd
= cwd
;
1546 mutex_exit(&p
->p_lock
);
1549 refstr_rele(oldcwd
);