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 2015 Nexenta Systems, Inc. All rights reserved.
24 * Copyright (c) 1988, 2010, Oracle and/or its affiliates. All rights reserved.
25 * Copyright 2016 Joyent, Inc.
28 /* Copyright (c) 1983, 1984, 1985, 1986, 1987, 1988, 1989 AT&T */
29 /* All Rights Reserved */
32 * University Copyright- Copyright (c) 1982, 1986, 1988
33 * The Regents of the University of California
36 * University Acknowledgment- Portions of this document are derived from
37 * software developed by the University of California, Berkeley, and its
41 #include <sys/types.h>
42 #include <sys/param.h>
43 #include <sys/systm.h>
44 #include <sys/cpuvar.h>
45 #include <sys/errno.h>
50 #include <sys/vnode.h>
51 #include <sys/pathname.h>
53 #include <sys/vtrace.h>
54 #include <sys/sysmacros.h>
55 #include <sys/debug.h>
56 #include <sys/dirent.h>
59 #include <sys/fs/snode.h>
61 /* Controls whether paths are stored with vnodes. */
62 int vfs_vnode_path
= 1;
72 return (lookupnameatcred(fnamep
, seg
, followlink
, dirvpp
, compvpp
, NULL
,
77 * Lookup the user file name,
78 * Handle allocation and freeing of pathname buffer, return error.
82 char *fnamep
, /* user pathname */
83 enum uio_seg seg
, /* addr space that name is in */
84 int followlink
, /* follow sym links */
85 vnode_t
**dirvpp
, /* ret for ptr to parent dir vnode */
86 vnode_t
**compvpp
, /* ret for ptr to component vnode */
87 vnode_t
*startvp
, /* start path search from vp */
88 cred_t
*cr
) /* credential */
90 char namebuf
[TYPICALMAXPATHLEN
];
91 struct pathname lookpn
;
94 error
= pn_get_buf(fnamep
, seg
, &lookpn
, namebuf
, sizeof (namebuf
));
96 error
= lookuppnatcred(&lookpn
, NULL
, followlink
,
97 dirvpp
, compvpp
, startvp
, cr
);
99 if (error
== ENAMETOOLONG
) {
101 * This thread used a pathname > TYPICALMAXPATHLEN bytes long.
103 if (error
= pn_get(fnamep
, seg
, &lookpn
))
105 error
= lookuppnatcred(&lookpn
, NULL
, followlink
,
106 dirvpp
, compvpp
, startvp
, cr
);
114 lookupnameat(char *fnamep
, enum uio_seg seg
, int followlink
,
115 vnode_t
**dirvpp
, vnode_t
**compvpp
, vnode_t
*startvp
)
117 return (lookupnameatcred(fnamep
, seg
, followlink
, dirvpp
, compvpp
,
123 struct pathname
*pnp
,
124 struct pathname
*rpnp
,
129 return (lookuppnatcred(pnp
, rpnp
, followlink
, dirvpp
, compvpp
, NULL
,
134 * Lookup the user file name from a given vp, using a specific credential.
138 struct pathname
*pnp
, /* pathname to lookup */
139 struct pathname
*rpnp
, /* if non-NULL, return resolved path */
140 int followlink
, /* (don't) follow sym links */
141 vnode_t
**dirvpp
, /* ptr for parent vnode */
142 vnode_t
**compvpp
, /* ptr for entry vnode */
143 vnode_t
*startvp
, /* start search from this vp */
144 cred_t
*cr
) /* user credential */
146 vnode_t
*vp
; /* current directory vp */
150 if (pnp
->pn_pathlen
== 0)
153 mutex_enter(&p
->p_lock
); /* for u_rdir and u_cdir */
154 if ((rootvp
= PTOU(p
)->u_rdir
) == NULL
)
156 else if (rootvp
!= rootdir
) /* no need to VN_HOLD rootdir */
159 if (pnp
->pn_path
[0] == '/') {
162 vp
= (startvp
== NULL
) ? PTOU(p
)->u_cdir
: startvp
;
165 mutex_exit(&p
->p_lock
);
168 * Skip over leading slashes
170 if (pnp
->pn_path
[0] == '/') {
174 } while (pnp
->pn_path
[0] == '/');
177 return (lookuppnvp(pnp
, rpnp
, followlink
, dirvpp
,
178 compvpp
, rootvp
, vp
, cr
));
182 lookuppnat(struct pathname
*pnp
, struct pathname
*rpnp
,
183 int followlink
, vnode_t
**dirvpp
, vnode_t
**compvpp
,
186 return (lookuppnatcred(pnp
, rpnp
, followlink
, dirvpp
, compvpp
, startvp
,
190 /* Private flag to do our getcwd() dirty work */
191 #define LOOKUP_CHECKREAD 0x10
192 #define LOOKUP_MASK (~LOOKUP_CHECKREAD)
195 * Starting at current directory, translate pathname pnp to end.
196 * Leave pathname of final component in pnp, return the vnode
197 * for the final component in *compvpp, and return the vnode
198 * for the parent of the final component in dirvpp.
200 * This is the central routine in pathname translation and handles
201 * multiple components in pathnames, separating them at /'s. It also
202 * implements mounted file systems and processes symbolic links.
204 * vp is the vnode where the directory search should start.
206 * Reference counts: vp must be held prior to calling this function. rootvp
207 * should only be held if rootvp != rootdir.
211 struct pathname
*pnp
, /* pathname to lookup */
212 struct pathname
*rpnp
, /* if non-NULL, return resolved path */
213 int flags
, /* follow symlinks */
214 vnode_t
**dirvpp
, /* ptr for parent vnode */
215 vnode_t
**compvpp
, /* ptr for entry vnode */
216 vnode_t
*rootvp
, /* rootvp */
217 vnode_t
*vp
, /* directory to start search at */
218 cred_t
*cr
) /* user's credential */
220 vnode_t
*cvp
; /* current component vp */
221 char component
[MAXNAMELEN
]; /* buffer for component (incl null) */
225 struct pathname presrvd
; /* case preserved name */
226 struct pathname
*pp
= NULL
;
228 vnode_t
*zonevp
= curproc
->p_zone
->zone_rootvp
; /* zone root */
229 int must_be_directory
= 0;
230 boolean_t retry_with_kcred
;
232 CPU_STATS_ADDQ(CPU
, sys
, namei
, 1);
236 rpnp
->pn_pathlen
= 0;
238 lookup_flags
= dirvpp
? LOOKUP_DIR
: 0;
239 if (flags
& FIGNORECASE
) {
240 lookup_flags
|= FIGNORECASE
;
246 * Eliminate any trailing slashes in the pathname.
247 * If there are any, we must follow all symlinks.
248 * Also, we must guarantee that the last component is a directory.
250 if (pn_fixslash(pnp
)) {
252 must_be_directory
= 1;
257 retry_with_kcred
= B_FALSE
;
260 * Make sure we have a directory.
262 if (vp
->v_type
!= VDIR
) {
267 if (rpnp
&& VN_CMP(vp
, rootvp
))
268 (void) pn_set(rpnp
, "/");
271 * Process the next component of the pathname.
273 if (error
= pn_getcomponent(pnp
, component
)) {
278 * Handle "..": two special cases.
279 * 1. If we're at the root directory (e.g. after chroot or
280 * zone_enter) then change ".." to "." so we can't get
281 * out of this subtree.
282 * 2. If this vnode is the root of a mounted file system,
283 * then replace it with the vnode that was mounted on
284 * so that we take the ".." in the other file system.
286 if (component
[0] == '.' && component
[1] == '.' && component
[2] == 0) {
288 if (VN_CMP(vp
, rootvp
) || VN_CMP(vp
, zonevp
)) {
290 } else if (vp
->v_flag
& VROOT
) {
295 * While we deal with the vfs pointer from the vnode
296 * the filesystem could have been forcefully unmounted
297 * and the vnode's v_vfsp could have been invalidated
298 * by VFS_UNMOUNT. Hence, we cache v_vfsp and use it
299 * with vfs_rlock_wait/vfs_unlock.
300 * It is safe to use the v_vfsp even it is freed by
301 * VFS_UNMOUNT because vfs_rlock_wait/vfs_unlock
302 * do not dereference v_vfsp. It is just used as a
304 * One more corner case here is the memory getting
305 * reused for another vfs structure. In this case
306 * lookuppnvp's vfs_rlock_wait will succeed, domount's
307 * vfs_lock will fail and domount will bail out with an
313 * This lock is used to synchronize
314 * mounts/unmounts and lookups.
315 * Threads doing mounts/unmounts hold the
316 * writers version vfs_lock_wait().
319 vfs_rlock_wait(vfsp
);
322 * If this vnode is on a file system that
323 * has been forcibly unmounted,
324 * we can't proceed. Cancel this operation
327 * vfs_vnodecovered is NULL if unmounted.
328 * Currently, nfs uses VFS_UNMOUNTED to
329 * check if it's a forced-umount. Keep the
330 * same checking here as well even though it
333 if (((vp
= cvp
->v_vfsp
->vfs_vnodecovered
) == NULL
) ||
334 (cvp
->v_vfsp
->vfs_flag
& VFS_UNMOUNTED
)) {
346 * Crossing mount points. For eg: We are doing
347 * a lookup of ".." for file systems root vnode
348 * mounted here, and fop_lookup() (with covered vnode)
349 * will be on underlying file systems mount point
350 * vnode. Set retry_with_kcred flag as we might end
351 * up doing fop_lookup() with kcred if required.
353 retry_with_kcred
= B_TRUE
;
359 * LOOKUP_CHECKREAD is a private flag used by vnodetopath() to indicate
360 * that we need to have read permission on every directory in the entire
361 * path. This is used to ensure that a forward-lookup of a cached value
362 * has the same effect as a reverse-lookup when the cached value cannot
365 if ((flags
& LOOKUP_CHECKREAD
) &&
366 (error
= fop_access(vp
, VREAD
, 0, cr
, NULL
)) != 0)
370 * Perform a lookup in the current directory.
372 error
= fop_lookup(vp
, component
, &cvp
, pnp
, lookup_flags
,
373 rootvp
, cr
, NULL
, NULL
, pp
);
376 * Retry with kcred - If crossing mount points & error is EACCES.
378 * If we are crossing mount points here and doing ".." lookup,
379 * fop_lookup() might fail if the underlying file systems
380 * mount point has no execute permission. In cases like these,
381 * we retry fop_lookup() by giving as much privilage as possible
382 * by passing kcred credentials.
384 * In case of hierarchical file systems, passing kcred still may
386 * For eg: UFS FS --> Mount NFS FS --> Again mount UFS on some
387 * directory inside NFS FS.
389 if ((error
== EACCES
) && retry_with_kcred
)
390 error
= fop_lookup(vp
, component
, &cvp
, pnp
, lookup_flags
,
391 rootvp
, zone_kcred(), NULL
, NULL
, pp
);
396 * On error, return hard error if
397 * (a) we're not at the end of the pathname yet, or
398 * (b) the caller didn't want the parent directory, or
399 * (c) we failed for some reason other than a missing entry.
401 if (pn_pathleft(pnp
) || dirvpp
== NULL
|| error
!= ENOENT
)
406 * We inform the caller that the desired entry must be
407 * a directory by adding a '/' to the component name.
409 if (must_be_directory
&& (error
= pn_addslash(pnp
)) != 0)
414 if (rootvp
!= rootdir
)
422 * Traverse mount points.
423 * XXX why don't we need to hold a read lock here (call vn_vfsrlock)?
424 * What prevents a concurrent update to v_vfsmountedhere?
425 * Possible answer: if mounting, we might not see the mount
426 * if it is concurrently coming into existence, but that's
427 * really not much different from the thread running a bit slower.
428 * If unmounting, we may get into traverse() when we shouldn't,
429 * but traverse() will catch this case for us.
430 * (For this to work, fetching v_vfsmountedhere had better
433 if (vn_mountedvfs(cvp
) != NULL
) {
434 if ((error
= traverse(&cvp
)) != 0)
439 * If we hit a symbolic link and there is more path to be
440 * translated or this operation does not wish to apply
441 * to a link, then place the contents of the link at the
442 * front of the remaining pathname.
444 if (cvp
->v_type
== VLNK
&& ((flags
& FOLLOW
) || pn_pathleft(pnp
))) {
445 struct pathname linkpath
;
447 if (++nlink
> MAXSYMLINKS
) {
452 if (error
= pn_getsymlink(cvp
, &linkpath
, cr
)) {
457 if (pn_pathleft(&linkpath
) == 0)
458 (void) pn_set(&linkpath
, ".");
459 error
= pn_insert(pnp
, &linkpath
, strlen(component
));
465 if (pnp
->pn_pathlen
== 0) {
469 if (pnp
->pn_path
[0] == '/') {
473 } while (pnp
->pn_path
[0] == '/');
478 if (pn_fixslash(pnp
)) {
480 must_be_directory
= 1;
486 * If rpnp is non-NULL, remember the resolved path name therein.
487 * Do not include "." components. Collapse occurrences of
488 * "previous/..", so long as "previous" is not itself "..".
489 * Exhausting rpnp results in error ENAMETOOLONG.
491 if (rpnp
&& strcmp(component
, ".") != 0) {
494 if (strcmp(component
, "..") == 0 &&
495 rpnp
->pn_pathlen
!= 0 &&
496 !((rpnp
->pn_pathlen
> 2 &&
497 strncmp(rpnp
->pn_path
+rpnp
->pn_pathlen
-3, "/..", 3) == 0) ||
498 (rpnp
->pn_pathlen
== 2 &&
499 strncmp(rpnp
->pn_path
, "..", 2) == 0))) {
500 while (rpnp
->pn_pathlen
&&
501 rpnp
->pn_path
[rpnp
->pn_pathlen
-1] != '/')
503 if (rpnp
->pn_pathlen
> 1)
505 rpnp
->pn_path
[rpnp
->pn_pathlen
] = '\0';
507 if (rpnp
->pn_pathlen
!= 0 &&
508 rpnp
->pn_path
[rpnp
->pn_pathlen
-1] != '/')
509 rpnp
->pn_path
[rpnp
->pn_pathlen
++] = '/';
510 if (flags
& FIGNORECASE
) {
512 * Return the case-preserved name
513 * within the resolved path.
515 error
= copystr(pp
->pn_buf
,
516 rpnp
->pn_path
+ rpnp
->pn_pathlen
,
517 rpnp
->pn_bufsize
- rpnp
->pn_pathlen
, &len
);
519 error
= copystr(component
,
520 rpnp
->pn_path
+ rpnp
->pn_pathlen
,
521 rpnp
->pn_bufsize
- rpnp
->pn_pathlen
, &len
);
523 if (error
) /* copystr() returns ENAMETOOLONG */
525 rpnp
->pn_pathlen
+= (len
- 1);
526 ASSERT(rpnp
->pn_bufsize
> rpnp
->pn_pathlen
);
531 * If no more components, return last directory (if wanted) and
532 * last component (if wanted).
534 if (pn_pathleft(pnp
) == 0) {
536 * If there was a trailing slash in the pathname,
537 * make sure the last component is a directory.
539 if (must_be_directory
&& cvp
->v_type
!= VDIR
) {
543 if (dirvpp
!= NULL
) {
545 * Check that we have the real parent and not
546 * an alias of the last component.
548 if (vn_compare(vp
, cvp
)) {
552 if (rootvp
!= rootdir
)
561 if (pnp
->pn_path
== pnp
->pn_buf
)
562 (void) pn_set(pnp
, ".");
566 if (VN_CMP(cvp
, rootvp
))
567 (void) pn_set(rpnp
, "/");
568 else if (rpnp
->pn_pathlen
== 0)
569 (void) pn_set(rpnp
, ".");
576 if (rootvp
!= rootdir
)
584 * Skip over slashes from end of last component.
586 while (pnp
->pn_path
[0] == '/') {
592 * Searched through another level of directory:
593 * release previous directory handle and save new (result
594 * of lookup) as current directory.
603 * Error. Release vnodes and return.
608 * If the error was ESTALE and the current directory to look in
609 * was the root for this lookup, the root for a mounted file
610 * system, or the starting directory for lookups, then
611 * return ENOENT instead of ESTALE. In this case, no recovery
612 * is possible by the higher level. If ESTALE was returned for
613 * some intermediate directory along the path, then recovery
614 * is potentially possible and retrying from the higher level
615 * will either correct the situation by purging stale cache
616 * entries or eventually get back to the point where no recovery
619 if (error
== ESTALE
&&
620 (VN_CMP(vp
, rootvp
) || (vp
->v_flag
& VROOT
) || vp
== startvp
))
623 if (rootvp
!= rootdir
)
631 * Traverse a mount point. Routine accepts a vnode pointer as a reference
632 * parameter and performs the indirection, releasing the original vnode.
635 traverse(vnode_t
**cvpp
)
645 * If this vnode is mounted on, then we transparently indirect
646 * to the vnode which is the root of the mounted file system.
647 * Before we do this we must check that an unmount is not in
648 * progress on this vnode.
653 * Try to read lock the vnode. If this fails because
654 * the vnode is already write locked, then check to
655 * see whether it is the current thread which locked
656 * the vnode. If it is not, then read lock the vnode
657 * by waiting to acquire the lock.
659 * The code path in domount() is an example of support
660 * which needs to look up two pathnames and locks one
661 * of them in between the two lookups.
663 error
= vn_vfsrlock(cvp
);
665 if (!vn_vfswlock_held(cvp
))
666 error
= vn_vfsrlock_wait(cvp
);
669 * lookuppn() expects a held vnode to be
670 * returned because it promptly calls
671 * VN_RELE after the error return
679 * Reached the end of the mount chain?
681 vfsp
= vn_mountedvfs(cvp
);
688 * The read lock must be held across the call to VFS_ROOT() to
689 * prevent a concurrent unmount from destroying the vfs.
691 error
= VFS_ROOT(vfsp
, &tvp
);
707 * Return the lowermost vnode if this is a mountpoint.
710 vn_under(vnode_t
*vp
)
715 while (vp
->v_flag
& VROOT
) {
718 vfs_rlock_wait(vfsp
);
719 if ((uvp
= vfsp
->vfs_vnodecovered
) == NULL
||
720 (vfsp
->vfs_flag
& VFS_UNMOUNTED
)) {
734 vnode_match(vnode_t
*v1
, vnode_t
*v2
, cred_t
*cr
)
736 vattr_t v1attr
, v2attr
;
739 * If we have a device file, check to see if is a cloned open of the
740 * same device. For self-cloning devices, the major numbers will match.
741 * For devices cloned through the 'clone' driver, the minor number of
742 * the source device will be the same as the major number of the cloned
745 if ((v1
->v_type
== VCHR
|| v1
->v_type
== VBLK
) &&
746 v1
->v_type
== v2
->v_type
) {
747 if ((spec_is_selfclone(v1
) || spec_is_selfclone(v2
)) &&
748 getmajor(v1
->v_rdev
) == getmajor(v2
->v_rdev
))
751 if (spec_is_clone(v1
) &&
752 getmajor(v1
->v_rdev
) == getminor(v2
->v_rdev
))
755 if (spec_is_clone(v2
) &&
756 getmajor(v2
->v_rdev
) == getminor(v1
->v_rdev
))
760 v1attr
.va_mask
= v2attr
.va_mask
= VATTR_TYPE
;
763 * This check for symbolic links handles the pseudo-symlinks in procfs.
764 * These particular links have v_type of VDIR, but the attributes have a
765 * type of VLNK. We need to avoid these links because otherwise if we
766 * are currently in '/proc/self/fd', then '/proc/self/cwd' will compare
769 if (fop_getattr(v1
, &v1attr
, 0, cr
, NULL
) != 0 ||
770 fop_getattr(v2
, &v2attr
, 0, cr
, NULL
) != 0 ||
771 v1attr
.va_type
== VLNK
|| v2attr
.va_type
== VLNK
)
774 v1attr
.va_mask
= v2attr
.va_mask
= VATTR_TYPE
| VATTR_FSID
| VATTR_NODEID
;
776 if (fop_getattr(v1
, &v1attr
, ATTR_REAL
, cr
, NULL
) != 0 ||
777 fop_getattr(v2
, &v2attr
, ATTR_REAL
, cr
, NULL
) != 0)
780 return (v1attr
.va_fsid
== v2attr
.va_fsid
&&
781 v1attr
.va_nodeid
== v2attr
.va_nodeid
);
786 * Find the entry in the directory corresponding to the target vnode.
789 dirfindvp(vnode_t
*vrootp
, vnode_t
*dvp
, vnode_t
*tvp
, cred_t
*cr
, char *dbuf
,
790 size_t dlen
, dirent64_t
**rdp
)
801 ASSERT(dvp
->v_type
== VDIR
);
804 * This is necessary because of the strange semantics of fop_lookup().
806 bzero(&pnp
, sizeof (pnp
));
812 uio
.uio_segflg
= UIO_SYSSPACE
;
814 uio
.uio_extflg
= UIO_COPY_CACHED
;
817 if ((error
= fop_access(dvp
, VREAD
, 0, cr
, NULL
)) != 0)
821 uio
.uio_resid
= dlen
;
825 (void) fop_rwlock(dvp
, V_WRITELOCK_FALSE
, NULL
);
826 error
= fop_readdir(dvp
, &uio
, cr
, &eof
, NULL
, 0);
827 fop_rwunlock(dvp
, V_WRITELOCK_FALSE
, NULL
);
829 dbuflen
= dlen
- uio
.uio_resid
;
831 if (error
|| dbuflen
== 0)
834 dp
= (dirent64_t
*)dbuf
;
835 while ((intptr_t)dp
< (intptr_t)dbuf
+ dbuflen
) {
837 * Ignore '.' and '..' entries
839 if (strcmp(dp
->d_name
, ".") == 0 ||
840 strcmp(dp
->d_name
, "..") == 0) {
841 dp
= (dirent64_t
*)((intptr_t)dp
+
846 error
= fop_lookup(dvp
, dp
->d_name
, &cmpvp
, &pnp
, 0,
847 vrootp
, cr
, NULL
, NULL
, NULL
);
850 * We only want to bail out if there was an error other
851 * than ENOENT. Otherwise, it could be that someone
852 * just removed an entry since the readdir() call, and
853 * the entry we want is further on in the directory.
856 if (vnode_match(tvp
, cmpvp
, cr
)) {
863 } else if (error
!= ENOENT
) {
867 dp
= (dirent64_t
*)((intptr_t)dp
+ dp
->d_reclen
);
872 * Something strange has happened, this directory does not contain the
873 * specified vnode. This should never happen in the normal case, since
874 * we ensured that dvp is the parent of vp. This is possible in some
875 * rare conditions (races and the special .zfs directory).
878 error
= fop_lookup(dvp
, ".zfs", &cmpvp
, &pnp
, 0, vrootp
, cr
,
881 if (vnode_match(tvp
, cmpvp
, cr
)) {
882 (void) strcpy(dp
->d_name
, ".zfs");
883 dp
->d_reclen
= strlen(".zfs");
898 * Given a global path (from rootdir), and a vnode that is the current root,
899 * return the portion of the path that is beneath the current root or NULL on
900 * failure. The path MUST be a resolved path (no '..' entries or symlinks),
901 * otherwise this function will fail.
904 localpath(char *path
, struct vnode
*vrootp
, cred_t
*cr
)
908 char component
[MAXNAMELEN
];
913 * We use vn_compare() instead of VN_CMP() in order to detect lofs
914 * mounts and stacked vnodes.
916 if (vn_compare(vrootp
, rootdir
))
919 if (pn_get(path
, UIO_SYSSPACE
, &pn
) != 0)
925 if (vn_ismntpt(vp
) && traverse(&vp
) != 0) {
931 while (pn_pathleft(&pn
)) {
934 if (pn_getcomponent(&pn
, component
) != 0)
937 if (fop_lookup(vp
, component
, &cvp
, &pn
, 0, rootdir
, cr
,
938 NULL
, NULL
, NULL
) != 0)
943 if (vn_ismntpt(vp
) && traverse(&vp
) != 0)
946 if (vn_compare(vp
, vrootp
)) {
947 ret
= path
+ (pn
.pn_path
- pn
.pn_buf
);
959 * Clean a stale v_path from a vnode. This is only performed if the v_path has
960 * not been altered since it was found to be stale
963 vnode_clear_vpath(vnode_t
*vp
, char *vpath_old
)
965 mutex_enter(&vp
->v_lock
);
966 if (vp
->v_path
!= vn_vpath_empty
&& vp
->v_path
== vpath_old
) {
967 vp
->v_path
= vn_vpath_empty
;
968 mutex_exit(&vp
->v_lock
);
969 kmem_free(vpath_old
, strlen(vpath_old
) + 1);
971 mutex_exit(&vp
->v_lock
);
976 * Validate that a pathname refers to a given vnode.
979 vnode_valid_pn(vnode_t
*vp
, vnode_t
*vrootp
, pathname_t
*pn
, pathname_t
*rpn
,
980 int flags
, cred_t
*cr
)
984 * If we are in a zone or a chroot environment, then we have to
985 * take additional steps, since the path to the root might not
986 * be readable with the current credentials, even though the
987 * process can legitmately access the file. In this case, we
990 * lookuppnvp() with all privileges to get the resolved path.
991 * call localpath() to get the local portion of the path, and
992 * continue as normal.
994 * If the the conversion to a local path fails, then we continue
995 * as normal. This is a heuristic to make process object file
996 * paths available from within a zone. Because lofs doesn't
997 * support page operations, the vnode stored in the seg_t is
998 * actually the underlying real vnode, not the lofs node itself.
999 * Most of the time, the lofs path is the same as the underlying
1000 * vnode (for example, /usr/lib/libc.so.1).
1002 if (vrootp
!= rootdir
) {
1006 if (lookuppnvp(pn
, rpn
, FOLLOW
, NULL
, &compvp
, rootdir
,
1007 rootdir
, kcred
) == 0) {
1008 local
= localpath(rpn
->pn_path
, vrootp
, kcred
);
1013 * The original pn was changed through lookuppnvp().
1014 * Set it to local for next validation attempt.
1017 (void) pn_set(pn
, local
);
1024 * We should have a local path at this point, so start the search from
1025 * the root of the current process.
1028 if (vrootp
!= rootdir
)
1030 if (lookuppnvp(pn
, rpn
, FOLLOW
| flags
, NULL
, &compvp
, vrootp
, vrootp
,
1033 * Check to see if the returned vnode is the same as the one we
1036 if (vn_compare(vp
, compvp
) ||
1037 vnode_match(vp
, compvp
, cr
)) {
1049 * Struct for tracking vnodes with invalidated v_path entries during a
1050 * dirtopath reverse lookup. By keeping adequate state, those vnodes can be
1051 * revisted to populate v_path.
1053 struct dirpath_walk
{
1054 struct dirpath_walk
*dw_next
;
1062 * Given a directory, return the full, resolved path. This looks up "..",
1063 * searches for the given vnode in the parent, appends the component, etc. It
1064 * is used to implement vnodetopath() and getcwd() when the cached path fails.
1067 dirtopath(vnode_t
*vrootp
, vnode_t
*vp
, char *buf
, size_t buflen
, int flags
,
1070 pathname_t pn
, rpn
, emptypn
;
1071 vnode_t
*pvp
= NULL
, *startvp
= vp
;
1075 char *bufloc
, *dbuf
;
1076 const size_t dlen
= DIRENT64_RECLEN(MAXPATHLEN
);
1077 struct dirpath_walk
*dw_chain
= NULL
, *dw_entry
;
1079 /* Operation only allowed on directories */
1080 ASSERT(vp
->v_type
== VDIR
);
1082 /* We must have at least enough space for "/" */
1084 return (ENAMETOOLONG
);
1086 /* Start at end of string with terminating null */
1087 bufloc
= &buf
[buflen
- 1];
1092 dbuf
= kmem_alloc(dlen
, KM_SLEEP
);
1093 bzero(&emptypn
, sizeof (emptypn
));
1096 * Begin with an additional reference on vp. This will be decremented
1103 hrtime_t cached_stamp
;
1106 * Return if we've reached the root. If the buffer is empty,
1107 * return '/'. We explicitly don't use vn_compare(), since it
1108 * compares the real vnodes. A lofs mount of '/' would produce
1109 * incorrect results otherwise.
1111 if (VN_CMP(vrootp
, vp
)) {
1112 if (*bufloc
== '\0')
1118 * If we've reached the VFS root, something has gone wrong. We
1119 * should have reached the root in the above check. The only
1120 * explantation is that 'vp' is not contained withing the given
1121 * root, in which case we return EPERM.
1123 if (VN_CMP(rootdir
, vp
)) {
1129 * Shortcut: see if this vnode has correct v_path. If so,
1130 * we have the work done.
1132 mutex_enter(&vp
->v_lock
);
1133 if (vp
->v_path
!= vn_vpath_empty
&&
1134 pn_set(&pn
, vp
->v_path
) == 0) {
1135 cached_stamp
= vp
->v_path_stamp
;
1136 mutex_exit(&vp
->v_lock
);
1137 rpn
.pn_path
= rpn
.pn_buf
;
1139 /* Ensure the v_path pointing to correct vnode */
1140 if (vnode_valid_pn(vp
, vrootp
, &pn
, &rpn
, flags
,
1142 complen
= strlen(rpn
.pn_path
);
1148 bcopy(rpn
.pn_path
, bufloc
, complen
);
1152 * Immediately nuke cached v_path entries known
1155 vn_clearpath(vp
, cached_stamp
);
1158 mutex_exit(&vp
->v_lock
);
1162 * Shortcuts failed, search for this vnode in its parent. If
1163 * this is a mountpoint, then get the vnode underneath.
1165 if (vp
->v_flag
& VROOT
)
1167 if ((err
= fop_lookup(vp
, "..", &pvp
, &emptypn
, 0, vrootp
, cr
,
1168 NULL
, NULL
, NULL
)) != 0)
1172 * With extended attributes, it's possible for a directory to
1173 * have a parent that is a regular file. Check for that here.
1175 if (pvp
->v_type
!= VDIR
) {
1181 * If this is true, something strange has happened. This is
1182 * only true if we are the root of a filesystem, which should
1183 * have been caught by the check above.
1185 if (VN_CMP(pvp
, vp
)) {
1191 * Check if we have read and search privilege so, that
1192 * we can lookup the path in the directory
1194 vprivs
= (flags
& LOOKUP_CHECKREAD
) ? VREAD
| VEXEC
: VEXEC
;
1195 if ((err
= fop_access(pvp
, vprivs
, 0, cr
, NULL
)) != 0) {
1200 * Search the parent directory for the entry corresponding to
1203 if ((err
= dirfindvp(vrootp
, pvp
, vp
, cr
, dbuf
, dlen
, &dp
))
1206 complen
= strlen(dp
->d_name
);
1208 if (bufloc
<= buf
) {
1212 bcopy(dp
->d_name
, bufloc
, complen
);
1214 /* Prepend a slash to the current path. */
1218 * Record the name and directory for later reconstruction and
1219 * link it up with the others.
1221 dw_entry
= kmem_alloc(sizeof (*dw_entry
), KM_SLEEP
);
1222 dw_entry
->dw_name
= kmem_alloc(complen
+ 1, KM_SLEEP
);
1223 VN_HOLD(dw_entry
->dw_vnode
= vp
);
1224 VN_HOLD(dw_entry
->dw_pvnode
= pvp
);
1225 bcopy(dp
->d_name
, dw_entry
->dw_name
, complen
+ 1);
1226 dw_entry
->dw_len
= complen
;
1227 dw_entry
->dw_next
= dw_chain
;
1228 dw_chain
= dw_entry
;
1230 /* And continue with the next component */
1237 * Place the path at the beginning of the buffer.
1240 ovbcopy(bufloc
, buf
, buflen
- (bufloc
- buf
));
1244 * Walk over encountered directory entries which were afflicted with a
1245 * stale or absent v_path. If the dirtopath was successful, we should
1246 * possess the necessary information to populate all of them with a
1249 * While processing this list, it is safe to call vn_setpath despite
1250 * the fact that racing vnode actions may have altered v_path entries
1251 * while the above loopwas still executing. Any updated entries will
1252 * have a newer v_path_stamp value which prevents an invalid overwrite.
1254 * If an error was encountered during the search, freeing the chain is
1257 dw_entry
= dw_chain
;
1258 while (dw_entry
!= NULL
) {
1259 struct dirpath_walk
*next
= dw_entry
->dw_next
;
1262 vn_setpath(NULL
, dw_entry
->dw_pvnode
,
1263 dw_entry
->dw_vnode
, dw_entry
->dw_name
,
1267 VN_RELE(dw_entry
->dw_vnode
);
1268 VN_RELE(dw_entry
->dw_pvnode
);
1269 kmem_free(dw_entry
->dw_name
, dw_entry
->dw_len
+ 1);
1270 kmem_free(dw_entry
, sizeof (*dw_entry
));
1275 * If the error was ESTALE and the current directory to look in
1276 * was the root for this lookup, the root for a mounted file
1277 * system, or the starting directory for lookups, then
1278 * return ENOENT instead of ESTALE. In this case, no recovery
1279 * is possible by the higher level. If ESTALE was returned for
1280 * some intermediate directory along the path, then recovery
1281 * is potentially possible and retrying from the higher level
1282 * will either correct the situation by purging stale cache
1283 * entries or eventually get back to the point where no recovery
1286 if (err
== ESTALE
&&
1287 (VN_CMP(vp
, vrootp
) || (vp
->v_flag
& VROOT
) || vp
== startvp
))
1290 kmem_free(dbuf
, dlen
);
1301 * The additional flag, LOOKUP_CHECKREAD, is used to enforce artificial
1302 * constraints in order to be standards compliant. For example, if we have
1303 * the cached path of '/foo/bar', and '/foo' has permissions 100 (execute
1304 * only), then we can legitimately look up the path to the current working
1305 * directory without needing read permission. Existing standards tests,
1306 * however, assume that we are determining the path by repeatedly looking up
1307 * "..". We need to keep this behavior in order to maintain backwards
1311 vnodetopath_common(vnode_t
*vrootp
, vnode_t
*vp
, char *buf
, size_t buflen
,
1312 cred_t
*cr
, int flags
)
1317 boolean_t doclose
= B_FALSE
;
1320 * If vrootp is NULL, get the root for curproc. Callers with any other
1321 * requirements should pass in a different vrootp.
1323 if (vrootp
== NULL
) {
1324 proc_t
*p
= curproc
;
1326 mutex_enter(&p
->p_lock
);
1327 if ((vrootp
= PTOU(p
)->u_rdir
) == NULL
)
1330 mutex_exit(&p
->p_lock
);
1336 * This is to get around an annoying artifact of the /proc filesystem,
1337 * which is the behavior of {cwd/root}. Trying to resolve this path
1338 * will result in /proc/pid/cwd instead of whatever the real working
1339 * directory is. We can't rely on fop_realvp(), since that will break
1340 * lofs. The only difference between procfs and lofs is that opening
1341 * the file will return the underling vnode in the case of procfs.
1343 if (vp
->v_type
== VDIR
&& fop_realvp(vp
, &realvp
, NULL
) == 0 &&
1346 if (fop_open(&vp
, FREAD
, cr
, NULL
) == 0)
1353 * Check to see if we have a valid cached path in the vnode.
1356 mutex_enter(&vp
->v_lock
);
1357 if (vp
->v_path
!= vn_vpath_empty
) {
1358 hrtime_t cached_stamp
;
1361 cached_stamp
= vp
->v_path_stamp
;
1362 (void) pn_set(&pn
, vp
->v_path
);
1363 mutex_exit(&vp
->v_lock
);
1365 /* We should only cache absolute paths */
1366 ASSERT(pn
.pn_buf
[0] == '/');
1369 if (vnode_valid_pn(vp
, vrootp
, &pn
, &rpn
, flags
, cr
) == 0) {
1370 /* Return the result, if we're able. */
1371 if (buflen
> rpn
.pn_pathlen
) {
1372 bcopy(rpn
.pn_path
, buf
, rpn
.pn_pathlen
+ 1);
1381 vn_clearpath(vp
, cached_stamp
);
1383 mutex_exit(&vp
->v_lock
);
1387 if (vp
->v_type
!= VDIR
) {
1389 * The reverse lookup tricks used by dirtopath aren't possible
1390 * for non-directory entries. The best which can be done is
1391 * clearing any stale v_path so later lookups can potentially
1392 * repopulate it with a valid path.
1396 ret
= dirtopath(vrootp
, vp
, buf
, buflen
, flags
, cr
);
1402 (void) fop_close(vp
, FREAD
, 1, 0, cr
, NULL
);
1410 vnodetopath(vnode_t
*vrootp
, vnode_t
*vp
, char *buf
, size_t buflen
, cred_t
*cr
)
1412 return (vnodetopath_common(vrootp
, vp
, buf
, buflen
, cr
, 0));
1416 dogetcwd(char *buf
, size_t buflen
)
1421 refstr_t
*cwd
, *oldcwd
;
1423 pathname_t rpnp
, pnp
;
1424 proc_t
*p
= curproc
;
1427 * Check to see if there is a cached version of the cwd. If so, lookup
1428 * the cached value and make sure it is the same vnode.
1430 mutex_enter(&p
->p_lock
);
1431 if ((cwd
= PTOU(p
)->u_cwd
) != NULL
)
1433 vp
= PTOU(p
)->u_cdir
;
1435 mutex_exit(&p
->p_lock
);
1438 * Make sure we have permission to access the current directory.
1440 if ((ret
= fop_access(vp
, VEXEC
, 0, CRED(), NULL
)) != 0) {
1448 value
= refstr_value(cwd
);
1449 if ((ret
= pn_get((char *)value
, UIO_SYSSPACE
, &pnp
)) != 0) {
1457 if (lookuppn(&pnp
, &rpnp
, NO_FOLLOW
, NULL
, &compvp
) == 0) {
1459 if (VN_CMP(vp
, compvp
) &&
1460 strcmp(value
, rpnp
.pn_path
) == 0) {
1465 if (strlen(value
) + 1 > buflen
) {
1467 return (ENAMETOOLONG
);
1469 bcopy(value
, buf
, strlen(value
) + 1);
1483 ret
= vnodetopath_common(NULL
, vp
, buf
, buflen
, CRED(),
1489 * Store the new cwd and replace the existing cached copy.
1492 cwd
= refstr_alloc(buf
);
1496 mutex_enter(&p
->p_lock
);
1497 oldcwd
= PTOU(p
)->u_cwd
;
1498 PTOU(p
)->u_cwd
= cwd
;
1499 mutex_exit(&p
->p_lock
);
1502 refstr_rele(oldcwd
);