Avoid a race condition vulnerability in chown, when used with

--from=O:G and without the (-h) --no-dereference option.

(restricted_chown): New function.
(change_file_owner): Call it.
Reported by Ulrich Drepper.
This commit is contained in:
Jim Meyering committed 2004-12-11 10:24:12 +00:00
1 parent 9278bc44cd
commit cbff85ac09
1 file changed
+151 -14
+151 -14
View File
@@ -24,19 +24,34 @@
#include <grp.h>
#include "system.h"
#include "chown-core.h"
#include "error.h"
#include "xfts.h"
#include "inttostr.h"
#include "lchown.h"
#include "quote.h"
#include "root-dev-ino.h"
#include "chown-core.h"
#include "xfts.h"
#ifndef _POSIX_VERSION
struct group *getgrnam ();
struct group *getgrgid ();
#endif
enum RCH_status
{
/* we called fchown and close, and both succeeded */
RC_ok = 2,
/* required_uid and/or required_gid are specified, but don't match */
RC_excluded,
/* SAME_INODE check failed */
RC_inode_changed,
/* open, fstat, fchown, or close failed */
RC_error,
};
extern void
chopt_init (struct Chown_option *chopt)
{
@@ -147,6 +162,86 @@ describe_change (const char *file, enum Change_status changed,
free (spec_allocated);
}
/* Change the owner and/or group of the FILE to UID and/or GID (safely)
only if REQUIRED_UID and REQUIRED_GID match the owner and group IDs
of FILE. ORIG_ST must be the result of `stat'ing FILE. If this
function ends up being called with a FILE that is a symlink and when
we are *not* dereferencing symlinks, we'll detect the problem via
the SAME_INODE test below.
The `safely' part above means that we can't simply use chown(2),
since FILE might be replaced with some other file between the time
of the preceding stat/lstat and this chown call. So here we open
FILE and do everything else via the resulting file descriptor.
We first call fstat and verify that the dev/inode match those from
the preceding stat call, and only then, if appropriate (given the
required_uid and required_gid constraints) do we call fchmod.
A minor problem:
This function fails when FILE cannot be opened, but chown/lchown have
no such limitation. But this may not be a problem for chown(1),
since chown is useful mainly to root, and since root seems to have
no problem opening `unreadable' files (on Linux). However, this can
cause trouble when non-root users apply chgrp to files they own but
to which they have neither read nor write access. For now, that
isn't a problem since chgrp doesn't have a --from=O:G option.
Return one of the RCH_status values. */
static enum RCH_status
restricted_chown (char const *file,
struct stat const *orig_st,
uid_t uid, gid_t gid,
uid_t required_uid, gid_t required_gid)
{
enum RCH_status status = RC_ok;
struct stat st;
int o_flags = (O_NONBLOCK | O_NOCTTY);
int fd = open (file, O_RDONLY | o_flags);
if (fd < 0)
{
fd = open (file, O_WRONLY | o_flags);
if (fd < 0)
return RC_error;
}
if (fstat (fd, &st) != 0)
{
status = RC_error;
goto Lose;
}
if ( ! SAME_INODE (*orig_st, st))
{
status = RC_inode_changed;
goto Lose;
}
if ((required_uid == (uid_t) -1 || required_uid == st.st_uid)
&& (required_gid == (gid_t) -1 || required_gid == st.st_gid))
{
if (fchown (fd, uid, gid) == 0)
{
status = (close (fd) == 0
? RC_ok : RC_error);
return status;
}
else
{
status = RC_error;
}
}
Lose:
{ /* FIXME: remove these curly braces when we assume C99. */
int saved_errno = errno;
close (fd);
errno = saved_errno;
return status;
}
}
/* Change the owner and/or group of the file specified by FTS and ENT
to UID and/or GID as appropriate.
If REQUIRED_UID is not -1, then skip files with any other user ID.
@@ -161,7 +256,7 @@ change_file_owner (FTS *fts, FTSENT *ent,
{
char const *file_full_name = ent->fts_path;
char const *file = ent->fts_accpath;
struct stat const *file_stats IF_LINT (= NULL);
struct stat const *file_stats;
struct stat stat_buf;
bool ok = true;
bool do_chown;
@@ -200,16 +295,22 @@ change_file_owner (FTS *fts, FTSENT *ent,
}
if (!ok)
do_chown = false;
{
do_chown = false;
file_stats = NULL;
}
else if (required_uid == (uid_t) -1 && required_gid == (gid_t) -1
&& chopt->verbosity == V_off && ! chopt->root_dev_ino)
do_chown = true;
{
do_chown = true;
file_stats = ent->fts_statp;
}
else
{
file_stats = ent->fts_statp;
/* If this is a symlink and we're dereferencing them,
stat it to get the permissions of the referent. */
stat it to get info on the referent. */
if (S_ISLNK (file_stats->st_mode) && chopt->affect_symlink_referent)
{
if (stat (file, &stat_buf) != 0)
@@ -237,14 +338,7 @@ change_file_owner (FTS *fts, FTSENT *ent,
if (do_chown)
{
if (chopt->affect_symlink_referent)
{
/* Applying chown to a symlink and expecting it to affect
the referent is not portable, but here we may be using a
wrapper that tries to correct for unconforming chown. */
ok = (chown (file, uid, gid) == 0);
}
else
if ( ! chopt->affect_symlink_referent)
{
ok = (lchown (file, uid, gid) == 0);
@@ -257,6 +351,47 @@ change_file_owner (FTS *fts, FTSENT *ent,
symlink_changed = false;
}
}
else
{
if ( required_uid == (uid_t) -1 && required_gid == (gid_t) -1)
{
ok = (chown (file, uid, gid) == 0);
}
else
{
/* Avoid a race condition with --from=O:G and without the
(-h) --no-dereference option. If fts' stat call determined
that the uid/gid of FILE matched the --from=O:G-selected
owner and group IDs, blindly using chown(2) here could lead
chown(1) or chgrp(1) mistakenly to dereference a *symlink*
to an arbitrary file that an attacker had moved into the
place of FILE during the window between the stat and
chown(2) calls. */
enum RCH_status err
= restricted_chown (file, file_stats, uid, gid,
required_uid, required_gid);
switch (err)
{
case RC_ok:
ok = true;
break;
case RC_error:
ok = false;
break;
case RC_inode_changed:
/* FIXME: give a diagnostic in this case? */
case RC_excluded:
do_chown = false;
ok = false;
goto Skip_chown;
default:
abort ();
}
}
}
/* On some systems (e.g., Linux-2.4.x),
the chown function resets the `special' permission bits.
@@ -272,6 +407,8 @@ change_file_owner (FTS *fts, FTSENT *ent,
quote (file_full_name));
}
Skip_chown:;
if (chopt->verbosity != V_off)
{
bool changed =