Openwall GNU/*/Linux - a small security-enhanced Linux distro for servers
[<prev] [next>] [<thread-prev] [thread-next>] [day] [month] [year] [list]
Date: Wed, 30 Apr 2014 21:23:11 -0700
From: Andy Lutomirski <>
To: Solar Designer <>
Cc: Steve Grubb <>,
Subject: Re: local privilege escalation due to capng_lock as
 used in seunshare

On Wed, Apr 30, 2014 at 8:06 PM, Solar Designer <> wrote:
> On Thu, May 01, 2014 at 06:43:10AM +0400, Solar Designer wrote:
>> On Wed, Apr 30, 2014 at 09:27:10PM -0400, Steve Grubb wrote:
>> > And switching to NO_NEW_PRIVS broke the sandbox:
>> >
>> >
>> > So, perhaps fixing SECURE_NOROOT is the safest bet? Are there any other
>> > opinions on this?
>> If SECURE_NOROOT is meant to be usable to run entire Linux distros
>> (whether "on host" or/and "in containers"),
> Actually, I think it won't work well for that unless the distro in
> question doesn't use any SUID root programs that need capabilities,
> because SECURE_NOROOT breaks the raising of capabilities for SUID root
> exec (on purpose).  So generic implementations of containers capable of
> running arbitrary Linux distro userlands are probably not making use of
>> then it must not have an
>> effect of excluding UID 0 from "appropriate privileges" for setuid(2).
>> Do we know reliably that in this case excluding UID 0 from "appropriate
>> privileges" for setuid(2) was an effect specifically of SECURE_NOROOT?
> Per my quick greps, this does not appear to be the case.  The only
> checks for SECURE_NOROOT that I could find are in cap_bprm_set_creds(),
> so SECURE_NOROOT should affect execve(2), but not setuid(2).
> Why are we talking about it in this context, then?

I think that SECURE_NO_SETUID_FIXUP is actually at fault here.  And I
don't see how changing its semantics would help -- it's not safe to
run setuid programs without granting them capabilities, and it's not
really safe to grant them capabilities without setting euid == 0, and
it seems like it's unsafe to grant capabilities and euid == 0.

This leaves granting no extra privileges at all to setuid programs,
which is exactly what no_new_privs does.  seunshare sets up a weird
mount namespace, and anyone can use it and *configure* things about
that namespace.  no_new_privs blocks anything that causes execve to
grant new privileges, so the whole point is that it's safe to do
non-posixy things that are inherited by children as long as
no_new_privs is set.

I think I correctly analyzed exactly how no_new_privs broke sandbox in
the rhbz bug:

The short answer is that selinux doesn't currently distinguish whether
labels on executables are granting or removing privilege, and selinux
also fails to distinguish between the right to change labels on
request and the right to change labels because the policy said so [1],
so no_new_privs takes the conservative approach and blocks the whole
transition-on-exec mechanism.  And sandbox fails.  I suspect that
sandbox is already broken in the case where the program being
sandboxed is on a nosuid mount, because selinux's nosuid behavior is

In any event, I think that seunshare can be fixed by using dyntransition. Ugh.

A better fix might be to rewrite seunshare to use user namespaces
instead of requiring permissions.  This won't fly on RHEL5/6, though.


Powered by blists - more mailing lists

Your e-mail address:

Please check out the Open Source Software Security Wiki, which is counterpart to this mailing list.

Confused about mailing lists and their use? Read about mailing lists on Wikipedia and check out these guidelines on proper formatting of your messages.