Healing the bash fork (was: Re: [oss-security] CVE-2014-6271: remote code execution through bash)

[Florian Weimer <fweimer () redhat com>]

On 09/28/2014 03:39 AM, Chet Ramey wrote:
> OK, here are the more-or-less final versions of the patches for bash-2.05b
> through bash-4.3.  I made two changes from earlier today: the function
> export suffix is now `%%', which is not part of a the set of valid variable
> name characters but avoids any potential problems with including
> shell metacharacters in the name; and this version refuses to import shell
> functions whose name contains a slash, for reasons I discussed earlier.

Chet, thanks for posting an official version of the prefix/suffix patch.

I looked at how the “%%” encoding works with Debian's “at” (which is 
also used by Fedora and downstreams).  Unfortunately, it does not 
address the issue, “at” still prints error messages, both with dash as 
/bin/sh and bash.  As a result, I wonder if a suffix which is actually 
within the shell variable syntax wouldn't be a better choice (e.g., 
three randomly chosen alphanumerics), as that would make the “at” 
environment serialization code work again.  (I'm not concerned about 
“at” specifically, we'll change it anyway, it's about similar code out 
there which we don't know about it yet.)

Eric, does “%%” even work for Cygwin, or does it cause strange effects 
there?  (For the Windows shell, “%” is the variable starter character, a 
bit like “$” in sh-type shells.)

Related to that is that we should try to converge back to uniform bash 
behavior across distributions.  Right now, the majority seems to use 
“()” as the suffix (which is problematic, per the above), and they also 
reject characters such as “.:-” in import function names (a restriction 
which was inherited from the first patch which only tried to block 
command execution).  The latest upstream patch uses “%%”, and allows 
anything allowed in a regular function definition, except absolute 
pathnames.

I'm not sure how to move towards a common solution.  I think avoiding 
non-serializable environments could be a compelling reason to switch the 
suffix, but “%%” does not provide that.

(From a security POV, *requiring* that imported functions contain at 
least one special character would actually be best, but obviously, 
that's not backwards-compatible.)

--
Florian Weimer / Red Hat Product Security