[CVE Request] glibc iconv_open buffer overflow (was: Re: [oss-security] Re: glibc locale issues)

[Florian Weimer <fweimer () redhat com>]

On 07/21/2014 02:17 PM, Florian Weimer wrote:
> On 07/14/2014 04:15 AM, Tavis Ormandy wrote:
> > Tavis Ormandy <taviso () cmpxchg8b com> wrote:
> >
> > > I just remembered another charset issues I had looked into but
> > > abandoned.
> > >
> > > First of all, I think the need_so logic in gconv_trans is broken, but
> > > even
> > > if it worked there is an off by one error in __gconv_translit_find() (it
> > > does + 3 instead of + 3 + 1 in the allocation.
> >
> > To be clear, I suspect this is exploitable. It would be nice if you could
> > modify the buffer such that gconv will open a path with a string you've
> > appended it (e.g. CHARSET=//. pkexec ./../../../../tmp/foo.so),
>
> This is about the glib part and the alias processing, right?
>
> iconv/gconv_charset.h:strip() normalizes the transliteration argument to
> iconv_open, so the resulting file names follow a particular pattern, and
> there cannot be enough slashes to ascend to a writable directory.
>
> > if not maybe the one byte overflow is still exploitable.
>
> Hmm.  How likely is that?  It overflows in to malloc metadata, and the
> glibc malloc hardening should catch that these days.

Not necessarily on 32-bit architectures, so I agree with Tavis now, and 
we need a CVE.  The upstream bug is:

  <https://sourceware.org/bugzilla/show_bug.cgi?id=17187>

The discussion on the libc-alpha mailing list about the fix is still 
ongoing, and nothing has been committed yet.

--
Florian Weimer / Red Hat Product Security