Re: Linux peer_cred Mischmasch

[Andy Lutomirski <luto () amacapital net>]

On 07/22/2014 11:43 PM, Sebastian Krahmer wrote:
> On Tue, Jul 22, 2014 at 12:22:30PM -0700, Andy Lutomirski wrote:
> > On 07/22/2014 04:17 AM, Florian Weimer wrote:
> > > On 07/22/2014 12:15 PM, Sebastian Krahmer wrote:
> > > > While maybe_add_creds() (via SOCK_PASSCRED) and scm_send()
> > > > (via unix_{stream,dgram}_sendmsg()) use the real UID,
> > > >
> > > > cred_to_ucred() (via SO_PEERCRED) passes the EUID (this time
> > > > also kuid_munged()).
> > >
> > > There should also be a discrepancy regarding when the credentials are
> > > captured (time of send for SOCK_PASSCRED, time of socket creation for
> > > SO_PEERCRED).  The latter is required because privileged processes
> > > assume that they can safely write to stderr, so picking the current
> > > process credentials may well introduce vulnerabilities.
>
> It does, and that should be ok.
>
> > >
> >
> > Indeed.  IMO both of these interfaces are flawed, but PASSCRED is
> > terminally broken and should never be used.  See, for example,
> > CVE-2013-1979, which is the immediate cause of the ruid thing.
>
> Thats what I was wondering whether CVE-2013-1979 only fixed SCM_CREDENTIALS
> case and missed to fix SO_PEERCRED.
> I am not fully convinced thats OK to get one time the euid and another time
> the uid (even though I liked the spy example:)

SO_PEERCRED is at least less bad: you'd need to convince someone to call
socket(2) (or maybe connect(2)) on your behalf, which is hopefully much
harder than getting them to call write(2) on your behalf.

I still don't like it, but that ship sailed a long, long time ago.

> Sebastian