Re: BUG #15520: PAM authentication + domain socket -> DNS query forsymbolic hostname [local]

On Sun, Nov 25, 2018 at 12:57 AM Thomas Munro
<thomas.munro@enterprisedb.com> wrote:
> It seems we shouldn't be passing a bogus hostname in PAM_RHOST.

I wonder if we should simply not set PAM_RHOST for Unix sockets, since
(in the words of the Linux PAM man page) "[i]n some applications,
PAM_RHOST may be NULL", or set it to the hostname of the local
machine, since in a general sense it is "the hostname of the machine
from which the PAM_RUSER entity is requesting service" (Linux PAM) and
"[t]he name of the applicant's host" (OpenPAM).  The latter doesn't
seem great because it means that a PAM module loses the ability to
distinguish this case from the IP case, so I think the former is
probably better.  I haven't managed to find anything explicit about
the expected value of PAM_RHOST for Unix sockets in either the Linux
PAM or OpenPAM projects.

I wonder if anyone out there has come to rely on the value "[local]"
that PostgreSQL generates for this case (ie in a custom PAM module or
script executed with pam_exec.so), and would get upset if we changed
it.  Seems pretty unlikely.

The comments for pg_getnameinfo_all() could probably do with a mention
of the special value written to "node" for Unix sockets.

Some relevant code:

We can see that linux-audit only starts trying to resolve host if you
didn't also pass in an address:

https://github.com/linux-audit/audit-userspace/blob/e42602b7b246ae62e7a12e9cd91f0ac37b1b1968/lib/audit_logging.c#L456

We can also see that linux-pam always passes NULL as an address:

https://github.com/linux-pam/linux-pam/blob/955b3e2f100205be2db4358e9c812de2ae453b8e/libpam/pam_audit.c#L41