On Mon, Jun 1, 2015 at 12:29 AM, Peter Geoghegan <pg@heroku.com> wrote:
> If you're using another well known MVCC database system that has RLS,
> I imagine when this happens the attacker similarly waits on the
> conflicting (privileged) xact to finish (in my example in the patch,
> Bob's xact). However, unlike with the Postgres READ COMMITTED mode,
> Mallory would then have her malicious UPDATE statement entirely rolled
> back, and her statement would acquire an entirely new MVCC snapshot,
> to be used by the USING security barrier qual (and everything else)
> from scratch. This other system would then re-run her UPDATE with the
> new MVCC snapshot. This would repeat until Mallory's UPDATE statement
> completes without encountering any concurrent UPDATEs/DELETEs to her
> would-be affected rows.
>
> In general, with this other database system, an UPDATE must run to
> completion without violating MVCC, even in READ COMMITTED mode. For
> that reason, I think we can take no comfort from the presumption that
> this flexibility in USING security barrier quals (allowing subqueries,
> etc) works securely in this other system. (I actually didn't check
> this out, but I imagine it's true).
Where are we on this?
Discussion during pgCon with Heikki and Andres led me to believe that
the issue is acceptable. The issue can be documented to help ensure
that user expectation is in line with actual user-visible behavior.
Unfortunately, I think that that will be a clunky documentation patch.
--
Peter Geoghegan