AFAIK, the authentication is per-database, which means it has to read the database name from the client before authenticating, even if you are using certificates.
The bug in question involved the client sending an invalid database name.
Posted Apr 7, 2013 22:40 UTC (Sun) by hummassa (subscriber, #307)
[Link]
IIRC correctly my SSL/TLS, the authentication (and the start of encrypting the whole session) comes before anything else. I suppose when the client asks for connection to a database, he is already autheticated via his cert (and the dbms can check if he can effectively connect to *that* particular db).
TLS authorization
Posted Apr 8, 2013 12:44 UTC (Mon) by tialaramex (subscriber, #21167)
[Link]
I don't know specifically about PostgreSQL but generally when you have a single SSL endpoint you make all the "authorization" decisions later, after the SSL handshake is done and you're "secure".
So very likely you can connect to any TLS-supporting PostgreSQL server with proof that you're "A Badman <evil@example.com>" and the server will go try to access the database to determine whether this "A Badman" person is authorized. With this particular bug, that's already too late.
In principle the server _could_ round up all the rules for authorization for all the databases, and then do early reject for connections that don't match, but that would be a lot of extra work to handle a slow path case, and (unless you know about this bug) there's no reason to approach it that way.
