Re: SCRAM authentication, take three

On 04/14/2017 10:33 PM, Peter Eisentraut wrote:
> On 4/11/17 01:10, Heikki Linnakangas wrote:
>> That question won't arise in practice. Firstly, if the server can do
>> scram-sha-256-plus, it presumably can also do scram-sha-512-plus. Unless
>> there's a change in the way the channel binding works, such that the
>> scram-sha-512-plus variant needs a newer version of OpenSSL or
>> something. Secondly, the user's pg_authid row will contain a
>> SCRAM-SHA-256 or SCRAM-SHA-512 verifier, not both, so that will dictate
>> which one to use.
>
> Right. So putting the actual password method in pg_hba.conf isn't going
> to be useful very often.
>
> I think the most practical thing that the user wants in pg_hba.conf is
> "best password method supported by what is in pg_authid". This is
> currently spelled "md5", which is of course pretty weird. And it will
> become weirder over time.
>
> I think we want to have a new keyword in pg_hba.conf for that, one which
> does not indicate any particular algorithm or method (so not "scram" or
> "sasl").
>
> We could use "password". If we think that "md5" can mean md5-or-beyond,
> then maybe "password" can mean password-or-md5-or-beyond.
>
> Or otherwise a completely new word.
>
> We also want to give users/admins a way to phase out old methods or set
> some policy. We could either make a global GUC setting
> password_methods='md5 scram-sha-256' and/or make that an option in
> pg_hba.conf past the method field.

Yeah, that would be reasonable. It can't be called just "password",
though, because there's no way to implement "password-or-md5-or-scram"
in a sensible way (see my reply to Simon at [1]). Unless we remove the
support for what "password" does today altogether, and redefine
"password" to mean just "md5-or-beyond". Which might not be a bad idea,
but that's a separate discussion.

In any case, I think we would probably still need more fine-grained
control, too, so we would still need to have "scram-sha-256" as a method
you can specify directly in pg_hba.conf. So I consider this as a
separate, new, feature that we can add in the future, if it seems worth
the effort.

I've committed a simple renaming of "scram" to "scram-sha-256", as the
pg_hba.conf and password_encryption option. I think that will do for v10.

[1]
https://www.postgresql.org/message-id/fa6cec54-4fa9-756d-53be-a5ba3d03d881@iki.fi

- Heikki