Re: TCP/IP with 7.4 beta2 broken?

Andrew Dunstan <andrew(at)dunslane(dot)net> writes:
> There's a lot of confusion around :-) Let me see if I can disentangle
> some of it.

> People seem to want two things:
> 1. if ip4 is being tunneled over ip6 as it is in most Linux
> distributions, match a corresponding 'host*' line with an ip4 address.
> 2. enable local connections of whatever flavor by default.

Sounds right to me.

> Andreas has addressed item 1. I suggested an approach to item 2. The
> only alternative I can see is to allow ip4-only postmasters to recognize
> and silently drop ip6 'host*' lines. I don't like the idea of silently
> ignoring config lines - it seems dangerous to me. Suggestions of having
> initdb or something similar conditionally set the default pg_hba.conf
> also strike me as impractical and fragile.

Bruce and I were just discussing this on the phone. It seems we have
two basic approaches to problem #2. Either we hack the postmaster so
that it will swallow IPv6 addresses in pg_hba.conf even without any real
IPv6 support, or we make the default pg_hba.conf contents different.
Neither of these is real pretty, but I am leaning to the second, because
I agree with your feeling that silently ignoring config lines is a bad
idea.

I do not believe that there's anything fragile about having initdb make
this adjustment. We can arrange for initdb to be aware of the HAVE_IPV6
compilation flag (its value can be inserted when initdb is made from
initdb.sh, the same way some other configuration items are already
inserted into the script). As far as I can see, HAVE_IPV6 is exactly
what we want to look at to decide whether to put "::1" into pg_hba.conf.
If we HAVE_IPV6, then the postmaster can parse "::1". Whether the
kernel has IPv6 enabled doesn't matter --- if not, it would only mean
that the postmaster will never actually see a connection from "::1";
so the pg_hba.conf entry will never be matched. But it won't hurt
anything. Conversely, if we don't HAVE_IPV6, we can't parse "::1"
... but we don't need to, even if the kernel has IPv6, because such a
postmaster won't try to listen for AF_INET6 connections.

regards, tom lane