Re: impending update of amanda packages to 2.5.2p1

[Greg Troxel <gdt%ir.bbn.com@localhost>]

Urban Boquist <urban%boquist.net@localhost> writes:

> Greg> I think chunker listens on a socket, and ends up on v6 instead
> Greg> of v4, and then the incoming v4 connections don't hit the v6
> Greg> socket (by default on BSD, seems to violate RFCs).
>
> Ahhhh, OK, that problem again! And yes, I can see that with netstat
> too, it is only listening on the v6 address.
>
> Greg> Do you have chunksize set?  Does your machine have v6 addresses?
>
> No chunksize set. I have the v6 address configured that you get "for
> free" with the v4 address, but nothing else.
>
> Greg> Try setting
> Greg> sysctl -w net.inet6.ip6.v6only=0
>
> Yup, that did it, thanks!

Glad to hear it - that raises my confidence in the diagnosis.

> I see in the mailing list archives that there has been a few quite
> heated discussions on the default for that sysctl, so I won't bring
> that up again... :-)
>
> But still, isn't this really an error in amanda? I mean it should not
> assume that this automatic v6-v4 mappings happens on all systems that
> support v6?

Well, my understanding is that there's a published specification for the
socket interface to IPv6 (RFC 2292), and that tihs specification says
that v4 connections match v6 listening sockets.  Amanda is relying on
this specification, and that's on one hand arguably reasonable.  On the
other hand, all of the BSDs have v6only=1 as a default, and it's not
reasonable of amanda to ignore that.

> It would be nice if amanda could work out of the box on NetBSD.

Yes, and we could do that by patching pkgsrc to skip v6, or to set
v6only=0 on the socket as it is created (there's a per-socket option).

I wrote to the amanda list when I first found the problem, and should
write again now that we understand it.  I would expect that they'd be
quite receptive to a patch that makes amanda listen on two sockets, one
v6 and one v4, or that sets v6only=0 for that socket.