[OpenAFS] multiple network interfaces in AFS

[Warren.Yenson@morganstanley.com]

> >In practise (on Solaris fileservers), we've seen that when although a
> >client will connect to qfe1:0, the return packets from the fileserver
> >have the source address of qfe1, which can fool the RX driver on the
> >client to continue the rest of the RX exchange to qfe1.  This is despite
> >using any combination of NetInfo and NetRestrict on the server.

> If I understand you correctly, true failover is not possible because the
> moment the standby machine takes over, some clients will want to "RX
> echange" with the "real" IP of the default active fileserver. Obviously,
> you have such a setup. What happens to those clients, do they recover
> somehow?

We use a kernel hack which post-sets the IP address of the socket for the
fileserver to the IP address for qfe1:0.  The kernel is forced into
thinking that the fileserver opened UDP 7000 with a specific IP address,
and will use that IP address as the source address on all outgoing UDP
7000 packets.

> Based on which criteria does the fileserver select the interface to
> return packets, closest IP to the client?

In our experience, unless we use the hack described above, the interface
selected is the one indicated by the kernel routing tables.