[openib-general] gen2 dev branch

[Roland Dreier]

> I think of this more as providing a reliable communication transport.  It seems that this sort of feature would be duplicated by any client that sends a MAD and wants to get a response.

> If you push RMPP down into the access layer (which I think is desirable, given its complexity - even if it's a user-mode piece), then retransmissions should probably be pushed down as well.

> Did you have a different idea of how to isolate retransmissions from RMPP?

I think it's fine to do retransmissions after timeouts in the context
of an RMPP implementation, since the spec defines the timeout
explicitly.  If a consumer issues an RMPP send request, I agree
completely that the RMPP implementation should handle any timeouts
that happen during the send (and similarly for receiving multi-packet
messages).

The case I'm talking about is when, for example, a consumer sends a
path record request.  If no response is received after the consumer's
specified timeout, then we should just return that timeout to the
consumer and let the consumer decide whether and when to resend the
request.

 - Roland