[openib-general] Re: [PATCH] CMA: allow/require bind beforeconnect

[James Lentini]

On Mon, 27 Mar 2006, Sean Hefty wrote:

> Roland Dreier wrote:
> > OK, fair enough.  I was really replying to the first sentence of:
> > 
> >      Caitlin> From the perspective of any given host, IP addresses are
> >      Caitlin> unique across all interface devices. A given connection
> >      Caitlin> can therefore be identified by just the 4-tuple, with no
> >      Caitlin> need to explicitly state "via this device".
> > 
> > IP addresses are not unique.  However, I do agree that a 4-tuple
> > uniquely identifies a TCP connection.
> 
> I agree with this as well.  But the CMA, when running over IB, does not
> establish TCP connections.  It's simply mapping addresses.  RDMA connections
> will end up being identified by QPs.

On IB, it also maps ports to service IDs.

This argues that the port values are within the scope of the CMA 
and that the active side behavior should be changed.

> If a user tries to establish a connection, the CMA will determine 
> which device that connection will go out on.  If it's an IB device, 
> there's no need for a local port number.  If it's an iWarp device, 
> then the iWarp CM will need to allocate a usable port number.
> 
> What I'm still trying to understand is why the CMA should allocate a 
> port number for active connections.  The port space over IB is 
> separate, and the port number is not needed for connecting or 
> routing data.  Are there specific applications that will run over 
> RDMA that will have a problem with this?

By my count, we've identified two applications: NFS-RDMA and SDP.