[openib-general] DAPL for openib

[Sean Hefty]

On Tue, 24 Aug 2004 08:54:34 -0700
Roland Dreier <roland at topspin.com> wrote:

>     Roland> - Finish new MAD API (stalled?)
> 
>     Hal> Can you elaborate on what you mean by this ? Are there API
>     Hal> issues (other than the ones deferred) or is this referring to
>     Hal> an implementation to go with the API ?
> 
> I haven't been following this very closely but I didn't feel we had
> reached a final form of the API, and Sean's ib_mad.h has not been
> updated for several weeks.  Also (correct me if I'm wrong) you seem to
> be working on an implementation of a different API.

I thought that we were pretty much in agreement with the API as defined in ib_mad.h.  There are some MAD related issues in trunk/contrib/intel/TODO, however.  We decided to defer some of the QP redirection issues on the client side, but that may or may not affect the API.
 
> Also, I can think of at least one issue with the ib_mad.h API, and
> other people probably have their own issues.  My question is how we
> should split SM class queries such as PortInfo, which go to the SMA,
> from queries such as SMInfo, which needs to be passed to userspace to
> be handled by the SM.  Right now the ib_mad_reg() call just takes
> class, version and method, but both PortInfo and SMInfo will be
> identical in those attributes.

We need to discuss how to route MADs to the SMA.  Are you suggesting that the SMA would register for MADs the same as any other client?  My plan was to extend the ib_mad_reg_req structure as needed.  I limited it to version, class, method currently, since it easily allows O(1) routing.  We could add in AttributeID and AttributeModifier if those are needed.