[openib-general] DAPL for openib

[Sean Hefty]

On Tue, 31 Aug 2004 18:23:59 -0400
Hal Rosenstock <halr at voltaire.com> wrote:

> > > This means that for any class, the IB MAD layer needs to know
> > > which methods are used for unsolicited MADs. 
> > 
> > I don't think that the MAD layer needs to know this information.  
> 
> Then can you remind me why we are passing in the ib_mad_reg_req
> structure which contains the method array bitmask ? (see below)
>... 
>
> Right and we don't go down to attribute right now but may need to. Isn't
> the method here the reason for the method bitmask ? This seems to be 
> at odds with the previous statement about the MAD layer not needing to
> know this. That would be true for responses but not unsolicited
> requests. Is that what you meant ?

I don't think we're matching our terminology.  The MAD layer needs to know how to route, and will operate based on class/version/method/etc., but doesn't need to know the specifics for any given class.  I.e. it routes based on the values of given fields, not their meaning, wherever possible.

The registration process is intended to provide the MAD layer a set of values that it uses to route with.  The MAD layer shouldn't care what those values mean.  This is what I meant when I said that the MAD layer doesn't need to "know" which methods are unsolicited.
 
> > Of course there are exceptions to the rule of simply checking the response bit, like RMPP.
> 
> Are there any exceptions you are aware of off the top of your head ?

I *think* just RMPP, which doesn't use the response bit in all cases (or has the bit flipped).  This is really an internal issue inside the MAD layer, but may affect the implementation of how MAD routing is done.