[openib-general] smpdump and current MAD layer

[Sean Hefty]

Hal Rosenstock wrote:

>>I guess filtering can be done above the MAD layer, 
> 
> That seems like the right way to go to me.

Same here.

> Then there are really only 2 snoopers at the MAD level (SMI, GSI). Any
> additional demux (snoopers) would be done above the MAD level.

I was referring to allowing multiple clients snoop QP0/1 traffic.  To 
implement this, it seems that we'd only need a single list per QP per port.

>>Also, would snooping include redirected QPs?  I think that 
>>we can support this.
> 
> Once a QP is redirected, is the MAD layer still involved in handing off
> the receive completions for that QP ? 

Currently, the API expects the user to call ib_process_mad_wc for 
*some* send or receive completions: those associated with RMPP and for 
requests/responses.  We can state that users of redirected QPs should 
always call ib_process_mad_wc for any MAD related work completion, but 
that isn't strictly enforceable as long as the user controls the CQ.