[openib-general] Handling SM class (SMInfo vs. other queries)

Fri Oct 22 14:32:07 PDT 2004

> From: Roland Dreier [mailto:roland at topspin.com]
> Sent: Friday, October 22, 2004 2:26 PM
> 
> All these special cases are starting to look pretty awkward.  Another
> solution is to get rid of the snoop_mad entry point and the SMA and
> PMA agents in ib_smi, and just pass every MAD to the low-level
> driver.  The low-level driver can decide to handle it (consume it
> and/or generate a response) or not.  Everything not consumed by the
> low-level driver gets put through agent dispatch as usual.
> 

This seems to imply that received MAD completion processing will never
happen in the context of the CQ notification callback, and always require a
context switch to a thread context that can block while the MADs are handed
off.  Am I following correctly?  Do we want to have such a context switch
for every received MAD?

This is where having a local mad interface that is asynchronous would help
quite a bit.

- Fab