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

Sat Oct 23 14:02:56 PDT 2004

Hello!
Quoting r. Roland Dreier (roland at topspin.com) "Re: [openib-general] Handling SM class (SMInfo vs. other queries)":
>     Fab> This seems to imply that received MAD completion processing
>     Fab> will never happen in the context of the CQ notification
>     Fab> callback, and always require a context switch to a thread
>     Fab> context that can block while the MADs are handed off.  Am I
>     Fab> following correctly?  Do we want to have such a context
>     Fab> switch for every received MAD?
> 
> This design decision (to handle MAD completions in thread context) was
> made a while ago.  However I agree that long term it would make sense
> to move at least the SMA processing back to interrupt/tasklet context
> (and switch the local MAD interface to be asynch, of course).
> 
>  - R.

I think the difficulty with the last one is that (at least for Tavor)
process local mad can block, since there is a limited
number of outstanding commands. Of course you could always 
make it non-blocking by dropping the MAD if the command interface
is busy ... no idea if this will improve or decease performance, though.

MST