Eitan Zahavi wrote:
> Maybe I did not explain myself right.
> The idea is not to implement it in the mad.c code but rather to 
> implement it at the lowest level:
> The problem with a new API is that a single ULP/applications which does 
> direct umad or QP1 access will break the reference count.
> 
> Implementing at the lowest level - I.e. by sniffing QP1 packets - would 
> be enforced for all applications/ULPs.

Implementing this in the MAD layer is not the right solution.  Something, 
somewhere must send MADs and control access to QP0/1.  Something, somewhere 
must 
send MADs to the SA, and must be layered over the previous something. 
Something, somewhere needs to perform reference counting on multicast groups 
and 
request join/leave requests to the SA, and must be layered over the previous 
something.

Pushing all of this functionality down into a single module does not remove the 
need for the different layers of functionality.  It just produces an overly 
complex module.

- Sean

_______________________________________________
openib-general mailing list
openib-general@openib.org
http://openib.org/mailman/listinfo/openib-general

To unsubscribe, please visit http://openib.org/mailman/listinfo/openib-general

Reply via email to