[openib-general] [RFC] [PATCH 2/7] ib_multicast 2.6.20: add ib_multicast module to track join requests from the same port

[Sean Hefty]

Eitan Zahavi wrote:
> The point is not the fact you need to layer. But you can not enforce ref 
> counting by adding a top layer everybody can bypass.
> It simply breaks on the first client that goes directly to the lower level.
> Do you have a solution for this problem?
> The layer you need to add is BELOW the current interface not ABOVE it.

I disagree that there's an issue here.  Should we push the snooping down into 
ib_post_send()?  Into each driver's post_send()?

We're trusting kernel clients to call the right interfaces.  If we want to start 
worrying about kernel clients trying to bypass layers, there's nothing that 
prevents some client from allocating QP0 or 1 for itself (provided that it gets 
loaded first), snooping MADs and modifying their contents, redirecting requests, 
registering with the MAD layer to receive MADs of all types, sending MADs out 
other QPs, etc.

We should focus on coming up with the right architecture.  And the functionality 
that needs to reference count when to send join/leave requests to the SA is 
above whatever functionality is needed to actually send the MAD.

- Sean