[openib-general][PATCH][RFC]: CMA IB implementation

[Caitlin Bestler]

> 
> >> We can avoid some synchronization by letting it exist 
> until the user 
> >> destroys the corresponding cma_id.  Otherwise, there's the 
> potential 
> >> of the user trying to destroy it twice.  Once from the 
> >> cma_connection_callback reporting an error, and then again here.
> > 
> > So you suggests this function will always return 0, then ?
> 
> I would recommend:  If the connection is reported to the 
> user, return 0.  If a new connection request cannot be 
> reported to the user (e.g. unable to allocate memory), return 
> a non-zero value to cleanup.
> 

Definitely, but put the emphasis on "cannot be reported
to the user". The reason could be "unable to allocate memory"
or any form of throttling on the maximum number of pending
connection requests. I'm not sure that the device SHOULD
forward all connection requests to the ULP without limit
until running out of allocatable memory.