[openib-general] RFC: detecting duplicate MAD requests

[Sean Hefty]

Hal Rosenstock wrote:
>>If everyone is okay with breaking the ABI, then I would add send completion 
>>notification to umad, and put the responsibility on callers not to generate 
>>duplicate responses.
> 
> Is this a better architectural solution ?

Not sure.  It doesn't solve supporting DS RMPP, which requires maintaining state 
between receiving a request and the generation of a response.

> I'm not sure I totally understand what the new ABI would be and its
> impact on existing applications. Is there an example of what this might
> look like ?

Currently, the only send MADs that are reported to the user are requests that 
time out waiting for a response.  We could probably change that to report all 
send completions.  Failed sends are reported using a status of timeout, with the 
MAD header copied to userspace.  So the length of the MAD indicates if it was a 
send or receive.

 From an implementation stand point, this approach likely requires only minor 
changes to the kernel code.  But any userspace applications that send MADs would 
need to change to handle this.  The list of application that do send MADs is 
likely fairly small however.

If we wanted to be more restrictive on which applications would be affected, we 
could only generate send completions for response MADs.

- Sean