[openib-general] Some Missing Features from mthca/user MAD ac cess

[Sean Hefty]

Hal Rosenstock wrote:

>>Since user-mode requires a data copy anyway, I'm wondering if it makes 
>>more sense for all user-mode clients to register in the kernel to snoop 
>>MADs, rather than to receive them.  The user-mode interface could then 
>>be modified to support more flexible filtering options.
> 
> Would this allow richer filtering options or also allow multiple
> receivers for the same MAD ?

Yes to both.  The filtering options would be dependent on the user-mode 
API and the user_mad implementation in the kernel.  I wouldn't think 
that the kernel MAD implementation would need to change.  I was just 
trying to point out that the main benefit to the API in the kernel 
doesn't apply in user-mode.

> Any idea on what the interface would look like ? Would it support
> attribute ID ? What else ?

I haven't given this more than two e-mails worth of thought, but it 
could be as simple as adding an attribute ID mask to something similar 
to the existing API.  A different approach could be to provide a mask 
over an entire MAD header.

> Are there any compelling reasons to do this ? 

*shrugs*  The disadvantage is that two applications would both receive 
the same request, so they would need to have some sort of coordination 
between them to know which one should generate a response.  (The 
"coordination" could be that they're coding to respond to different MADs.)

- Sean