[openib-general] Re: Some Initial RMPP Comments and Questions

Sean Hefty mshefty at ichips.intel.com
Fri Apr 29 15:09:48 PDT 2005


Hal Rosenstock wrote:
>>The receiving side doesn't perform a data copy.  It collects the separate
>>MAD buffers together in a list and hands those to the user.
> 
> 
> Yes. I meant user_mad.c needs a buffer to copy into on ib_umad_read and
> hence the size needs to be known ahead of time. That's where I think a
> peek might be useful (for RMPP, not for fixed size MADs).

By the time the kernel client gets the MAD, it's been reassembled, and 
the exact size that's needed is known.  So, I don't think that this is 
an issue for kernel clients.

Maybe you could add a peek for usermode or have read return the correct 
size if the requested size is too small.  Since you have to do a data 
copy for usermode anyway, I think it makes sense to just return the 
coalesced buffer.  This makes me think that I should implement 
ib_coalesce_recv_mad() now.

> I'm not sure how important streaming RMPP is. I would defer this.

I agree.  :)

- Sean



More information about the general mailing list