[openib-general] [RMPP] RMPP formatting assumptions

Hal Rosenstock halr at voltaire.com
Fri Apr 1 04:08:29 PST 2005


On Thu, 2005-03-31 at 19:16, Sean Hefty wrote:
> So far, here are my assumptions regarding the formatting of the RMPP MADs.
> 
> The following fields in the RMPP header are set by the user:
> Version, Type = DATA, RTime, Flags = ACTIVE, and Status = 0

Should RMPP set the status rather than the user or is this an efficiency
thing ?

> The RMPP code will set the SegNum and update the Flags, but uses the 
> ACTIVE bit to determine if the user requires RMPP for a given transfer. 
>   I could easily have the RMPP code set some of these fields, but 
> thought that the caller might be able to initialize them more efficiently.
> 
> The WR length of a transfer should equal the size of the MAD header, 
> the RMPP header, class specific header for SA or vendor, plus a data 
> buffer that is evenly divisible by the size of the class' Data field. 
> This requirement is needed to prevent the RMPP code from allocating and 
> copying data segments.
> 
> The payload field in the RMPP header should be set to the size of the 
> class specific header plus the number of valid bytes of user data in 
> the data buffer.  The RMPP code will adjust the payload value to 
> account for multiple headers.

So it sounds like the streaming mode of RMPPis not handled on transmit.
It is optional on that side.

What about receive ? Can it/will it handle streaming mode ?

-- Hal



More information about the general mailing list