[openib-general] Re: Dual Sided RMPP Support as well as OpenSM Implications

[Hal Rosenstock]

On Fri, 2006-04-07 at 17:26, Roland Dreier wrote:
>     Hal> The simplest way I can think of to add this from an ABI/API
>     Hal> perspective is to add an ioctl to user_mad for this. Prior to
>     Hal> the kernel supporting this, libibumad can just return not
>     Hal> supported for the "is dual sided" check for the current
>     Hal> user_mad ABI version (which is 5). Anyone have any better
>     Hal> ideas on how to accomplish this ?
> 
> Maybe it would be simpler to add a flag to the agent registration
> request structure?  If you wanted to, you could even steal the high
> order bit of the QPN field and do this in a backwards compatible way.

I thought about using the high order bit of the rmpp_version as this
would never get to needing 8 bits. It seems this is not a per
registration thing though although that approach would work. This has
other minor downsides. Is this due to adverseness to ioctls ?

>     Hal> Also, there is the question of what should the existing RMPP
>     Hal> code do if it receives dual sided RMPP request on the network
>     Hal> side. It could ABORT this request although RMPP currently has
>     Hal> no specific status code for this. This is likely an issue
>     Hal> with any other dual sided RMPP implementations and there
>     Hal> would be no guarantee of interoperability here. It's unclear
>     Hal> what RMPP implementations which do not support dual sided
>     Hal> would do: likely just handle it as if it weren't dual sided.
> 
> I don't see how we can change the behavior of the existing code
> without a time machine -- it does whatever it does...

I didn't explain this well. I was suggesting to make a minor
modification to the existing RMPP code prior to dual sided being
supported.

-- Hal