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

[Hal Rosenstock]

On Sun, 2006-04-09 at 18:08, Roland Dreier wrote:
>     Hal> I thought about using the high order bit of the rmpp_version
>     Hal> as this would never get to needing 8 bits. It seems this is
>     Hal> not a per registration thing though although that approach
>     Hal> would work. This has other minor downsides. Is this due to
>     Hal> adverseness to ioctls ?
> 
> Yeah, mostly adverseness to ioctls.  But it does seem to be a
> per-registration thing.  After all, I assume this is something that is
> a property of a MAD agent -- otherwise how does the MAD core know how
> to react to a two-sided RMPP when it gets one?

Right now, it needs to be burnt into the MAD core as there is no other
way to determine this (as it is class/method dependent). I have made a
spec comment on this lack of flexibility but don't have a good solution
for this at least yet...

I suppose one could argue it only applies to those classes/methods (e.g.
MAD agents which use those). It certainly can work that way. I think the
ioctl way is a little cleaner though but it's only a slight preference.
I started with the rmpp_version bit stealing approach.

-- Hal

>  - R.