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

[Hal Rosenstock]

On Mon, 2006-04-10 at 16:23, Roland Dreier wrote:
>     Hal> It certainly is more deterministic if there is an ABI version
>     Hal> change in that the errors can be disambiguated better but is
>     Hal> it required if an ioctl is added or alternatively
>     Hal> rmpp_version bit is "stolen" ?
> 
> I'm pretty confused.  What would the ioctl do?  Return a bit saying
> whether or not dual-sided RMPP is supported? 

Yes.

>  What would happen if an
> old app that didn't know about dual-sided RMPP ran and didn't perform
> the ioctl?

Nothing. An old app wouldn't do dual sided RMPP so it doesn't need to
know. It's only a new app which might want to do DS RMPP which needs to
know.

> I guess this isn't really an ABI change -- old binaries continue to
> run as long as they don't try dual-sided RMPP (which doesn't work now).
> So maybe the answer is /sys/class/infiniband_mad/dual_sided_rmpp?
> 
> An ioctl for something that is not attached to a file descriptor isn't
> really the right thing.

Yes, a file based approach is another way.

Sean made me realize that this capability may not be so binary per
machine due to vendor class 2. I need to think more on how this would be
handled (burned into the RMPP core or a loadable kernel table seems like
the only ways now; I would like something more flexible in the IBA spec
(self identifying dual sided operations)).

-- Hal

>  - R.