[openib-general] Re: roland-uverbs: CQ consumer index

[Michael S. Tsirkin]

Quoting r. Roland Dreier <roland at topspin.com>:
> Subject: Re: roland-uverbs: CQ consumer index
> 
>     Libor>   Without knowing how much code difference there will be
>     Libor> between Tavor and Arbel native mode, I'm wondering if it
>     Libor> makes sense at some point to have seperate hardware drivers
>     Libor> for the two parts? Just curious what the your thinking is
>     Libor> on the value of maintaining the same code base for the two.
> 
> I've gone back and forth between whether it's better to split mthca
> into two drivers or keep it as one driver that supports both Tavor and
> Arbel modes.  A lot of the code is common but there are also a lot of
> differences.  Right now I think keeping one driver is a little better
> because it's fairly easy to keep the Arbel-specific code in mthca, and
> I keep coming up with places where it would be very awkward to turn
> the common code into a library.  However I've gone back and forth in
> my mind a bunch of times so we may end up with two drivers before it's
> all over.
> 
>  - R.

So far the shared code looks good, so this simple approach seems to work out
well.
We are paying a bit in memory consumption, but nothing
drastic. I see no adverse effect on performance, at least as far as
IP over IB is concerned.


-- 
MST - Michael S. Tsirkin