[ofa-general] [PATCHv2] IB/ipoib: S/G and HW checksum support

[Jason Gunthorpe]

On Wed, Sep 05, 2007 at 11:35:06PM +0300, Or Gerlitz wrote:
> > Judging by the other comments in this thread, it still seems to me
> > this would be best as RC only, notionally with the idea that RC is
> > only used between hosts and not between gateways and hosts
> > (administratively configured). That way the end-to-end nature of the
> > checksum is retained. Gateways that want to support RC can negotiate
> > this feature off.

> I guess by "RC" you mean connected mode. The connected mode is now
> implemented over RC but as was discussed over this list few times, it
> should (and it would) move to use UC, which is also much easier to
> implement in hw based gateways. Anyway, your idea to allow this
> feature coming into play only under negotiation schem sounds fine to
> me, however:

Sure.. Though, I'm not sure what advantage UC/RC brings to a gateway app
when you can't pass 64k MTU onto ethernet...

> > You may also want to look at using the new TSO/GSO/LRO stuff in a RC
> > context. If you could send an entire GSO in one go and receive it as a
> > LRO that might be a big improvement too.
> 
> From Michael's and Eli's responses over the stateless offload related
> thread, I understood that these optimizations are supported only for
> UD QPs, which makes them irrelevant for the  connected mode.

Right, but I'm not suggesting using the chips offload.

Micheal has made it so you can use 'csum offload' (via disabling csum)
on any nic. You can also do the same kind of thing for TSO/GSO. If you
send jumbo TSO/GSO packets in a chunk the receiver can then do
LRO. Win all around. Sort of like jumbo MTU but without actually
changing the MTU.

This is all basically the same set of techniques we see between a
Linux guest and the linux host in a virtualization environment.

Jason