[openib-general] Re: [PATCH] RFC Verbs: add support for transport specific verbs

[Fabian Tillier]

On 3/1/06, Michael S. Tsirkin <mst at mellanox.co.il> wrote:
> Quoting Roland Dreier <rdreier at cisco.com>:
> > Subject: Re: [PATCH] RFC Verbs: add support for transport specific verbs
> >
> >     Michael> I agree.  Roland, what you might have missed is that the
> >     Michael> first solution posted added a requirement to check device
> >     Michael> type *in addition to checking the method is non-NULL and
> >     Michael> its return code* on each transport-specific call, so it
> >     Michael> had run-time penalty.
> >
> > I don't think I really agree with you that the device type checks are
> > necessary in most places.  For example, if we reach code that wants to
> > call the process_mad method, then we must already know that we're
> > dealing with an IB device.
>
> Sure. But still. There's already discussion in this thread on moving multicast
> stuff to ib-specific until iwarp defines it.

First, like Roland said, if code is going to try to attach a QP to a
multicast group, it likely already knows that it's using an IB device,
so the runtime check could just be a BUG_ON.

Second, things like multicast attach and process MAD are not speedpath
operations, so the extra check should really be inconsequential. 
Speed path operations, like post and poll verbs, would be in the
common device an not subject to such penalties.

- Fab