[openib-general] [PATCH v3 0/6] Tranport Neutral Verbs Proposal.

[James Lentini]

On Fri, 11 Aug 2006, Tom Tucker wrote:

> On Fri, 2006-08-11 at 10:20 -0400, James Lentini wrote:
> > 
> > On Wed, 9 Aug 2006, Krishna Kumar2 wrote:
> > 
> > > Hi James,
> > > 
> > > Sorry for the late response, my system was down and I just got it fixed.
> > > 
> > > > Is there a benefit to having rdmav_create_qp() take generic 
> > > > parameters if the application needs to understand the type of QP (IB, 
> > > > iWARP, etc.) created and the transport specific communication manager 
> > > > calls that are needed to manipulate it?
> > > > 
> > > > Would it make more sense if the QP create command was also transport 
> > > > specific?
> > > 
> > > My opinion is that the create_qp taking generic parameters is 
> > > correct, only subsequent calls may need to use transport specific 
> > > calls/arguments. Infact rdma_create_qp uses the ibv_create_qp (now 
> > > changed to rdmav_create_qp) call internally.
> > 
> > If you want to have a generic rdmav_create_qp() call, there needs to 
> > be programmatic way for the API consumer to determine what type of QP 
> > (iWARP vs. IB) was created.
> > 
> > I don't see any way to do that in your patch:
> 
> I think the QP is associated with the transport type indirectly through
> the context. It can be queried with ibv_get_transport_type verb.  A
> renamed rdma_get_transport type would probably suffice.

We don't have a userspace ibv_get_transport_type() verb. There is a 
kernel verb, but no userspace version.

> > http://openib.org/pipermail/openib-general/2006-August/024605.html
> > 
> > > PS : What is the opinion on this patchset ?
> > 
> > I like the new approach you are taking (keeping 1 verbs library and 
> > adding rdmav_ symbol names). This change to transport neutral names is 
> > long overdue.
> > 
> > When you finish with the userspace APIs, I hope you will update the 
> > kernel APIs as well.