[openib-general] RE: Making gen2 transport neutral

[Caitlin Bestler]

> -----Original Message-----

> 
> >This is a quick outline of the changes required to make ib_verbs.h 
> >transport neutral. Some of the changes were actually 
> required anyway to 
> >fully support IB 1.2.
> 
> Same comment on the following structure, it is not an RDMA 
> transport independent structure, it is an IB/iWarp only 
> structure that will require changes to support a new RDMA 
> capable device.
> 

The goal of having "transport neutral verbs" has to be understood.
My assumption is that it is not a "verb layer" interface if it
abstracts or limits the ability of system software to control
and understand what is happening. DAT/IT-API already provide
abstractions to allow *applications* to be unaware of these
details, the goal of a lower layer 'transport neutral' interface
is to preserve as much transport independence as possible *without*
the loss of information or control.

But of course to be useful it must be possible to do a substantial
number of operations in a transport neutral way. There is no point
to even having the interface if it is nothing more than endless
unions.

I believe this is achievable, and that is what is outlined. The
most commonly used operations can be invoked in a transport neutral
fashion. You can create PDs, virtual memory MRs, MWs, QPs, SRQs,
CQ and post to connected QPs without needing a single 'if(ib)'
statement.

You still must deal with connection management and many error
conditions in a transport neutral fashion.

I also believe that any third transport would proably be able
to define itself without the addition of any new connection
establishment controls, error returns or error completions
or asynchronous events. Any additions would be minor. The
large number of distinctions between IB and iWARP are a direct
result of two different models of how L4.5 (RDMA) interfaces
with L2-4. I doubt that a third transport would come up with
a third model. It will resemble iWARP or it will resemble IB.