[Rdma-developers] RE: [openib-general] OpenIB and OpenRDMA: Convergence on commonRDMA APIs and ULPs for Linux

[Roland Dreier]

    Caitlin> There isn't enough there to go farther.

I think there is.  In fact I outlined exactly what I would do if I
were working at an iWARP company:

    Roland> Then someone would have to start implementing a low-level
    Roland> driver for a specific RNIC, and find which modifications
    Roland> to the existing verbs are required.  For example, I
    Roland> believe the QP attribute structure passed into the QP
    Roland> modify verb probably has to become a union containing the
    Roland> IB attributes and the RNIC attributes.  However, most
    Roland> verbs should work fine with at most trivial modifications.

As for which option I was suggesting, it was c:

    Caitlin> c) The same methods but with struct/enums that have
    Caitlin> common and transport specific portions? That is doable,
    Caitlin> in fact that is what RNIC-PI is today. Repeating that
    Caitlin> work with the gen2 verbs will be time consuming.  I don't
    Caitlin> want to have to wait 4 months to debate the details of
    Caitlin> this before I can start working on my next generation of
    Caitlin> verbs.

OpenIB chose to focus on getting working code released quickly.
OpenRDMA chose to focus on writing architecture documents and API
specifications.  The results were completely unsurprising.

I have outlined what I believe are good and valid reasons why there
should not be two verbs layers in the Linux kernel.  If you think I'm
wrong and that you will be able to have RNIC-PI merged into the Linux
kernel alongside the existing IB midlayer, then implementing iWARP
support through RNIC-PI is a reasonable way forward.

If you believe that the single verbs layer in the kernel should be
RNIC-PI, then you should extend RNIC-PI to support the IB features
currently supported (eg datagrams), and port the existing IB code to
the new API.  Once that's done I would certainly be happy to review
the changes for merging.  However, I don't think any current OpenIB
contributors will be willing to do the work to port to RNIC-PI.  I see
no reason to spend a lot of effort to end up with a result that will
be at best equivalent to what we have today and will likely be worse
in some real ways (for example, how does RNIC-PI handle adapter hotplug?)

If neither of these options appeals to you, then the only alternative
left seems to be to work with the OpenIB community to evolve the
existing IB midlayer into an RDMA midlayer than can support both IB
and iWARP.

Without seeing some real patches from the iWARP side, it's hard for me
to see any value in continuing to participate in this debate.

 - R.