Re: [PATCH rdma-next v3 00/13] RDMA/core: Ultra Ethernet transport verbs
From: Eric Davis
Date: Wed Nov 18 2026 - 11:31:54 EST
On Mon, Nov 16, 2026 at 10:03 AM Jason Gunthorpe <jgg@xxxxxxxx> wrote:
>
> On Mon, Nov 16, 2026 at 03:26:52PM +0000, Michael Margolin wrote:
> > On Mon, Nov 16, 2026 at 10:48:30AM -0400, Jason Gunthorpe wrote:
> > > On Mon, Nov 16, 2026 at 02:36:56PM +0000, Michael Margolin wrote:
> > > > On Mon, Nov 09, 2026 at 08:00:33PM -0800, Eric Davis wrote:
> > > > > Ultra Ethernet defines a reliable unconnected transport. A queue pair
> > > > > sends to any peer in its job without a connection, incoming traffic is
> > > > > demultiplexed on a JobID rather than on a connection, and what a device
> > > > > orders is far weaker than IB or RoCE and varies between devices. None of
> > > > > that has a home in the current uverbs ABI.
> > > > >
> > > > > This series adds the generic half. The objects, methods, and attributes
> > > > > any UET device implements. No driver is part of it, though two already
> > > > > exist against this ABI. A reference device model that implements all of
> > > > > it, and HPE's Slingshot driver implementing a large subset. That is the
> > > > > argument for ib_core carrying this. What is device specific should come
> > > > > from private interfaces per vendor.
> > > >
> > > > Hey Eric,
> > > >
> > > > Thanks for posting this. EFA's SRD is also a reliable unconnected
> > > > transport (one QP reaches any peer, delivery isn't ordered), so we
> > > > looked at the series as a possible second implementor of IB_QPT_RU
> > > > that isn't a UET device. Other vendors might also want to support
> > > > this interface in the future.
> > >
> > > I'm not sure about this. If you have exactly UET programming
> > > semantics, including the job, then sure.
> > >
> > > But squeezing some EFA specific thing that doesn't follow the spec
> > > isn't a good way to build the APIs.. Persumably EFA has its own
> > > job-like mechanism for security that is different?
> > >
> >
> > I see your point about trying to fit an API to a different use case it's
> > not meant for, but I don't think this is the case here. EFA doesn't have
> > a job-like mechanism, isolation is handled by the network, so jobs just
> > don't apply there.
>
> That sounds like you are saying you have a job for the entire
> node. Which isn't the UET model at all.
>
> I haven't looked at these patches, but I understood job should come
> from outside verbs as it is something an external entitiy is creating
> and then sharing across alot of processes.
>
> If so then EFA could still have a singleton job object and that could
> be the deviation. One job only, and the job doesn't actually do
> anything to limit access.
>
> Jason
In the UET spec v1.1, Sean Hefty and I contributed a new section
defining the Verbs mapping over UET (if you're curious, it complements
this patch series and is a short read).
One key design point for RU QPs is that we tried to separate them from
the transport. Specifically the GID type selects the transport and an
RU QP should work over any transport that supports reliable
unconnected. Of course each supported transport may require different
attributes to come from the QP configuration. I'm totally fine with
making some attributes optional and requiring the provider driver to
verify the given attributes.
Based on Michael's porting effort, minor tweaks may be needed in some
cases. I will follow up in each of those patch threads.
Curious what others think about this direction...?
- eric
Attachment:
smime.p7s
Description: S/MIME Cryptographic Signature