[openib-general] Re: change Mellanox SDP workaround to a module parameter

[Michael S. Tsirkin]

Quoting Sean Hefty <mshefty at ichips.intel.com>:
> Can you clarify what implementation is missing?

Here's an executive summary of the thread, so far (unfortunately at some
point In-Reply-To got stripped from some messages and so the thread
is broken on reflector, but you can search for Mellanox SDP workaround).

We are talking about a situation where a passive side wants to downgrade
the connecton MTU from the value provided by the active side.

This is important for Tavor which supports 2K MTU, but performs in RC much
better with 1K MTU.

The only way to do this in the existing CM is by sending REJ with
code 26, this REJ packet includes the MTU that can be supported.
The active side is supposed to check this and retry the connection
with the MTU value returned.

Note CMA will have to do it too at least when CMA manages the QP states itself.

Of course extra roundtrips and retries affect connection rate significantly.
I wish we had an MTU field in the REP packet, but we dont.

HTH,

-- 
Michael S. Tsirkin
Staff Engineer, Mellanox Technologies