[openib-general] Mellanox request CQ notification _really_ compliant?

[Roland Dreier]

So I was reading the IB spec and I noticed the following compliance
statement:

    C10-113: A CI shall not generate a Completion Event for existing
    CQ entries on the specified CQ at the time the completion
    notification request is registered.

And I wonder if the well-known Mellanox behavior of generating an
event for existing CQEs is really compliant.  (Don't get me wrong --
the behavior is so useful that if it isn't compliant, then I think the
spec is what needs correction)

It seems to me that I could concoct a situation where a consumer could
actually prove that this compliance statement has been violated.  For
example I think the following example works:

 - create 2 CQs, A and B -- they are now known to be empty since they
   are freshly created.
 - create 2 QPs, X and Y, and connect them to each other (loopback).
   Use CQ A for QP X and CQ B for QP Y.
 - post a receive on QP X.
 - post a bind memory window operation on QP Y.
 - post an unsignaled send on QP Y.
 - poll CQ A until a receive completion appears (which it will, since
   the send from QP Y will be received on QP X)
 - We now know CQ B has exactly one completion, namely the bind
   completion, since the send was executed and C10-98.2.1 says that
   work requests after a bind may not begin execution until after the
   bind completes.
 - Request notification on CQ B.  We should never get an event, but we
   will on Mellanox HCAs.

I guess one way around this is to say that a work request may be
"completed" before the work completion has been entered in the
corresponding completion queue.  I don't see a rigorous definition of
what it means for a request to be completed in the IB spec.

 - R.