[openib-general] DAPL for openib

[Roland Dreier]

    Sean> My preference would be for the implementation to hand the
    Sean> MAD directly to the client who needs to process it, and for
    Sean> that algorithm to be O(1).

This is a reasonable goal.  However here's one more SMA wrinkle that I
just remembered.  When a Mellanox HCA generates a trap for the SM, it
doesn't send it directly -- the trap shows up on the receive queue of
QP0 locally with a source LID of 0 instead.  The low-level driver
needs to know about this and forward it on to the SM.

This leads to some difficulty on a node where the SM is running, since
the SM will also want to see SM traps received on QP0 (obviously).

The way we handled this in the Topspin drivers was to let the
SMA/low-level driver see all MADs before they got dispatched.  I guess
we need to decide what to do for OpenIB.

 - R.