[openib-general] [RFC] [PATCH] user_mad: Support RMPP on send side

[Roland Dreier]

    Sean> Why not copy down the ib_user_mad, or any necessary header
    Sean> information, separately from the actual MAD data itself?
    Sean> I.e. read in the mad_agent and MAD length first, allocate
    Sean> the send MAD, and then copy the data.  (I'll admit that I'm
    Sean> not overly familiar with this code, so if this is overly
    Sean> difficult just say so...)

    Sean> I'm concerned about having excessive copies on the opensm
    Sean> node, especially if it's sending a large number of RMPP
    Sean> packets.

I agree that it's better to avoid copies if it doesn't cost much in
complexity, but we should also keep in mind that a modern CPU has a
rather huge amount of memory bandwidth (gigabytes per second).  It's
hard to imagine that any remotely realistic RMPP workload could
consume even 1/100 of that.

In this case it does seem reasonable to copy the header part before
proceeding with the rest of the data.

 - R.