[openib-general] MAD receive work completion

[Roland Dreier]

    Sean> It could be done without an API change, likely changing 3-4
    Sean> lines of code, with the result that the work completion
    Sean> would be copied for all received MADs.  (The copy could be
    Sean> avoided with a more extensive change, but I would go with a
    Sean> simpler solution for now.)

Sorry, you're right.  I hadn't gone back and looked at the actual API,
and I didn't remember that ib_free_recv_mad() takes the whole
struct ib_mad_recv_wc (I thought it just took the ib_mad_recv_buf).

    Sean> To me, it seems that the behavior isn't what a user would
    Sean> expect given the current API.  The ib_mad_recv_wc belongs to
    Sean> the user until it is freed, but one of the fields in it
    Sean> exists only during the callback.  Is this the behavior that
    Sean> we want?

I agree, this is rather ugly.

Now I'm not sure whether it makes sense to change the wc member of
struct ib_mad_recv_wc from a struct ib_wc * to just a struct ib_wc.
On the one hand it makes the API slightly cleaner, but on the other
hand it is an incompatible change that may limit the internal
implementation of handling MAD receives.

 - Roland