[openib-general] [PATCH] agent: Fix agent_mad_send PCI mapping and gather address and length

[Roland Dreier]

    Roland> I guess the problem with calling smi_handle_dr_smp_recv()
    Roland> twice on the same packet is that the function may alter
    Roland> the packet.

    Hal> No, the second call to smi_handle_dr_smp_recv() was on the
    Hal> outgoing response and not the incoming request. The thought
    Hal> was that a packet coming from process_mad is much like an
    Hal> incoming received packet and hence the call to
    Hal> smi_handle_dr_smp_recv. The routine validates the packet but
    Hal> also can do some fixups depending on which case it falls
    Hal> into. Guess it's only dangerous to validate this and wrong to
    Hal> fix it up.

Maybe I'm misreading the code, but my patch deleted the call to
smi_handle_dr_smp_recv() before the call to agent_send.  agent_send()
eventually ends up in ib_post_send_mad(), which calls
handle_outgoing_smp() for directed route MADs, which ends up calling
smi_handle_dr_smp_recv() again.  Since smi_handle_dr_smp_recv() can
change the packet, calling it twice on the same packet seems to break things.

However I don't think it's a good idea to think of responses generated
by process_mad as an incoming received packet.  I think they should be
thought of as returning DR SMPs being passed to the SMI for sending
(as in section 14.2.2 of the IB spec).

    Hal> The key to me is the following: The split of responsibility
    Hal> on the DR header formation is a little unclear to me. In the
    Hal> case of the SM, are the DR headers fully formed before
    Hal> handing it to the MAD layer or is some DR fixup needed ?

My suggestion would be to follow the IB spec, and assume that the SM
follows the SMP initialization in 14.2.2.1 and have the MAD layer just
implement the SMI processing in 14.2.2.2.  (And I believe things
should work similarly for responses generated by the SMA -- the MAD
layer should just do SMI processing).

 - R.