[openib-general] Re: [PATCH] request/response matching in MAD code

Mon Oct 4 08:53:21 PDT 2004

On Fri, 2004-10-01 at 12:41, Sean Hefty wrote:
> On Fri, 01 Oct 2004 12:33:44 -0400
> Hal Rosenstock <halr at voltaire.com> wrote:
> > Also, should the TID be overwritten in the high 32 bits or do we trust
> > the client to set this properly ?
> 
> Based on our previous discussions on this, clients are responsible 
> for setting the upper 32-bits of the TID correctly.  

Yup, I recall this discussion.

> This should be more efficient, and only requires to use the TID provided 
> through registration.

It is a minor efficiency gain. The downside is the following:
If the client sets the hi_tid wrong (not corresponding to the mad_agent
supplied), the response either gets thrown away (if there is no hi_tid
match) or delivered to the "wrong" client callback.

A "similar" scenario exists when the client specifies a timeout and the
send is "unsolicited". In this case, the send although never getting a
response is timed out and not completed until then. It could be
completed when the send completion occurs ignoring the timeout.

With minor checks in both places, the MAD layer could protect against
these scenarios.

-- Hal