[openib-general] CM oops in cm_free_msg()

[Libor Michalek]

On Wed, Jan 26, 2005 at 08:49:27PM -0800, Sean Hefty wrote:
> >  I'm seeing the following Oops on both passive and active connects.
> >On active side it's around theend of the REP callback, or the sending
> >of the RTU. On the passive side it appears to be around the sending
> >of the REP or receiving of the RTU.
> 
> Thanks for the feedback.  I'll take a look at the oops first off in the
> morning.

  On the active connection side, I see two cm_alloc_msg() calls,
presumably one for the REQ and one for the RTU. However, I only
see a single cm_free_msg, which is the one that is causing the
oops, and the fields in the structure appear to be garbage. Is
there any info I can gather for you, if you want me to take a
closer look, let me know how alloc and free should work...

> >  On another note, calling ib_cm_establish from CQ callback context, I
> >would expect to see a callback from the CM that I've entered ESTABLISHED
> >state, but looking at the code, that does not appear to be the case. Is
> >this something that wasn't intended, or just isn't there yet?
> 
> This wasn't intended.  The assumption being that the user knows that the
> connection was already established.  I don't think that it would be too hard
> to add.  (The only issue that jumps to mind at the moment could be a
> potential race condition processing a DREQ.)  Is this something that you
> need?  Would you need the callback from one of the CM's threads?

  Once the connection is in established state, on the passive side, I
need to make one more QP state transition. So if the CM doesn't make
an established callback, the consumer needs to create a work thread to
perform the QP state transition. 

  If a DREP is received ahead of an RTU, in REP_SENT state, it means that
the RTU was sent and lost. So the DREP implies a established transition.
How about receiving a DREP in REP_SENT, generates two transitions and
callbacks, established and timewait? Likewise, a sent RTU in the REP_RCVD
state would generate and established callback as well.


-Libor