btl/openib: fix rdmacm

[hjelmn]

The rdma_disconnect function specifies that both the server and client
should call rdma_disconnect. The code was not calling rdma_disconnect
on an endpoint if the event came before the endpoint finalization.

Signed-off-by: Nathan Hjelm hjelmn@lanl.gov

[hjelmn]

This was the only error I could see in the disconnect path. Will see if this fixes jenkins.

[jladd-mlnx]

bot:retest

[hjelmn]

I think this is a correct change given the wording in the man page. Probably needs to be tested by chelsio.

[larrystevenwise]

@bharatpotnuri - please test this out and also perhaps it fixes the current stall we're seeing?

[bharatpotnuri]

commit 17ae1ac works fine for iWARP.
I don't see the stall at MPI_Finalize which was seen earlier today with commit 80e362d. @larrystevenwise coming to the stall in gather test with higher message size, I am testing it out, I see it with combination of ofed and iwpm core, that should be separate issue I believe.
Thanks Jeff.

[jsquyres]

@bharatpotnuri Github pro tip: if you mention a non-escaped git hash from this repo in a comment, Github's web UI will auto-link it. E.g., 17ae1ac (vs. 17ae1ac, which isn't auto-linked, because it's in verbatim mode).

[hppritcha]

bot:retest
looking for ucx unload hang again.

[hppritcha]

bot:retest
looks like NERSC shutdown over weekend has leftover problems.

[lanl-ompi]

Test FAILed.

[lanl-ompi]

Test FAILed.

[hppritcha]

bot:retest

[hjelmn]

Any update on whether this breaks anything? If it doesn't I would like to merge this.

[jsquyres]

@larrystevenwise @bharatpotnuri Can you guys comment?

[larrystevenwise]

@bharatpotnuri please verify this change is good. Thanks!

[bharatpotnuri]

Commit 17ae1ac works fine for Chelsio iWARP.

[hjelmn]

@bharatpotnuri Thanks! I will go ahead and commit this. Not sure how rdmacm was working when it wasn't following the spec... Not sure this fixes any issue but at least it doesn't hurt.