v2.x: fix pmix typos and timeout

[ggouaillardet]

No description provided.

[rhc54]

If we are going to handle the timeout up in the server, then we probably need to remove it from the pmix base function else we'll have a race to see who times out first.

[mellanox-github]

Test PASSed.
See http://bgate.mellanox.com/jenkins/job/gh-ompi-release-pr/1985/ for details.

[ggouaillardet]

i updated the PR,
now if you OPAL_PMIX_EXCHANGE(...,0) then MPI_Comm_accept() and MPI_Comm_connect() might hang forever.

under the hood, the timeout is triggered in pmix by libevent, and then it invokes the orted eviction callback.
if the timeout is 60 seconds, a timeout event will be generated every 2 seconds, and orted will check-in again into the opal hotel until the 60 seconds timeout expires.
if there is no pmix timeout, then the orted eviction callback is never called.

i do not understand what you mean by race condition.
the pmix timeout calls the orted "timeout" (simple function call, not through libevent)

[ggouaillardet]

Refs open-mpi/ompi#1905

[mellanox-github]

Test PASSed.
See http://bgate.mellanox.com/jenkins/job/gh-ompi-release-pr/1986/ for details.

[rhc54]

You are absolutely correct - my bad. I was mistakenly thinking that we also did a timeout event down in the opal layer as well.

Can you please bring the relevant parts back to master as well?

Thanks!

👍

[jsquyres]

Please comment back here on this PR after this stuff has been committed to master and gone through a night of MTT.

[rhc54]

went into master now

[jsquyres]

@hppritcha Good to go.