coll/libnbc: do not recursively call opal_progress()

[ggouaillardet]

instead of invoking ompi_request_test_all(), that will end up
calling opal_progress() recursively, manually check the status
of the requests.

the same method is used in ompi_comm_request_progress()

Refs #3901

Signed-off-by: Gilles Gouaillardet gilles@rist.or.jp

[ggouaillardet]

@bosilca per my comments on #3901, i had the feeling that ompi_coll_libnbc_progress() being called recursively was either sub-optimal or a bad design.
This PR should fix that.

Generally speaking, is there any legitimate use case in which ompi_coll_libnbc_progress() is called recursively ? an other way to put it, is do we still need libnbc_in_progress ?

[bosilca]

Your change to the code looks legit. I don't have enough knowledge about libNBC to answer your question, but keep in mind that the progress can be also called every time we reach a resource exhaustion (such as fragments for BTL/PML). So, if you post new communications from a progress function, there is an opportunity to be called recursively (if new communications were posted from a progress function).

[artpol84]

bot:mellanox:retest

[artpol84]

vader issue, try again
bot:mellanox:retest

[hjelmn]

Looks good to me.

[ggouaillardet]

@kawashima-fj FYI

as pointed by @bosilca, MPI_Isend() might invoke opal_progress(), so there is still the possibility the coll/libnbc progress function is invoked recursively.