btl/uct: clear related flags when there is no rdma transport layer

[ggouaillardet]

This fixes the use of btl/uct with the tcp memory_domain

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

[ggouaillardet]

@hjelmn can you please review this PR ?
@hoopoepg if no are using the tcp memory domain with btl/uct then you need this patch (otherwise Open MPI crashes). This is unrelated to #6137 though

[ggouaillardet]

:bot:retest

[ggouaillardet]

This patch is only good enough for eager message (e.g. small enough messages).

Function
main (osu_bw.c:117)
  PMPI_Isend (pisend.c:95)
    mca_pml_ob1_isend (pml_ob1_isend.c:198)
      mca_pml_ob1_send_request_start_seq (pml_ob1_sendreq.h:467)
        mca_pml_ob1_send_request_start_btl (pml_ob1_sendreq.h:432)
          mca_pml_ob1_send_request_start_rndv (pml_ob1_sendreq.c:811)
            mca_pml_ob1_rendezvous_hdr_prepare (pml_ob1_hdr.h:147)
              mca_pml_ob1_match_hdr_prepare (pml_ob1_hdr.h:102)
                mca_pml_ob1_common_hdr_prepare (pml_ob1_hdr.h:70)

MPI_Send() crashes with larger messages (a NULL header is usedin mca_pml_ob1_send_request_start_rndv() we use end up passing a NULL hdr to mca_pml_ob1_rendezvous_hdr_prepare().

[hjelmn]

I will see about checking this out in the next several weeks.

[hjelmn]

Note that I didn't really intend for btl/uct to be used with non-RDMA networks. It is worth fixing though.

[AboorvaDevarajan]

Can one of the admins verify this patch?

[ibm-ompi]

The IBM CI (GNU/Scale) build failed! Please review the log, linked below.

Gist: https://gist.github.com/1e23e7debe86706c7143d238a89b2b3c

[ibm-ompi]

The IBM CI (PGI) build failed! Please review the log, linked below.

Gist: https://gist.github.com/33d2372e313a2c9fba70b755f02dade8

[ibm-ompi]

The IBM CI (PGI) build failed! Please review the log, linked below.

Gist: https://gist.github.com/4eaaaa47f2a6244ce675ced588af4178