[Mvapich-discuss] Crash in mvapich-4h.1

[Alex]

Hi,

I think there is a bug in MPIDI_IPC_rndv_cts_cb (ipc_control.c). When you
do rendezvous cts you use sreg->status.MPI_SOURCE as destination (later
in MPIDI_IPC_remote_complete) but I'm afraid in 'sending' requests this
field might not be filled properly.
As the results mpi crashes if MPIR_CVAR_CH4_IPC_RNDV_PROTOCOL is set to
auto  (messagie_size>MPIR_CVAR_CH4_IPC_CMA_P2P_THRESHOLD)  | coop | rput
(with number of  ranks > 2 ).

I fixed it locally by propagating correct value to MPI_SOURCE: inserting
        sreq->status.MPI_SOURCE = MPIDIG_REQUEST(sreq, u.send.dest);
before
        MPIDI_IPC_remote_complete(sreq, hdr->ipc_hdr.ipc_type);
around ipc_control.c:139, but you can choose a more 'elegant' way to fix
this.

Cheers,
Alex
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.osu.edu/pipermail/mvapich-discuss/attachments/20260107/2d0681f1/attachment.html>