mca_bml_r2_del_procs() will do a double free if the BTL is presented on both btl_send and btl_rdma list, resulted in segfault in MPI_Finalize. (see both for loop at 620 and 637)
btl/uct and btl/ofi did a workaround on this issue by using hash table on del_proc. However, this will still be a problem for other BTL with both send and rdma capabilities. I just reproduced this on btl/openib. Should we fix this from r2 or the BTL has to implement their own workaround?
I can reproduce this on 4.0.0rc1.
(github pro tip: use permalinks to paste in code from a github repo)
I would say that this should be fixed in r2.
Most helpful comment
(github pro tip: use permalinks to paste in code from a github repo)
I would say that this should be fixed in r2.