Rolf, Something is fishy with this patch. Looking in the file ompi/mca/pml/bfo/pml_bfo_failover.c in the function mca_pml_bfo_recv_restart_completion (line 1306) I see that in the case MCA_PML_BFO_HDR_TYPE_RECVERRNOTIFY we're using the recvreq without initializing it.
george. On Jul 13, 2010, at 06:46 , ro...@osl.iu.edu wrote: > Author: rolfv > Date: 2010-07-13 06:46:20 EDT (Tue, 13 Jul 2010) > New Revision: 23385 > URL: https://svn.open-mpi.org/trac/ompi/changeset/23385 > > Log: > New PML to support failover between openib BTLs. > openib BTL changes coming soon. > > Added: > trunk/ompi/mca/pml/bfo/ > trunk/ompi/mca/pml/bfo/Makefile.am > trunk/ompi/mca/pml/bfo/check-diffs-ob1.sh (contents, props changed) > trunk/ompi/mca/pml/bfo/configure.params > trunk/ompi/mca/pml/bfo/help-mpi-pml-bfo.txt > trunk/ompi/mca/pml/bfo/pml_bfo.c > trunk/ompi/mca/pml/bfo/pml_bfo.h > trunk/ompi/mca/pml/bfo/pml_bfo_comm.c > trunk/ompi/mca/pml/bfo/pml_bfo_comm.h > trunk/ompi/mca/pml/bfo/pml_bfo_component.c > trunk/ompi/mca/pml/bfo/pml_bfo_component.h > trunk/ompi/mca/pml/bfo/pml_bfo_failover.c > trunk/ompi/mca/pml/bfo/pml_bfo_failover.h > trunk/ompi/mca/pml/bfo/pml_bfo_hdr.h > trunk/ompi/mca/pml/bfo/pml_bfo_iprobe.c > trunk/ompi/mca/pml/bfo/pml_bfo_irecv.c > trunk/ompi/mca/pml/bfo/pml_bfo_isend.c > trunk/ompi/mca/pml/bfo/pml_bfo_progress.c > trunk/ompi/mca/pml/bfo/pml_bfo_rdma.c > trunk/ompi/mca/pml/bfo/pml_bfo_rdma.h > trunk/ompi/mca/pml/bfo/pml_bfo_rdmafrag.c > trunk/ompi/mca/pml/bfo/pml_bfo_rdmafrag.h > trunk/ompi/mca/pml/bfo/pml_bfo_recvfrag.c > trunk/ompi/mca/pml/bfo/pml_bfo_recvfrag.h > trunk/ompi/mca/pml/bfo/pml_bfo_recvreq.c > trunk/ompi/mca/pml/bfo/pml_bfo_recvreq.h > trunk/ompi/mca/pml/bfo/pml_bfo_sendreq.c > trunk/ompi/mca/pml/bfo/pml_bfo_sendreq.h > trunk/ompi/mca/pml/bfo/pml_bfo_start.c > trunk/ompi/mca/pml/bfo/post_configure.sh