osc: Fix rdma component when not using ob1

When the ob1 PML was not eligible for selection (such as when the user
sets --mca pml cm), the BML and BTL frameworks are not initialized and
the rdma osc component will later fail as there are no BTLs available.
This patch resolves the issue by having the rdma osc component
initialize the BML interface.

Making this change required two additional, related changes.  First,
since the BTLs use the modex, the rdma initialization must be moved
before the modex point, so that putting data in the modex works as
expected.  Second, BTLs can require loading the entire world during
init (such as TCP when there are multiple threads and multiple NICs or
usnic), so we extend the world loading checks to include OSC.

Since the other Portals4 components say that they do require world
loading, we also assume the Portals4 osc component also requires
world loading.

Signed-off-by: Brian Barrett <[email protected]>

Author: bwbarrett

ompi/instance/instance.c

@@ -536,6 +536,10 @@ static int ompi_mpi_instance_init_common (int argc, char **argv)
         return ompi_instance_print_error ("mca_pml_base_select() failed", ret);
     }
 
+    if (OMPI_SUCCESS != (ret = ompi_osc_base_find_available (OPAL_ENABLE_PROGRESS_THREADS, ompi_mpi_thread_multiple))) {
+        return ompi_instance_print_error ("ompi_osc_base_find_available() failed", ret);
+    }
+
     OMPI_TIMING_IMPORT_OPAL("orte_init");
     OMPI_TIMING_NEXT("rte_init-commit");
 
@@ -617,10 +621,6 @@ static int ompi_mpi_instance_init_common (int argc, char **argv)
         return ompi_instance_print_error ("mca_coll_base_find_available() failed", ret);
     }
 
-    if (OMPI_SUCCESS != (ret = ompi_osc_base_find_available (OPAL_ENABLE_PROGRESS_THREADS, ompi_mpi_thread_multiple))) {
-        return ompi_instance_print_error ("ompi_osc_base_find_available() failed", ret);
-    }
-
     /* io and topo components are not selected here -- see comment
        above about the io and topo frameworks being loaded lazily */
 
@@ -654,7 +654,8 @@ static int ompi_mpi_instance_init_common (int argc, char **argv)
         return ompi_instance_print_error ("ompi_attr_create_predefined_keyvals() failed", ret);
     }
 
-    if (mca_pml_base_requires_world ()) {
+    if (mca_pml_base_requires_world() ||
+        mca_osc_base_requires_world()) {
         /* need to set up comm world for this instance -- XXX -- FIXME -- probably won't always
          * be the case. */
         if (OMPI_SUCCESS != (ret = ompi_comm_init_mpi3 ())) {
@@ -699,7 +700,8 @@ static int ompi_mpi_instance_init_common (int argc, char **argv)
     /* some btls/mtls require we call add_procs with all procs in the job.
      * since the btls/mtls have no visibility here it is up to the pml to
      * convey this requirement */
-    if (mca_pml_base_requires_world ()) {
+    if (mca_pml_base_requires_world() ||
+        mca_osc_base_requires_world()) {
         if (NULL == (procs = ompi_proc_world (&nprocs))) {
             return ompi_instance_print_error ("ompi_proc_get_allocated () failed", ret);
         }

ompi/mca/osc/base/osc_base_init.c

@@ -30,6 +30,8 @@
 #include "ompi/communicator/communicator.h"
 #include "ompi/win/win.h"
 
+bool ompi_osc_base_requires_world = false;
+
 int
 ompi_osc_base_select(ompi_win_t *win,
                      void **base,

ompi/mca/osc/osc.h

@@ -53,6 +53,9 @@ struct ompi_datatype_t;
 struct ompi_op_t;
 struct ompi_request_t;
 
+
+extern bool ompi_osc_base_requires_world;
+
 /* ******************************************************************** */
 
 
@@ -419,6 +422,11 @@ typedef ompi_osc_base_module_4_0_0_t ompi_osc_base_module_t;
 
 /* ******************************************************************** */
 
+static inline bool mca_osc_base_requires_world (void)
+{
+    return ompi_osc_base_requires_world;
+}
+
 
 END_C_DECLS
 

ompi/mca/osc/portals4/osc_portals4_component.c

@@ -349,6 +349,8 @@ component_init(bool enable_progress_threads, bool enable_mpi_threads)
         return ret;
     }
 
+    ompi_osc_base_requires_world = true;
+
     return OMPI_SUCCESS;
 }
 

ompi/mca/osc/rdma/osc_rdma_component.c

@@ -345,6 +345,27 @@ static int ompi_osc_rdma_component_init (bool enable_progress_threads,
                             __FILE__, __LINE__, ret);
     }
 
+    ret = mca_bml_base_init(enable_progress_threads, enable_mpi_threads);
+    if (OPAL_SUCCESS != ret) {
+        opal_output_verbose(1, ompi_osc_base_framework.framework_output,
+                            "%s:%d: bml_base_init() failed: %d",
+                            __FILE__, __LINE__, ret);
+        return ret;
+    }
+
+    /* check if any btls do not support dynamic add_procs */
+    mca_btl_base_selected_module_t* selected_btl;
+    OPAL_LIST_FOREACH(selected_btl, &mca_btl_base_modules_initialized,
+                      mca_btl_base_selected_module_t) {
+        mca_btl_base_module_t *btl = selected_btl->btl_module;
+
+        if (btl->btl_flags & MCA_BTL_FLAGS_SINGLE_ADD_PROCS) {
+            ompi_osc_base_requires_world = true;
+            break;
+        }
+
+    }
+
     return ret;
 }