Bug fixes

Signed-off-by: Joseph Antony <jajoseph.antony18@gmail.com>

Author: joe-explr

ompi/mca/osc/sm/osc_sm_comm.c

@@ -44,8 +44,8 @@ ompi_osc_sm_win_get_notify_value(struct ompi_win_t *win,
         return MPI_ERR_NOTIFY_IDX;
     }
 
-    *value = (OMPI_MPI_COUNT_TYPE) osc_sm_target_notify_base(module, rank)[notify];
     opal_atomic_rmb();
+    *value = (OMPI_MPI_COUNT_TYPE) osc_sm_target_notify_base(module, rank)[notify];
 
     return OMPI_SUCCESS;
 }
@@ -141,15 +141,15 @@ ompi_osc_sm_rput_notify(const void *origin_addr,
         return ret;
     }
 
+    if (notify < 0 || (uint32_t) notify >= module->node_states[target].notify_counter_count) {
+        return MPI_ERR_NOTIFY_IDX;
+    }
+
     /* the only valid field of RMA request status is the MPI_ERROR field.
      * ompi_request_empty has status MPI_SUCCESS and indicates the request is
      * complete. */
     *ompi_req = &ompi_request_empty;
 
-    if (notify < 0 || (uint32_t) notify >= module->node_states[target].notify_counter_count) {
-        return MPI_ERR_NOTIFY_IDX;
-    }
-
     opal_atomic_wmb();
     opal_atomic_add(&osc_sm_target_notify_base(module, target)[notify], 1);
 
@@ -483,7 +483,7 @@ ompi_osc_sm_get_notify(void *origin_addr,
         return ret;
     }
     if (notify < 0 || (uint32_t) notify >= module->node_states[target].notify_counter_count) {
-        return OMPI_ERR_BAD_PARAM;
+        return MPI_ERR_NOTIFY_IDX;
     }
 
     opal_atomic_rmb();

ompi/mca/osc/sm/osc_sm_component.c

@@ -312,7 +312,7 @@ component_select(struct ompi_win_t *win, void **base, size_t size, ptrdiff_t dis
             free(rbuf);
             goto error;
         }
-
+        /*  */
         total = 0;
         total_counters = 0;
         for (i = 0 ; i < comm_size ; ++i) {
@@ -391,9 +391,10 @@ component_select(struct ompi_win_t *win, void **base, size_t size, ptrdiff_t dis
         module->global_state = (ompi_osc_sm_global_state_t *) (module->posts[0] + comm_size * post_size);
         module->node_states = (ompi_osc_sm_node_state_t *) (module->global_state + 1);
 
-        /* set up notify counters in shared memory after node_states */
-        module->notify_counters = (uint64_t *) ((char *)(module->node_states + comm_size) +
-                                   OPAL_ALIGN_PAD_AMOUNT((uintptr_t)(module->node_states + comm_size), 64));
+        /* set up notify counters in shared memory after posts and state.
+         * posts_size and state_size are both padded to 64-byte multiples (lines 328, 330),
+         * and segment_base is page-aligned, so this is guaranteed 64-byte aligned. */
+        module->notify_counters = (uint64_t *)((char *) module->segment_base + posts_size + state_size);
         /* zero out notify counters */
         memset(module->notify_counters, 0, total_counters * sizeof(uint64_t));
 

ompi/mpi/c/Makefile.am

@@ -702,6 +702,7 @@ interface_profile_sources = \
 	get_elements_x_generated.c \
 	get_hw_resource_info_generated.c \
 	get_library_version_generated.c \
+	get_notify_generated.c \
 	get_processor_name_generated.c \
 	get_version_generated.c \
 	graph_create_generated.c \
@@ -815,6 +816,7 @@ interface_profile_sources = \
 	psend_init_generated.c \
 	publish_name_generated.c \
 	put_generated.c \
+	put_notify_generated.c \
 	query_thread_generated.c \
 	raccumulate_generated.c \
 	recv_generated.c \