OSC rdma: make sure accumulating in shared memory is safe

Signed-off-by: Joseph Schuchart <schuchart@hlrs.de>
(cherry picked from commit c67e229193)

ompi/mca/osc/rdma/osc_rdma.h

@@ -145,6 +145,9 @@ struct ompi_osc_rdma_module_t {
 
     bool acc_use_amo;
 
+    /** whether the group is located on a single node */
+    bool single_node;
+
     /** flavor of this window */
     int flavor;
 

ompi/mca/osc/rdma/osc_rdma_accumulate.c

@@ -889,10 +889,19 @@ int ompi_osc_rdma_rget_accumulate_internal (ompi_osc_rdma_sync_t *sync, const vo
         (void) ompi_osc_rdma_lock_acquire_exclusive (module, peer, offsetof (ompi_osc_rdma_state_t, accumulate_lock));
     }
 
+    /* accumulate in (shared) memory if there is only a single node
+     * OR if we have an exclusive lock
+     * OR if other processes won't try to use the network either */
+    bool use_shared_mem = module->single_node ||
+                          (ompi_osc_rdma_peer_local_base (peer) &&
+                              (ompi_osc_rdma_peer_is_exclusive (peer) ||
+                                  !module->acc_single_intrinsic));
+
     /* if the datatype is small enough (and the count is 1) then try to directly use the hardware to execute
      * the atomic operation. this should be safe in all cases as either 1) the user has assured us they will
-     * never use atomics with count > 1, 2) we have the accumulate lock, or 3) we have an exclusive lock */
-    if (origin_extent <= 8 && 1 == origin_count && !ompi_osc_rdma_peer_local_base (peer)) {
+     * never use atomics with count > 1, 2) we have the accumulate lock, or 3) we have an exclusive lock.
+     * avoid using the NIC if the operation can be done directly in shared memory. */
+    if (origin_extent <= 8 && 1 == origin_count && !use_shared_mem) {
         if (module->acc_use_amo && ompi_datatype_is_predefined (origin_datatype)) {
             if (NULL == result_addr) {
                 ret = ompi_osc_rdma_acc_single_atomic (sync, origin_addr, origin_datatype, origin_extent, peer, target_address,

ompi/mca/osc/rdma/osc_rdma_component.c

@@ -543,7 +543,8 @@ static int allocate_state_shared (ompi_osc_rdma_module_t *module, void **base, s
     local_size = ompi_comm_size (shared_comm);
 
     /* CPU atomics can be used if every process is on the same node or the NIC allows mixing CPU and NIC atomics */
-    module->use_cpu_atomics = local_size == global_size || (module->selected_btl->btl_flags & MCA_BTL_ATOMIC_SUPPORTS_GLOB);
+    module->single_node     = local_size == global_size;
+    module->use_cpu_atomics = module->single_node || (module->selected_btl->btl_flags & MCA_BTL_ATOMIC_SUPPORTS_GLOB);
 
     if (1 == local_size) {
         /* no point using a shared segment if there are no other processes on this node */