Merge pull request #7004 from mwheinz/REFS6976-master

REF6976 Silent failure of OMPI over OFI with large messages sizes

ompi/mca/mtl/ofi/help-mtl-ofi.txt

@@ -75,3 +75,5 @@ recoverable and your application is likely to abort.
   Local host: %s
   Remote host: %s
   Error: %s (%d)
+[message too big]
+Message size %llu bigger than supported by selected transport. Max = %llu

ompi/mca/mtl/ofi/mtl_ofi.h

@@ -421,13 +421,22 @@ ompi_mtl_ofi_send_generic(struct mca_mtl_base_module_t *mtl,
     sep_peer_fiaddr = fi_rx_addr(endpoint->peer_fiaddr, ctxt_id, ompi_mtl_ofi.rx_ctx_bits);
 
     ompi_ret = ompi_mtl_datatype_pack(convertor, &start, &length, &free_after);
-    if (OMPI_SUCCESS != ompi_ret) return ompi_ret;
+    if (OPAL_UNLIKELY(OMPI_SUCCESS != ompi_ret)) {
+        return ompi_ret;
+    }
 
     ofi_req.buffer = (free_after) ? start : NULL;
     ofi_req.length = length;
     ofi_req.status.MPI_ERROR = OMPI_SUCCESS;
     ofi_req.completion_count = 0;
 
+    if (OPAL_UNLIKELY(length > endpoint->mtl_ofi_module->max_msg_size)) {
+        opal_show_help("help-mtl-ofi.txt",
+            "message too big", false,
+            length, endpoint->mtl_ofi_module->max_msg_size);
+        return OMPI_ERROR;
+    }
+
     if (ofi_cq_data) {
         match_bits = mtl_ofi_create_send_tag_CQD(comm->c_contextid, tag);
         src_addr = sep_peer_fiaddr;
@@ -553,13 +562,20 @@ ompi_mtl_ofi_isend_generic(struct mca_mtl_base_module_t *mtl,
     sep_peer_fiaddr = fi_rx_addr(endpoint->peer_fiaddr, ctxt_id, ompi_mtl_ofi.rx_ctx_bits);
 
     ompi_ret = ompi_mtl_datatype_pack(convertor, &start, &length, &free_after);
-    if (OMPI_SUCCESS != ompi_ret) return ompi_ret;
+    if (OMPI_UNLIKELY(OMPI_SUCCESS != ompi_ret)) return ompi_ret;
 
     ofi_req->buffer = (free_after) ? start : NULL;
     ofi_req->length = length;
     ofi_req->status.MPI_ERROR = OMPI_SUCCESS;
     ofi_req->completion_count = 1;
 
+    if (OPAL_UNLIKELY(length > endpoint->mtl_ofi_module->max_msg_size)) {
+        opal_show_help("help-mtl-ofi.txt",
+            "message too big", false,
+            length, endpoint->mtl_ofi_module->max_msg_size);
+        return OMPI_ERROR;
+    }
+
     if (ofi_cq_data) {
         match_bits = mtl_ofi_create_send_tag_CQD(comm->c_contextid, tag);
     } else {

ompi/mca/mtl/ofi/mtl_ofi_component.c

@@ -881,9 +881,10 @@ ompi_mtl_ofi_component_init(bool enable_progress_threads,
     }
 
     /**
-     * Save the maximum inject size.
+     * Save the maximum sizes.
      */
     ompi_mtl_ofi.max_inject_size = prov->tx_attr->inject_size;
+    ompi_mtl_ofi.max_msg_size = prov->ep_attr->max_msg_size;
 
     /**
      * The user is not allowed to exceed MTL_OFI_MAX_PROG_EVENT_COUNT.

ompi/mca/mtl/ofi/mtl_ofi_types.h

@@ -70,6 +70,9 @@ typedef struct mca_mtl_ofi_module_t {
     /** Maximum inject size */
     size_t max_inject_size;
 
+    /** Largest message that can be sent in a single send. */
+    size_t max_msg_size;
+
     /** Maximum number of CQ events to read in OFI Progress */
     int ofi_progress_event_count;