btl/uct: keep UCT component list around after initial scan

From what I can tell it is not ok to be releasing this list while the components
within the list are still in use. To be safe keep the list around until the uct
component is closed.

Signed-off-by: Nathan Hjelm <[email protected]>

Author: hjelmn

opal/mca/btl/uct/btl_uct.h

@@ -162,6 +162,11 @@ struct mca_btl_uct_component_t {
     /** alternate connection-only module that can be used if no suitable
      * connection tl is found. this is usually a tcp tl. */
     mca_btl_uct_module_t *conn_module;
+
+#if UCT_API >= UCT_VERSION(1, 7)
+    uct_component_h *uct_components;
+    unsigned num_uct_components;
+#endif
 };
 typedef struct mca_btl_uct_component_t mca_btl_uct_component_t;
 

opal/mca/btl/uct/btl_uct_component.c

@@ -239,6 +239,14 @@ static int mca_btl_uct_component_close(void)
     mca_btl_uct_include_list_free (&mca_btl_uct_component.allowed_transport_list);
     mca_btl_uct_include_list_free (&mca_btl_uct_component.connection_domain_list);
 
+#if UCT_API >= UCT_VERSION(1, 7)
+    if (NULL != mca_btl_uct_component.uct_components) {
+        uct_release_component_list(mca_btl_uct_component.uct_components);
+        mca_btl_uct_component.uct_components = NULL;
+        mca_btl_uct_component.num_uct_components = 0;
+    }
+#endif
+
     return OPAL_SUCCESS;
 }
 
@@ -690,25 +698,20 @@ static mca_btl_base_module_t **mca_btl_uct_component_init(int *num_btl_modules,
     mca_btl_uct_component.module_count = 0;
 
 #if UCT_API >= UCT_VERSION(1, 7)
-    uct_component_h *components;
-    unsigned num_components;
-
-    ucs_status = uct_query_components(&components, &num_components);
+    ucs_status = uct_query_components(&mca_btl_uct_component.uct_components,
+                                      &mca_btl_uct_component.num_uct_components);
     if (UCS_OK != ucs_status) {
         BTL_ERROR(("could not query UCT components"));
         return NULL;
     }
 
     /* generate all suitable btl modules */
-    for (unsigned i = 0; i < num_components; ++i) {
-        rc = mca_btl_uct_component_process_uct_component(components[i]);
+    for (unsigned i = 0; i < mca_btl_uct_component.num_uct_components; ++i) {
+        rc = mca_btl_uct_component_process_uct_component(mca_btl_uct_component.uct_components[i]);
         if (OPAL_SUCCESS != rc) {
             break;
         }
     }
-
-    uct_release_component_list(components);
-
 #else /* UCT 1.6 and older */
     uct_md_resource_desc_t *resources;
     unsigned resource_count;