Merge pull request #1459 from alinask/topic/openib_diff_subnets

btl/openib: enable connecting processes from different subnets.

Author: mike-dubman

opal/mca/btl/openib/btl_openib.c

@@ -849,6 +849,14 @@ static int init_ib_proc_nolock(mca_btl_openib_module_t* openib_btl, mca_btl_open
                 matching_port = j;
             }
             rem_port_cnt++;
+        } else {
+            if (mca_btl_openib_component.allow_different_subnets) {
+                BTL_VERBOSE(("Using different subnets!"));
+                if (rem_port_cnt == btl_rank) {
+                    matching_port = j;
+                }
+                rem_port_cnt++;
+            }
         }
     }
 
@@ -915,6 +923,13 @@ static int init_ib_proc_nolock(mca_btl_openib_module_t* openib_btl, mca_btl_open
                     break;
                 else
                     rem_port_cnt ++;
+            } else {
+                if (mca_btl_openib_component.allow_different_subnets) {
+                    if (rem_port_cnt == btl_rank)
+                        break;
+                    else
+                        rem_port_cnt ++;
+                }
             }
         }
 
@@ -981,6 +996,13 @@ static int get_openib_btl_params(mca_btl_openib_module_t* openib_btl, int *port_
                 rank = port_cnt;
             }
             port_cnt++;
+        } else {
+            if (mca_btl_openib_component.allow_different_subnets) {
+                if (openib_btl == mca_btl_openib_component.openib_btls[j]) {
+                    rank = port_cnt;
+                }
+                port_cnt++;
+            }
         }
     }
     *port_cnt_ptr = port_cnt;

opal/mca/btl/openib/btl_openib.h

@@ -295,6 +295,9 @@ struct mca_btl_openib_component_t {
     char* default_recv_qps;
     /** GID index to use */
     int gid_index;
+    /*  Whether we want to allow connecting processes from different subnets.
+     *  set to 'no' by default */
+    bool allow_different_subnets;
     /** Whether we want a dynamically resizing srq, enabled by default */
     bool enable_srq_resize;
     bool allow_max_memory_registration;

opal/mca/btl/openib/btl_openib_mca.c

@@ -706,6 +706,11 @@ int btl_openib_register_mca_params(void)
                   0, &mca_btl_openib_component.gid_index,
                   REGINT_GE_ZERO));
 
+    CHECK(reg_bool("allow_different_subnets", NULL,
+                   "Allow connecting processes from different IB subnets."
+                   "(0 = do not allow; 1 = allow)",
+                   false, &mca_btl_openib_component.allow_different_subnets));
+
 #if MEMORY_LINUX_MALLOC_ALIGN_ENABLED
     tmp = mca_base_var_find ("opal", "memory", "linux", "memalign");
     if (0 <= tmp) {

opal/mca/btl/openib/connect/btl_openib_connect_rdmacm.c

@@ -2109,7 +2109,8 @@ static int rdmacm_component_query(mca_btl_openib_module_t *openib_btl, opal_btl_
     sin.sin_addr.s_addr = rdmacm_addr;
     sin.sin_port = (uint16_t) rdmacm_port;
 #else
-    rc = ibv_query_gid(openib_btl->device->ib_pd->context, openib_btl->port_num, 0, &server->gid);
+    rc = ibv_query_gid(openib_btl->device->ib_pd->context, openib_btl->port_num,
+                       mca_btl_openib_component.gid_index, &server->gid);
     if (0 != rc) {
         BTL_ERROR(("local gid query failed"));
         goto out4;