btl/openib: immediately release the device when no port is allowed

Many thanks to Sergey Oblomov for reporting this issue
and the countless traces provided when troubleshooting it.

This is a one-off commit for the v4.0.x branch since btl/openib has been removed
 from master.

Refs. #6137

Signed-off-by: Gilles Gouaillardet <[email protected]>

Author: ggouaillardet

opal/mca/btl/openib/btl_openib.c

@@ -1045,7 +1045,7 @@ int mca_btl_openib_add_procs(
         opal_bitmap_clear_all_bits(reachable);
         opal_show_help("help-mpi-btl-openib.txt", "ib port not selected",
                        true, opal_process_info.nodename,
-                       ibv_get_device_name(openib_btl->device->ib_dev), openib_btl->port_num);
+                       openib_btl->device_name, openib_btl->port_num);
         return OPAL_SUCCESS;
     }
 
@@ -1718,11 +1718,11 @@ static int mca_btl_openib_finalize_resources(struct mca_btl_base_module_t* btl)
             free(openib_btl->cpcs[i]);
         }
         free(openib_btl->cpcs);
-    }
 
-    /* Release device if there are no more users */
-    if(!(--openib_btl->device->btls)) {
-        OBJ_RELEASE(openib_btl->device);
+        /* Release device if there are no more users */
+        if(!(--openib_btl->device->allowed_btls)) {
+            OBJ_RELEASE(openib_btl->device);
+        }
     }
 
     if (NULL != openib_btl->qps) {

opal/mca/btl/openib/btl_openib.h

@@ -392,6 +392,7 @@ typedef struct mca_btl_openib_device_t {
     /* Whether this device supports eager RDMA */
     uint8_t use_eager_rdma;
     uint8_t btls;              /** < number of btls using this device */
+    uint8_t allowed_btls;      /** < number of allowed btls using this device */
     opal_pointer_array_t *endpoints;
     opal_pointer_array_t *device_btls;
     uint16_t hp_cq_polls;
@@ -483,6 +484,7 @@ struct mca_btl_openib_module_t {
     uint8_t num_cpcs;
 
     mca_btl_openib_device_t *device;
+    char * device_name;
     uint8_t port_num;                  /**< ID of the PORT */
     uint16_t pkey_index;
     struct ibv_port_attr ib_port_attr;

opal/mca/btl/openib/btl_openib_component.c

@@ -648,9 +648,10 @@ static int init_one_port(opal_list_t *btl_list, mca_btl_openib_device_t *device,
                 sizeof(mca_btl_openib_module));
         ib_selected = OBJ_NEW(mca_btl_base_selected_module_t);
         ib_selected->btl_module = (mca_btl_base_module_t*) openib_btl;
-        openib_btl->device = device;
         openib_btl->port_num = (uint8_t) port_num;
         openib_btl->allowed = false;
+        openib_btl->device = NULL;
+        openib_btl->device_name = strdup(ibv_get_device_name(device->ib_dev));
         OBJ_CONSTRUCT(&openib_btl->ib_lock, opal_mutex_t);
         opal_list_append(btl_list, (opal_list_item_t*) ib_selected);
         opal_pointer_array_add(device->device_btls, (void*) openib_btl);
@@ -784,6 +785,7 @@ static int init_one_port(opal_list_t *btl_list, mca_btl_openib_device_t *device,
             ib_selected = OBJ_NEW(mca_btl_base_selected_module_t);
             ib_selected->btl_module = (mca_btl_base_module_t*) openib_btl;
             openib_btl->device = device;
+            openib_btl->device_name = NULL;
             openib_btl->port_num = (uint8_t) port_num;
             openib_btl->pkey_index = pkey_index;
             openib_btl->lid = lid;
@@ -904,6 +906,7 @@ static int init_one_port(opal_list_t *btl_list, mca_btl_openib_device_t *device,
             opal_list_append(btl_list, (opal_list_item_t*) ib_selected);
             opal_pointer_array_add(device->device_btls, (void*) openib_btl);
             ++device->btls;
+            ++device->allowed_btls;
             ++mca_btl_openib_component.ib_num_btls;
             ++mca_btl_openib_component.ib_allowed_btls;
             if (-1 != mca_btl_openib_component.ib_max_btls &&
@@ -1933,7 +1936,7 @@ static int init_one_device(opal_list_t *btl_list, struct ibv_device* ib_dev)
             if (ib_port_attr.active_mtu < device->mtu){
                 device->mtu = ib_port_attr.active_mtu;
             }
-            if (mca_btl_openib_component.apm_ports && device->btls > 0) {
+            if (mca_btl_openib_component.apm_ports && device->allowed_btls > 0) {
                 init_apm_port(device, i, ib_port_attr.lid);
                 break;
             }
@@ -1969,7 +1972,7 @@ static int init_one_device(opal_list_t *btl_list, struct ibv_device* ib_dev)
 
     /* If we made a BTL, check APM status and return.  Otherwise, fall
        through and destroy everything */
-    if (device->btls > 0) {
+    if (device->allowed_btls > 0) {
         /* if apm was enabled it should be > 1 */
         if (1 == mca_btl_openib_component.apm_ports) {
             opal_show_help("help-mpi-btl-openib.txt",
@@ -2290,6 +2293,11 @@ static int init_one_device(opal_list_t *btl_list, struct ibv_device* ib_dev)
     good:
         mca_btl_openib_component.devices_count++;
         return OPAL_SUCCESS;
+    } else if (device->btls > 0) {
+        /* no port is allowed to be used by btl/openib,
+         * so release the device right away */
+        OBJ_RELEASE(device);
+        return OPAL_SUCCESS;
     }
 
 error: