Merge pull request #3301 from rhc54/topic/faults

Correctly identify the source of the event when notifying of abnormal termination by a proces

orte/mca/state/base/state_base_fns.c

@@ -460,6 +460,7 @@ void orte_state_base_report_progress(int fd, short argc, void *cbdata)
 }
 
 static void _send_notification(int status,
+                               orte_proc_state_t state,
                                orte_process_name_t *proc,
                                orte_process_name_t *target)
 {
@@ -485,13 +486,36 @@ static void _send_notification(int status,
         return;
     }
 
-    /* the source is me */
+    /* the source is the proc */
-    if (ORTE_SUCCESS != (rc = opal_dss.pack(buf, ORTE_PROC_MY_NAME, 1, ORTE_NAME))) {
+    if (ORTE_SUCCESS != (rc = opal_dss.pack(buf, proc, 1, ORTE_NAME))) {
         ORTE_ERROR_LOG(rc);
         OBJ_RELEASE(buf);
         return;
     }
 
+    if (OPAL_ERR_PROC_ABORTED == status) {
+        /* we will pass four opal_value_t's */
+        rc = 4;
+        if (ORTE_SUCCESS != (rc = opal_dss.pack(buf, &rc, 1, OPAL_INT))) {
+            ORTE_ERROR_LOG(rc);
+            OBJ_RELEASE(buf);
+            return;
+        }
+        /* pass along the affected proc(s) */
+        OBJ_CONSTRUCT(&kv, opal_value_t);
+        kv.key = strdup(OPAL_PMIX_EVENT_AFFECTED_PROC);
+        kv.type = OPAL_NAME;
+        kv.data.name.jobid = proc->jobid;
+        kv.data.name.vpid = proc->vpid;
+        kvptr = &kv;
+        if (ORTE_SUCCESS != (rc = opal_dss.pack(buf, &kvptr, 1, OPAL_VALUE))) {
+            ORTE_ERROR_LOG(rc);
+            OBJ_DESTRUCT(&kv);
+            OBJ_RELEASE(buf);
+            return;
+        }
+        OBJ_DESTRUCT(&kv);
+    } else {
         /* we are going to pass three opal_value_t's */
         rc = 3;
         if (ORTE_SUCCESS != (rc = opal_dss.pack(buf, &rc, 1, OPAL_INT))) {
@@ -499,6 +523,7 @@ static void _send_notification(int status,
             OBJ_RELEASE(buf);
             return;
         }
+    }
 
     /* pass along the affected proc(s) */
     OBJ_CONSTRUCT(&kv, opal_value_t);
@@ -699,11 +724,11 @@ void orte_state_base_track_procs(int fd, short argc, void *cbdata)
                     /* notify everyone who asked for it */
                     target.jobid = jdata->jobid;
                     target.vpid = ORTE_VPID_WILDCARD;
-                    _send_notification(OPAL_ERR_JOB_TERMINATED, &target, ORTE_NAME_WILDCARD);
+                    _send_notification(OPAL_ERR_JOB_TERMINATED, pdata->state, &target, ORTE_NAME_WILDCARD);
                 } else {
                     target.jobid = jdata->jobid;
                     target.vpid = ORTE_VPID_WILDCARD;
-                    _send_notification(OPAL_ERR_JOB_TERMINATED, &target, &parent);
+                    _send_notification(OPAL_ERR_JOB_TERMINATED, pdata->state, &target, &parent);
                 }
             }
         } else if (ORTE_PROC_STATE_TERMINATED < pdata->state &&
@@ -711,7 +736,7 @@ void orte_state_base_track_procs(int fd, short argc, void *cbdata)
             /* if this was an abnormal term, notify the other procs of the termination */
             parent.jobid = jdata->jobid;
             parent.vpid = ORTE_VPID_WILDCARD;
-            _send_notification(OPAL_ERR_PROC_ABORTED, &pdata->name, &parent);
+            _send_notification(OPAL_ERR_PROC_ABORTED, pdata->state, &pdata->name, &parent);
         }
     }
 

orte/util/nidmap.c

@@ -641,6 +641,13 @@ int orte_util_encode_nodemap(opal_buffer_t *buffer)
         OBJ_CONSTRUCT(&bucket, opal_buffer_t);
         while (NULL != (item = opal_list_remove_first(&topos))) {
             rng = (orte_regex_range_t*)item;
+            if (NULL == rng->t) {
+                /* when we pass thru here prior to launching the daemons, we
+                 * won't have topologies for them and so this entry might
+                 * be NULL - protect ourselves */
+                OBJ_RELEASE(item);
+                continue;
+            }
             if (NULL == tmp) {
                 asprintf(&tmp, "%d", rng->cnt);
             } else {