Merge pull request #892 from rhc54/topic/pmix

Fix the no-disconnect test by resolving a segfault on free - opal_dss…

opal/dss/dss_load_unload.c

@@ -49,19 +49,30 @@ int opal_dss_unload(opal_buffer_t *buffer, void **payload,
         return OPAL_SUCCESS;
     }
 
+    /* if nothing has been unpacked, we can pass the entire
+     * region back and protect it - no need to copy. This is
+     * an optimization */
+    if (buffer->unpack_ptr == buffer->base_ptr) {
+        *payload = buffer->base_ptr;
+        *bytes_used = buffer->bytes_used;
+        buffer->base_ptr = NULL;
+        buffer->unpack_ptr = NULL;
+        buffer->pack_ptr = NULL;
+        buffer->bytes_used = 0;
+        return OPAL_SUCCESS;
+    }
+
     /* okay, we have something to provide - pass it back */
     *bytes_used = buffer->bytes_used - (buffer->unpack_ptr - buffer->base_ptr);
     if (0 == (*bytes_used)) {
         *payload = NULL;
     } else {
-        *payload = buffer->unpack_ptr;
+        /* we cannot just set the pointer as it might be
+         * partway in a malloc'd region */
+        *payload = (void*)malloc(*bytes_used);
+        memcpy(*payload, buffer->unpack_ptr, *bytes_used);
     }
 
-    /* dereference everything in buffer */
-    buffer->base_ptr = NULL;
-    buffer->pack_ptr = buffer->unpack_ptr = NULL;
-    buffer->bytes_allocated = buffer->bytes_used = 0;
-
     /* All done */
 
     return OPAL_SUCCESS;

opal/mca/pmix/pmix1xx/pmix/src/server/pmix_server_ops.c

@@ -857,7 +857,10 @@ static void _process_dmdx_reply(int fd, short args, void *cbdata)
         kp->key = strdup("modex");
         PMIX_VALUE_CREATE(kp->value, 1);
         kp->value->type = PMIX_BYTE_OBJECT;
-        kp->value->data.bo.bytes = (char*)caddy->data;
+        /* we don't know if the host is going to save this data
+         * or not, so we have to copy it */
+        kp->value->data.bo.bytes = (char*)malloc(caddy->ndata);
+        memcpy(kp->value->data.bo.bytes, caddy->data, caddy->ndata);
         kp->value->data.bo.size = caddy->ndata;
         /* store it in the appropriate hash */
         if (PMIX_SUCCESS != (rc = pmix_hash_store(&nptr->server->remote, caddy->lcd->proc.rank, kp))) {

orte/orted/pmix/pmix_server.c

@@ -537,7 +537,10 @@ static void pmix_server_dmdx_resp(int status, orte_process_name_t* sender,
     }
 
     /* unload the remainder of the buffer */
-    opal_dss.unload(buffer, (void**)&data, &ndata);
+    if (OPAL_SUCCESS != (rc = opal_dss.unload(buffer, (void**)&data, &ndata))) {
+        ORTE_ERROR_LOG(rc);
+        return;
+    }
 
     /* if we got something, store the blobs locally so we can
      * meet any further requests without doing a remote fetch.
@@ -565,9 +568,6 @@ static void pmix_server_dmdx_resp(int status, orte_process_name_t* sender,
         }
         OBJ_RELEASE(req);
     }
-    if (NULL != data) {
-        free(data);
-    }
 }
 
 static void opcon(orte_pmix_server_op_caddy_t *p)