Better clean-up of pvmove failures in a cluster

Hi,

I don't see any reason why the following patch would be bad, but I'd like to hear from others first. In essence, I'm just trying to improve the error code path. This seems to improve things for pvmove when a node in the cluster is not running 'cmirrord' as it should. However, it doesn't clean everything up when attempting to create a full cluster mirror LV. So there might still be some work left after this.

brassow

activation: Remove empty DM devices when table fails to load.

As part of better error handling, remove DM devices that have been successfully created but failed to load a table. This can happen when pvmove'ing in a cluster and the cluster mirror daemon is not running on a remote node - the mapping table failing to load as a result. In this case, any revert would work on other nodes running cmirrord because the DM devices on those nodes did succeed in loading. However, because no table was able to load on the non-cmirrord nodes, there is no table present that points to what needs to be reverted. This causes the empty DM device to remain on the system without being present in any LVM representation.

Index: lvm2/libdm/libdm-deptree.c
===================================================================
@@ -1929,6 +1929,26 @@ out:
 	return r;
 }
 
+/*
+ * _remove_node
+ *
+ * This function is only used to remove a DM device that has failed
+ * to load any table.
+ */
+static int _remove_node(struct dm_tree_node *dnode)
+{
+	if (!dnode->info.exists)
+		return 1;
+
+	if (dnode->info.live_table || dnode->info.inactive_table) {
+		log_error(INTERNAL_ERROR
+			  "_remove_node called on device with loaded table(s).");
+		return 0;
+	}
+
+	return _deactivate_node(dnode->name, dnode->info.major,
+				dnode->info.minor, NULL, 0, 0);
+}
 
 static int _build_dev_string(char *devbuf, size_t bufsize, struct dm_tree_node *node)
 {

@@ -2662,7 +2682,7 @@ int dm_tree_preload_children(struct dm_t
 			     const char *uuid_prefix,
 			     size_t uuid_prefix_len)
 {
-	int r = 1;
+	int r = 1, node_created = 0;
 	void *handle = NULL;
 	struct dm_tree_node *child;
 	struct dm_info newinfo;

@@ -2684,13 +2704,16 @@ int dm_tree_preload_children(struct dm_t
 				return_0;
 
 		/* FIXME Cope if name exists with no uuid? */
-		if (!child->info.exists && !_create_node(child))
+		if (!child->info.exists && !(node_created = _create_node(child)))
 			return_0;
 
 		if (!child->info.inactive_table &&
 		    child->props.segment_count &&
-		    !_load_node(child))
+		    !_load_node(child)) {
+			if (node_created && !_remove_node(child))
+				return_0;
 			return_0;
+		}
 
 		/* Propagate device size change change */
 		if (child->props.size_changed)

Dne 5.5.2014 23:38, Jonathan Brassow napsal(a):

I've already tried to add some initial recovery support for error path.

If you take a look into the upstream libdm-deptee code - there is already dnode->activated list for recovery path - which track just newly introduced devices.

The problem is however much more complicated - it should work as transaction, since multiple device may get changed - and all changes should be reverted properly.

There is postprocessing routine which should cleanup any left empty entries in the table.

I'm currently working on more generic solution.

Zdenek