From Wolfgang Mueller

Backport a patch from head for updating nodes flags

also make sure that when a node repotrs incinsistent node flags, we mark it as a culprit so that if it insists in retaining an opinion not accepted in the cluster groupthink we ban it quickly.

diff --git a/server/ctdb_monitor.c b/server/ctdb_monitor.c
index 14da3b77e4bc2ecce4eedd3956076eb332721f32..25d2226f9a23bf41a96efcc79d5fb54b716fa719 100644 (file)
--- a/server/ctdb_monitor.c
+++ b/server/ctdb_monitor.c
@@ -293,6 +293,9 @@ int32_t ctdb_control_modflags(struct ctdb_context *ctdb, TDB_DATA indata)
 
        node         = ctdb->nodes[c->pnn];
        c->old_flags  = node->flags;
+       if (c->pnn != ctdb->pnn) {
+               c->old_flags = node->flags;
+       }
        node->flags   = c->new_flags & ~NODE_FLAGS_DISCONNECTED;
        node->flags  |= (c->old_flags & NODE_FLAGS_DISCONNECTED);
 

diff --git a/server/ctdb_recoverd.c b/server/ctdb_recoverd.c
index e065ed519ff94bc6f8b8171e58a74cc1d5adef22..31335d9065329b7e48747236ac4216c7c389326d 100644 (file)
--- a/server/ctdb_recoverd.c
+++ b/server/ctdb_recoverd.c
@@ -1992,11 +1992,46 @@ static void push_flags_handler(struct ctdb_context *ctdb, uint64_t srvid,
 {
        int ret;
        struct ctdb_node_flag_change *c = (struct ctdb_node_flag_change *)data.dptr;
+       struct ctdb_node_map *nodemap=NULL;
+       TALLOC_CTX *tmp_ctx = talloc_new(ctdb);
+       uint32_t recmaster;
+       uint32_t *nodes;
 
-       ret = ctdb_ctrl_modflags(ctdb, CONTROL_TIMEOUT(), c->pnn, c->new_flags, ~c->new_flags);
+       /* find the recovery master */
+       ret = ctdb_ctrl_getrecmaster(ctdb, tmp_ctx, CONTROL_TIMEOUT(), CTDB_CURRENT_NODE, &recmaster);
        if (ret != 0) {
-               DEBUG(DEBUG_ERR, (__location__ " Unable to update nodeflags on remote nodes\n"));
+               DEBUG(DEBUG_ERR, (__location__ " Unable to get recmaster from local node\n"));
+               talloc_free(tmp_ctx);
+               return;
        }
+       /* read the node flags from the recmaster */
+ 
+       ret = ctdb_ctrl_getnodemap(ctdb, CONTROL_TIMEOUT(), recmaster, tmp_ctx, &nodemap);
+       if (ret != 0) {
+               DEBUG(DEBUG_ERR, (__location__ " Unable to get nodemap from node %u\n", c->pnn));
+               talloc_free(tmp_ctx);
+               return;
+       }
+
+       if (c->pnn >= nodemap->num) {
+               DEBUG(DEBUG_ERR,(__location__ " Nodemap from recmaster does not contain node %d\n", c->pnn));
+               talloc_free(tmp_ctx);
+               return;
+       }
+
+       /* send the flags update to all connected nodes */
+       nodes = list_of_connected_nodes(ctdb, nodemap, tmp_ctx, true);
+
+       if (ctdb_client_async_control(ctdb, CTDB_CONTROL_MODIFY_FLAGS,
+                                                                 nodes, CONTROL_TIMEOUT(),
+                                                                 false, data,
+                                                                 NULL, NULL,
+                                                                 NULL) != 0) {
+               DEBUG(DEBUG_ERR, (__location__ " ctdb_control to modify node flags failed\n"));
+               talloc_free(tmp_ctx);
+               return;
+       }
+       talloc_free(tmp_ctx);
 }
 
 
@@ -2739,6 +2774,9 @@ again:
                                  nodemap->nodes[i].pnn, 
                                  remote_nodemaps[j]->nodes[i].flags,
                                  nodemap->nodes[j].flags));
+
+                               ctdb_set_culprit_count(rec, nodemap->nodes[j].pnn, 4);
+
                                if (i == j) {
                                        DEBUG(DEBUG_ERR,("Use flags 0x%02x from remote node %d for cluster update of its own flags\n", remote_nodemaps[j]->nodes[i].flags, j));
                                        update_flags_on_all_nodes(ctdb, nodemap, nodemap->nodes[i].pnn, remote_nodemaps[j]->nodes[i].flags);