luminous: ceph mgr versions shows active mgr as Unknown

qa/tasks/mgr/test_failover.py

@@ -1,5 +1,6 @@
 
 import logging
+import json
 
 from tasks.mgr.mgr_test_case import MgrTestCase
 
@@ -92,6 +93,15 @@ def test_explicit_fail(self):
             timeout=10
         )
 
+        # Both daemons should have fully populated metadata
+        # (regression test for http://tracker.ceph.com/issues/21260)
+        meta = json.loads(self.mgr_cluster.mon_manager.raw_cluster_cmd(
+            "mgr", "metadata"))
+        id_to_meta = dict([(i['id'], i) for i in meta])
+        for i in [original_active] + original_standbys:
+            self.assertIn(i, id_to_meta)
+            self.assertIn('ceph_version', id_to_meta[i])
+
         # We should be able to fail back over again: the exercises
         # our re-initialization of the python runtime within
         # a single process lifetime.

src/mon/MgrMonitor.cc

@@ -332,7 +332,7 @@ bool MgrMonitor::prepare_beacon(MonOpRequestRef op)
   } else if (pending_map.active_gid == 0) {
     // There is no currently active daemon, select this one.
     if (pending_map.standbys.count(m->get_gid())) {
-      drop_standby(m->get_gid());
+      drop_standby(m->get_gid(), false);
     }
     dout(4) << "selecting new active " << m->get_gid()
 	    << " " << m->get_name()
@@ -599,7 +599,8 @@ bool MgrMonitor::promote_standby()
     pending_map.available = false;
     pending_map.active_addr = entity_addr_t();
 
-    drop_standby(replacement_gid);
+    drop_standby(replacement_gid, false);
+
     return true;
   } else {
     return false;
@@ -624,10 +625,12 @@ void MgrMonitor::drop_active()
   cancel_timer();
 }
 
-void MgrMonitor::drop_standby(uint64_t gid)
+void MgrMonitor::drop_standby(uint64_t gid, bool drop_meta)
 {
-  pending_metadata_rm.insert(pending_map.standbys[gid].name);
-  pending_metadata.erase(pending_map.standbys[gid].name);
+  if (drop_meta) {
+    pending_metadata_rm.insert(pending_map.standbys[gid].name);
+    pending_metadata.erase(pending_map.standbys[gid].name);
+  }
   pending_map.standbys.erase(gid);
   if (last_beacon.count(gid) > 0) {
     last_beacon.erase(gid);

src/mon/MgrMonitor.h

@@ -43,7 +43,16 @@ class MgrMonitor: public PaxosService
    */
   bool promote_standby();
   void drop_active();
-  void drop_standby(uint64_t gid);
+
+  /**
+   * Remove this gid from the list of standbys.  By default,
+   * also remove metadata (i.e. forget the daemon entirely).
+   *
+   * Set `drop_meta` to false if you would like to keep
+   * the daemon's metadata, for example if you're dropping
+   * it as a standby before reinstating it as the active daemon.
+   */
+  void drop_standby(uint64_t gid, bool drop_meta=true);
 
   Context *digest_event = nullptr;
   void cancel_timer();