From e7158e8e41e89eb845ecd7a888a5bfc6a008cc75 Mon Sep 17 00:00:00 2001
From: John Spray <john.spray@redhat.com>
Date: Thu, 7 Sep 2017 09:44:36 -0400
Subject: [PATCH] mon: fix dropping mgr metadata for active mgr

drop_standby() was killing it and it was only getting added
back in certain locations.  Instead, make the metadata
drop conditional and only do it in the places we're
really dropping the daemon, not when we're promoting
it to active.

Fixes: http://tracker.ceph.com/issues/21260
Signed-off-by: John Spray <john.spray@redhat.com>
(cherry picked from commit 29c6f9adf178f6611a625740f395e397cad9147b)
---
 src/mon/MgrMonitor.cc | 13 ++++++++-----
 src/mon/MgrMonitor.h  | 11 ++++++++++-
 2 files changed, 18 insertions(+), 6 deletions(-)

diff --git a/src/mon/MgrMonitor.cc b/src/mon/MgrMonitor.cc
index fc4f08d8d70c5..3840b642b956c 100644
--- a/src/mon/MgrMonitor.cc
+++ b/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);
diff --git a/src/mon/MgrMonitor.h b/src/mon/MgrMonitor.h
index 65451633dbefc..563ae7c5de8de 100644
--- a/src/mon/MgrMonitor.h
+++ b/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();
-- 
2.39.5