Also avoid partition-related hangs here.bug26404

1 files changed, 8 insertions, 2 deletions

diff --git a/src/rabbit_mirror_queue_master.erl b/src/rabbit_mirror_queue_master.erl
index f1b1a8a5..aa1e1ab9 100644
--- a/src/rabbit_mirror_queue_master.erl
+++ b/src/rabbit_mirror_queue_master.erl
@@ -201,9 +201,15 @@ delete_and_terminate(Reason, State = #state { backing_queue       = BQ,
 
 stop_all_slaves(Reason, #state{name = QName, gm = GM}) ->
     {ok, #amqqueue{slave_pids = SPids}} = rabbit_amqqueue:lookup(QName),
-    MRefs = [erlang:monitor(process, Pid) || Pid <- [GM | SPids]],
+    PidsMRefs = [{Pid, erlang:monitor(process, Pid)} || Pid <- [GM | SPids]],
     ok = gm:broadcast(GM, {delete_and_terminate, Reason}),
-    [receive {'DOWN', MRef, process, _Pid, _Info} -> ok end || MRef <- MRefs],
+    %% It's possible that we could be partitioned from some slaves
+    %% between the lookup and the broadcast, in which case we could
+    %% monitor them but they would not have received the GM
+    %% message. So only wait for slaves which are still
+    %% not-partitioned.
+    [receive {'DOWN', MRef, process, _Pid, _Info} -> ok end
+     || {Pid, MRef} <- PidsMRefs, rabbit_mnesia:on_running_node(Pid)],
     %% Normally when we remove a slave another slave or master will
     %% notice and update Mnesia. But we just removed them all, and
     %% have stopped listening ourselves. So manually clean up.