crimson/monc: discard active/pending connections when reopening

Otherwise pending_conns vector just keeps growing with redundant
connections all trying to reach the same set of monitors.  When one
of the attempts finally succeeds, _finish_auth() will pick the first
connection with a matching entity_addr_t, designate it as active and
close all others.  The match is very likely to be wrong and hence the
actual authenticated connection gets closed, leaving the OSD with
a bogus active_con and no monitor session.

Signed-off-by: Ilya Dryomov <idryomov@gmail.com>

diff --git a/src/crimson/mon/MonClient.cc b/src/crimson/mon/MonClient.cc
index 2d82aadc4fc053f0badc8e6ca9a5662316bcb8b7..2628998eb4a18d57856984f54ced306e00ee4e53 100644 (file)
--- a/src/crimson/mon/MonClient.cc
+++ b/src/crimson/mon/MonClient.cc
@@ -515,8 +515,6 @@ void Client::ms_handle_reset(crimson::net::ConnectionRef conn, bool /* is_replac
       return seastar::now();
     } else if (active_con && active_con->is_my_peer(conn->get_peer_addr())) {
       logger().warn("active conn reset {}", conn->get_peer_addr());
-      active_con->close();
-      active_con.reset();
       return reopen_session(-1).then([this](bool opened) {
         if (opened) {
           return on_session_opened();
@@ -929,6 +927,16 @@ static entity_addr_t choose_client_addr(
 seastar::future<bool> Client::reopen_session(int rank)
 {
   logger().info("{} to mon.{}", __func__, rank);
+  if (active_con) {
+    active_con->close();
+    active_con.reset();
+    ceph_assert(pending_conns.empty());
+  } else {
+    for (auto& pending_con : pending_conns) {
+      pending_con->close();
+    }
+    pending_conns.clear();
+  }
   vector<unsigned> mons;
   if (rank >= 0) {
     mons.push_back(rank);