osd: take heartbeat_lock before checking for session

When we open a connection, there is a short window before we attach
the session.  If a fault happens quickly, we won't get the reset, and
will persistently fail to send osd pings.

Move the lock up to avoid this.  Note that we should rarely really see
connections without sessions here anyway (except when this specific
race happens), so this should have no negative impact (by taking the lock
when we weren't before).

Fixes: http://tracker.ceph.com/issues/36602
Signed-off-by: Sage Weil <sage@redhat.com>

diff --git a/src/osd/OSD.cc b/src/osd/OSD.cc
index 0c07b1b2945b1fa9426c50c30f0e856894142873..b7168be921310641a192e96c41c66bb2d9d61e5a 100644 (file)
--- a/src/osd/OSD.cc
+++ b/src/osd/OSD.cc
@@ -4957,9 +4957,9 @@ void OSD::heartbeat()
 
 bool OSD::heartbeat_reset(Connection *con)
 {
+  std::lock_guard l(heartbeat_lock);
   auto s = con->get_priv();
   if (s) {
-    heartbeat_lock.Lock();
     if (is_stopping()) {
       heartbeat_lock.Unlock();
       return true;
@@ -4996,7 +4996,6 @@ bool OSD::heartbeat_reset(Connection *con)
     } else {
       dout(10) << "heartbeat_reset closing (old) failed hb con " << con << dendl;
     }
-    heartbeat_lock.Unlock();
   }
   return true;
 }