Fix a rare deadlock situation in Replication sync phase (arangodb#7759)

mchacki · web-flow · commit 647a29d1aeab · 2018-12-12T22:05:01.000+01:00
diff --git a/CHANGELOG b/CHANGELOG
@@ -1,6 +1,11 @@
 devel
 -----
 
+* fixed a rare thread local dead lock situation in replication:
+  If a follower tries to get in sync in the last steps it requires
+  a lock on master, if he cancels the lock before the master has
+  succeeded with locking we can end up with one thread beeing deadlocked.
+
 * allow usage of floating point values in AQL without leading zeros, e.g.
   `.1234`. Previous versions of ArangoDB required a leading zero in front of
   the decimal separator, i.e `0.1234`.
diff --git a/arangod/RestHandler/RestReplicationHandler.cpp b/arangod/RestHandler/RestReplicationHandler.cpp
@@ -2872,7 +2872,9 @@ bool RestReplicationHandler::isTombstoned(aql::QueryId id) const {
 
 void RestReplicationHandler::registerTombstone(aql::QueryId id) const {
   std::string key = IdToTombstoneKey(_vocbase, id);
-  WRITE_LOCKER(writeLocker, RestReplicationHandler::_tombLock);
-  RestReplicationHandler::_tombstones.emplace(key, std::chrono::steady_clock::now() + RestReplicationHandler::_tombstoneTimeout); 
+  {
+    WRITE_LOCKER(writeLocker, RestReplicationHandler::_tombLock);
+    RestReplicationHandler::_tombstones.emplace(key, std::chrono::steady_clock::now() + RestReplicationHandler::_tombstoneTimeout); 
+  }
   timeoutTombstones();
 }
