MDEV-7123 : MariaDB 10.0.14 Galera node shutdown with signal 11

wsrep-patch uses same connection name for constructing Master_info
objects. As a result all existing wsrep Master_info objects refer
to same rpl_filter object. This could lead to race when multiple
threads try to delete/destruct Master_info object, as they would
all try to delete the same relay_log object.

Fixed by adding a check in Master_info's destructor to not free
the "wsrep" rpl_filter, so that its reused by current & subsequent
wsrep threads and later reclaimed by free_all_rpl_filters() during
server shutdown.

1 files changed, 7 insertions, 0 deletions

diff --git a/sql/rpl_mi.cc b/sql/rpl_mi.cc
index 78218e0418e..e513297e99f 100644
--- a/sql/rpl_mi.cc
+++ b/sql/rpl_mi.cc
@@ -89,6 +89,13 @@ Master_info::Master_info(LEX_STRING *connection_name_arg,
 
 Master_info::~Master_info()
 {
+#ifdef WITH_WSREP
+  /*
+    Do not free "wsrep" rpl_filter. It will eventually be freed by
+    free_all_rpl_filters() when server terminates.
+  */
+  if (strncmp(connection_name.str, STRING_WITH_LEN("wsrep")))
+#endif
   rpl_filters.delete_element(connection_name.str, connection_name.length,
                              (void (*)(const char*, uchar*)) free_rpl_filter);
   my_free(connection_name.str);