The fix for BUG 21136 (ChangeSet@1.2611.1.1) introduced a regression that

caused a few tests to fail because the thd->extra_lock wasn't being set to
NULL after the table was unlocked. This poses a serious problem because later
attempts to access thd->extra_lock (now a dangling pointer) will probably
result in a crash (undefined behavior) -- and that's what actually happens
in some test cases.

The solution is to set the select_create::m_plock pointee to NULL, which
means that thd->extra_lock is set to NULL when the lock data is not for a
temporary table.

sql/sql_insert.cc

@@ -3653,7 +3653,8 @@ bool select_create::send_eof()
     if (m_plock)
     {
       mysql_unlock_tables(thd, *m_plock);
-      m_plock= 0;
+      *m_plock= NULL;
+      m_plock= NULL;
     }
   }
   return tmp;
@@ -3691,7 +3692,8 @@ void select_create::abort()
   if (m_plock)
   {
     mysql_unlock_tables(thd, *m_plock);
-    m_plock= 0;
+    *m_plock= NULL;
+    m_plock= NULL;
   }
 
   if (table)