A follow up after the fix for Bug#21074 - fix NDB tests breaking on

asserts.
The patch for Bug#21074 replaces acquisition of the global LOCK_open lock
with exclusive locks on table names in such operations ad DROP TABLE
and RENAME TABLE.
Unfortunately, NDB internally assumes that LOCK_open is acquired and
tries to release it.
This dependency should be fixed by a separate (and significant in size)
patch. For now we just satisfy it - after all, the original
goal of the patch for Bug#21074 was to move query_cache_invalidate
outside of the scope of LOCK_open, and we still can do that.

This fixes some failing NDB tests in the runtime tree.


sql/sql_rename.cc:
  Move release of LOCK_open after ha_ndbcluster::rename_tables to
  satisfy an assert in ndb_log_schema_op.

1 files changed, 11 insertions, 1 deletions

diff --git a/sql/sql_rename.cc b/sql/sql_rename.cc
index e68a360f2d4..f5e1b8988f3 100644
--- a/sql/sql_rename.cc
+++ b/sql/sql_rename.cc
@@ -150,7 +150,6 @@ bool mysql_rename_tables(THD *thd, TABLE_LIST *table_list, bool silent)
     pthread_mutex_unlock(&LOCK_open);
     goto err;
   }
-  pthread_mutex_unlock(&LOCK_open);
 
   error=0;
   if ((ren_table=rename_tables(thd,table_list,0)))
@@ -174,6 +173,17 @@ bool mysql_rename_tables(THD *thd, TABLE_LIST *table_list, bool silent)
 
     error= 1;
   }
+  /*
+    An exclusive lock on table names is satisfactory to ensure
+    no other thread accesses this table.
+    However, NDB assumes that handler::rename_tables is called under
+    LOCK_open. And it indeed is, from ALTER TABLE.
+    TODO: remove this limitation.
+    We still should unlock LOCK_open as early as possible, to provide
+    higher concurrency - query_cache_invalidate can take minutes to
+    complete.
+  */
+  pthread_mutex_unlock(&LOCK_open);
 
   /* Lets hope this doesn't fail as the result will be messy */ 
   if (!silent && !error)