Bug#16581: deadlock: server and client both read from connection in
'conc_sys' test Concurrent execution of SELECT involing at least two INFORMATION_SCHEMA tables, DROP DATABASE statement and DROP TABLE statement could have resulted in stalled connection for this SELECT statement. The problem was that for the first query of a join there was a race between select from I_S.TABLES and DROP DATABASE, and the error (no such database) was prepared to be send to the client, but the join processing was continued. On second query to I_S.COLUMNS there was a race with DROP TABLE, but this error (no such table) was downgraded to warning, and thd->net.report_error was reset. And so neither result nor error was sent to the client. The solution is to stop join processing once it is clear we are going to report a error, and also to downgrade to warnings file system errors like 'no such database' (unless we are in the 'SHOW' command), because I_S is designed not to use locks and the query to I_S should not abort if something is dropped in the middle. No test case is provided since this bug is a result of a race, and is timing dependant. But we test that plain SHOW TABLES and SHOW COLUMNS give a error if there is no such database or a table respectively.
Showing
Please register or sign in to comment