Shutdown hangs if redo log tracker thread stopped

[Laurynas Biveinis]

The following bit of percona_changed_page_bmp

#
# Test that impossible to track log is handled gracefully (bug 1108613)
#
CREATE TABLE t3 (a MEDIUMBLOB) ENGINE=InnoDB;

call mtr.add_suppression("InnoDB: Error: the age of the oldest untracked record exceeds the log group capacity!");
call mtr.add_suppression("InnoDB: Error: stopping the log tracking thread at LSN");

INSERT INTO t3 VALUES (REPEAT('a', 12582912));

--echo 10th restart
--source include/restart_mysqld.inc

always results in server shutdown hanging and MTR issuing kill -9.

This is caused by the tracking margin code setting srv_track_changed_pages = FALSE as tracking cannot continue, and then never raising the tracking event so that the tracker thread can wake up and quit.

The root cause is srv_track_changed_pages attempting to handle two different purposes: 1) whether the tracking is currently active; 2) whether the tracker thread has been started.

By code review this mismatch could cause issues elsewhere too, e.g. in log_online_purge_changed_page_bitmaps, which handles PURGE CHANGED_PAGE_BITMAPS:

        if (srv_track_changed_pages) {
                /* User requests might happen with both enabled and disabled
                tracking */
                mutex_enter(&log_bmp_sys->mutex);
        }
...
        if (srv_track_changed_pages) {
...
                mutex_exit(&log_bmp_sys->mutex);
        }

If srv_track_changed_pages would be reset from TRUE to FALSE in parallel, this would result in an unreleased mutex.

[Laurynas Biveinis (laurynas-biveinis)]

    https://github.com/percona/percona-server/pull/711
    https://github.com/percona/percona-server/pull/712
    https://github.com/percona/percona-server/pull/713

[Shahriyar Rzayev (rzayev-sehriyar)]

Percona now uses JIRA for bug reports so this bug report is migrated to: https://jira.percona.com/browse/PS-1730