• Alfranio Correia's avatar
    BUG#50038 Deadlock on flush logs with concurrent DML and RBR · e264ff39
    Alfranio Correia authored
    In auto-commit mode, updating both trx and non-trx tables (i.e. issuing a mixed
    statement) causes the following sequence of events:
    
    1 - "Flush trx changes" (MYSQL_BIN_LOG::write) - T1:
      1.1 - mutex_lock (&LOCK_log)
      1.2 - mutex_lock (&LOCK_prep_xids)
      1.3 - increase prepared_xids
      1.4 - mutex_unlock (&LOCK_prep_xids)
      1.5 - mutex_unlock (&LOCK_log)
    
    2 - "Flush non-trx changes" (MYSQL_BIN_LOG::write) - T1:
      2.1 - mutex_lock (&LOCK_log)
      2.2 - mutex_unlock (&LOCK_log)
    
    3. "unlog" - T1
      3.1 - mutex_lock (&LOCK_prep_xids)
      3.2 - decrease prepared xids
      3.3 - pthread_cond_signal(&COND_prep_xids);
      3.4 - mutex_unlock (&LOCK_prep_xids)
    
    The "FLUSH logs" command produces the following sequence of events:
    
    1 - "FLUSH logs" command (MYSQL_BIN_LOG::new_file_impl) - user thread:
      1.1 - mutex_lock (&LOCK_log)
      1.2 - mutex_lock (&LOCK_prep_xids)
      1.3 - while (prepared_xids)  pthread_cond_wait(..., &LOCK_prep_xids);
      1.4 - mutex_unlock (&LOCK_prep_xids)
      1.5 - mutex_unlock (&LOCK_log)
    
    A deadlock will arise if T1 flushes the trx changes and thus increases
    prepared_xids but before it is able to continue the execution and flush the
    non-trx changes, an user thread calls the "FLUSH logs" command and wait that
    the prepared_xids is decreased and gets to zero. However, T1 cannot proceed
    with the call to "Flush non-trx changes" because it will block in the mutex
    "LOCK_log" and by consequence cannot complete the execution and call the
    unlog to decrease the prepared_xids.
    
    To fix the problem, we ensure that the non-trx changes are always flushed
    before the trx changes.
    
    Note that if you call "Flush non-trx changes" and a concurrent "FLUSH logs" is
    issued, the "Flush non-trx changes" may block, but a deadlock will never happen
    because the prepared_xids will eventually get to zero. Bottom line, there will
    not be any transaction able to increase the prepared_xids because they will
    block in the mutex "LOCK_log" (MYSQL_BIN_LOG::write) and those that increased
    the prepared_xids will eventually commit and decrease the prepared_xids.
    e264ff39
binlog_stm_innodb_stat.result 610 Bytes