maria-developers team mailing list archive

[Kristian Nielsen <knielsen@xxxxxxxxxxxxxxx>]

Jan Lindström <jplindst@xxxxxxxxxxx> writes:

> static int binlog_savepoint_set(handlerton *hton, THD *thd, void *sv)
> {

> #ifdef WITH_WSREP
>   if (wsrep_emulate_bin_log) DBUG_RETURN(0);
> #endif /* WITH_WSREP */
>
> wsrep_emulate_bin_log = 1, thus no savepoint is written to bin log

Ah, I missed that.

Right, then I think the problem is clear.

In binlog_savepoint_set(), we do at the end to initialise *sv:

  binlog_trans_log_savepos(thd, (my_off_t*) sv);

Then in binlog_savepoint_rollback() we read back the value that
binlog_savepoint_set() put in:

  binlog_trans_log_truncate(thd, *(my_off_t*)sv);

But the handling of WITH_WSREP and wsrep_emulate_bin_log=1 is done
incorrectly. Because in this case the initialisation in binlog_savepoint_set()
is disabled, but the value is still used in binlog_savepoint_rollback() (the
access is outside of if (!wsrep_emulate_bin_log && ...).

So this is clearly a problem with the WSREP patch related to savepoints. I am
not familiar with how that is supposed to work, so now you need to ask the
Codership people how this is supposed to work. Either
binlog_savepoint_rollback() should not access *sv in the wsrep_emulate_bin_log
case, or alternatively binlog_savepoint_set() _should_ initialise *sv in this
case.

> Thanks for the help, I do not seem to be able to run with that --debug=d,*:t:i 
> setup, it does not seem to progress.

Maybe it is just slow (especially if you also had Valgrind enabled), it will
generate lots and lots of output.

 - Kristian.