maria-developers team mailing list archive

[Michael Widenius <monty@xxxxxxxxxxxx>]

Hi!

>>>>> "Alex" == Alex Budovski <abudovski@xxxxxxxxx> writes:

Alex> Re: my_thread_init
Alex> I tried reverting my change and looking at the code again -- perhaps
Alex> we have a different bug.

Alex> The function does indeed set 'id' to a non-zero value (tmp->id =
Alex> ++thread_count), but then it is reset to 0 by someone else.

Alex> Let me show you:

Alex> First of all, we hit a verifier stop telling us a CS was already initialized:

<cut>

Alex> So it was reset by someone else, after being initialized, causing the
Alex> "if (THR_KEY_mysys.id)" to fail and re-initialize the CSs (there's
Alex> more than one).

Alex> So let's see who is writing to it: (by setting a data breakpoint on
Alex> 0x03338f70+0x054)

Alex> ChildEBP RetAddr
Alex> 0012f608 008967fa mysqld!THD::store_globals+0xa4 <-- here
Alex> 0012f624 008745f9 mysqld!myxt_create_thread+0x10a
Alex> 0012f648 004d46f5 mysqld!pbxt_init+0x3f9
Alex> 0012f6a0 00471f97 mysqld!ha_initialize_handlerton+0xa5
Alex> 0012f6e4 00471cf1 mysqld!plugin_initialize+0x67
Alex> 0012f86c 00447d44 mysqld!plugin_init+0x541
Alex> 0012fb7c 0044595c mysqld!init_server_components+0x5c4
Alex> 0012fd30 00448a08 mysqld!win_main+0x1cc
Alex> 0012fd40 00448d7a mysqld!mysql_service+0x38
Alex> 0012ff6c 00401137 mysqld!main+0x35a
Alex> 0012ffb8 0040100f mysqld!__tmainCRTStartup+0x117
Alex> 0012ffc0 7c817077 mysqld!mainCRTStartup+0xf
Alex> 0012fff0 00000000 kernel32!BaseProcessStart+0x23

Alex> THD::store_globals() does the following:

Alex>   mysys_var=my_thread_var;
Alex>   mysys_var->id= thread_id;  // Let's see what this->thread_id is.

Alex> 2:005> ?? this->thread_id
Alex> unsigned long 0

Alex> So here it gets set to 0, and hence our code fails.

Alex> Do you know why this might be?

Yes, this explain it.

We set the mysys_var->id to thread_id in store_globals() to allow
MySQL to bind a user connection to different threads during execution
of one query.  We need to change mysys->thread_id so that all reports
(like dbug_print) are printed with the same logical thread id.

The bug in question after this change was that thread_id was not
initialized to 1 and the above call to store_global() was done before
any creating of a user connection.

Thanks to sort this out.  This makes it clear that it was a correct
change to check for 'init'.

Regards,
Monty