Debian Squeeze amd64

Hi,

Is the table MyISAM?
Could you please attach your cnf file or paste SHOW VARIABLES?

Thanks.

The tables are in MyISAM and i attached the vairables list.

Notice especially these values :
myisam_repair_threads = 8
key_cache_segments = 4

I can easily trigger it with a stress test. By far the most frequent value that the field gets populated with, at least in my env, is CR (13). Here are other examples in decimal codes that I was getting on the client side while executing SHOW PROCESSLIST.

63 63 1 16 63 127
13 0 0 63 0 0 0 0 38 0 0 0 0 0 0 0 0 0 63 0 0
etc.

Unfortunately I wasn't able to produce a syncpoint-based test case in any reasonable time. I think the problem is not so much the synchronization itself (it seems fairly simple – let the field value be generated, then wait till the query it belongs to finishes, proceed), but the fact that the problem is only visible when the old data is actually corrupted, which I guess is sporadic and depends on the memory layout, and I couldn't figure how to make it reliable.

So instead of a sync point I added a check and assertion in the SHOW PROCESSLIST code where it gets corrupted (in mysqld_list_processes(THD *thd,const char *user, bool verbose)):

=== modified file 'sql/sql_show.cc'
— sql/sql_show.cc 2013-05-11 17:31:50 +0000
+++ sql/sql_show.cc 2013-07-02 23:03:21 +0000
@@ -2282,6 +2282,12 @@
thd_info->query_string.charset());
if (!thd->variables.old_mode)
protocol->store(thd_info->progress, 3, &store_buffer);
+
+ if ( thd_info && thd_info->state_info && (strlen(thd_info->state_info) > 0) && !my_isalpha(system_charset_info,(thd_info->state_info)[0]) )
+

if (protocol->write())
break; /* purecov: inspected */
}

(sorry if it's not elegant, but I'm not a practicing developer, my code is supposed to be ugly to make the real coders' code look better).
The idea was to make sure that if the state_info exists and isn't empty, then it starts with a letter. Hopefully I did it right.

In fact, the value gets broken any time after
2223: thd_info->state_info= thread_state_info(tmp);
2224: if (mysys_var)
2225: mysql_mutex_unlock(&mysys_var->mutex);
2226:
2227: /* Lock THD mutex that protects its data when looking at it. */

I just chose the downmost point, because once broken, it remains broken, so the further down the higher probability.

With that and a primitive RQG grammar below, the assertion gets hit quite quickly, at least on my machine and on perro. Obviously, instead of the assertion you can do something smart in there.

1. RQG grammar:
   $ cat mdev4723.yy

thread1_init:
SET myisam_repair_threads = 8 ; SET GLOBAL max_allowed_packet = 1024*1024*1024
; DROP TABLE IF EXISTS `t` ; CREATE TABLE `t1` ( `p` INT, KEY (`p`) ) ENGINE=MyISAM ;

thread2:
SHOW PROCESSLIST;

query:
INSERT INTO `t1` SELECT `p` FROM `t1` LIMIT 100 ;

1. End of RQG grammar

runall command line:

perl ./runall-new.pl --grammar=mdev4723.yy --skip-gendata --threads=4 --queries=100M --duration=600 --basedir=<basedir> --vardir=<vardir>

gentest command line:

perl ./gentest.pl --grammar=mdev4723.yy --threads=4 --queries=100M --duration=600 --dsn='dbi:mysql:host=127.0.0.1:port=3306:user=root:database=test'

You can also try to increase the number of threads.

No specific server options are needed, server can be start with all defaults.
I built server with BUILD/compile-pentium-valgrind-max-no-ndb, ran tests on
bzr version-info
revision-id: monty@askmonty.org-20130611104943-se0nzq67eexlhd9z
revno: 3793
branch-nick: 5.5

If it's not reproducible in your environment, I can set it up on perro, I already checked it fails there.

I ran tests on 5.5-serg revision-id: sergii@pisem.net-20131111090500-ur3361v8k40y6v00 / revno: 3960.
The failure triggered by the injected assertion from the provided test case seems to be gone; but the field still happens to contain garbage, although more rarely. I modified the injection this way:

=== modified file 'sql/sql_show.cc'
— sql/sql_show.cc 2013-11-11 08:31:17 +0000
+++ sql/sql_show.cc 2013-11-11 15:42:20 +0000
@@ -2282,6 +2282,13 @@
thd_info->query_string.charset());
if (!thd->variables.old_mode)
protocol->store(thd_info->progress, 3, &store_buffer);
+
+ if ( thd_info && thd_info->state_info && (strlen(thd_info->state_info) > 0) )

+
if (protocol->write())
break; /* purecov: inspected */
}

Now it fails on the 2nd DBUG_ASSERT (while on a tree before the fix the first one fails).
The grammar is the same.

I've set up the test on perro under mariadb.

• log in as usual;
• cd mdev4723
• . ./run

It starts server on port 4723 and runs the RQG test. The basedir is ./5.5-serg (symlink). The logs go to ./vardir.
The patch above is already applied there.
Logs from the last failure are in ./vardir (you can see the garbage value printed before the assertion failure in vardir/mysql.err