[I Cc'ed Paul, let me know if there is another prefered way of
communicating,
like using Launchpad bugs or some mailing list]
Vladimir Kolesnikov <vladimir@xxxxxxxxxxxxx> writes:
"Crash during shutdown (need help from PBXT team)"
can you provide any details?
Yes, of course! Just been busy with other stuff, but very happy
that you
mention it on your own initiative!
Basically, after merging PBXT I see crashes in the
main.mysqld_option_err test
case. Here is an example build from Buildbot showing this:
http://askmonty.org/buildbot/builders/debian5-i386-fulltest/builds/52
http://askmonty.org/buildbot/builders/debian5-i386-fulltest/builds/52/steps/test_4/logs/stdio
The failure appears to be somewhat random, but I was able to
repeat it easily
enough locally using a release build. Debug build did not seem to
show the
problem.
The issue is with the fix that Paul pushed for Bug#489088:
https://bugs.launchpad.net/pbxt/+bug/489088
=== modified file 'storage/pbxt/src/myxt_xt.cc'
--- storage/pbxt/src/myxt_xt.cc 2009-11-27 15:37:02 +0000
+++ storage/pbxt/src/myxt_xt.cc 2009-12-01 09:50:46 +0000
@@ -3041,6 +3041,14 @@ xtPublic MX_CHARSET_INFO *myxt_getcharse
return (MX_CHARSET_INFO *)&my_charset_utf8_general_ci;
}
+#ifdef DBUG_OFF
+//typedef struct st_plugin_int *plugin_ref;
+#define REF_MYSQL_PLUGIN(x) (x)
+#else
+//typedef struct st_plugin_int **plugin_ref;
+#define REF_MYSQL_PLUGIN(x) (*(x))
+#endif
+
xtPublic void *myxt_create_thread()
{
#ifdef DRIZZLED
+ LEX_STRING& plugin_name = REF_MYSQL_PLUGIN(new_thd-
>variables.table_plugin)->name;
+ if ((plugin_name.length == 4) && (strncmp(plugin_name.str,
"PBXT", plugin_name.length) == 0)) {
+ REF_MYSQL_PLUGIN(new_thd->variables.table_plugin)->ref_count--;
+ }
new_thd->thread_stack = (char *) &new_thd;
new_thd->store_globals();
lex_start(new_thd);
This code crashes because new_thd->variables.table_plugin is NULL
at this
point in the code (or at least sometimes is).
So this problem is similar to the previous problem handled just
above in the
code with global_system_variables.table_plugin being NULL.
I tried a patch like this:
=== modified file 'storage/pbxt/src/myxt_xt.cc'
--- storage/pbxt/src/myxt_xt.cc 2009-12-01 09:50:46 +0000
+++ storage/pbxt/src/myxt_xt.cc 2009-12-04 21:41:25 +0000
@@ -3112,10 +3112,15 @@ xtPublic void *myxt_create_thread()
* references to the PBXT plugin object and will effectively
deadlock the plugin so * that server will have to force
plugin shutdown. To avoid deadlocking and forced shutdown * we
must dereference the plugin after creating THD objects.
+ * Similarly to global_system_variables.table_plugin as
described above,
+ * new_thd->valriables.table_plugin can also become NULL due to
shutdown.
*/
- LEX_STRING& plugin_name = REF_MYSQL_PLUGIN(new_thd-
>variables.table_plugin)->name;
- if ((plugin_name.length == 4) && (strncmp(plugin_name.str,
"PBXT", plugin_name.length) == 0)) {
- REF_MYSQL_PLUGIN(new_thd->variables.table_plugin)->ref_count--;
+ plugin_ref table_plugin = new_thd->variables.table_plugin;
+ if (table_plugin) {
+ LEX_STRING& plugin_name = REF_MYSQL_PLUGIN(table_plugin)->name;
+ if ((plugin_name.length == 4) && (strncmp(plugin_name.str,
"PBXT", plugin_name.length) == 0)) {
+ REF_MYSQL_PLUGIN(table_plugin)->ref_count--;
+ }
}
new_thd->thread_stack = (char *) &new_thd;
new_thd->store_globals();
So this fixes the crashes, but of course re-introduces the
original problem in
Bug#489088 with warnings about forced shutdown.
I think you should be able to repeat the problem easily enough on
Linux with a
non-debug build running the test main.mysqld_option_err, but if
not let me
know and I will help set something up.
- Kristian.
