On Tue, Aug 20, 2013 at 1:09 AM, Sergei Golubchik <serg@mariadb.org> wrote:
Hi, Pavel!
On Aug 19, Pavel Ivanov wrote:
No, it's not reasonable for semi-sync to lock/unlock LOCK_plugin. It's plugin infrastructure that does that.
I've actually was terrified to learn that each call into semi-sync plugin is surrounded with pthread_rwlock_rdlock/pthread_rwlock_unlock (which is not cheap I believe). And also for each such call it "locks"/"unlocks" the semi-sync plugin. And although "locking" plugin avoids locking LOCK_plugin when plugin is linked statically, "unlocking" doesn't do that.
Unlocking does that too, it's completely symmetrical. See plugin_unlock() function, it starts from
/* built-in plugins don't need ref counting */ if (!plugin_dlib(plugin)) DBUG_VOID_RETURN;
Sure, but macro FOREACH_OBSERVER inside rpl_handler.cc doesn't use this function. It uses plugin_unlock_list() which always locks LOCK_plugin. BTW, MariaDB still supports compiling semi-sync plugins dynamically, but it seems that it doesn't do anything against unloading semi-sync plugins in the middle of transactions. Did anyone think about this? Pavel