[Gluster-devel] add-brick crashes client

[Emmanuel Dreyfus]

On Fri, Aug 03, 2012 at 04:52:11AM +0000, Emmanuel Dreyfus wrote:
> This means pthread_getspecific(synctask_key) in synctask_get() returned 
> NULL, something I cannot explain. I see the logs complain about a volume
> being down. This may be the cause of the problem since I have been able
> to do a live brick-add later, once I restarterd glusterd/glusterfsd on 
> all bricks.

Wrong, it did not succeed after the restart. I thought it did because
it did not crashed, but it just failed for another reason (the brick
already has glusterfs extended attributes and is considered already
in use, I had to cleanup extended attrbiute to crash with add-brick 
again).

Reading the code further, it seems that I crash here because 
syncenv_processor/synctask_switchto/synctask_set/pthread_setspecific
is not called before 
synctask_new/synctask_wrap/synctask_get/pthread_getspecific

It seems there is a race condition here. Someone knowledgable can
confirm?

-- 
Emmanuel Dreyfus
manu at netbsd.org