Re: [Bug] premature worker thread wakeup

[Valentin Schneider]

On 20/09/24 15:25, Yan, Zheng wrote:
>
> Looks like the newly created kthread got woke up prematurely. Checking
> the source code, path "workqueue: Unbind kworkers before sending them
> to exit()" looks suspicious. The patch wakes up dying worker's task up
> without holding pool->lock.  Is it possible that another worker thread
> recently died, the newly create worker thread reused the dead thread's
> task_struct and wrongly  got woke up.
>

Per the comment in idle_cull_fn():

         * Grabbing wq_pool_attach_mutex here ensures an already-running worker
         * cannot proceed beyong set_pf_worker() in its self-destruct path.
         * This is required as a previously-preempted worker could run after
         * set_worker_dying() has happened but before detach_dying_workers() did.

the task shouldn't be able to exit() even if we don't hold &pool->lock.

Also, things have changed a little in this area, for instance see
  f4b7b53c94af ("workqueue: Detach workers directly in idle_cull_fn()")
Could you try to reproduce this using a more recent upstream kernel?


> Regards
> Yan, Zheng