Re: [PATCH v3 1/1] workqueue: ignore dead tasks in a workqueue sleep hook

[Oleg Nesterov]

On 10/25, Roman Pen wrote:
>
>  struct task_struct *wq_worker_sleeping(struct task_struct *task)
>  {
> -	struct worker *worker = kthread_data(task), *to_wakeup = NULL;
> +	struct worker *worker, *to_wakeup = NULL;
>  	struct worker_pool *pool;
>  
> +
> +	if (task->state == TASK_DEAD) {
> +		/*
> +		 * Here we try to catch the following path before
> +		 * accessing NULL kthread->vfork_done ptr thru
> +		 * kthread_data():
> +		 *
> +		 *    oops_end()
> +		 *    do_exit()
> +		 *    schedule()
> +		 *
> +		 * If panic_on_oops is not set and oops happens on
> +		 * a workqueue execution path, thread will be killed.
> +		 * That is definitly sad, but not to make the situation
> +		 * even worse we have to ignore dead tasks in order not
> +		 * to step on zeroed out members (e.g. t->vfork_done is
> +		 * already NULL on that path, since we were called by
> +		 * do_exit())).
> +		 */
> +		return NULL;
> +	}

I still think that PF_EXITING check makes more sense than TASK_DEAD,
but I won't insist.

Oleg.