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

From: Oleg Nesterov
Date: Tue Oct 25 2016 - 08:57:54 EST


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.