Re: [PATCH] kthreads: Fix startup synchronization boot crash

[Oleg Nesterov <oleg@redhat.com>]

On 09/01, Ingo Molnar wrote:
>
> In fact i dont see any proper serialization here: there appears to
> be a race between the initial task and the init task (which are not
> one and the same). The race is possibly timing dependent as well,
> hence the (in hindsight, false) dependency on the stackprotector
> commit.

Yes, this looks racy, and I think this was always racy.

> I think the bug was introduced
> via:
>
>   cdd140b: kthreads: simplify the startup synchronization

Cough ;) No, I don't think this patch introduced this bug. With or without
this patch, kthread_create() assumes kthreadd_task != NULL, otherwise
wake_up_process(kthreadd_task) is obviously can crash.

>  static DEFINE_SPINLOCK(kthread_create_lock);
>  static LIST_HEAD(kthread_create_list);
> +
>  struct task_struct *kthreadd_task;
> +DECLARE_COMPLETION(kthreadd_task_init_done);
>
>  struct kthread_create_info
>  {
> @@ -129,6 +131,9 @@ struct task_struct *kthread_create(int (*threadfn)(void *data),
>  	list_add_tail(&create.list, &kthread_create_list);
>  	spin_unlock(&kthread_create_lock);
>
> +	if (unlikely(!kthreadd_task))
> +		wait_for_completion(&kthreadd_task_init_done);
> +

Yes, this should work. But I _think_ we can make the better fix...

I'll try to make the patch soon. Afaics we don't need kthreadd_task_init_done.

Oleg.