Re: [BUG] TASK_DEAD task is able to be woken up in specialcondition

[Oleg Nesterov]

On 01/06, Yasunori Goto wrote:
>
> +change_task_state_cmpxchg(struct task_struct *p, unsigned int target_state,
> +			  unsigned int new)
>  {
> -	trace_sched_wakeup(p, true);
> -	check_preempt_curr(rq, p, wake_flags);
> +	unsigned int old, tmp;
>
> -	p->state = TASK_RUNNING;
> +	old = p->state;
> +
> +	while (1) {
> +		if (old == new)
> +			return 1;
> +
> +		if (unlikely(!(old & target_state)))
> +			return 0;
> +
> +		tmp = cmpxchg(&p->state, old, new);
> +		if (likely(tmp == old))
> +			return 1;
> +
> +		old = tmp;
> +	}

I do not really think we should retry if cmpxchg fails. The state was
changed after initial check, we can pretend it was changed after we
change it succesfully.

> @@ -2828,11 +2883,18 @@ try_to_wake_up(struct task_struct *p, un
>  	if (!(p->state & state))
>  		goto out;
>
> -	success = 1; /* we're going to change ->state */
>  	cpu = task_cpu(p);
>
> -	if (p->on_rq && ttwu_remote(p, wake_flags))
> -		goto stat;
> +	if (p->on_rq) {
> +		int ret;
> +		ret = ttwu_remote(p, state, wake_flags);
> +		if (likely(ret == 1)) {
> +			success = 1;
> +			goto stat;
> +		} else if (unlikely(ret < 0))
> +			goto out;
> +	}
> +	success = 1; /* we're going to change ->state */

But this is not enough, you should also recheck the state below. Although
in this case you do not need atomic ops, the target can do nothing with
its ->state.

Anyway. I have to agree with Peter even if I suggested this initially.
This adds the unpleasant complications. Hopefully you found the only
case when the spurious wakeup really hurts, probably it is better to
fix do_exit().

Oleg.

--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/