Re: [PATCH 1/3] do_wait reorganization

[Oleg Nesterov <oleg@tv-sign.ru>]

(I apologize in advance if I missed something, I'm not able to study this series
 carefully now. But I hope that a false alarm is better than a probable bug.)

On 03/30, Roland McGrath wrote:
> +static int wait_consider_task(struct task_struct *parent,
> +			      struct task_struct *p, int *retval,
> +			      enum pid_type type, struct pid *pid, int options,
> +			      struct siginfo __user *infop,
> +			      int __user *stat_addr, struct rusage __user *ru)
> +{
> +	int ret = eligible_child(type, pid, options, p);
> +	if (ret <= 0)
> +		return ret;
> +
> +	/*
> +	 * We don't reap group leaders with subthreads.
> +	 */
> +	if (p->exit_state == EXIT_ZOMBIE && !delay_group_leader(p))
> +		return wait_task_zombie(p, options, infop, stat_addr, ru);
> +
> +	/*
> +	 * It's stopped or running now, so it might
> +	 * later continue, exit, or stop again.
> +	 */
> +	*retval = 0;
> +
> +	if (task_is_stopped_or_traced(p))
> +		return wait_task_stopped(p, options, infop, stat_addr, ru);
> +
> +	if (p->exit_state != EXIT_DEAD)
> +		return wait_task_continued(p, options, infop, stat_addr, ru);
> +
> +	return 0;
> +}

This looks wrong, we shouldn't clear *retval if the child is EXIT_DEAD.

We can race with another thread which changed EXIT_ZOMBIE->EXIT_DEAD but
hasn't released the child yet. Suppose we don't have other childs, in that
case do_wait() will falsely sleep on ->wait_chldexit (unless WHOHANG).

Oleg.