Re: [PATCH 2/2] posix-cpu-timers: remove tasklist_lock where we can

[Oleg Nesterov <oleg@redhat.com>]

On 06/12, Stanislaw Gruszka wrote:
>
> tasklist_lock is not needed to protect find_task_by_vpid() nor
> thread_group_leader()

It does protect thread_group_leader(), unless we use current.
Please see below.

> @@ -398,7 +398,7 @@ int posix_cpu_timer_create(struct k_itimer *new_timer)
>  	new_timer->it.cpu.incr.sched = 0;
>  	new_timer->it.cpu.expires.sched = 0;
>
> -	read_lock(&tasklist_lock);
> +	rcu_read_lock();
>  	if (CPUCLOCK_PERTHREAD(new_timer->it_clock)) {
>  		if (pid == 0) {
>  			p = current;
> @@ -422,7 +422,7 @@ int posix_cpu_timer_create(struct k_itimer *new_timer)
>  	} else {
>  		ret = -EINVAL;
>  	}
> -	read_unlock(&tasklist_lock);
> +	rcu_read_unlock();

Suppose that the non-main thread execs. de_thread() does

	attach_pid(tsk, PIDTYPE_PID, task_pid(leader));
	leader->group_leader = tsk;

under write_lock(tasklist). This means posix_cpu_timer_create() can return
-EINVAL if it is called in between, or if we found the group leader but
thread_group_leader() is called after "tsk->group_leader = tsk".

I think the patch is fine, but you should also replace thread_group_leader()
with has_group_leader_pid().



This reminds me. !CPUCLOCK_PERTHREAD timers survive after exec, but only
if ->group_leader execs. Not good.

Oleg.