Re: [RESEND PATCH] sched: restore the behavior of put_task_struct() for non-rt

[Oleg Nesterov]

On 09/16, Peter Zijlstra wrote:
>
> On Mon, Sep 15, 2025 at 02:35:40PM +0200, Oleg Nesterov wrote:
> > On 09/15, Peter Zijlstra wrote:
> > >
> > > Why have !RT behave differently? That is, why isn't this simply a
> > > 'buggy' comment/changelog issue?
> >
> > Well, this was discussed several times, in particular see
> > https://lore.kernel.org/all/CAHk-=whtj+aSYftniMRG4xvFE8dmmYyrqcJyPmzStsfj5w9r=w@xxxxxxxxxxxxxx/
> >
> > And task_struct->rcu_users was introduced to avoid RCU call in put_task_struct() ...
>
> Ah, I forgot about that thing.. Although I had vague memories of that
> argument on rcu_assign_pointer() vs RCU_INIT_POINTER().
>
> > But I won't really argue if you decide to remove this !RT optimization, although
> > I think it would be better to do this in a separate patch.
>
> Right. So when they wanted to remove that preemptible() clause, I was
> like why again do we have this !RT exception at all, and can't we get
> rid of that.
>
> If git isn't confusing me again, this got merged in this cycle. But so
> far no benchmark came and told us this was a bad idea.

I still think it would be safer to merge this patch from Luis before
v6.17, then possibly remove this special case in a separate patch...

> So what do we do now... do we restore the !RT exception (so far there
> aren't any numbers to suggest this mattered) or do we let it be for a
> bit and then go and clean things up?

It is not that simple. Please note that put_task_struct_rcu_user()
delays put(tsk->usage), not free(tsk).

So for example with this change

> @@ -305,7 +288,7 @@ void release_task(struct task_struct *p)
>  	if (thread_group_leader(p))
>  		flush_sigqueue(&p->signal->shared_pending);
>
> -	put_task_struct_rcu_user(p);
> +	put_task_struct(p);
>
>  	p = leader;
>  	if (unlikely(zap_leader))

This code

	rcu_read_lock();
	tsk = find_task_by_vpid(...);
	if (tsk)
		get_task_struct(tsk);
	rcu_read_unlock();

becomes wrong, get_task_struct(tsk) can increment tsk->usage == 0.

Oleg.