Re: [RFC,PATCH 2/2] change __wake_up_parent() to use filtered wakeup

[Oleg Nesterov <oleg@redhat.com>]

On 06/24, Roland McGrath wrote:
>
> Looks good, I'm glad to see this revived.
>
> I note that even simpler than eligible_child() is just:

I think the check below is orthogonal to eligible_child(). Not sure
eligible_child() can really help, but otoh it is cheap and doesn't hurt.
But perhaps we can kill it later.


> 	if ((wo->wo_flags & __WNOTHREAD) && wo->child_wait.private != p->parent)
> 		return 0;
>
> IIRC that is the test that Ratan's original patch used to address the
> particular application usage that first troubled him.

Aha, now I see what was the problem with Ratan's workload.

> But probably this
> is already what you meant by "more clever later"

I didn't mean this particular optimization, but it looks good to me.

> (and ->parent is perhaps
> not right in all cases there).

I think this is right... Except I'd like to avoid using ->parent.

> Your two patches as they are look safe and useful to me and I hope they can
> go in soon.

Thanks.

Yes I think these 2 patches should be applied first, even if eligible_child()
itself doesn't buy much. It will be cleaner if we add "real" checks on top.

Oleg.