Re: [linux-next:master] [pipe_read] aaec5a95d5: stress-ng.poll.ops_per_sec 11.1% regression

[Oleg Nesterov <oleg@redhat.com>]

Sorry for delay,

On 01/20, Mateusz Guzik wrote:
>
> On Mon, Jan 20, 2025 at 9:31 PM Oleg Nesterov <oleg@redhat.com> wrote:
> > I'm afraid my emails can look as if I am trying to deny the problem.
> > No. Just I think we need to understand why exactly this patch makes
> > a difference.
> >
>
> I agree.
>
> I was going to state there is 0 urgency as long as the patch does not
> make the merge window, but it just did.

Yes...

> So one would preferably survey a bunch of real workloads, see what
> happens with real pipes with both policies -- the early wake up is
> basically a tradeoff and it very well may be it is worth it in the
> real world.

The problem is that this early wakeup is not intended, the code is
not supposed to do this. So in some sense this patch fixes the
intended/documented "avoid unnecessary wakeups" logic.

Now I can reproduce the hackbench's slowdown on my laptop, but still
don't understand it... I'll try to think more on Weekend, then I'll
discuss the possible revert with Linus who wrote that code and
reviewed this patch.

Thanks for your investigations,

Oleg.