Re: [pipe] 3b844826b6: stress-ng.sigio.ops_per_sec -99.3% regression

[Linus Torvalds]

On Tue, Aug 24, 2021 at 7:56 AM kernel test robot <oliver.sang@xxxxxxxxx> wrote:
>
> FYI, we noticed a -99.3% regression of stress-ng.sigio.ops_per_sec due to commit:

Well, that's bad.

> commit: 3b844826b6c6 ("pipe: avoid unnecessary EPOLLET wakeups under normal loads")

You fix one benchmark, you break another..

What's a bit odd is that this commit basically reverts commit
3a34b13a88ca ("pipe: make pipe writes always wake up readers") which
did *not* result in any kernel test robot report.

It's not a pure revert, because it adds that 'poll_usage' case (for
EPOLLET), but the stress-ng.sigio test doesn't even use select or poll
(ok, there's a select() call with an empty file descriptor set, which
seems to be just an odd way to spell "usleep()").

So it _looks_ to me like it's a 100% revert in practice for that test.
I strace'd the "stress-ng --sigio" case just to make sure I didn't
miss anything.

But I'm clearly missing something. Can anybody see what I'm missing,
and hit me over the head with the clue-bat?

                      Linus