Re: [PATCH 0/2] pipe: Fixes [ver #2]

[Linus Torvalds <torvalds@linux-foundation.org>]

[-- Attachment #1: Type: text/plain, Size: 1665 bytes --]

On Sun, Dec 8, 2019 at 10:04 AM Linus Torvalds
<torvalds@linux-foundation.org> wrote:
>
> On Sun, Dec 8, 2019 at 8:46 AM Akemi Yagi <toracat@elrepo.org> wrote:
> >
> > I forgot to mention that the aforementioned bug in make was originally
> > reported for Fedora. They now have a patched version of make
> > (make-4.2.1-15.fc32) in Rawhide.
>
> Yes, as mentioned I do expect it's some jobserver interaction - I've
> seen bad jobserver behavior before. But my 'make' hasn't been upgraded
> since May, as far as I can tell from my logs, so the huge performance
> regression was new.

Interestingly, when I'm trying to do the non-thundering-herd pipe
reads and writes, a side effect of that is that the read target is
"fair" (ie when there are multiple concurrent waiting read() calls,
it's always the first one that gets it).

And that seems to really mess up the jobserver bug, and triggers it
every single time.

Fairness is often bad for locking throughput, but this is on another
level entirely, and yeah, I see a lot of "defunct" shell and make
processes, so it does seem to be the make bug.

(Adding scheduler people to the participants list, because this patch
is the "avoid using the return value of
wait_event_interruptible_exclusive()" version because I think
wait_event_interruptible_exclusive() is itself buggy).

I built a new version of 'make' from source, and that one seems to
work fine with the attached patch. But the standard Fedora 'make'
binary completely hates this patch.

Sad. The patch does seem to be the RightThing(tm) to do, but this make
bug makes it inadvisable to apply it unless you want to play with
things.

                   Linus

[-- Attachment #2: patch.diff --]
[-- Type: application/x-patch, Size: 9612 bytes --]