Please add 127b90315ca0 ("sched/proxy: Yield the donor task") to 6.12.y / 6.18.y

Please add 127b90315ca0 ("sched/proxy: Yield the donor task") to 6.12.y / 6.18.y

[Holger Hoffstätte]

Hi -

a Gentoo user recently found that 6.18.2 started to reproducuibly
crash when building their go toolchain [1].

Apparently the addition of "sched/fair: Forfeit vruntime on yield"
(mainline 79104becf42b) can result in the infamous NULL returned from
pick_eevdf(), which is not supposed to happen.

It turned out that the mentioned commit triggered a bug related
to the recently added proxy execution feature, which was already
fixed in mainline by "sched/proxy: Yield the donor task"
(127b90315ca0), though not marked for stable.

Applying this to 6.18.2/.3-rc1 (and probably 6.12 as well)
has reproducibly fixed the problem. A possible reason the crash
was triggered by the Go runtime could be its specific use of yield(),
though that's just speculation on my part.

So please add 127b90315ca0 ("sched/proxy: Yield the donor task")
to 6.18.y/6.12.y. I know we're already in 6.18.3-rc1, but the
crasher seems reproducible.

Fernand, please correct me if I got the explanations wrong.

Thanks!
Holger

[1] https://bugs.gentoo.org/968116 starting at #8

Re: Please add 127b90315ca0 ("sched/proxy: Yield the donor task") to 6.12.y / 6.18.y

[Wang Yugui]

Hi,

> Hi -
> 
> a Gentoo user recently found that 6.18.2 started to reproducuibly
> crash when building their go toolchain [1].
> 
> Apparently the addition of "sched/fair: Forfeit vruntime on yield"
> (mainline 79104becf42b) can result in the infamous NULL returned from
> pick_eevdf(), which is not supposed to happen.
> 
> It turned out that the mentioned commit triggered a bug related
> to the recently added proxy execution feature, which was already
> fixed in mainline by "sched/proxy: Yield the donor task"
> (127b90315ca0), though not marked for stable.
> 
> Applying this to 6.18.2/.3-rc1 (and probably 6.12 as well)
> has reproducibly fixed the problem. A possible reason the crash
> was triggered by the Go runtime could be its specific use of yield(),
> though that's just speculation on my part.
> 
> So please add 127b90315ca0 ("sched/proxy: Yield the donor task")
> to 6.18.y/6.12.y. I know we're already in 6.18.3-rc1, but the
> crasher seems reproducible.

Failed to apply 127b90315ca0 ("sched/proxy: Yield the donor task")
to 6.12.y, because  the 'donor task‘ is a feature of 6.13.

commit af0c8b2bf67b25756f27644936e74fd9a6273bd2
Author: Peter Zijlstra <peterz@infradead.org>
Date:   Wed Oct 9 16:53:40 2024 -0700

    sched: Split scheduler and execution contexts

Best Regards
Wang Yugui (wangyugui@e16-tech.com)
2025/12/31

Re: Please add 127b90315ca0 ("sched/proxy: Yield the donor task") to 6.12.y / 6.18.y

[Holger Hoffstätte]

On 2025-12-31 00:45, Wang Yugui wrote:
>> So please add 127b90315ca0 ("sched/proxy: Yield the donor task")
>> to 6.18.y/6.12.y. I know we're already in 6.18.3-rc1, but the
>> crasher seems reproducible.
> 
> Failed to apply 127b90315ca0 ("sched/proxy: Yield the donor task")
> to 6.12.y, because  the 'donor task‘ is a feature of 6.13.
> 
> commit af0c8b2bf67b25756f27644936e74fd9a6273bd2
> Author: Peter Zijlstra <peterz@infradead.org>
> Date:   Wed Oct 9 16:53:40 2024 -0700
> 
>      sched: Split scheduler and execution contexts
> 
> Best Regards
> Wang Yugui (wangyugui@e16-tech.com)
> 2025/12/31

Since I no longer use 6.12 I was not sure if 6.12 is really affected due
to gradual proxy execution rollout and possible backports, but I guess
now we know. Thank you for the clarification. :)

cheers
Holger

Re: Please add 127b90315ca0 ("sched/proxy: Yield the donor task") to 6.12.y / 6.18.y

[Greg KH]

On Tue, Dec 30, 2025 at 03:38:34PM +0100, Holger Hoffstätte wrote:
> Hi -
> 
> a Gentoo user recently found that 6.18.2 started to reproducuibly
> crash when building their go toolchain [1].
> 
> Apparently the addition of "sched/fair: Forfeit vruntime on yield"
> (mainline 79104becf42b) can result in the infamous NULL returned from
> pick_eevdf(), which is not supposed to happen.
> 
> It turned out that the mentioned commit triggered a bug related
> to the recently added proxy execution feature, which was already
> fixed in mainline by "sched/proxy: Yield the donor task"
> (127b90315ca0), though not marked for stable.
> 
> Applying this to 6.18.2/.3-rc1 (and probably 6.12 as well)
> has reproducibly fixed the problem. A possible reason the crash
> was triggered by the Go runtime could be its specific use of yield(),
> though that's just speculation on my part.
> 
> So please add 127b90315ca0 ("sched/proxy: Yield the donor task")
> to 6.18.y/6.12.y. I know we're already in 6.18.3-rc1, but the
> crasher seems reproducible.

Now queued up to 6.18.y, thanks.

greg k-h