Re: [PATCH] eventpoll: Fix priority inversion problem

[Valentin Schneider]

On 19/05/25 11:45, Nam Cao wrote:
> On Mon, May 19, 2025 at 11:25:51AM +0200, Florian Bezdeka wrote:
>> Hi all,
>>
>> sorry for top-posting, but I think it makes sense in this case as I'm
>> trying to merge different workstreams, likely working on the same
>> problem showing up in different colors.
>>
>> Main goal is to make everybody aware of the other stream / patch
>> series.
>>
>> We have colleagues from Bytedance working on non-RT performance
>> optimizations related to CONFIG_CFS_BANDWIDTH at [1].
>>
>> This series came to light while searching for a solution for a RT
>> lockup, reported at [2].
>>
>> We heavily tested [1] during the last month on RT and can report
>> success now. In our tests we saw read-lock holder preemption only
>> within the epoll interface. It might be that [1] fixes more potential
>> issues in this regard.
>>
>> Today [3] (= the patch I'm replying to, see below) got posted.
>> Linutronix reworking the epoll infrastructure.
>>
>> I would love to learn how/if the combination, basically [1] and [3] fit
>> together.
>
> [1] fixes stall problem involving rw semaphore which epoll uses, but it
> doesn't fix the possible priority inversion with epoll
>
> [3] fixes priority inversion problem with epoll by stop using rw semaphore,
> but it doesn't do anything about rw semaphore
>
> So I propose we keep both.
>

FWIW I'm thinking the same.

[1] fixes the CFS throttling circular dependency on RT but also improves
tail latency when CFS throttling is in use (so not just for RT).
[3] is specific to epoll but gets rid of the rwsem which is a win overall.

I need to go get my reviewing hat back on and stare at both...

> Best regards,
> Nam
>
>> My understanding right now is, that [1] fixes a CFS issue, throttling
>> while holding a lock is not ideal on !RT - but might cause a critical
>> lockup on RT - while [3] is addressing a similar (RT) problem in epoll.
>>
>> Best regards,
>> Florian
>>
>> [1] https://lore.kernel.org/all/20250409120746.635476-1-ziqianlu@xxxxxxxxxxxxx/
>> [2] https://lore.kernel.org/linux-rt-users/xhsmhttqvnall.mognet@xxxxxxxxxxxxxxxxxxx/
>> [3] https://lore.kernel.org/linux-rt-users/20250519074016.3337326-1-namcao@xxxxxxxxxxxxx/T/#u