Groups | Search | Server Info | Keyboard shortcuts | Login | Register [http] [https] [nntp] [nntps]

Groups > linux.kernel > #1646858 > unrolled thread

Core isolation, cpu hotplug and stop_machine

Back to article view | Back to linux.kernel

Contents

  Core isolation, cpu hotplug and stop_machine Fredrik Markström <fredrik.markstrom@gmail.com> - 2017-05-22 15:10 +0200
    Re: Core isolation, cpu hotplug and stop_machine Thomas Gleixner <tglx@linutronix.de> - 2017-05-22 21:10 +0200

#1646858 — Core isolation, cpu hotplug and stop_machine

Hello,

As I understand it the recommended way to force pending timers of a
core when trying to isolate it (runtime) is to use cpu hotplug. The
current implementation of cpu hotplug has a serious drawback in that
it uses stop_machine() which disturbs already isolated cores. This is
a serious problem in our usecase.

I would guess we are not the only ones with this problem, but as far
as I can see there is currently no solution for this in the latest
next. There was a patchset back in 2013 that implemented
stop_machine-free hotplug, but that was never accepted.
(https://lwn.net/Articles/538819/)

So is there any ongoing work to solve this problem and what is the
preferred way to solve it, is it:

1) stop_machine() free hotplug. It seems to be quite complex but
probably solves more problems
    then the one above.
2) Invent some new mechanism/interface separate from hotplug to
migrate timers and hrtimers.
3) Don't, because this isn't a use-case that linux should support.

/Fredrik

[toc] | [next] | [standalone]

#1647272

[Multipart message — attachments visible in raw view] — view raw

On Mon, 22 May 2017, Fredrik Markström wrote:
> So is there any ongoing work to solve this problem and what is the
> preferred way to solve it, is it:
> 
> 1) stop_machine() free hotplug. It seems to be quite complex but
> probably solves more problems
>     then the one above.

Yes, but it's a non-trivial problem. I looked into it not too long ago and
it just refuses to fall in place due to circular dependencies. It's on my
todo list, but certainly not on the top.

And aside of that cpu hotplug is a vehicle and not a solution.

> 2) Invent some new mechanism/interface separate from hotplug to
> migrate timers and hrtimers.

There is ongoing work to make this nonsensical timer placement stuff
sane. It's buggy atm, but we are working on a fix. We have a similar plan
for hrtimers.

  https://marc.info/?i=20170418111102.490432548%40linutronix.de

> 3) Don't, because this isn't a use-case that linux should support.

Well, we want full isolation to work, but we want proper solutions for it.

Thanks,

	tglx

[toc] | [prev] | [standalone]

Back to top | Article view | linux.kernel

csiph-web