No Linux logs when doing `ppc64_cpu --smt=off/8`

[Joel Stanley]

On Thu, 17 Feb 2022 at 01:07, Michael Ellerman <mpe at ellerman.id.au> wrote:
>
> Michal Suchánek <msuchanek at suse.de> writes:
> > On Mon, Feb 14, 2022 at 01:33:24PM +0100, Paul Menzel wrote:

> >> Yes, simple `nproc` suffice, but I was more thinking about, that the Linux
> >> log is often used for debugging and the changes of amount of processing
> >> units might be good to have. `ppc64_cpu --smt=off` or `=8` seems to block
> >> for quite some time, and each thread/processing unit seems to powered
> >> down/on sequentially, so it takes quite some time and it blocks. So 140
> >> messages would indeed be quite noise. No idea how `ppc64_cpu` works, and if
> >> it could log a message at the beginning and end.
> >
> > Yes, it enables/disables threads one by one. AFAICT the kernel cannot know that
> > ppc64_cpu will enable/disable more threads later, it can either log each
> > or none. Rate limiting would not show the whole picture so it's not
> > great solution either.
>
> Right, ppc64_cpu just uses the sysfs online files, so it's doing them one at a
> time. The kernel has no knowledge that ppc64_cpu is turning all
> secondaries on/off so there's no easy way for the kernel to do a summary
> message.
>
> An easy solution would be for ppc64_cpu to log something via syslog(3).

Another option would be to implement this sysfs knob for powerpc:

 /sys/devices/system/cpu/smt/control

 https://github.com/torvalds/linux/commit/05736e4ac13c08a4a9b1ef2de26dd31a32cbee57

It currently returns "notimplemented" on power, but it's designed (for
x86) to turn SMT on and off.

With this implemented we could offline all CPUs at once before
returning to userspace. It "just" needs someone with some time to work
on it.

...which I spent the afternoon attempting to do:

 https://lore.kernel.org/linuxppc-dev/20220217070419.351683-1-joel@jms.id.au/

It does what it says on the tin, but there's a few rough edges that
need working out. It won't solve the problem you had of wanting a
message in the kernel logs (I notice that the x86 cpu hotplug code
does print something; I wonder how that goes on a large box). If you
or anyone else wants to run with the idea then free free to use my
patch as a starting point.

Cheers,

Joel