Re: [PATCH] x86/mce/therm_throt: Handle case where throttle_active_work() is called on behalf of an offline CPU

From: Pandruvada, Srinivas
Date: Mon Feb 24 2020 - 16:05:12 EST


On Mon, 2020-02-24 at 20:25 +0100, Thomas Gleixner wrote:
> Thomas Gleixner <tglx@xxxxxxxxxxxxx> writes:
> > Which is wrong as well. Trying to "fix" it in the work queue
> > callback is
> > papering over the root cause.
> >
> > Why is any work scheduled on an outgoing CPU after this CPU
> > executed
> > thermal_throttle_offline()?
> >
> > When thermal_throttle_offline() is invoked the cpu bound work
> > queues are
> > still functional and thermal_throttle_offline() cancels outstanding
> > work.
> >
> > So no, please fix the root cause not the symptom.
>
> And if you look at thermal_throttle_online() then you'll notice that
> it
> is asymetric vs. thermal_throttle_offline().
>
> Also you want to do cancel_delayed_work_sync() and not just
> cancel_delayed_work() because only the latter guarantees that the
> work
> is not enqueued anymore while the former does not take running or
> self
> requeueing work into account.
>
> Something like the untested patch below.
Thanks Thomas. I am testing this patch.

-Srinivas

>
> Thanks,
>
> tglx
> ---
> --- a/arch/x86/kernel/cpu/mce/therm_throt.c
> +++ b/arch/x86/kernel/cpu/mce/therm_throt.c
> @@ -487,8 +487,12 @@ static int thermal_throttle_offline(unsi
> struct thermal_state *state = &per_cpu(thermal_state, cpu);
> struct device *dev = get_cpu_device(cpu);
>
> - cancel_delayed_work(&state->package_throttle.therm_work);
> - cancel_delayed_work(&state->core_throttle.therm_work);
> + /* Mask the thermal vector before draining evtl. pending work
> */
> + l = apic_read(APIC_LVTTHMR);
> + apic_write(APIC_LVTTHMR, l | APIC_LVT_MASKED);
> +
> + cancel_delayed_work_sync(&state->package_throttle.therm_work);
> + cancel_delayed_work_sync(&state->core_throttle.therm_work);
>
> state->package_throttle.rate_control_active = false;
> state->core_throttle.rate_control_active = false;