Re: [PATCH v3 15/24] thermal: intel: hfi: Report the IPC class score of a CPU

[Rafael J. Wysocki]

On Wed, Mar 29, 2023 at 1:30 AM Ricardo Neri
<ricardo.neri-calderon@xxxxxxxxxxxxxxx> wrote:
>
> On Mon, Mar 27, 2023 at 06:50:13PM +0200, Rafael J. Wysocki wrote:
> > On Tue, Feb 7, 2023 at 6:02 AM Ricardo Neri
> > <ricardo.neri-calderon@xxxxxxxxxxxxxxx> wrote:
> > >
> > > Implement the arch_get_ipcc_score() interface of the scheduler. Use the
> > > performance capabilities of the extended Hardware Feedback Interface table
> > > as the IPC score.
> > >
> > > Cc: Ben Segall <bsegall@xxxxxxxxxx>
> > > Cc: Daniel Bristot de Oliveira <bristot@xxxxxxxxxx>
> > > Cc: Dietmar Eggemann <dietmar.eggemann@xxxxxxx>
> > > Cc: Ionela Voinescu <ionela.voinescu@xxxxxxx>
> > > Cc: Joel Fernandes (Google) <joel@xxxxxxxxxxxxxxxxx>
> > > Cc: Len Brown <len.brown@xxxxxxxxx>
> > > Cc: Lukasz Luba <lukasz.luba@xxxxxxx>
> > > Cc: Mel Gorman <mgorman@xxxxxxx>
> > > Cc: Rafael J. Wysocki <rafael.j.wysocki@xxxxxxxxx>
> > > Cc: Srinivas Pandruvada <srinivas.pandruvada@xxxxxxxxxxxxxxx>
> > > Cc: Steven Rostedt <rostedt@xxxxxxxxxxx>
> > > Cc: Tim C. Chen <tim.c.chen@xxxxxxxxx>
> > > Cc: Valentin Schneider <vschneid@xxxxxxxxxx>
> > > Cc: x86@xxxxxxxxxx
> > > Cc: linux-pm@xxxxxxxxxxxxxxx
> > > Cc: linux-kernel@xxxxxxxxxxxxxxx
> > > Signed-off-by: Ricardo Neri <ricardo.neri-calderon@xxxxxxxxxxxxxxx>
> > > ---
> > > Changes since v2:
> > >  * None
> > >
> > > Changes since v1:
> > >  * Adjusted the returned HFI class (which starts at 0) to match the
> > >    scheduler IPCC class (which starts at 1). (PeterZ)
> > >  * Used the new interface names.
> > > ---
> > >  arch/x86/include/asm/topology.h   |  2 ++
> > >  drivers/thermal/intel/intel_hfi.c | 27 +++++++++++++++++++++++++++
> > >  2 files changed, 29 insertions(+)
> > >
> > > diff --git a/arch/x86/include/asm/topology.h b/arch/x86/include/asm/topology.h
> > > index ffcdac3f398f..c4fcd9c3c634 100644
> > > --- a/arch/x86/include/asm/topology.h
> > > +++ b/arch/x86/include/asm/topology.h
> > > @@ -229,8 +229,10 @@ void init_freq_invariance_cppc(void);
> > >
> > >  #if defined(CONFIG_IPC_CLASSES) && defined(CONFIG_INTEL_HFI_THERMAL)
> > >  void intel_hfi_update_ipcc(struct task_struct *curr);
> > > +unsigned long intel_hfi_get_ipcc_score(unsigned short ipcc, int cpu);
> > >
> > >  #define arch_update_ipcc intel_hfi_update_ipcc
> > > +#define arch_get_ipcc_score intel_hfi_get_ipcc_score
> > >  #endif /* defined(CONFIG_IPC_CLASSES) && defined(CONFIG_INTEL_HFI_THERMAL) */
> > >
> > >  #endif /* _ASM_X86_TOPOLOGY_H */
> > > diff --git a/drivers/thermal/intel/intel_hfi.c b/drivers/thermal/intel/intel_hfi.c
> > > index 530dcf57e06e..fa9b4a678d92 100644
> > > --- a/drivers/thermal/intel/intel_hfi.c
> > > +++ b/drivers/thermal/intel/intel_hfi.c
> > > @@ -206,6 +206,33 @@ void intel_hfi_update_ipcc(struct task_struct *curr)
> > >         curr->ipcc = msr.split.classid + 1;
> > >  }
> > >
> > > +unsigned long intel_hfi_get_ipcc_score(unsigned short ipcc, int cpu)
> > > +{
> > > +       unsigned short hfi_class;
> >
> > It looks like the variable above is only used to save a subtraction or
> > addition of 1 to something going forward.
> >
> > > +       int *scores;
> > > +
> > > +       if (cpu < 0 || cpu >= nr_cpu_ids)
> > > +               return -EINVAL;
> > > +
> > > +       if (ipcc == IPC_CLASS_UNCLASSIFIED)
> > > +               return -EINVAL;
> > > +
> > > +       /*
> > > +        * Scheduler IPC classes start at 1. HFI classes start at 0.
> > > +        * See note intel_hfi_update_ipcc().
> > > +        */
> > > +       hfi_class = ipcc - 1;
> > > +
> > > +       if (hfi_class >= hfi_features.nr_classes)
> >
> > Personally, I would do
> >
> > if (ipcc >= hfi_features.nr_classes + 1)
> >
> > here and ->
> >
> > > +               return -EINVAL;
> > > +
> > > +       scores = per_cpu_ptr(hfi_ipcc_scores, cpu);
> > > +       if (!scores)
> > > +               return -ENODEV;
> > > +
> >
> > -> scores[ipcc - 1]
>
> Sure, I can get rid of hfi_class.
>
> >
> > below.
> >
> > > +       return READ_ONCE(scores[hfi_class]);
> >
> > And why does this need to use READ_ONCE()?
>
> This is the corresponding read of the WRITE_ONCE in patch 13. The CPU
> handling the HFI interrupt, very likely a different CPU, updates
> scores[hfi_class]. My intention is to let that write to complete before
> reading the score here.

However, READ_ONCE()/WRITE_ONCE() only affect compiler optimizations
AFAICS.  What if the CPUs running the code reorder the instructions?

In any case, IMV the reason why READ_ONCE() is used needs to be clear
to the reviewers from the patch itself (and to a casual reader of the
code from the code itself).

> >
> > > +}
> > > +
> > >  static int alloc_hfi_ipcc_scores(void)
> > >  {
> > >         if (!cpu_feature_enabled(X86_FEATURE_ITD))
> > > --