Re: [PATCH] cpufreq: intel_pstate: Use the latest guaranteed freq during verify

From: Srinivas Pandruvada
Date: Thu Dec 17 2020 - 10:23:12 EST


On Thu, 2020-12-17 at 16:12 +0100, Rafael J. Wysocki wrote:
> On Thursday, December 17, 2020 3:23:44 PM CET Srinivas Pandruvada
> wrote:
> > On Thu, 2020-12-17 at 06:19 -0800, Srinivas Pandruvada wrote:
> > > On Thu, 2020-12-17 at 14:58 +0100, Rafael J. Wysocki wrote:
> > > > On Thu, Dec 17, 2020 at 11:44 AM Srinivas Pandruvada
> > > > <srinivas.pandruvada@xxxxxxxxxxxxxxx> wrote:
> > > > >
> > > > > This change tries to address an issue, when BIOS disabled
> > > > > turbo
> > > > > but HWP_CAP guaranteed is changed later and user space wants
> > > > > to
> > > > > take
> > > > > advantage of this increased guaranteed performance.
> > > > >
> > > > > The HWP_CAP.GUARANTEED value is not a static value. It can be
> > > > > changed
> > > > > by some out of band agent or during Intel Speed Select
> > > > > performance
> > > > > level change. The HWP_CAP.MAX still shows max possible
> > > > > performance
> > > > > when
> > > > > BIOS disabled turbo. So guaranteed can still change as long
> > > > > as
> > > > > this
> > > > > is
> > > > > same or below HWP_CAP.MAX.
> > > > >
> > > > > When guaranteed is changed, the sysfs base_frequency
> > > > > attributes
> > > > > shows
> > > > > the latest guaranteed frequency. This attribute can be used
> > > > > by
> > > > > user
> > > > > space software to update scaling min/max frequency.
> > > > >
> > > > > Currently the setpolicy callback already uses the latest
> > > > > HWP_CAP
> > > > > values when setting HWP_REQ. But the verify callback will
> > > > > still
> > > > > restrict
> > > > > the user settings to the to old guaranteed value. So if the
> > > > > guaranteed
> > > > > is increased, user space can't take advantage of it.
> > > > >
> > > > > To solve this similar to setpolicy callback, read the latest
> > > > > HWP_CAP
> > > > > values and use it to restrict the maximum setting. This is
> > > > > done
> > > > > by
> > > > > calling intel_pstate_get_hwp_max(), which already accounts
> > > > > for
> > > > > user
> > > > > and BIOS turbo disable to get the current max performance.
> > > > >
> > > > > This issue is side effect of fixing the issue of scaling
> > > > > frequency
> > > > > limits by the
> > > > >  'commit eacc9c5a927e ("cpufreq: intel_pstate:
> > > > >  Fix intel_pstate_get_hwp_max() for turbo disabled")'
> > > > > The fix resulted in correct setting of reduced scaling
> > > > > frequencies,
> > > > > but this resulted in capping HWP.REQ to HWP_CAP.GUARANTEED in
> > > > > this
> > > > > case.
> > > > >
> > > > > Cc: 5.8+ <stable@xxxxxxxxxxxxxxx> # 5.8+
> > > > > Signed-off-by: Srinivas Pandruvada <
> > > > > srinivas.pandruvada@xxxxxxxxxxxxxxx>
> > > > > ---
> > > > >  drivers/cpufreq/intel_pstate.c | 6 ++++++
> > > > >  1 file changed, 6 insertions(+)
> > > > >
> > > > > diff --git a/drivers/cpufreq/intel_pstate.c
> > > > > b/drivers/cpufreq/intel_pstate.c
> > > > > index 2a4db856222f..7081d1edb22b 100644
> > > > > --- a/drivers/cpufreq/intel_pstate.c
> > > > > +++ b/drivers/cpufreq/intel_pstate.c
> > > > > @@ -2199,6 +2199,12 @@ static void
> > > > > intel_pstate_clear_update_util_hook(unsigned int cpu)
> > > > >
> > > > >  static int intel_pstate_get_max_freq(struct cpudata *cpu)
> > > > >  {
> > > > > +       if (hwp_active) {
> > > > > +               int turbo_max, max_state;
> > > > > +
> > > > > +               intel_pstate_get_hwp_max(cpu->cpu,
> > > > > &turbo_max,
> > > > > &max_state);
> > > >
> > > > This would cause intel_pstate_get_hwp_max() to be called twice
> > > > in
> > > > intel_pstate_update_perf_limits() which is not perfect.
> > >
> > > We can optimize by using cached value.
> > >
> > >
> > > diff --git a/drivers/cpufreq/intel_pstate.c
> > > b/drivers/cpufreq/intel_pstate.c
> > > index 7081d1edb22b..d345c9ef240c 100644
> > > --- a/drivers/cpufreq/intel_pstate.c
> > > +++ b/drivers/cpufreq/intel_pstate.c
> > > @@ -2223,7 +2223,11 @@ static void
> > > intel_pstate_update_perf_limits(struct cpudata *cpu,
> > >          * rather than pure ratios.
> > >          */
> > >         if (hwp_active) {
> > > -               intel_pstate_get_hwp_max(cpu->cpu, &turbo_max,
> > > &max_state);
> > > +               if (global.no_turbo || global.turbo_disabled)
> > > +                       max_state = HWP_GUARANTEED_PERF(cpu-
> > > > hwp_cap_cached);
> > > +               else
> > > +                       max_state = HWP_HIGHEST_PERF(cpu-
> > > > hwp_cap_cached);
> > Can use  ternary operator instead of if..else. to further simplify.
> >
> > > +               turbo_max = HWP_HIGHEST_PERF(cpu->hwp_cached);
> > >         } else {
> > >                 max_state = global.no_turbo ||
> > > global.turbo_disabled
> > > ?
> > >                         cpu->pstate.max_pstate : cpu-
> > > > pstate.turbo_pstate;
>
> Well, would something like the patch below work?
>
> ---
>  drivers/cpufreq/intel_pstate.c |   16 +++++++++++++---
>  1 file changed, 13 insertions(+), 3 deletions(-)
>
> Index: linux-pm/drivers/cpufreq/intel_pstate.c
> ===================================================================
> --- linux-pm.orig/drivers/cpufreq/intel_pstate.c
> +++ linux-pm/drivers/cpufreq/intel_pstate.c
> @@ -2207,9 +2207,9 @@ static void intel_pstate_update_perf_lim
>                                             unsigned int policy_min,
>                                             unsigned int policy_max)
>  {
> -       int max_freq = intel_pstate_get_max_freq(cpu);
>         int32_t max_policy_perf, min_policy_perf;
>         int max_state, turbo_max;
> +       int max_freq;
>  
>         /*
>          * HWP needs some special consideration, because on BDX the
> @@ -2223,6 +2223,7 @@ static void intel_pstate_update_perf_lim
>                         cpu->pstate.max_pstate : cpu-
> >pstate.turbo_pstate;
>                 turbo_max = cpu->pstate.turbo_pstate;
>         }
> +       max_freq = max_state * cpu->pstate.scaling;
>  
>         max_policy_perf = max_state * policy_max / max_freq;
>         if (policy_max == policy_min) {
> @@ -2325,9 +2326,18 @@ static void intel_pstate_adjust_policy_m
>  static void intel_pstate_verify_cpu_policy(struct cpudata *cpu,
>                                            struct cpufreq_policy_data
> *policy)
>  {
> +       int max_freq;
> +
>         update_turbo_state();
> -       cpufreq_verify_within_limits(policy, policy-
> >cpuinfo.min_freq,
> -                                    intel_pstate_get_max_freq(cpu));
> +       if (hwp_active) {
> +               int max_state, turbo_max;
> +
> +               intel_pstate_get_hwp_max(cpu->cpu, &turbo_max,
> &max_state);
> +               max_freq = max_state * cpu->pstate.scaling;
> +       } else {
> +               max_freq = intel_pstate_get_max_freq(cpu);
> +       }
> +       cpufreq_verify_within_limits(policy, policy-
> >cpuinfo.min_freq, max_freq);
>  
>         intel_pstate_adjust_policy_max(cpu, policy);
>  }
>
Should work.
I will test this patch and let you know once I get the system.

Thanks,
Srinivas

>
>