RE: [PATCH] [PATCH] amd_pstate: fix erroneous highest_perf value on some CPUs

[Meng, Li (Jassmine)]

[AMD Official Use Only - General]

Hi Lucas:

> -----Original Message-----
> From: Lucas Lee Jing Yi <lucasleeeeeeeee@xxxxxxxxx>
> Sent: Monday, February 19, 2024 12:11 AM
> To: rafael@xxxxxxxxxx
> Cc: Yuan, Perry <Perry.Yuan@xxxxxxx>; Du, Xiaojian
> <Xiaojian.Du@xxxxxxx>; Deucher, Alexander
> <Alexander.Deucher@xxxxxxx>; bp@xxxxxxxxx; Sharma, Deepak
> <Deepak.Sharma@xxxxxxx>; Meng, Li (Jassmine) <Li.Meng@xxxxxxx>;
> linux-acpi@xxxxxxxxxxxxxxx; linux-kernel@xxxxxxxxxxxxxxx; linux-
> kselftest@xxxxxxxxxxxxxxx; linux-pm@xxxxxxxxxxxxxxx; Limonciello, Mario
> <Mario.Limonciello@xxxxxxx>; Fontenot, Nathan
> <Nathan.Fontenot@xxxxxxx>; oleksandr@xxxxxxxxxxxxxx;
> rafael.j.wysocki@xxxxxxxxx; Huang, Ray <Ray.Huang@xxxxxxx>; Huang,
> Shimmer <Shimmer.Huang@xxxxxxx>; skhan@xxxxxxxxxxxxxxxxxxx;
> viresh.kumar@xxxxxxxxxx; x86@xxxxxxxxxx; Lucas Lee Jing Yi
> <lucasleeeeeeeee@xxxxxxxxx>
> Subject: [PATCH] [PATCH] amd_pstate: fix erroneous highest_perf value on
> some CPUs
>
> Caution: This message originated from an External Source. Use proper
> caution when opening attachments, clicking links, or responding.
>
>
> On a Ryzen 7840HS the highest_perf value is 196, not 166 as AMD assumed.
> This leads to the advertised max clock speed to only be 4.35ghz instead of
> 5.14ghz , leading to a large degradation in performance.
>
> Fix the broken assumption and revert back to the old logic for getting
> highest_perf.
>
> TEST:
> Geekbench 6 Before Patch:
> Single Core:    2325 (-22%)!
> Multi Core:     11335 (-10%)
>
> Geekbench 6 AFTER Patch:
> Single Core:    2635
> Multi Core:     12487
>
> Signed-off-by: Lucas Lee Jing Yi <lucasleeeeeeeee@xxxxxxxxx>
> ---
>  drivers/cpufreq/amd-pstate.c | 22 ++++++++++------------
>  1 file changed, 10 insertions(+), 12 deletions(-)
>
> diff --git a/drivers/cpufreq/amd-pstate.c b/drivers/cpufreq/amd-pstate.c
> index 08e112444c27..54df68773620 100644
> --- a/drivers/cpufreq/amd-pstate.c
> +++ b/drivers/cpufreq/amd-pstate.c
> @@ -50,7 +50,6 @@
>
>  #define AMD_PSTATE_TRANSITION_LATENCY  20000
>  #define AMD_PSTATE_TRANSITION_DELAY    1000
> -#define AMD_PSTATE_PREFCORE_THRESHOLD  166
>
>  /*
>   * TODO: We need more time to fine tune processors with shared memory
> solution @@ -299,15 +298,12 @@ static int pstate_init_perf(struct
> amd_cpudata *cpudata)
>                                      &cap1);
>         if (ret)
>                 return ret;
> -
> -       /* For platforms that do not support the preferred core feature, the
> -        * highest_pef may be configured with 166 or 255, to avoid max
> frequency
> -        * calculated wrongly. we take the AMD_CPPC_HIGHEST_PERF(cap1)
> value as
> -        * the default max perf.
> +
> +       /* Some CPUs have different highest_perf from others, it is safer
> +        * to read it than to assume some erroneous value, leading to
> performance issues.
>          */
> -       if (cpudata->hw_prefcore)
> -               highest_perf = AMD_PSTATE_PREFCORE_THRESHOLD;
> -       else
> +       highest_perf = amd_get_highest_perf();
> +       if(highest_perf > AMD_CPPC_HIGHEST_PERF(cap1))
>                 highest_perf = AMD_CPPC_HIGHEST_PERF(cap1);
>
>         WRITE_ONCE(cpudata->highest_perf, highest_perf); @@ -329,9 +325,11
> @@ static int cppc_init_perf(struct amd_cpudata *cpudata)
>         if (ret)
>                 return ret;
>
> -       if (cpudata->hw_prefcore)
> -               highest_perf = AMD_PSTATE_PREFCORE_THRESHOLD;
> -       else
> +       /* Some CPUs have different highest_perf from others, it is safer
> +        * to read it than to assume some erroneous value, leading to
> performance issues.
> +        */
> +       highest_perf = amd_get_highest_perf();
> +       if(highest_perf > cppc_perf.highest_perf)
>                 highest_perf = cppc_perf.highest_perf;
>
>         WRITE_ONCE(cpudata->highest_perf, highest_perf);
> --
> 2.43.2
[Meng, Li (Jassmine)]
Reviewed-by: Li Meng < li.meng@xxxxxxx>