Re: [PATCH 3/3] ACPI: cpufreq: Add ITMT support when CPPC enabled for Zhaoxin CPUs

[Rafael J. Wysocki]

On Thu, Dec 28, 2023 at 8:57 AM Tony W Wang-oc <TonyWWang-oc@xxxxxxxxxxx> wrote:
>
> For Zhaoxin CPUs, the cores' highest frequencies may be different, which
> means that cores may run at different max frequencies,
>
> According to ACPI-spec6 chapter 8.4.7, the per-core highest frequency
> value can be obtained via cppc.
>
> The core with the higher frequency have better performance, which can be
> called as preferred core. And better performance can be achieved by
> making the scheduler to run tasks on these preferred cores.
>
> The cpufreq driver can use the highest frequency value as the prioriy of
> core to make the scheduler try to get better performace. More specifically,
> in the acpi-cpufreq driver use cppc_get_highest_perf() to get highest
> frequency value of each core, use sched_set_itmt_core_prio() to set
> highest frequency value as core priority, and use sched_set_itmt_support()
> provided by ITMT to tell the scheduler to favor on the preferred cores.
>
> Signed-off-by: Tony W Wang-oc <TonyWWang-oc@xxxxxxxxxxx>
> ---
>  drivers/cpufreq/acpi-cpufreq.c | 56 +++++++++++++++++++++++++++++++++-
>  1 file changed, 55 insertions(+), 1 deletion(-)
>
> diff --git a/drivers/cpufreq/acpi-cpufreq.c b/drivers/cpufreq/acpi-cpufreq.c
> index 37f1cdf46d29..f4c1ff9e4bb0 100644
> --- a/drivers/cpufreq/acpi-cpufreq.c
> +++ b/drivers/cpufreq/acpi-cpufreq.c
> @@ -663,8 +663,56 @@ static u64 get_max_boost_ratio(unsigned int cpu)
>
>         return div_u64(highest_perf << SCHED_CAPACITY_SHIFT, nominal_perf);
>  }
> +
> +/* The work item is needed to avoid CPU hotplug locking issues */
> +static void sched_itmt_work_fn(struct work_struct *work)
> +{
> +       sched_set_itmt_support();
> +}
> +
> +static DECLARE_WORK(sched_itmt_work, sched_itmt_work_fn);
> +
> +static void set_itmt_prio(int cpu)
> +{
> +       static bool cppc_highest_perf_diff;
> +       static struct cpumask core_prior_mask;
> +       u64 highest_perf;
> +       static u64 max_highest_perf = 0, min_highest_perf = U64_MAX;
> +       int ret;
> +
> +       ret = cppc_get_highest_perf(cpu, &highest_perf);
> +       if (ret)
> +               return;
> +
> +       sched_set_itmt_core_prio(highest_perf, cpu);
> +       cpumask_set_cpu(cpu, &core_prior_mask);
> +
> +       if (max_highest_perf <= min_highest_perf) {
> +               if (highest_perf > max_highest_perf)
> +                       max_highest_perf = highest_perf;
> +
> +               if (highest_perf < min_highest_perf)
> +                       min_highest_perf = highest_perf;
> +
> +               if (max_highest_perf > min_highest_perf) {
> +                       /*
> +                        * This code can be run during CPU online under the
> +                        * CPU hotplug locks, so sched_set_itmt_support()
> +                        * cannot be called from here.  Queue up a work item
> +                        * to invoke it.
> +                        */
> +                       cppc_highest_perf_diff = true;
> +               }
> +       }
> +
> +       if (cppc_highest_perf_diff && cpumask_equal(&core_prior_mask, cpu_online_mask)) {
> +               pr_debug("queue a work to set itmt enabled\n");
> +               schedule_work(&sched_itmt_work);
> +       }
> +}
>  #else
>  static inline u64 get_max_boost_ratio(unsigned int cpu) { return 0; }
> +static void set_itmt_prio(int cpu) { }
>  #endif
>
>  static int acpi_cpufreq_cpu_init(struct cpufreq_policy *policy)
> @@ -677,7 +725,7 @@ static int acpi_cpufreq_cpu_init(struct cpufreq_policy *policy)
>         unsigned int valid_states = 0;
>         unsigned int result = 0;
>         u64 max_boost_ratio;
> -       unsigned int i;
> +       unsigned int i, j;
>  #ifdef CONFIG_SMP
>         static int blacklisted;
>  #endif
> @@ -742,6 +790,12 @@ static int acpi_cpufreq_cpu_init(struct cpufreq_policy *policy)
>         }
>  #endif
>
> +       if (c->x86_vendor == X86_VENDOR_CENTAUR || c->x86_vendor == X86_VENDOR_ZHAOXIN) {
> +               for_each_cpu(j, policy->cpus) {
> +                       set_itmt_prio(j);
> +               }
> +       }
> +
>         /* capability check */
>         if (perf->state_count <= 1) {
>                 pr_debug("No P-States\n");
> --

Have you considered using the CPPC cpufreq driver on those platforms?