Message-ID: <364029b8-3a73-60a3-b875-0d7323dcdd27@os.amperecomputing.com>
Date:   Wed, 24 May 2023 11:08:47 -0700
From:   Yang Shi <yang@...amperecomputing.com>
To:     Sumit Gupta <sumitg@...dia.com>,
        Pierre Gondois <pierre.gondois@....com>,
        Zeng Heng <zengheng4@...wei.com>,
        Ionela Voinescu <Ionela.Voinescu@....com>
Cc:     linux-kernel@...r.kernel.org, linux-pm@...r.kernel.org,
        wangxiongfeng2@...wei.com, xiexiuqi@...wei.com,
        liwei391@...wei.com, linux-acpi@...r.kernel.org,
        weiyongjun1@...wei.com, lenb@...nel.org, viresh.kumar@...aro.org,
        rafael@...nel.org
Subject: Re: [PATCH v2 1/2] cpufreq: CPPC: keep target core awake when reading
 its cpufreq rate



On 5/18/23 7:40 AM, Sumit Gupta wrote:
>>>>
>>>> +Ionela, Sumit, Yang,
>>>>
>>>> Hello Zeng,
>>>>
>>>> I think solutions around related issues were suggested at:
>>>>
>>>> [1]
>>>> https://lore.kernel.org/all/20230418113459.12860-7-sumitg@nvidia.com/
>>>> [2]
>>>> https://lore.kernel.org/all/20230328193846.8757-1-yang@os.amperecomputing.com/ 
>>>>
>>>> [3] https://lore.kernel.org/all/ZEl1Fms%2FJmdEZsVn@arm.com/
>>>>
>>>> About this patch, it seems to mean that CPPC counters of CPUx are 
>>>> always
>>>> accessed from CPUx, even when they are not AMUs. For instance CPPC
>>>> counters could be memory mapped and accessible from any CPU.
>>>> cpu_has_amu_feat() should allow to probe if a CPU uses AMUs or not,
>>>> and [2] had an implementation using it.
>>>>
>>>> Another comment about PATCH 2/2 is that if the counters are accessed
>>>> through FFH, arm64 version of cpc_read_ffh() is calling
>>>> counters_read_on_cpu(), and a comment in counters_read_on_cpu() seems
>>>> to specify the function must be called with interrupt enabled.
>>>>
>>>> I think the best solution so far was the one at [3], suggested by
>>>> Ionela,
>>>> but it doesn't seem to solve your issue. Indeed, it is not checked
>>>> whether
>>>> the counters are AMU counters and that they must be remotely read (to
>>>> have the CPU awake),
>>>>
>>>> Regards,
>>>> Pierre
>>>>
>>>
>>> I think the solution in [1] is simple and solves all the three cases.
>>> Also, it provides better accuracy between the set and get frequency as
>>> compared to [3].
>>
>> I don't think [1] patches work for our case. We use mmio instead of AMU.
>> Increasing delay could help to mitigate it somehow, buyt 25us is not
>> good enough for our case. IIRC the fix proposed by Ionela works for both
>> yours and mine.
>>
>
> I have added the CPC_IN_SYSTEM_MEMORY check from [2] in [1].
> Could you please test if the below change works for you.

Thanks, Sumit. Sorry for the late reply. I didn't find time to test your 
patch, but the visual inspection shows it should work for my case. The 
deviation in our case is mainly caused by:
   - interrupt
   - congestion on interconnect

So disabling interrupt when reading the counters + longer delay (e.g. 25 
us) could mitigate it significantly. So as long as the two are covered 
by the patch, it should work for us and your patch does.

>
> -----------------------------------------
> diff --git a/drivers/acpi/cppc_acpi.c b/drivers/acpi/cppc_acpi.c
> index 7ff269a78c20..67aa09b5f15c 100644
> --- a/drivers/acpi/cppc_acpi.c
> +++ b/drivers/acpi/cppc_acpi.c
> @@ -1315,6 +1315,7 @@ int cppc_get_perf_ctrs(int cpunum, struct 
> cppc_perf_fb_ctrs *perf_fb_ctrs)
>      struct cppc_pcc_data *pcc_ss_data = NULL;
>      u64 delivered, reference, ref_perf, ctr_wrap_time;
>      int ret = 0, regs_in_pcc = 0;
> +    unsigned long flags;
>
>      if (!cpc_desc) {
>          pr_debug("No CPC descriptor for CPU:%d\n", cpunum);
> @@ -1350,8 +1351,17 @@ int cppc_get_perf_ctrs(int cpunum, struct 
> cppc_perf_fb_ctrs *perf_fb_ctrs)
>          }
>      }
>
> +    if (CPC_IN_SYSTEM_MEMORY(delivered_reg) &&
> +        CPC_IN_SYSTEM_MEMORY(reference_reg))
> +        local_irq_save(flags);
> +
>      cpc_read(cpunum, delivered_reg, &delivered);
>      cpc_read(cpunum, reference_reg, &reference);
> +
> +    if (CPC_IN_SYSTEM_MEMORY(delivered_reg) &&
> +        CPC_IN_SYSTEM_MEMORY(reference_reg))
> +        local_irq_restore(flags);
> +
>      cpc_read(cpunum, ref_perf_reg, &ref_perf);
>
>      /*
> diff --git a/drivers/cpufreq/cppc_cpufreq.c 
> b/drivers/cpufreq/cppc_cpufreq.c
> index 5e6a132a525e..23e690854459 100644
> --- a/drivers/cpufreq/cppc_cpufreq.c
> +++ b/drivers/cpufreq/cppc_cpufreq.c
> @@ -46,6 +46,8 @@ static bool boost_supported;
>  /* default 2usec delay between sampling */
>  static unsigned int sampling_delay_us = 2;
>
> +static bool get_rate_use_wq;
> +
>  static void cppc_check_hisi_workaround(void);
>  static void cppc_nvidia_workaround(void);
>
> @@ -99,6 +101,12 @@ struct cppc_freq_invariance {
>  static DEFINE_PER_CPU(struct cppc_freq_invariance, cppc_freq_inv);
>  static struct kthread_worker *kworker_fie;
>
> +struct feedback_ctrs {
> +    u32 cpu;
> +    struct cppc_perf_fb_ctrs fb_ctrs_t0;
> +    struct cppc_perf_fb_ctrs fb_ctrs_t1;
> +};
> +
>  static unsigned int hisi_cppc_cpufreq_get_rate(unsigned int cpu);
>  static int cppc_perf_from_fbctrs(struct cppc_cpudata *cpu_data,
>                   struct cppc_perf_fb_ctrs *fb_ctrs_t0,
> @@ -851,28 +859,44 @@ static int cppc_perf_from_fbctrs(struct 
> cppc_cpudata *cpu_data,
>      return (reference_perf * delta_delivered) / delta_reference;
>  }
>
> +static int cppc_get_perf_ctrs_sync(void *fb_ctrs)
> +{
> +    struct feedback_ctrs *ctrs = fb_ctrs;
> +    int ret;
> +
> +    ret = cppc_get_perf_ctrs(ctrs->cpu, &(ctrs->fb_ctrs_t0));
> +    if (ret)
> +        return ret;
> +
> +    udelay(sampling_delay_us);
> +
> +    ret = cppc_get_perf_ctrs(ctrs->cpu, &(ctrs->fb_ctrs_t1));
> +    if (ret)
> +        return ret;
> +
> +    return ret;
> +}
> +
>  static unsigned int cppc_cpufreq_get_rate(unsigned int cpu)
>  {
> -    struct cppc_perf_fb_ctrs fb_ctrs_t0 = {0}, fb_ctrs_t1 = {0};
>      struct cpufreq_policy *policy = cpufreq_cpu_get(cpu);
>      struct cppc_cpudata *cpu_data = policy->driver_data;
> +    struct feedback_ctrs fb_ctrs = {0};
>      u64 delivered_perf;
>      int ret;
>
>      cpufreq_cpu_put(policy);
> +    fb_ctrs.cpu = cpu;
>
> -    ret = cppc_get_perf_ctrs(cpu, &fb_ctrs_t0);
> -    if (ret)
> -        return ret;
> -
> -    udelay(sampling_delay_us);
> -
> -    ret = cppc_get_perf_ctrs(cpu, &fb_ctrs_t1);
> +    if (get_rate_use_wq)
> +        ret = smp_call_on_cpu(cpu, cppc_get_perf_ctrs_sync, &fb_ctrs, 
> false);
> +    else
> +        ret = cppc_get_perf_ctrs_sync(&fb_ctrs);
>      if (ret)
>          return ret;
>
> -    delivered_perf = cppc_perf_from_fbctrs(cpu_data, &fb_ctrs_t0,
> -                           &fb_ctrs_t1);
> +    delivered_perf = cppc_perf_from_fbctrs(cpu_data, 
> &(fb_ctrs.fb_ctrs_t0),
> +                           &(fb_ctrs.fb_ctrs_t1));
>
>      return cppc_cpufreq_perf_to_khz(cpu_data, delivered_perf);
>  }
> @@ -1002,10 +1026,18 @@ static void cppc_apply_workarounds(void)
>  static int __init cppc_cpufreq_init(void)
>  {
>      int ret;
> +    int cpu;
>
>      if (!acpi_cpc_valid())
>          return -ENODEV;
>
> +#ifdef CONFIG_ARM64_AMU_EXTN
> +    cpu = get_cpu_with_amu_feat();
> +
> +    if (cpu < nr_cpu_ids)
> +        get_rate_use_wq = true;
> +#endif
>
> --------------------------------------------
>
> We can add additional check to call smp_call_on_cpu() only when 
> CPC_IN_FFH if we want to reduce the scope of calling smp_call_on_cpu.
>
> diff --git a/drivers/acpi/cppc_acpi.c b/drivers/acpi/cppc_acpi.c
> index 67aa09b5f15c..3d8348911403 100644
> --- a/drivers/acpi/cppc_acpi.c
> +++ b/drivers/acpi/cppc_acpi.c
> @@ -110,6 +110,11 @@ static DEFINE_PER_CPU(struct cpc_desc *, 
> cpc_desc_ptr);
>                                 (cpc)->cpc_entry.reg.space_id ==        \
>                                 ACPI_ADR_SPACE_SYSTEM_IO)
>
> +/* Check if a CPC register is in FFH */
> +#define CPC_IN_FFH(cpc) ((cpc)->type == ACPI_TYPE_BUFFER &&    \
> +                               (cpc)->cpc_entry.reg.space_id ==        \
> +                               ACPI_ADR_SPACE_FIXED_HARDWARE)
> +
>  /* Evaluates to True if reg is a NULL register descriptor */
>  #define IS_NULL_REG(reg) ((reg)->space_id == 
> ACPI_ADR_SPACE_SYSTEM_MEMORY && \
>                                 (reg)->address == 0 &&                  \
> @@ -437,6 +442,29 @@ bool acpi_cpc_valid(void)
>  }
>  EXPORT_SYMBOL_GPL(acpi_cpc_valid);
>
> +bool acpi_cpc_in_ffh(void)
> +{
> +       struct cpc_register_resource *delivered_reg, *reference_reg;
> +       struct cpc_desc *cpc_ptr;
> +       int cpu;
> +
> +       if (acpi_disabled)
> +               return false;
> +
> +       for_each_possible_cpu(cpu) {
> +               cpc_ptr = per_cpu(cpc_desc_ptr, cpu);
> +               delivered_reg = &cpc_ptr->cpc_regs[DELIVERED_CTR];
> +               reference_reg = &cpc_ptr->cpc_regs[REFERENCE_CTR];
> +
> +               if (!CPC_IN_FFH(delivered_reg) ||
> +                   !CPC_IN_FFH(reference_reg))
> +                       return false;
> +       }
> +
> +       return true;
> +}
> +EXPORT_SYMBOL_GPL(acpi_cpc_in_ffh);
> +
>  bool cppc_allow_fast_switch(void)
>  {
>         struct cpc_register_resource *desired_reg;
> diff --git a/drivers/cpufreq/cppc_cpufreq.c 
> b/drivers/cpufreq/cppc_cpufreq.c
> index 23e690854459..4109e00b957e 100644
> --- a/drivers/cpufreq/cppc_cpufreq.c
> +++ b/drivers/cpufreq/cppc_cpufreq.c
> @@ -1034,7 +1034,7 @@ static int __init cppc_cpufreq_init(void)
>  #ifdef CONFIG_ARM64_AMU_EXTN
>         cpu = get_cpu_with_amu_feat();
>
> -       if (cpu < nr_cpu_ids)
> +       if ((cpu < nr_cpu_ids) && acpi_cpc_in_ffh())
>                 get_rate_use_wq = true;
>  #endif
>
>>>
>>> This can be merged and can later still be improved in Upstream.
>>>
>>> If OK, I can send new version by changing the patch to apply for all
>>> ARM SoC's with AMU and not specific to Tegra.
>>>
>>> Thank you,
>>> Sumit Gupta

Powered by blists - more mailing lists