diff mbox series

cpufreq: CPPC: Resolve the large frequency discrepancy from cpuinfo_cur_freq

Message ID 20231212072617.14756-1-lihuisong@huawei.com (mailing list archive)
State New, archived
Headers show
Series cpufreq: CPPC: Resolve the large frequency discrepancy from cpuinfo_cur_freq | expand

Commit Message

lihuisong (C) Dec. 12, 2023, 7:26 a.m. UTC
Many developers found that the cpu current frequency is greater than
the maximum frequency of the platform, please see [1], [2] and [3].

In the scenarios with high memory access pressure, the patch [1] has
proved the significant latency of cpc_read() which is used to obtain
delivered and reference performance counter cause an absurd frequency.
The sampling interval for this counters is very critical and is expected
to be equal. However, the different latency of cpc_read() has a direct
impact on their sampling interval.

This patch adds a interface, cpc_read_arch_counters_on_cpu, to read
delivered and reference performance counter together. According to my
test[4], the discrepancy of cpu current frequency in the scenarios with
high memory access pressure is lower than 0.2% by stress-ng application.

[1] https://lore.kernel.org/all/20231025093847.3740104-4-zengheng4@huawei.com/
[2] https://lore.kernel.org/all/20230328193846.8757-1-yang@os.amperecomputing.com/
[3] https://lore.kernel.org/all/20230418113459.12860-7-sumitg@nvidia.com/

[4] My local test:
The testing platform enable SMT and include 128 logical CPU in total,
and CPU base frequency is 2.7GHz. Reading "cpuinfo_cur_freq" for each
physical core on platform during the high memory access pressure from
stress-ng, and the output is as follows:
  0: 2699133     2: 2699942     4: 2698189     6: 2704347
  8: 2704009    10: 2696277    12: 2702016    14: 2701388
 16: 2700358    18: 2696741    20: 2700091    22: 2700122
 24: 2701713    26: 2702025    28: 2699816    30: 2700121
 32: 2700000    34: 2699788    36: 2698884    38: 2699109
 40: 2704494    42: 2698350    44: 2699997    46: 2701023
 48: 2703448    50: 2699501    52: 2700000    54: 2699999
 56: 2702645    58: 2696923    60: 2697718    62: 2700547
 64: 2700313    66: 2700000    68: 2699904    70: 2699259
 72: 2699511    74: 2700644    76: 2702201    78: 2700000
 80: 2700776    82: 2700364    84: 2702674    86: 2700255
 88: 2699886    90: 2700359    92: 2699662    94: 2696188
 96: 2705454    98: 2699260   100: 2701097   102: 2699630
104: 2700463   106: 2698408   108: 2697766   110: 2701181
112: 2699166   114: 2701804   116: 2701907   118: 2701973
120: 2699584   122: 2700474   124: 2700768   126: 2701963

Signed-off-by: Huisong Li <lihuisong@huawei.com>
---
 arch/arm64/kernel/topology.c | 43 ++++++++++++++++++++++++++++++++++--
 drivers/acpi/cppc_acpi.c     | 22 +++++++++++++++---
 include/acpi/cppc_acpi.h     |  5 +++++
 3 files changed, 65 insertions(+), 5 deletions(-)

Comments

Rafael J. Wysocki Dec. 14, 2023, 7:31 p.m. UTC | #1
On Tue, Dec 12, 2023 at 8:26 AM Huisong Li <lihuisong@huawei.com> wrote:
>
> Many developers found that the cpu current frequency is greater than
> the maximum frequency of the platform, please see [1], [2] and [3].
>
> In the scenarios with high memory access pressure, the patch [1] has
> proved the significant latency of cpc_read() which is used to obtain
> delivered and reference performance counter cause an absurd frequency.
> The sampling interval for this counters is very critical and is expected
> to be equal. However, the different latency of cpc_read() has a direct
> impact on their sampling interval.
>
> This patch adds a interface, cpc_read_arch_counters_on_cpu, to read
> delivered and reference performance counter together. According to my
> test[4], the discrepancy of cpu current frequency in the scenarios with
> high memory access pressure is lower than 0.2% by stress-ng application.
>
> [1] https://lore.kernel.org/all/20231025093847.3740104-4-zengheng4@huawei.com/
> [2] https://lore.kernel.org/all/20230328193846.8757-1-yang@os.amperecomputing.com/
> [3] https://lore.kernel.org/all/20230418113459.12860-7-sumitg@nvidia.com/
>
> [4] My local test:
> The testing platform enable SMT and include 128 logical CPU in total,
> and CPU base frequency is 2.7GHz. Reading "cpuinfo_cur_freq" for each
> physical core on platform during the high memory access pressure from
> stress-ng, and the output is as follows:
>   0: 2699133     2: 2699942     4: 2698189     6: 2704347
>   8: 2704009    10: 2696277    12: 2702016    14: 2701388
>  16: 2700358    18: 2696741    20: 2700091    22: 2700122
>  24: 2701713    26: 2702025    28: 2699816    30: 2700121
>  32: 2700000    34: 2699788    36: 2698884    38: 2699109
>  40: 2704494    42: 2698350    44: 2699997    46: 2701023
>  48: 2703448    50: 2699501    52: 2700000    54: 2699999
>  56: 2702645    58: 2696923    60: 2697718    62: 2700547
>  64: 2700313    66: 2700000    68: 2699904    70: 2699259
>  72: 2699511    74: 2700644    76: 2702201    78: 2700000
>  80: 2700776    82: 2700364    84: 2702674    86: 2700255
>  88: 2699886    90: 2700359    92: 2699662    94: 2696188
>  96: 2705454    98: 2699260   100: 2701097   102: 2699630
> 104: 2700463   106: 2698408   108: 2697766   110: 2701181
> 112: 2699166   114: 2701804   116: 2701907   118: 2701973
> 120: 2699584   122: 2700474   124: 2700768   126: 2701963
>
> Signed-off-by: Huisong Li <lihuisong@huawei.com>

First off, please Cc ACPI-related patches to linux-acpi.

> ---
>  arch/arm64/kernel/topology.c | 43 ++++++++++++++++++++++++++++++++++--
>  drivers/acpi/cppc_acpi.c     | 22 +++++++++++++++---
>  include/acpi/cppc_acpi.h     |  5 +++++
>  3 files changed, 65 insertions(+), 5 deletions(-)
>
> diff --git a/arch/arm64/kernel/topology.c b/arch/arm64/kernel/topology.c
> index 7d37e458e2f5..c3122154d738 100644
> --- a/arch/arm64/kernel/topology.c
> +++ b/arch/arm64/kernel/topology.c
> @@ -299,6 +299,11 @@ core_initcall(init_amu_fie);
>  #ifdef CONFIG_ACPI_CPPC_LIB
>  #include <acpi/cppc_acpi.h>
>
> +struct amu_counters {
> +       u64 corecnt;
> +       u64 constcnt;
> +};
> +
>  static void cpu_read_corecnt(void *val)
>  {
>         /*
> @@ -322,8 +327,27 @@ static void cpu_read_constcnt(void *val)
>                       0UL : read_constcnt();
>  }
>
> +static void cpu_read_amu_counters(void *data)
> +{
> +       struct amu_counters *cnt = (struct amu_counters *)data;
> +
> +       /*
> +        * The running time of the this_cpu_has_cap() might have a couple of
> +        * microseconds and is significantly increased to tens of microseconds.
> +        * But AMU core and constant counter need to be read togeter without any
> +        * time interval to reduce the calculation discrepancy using this counters.
> +        */
> +       if (this_cpu_has_cap(ARM64_WORKAROUND_2457168)) {
> +               cnt->corecnt = read_corecnt();

This statement is present in both branches, so can it be moved before the if ()?

> +               cnt->constcnt = 0;
> +       } else {
> +               cnt->corecnt = read_corecnt();
> +               cnt->constcnt = read_constcnt();
> +       }
> +}
> +
>  static inline
> -int counters_read_on_cpu(int cpu, smp_call_func_t func, u64 *val)
> +int counters_read_on_cpu(int cpu, smp_call_func_t func, void *data)
>  {
>         /*
>          * Abort call on counterless CPU or when interrupts are
> @@ -335,7 +359,7 @@ int counters_read_on_cpu(int cpu, smp_call_func_t func, u64 *val)
>         if (WARN_ON_ONCE(irqs_disabled()))
>                 return -EPERM;
>
> -       smp_call_function_single(cpu, func, val, 1);
> +       smp_call_function_single(cpu, func, data, 1);
>
>         return 0;
>  }
> @@ -364,6 +388,21 @@ bool cpc_ffh_supported(void)
>         return true;
>  }
>
> +int cpc_read_arch_counters_on_cpu(int cpu, u64 *delivered, u64 *reference)
> +{
> +       struct amu_counters cnts = {0};
> +       int ret;
> +
> +       ret = counters_read_on_cpu(cpu, cpu_read_amu_counters, &cnts);
> +       if (ret)
> +               return ret;
> +
> +       *delivered = cnts.corecnt;
> +       *reference = cnts.constcnt;
> +
> +       return 0;
> +}
> +
>  int cpc_read_ffh(int cpu, struct cpc_reg *reg, u64 *val)
>  {
>         int ret = -EOPNOTSUPP;
> diff --git a/drivers/acpi/cppc_acpi.c b/drivers/acpi/cppc_acpi.c
> index 7ff269a78c20..f303fabd7cfe 100644
> --- a/drivers/acpi/cppc_acpi.c
> +++ b/drivers/acpi/cppc_acpi.c
> @@ -1299,6 +1299,11 @@ bool cppc_perf_ctrs_in_pcc(void)
>  }
>  EXPORT_SYMBOL_GPL(cppc_perf_ctrs_in_pcc);
>
> +int __weak cpc_read_arch_counters_on_cpu(int cpu, u64 *delivered, u64 *reference)
> +{
> +       return 0;
> +}
> +
>  /**
>   * cppc_get_perf_ctrs - Read a CPU's performance feedback counters.
>   * @cpunum: CPU from which to read counters.
> @@ -1313,7 +1318,8 @@ int cppc_get_perf_ctrs(int cpunum, struct cppc_perf_fb_ctrs *perf_fb_ctrs)
>                 *ref_perf_reg, *ctr_wrap_reg;
>         int pcc_ss_id = per_cpu(cpu_pcc_subspace_idx, cpunum);
>         struct cppc_pcc_data *pcc_ss_data = NULL;
> -       u64 delivered, reference, ref_perf, ctr_wrap_time;
> +       u64 delivered = 0, reference = 0;
> +       u64 ref_perf, ctr_wrap_time;
>         int ret = 0, regs_in_pcc = 0;
>
>         if (!cpc_desc) {
> @@ -1350,8 +1356,18 @@ int cppc_get_perf_ctrs(int cpunum, struct cppc_perf_fb_ctrs *perf_fb_ctrs)
>                 }
>         }
>
> -       cpc_read(cpunum, delivered_reg, &delivered);
> -       cpc_read(cpunum, reference_reg, &reference);
> +       if (cpc_ffh_supported()) {
> +               ret = cpc_read_arch_counters_on_cpu(cpunum, &delivered, &reference);
> +               if (ret) {
> +                       pr_debug("read arch counters failed, ret=%d.\n", ret);
> +                       ret = 0;
> +               }
> +       }

The above is surely not applicable to every platform using CPPC.  Also
it looks like in the ARM64_WORKAROUND_2457168 enabled case it is just
pointless overhead, because "reference" is always going to be 0 here
then.

Please clean that up.

> +       if (!delivered || !reference) {
> +               cpc_read(cpunum, delivered_reg, &delivered);
> +               cpc_read(cpunum, reference_reg, &reference);
> +       }
> +
>         cpc_read(cpunum, ref_perf_reg, &ref_perf);
>
>         /*
> diff --git a/include/acpi/cppc_acpi.h b/include/acpi/cppc_acpi.h
> index 6126c977ece0..07d4fd82d499 100644
> --- a/include/acpi/cppc_acpi.h
> +++ b/include/acpi/cppc_acpi.h
> @@ -152,6 +152,7 @@ extern bool cpc_ffh_supported(void);
>  extern bool cpc_supported_by_cpu(void);
>  extern int cpc_read_ffh(int cpunum, struct cpc_reg *reg, u64 *val);
>  extern int cpc_write_ffh(int cpunum, struct cpc_reg *reg, u64 val);
> +extern int cpc_read_arch_counters_on_cpu(int cpu, u64 *delivered, u64 *reference);
>  extern int cppc_get_epp_perf(int cpunum, u64 *epp_perf);
>  extern int cppc_set_epp_perf(int cpu, struct cppc_perf_ctrls *perf_ctrls, bool enable);
>  extern int cppc_get_auto_sel_caps(int cpunum, struct cppc_perf_caps *perf_caps);
> @@ -209,6 +210,10 @@ static inline int cpc_write_ffh(int cpunum, struct cpc_reg *reg, u64 val)
>  {
>         return -ENOTSUPP;
>  }
> +static inline int cpc_read_arch_counters_on_cpu(int cpu, u64 *delivered, u64 *reference)
> +{
> +       return -EOPNOTSUPP;
> +}
>  static inline int cppc_set_epp_perf(int cpu, struct cppc_perf_ctrls *perf_ctrls, bool enable)
>  {
>         return -ENOTSUPP;
> --
lihuisong (C) Dec. 15, 2023, 2:41 a.m. UTC | #2
Hi Rafael,

Thanks for your review.
lihuisong (C) Dec. 18, 2023, 2:15 a.m. UTC | #3
在 2023/12/15 10:41, lihuisong (C) 写道:
> Hi Rafael,
>
> Thanks for your review.
lihuisong (C) Jan. 3, 2024, 3:59 a.m. UTC | #4
kindly ping for review.

在 2023/12/12 15:26, Huisong Li 写道:
> Many developers found that the cpu current frequency is greater than
> the maximum frequency of the platform, please see [1], [2] and [3].
>
> In the scenarios with high memory access pressure, the patch [1] has
> proved the significant latency of cpc_read() which is used to obtain
> delivered and reference performance counter cause an absurd frequency.
> The sampling interval for this counters is very critical and is expected
> to be equal. However, the different latency of cpc_read() has a direct
> impact on their sampling interval.
>
> This patch adds a interface, cpc_read_arch_counters_on_cpu, to read
> delivered and reference performance counter together. According to my
> test[4], the discrepancy of cpu current frequency in the scenarios with
> high memory access pressure is lower than 0.2% by stress-ng application.
>
> [1] https://lore.kernel.org/all/20231025093847.3740104-4-zengheng4@huawei.com/
> [2] https://lore.kernel.org/all/20230328193846.8757-1-yang@os.amperecomputing.com/
> [3] https://lore.kernel.org/all/20230418113459.12860-7-sumitg@nvidia.com/
>
> [4] My local test:
> The testing platform enable SMT and include 128 logical CPU in total,
> and CPU base frequency is 2.7GHz. Reading "cpuinfo_cur_freq" for each
> physical core on platform during the high memory access pressure from
> stress-ng, and the output is as follows:
>    0: 2699133     2: 2699942     4: 2698189     6: 2704347
>    8: 2704009    10: 2696277    12: 2702016    14: 2701388
>   16: 2700358    18: 2696741    20: 2700091    22: 2700122
>   24: 2701713    26: 2702025    28: 2699816    30: 2700121
>   32: 2700000    34: 2699788    36: 2698884    38: 2699109
>   40: 2704494    42: 2698350    44: 2699997    46: 2701023
>   48: 2703448    50: 2699501    52: 2700000    54: 2699999
>   56: 2702645    58: 2696923    60: 2697718    62: 2700547
>   64: 2700313    66: 2700000    68: 2699904    70: 2699259
>   72: 2699511    74: 2700644    76: 2702201    78: 2700000
>   80: 2700776    82: 2700364    84: 2702674    86: 2700255
>   88: 2699886    90: 2700359    92: 2699662    94: 2696188
>   96: 2705454    98: 2699260   100: 2701097   102: 2699630
> 104: 2700463   106: 2698408   108: 2697766   110: 2701181
> 112: 2699166   114: 2701804   116: 2701907   118: 2701973
> 120: 2699584   122: 2700474   124: 2700768   126: 2701963
>
> Signed-off-by: Huisong Li <lihuisong@huawei.com>
> ---
>   arch/arm64/kernel/topology.c | 43 ++++++++++++++++++++++++++++++++++--
>   drivers/acpi/cppc_acpi.c     | 22 +++++++++++++++---
>   include/acpi/cppc_acpi.h     |  5 +++++
>   3 files changed, 65 insertions(+), 5 deletions(-)
>
> diff --git a/arch/arm64/kernel/topology.c b/arch/arm64/kernel/topology.c
> index 7d37e458e2f5..c3122154d738 100644
> --- a/arch/arm64/kernel/topology.c
> +++ b/arch/arm64/kernel/topology.c
> @@ -299,6 +299,11 @@ core_initcall(init_amu_fie);
>   #ifdef CONFIG_ACPI_CPPC_LIB
>   #include <acpi/cppc_acpi.h>
>   
> +struct amu_counters {
> +	u64 corecnt;
> +	u64 constcnt;
> +};
> +
>   static void cpu_read_corecnt(void *val)
>   {
>   	/*
> @@ -322,8 +327,27 @@ static void cpu_read_constcnt(void *val)
>   		      0UL : read_constcnt();
>   }
>   
> +static void cpu_read_amu_counters(void *data)
> +{
> +	struct amu_counters *cnt = (struct amu_counters *)data;
> +
> +	/*
> +	 * The running time of the this_cpu_has_cap() might have a couple of
> +	 * microseconds and is significantly increased to tens of microseconds.
> +	 * But AMU core and constant counter need to be read togeter without any
> +	 * time interval to reduce the calculation discrepancy using this counters.
> +	 */
> +	if (this_cpu_has_cap(ARM64_WORKAROUND_2457168)) {
> +		cnt->corecnt = read_corecnt();
> +		cnt->constcnt = 0;
> +	} else {
> +		cnt->corecnt = read_corecnt();
> +		cnt->constcnt = read_constcnt();
> +	}
> +}
> +
>   static inline
> -int counters_read_on_cpu(int cpu, smp_call_func_t func, u64 *val)
> +int counters_read_on_cpu(int cpu, smp_call_func_t func, void *data)
>   {
>   	/*
>   	 * Abort call on counterless CPU or when interrupts are
> @@ -335,7 +359,7 @@ int counters_read_on_cpu(int cpu, smp_call_func_t func, u64 *val)
>   	if (WARN_ON_ONCE(irqs_disabled()))
>   		return -EPERM;
>   
> -	smp_call_function_single(cpu, func, val, 1);
> +	smp_call_function_single(cpu, func, data, 1);
>   
>   	return 0;
>   }
> @@ -364,6 +388,21 @@ bool cpc_ffh_supported(void)
>   	return true;
>   }
>   
> +int cpc_read_arch_counters_on_cpu(int cpu, u64 *delivered, u64 *reference)
> +{
> +	struct amu_counters cnts = {0};
> +	int ret;
> +
> +	ret = counters_read_on_cpu(cpu, cpu_read_amu_counters, &cnts);
> +	if (ret)
> +		return ret;
> +
> +	*delivered = cnts.corecnt;
> +	*reference = cnts.constcnt;
> +
> +	return 0;
> +}
> +
>   int cpc_read_ffh(int cpu, struct cpc_reg *reg, u64 *val)
>   {
>   	int ret = -EOPNOTSUPP;
> diff --git a/drivers/acpi/cppc_acpi.c b/drivers/acpi/cppc_acpi.c
> index 7ff269a78c20..f303fabd7cfe 100644
> --- a/drivers/acpi/cppc_acpi.c
> +++ b/drivers/acpi/cppc_acpi.c
> @@ -1299,6 +1299,11 @@ bool cppc_perf_ctrs_in_pcc(void)
>   }
>   EXPORT_SYMBOL_GPL(cppc_perf_ctrs_in_pcc);
>   
> +int __weak cpc_read_arch_counters_on_cpu(int cpu, u64 *delivered, u64 *reference)
> +{
> +	return 0;
> +}
> +
>   /**
>    * cppc_get_perf_ctrs - Read a CPU's performance feedback counters.
>    * @cpunum: CPU from which to read counters.
> @@ -1313,7 +1318,8 @@ int cppc_get_perf_ctrs(int cpunum, struct cppc_perf_fb_ctrs *perf_fb_ctrs)
>   		*ref_perf_reg, *ctr_wrap_reg;
>   	int pcc_ss_id = per_cpu(cpu_pcc_subspace_idx, cpunum);
>   	struct cppc_pcc_data *pcc_ss_data = NULL;
> -	u64 delivered, reference, ref_perf, ctr_wrap_time;
> +	u64 delivered = 0, reference = 0;
> +	u64 ref_perf, ctr_wrap_time;
>   	int ret = 0, regs_in_pcc = 0;
>   
>   	if (!cpc_desc) {
> @@ -1350,8 +1356,18 @@ int cppc_get_perf_ctrs(int cpunum, struct cppc_perf_fb_ctrs *perf_fb_ctrs)
>   		}
>   	}
>   
> -	cpc_read(cpunum, delivered_reg, &delivered);
> -	cpc_read(cpunum, reference_reg, &reference);
> +	if (cpc_ffh_supported()) {
> +		ret = cpc_read_arch_counters_on_cpu(cpunum, &delivered, &reference);
> +		if (ret) {
> +			pr_debug("read arch counters failed, ret=%d.\n", ret);
> +			ret = 0;
> +		}
> +	}
> +	if (!delivered || !reference) {
> +		cpc_read(cpunum, delivered_reg, &delivered);
> +		cpc_read(cpunum, reference_reg, &reference);
> +	}
> +
>   	cpc_read(cpunum, ref_perf_reg, &ref_perf);
>   
>   	/*
> diff --git a/include/acpi/cppc_acpi.h b/include/acpi/cppc_acpi.h
> index 6126c977ece0..07d4fd82d499 100644
> --- a/include/acpi/cppc_acpi.h
> +++ b/include/acpi/cppc_acpi.h
> @@ -152,6 +152,7 @@ extern bool cpc_ffh_supported(void);
>   extern bool cpc_supported_by_cpu(void);
>   extern int cpc_read_ffh(int cpunum, struct cpc_reg *reg, u64 *val);
>   extern int cpc_write_ffh(int cpunum, struct cpc_reg *reg, u64 val);
> +extern int cpc_read_arch_counters_on_cpu(int cpu, u64 *delivered, u64 *reference);
>   extern int cppc_get_epp_perf(int cpunum, u64 *epp_perf);
>   extern int cppc_set_epp_perf(int cpu, struct cppc_perf_ctrls *perf_ctrls, bool enable);
>   extern int cppc_get_auto_sel_caps(int cpunum, struct cppc_perf_caps *perf_caps);
> @@ -209,6 +210,10 @@ static inline int cpc_write_ffh(int cpunum, struct cpc_reg *reg, u64 val)
>   {
>   	return -ENOTSUPP;
>   }
> +static inline int cpc_read_arch_counters_on_cpu(int cpu, u64 *delivered, u64 *reference)
> +{
> +	return -EOPNOTSUPP;
> +}
>   static inline int cppc_set_epp_perf(int cpu, struct cppc_perf_ctrls *perf_ctrls, bool enable)
>   {
>   	return -ENOTSUPP;
Rafael J. Wysocki Jan. 3, 2024, 10:59 a.m. UTC | #5
On Mon, Dec 18, 2023 at 3:15 AM lihuisong (C) <lihuisong@huawei.com> wrote:
>
>
> 在 2023/12/15 10:41, lihuisong (C) 写道:
> > Hi Rafael,
> >
> > Thanks for your review.
Ionela Voinescu Jan. 3, 2024, 5:53 p.m. UTC | #6
Hi,

On Tuesday 12 Dec 2023 at 15:26:17 (+0800), Huisong Li wrote:
> Many developers found that the cpu current frequency is greater than
> the maximum frequency of the platform, please see [1], [2] and [3].
> 
> In the scenarios with high memory access pressure, the patch [1] has
> proved the significant latency of cpc_read() which is used to obtain
> delivered and reference performance counter cause an absurd frequency.
> The sampling interval for this counters is very critical and is expected
> to be equal. However, the different latency of cpc_read() has a direct
> impact on their sampling interval.
> 

Would this [1] alternative solution work for you?

[1] https://lore.kernel.org/lkml/20231127160838.1403404-1-beata.michalska@arm.com/

Thanks,
Ionela.

> This patch adds a interface, cpc_read_arch_counters_on_cpu, to read
> delivered and reference performance counter together. According to my
> test[4], the discrepancy of cpu current frequency in the scenarios with
> high memory access pressure is lower than 0.2% by stress-ng application.
> 
> [1] https://lore.kernel.org/all/20231025093847.3740104-4-zengheng4@huawei.com/
> [2] https://lore.kernel.org/all/20230328193846.8757-1-yang@os.amperecomputing.com/
> [3] https://lore.kernel.org/all/20230418113459.12860-7-sumitg@nvidia.com/
> 
> [4] My local test:
> The testing platform enable SMT and include 128 logical CPU in total,
> and CPU base frequency is 2.7GHz. Reading "cpuinfo_cur_freq" for each
> physical core on platform during the high memory access pressure from
> stress-ng, and the output is as follows:
>   0: 2699133     2: 2699942     4: 2698189     6: 2704347
>   8: 2704009    10: 2696277    12: 2702016    14: 2701388
>  16: 2700358    18: 2696741    20: 2700091    22: 2700122
>  24: 2701713    26: 2702025    28: 2699816    30: 2700121
>  32: 2700000    34: 2699788    36: 2698884    38: 2699109
>  40: 2704494    42: 2698350    44: 2699997    46: 2701023
>  48: 2703448    50: 2699501    52: 2700000    54: 2699999
>  56: 2702645    58: 2696923    60: 2697718    62: 2700547
>  64: 2700313    66: 2700000    68: 2699904    70: 2699259
>  72: 2699511    74: 2700644    76: 2702201    78: 2700000
>  80: 2700776    82: 2700364    84: 2702674    86: 2700255
>  88: 2699886    90: 2700359    92: 2699662    94: 2696188
>  96: 2705454    98: 2699260   100: 2701097   102: 2699630
> 104: 2700463   106: 2698408   108: 2697766   110: 2701181
> 112: 2699166   114: 2701804   116: 2701907   118: 2701973
> 120: 2699584   122: 2700474   124: 2700768   126: 2701963
> 
> Signed-off-by: Huisong Li <lihuisong@huawei.com>
> ---
>  arch/arm64/kernel/topology.c | 43 ++++++++++++++++++++++++++++++++++--
>  drivers/acpi/cppc_acpi.c     | 22 +++++++++++++++---
>  include/acpi/cppc_acpi.h     |  5 +++++
>  3 files changed, 65 insertions(+), 5 deletions(-)
> 
> diff --git a/arch/arm64/kernel/topology.c b/arch/arm64/kernel/topology.c
> index 7d37e458e2f5..c3122154d738 100644
> --- a/arch/arm64/kernel/topology.c
> +++ b/arch/arm64/kernel/topology.c
> @@ -299,6 +299,11 @@ core_initcall(init_amu_fie);
>  #ifdef CONFIG_ACPI_CPPC_LIB
>  #include <acpi/cppc_acpi.h>
>  
> +struct amu_counters {
> +	u64 corecnt;
> +	u64 constcnt;
> +};
> +
>  static void cpu_read_corecnt(void *val)
>  {
>  	/*
> @@ -322,8 +327,27 @@ static void cpu_read_constcnt(void *val)
>  		      0UL : read_constcnt();
>  }
>  
> +static void cpu_read_amu_counters(void *data)
> +{
> +	struct amu_counters *cnt = (struct amu_counters *)data;
> +
> +	/*
> +	 * The running time of the this_cpu_has_cap() might have a couple of
> +	 * microseconds and is significantly increased to tens of microseconds.
> +	 * But AMU core and constant counter need to be read togeter without any
> +	 * time interval to reduce the calculation discrepancy using this counters.
> +	 */
> +	if (this_cpu_has_cap(ARM64_WORKAROUND_2457168)) {
> +		cnt->corecnt = read_corecnt();
> +		cnt->constcnt = 0;
> +	} else {
> +		cnt->corecnt = read_corecnt();
> +		cnt->constcnt = read_constcnt();
> +	}
> +}
> +
>  static inline
> -int counters_read_on_cpu(int cpu, smp_call_func_t func, u64 *val)
> +int counters_read_on_cpu(int cpu, smp_call_func_t func, void *data)
>  {
>  	/*
>  	 * Abort call on counterless CPU or when interrupts are
> @@ -335,7 +359,7 @@ int counters_read_on_cpu(int cpu, smp_call_func_t func, u64 *val)
>  	if (WARN_ON_ONCE(irqs_disabled()))
>  		return -EPERM;
>  
> -	smp_call_function_single(cpu, func, val, 1);
> +	smp_call_function_single(cpu, func, data, 1);
>  
>  	return 0;
>  }
> @@ -364,6 +388,21 @@ bool cpc_ffh_supported(void)
>  	return true;
>  }
>  
> +int cpc_read_arch_counters_on_cpu(int cpu, u64 *delivered, u64 *reference)
> +{
> +	struct amu_counters cnts = {0};
> +	int ret;
> +
> +	ret = counters_read_on_cpu(cpu, cpu_read_amu_counters, &cnts);
> +	if (ret)
> +		return ret;
> +
> +	*delivered = cnts.corecnt;
> +	*reference = cnts.constcnt;
> +
> +	return 0;
> +}
> +
>  int cpc_read_ffh(int cpu, struct cpc_reg *reg, u64 *val)
>  {
>  	int ret = -EOPNOTSUPP;
> diff --git a/drivers/acpi/cppc_acpi.c b/drivers/acpi/cppc_acpi.c
> index 7ff269a78c20..f303fabd7cfe 100644
> --- a/drivers/acpi/cppc_acpi.c
> +++ b/drivers/acpi/cppc_acpi.c
> @@ -1299,6 +1299,11 @@ bool cppc_perf_ctrs_in_pcc(void)
>  }
>  EXPORT_SYMBOL_GPL(cppc_perf_ctrs_in_pcc);
>  
> +int __weak cpc_read_arch_counters_on_cpu(int cpu, u64 *delivered, u64 *reference)
> +{
> +	return 0;
> +}
> +
>  /**
>   * cppc_get_perf_ctrs - Read a CPU's performance feedback counters.
>   * @cpunum: CPU from which to read counters.
> @@ -1313,7 +1318,8 @@ int cppc_get_perf_ctrs(int cpunum, struct cppc_perf_fb_ctrs *perf_fb_ctrs)
>  		*ref_perf_reg, *ctr_wrap_reg;
>  	int pcc_ss_id = per_cpu(cpu_pcc_subspace_idx, cpunum);
>  	struct cppc_pcc_data *pcc_ss_data = NULL;
> -	u64 delivered, reference, ref_perf, ctr_wrap_time;
> +	u64 delivered = 0, reference = 0;
> +	u64 ref_perf, ctr_wrap_time;
>  	int ret = 0, regs_in_pcc = 0;
>  
>  	if (!cpc_desc) {
> @@ -1350,8 +1356,18 @@ int cppc_get_perf_ctrs(int cpunum, struct cppc_perf_fb_ctrs *perf_fb_ctrs)
>  		}
>  	}
>  
> -	cpc_read(cpunum, delivered_reg, &delivered);
> -	cpc_read(cpunum, reference_reg, &reference);
> +	if (cpc_ffh_supported()) {
> +		ret = cpc_read_arch_counters_on_cpu(cpunum, &delivered, &reference);
> +		if (ret) {
> +			pr_debug("read arch counters failed, ret=%d.\n", ret);
> +			ret = 0;
> +		}
> +	}
> +	if (!delivered || !reference) {
> +		cpc_read(cpunum, delivered_reg, &delivered);
> +		cpc_read(cpunum, reference_reg, &reference);
> +	}
> +
>  	cpc_read(cpunum, ref_perf_reg, &ref_perf);
>  
>  	/*
> diff --git a/include/acpi/cppc_acpi.h b/include/acpi/cppc_acpi.h
> index 6126c977ece0..07d4fd82d499 100644
> --- a/include/acpi/cppc_acpi.h
> +++ b/include/acpi/cppc_acpi.h
> @@ -152,6 +152,7 @@ extern bool cpc_ffh_supported(void);
>  extern bool cpc_supported_by_cpu(void);
>  extern int cpc_read_ffh(int cpunum, struct cpc_reg *reg, u64 *val);
>  extern int cpc_write_ffh(int cpunum, struct cpc_reg *reg, u64 val);
> +extern int cpc_read_arch_counters_on_cpu(int cpu, u64 *delivered, u64 *reference);
>  extern int cppc_get_epp_perf(int cpunum, u64 *epp_perf);
>  extern int cppc_set_epp_perf(int cpu, struct cppc_perf_ctrls *perf_ctrls, bool enable);
>  extern int cppc_get_auto_sel_caps(int cpunum, struct cppc_perf_caps *perf_caps);
> @@ -209,6 +210,10 @@ static inline int cpc_write_ffh(int cpunum, struct cpc_reg *reg, u64 val)
>  {
>  	return -ENOTSUPP;
>  }
> +static inline int cpc_read_arch_counters_on_cpu(int cpu, u64 *delivered, u64 *reference)
> +{
> +	return -EOPNOTSUPP;
> +}
>  static inline int cppc_set_epp_perf(int cpu, struct cppc_perf_ctrls *perf_ctrls, bool enable)
>  {
>  	return -ENOTSUPP;
> -- 
> 2.33.0
>
lihuisong (C) Jan. 4, 2024, 9:07 a.m. UTC | #7
在 2024/1/3 18:59, Rafael J. Wysocki 写道:
> On Mon, Dec 18, 2023 at 3:15 AM lihuisong (C) <lihuisong@huawei.com> wrote:
>>
>> 在 2023/12/15 10:41, lihuisong (C) 写道:
>>> Hi Rafael,
>>>
>>> Thanks for your review.
lihuisong (C) Jan. 4, 2024, 9:36 a.m. UTC | #8
在 2024/1/4 1:53, Ionela Voinescu 写道:
> Hi,
>
> On Tuesday 12 Dec 2023 at 15:26:17 (+0800), Huisong Li wrote:
>> Many developers found that the cpu current frequency is greater than
>> the maximum frequency of the platform, please see [1], [2] and [3].
>>
>> In the scenarios with high memory access pressure, the patch [1] has
>> proved the significant latency of cpc_read() which is used to obtain
>> delivered and reference performance counter cause an absurd frequency.
>> The sampling interval for this counters is very critical and is expected
>> to be equal. However, the different latency of cpc_read() has a direct
>> impact on their sampling interval.
>>
> Would this [1] alternative solution work for you?
It would work for me AFAICS.
Because the "arch_freq_scale" is also from AMU core and constant 
counter, and read together.
But, from their discuss line, it seems that there are some tricky points 
to clarify or consider.
>
> [1] https://lore.kernel.org/lkml/20231127160838.1403404-1-beata.michalska@arm.com/
>
> Thanks,
> Ionela.
>
>> This patch adds a interface, cpc_read_arch_counters_on_cpu, to read
>> delivered and reference performance counter together. According to my
>> test[4], the discrepancy of cpu current frequency in the scenarios with
>> high memory access pressure is lower than 0.2% by stress-ng application.
>>
>> [1] https://lore.kernel.org/all/20231025093847.3740104-4-zengheng4@huawei.com/
>> [2] https://lore.kernel.org/all/20230328193846.8757-1-yang@os.amperecomputing.com/
>> [3] https://lore.kernel.org/all/20230418113459.12860-7-sumitg@nvidia.com/
>>
>> [4] My local test:
>> The testing platform enable SMT and include 128 logical CPU in total,
>> and CPU base frequency is 2.7GHz. Reading "cpuinfo_cur_freq" for each
>> physical core on platform during the high memory access pressure from
>> stress-ng, and the output is as follows:
>>    0: 2699133     2: 2699942     4: 2698189     6: 2704347
>>    8: 2704009    10: 2696277    12: 2702016    14: 2701388
>>   16: 2700358    18: 2696741    20: 2700091    22: 2700122
>>   24: 2701713    26: 2702025    28: 2699816    30: 2700121
>>   32: 2700000    34: 2699788    36: 2698884    38: 2699109
>>   40: 2704494    42: 2698350    44: 2699997    46: 2701023
>>   48: 2703448    50: 2699501    52: 2700000    54: 2699999
>>   56: 2702645    58: 2696923    60: 2697718    62: 2700547
>>   64: 2700313    66: 2700000    68: 2699904    70: 2699259
>>   72: 2699511    74: 2700644    76: 2702201    78: 2700000
>>   80: 2700776    82: 2700364    84: 2702674    86: 2700255
>>   88: 2699886    90: 2700359    92: 2699662    94: 2696188
>>   96: 2705454    98: 2699260   100: 2701097   102: 2699630
>> 104: 2700463   106: 2698408   108: 2697766   110: 2701181
>> 112: 2699166   114: 2701804   116: 2701907   118: 2701973
>> 120: 2699584   122: 2700474   124: 2700768   126: 2701963
>>
>> Signed-off-by: Huisong Li <lihuisong@huawei.com>
>> ---
>>   arch/arm64/kernel/topology.c | 43 ++++++++++++++++++++++++++++++++++--
>>   drivers/acpi/cppc_acpi.c     | 22 +++++++++++++++---
>>   include/acpi/cppc_acpi.h     |  5 +++++
>>   3 files changed, 65 insertions(+), 5 deletions(-)
>>
>> diff --git a/arch/arm64/kernel/topology.c b/arch/arm64/kernel/topology.c
>> index 7d37e458e2f5..c3122154d738 100644
>> --- a/arch/arm64/kernel/topology.c
>> +++ b/arch/arm64/kernel/topology.c
>> @@ -299,6 +299,11 @@ core_initcall(init_amu_fie);
>>   #ifdef CONFIG_ACPI_CPPC_LIB
>>   #include <acpi/cppc_acpi.h>
>>   
>> +struct amu_counters {
>> +	u64 corecnt;
>> +	u64 constcnt;
>> +};
>> +
>>   static void cpu_read_corecnt(void *val)
>>   {
>>   	/*
>> @@ -322,8 +327,27 @@ static void cpu_read_constcnt(void *val)
>>   		      0UL : read_constcnt();
>>   }
>>   
>> +static void cpu_read_amu_counters(void *data)
>> +{
>> +	struct amu_counters *cnt = (struct amu_counters *)data;
>> +
>> +	/*
>> +	 * The running time of the this_cpu_has_cap() might have a couple of
>> +	 * microseconds and is significantly increased to tens of microseconds.
>> +	 * But AMU core and constant counter need to be read togeter without any
>> +	 * time interval to reduce the calculation discrepancy using this counters.
>> +	 */
>> +	if (this_cpu_has_cap(ARM64_WORKAROUND_2457168)) {
>> +		cnt->corecnt = read_corecnt();
>> +		cnt->constcnt = 0;
>> +	} else {
>> +		cnt->corecnt = read_corecnt();
>> +		cnt->constcnt = read_constcnt();
>> +	}
>> +}
>> +
>>   static inline
>> -int counters_read_on_cpu(int cpu, smp_call_func_t func, u64 *val)
>> +int counters_read_on_cpu(int cpu, smp_call_func_t func, void *data)
>>   {
>>   	/*
>>   	 * Abort call on counterless CPU or when interrupts are
>> @@ -335,7 +359,7 @@ int counters_read_on_cpu(int cpu, smp_call_func_t func, u64 *val)
>>   	if (WARN_ON_ONCE(irqs_disabled()))
>>   		return -EPERM;
>>   
>> -	smp_call_function_single(cpu, func, val, 1);
>> +	smp_call_function_single(cpu, func, data, 1);
>>   
>>   	return 0;
>>   }
>> @@ -364,6 +388,21 @@ bool cpc_ffh_supported(void)
>>   	return true;
>>   }
>>   
>> +int cpc_read_arch_counters_on_cpu(int cpu, u64 *delivered, u64 *reference)
>> +{
>> +	struct amu_counters cnts = {0};
>> +	int ret;
>> +
>> +	ret = counters_read_on_cpu(cpu, cpu_read_amu_counters, &cnts);
>> +	if (ret)
>> +		return ret;
>> +
>> +	*delivered = cnts.corecnt;
>> +	*reference = cnts.constcnt;
>> +
>> +	return 0;
>> +}
>> +
>>   int cpc_read_ffh(int cpu, struct cpc_reg *reg, u64 *val)
>>   {
>>   	int ret = -EOPNOTSUPP;
>> diff --git a/drivers/acpi/cppc_acpi.c b/drivers/acpi/cppc_acpi.c
>> index 7ff269a78c20..f303fabd7cfe 100644
>> --- a/drivers/acpi/cppc_acpi.c
>> +++ b/drivers/acpi/cppc_acpi.c
>> @@ -1299,6 +1299,11 @@ bool cppc_perf_ctrs_in_pcc(void)
>>   }
>>   EXPORT_SYMBOL_GPL(cppc_perf_ctrs_in_pcc);
>>   
>> +int __weak cpc_read_arch_counters_on_cpu(int cpu, u64 *delivered, u64 *reference)
>> +{
>> +	return 0;
>> +}
>> +
>>   /**
>>    * cppc_get_perf_ctrs - Read a CPU's performance feedback counters.
>>    * @cpunum: CPU from which to read counters.
>> @@ -1313,7 +1318,8 @@ int cppc_get_perf_ctrs(int cpunum, struct cppc_perf_fb_ctrs *perf_fb_ctrs)
>>   		*ref_perf_reg, *ctr_wrap_reg;
>>   	int pcc_ss_id = per_cpu(cpu_pcc_subspace_idx, cpunum);
>>   	struct cppc_pcc_data *pcc_ss_data = NULL;
>> -	u64 delivered, reference, ref_perf, ctr_wrap_time;
>> +	u64 delivered = 0, reference = 0;
>> +	u64 ref_perf, ctr_wrap_time;
>>   	int ret = 0, regs_in_pcc = 0;
>>   
>>   	if (!cpc_desc) {
>> @@ -1350,8 +1356,18 @@ int cppc_get_perf_ctrs(int cpunum, struct cppc_perf_fb_ctrs *perf_fb_ctrs)
>>   		}
>>   	}
>>   
>> -	cpc_read(cpunum, delivered_reg, &delivered);
>> -	cpc_read(cpunum, reference_reg, &reference);
>> +	if (cpc_ffh_supported()) {
>> +		ret = cpc_read_arch_counters_on_cpu(cpunum, &delivered, &reference);
>> +		if (ret) {
>> +			pr_debug("read arch counters failed, ret=%d.\n", ret);
>> +			ret = 0;
>> +		}
>> +	}
>> +	if (!delivered || !reference) {
>> +		cpc_read(cpunum, delivered_reg, &delivered);
>> +		cpc_read(cpunum, reference_reg, &reference);
>> +	}
>> +
>>   	cpc_read(cpunum, ref_perf_reg, &ref_perf);
>>   
>>   	/*
>> diff --git a/include/acpi/cppc_acpi.h b/include/acpi/cppc_acpi.h
>> index 6126c977ece0..07d4fd82d499 100644
>> --- a/include/acpi/cppc_acpi.h
>> +++ b/include/acpi/cppc_acpi.h
>> @@ -152,6 +152,7 @@ extern bool cpc_ffh_supported(void);
>>   extern bool cpc_supported_by_cpu(void);
>>   extern int cpc_read_ffh(int cpunum, struct cpc_reg *reg, u64 *val);
>>   extern int cpc_write_ffh(int cpunum, struct cpc_reg *reg, u64 val);
>> +extern int cpc_read_arch_counters_on_cpu(int cpu, u64 *delivered, u64 *reference);
>>   extern int cppc_get_epp_perf(int cpunum, u64 *epp_perf);
>>   extern int cppc_set_epp_perf(int cpu, struct cppc_perf_ctrls *perf_ctrls, bool enable);
>>   extern int cppc_get_auto_sel_caps(int cpunum, struct cppc_perf_caps *perf_caps);
>> @@ -209,6 +210,10 @@ static inline int cpc_write_ffh(int cpunum, struct cpc_reg *reg, u64 val)
>>   {
>>   	return -ENOTSUPP;
>>   }
>> +static inline int cpc_read_arch_counters_on_cpu(int cpu, u64 *delivered, u64 *reference)
>> +{
>> +	return -EOPNOTSUPP;
>> +}
>>   static inline int cppc_set_epp_perf(int cpu, struct cppc_perf_ctrls *perf_ctrls, bool enable)
>>   {
>>   	return -ENOTSUPP;
>> -- 
>> 2.33.0
>>
> .
Vanshidhar Konda Jan. 5, 2024, 12:48 a.m. UTC | #9
On Thu, Jan 04, 2024 at 05:36:51PM +0800, lihuisong (C) wrote:
>
>在 2024/1/4 1:53, Ionela Voinescu 写道:
>>Hi,
>>
>>On Tuesday 12 Dec 2023 at 15:26:17 (+0800), Huisong Li wrote:
>>>Many developers found that the cpu current frequency is greater than
>>>the maximum frequency of the platform, please see [1], [2] and [3].
>>>
>>>In the scenarios with high memory access pressure, the patch [1] has
>>>proved the significant latency of cpc_read() which is used to obtain
>>>delivered and reference performance counter cause an absurd frequency.
>>>The sampling interval for this counters is very critical and is expected
>>>to be equal. However, the different latency of cpc_read() has a direct
>>>impact on their sampling interval.
>>>
>>Would this [1] alternative solution work for you?
>It would work for me AFAICS.
>Because the "arch_freq_scale" is also from AMU core and constant 
>counter, and read together.
>But, from their discuss line, it seems that there are some tricky 
>points to clarify or consider.

I think the changes in [1] would work better when CPUs may be idle. With this
patch we would have to wake any core that is in idle state to read the AMU
counters. Worst case, if core 0 is trying to read the CPU frequency of all
cores, it may need to wake up all the other cores to read the AMU counters.
For systems with 128 cores or more, this could be very expensive and happen
very frequently.

AFAICS, the approach in [1] would avoid this cost.

Thanks,
Vanshi

>>
>>[1] https://lore.kernel.org/lkml/20231127160838.1403404-1-beata.michalska@arm.com/
>>
>>Thanks,
>>Ionela.
>>
>>>This patch adds a interface, cpc_read_arch_counters_on_cpu, to read
>>>delivered and reference performance counter together. According to my
>>>test[4], the discrepancy of cpu current frequency in the scenarios with
>>>high memory access pressure is lower than 0.2% by stress-ng application.
>>>
>>>[1] https://lore.kernel.org/all/20231025093847.3740104-4-zengheng4@huawei.com/
>>>[2] https://lore.kernel.org/all/20230328193846.8757-1-yang@os.amperecomputing.com/
>>>[3] https://lore.kernel.org/all/20230418113459.12860-7-sumitg@nvidia.com/
>>>
>>>[4] My local test:
>>>The testing platform enable SMT and include 128 logical CPU in total,
>>>and CPU base frequency is 2.7GHz. Reading "cpuinfo_cur_freq" for each
>>>physical core on platform during the high memory access pressure from
>>>stress-ng, and the output is as follows:
>>>   0: 2699133     2: 2699942     4: 2698189     6: 2704347
>>>   8: 2704009    10: 2696277    12: 2702016    14: 2701388
>>>  16: 2700358    18: 2696741    20: 2700091    22: 2700122
>>>  24: 2701713    26: 2702025    28: 2699816    30: 2700121
>>>  32: 2700000    34: 2699788    36: 2698884    38: 2699109
>>>  40: 2704494    42: 2698350    44: 2699997    46: 2701023
>>>  48: 2703448    50: 2699501    52: 2700000    54: 2699999
>>>  56: 2702645    58: 2696923    60: 2697718    62: 2700547
>>>  64: 2700313    66: 2700000    68: 2699904    70: 2699259
>>>  72: 2699511    74: 2700644    76: 2702201    78: 2700000
>>>  80: 2700776    82: 2700364    84: 2702674    86: 2700255
>>>  88: 2699886    90: 2700359    92: 2699662    94: 2696188
>>>  96: 2705454    98: 2699260   100: 2701097   102: 2699630
>>>104: 2700463   106: 2698408   108: 2697766   110: 2701181
>>>112: 2699166   114: 2701804   116: 2701907   118: 2701973
>>>120: 2699584   122: 2700474   124: 2700768   126: 2701963
>>>
>>>Signed-off-by: Huisong Li <lihuisong@huawei.com>
>>>---
>>>  arch/arm64/kernel/topology.c | 43 ++++++++++++++++++++++++++++++++++--
>>>  drivers/acpi/cppc_acpi.c     | 22 +++++++++++++++---
>>>  include/acpi/cppc_acpi.h     |  5 +++++
>>>  3 files changed, 65 insertions(+), 5 deletions(-)
>>>
>>>diff --git a/arch/arm64/kernel/topology.c b/arch/arm64/kernel/topology.c
>>>index 7d37e458e2f5..c3122154d738 100644
>>>--- a/arch/arm64/kernel/topology.c
>>>+++ b/arch/arm64/kernel/topology.c
>>>@@ -299,6 +299,11 @@ core_initcall(init_amu_fie);
>>>  #ifdef CONFIG_ACPI_CPPC_LIB
>>>  #include <acpi/cppc_acpi.h>
>>>+struct amu_counters {
>>>+	u64 corecnt;
>>>+	u64 constcnt;
>>>+};
>>>+
>>>  static void cpu_read_corecnt(void *val)
>>>  {
>>>  	/*
>>>@@ -322,8 +327,27 @@ static void cpu_read_constcnt(void *val)
>>>  		      0UL : read_constcnt();
>>>  }
>>>+static void cpu_read_amu_counters(void *data)
>>>+{
>>>+	struct amu_counters *cnt = (struct amu_counters *)data;
>>>+
>>>+	/*
>>>+	 * The running time of the this_cpu_has_cap() might have a couple of
>>>+	 * microseconds and is significantly increased to tens of microseconds.
>>>+	 * But AMU core and constant counter need to be read togeter without any
>>>+	 * time interval to reduce the calculation discrepancy using this counters.
>>>+	 */
>>>+	if (this_cpu_has_cap(ARM64_WORKAROUND_2457168)) {
>>>+		cnt->corecnt = read_corecnt();
>>>+		cnt->constcnt = 0;
>>>+	} else {
>>>+		cnt->corecnt = read_corecnt();
>>>+		cnt->constcnt = read_constcnt();
>>>+	}
>>>+}
>>>+
>>>  static inline
>>>-int counters_read_on_cpu(int cpu, smp_call_func_t func, u64 *val)
>>>+int counters_read_on_cpu(int cpu, smp_call_func_t func, void *data)
>>>  {
>>>  	/*
>>>  	 * Abort call on counterless CPU or when interrupts are
>>>@@ -335,7 +359,7 @@ int counters_read_on_cpu(int cpu, smp_call_func_t func, u64 *val)
>>>  	if (WARN_ON_ONCE(irqs_disabled()))
>>>  		return -EPERM;
>>>-	smp_call_function_single(cpu, func, val, 1);
>>>+	smp_call_function_single(cpu, func, data, 1);
>>>  	return 0;
>>>  }
>>>@@ -364,6 +388,21 @@ bool cpc_ffh_supported(void)
>>>  	return true;
>>>  }
>>>+int cpc_read_arch_counters_on_cpu(int cpu, u64 *delivered, u64 *reference)
>>>+{
>>>+	struct amu_counters cnts = {0};
>>>+	int ret;
>>>+
>>>+	ret = counters_read_on_cpu(cpu, cpu_read_amu_counters, &cnts);
>>>+	if (ret)
>>>+		return ret;
>>>+
>>>+	*delivered = cnts.corecnt;
>>>+	*reference = cnts.constcnt;
>>>+
>>>+	return 0;
>>>+}
>>>+
>>>  int cpc_read_ffh(int cpu, struct cpc_reg *reg, u64 *val)
>>>  {
>>>  	int ret = -EOPNOTSUPP;
>>>diff --git a/drivers/acpi/cppc_acpi.c b/drivers/acpi/cppc_acpi.c
>>>index 7ff269a78c20..f303fabd7cfe 100644
>>>--- a/drivers/acpi/cppc_acpi.c
>>>+++ b/drivers/acpi/cppc_acpi.c
>>>@@ -1299,6 +1299,11 @@ bool cppc_perf_ctrs_in_pcc(void)
>>>  }
>>>  EXPORT_SYMBOL_GPL(cppc_perf_ctrs_in_pcc);
>>>+int __weak cpc_read_arch_counters_on_cpu(int cpu, u64 *delivered, u64 *reference)
>>>+{
>>>+	return 0;
>>>+}
>>>+
>>>  /**
>>>   * cppc_get_perf_ctrs - Read a CPU's performance feedback counters.
>>>   * @cpunum: CPU from which to read counters.
>>>@@ -1313,7 +1318,8 @@ int cppc_get_perf_ctrs(int cpunum, struct cppc_perf_fb_ctrs *perf_fb_ctrs)
>>>  		*ref_perf_reg, *ctr_wrap_reg;
>>>  	int pcc_ss_id = per_cpu(cpu_pcc_subspace_idx, cpunum);
>>>  	struct cppc_pcc_data *pcc_ss_data = NULL;
>>>-	u64 delivered, reference, ref_perf, ctr_wrap_time;
>>>+	u64 delivered = 0, reference = 0;
>>>+	u64 ref_perf, ctr_wrap_time;
>>>  	int ret = 0, regs_in_pcc = 0;
>>>  	if (!cpc_desc) {
>>>@@ -1350,8 +1356,18 @@ int cppc_get_perf_ctrs(int cpunum, struct cppc_perf_fb_ctrs *perf_fb_ctrs)
>>>  		}
>>>  	}
>>>-	cpc_read(cpunum, delivered_reg, &delivered);
>>>-	cpc_read(cpunum, reference_reg, &reference);
>>>+	if (cpc_ffh_supported()) {
>>>+		ret = cpc_read_arch_counters_on_cpu(cpunum, &delivered, &reference);
>>>+		if (ret) {
>>>+			pr_debug("read arch counters failed, ret=%d.\n", ret);
>>>+			ret = 0;
>>>+		}
>>>+	}
>>>+	if (!delivered || !reference) {
>>>+		cpc_read(cpunum, delivered_reg, &delivered);
>>>+		cpc_read(cpunum, reference_reg, &reference);
>>>+	}
>>>+
>>>  	cpc_read(cpunum, ref_perf_reg, &ref_perf);
>>>  	/*
>>>diff --git a/include/acpi/cppc_acpi.h b/include/acpi/cppc_acpi.h
>>>index 6126c977ece0..07d4fd82d499 100644
>>>--- a/include/acpi/cppc_acpi.h
>>>+++ b/include/acpi/cppc_acpi.h
>>>@@ -152,6 +152,7 @@ extern bool cpc_ffh_supported(void);
>>>  extern bool cpc_supported_by_cpu(void);
>>>  extern int cpc_read_ffh(int cpunum, struct cpc_reg *reg, u64 *val);
>>>  extern int cpc_write_ffh(int cpunum, struct cpc_reg *reg, u64 val);
>>>+extern int cpc_read_arch_counters_on_cpu(int cpu, u64 *delivered, u64 *reference);
>>>  extern int cppc_get_epp_perf(int cpunum, u64 *epp_perf);
>>>  extern int cppc_set_epp_perf(int cpu, struct cppc_perf_ctrls *perf_ctrls, bool enable);
>>>  extern int cppc_get_auto_sel_caps(int cpunum, struct cppc_perf_caps *perf_caps);
>>>@@ -209,6 +210,10 @@ static inline int cpc_write_ffh(int cpunum, struct cpc_reg *reg, u64 val)
>>>  {
>>>  	return -ENOTSUPP;
>>>  }
>>>+static inline int cpc_read_arch_counters_on_cpu(int cpu, u64 *delivered, u64 *reference)
>>>+{
>>>+	return -EOPNOTSUPP;
>>>+}
>>>  static inline int cppc_set_epp_perf(int cpu, struct cppc_perf_ctrls *perf_ctrls, bool enable)
>>>  {
>>>  	return -ENOTSUPP;
>>>-- 
>>>2.33.0
>>>
>>.
>
>_______________________________________________
>linux-arm-kernel mailing list
>linux-arm-kernel@lists.infradead.org
>http://lists.infradead.org/mailman/listinfo/linux-arm-kernel
lihuisong (C) Jan. 5, 2024, 7:04 a.m. UTC | #10
Hi Vanshi,

在 2024/1/5 8:48, Vanshidhar Konda 写道:
> On Thu, Jan 04, 2024 at 05:36:51PM +0800, lihuisong (C) wrote:
>>
>> 在 2024/1/4 1:53, Ionela Voinescu 写道:
>>> Hi,
>>>
>>> On Tuesday 12 Dec 2023 at 15:26:17 (+0800), Huisong Li wrote:
>>>> Many developers found that the cpu current frequency is greater than
>>>> the maximum frequency of the platform, please see [1], [2] and [3].
>>>>
>>>> In the scenarios with high memory access pressure, the patch [1] has
>>>> proved the significant latency of cpc_read() which is used to obtain
>>>> delivered and reference performance counter cause an absurd frequency.
>>>> The sampling interval for this counters is very critical and is 
>>>> expected
>>>> to be equal. However, the different latency of cpc_read() has a direct
>>>> impact on their sampling interval.
>>>>
>>> Would this [1] alternative solution work for you?
>> It would work for me AFAICS.
>> Because the "arch_freq_scale" is also from AMU core and constant 
>> counter, and read together.
>> But, from their discuss line, it seems that there are some tricky 
>> points to clarify or consider.
>
> I think the changes in [1] would work better when CPUs may be idle. 
> With this
> patch we would have to wake any core that is in idle state to read the 
> AMU
> counters. Worst case, if core 0 is trying to read the CPU frequency of 
> all
> cores, it may need to wake up all the other cores to read the AMU 
> counters.
 From the approach in [1], if all CPUs (one or more cores) under one 
policy are idle, they still cannot be obtained the CPU frequency, right?
In this case, the [1] API will return 0 and have to back to call 
cpufreq_driver->get() for cpuinfo_cur_freq.
Then we still need to face the issue this patch mentioned.
> For systems with 128 cores or more, this could be very expensive and 
> happen
> very frequently.
>
> AFAICS, the approach in [1] would avoid this cost.
But the CPU frequency is just an average value for the last tick period 
instead of the current one the CPU actually runs at.
In addition, there are some conditions to use 'arch_freq_scale' in this 
approach.
So I'm not sure if this approach can entirely cover the frequency 
discrepancy issue.

/Huisong

>>>
>>> [1] 
>>> https://lore.kernel.org/lkml/20231127160838.1403404-1-beata.michalska@arm.com/
>>>
>>> Thanks,
>>> Ionela.
>>>
>>>> This patch adds a interface, cpc_read_arch_counters_on_cpu, to read
>>>> delivered and reference performance counter together. According to my
>>>> test[4], the discrepancy of cpu current frequency in the scenarios 
>>>> with
>>>> high memory access pressure is lower than 0.2% by stress-ng 
>>>> application.
>>>>
>>>> [1] 
>>>> https://lore.kernel.org/all/20231025093847.3740104-4-zengheng4@huawei.com/
>>>> [2] 
>>>> https://lore.kernel.org/all/20230328193846.8757-1-yang@os.amperecomputing.com/
>>>> [3] 
>>>> https://lore.kernel.org/all/20230418113459.12860-7-sumitg@nvidia.com/
>>>>
>>>> [4] My local test:
>>>> The testing platform enable SMT and include 128 logical CPU in total,
>>>> and CPU base frequency is 2.7GHz. Reading "cpuinfo_cur_freq" for each
>>>> physical core on platform during the high memory access pressure from
>>>> stress-ng, and the output is as follows:
>>>>   0: 2699133     2: 2699942     4: 2698189     6: 2704347
>>>>   8: 2704009    10: 2696277    12: 2702016    14: 2701388
>>>>  16: 2700358    18: 2696741    20: 2700091    22: 2700122
>>>>  24: 2701713    26: 2702025    28: 2699816    30: 2700121
>>>>  32: 2700000    34: 2699788    36: 2698884    38: 2699109
>>>>  40: 2704494    42: 2698350    44: 2699997    46: 2701023
>>>>  48: 2703448    50: 2699501    52: 2700000    54: 2699999
>>>>  56: 2702645    58: 2696923    60: 2697718    62: 2700547
>>>>  64: 2700313    66: 2700000    68: 2699904    70: 2699259
>>>>  72: 2699511    74: 2700644    76: 2702201    78: 2700000
>>>>  80: 2700776    82: 2700364    84: 2702674    86: 2700255
>>>>  88: 2699886    90: 2700359    92: 2699662    94: 2696188
>>>>  96: 2705454    98: 2699260   100: 2701097   102: 2699630
>>>> 104: 2700463   106: 2698408   108: 2697766   110: 2701181
>>>> 112: 2699166   114: 2701804   116: 2701907   118: 2701973
>>>> 120: 2699584   122: 2700474   124: 2700768   126: 2701963
>>>>
>>>> Signed-off-by: Huisong Li <lihuisong@huawei.com>
>>>> ---
>>>>  arch/arm64/kernel/topology.c | 43 
>>>> ++++++++++++++++++++++++++++++++++--
>>>>  drivers/acpi/cppc_acpi.c     | 22 +++++++++++++++---
>>>>  include/acpi/cppc_acpi.h     |  5 +++++
>>>>  3 files changed, 65 insertions(+), 5 deletions(-)
>>>>
>>>> diff --git a/arch/arm64/kernel/topology.c 
>>>> b/arch/arm64/kernel/topology.c
>>>> index 7d37e458e2f5..c3122154d738 100644
>>>> --- a/arch/arm64/kernel/topology.c
>>>> +++ b/arch/arm64/kernel/topology.c
>>>> @@ -299,6 +299,11 @@ core_initcall(init_amu_fie);
>>>>  #ifdef CONFIG_ACPI_CPPC_LIB
>>>>  #include <acpi/cppc_acpi.h>
>>>> +struct amu_counters {
>>>> +    u64 corecnt;
>>>> +    u64 constcnt;
>>>> +};
>>>> +
>>>>  static void cpu_read_corecnt(void *val)
>>>>  {
>>>>      /*
>>>> @@ -322,8 +327,27 @@ static void cpu_read_constcnt(void *val)
>>>>                0UL : read_constcnt();
>>>>  }
>>>> +static void cpu_read_amu_counters(void *data)
>>>> +{
>>>> +    struct amu_counters *cnt = (struct amu_counters *)data;
>>>> +
>>>> +    /*
>>>> +     * The running time of the this_cpu_has_cap() might have a 
>>>> couple of
>>>> +     * microseconds and is significantly increased to tens of 
>>>> microseconds.
>>>> +     * But AMU core and constant counter need to be read togeter 
>>>> without any
>>>> +     * time interval to reduce the calculation discrepancy using 
>>>> this counters.
>>>> +     */
>>>> +    if (this_cpu_has_cap(ARM64_WORKAROUND_2457168)) {
>>>> +        cnt->corecnt = read_corecnt();
>>>> +        cnt->constcnt = 0;
>>>> +    } else {
>>>> +        cnt->corecnt = read_corecnt();
>>>> +        cnt->constcnt = read_constcnt();
>>>> +    }
>>>> +}
>>>> +
>>>>  static inline
>>>> -int counters_read_on_cpu(int cpu, smp_call_func_t func, u64 *val)
>>>> +int counters_read_on_cpu(int cpu, smp_call_func_t func, void *data)
>>>>  {
>>>>      /*
>>>>       * Abort call on counterless CPU or when interrupts are
>>>> @@ -335,7 +359,7 @@ int counters_read_on_cpu(int cpu, 
>>>> smp_call_func_t func, u64 *val)
>>>>      if (WARN_ON_ONCE(irqs_disabled()))
>>>>          return -EPERM;
>>>> -    smp_call_function_single(cpu, func, val, 1);
>>>> +    smp_call_function_single(cpu, func, data, 1);
>>>>      return 0;
>>>>  }
>>>> @@ -364,6 +388,21 @@ bool cpc_ffh_supported(void)
>>>>      return true;
>>>>  }
>>>> +int cpc_read_arch_counters_on_cpu(int cpu, u64 *delivered, u64 
>>>> *reference)
>>>> +{
>>>> +    struct amu_counters cnts = {0};
>>>> +    int ret;
>>>> +
>>>> +    ret = counters_read_on_cpu(cpu, cpu_read_amu_counters, &cnts);
>>>> +    if (ret)
>>>> +        return ret;
>>>> +
>>>> +    *delivered = cnts.corecnt;
>>>> +    *reference = cnts.constcnt;
>>>> +
>>>> +    return 0;
>>>> +}
>>>> +
>>>>  int cpc_read_ffh(int cpu, struct cpc_reg *reg, u64 *val)
>>>>  {
>>>>      int ret = -EOPNOTSUPP;
>>>> diff --git a/drivers/acpi/cppc_acpi.c b/drivers/acpi/cppc_acpi.c
>>>> index 7ff269a78c20..f303fabd7cfe 100644
>>>> --- a/drivers/acpi/cppc_acpi.c
>>>> +++ b/drivers/acpi/cppc_acpi.c
>>>> @@ -1299,6 +1299,11 @@ bool cppc_perf_ctrs_in_pcc(void)
>>>>  }
>>>>  EXPORT_SYMBOL_GPL(cppc_perf_ctrs_in_pcc);
>>>> +int __weak cpc_read_arch_counters_on_cpu(int cpu, u64 *delivered, 
>>>> u64 *reference)
>>>> +{
>>>> +    return 0;
>>>> +}
>>>> +
>>>>  /**
>>>>   * cppc_get_perf_ctrs - Read a CPU's performance feedback counters.
>>>>   * @cpunum: CPU from which to read counters.
>>>> @@ -1313,7 +1318,8 @@ int cppc_get_perf_ctrs(int cpunum, struct 
>>>> cppc_perf_fb_ctrs *perf_fb_ctrs)
>>>>          *ref_perf_reg, *ctr_wrap_reg;
>>>>      int pcc_ss_id = per_cpu(cpu_pcc_subspace_idx, cpunum);
>>>>      struct cppc_pcc_data *pcc_ss_data = NULL;
>>>> -    u64 delivered, reference, ref_perf, ctr_wrap_time;
>>>> +    u64 delivered = 0, reference = 0;
>>>> +    u64 ref_perf, ctr_wrap_time;
>>>>      int ret = 0, regs_in_pcc = 0;
>>>>      if (!cpc_desc) {
>>>> @@ -1350,8 +1356,18 @@ int cppc_get_perf_ctrs(int cpunum, struct 
>>>> cppc_perf_fb_ctrs *perf_fb_ctrs)
>>>>          }
>>>>      }
>>>> -    cpc_read(cpunum, delivered_reg, &delivered);
>>>> -    cpc_read(cpunum, reference_reg, &reference);
>>>> +    if (cpc_ffh_supported()) {
>>>> +        ret = cpc_read_arch_counters_on_cpu(cpunum, &delivered, 
>>>> &reference);
>>>> +        if (ret) {
>>>> +            pr_debug("read arch counters failed, ret=%d.\n", ret);
>>>> +            ret = 0;
>>>> +        }
>>>> +    }
>>>> +    if (!delivered || !reference) {
>>>> +        cpc_read(cpunum, delivered_reg, &delivered);
>>>> +        cpc_read(cpunum, reference_reg, &reference);
>>>> +    }
>>>> +
>>>>      cpc_read(cpunum, ref_perf_reg, &ref_perf);
>>>>      /*
>>>> diff --git a/include/acpi/cppc_acpi.h b/include/acpi/cppc_acpi.h
>>>> index 6126c977ece0..07d4fd82d499 100644
>>>> --- a/include/acpi/cppc_acpi.h
>>>> +++ b/include/acpi/cppc_acpi.h
>>>> @@ -152,6 +152,7 @@ extern bool cpc_ffh_supported(void);
>>>>  extern bool cpc_supported_by_cpu(void);
>>>>  extern int cpc_read_ffh(int cpunum, struct cpc_reg *reg, u64 *val);
>>>>  extern int cpc_write_ffh(int cpunum, struct cpc_reg *reg, u64 val);
>>>> +extern int cpc_read_arch_counters_on_cpu(int cpu, u64 *delivered, 
>>>> u64 *reference);
>>>>  extern int cppc_get_epp_perf(int cpunum, u64 *epp_perf);
>>>>  extern int cppc_set_epp_perf(int cpu, struct cppc_perf_ctrls 
>>>> *perf_ctrls, bool enable);
>>>>  extern int cppc_get_auto_sel_caps(int cpunum, struct 
>>>> cppc_perf_caps *perf_caps);
>>>> @@ -209,6 +210,10 @@ static inline int cpc_write_ffh(int cpunum, 
>>>> struct cpc_reg *reg, u64 val)
>>>>  {
>>>>      return -ENOTSUPP;
>>>>  }
>>>> +static inline int cpc_read_arch_counters_on_cpu(int cpu, u64 
>>>> *delivered, u64 *reference)
>>>> +{
>>>> +    return -EOPNOTSUPP;
>>>> +}
>>>>  static inline int cppc_set_epp_perf(int cpu, struct 
>>>> cppc_perf_ctrls *perf_ctrls, bool enable)
>>>>  {
>>>>      return -ENOTSUPP;
>>>> -- 
>>>> 2.33.0
>>>>
>>> .
>>
>> _______________________________________________
>> linux-arm-kernel mailing list
>> linux-arm-kernel@lists.infradead.org
>> http://lists.infradead.org/mailman/listinfo/linux-arm-kernel
>
> .
Ionela Voinescu Jan. 8, 2024, 2:03 p.m. UTC | #11
Hi,

On Friday 05 Jan 2024 at 15:04:47 (+0800), lihuisong (C) wrote:
> Hi Vanshi,
> 
> 在 2024/1/5 8:48, Vanshidhar Konda 写道:
> > On Thu, Jan 04, 2024 at 05:36:51PM +0800, lihuisong (C) wrote:
> > > 
> > > 在 2024/1/4 1:53, Ionela Voinescu 写道:
> > > > Hi,
> > > > 
> > > > On Tuesday 12 Dec 2023 at 15:26:17 (+0800), Huisong Li wrote:
> > > > > Many developers found that the cpu current frequency is greater than
> > > > > the maximum frequency of the platform, please see [1], [2] and [3].
> > > > > 
> > > > > In the scenarios with high memory access pressure, the patch [1] has
> > > > > proved the significant latency of cpc_read() which is used to obtain
> > > > > delivered and reference performance counter cause an absurd frequency.
> > > > > The sampling interval for this counters is very critical and
> > > > > is expected
> > > > > to be equal. However, the different latency of cpc_read() has a direct
> > > > > impact on their sampling interval.
> > > > > 
> > > > Would this [1] alternative solution work for you?
> > > It would work for me AFAICS.
> > > Because the "arch_freq_scale" is also from AMU core and constant
> > > counter, and read together.
> > > But, from their discuss line, it seems that there are some tricky
> > > points to clarify or consider.
> > 
> > I think the changes in [1] would work better when CPUs may be idle. With
> > this
> > patch we would have to wake any core that is in idle state to read the
> > AMU
> > counters. Worst case, if core 0 is trying to read the CPU frequency of
> > all
> > cores, it may need to wake up all the other cores to read the AMU
> > counters.
> From the approach in [1], if all CPUs (one or more cores) under one policy
> are idle, they still cannot be obtained the CPU frequency, right?
> In this case, the [1] API will return 0 and have to back to call
> cpufreq_driver->get() for cpuinfo_cur_freq.
> Then we still need to face the issue this patch mentioned.

With the implementation at [1], arch_freq_get_on_cpu() will not return 0
for idle CPUs and the get() callback will not be called to wake up the
CPUs.

Worst case, arch_freq_get_on_cpu() will return a frequency based on the
AMU counter values obtained on the last tick on that CPU. But if that CPU
is not a housekeeping CPU, a housekeeping CPU in the same policy will be
selected, as it would have had a more recent tick, and therefore a more
recent frequency value for the domain.

I understand that the frequency returned here will not be up to date,
but there's no proper frequency feedback for an idle CPU. If one only
wakes up a CPU to sample counters, before the CPU goes back to sleep,
the obtained frequency feedback is meaningless.

> > For systems with 128 cores or more, this could be very expensive and
> > happen
> > very frequently.
> > 
> > AFAICS, the approach in [1] would avoid this cost.
> But the CPU frequency is just an average value for the last tick period
> instead of the current one the CPU actually runs at.
> In addition, there are some conditions to use 'arch_freq_scale' in this
> approach.

What are the conditions you are referring to?

> So I'm not sure if this approach can entirely cover the frequency
> discrepancy issue.

Unfortunately there is no perfect frequency feedback. By the time you
observe/use the value of scaling_cur_freq/cpuinfo_cur_freq, the frequency
of the CPU might have already changed. Therefore, an average value might
be a better indication of the recent performance level of a CPU.

Would you be able to test [1] on your platform and usecase?

Many thanks,
Ionela.

> 
> /Huisong
> 
> > > > 
> > > > [1] https://lore.kernel.org/lkml/20231127160838.1403404-1-beata.michalska@arm.com/
> > > > 
> > > > Thanks,
> > > > Ionela.
> > > > 
> > > > > This patch adds a interface, cpc_read_arch_counters_on_cpu, to read
> > > > > delivered and reference performance counter together. According to my
> > > > > test[4], the discrepancy of cpu current frequency in the
> > > > > scenarios with
> > > > > high memory access pressure is lower than 0.2% by stress-ng
> > > > > application.
> > > > > 
> > > > > [1] https://lore.kernel.org/all/20231025093847.3740104-4-zengheng4@huawei.com/
> > > > > [2] https://lore.kernel.org/all/20230328193846.8757-1-yang@os.amperecomputing.com/
> > > > > [3]
> > > > > https://lore.kernel.org/all/20230418113459.12860-7-sumitg@nvidia.com/
> > > > > 
> > > > > [4] My local test:
> > > > > The testing platform enable SMT and include 128 logical CPU in total,
> > > > > and CPU base frequency is 2.7GHz. Reading "cpuinfo_cur_freq" for each
> > > > > physical core on platform during the high memory access pressure from
> > > > > stress-ng, and the output is as follows:
> > > > >   0: 2699133     2: 2699942     4: 2698189     6: 2704347
> > > > >   8: 2704009    10: 2696277    12: 2702016    14: 2701388
> > > > >  16: 2700358    18: 2696741    20: 2700091    22: 2700122
> > > > >  24: 2701713    26: 2702025    28: 2699816    30: 2700121
> > > > >  32: 2700000    34: 2699788    36: 2698884    38: 2699109
> > > > >  40: 2704494    42: 2698350    44: 2699997    46: 2701023
> > > > >  48: 2703448    50: 2699501    52: 2700000    54: 2699999
> > > > >  56: 2702645    58: 2696923    60: 2697718    62: 2700547
> > > > >  64: 2700313    66: 2700000    68: 2699904    70: 2699259
> > > > >  72: 2699511    74: 2700644    76: 2702201    78: 2700000
> > > > >  80: 2700776    82: 2700364    84: 2702674    86: 2700255
> > > > >  88: 2699886    90: 2700359    92: 2699662    94: 2696188
> > > > >  96: 2705454    98: 2699260   100: 2701097   102: 2699630
> > > > > 104: 2700463   106: 2698408   108: 2697766   110: 2701181
> > > > > 112: 2699166   114: 2701804   116: 2701907   118: 2701973
> > > > > 120: 2699584   122: 2700474   124: 2700768   126: 2701963
> > > > > 
> > > > > Signed-off-by: Huisong Li <lihuisong@huawei.com>
> > > > > ---
> > > > >  arch/arm64/kernel/topology.c | 43
> > > > > ++++++++++++++++++++++++++++++++++--
> > > > >  drivers/acpi/cppc_acpi.c     | 22 +++++++++++++++---
> > > > >  include/acpi/cppc_acpi.h     |  5 +++++
> > > > >  3 files changed, 65 insertions(+), 5 deletions(-)
> > > > > 
> > > > > diff --git a/arch/arm64/kernel/topology.c
> > > > > b/arch/arm64/kernel/topology.c
> > > > > index 7d37e458e2f5..c3122154d738 100644
> > > > > --- a/arch/arm64/kernel/topology.c
> > > > > +++ b/arch/arm64/kernel/topology.c
> > > > > @@ -299,6 +299,11 @@ core_initcall(init_amu_fie);
> > > > >  #ifdef CONFIG_ACPI_CPPC_LIB
> > > > >  #include <acpi/cppc_acpi.h>
> > > > > +struct amu_counters {
> > > > > +    u64 corecnt;
> > > > > +    u64 constcnt;
> > > > > +};
> > > > > +
> > > > >  static void cpu_read_corecnt(void *val)
> > > > >  {
> > > > >      /*
> > > > > @@ -322,8 +327,27 @@ static void cpu_read_constcnt(void *val)
> > > > >                0UL : read_constcnt();
> > > > >  }
> > > > > +static void cpu_read_amu_counters(void *data)
> > > > > +{
> > > > > +    struct amu_counters *cnt = (struct amu_counters *)data;
> > > > > +
> > > > > +    /*
> > > > > +     * The running time of the this_cpu_has_cap() might
> > > > > have a couple of
> > > > > +     * microseconds and is significantly increased to tens
> > > > > of microseconds.
> > > > > +     * But AMU core and constant counter need to be read
> > > > > togeter without any
> > > > > +     * time interval to reduce the calculation discrepancy
> > > > > using this counters.
> > > > > +     */
> > > > > +    if (this_cpu_has_cap(ARM64_WORKAROUND_2457168)) {
> > > > > +        cnt->corecnt = read_corecnt();
> > > > > +        cnt->constcnt = 0;
> > > > > +    } else {
> > > > > +        cnt->corecnt = read_corecnt();
> > > > > +        cnt->constcnt = read_constcnt();
> > > > > +    }
> > > > > +}
> > > > > +
> > > > >  static inline
> > > > > -int counters_read_on_cpu(int cpu, smp_call_func_t func, u64 *val)
> > > > > +int counters_read_on_cpu(int cpu, smp_call_func_t func, void *data)
> > > > >  {
> > > > >      /*
> > > > >       * Abort call on counterless CPU or when interrupts are
> > > > > @@ -335,7 +359,7 @@ int counters_read_on_cpu(int cpu,
> > > > > smp_call_func_t func, u64 *val)
> > > > >      if (WARN_ON_ONCE(irqs_disabled()))
> > > > >          return -EPERM;
> > > > > -    smp_call_function_single(cpu, func, val, 1);
> > > > > +    smp_call_function_single(cpu, func, data, 1);
> > > > >      return 0;
> > > > >  }
> > > > > @@ -364,6 +388,21 @@ bool cpc_ffh_supported(void)
> > > > >      return true;
> > > > >  }
> > > > > +int cpc_read_arch_counters_on_cpu(int cpu, u64 *delivered,
> > > > > u64 *reference)
> > > > > +{
> > > > > +    struct amu_counters cnts = {0};
> > > > > +    int ret;
> > > > > +
> > > > > +    ret = counters_read_on_cpu(cpu, cpu_read_amu_counters, &cnts);
> > > > > +    if (ret)
> > > > > +        return ret;
> > > > > +
> > > > > +    *delivered = cnts.corecnt;
> > > > > +    *reference = cnts.constcnt;
> > > > > +
> > > > > +    return 0;
> > > > > +}
> > > > > +
> > > > >  int cpc_read_ffh(int cpu, struct cpc_reg *reg, u64 *val)
> > > > >  {
> > > > >      int ret = -EOPNOTSUPP;
> > > > > diff --git a/drivers/acpi/cppc_acpi.c b/drivers/acpi/cppc_acpi.c
> > > > > index 7ff269a78c20..f303fabd7cfe 100644
> > > > > --- a/drivers/acpi/cppc_acpi.c
> > > > > +++ b/drivers/acpi/cppc_acpi.c
> > > > > @@ -1299,6 +1299,11 @@ bool cppc_perf_ctrs_in_pcc(void)
> > > > >  }
> > > > >  EXPORT_SYMBOL_GPL(cppc_perf_ctrs_in_pcc);
> > > > > +int __weak cpc_read_arch_counters_on_cpu(int cpu, u64
> > > > > *delivered, u64 *reference)
> > > > > +{
> > > > > +    return 0;
> > > > > +}
> > > > > +
> > > > >  /**
> > > > >   * cppc_get_perf_ctrs - Read a CPU's performance feedback counters.
> > > > >   * @cpunum: CPU from which to read counters.
> > > > > @@ -1313,7 +1318,8 @@ int cppc_get_perf_ctrs(int cpunum,
> > > > > struct cppc_perf_fb_ctrs *perf_fb_ctrs)
> > > > >          *ref_perf_reg, *ctr_wrap_reg;
> > > > >      int pcc_ss_id = per_cpu(cpu_pcc_subspace_idx, cpunum);
> > > > >      struct cppc_pcc_data *pcc_ss_data = NULL;
> > > > > -    u64 delivered, reference, ref_perf, ctr_wrap_time;
> > > > > +    u64 delivered = 0, reference = 0;
> > > > > +    u64 ref_perf, ctr_wrap_time;
> > > > >      int ret = 0, regs_in_pcc = 0;
> > > > >      if (!cpc_desc) {
> > > > > @@ -1350,8 +1356,18 @@ int cppc_get_perf_ctrs(int cpunum,
> > > > > struct cppc_perf_fb_ctrs *perf_fb_ctrs)
> > > > >          }
> > > > >      }
> > > > > -    cpc_read(cpunum, delivered_reg, &delivered);
> > > > > -    cpc_read(cpunum, reference_reg, &reference);
> > > > > +    if (cpc_ffh_supported()) {
> > > > > +        ret = cpc_read_arch_counters_on_cpu(cpunum,
> > > > > &delivered, &reference);
> > > > > +        if (ret) {
> > > > > +            pr_debug("read arch counters failed, ret=%d.\n", ret);
> > > > > +            ret = 0;
> > > > > +        }
> > > > > +    }
> > > > > +    if (!delivered || !reference) {
> > > > > +        cpc_read(cpunum, delivered_reg, &delivered);
> > > > > +        cpc_read(cpunum, reference_reg, &reference);
> > > > > +    }
> > > > > +
> > > > >      cpc_read(cpunum, ref_perf_reg, &ref_perf);
> > > > >      /*
> > > > > diff --git a/include/acpi/cppc_acpi.h b/include/acpi/cppc_acpi.h
> > > > > index 6126c977ece0..07d4fd82d499 100644
> > > > > --- a/include/acpi/cppc_acpi.h
> > > > > +++ b/include/acpi/cppc_acpi.h
> > > > > @@ -152,6 +152,7 @@ extern bool cpc_ffh_supported(void);
> > > > >  extern bool cpc_supported_by_cpu(void);
> > > > >  extern int cpc_read_ffh(int cpunum, struct cpc_reg *reg, u64 *val);
> > > > >  extern int cpc_write_ffh(int cpunum, struct cpc_reg *reg, u64 val);
> > > > > +extern int cpc_read_arch_counters_on_cpu(int cpu, u64
> > > > > *delivered, u64 *reference);
> > > > >  extern int cppc_get_epp_perf(int cpunum, u64 *epp_perf);
> > > > >  extern int cppc_set_epp_perf(int cpu, struct
> > > > > cppc_perf_ctrls *perf_ctrls, bool enable);
> > > > >  extern int cppc_get_auto_sel_caps(int cpunum, struct
> > > > > cppc_perf_caps *perf_caps);
> > > > > @@ -209,6 +210,10 @@ static inline int cpc_write_ffh(int
> > > > > cpunum, struct cpc_reg *reg, u64 val)
> > > > >  {
> > > > >      return -ENOTSUPP;
> > > > >  }
> > > > > +static inline int cpc_read_arch_counters_on_cpu(int cpu,
> > > > > u64 *delivered, u64 *reference)
> > > > > +{
> > > > > +    return -EOPNOTSUPP;
> > > > > +}
> > > > >  static inline int cppc_set_epp_perf(int cpu, struct
> > > > > cppc_perf_ctrls *perf_ctrls, bool enable)
> > > > >  {
> > > > >      return -ENOTSUPP;
> > > > > -- 
> > > > > 2.33.0
> > > > > 
> > > > .
> > > 
> > > _______________________________________________
> > > linux-arm-kernel mailing list
> > > linux-arm-kernel@lists.infradead.org
> > > http://lists.infradead.org/mailman/listinfo/linux-arm-kernel
> > 
> > .
lihuisong (C) Jan. 10, 2024, 7:09 a.m. UTC | #12
Hi Ionela,

在 2024/1/8 22:03, Ionela Voinescu 写道:
> Hi,
>
> On Friday 05 Jan 2024 at 15:04:47 (+0800), lihuisong (C) wrote:
>> Hi Vanshi,
>>
>> 在 2024/1/5 8:48, Vanshidhar Konda 写道:
>>> On Thu, Jan 04, 2024 at 05:36:51PM +0800, lihuisong (C) wrote:
>>>> 在 2024/1/4 1:53, Ionela Voinescu 写道:
>>>>> Hi,
>>>>>
>>>>> On Tuesday 12 Dec 2023 at 15:26:17 (+0800), Huisong Li wrote:
>>>>>> Many developers found that the cpu current frequency is greater than
>>>>>> the maximum frequency of the platform, please see [1], [2] and [3].
>>>>>>
>>>>>> In the scenarios with high memory access pressure, the patch [1] has
>>>>>> proved the significant latency of cpc_read() which is used to obtain
>>>>>> delivered and reference performance counter cause an absurd frequency.
>>>>>> The sampling interval for this counters is very critical and
>>>>>> is expected
>>>>>> to be equal. However, the different latency of cpc_read() has a direct
>>>>>> impact on their sampling interval.
>>>>>>
>>>>> Would this [1] alternative solution work for you?
>>>> It would work for me AFAICS.
>>>> Because the "arch_freq_scale" is also from AMU core and constant
>>>> counter, and read together.
>>>> But, from their discuss line, it seems that there are some tricky
>>>> points to clarify or consider.
>>> I think the changes in [1] would work better when CPUs may be idle. With
>>> this
>>> patch we would have to wake any core that is in idle state to read the
>>> AMU
>>> counters. Worst case, if core 0 is trying to read the CPU frequency of
>>> all
>>> cores, it may need to wake up all the other cores to read the AMU
>>> counters.
>>  From the approach in [1], if all CPUs (one or more cores) under one policy
>> are idle, they still cannot be obtained the CPU frequency, right?
>> In this case, the [1] API will return 0 and have to back to call
>> cpufreq_driver->get() for cpuinfo_cur_freq.
>> Then we still need to face the issue this patch mentioned.
> With the implementation at [1], arch_freq_get_on_cpu() will not return 0
> for idle CPUs and the get() callback will not be called to wake up the
> CPUs.
Right, arch_freq_get_on_cpu() will not return 0 for idle CPUs.
However, for no-housekeeping CPUs, it will return 0 and have to call 
get() callback, right?
>
> Worst case, arch_freq_get_on_cpu() will return a frequency based on the
> AMU counter values obtained on the last tick on that CPU. But if that CPU
> is not a housekeeping CPU, a housekeeping CPU in the same policy will be
> selected, as it would have had a more recent tick, and therefore a more
> recent frequency value for the domain.
But this frequency is from the last tick,
this last tick is probably a long time ago and it doesn't update 
'arch_freq_scale' for some reasons like CPU dile.
In addition, I'm not sure if there is possible that 
amu_scale_freq_tick() is executed delayed under high stress case.
It also have an impact on the accuracy of the cpu frequency we query.
>
> I understand that the frequency returned here will not be up to date,
> but there's no proper frequency feedback for an idle CPU. If one only
> wakes up a CPU to sample counters, before the CPU goes back to sleep,
> the obtained frequency feedback is meaningless.
>
>>> For systems with 128 cores or more, this could be very expensive and
>>> happen
>>> very frequently.
>>>
>>> AFAICS, the approach in [1] would avoid this cost.
>> But the CPU frequency is just an average value for the last tick period
>> instead of the current one the CPU actually runs at.
>> In addition, there are some conditions to use 'arch_freq_scale' in this
>> approach.
> What are the conditions you are referring to?
It depends on the housekeeping CPUs.
>
>> So I'm not sure if this approach can entirely cover the frequency
>> discrepancy issue.
> Unfortunately there is no perfect frequency feedback. By the time you
> observe/use the value of scaling_cur_freq/cpuinfo_cur_freq, the frequency
> of the CPU might have already changed. Therefore, an average value might
> be a better indication of the recent performance level of a CPU.
An average value for CPU frequency is ok. It may be better if it has not 
any delaying.

The original implementation for cpuinfo_cur_freq can more reflect their
meaning in the user-guide [1]. The user-guide said:
"cpuinfo_cur_freq : Current frequency of the CPU as obtained from the 
hardware, in KHz.
This is the frequency the CPU actually runs at."


[1]https://www.kernel.org/doc/Documentation/cpu-freq/user-guide.txt

>
> Would you be able to test [1] on your platform and usecase?
I has tested it on my platform (CPU number: 64, SMT: off and CPU base 
frequency: 2.7GHz).
Accoding to the testing result,
1> I found that patch [1] and [2] cannot cover the no housekeeping CPUs. 
They still have to face the large frequency discrepancy issue my patch 
mentioned.
2> Additionally, the frequency value of all CPUs are almost the same by 
using the 'arch_freq_scale' factor way. I'm not sure if it is ok.

The patch [1] has been modified silightly as below:
-->
@@ -1756,7 +1756,10 @@ static unsigned int 
cpufreq_verify_current_freq(struct cpufreq_policy *policy, b
  {
         unsigned int new_freq;

-       new_freq = cpufreq_driver->get(policy->cpu);
+       new_freq = arch_freq_get_on_cpu(policy->cpu);
+       if (!new_freq)
+               new_freq = cpufreq_driver->get(policy->cpu);
+
         if (!new_freq)
                 return 0;

And the result is as follows:
*case 1:**No setting the nohz_full and cpufreq use performance governor*
*--> Step1: *read 'cpuinfo_cur_freq' in no pressure
   0: 2699264     2: 2699264     4: 2699264     6: 2699264
   8: 2696628    10: 2696628    12: 2696628    14: 2699264
  16: 2699264    18: 2696628    20: 2699264    22: 2696628
  24: 2699264    26: 2696628    28: 2699264    30: 2696628
  32: 2696628    34: 2696628    36: 2696628    38: 2696628
  40: 2699264    42: 2699264    44: 2696628    46: 2696628
  48: 2696628    50: 2699264    52: 2699264    54: 2696628
  56: 2696628    58: 2696628    60: 2696628    62: 2696628
  64: 2696628    66: 2699264    68: 2696628    70: 2696628
  72: 2699264    74: 2696628    76: 2696628    78: 2699264
  80: 2696628    82: 2696628    84: 2699264    86: 2696628
  88: 2696628    90: 2696628    92: 2696628    94: 2699264
  96: 2696628    98: 2699264   100: 2699264   102: 2696628
104: 2699264   106: 2699264   108: 2699264   110: 2696628
112: 2699264   114: 2699264   116: 2699264   118: 2699264
120: 2696628   122: 2699264   124: 2696628   126: 2699264
Note: the frequency of all CPUs are almost the same.

*--> Step 2: *read 'cpuinfo_cur_freq' in the high memory access pressure.
   0: 2696628     2: 2696628     4: 2696628     6: 2696628
   8: 2696628    10: 2696628    12: 2696628    14: 2696628
  16: 2696628    18: 2696628    20: 2696628    22: 2696628
  24: 2696628    26: 2696628    28: 2696628    30: 2696628
  32: 2696628    34: 2696628    36: 2696628    38: 2696628
  40: 2696628    42: 2696628    44: 2696628    46: 2696628
  48: 2696628    50: 2696628    52: 2696628    54: 2696628
  56: 2696628    58: 2696628    60: 2696628    62: 2696628
  64: 2696628    66: 2696628    68: 2696628    70: 2696628
  72: 2696628    74: 2696628    76: 2696628    78: 2696628
  80: 2696628    82: 2696628    84: 2696628    86: 2696628
  88: 2696628    90: 2696628    92: 2696628    94: 2696628
  96: 2696628    98: 2696628   100: 2696628   102: 2696628
104: 2696628   106: 2696628   108: 2696628   110: 2696628
112: 2696628   114: 2696628   116: 2696628   118: 2696628
120: 2696628   122: 2696628   124: 2696628   126: 2696628

*Case 2: setting nohz_full and cpufreq use ondemand governor*
There is "isolcpus=1-10,41-50 nohz_full=1-10,41-50 rcu_nocbs=1-10,41-50" 
in /proc/cmdline.
*--> Step 1: *setting ondemand governor to all policy and query 
'cpuinfo_cur_freq' in no pressure case.
And the frequency of CPUs all are about 400MHz.
*--> Step 2:* read 'cpuinfo_cur_freq' in the high memory access pressure.
The high memory access pressure is from the command: "stress-ng -c 64 
--cpu-load 100% --taskset 0-63"
The result:
  0: 2696628     1:  400000     2:  400000     3:  400909
  4:  400000     5:  400000     6:  400000     7:  400000
  8:  400000     9:  400000    10:  400600    11: 2696628
12: 2696628    13: 2696628    14: 2696628    15: 2696628
16: 2696628    17: 2696628    18: 2696628    19: 2696628
20: 2696628    21: 2696628    22: 2696628    23: 2696628
24: 2696628    25: 2696628    26: 2696628    27: 2696628
28: 2696628    29: 2696628    30: 2696628    31: 2696628
32: 2696628    33: 2696628    34: 2696628    35: 2696628
36: 2696628    37: 2696628    38: 2696628    39: 2696628
40: 2696628    41:  400000    42:  400000    43:  400000
44:  400000    45:  398847    46:  400000    47:  400000
48:  400000    49:  400000    50:  400000    51: 2696628
52: 2696628    53: 2696628    54: 2696628    55: 2696628
56: 2696628    57: 2696628    58: 2696628    59: 2696628
60: 2696628    61: 2696628    62: 2696628    63: 2699264

Note:
(1) The frequency of 1-10 and 41-50 CPUs work on the lowest frequency.
      It turned out that nohz full was already work.
      I guess that stress-ng cannot use the CPU in the range of nohz full.
      Because the CPU frequency will be increased to 2.7G by binding CPU 
to other application.
(2) The frequency of the nohz full core is calculated by get() callback 
according to ftrace.

[1] https://lore.kernel.org/lkml/20230418113459.12860-7-sumitg@nvidia.com/
[2] 
https://lore.kernel.org/lkml/20231127160838.1403404-3-beata.michalska@arm.com/
>
> Many thanks,
> Ionela.
>
>> /Huisong
>>
>>>>> [1] https://lore.kernel.org/lkml/20231127160838.1403404-1-beata.michalska@arm.com/
>>>>>
>>>>> Thanks,
>>>>> Ionela.
>>>>>
>>>>>> This patch adds a interface, cpc_read_arch_counters_on_cpu, to read
>>>>>> delivered and reference performance counter together. According to my
>>>>>> test[4], the discrepancy of cpu current frequency in the
>>>>>> scenarios with
>>>>>> high memory access pressure is lower than 0.2% by stress-ng
>>>>>> application.
>>>>>>
>>>>>> [1] https://lore.kernel.org/all/20231025093847.3740104-4-zengheng4@huawei.com/
>>>>>> [2] https://lore.kernel.org/all/20230328193846.8757-1-yang@os.amperecomputing.com/
>>>>>> [3]
>>>>>> https://lore.kernel.org/all/20230418113459.12860-7-sumitg@nvidia.com/
>>>>>>
>>>>>> [4] My local test:
>>>>>> The testing platform enable SMT and include 128 logical CPU in total,
>>>>>> and CPU base frequency is 2.7GHz. Reading "cpuinfo_cur_freq" for each
>>>>>> physical core on platform during the high memory access pressure from
>>>>>> stress-ng, and the output is as follows:
>>>>>>    0: 2699133     2: 2699942     4: 2698189     6: 2704347
>>>>>>    8: 2704009    10: 2696277    12: 2702016    14: 2701388
>>>>>>   16: 2700358    18: 2696741    20: 2700091    22: 2700122
>>>>>>   24: 2701713    26: 2702025    28: 2699816    30: 2700121
>>>>>>   32: 2700000    34: 2699788    36: 2698884    38: 2699109
>>>>>>   40: 2704494    42: 2698350    44: 2699997    46: 2701023
>>>>>>   48: 2703448    50: 2699501    52: 2700000    54: 2699999
>>>>>>   56: 2702645    58: 2696923    60: 2697718    62: 2700547
>>>>>>   64: 2700313    66: 2700000    68: 2699904    70: 2699259
>>>>>>   72: 2699511    74: 2700644    76: 2702201    78: 2700000
>>>>>>   80: 2700776    82: 2700364    84: 2702674    86: 2700255
>>>>>>   88: 2699886    90: 2700359    92: 2699662    94: 2696188
>>>>>>   96: 2705454    98: 2699260   100: 2701097   102: 2699630
>>>>>> 104: 2700463   106: 2698408   108: 2697766   110: 2701181
>>>>>> 112: 2699166   114: 2701804   116: 2701907   118: 2701973
>>>>>> 120: 2699584   122: 2700474   124: 2700768   126: 2701963
>>>>>>
>>>>>> Signed-off-by: Huisong Li <lihuisong@huawei.com>
>>>>>> ---
>>>>>>   
[snip]
> .
Vanshidhar Konda Jan. 12, 2024, 6:33 p.m. UTC | #13
On Wed, Jan 10, 2024 at 03:09:48PM +0800, lihuisong (C) wrote:
>Hi Ionela,
>
>在 2024/1/8 22:03, Ionela Voinescu 写道:
>>Hi,
>>
>>On Friday 05 Jan 2024 at 15:04:47 (+0800), lihuisong (C) wrote:
>>>Hi Vanshi,
>>>
>>>在 2024/1/5 8:48, Vanshidhar Konda 写道:
>>>>On Thu, Jan 04, 2024 at 05:36:51PM +0800, lihuisong (C) wrote:
>>>>>在 2024/1/4 1:53, Ionela Voinescu 写道:
>>>>>>Hi,
>>>>>>
>>>>>>On Tuesday 12 Dec 2023 at 15:26:17 (+0800), Huisong Li wrote:
>>>>>>>Many developers found that the cpu current frequency is greater than
>>>>>>>the maximum frequency of the platform, please see [1], [2] and [3].
>>>>>>>
>>>>>>>In the scenarios with high memory access pressure, the patch [1] has
>>>>>>>proved the significant latency of cpc_read() which is used to obtain
>>>>>>>delivered and reference performance counter cause an absurd frequency.
>>>>>>>The sampling interval for this counters is very critical and
>>>>>>>is expected
>>>>>>>to be equal. However, the different latency of cpc_read() has a direct
>>>>>>>impact on their sampling interval.
>>>>>>>
>>>>>>Would this [1] alternative solution work for you?
>>>>>It would work for me AFAICS.
>>>>>Because the "arch_freq_scale" is also from AMU core and constant
>>>>>counter, and read together.
>>>>>But, from their discuss line, it seems that there are some tricky
>>>>>points to clarify or consider.
>>>>I think the changes in [1] would work better when CPUs may be idle. With
>>>>this
>>>>patch we would have to wake any core that is in idle state to read the
>>>>AMU
>>>>counters. Worst case, if core 0 is trying to read the CPU frequency of
>>>>all
>>>>cores, it may need to wake up all the other cores to read the AMU
>>>>counters.
>>> From the approach in [1], if all CPUs (one or more cores) under one policy
>>>are idle, they still cannot be obtained the CPU frequency, right?
>>>In this case, the [1] API will return 0 and have to back to call
>>>cpufreq_driver->get() for cpuinfo_cur_freq.
>>>Then we still need to face the issue this patch mentioned.
>>With the implementation at [1], arch_freq_get_on_cpu() will not return 0
>>for idle CPUs and the get() callback will not be called to wake up the
>>CPUs.
>Right, arch_freq_get_on_cpu() will not return 0 for idle CPUs.
>However, for no-housekeeping CPUs, it will return 0 and have to call 
>get() callback, right?
>>
>>Worst case, arch_freq_get_on_cpu() will return a frequency based on the
>>AMU counter values obtained on the last tick on that CPU. But if that CPU
>>is not a housekeeping CPU, a housekeeping CPU in the same policy will be
>>selected, as it would have had a more recent tick, and therefore a more
>>recent frequency value for the domain.
>But this frequency is from the last tick,
>this last tick is probably a long time ago and it doesn't update 
>'arch_freq_scale' for some reasons like CPU dile.
>In addition, I'm not sure if there is possible that 
>amu_scale_freq_tick() is executed delayed under high stress case.
>It also have an impact on the accuracy of the cpu frequency we query.
>>
>>I understand that the frequency returned here will not be up to date,
>>but there's no proper frequency feedback for an idle CPU. If one only
>>wakes up a CPU to sample counters, before the CPU goes back to sleep,
>>the obtained frequency feedback is meaningless.
>>
>>>>For systems with 128 cores or more, this could be very expensive and
>>>>happen
>>>>very frequently.
>>>>
>>>>AFAICS, the approach in [1] would avoid this cost.
>>>But the CPU frequency is just an average value for the last tick period
>>>instead of the current one the CPU actually runs at.
>>>In addition, there are some conditions to use 'arch_freq_scale' in this
>>>approach.
>>What are the conditions you are referring to?
>It depends on the housekeeping CPUs.
>>
>>>So I'm not sure if this approach can entirely cover the frequency
>>>discrepancy issue.
>>Unfortunately there is no perfect frequency feedback. By the time you
>>observe/use the value of scaling_cur_freq/cpuinfo_cur_freq, the frequency
>>of the CPU might have already changed. Therefore, an average value might
>>be a better indication of the recent performance level of a CPU.
>An average value for CPU frequency is ok. It may be better if it has 
>not any delaying.
>
>The original implementation for cpuinfo_cur_freq can more reflect their
>meaning in the user-guide [1]. The user-guide said:
>"cpuinfo_cur_freq : Current frequency of the CPU as obtained from the 
>hardware, in KHz.
>This is the frequency the CPU actually runs at."
>
>
>[1]https://www.kernel.org/doc/Documentation/cpu-freq/user-guide.txt
>
>>
>>Would you be able to test [1] on your platform and usecase?
>I has tested it on my platform (CPU number: 64, SMT: off and CPU base 
>frequency: 2.7GHz).
>Accoding to the testing result,
>1> I found that patch [1] and [2] cannot cover the no housekeeping 
>CPUs. They still have to face the large frequency discrepancy issue my 
>patch mentioned.
>2> Additionally, the frequency value of all CPUs are almost the same 
>by using the 'arch_freq_scale' factor way. I'm not sure if it is ok.
>
>The patch [1] has been modified silightly as below:
>-->
>@@ -1756,7 +1756,10 @@ static unsigned int 
>cpufreq_verify_current_freq(struct cpufreq_policy *policy, b
> {
>        unsigned int new_freq;
>
>-       new_freq = cpufreq_driver->get(policy->cpu);
>+       new_freq = arch_freq_get_on_cpu(policy->cpu);
>+       if (!new_freq)
>+               new_freq = cpufreq_driver->get(policy->cpu);
>+
>        if (!new_freq)
>                return 0;
>
>And the result is as follows:
>*case 1:**No setting the nohz_full and cpufreq use performance governor*
>*--> Step1: *read 'cpuinfo_cur_freq' in no pressure
>  0: 2699264     2: 2699264     4: 2699264     6: 2699264
>  8: 2696628    10: 2696628    12: 2696628    14: 2699264
> 16: 2699264    18: 2696628    20: 2699264    22: 2696628
> 24: 2699264    26: 2696628    28: 2699264    30: 2696628
> 32: 2696628    34: 2696628    36: 2696628    38: 2696628
> 40: 2699264    42: 2699264    44: 2696628    46: 2696628
> 48: 2696628    50: 2699264    52: 2699264    54: 2696628
> 56: 2696628    58: 2696628    60: 2696628    62: 2696628
> 64: 2696628    66: 2699264    68: 2696628    70: 2696628
> 72: 2699264    74: 2696628    76: 2696628    78: 2699264
> 80: 2696628    82: 2696628    84: 2699264    86: 2696628
> 88: 2696628    90: 2696628    92: 2696628    94: 2699264
> 96: 2696628    98: 2699264   100: 2699264   102: 2696628
>104: 2699264   106: 2699264   108: 2699264   110: 2696628
>112: 2699264   114: 2699264   116: 2699264   118: 2699264
>120: 2696628   122: 2699264   124: 2696628   126: 2699264
>Note: the frequency of all CPUs are almost the same.
>
>*--> Step 2: *read 'cpuinfo_cur_freq' in the high memory access pressure.
>  0: 2696628     2: 2696628     4: 2696628     6: 2696628
>  8: 2696628    10: 2696628    12: 2696628    14: 2696628
> 16: 2696628    18: 2696628    20: 2696628    22: 2696628
> 24: 2696628    26: 2696628    28: 2696628    30: 2696628
> 32: 2696628    34: 2696628    36: 2696628    38: 2696628
> 40: 2696628    42: 2696628    44: 2696628    46: 2696628
> 48: 2696628    50: 2696628    52: 2696628    54: 2696628
> 56: 2696628    58: 2696628    60: 2696628    62: 2696628
> 64: 2696628    66: 2696628    68: 2696628    70: 2696628
> 72: 2696628    74: 2696628    76: 2696628    78: 2696628
> 80: 2696628    82: 2696628    84: 2696628    86: 2696628
> 88: 2696628    90: 2696628    92: 2696628    94: 2696628
> 96: 2696628    98: 2696628   100: 2696628   102: 2696628
>104: 2696628   106: 2696628   108: 2696628   110: 2696628
>112: 2696628   114: 2696628   116: 2696628   118: 2696628
>120: 2696628   122: 2696628   124: 2696628   126: 2696628
>
>*Case 2: setting nohz_full and cpufreq use ondemand governor*
>There is "isolcpus=1-10,41-50 nohz_full=1-10,41-50 
>rcu_nocbs=1-10,41-50" in /proc/cmdline.
>*--> Step 1: *setting ondemand governor to all policy and query 
>'cpuinfo_cur_freq' in no pressure case.
>And the frequency of CPUs all are about 400MHz.
>*--> Step 2:* read 'cpuinfo_cur_freq' in the high memory access pressure.
>The high memory access pressure is from the command: "stress-ng -c 64 
>--cpu-load 100% --taskset 0-63"
>The result:
> 0: 2696628     1:  400000     2:  400000     3:  400909
> 4:  400000     5:  400000     6:  400000     7:  400000
> 8:  400000     9:  400000    10:  400600    11: 2696628
>12: 2696628    13: 2696628    14: 2696628    15: 2696628
>16: 2696628    17: 2696628    18: 2696628    19: 2696628
>20: 2696628    21: 2696628    22: 2696628    23: 2696628
>24: 2696628    25: 2696628    26: 2696628    27: 2696628
>28: 2696628    29: 2696628    30: 2696628    31: 2696628
>32: 2696628    33: 2696628    34: 2696628    35: 2696628
>36: 2696628    37: 2696628    38: 2696628    39: 2696628
>40: 2696628    41:  400000    42:  400000    43:  400000
>44:  400000    45:  398847    46:  400000    47:  400000
>48:  400000    49:  400000    50:  400000    51: 2696628
>52: 2696628    53: 2696628    54: 2696628    55: 2696628
>56: 2696628    57: 2696628    58: 2696628    59: 2696628
>60: 2696628    61: 2696628    62: 2696628    63: 2699264
>
>Note:
>(1) The frequency of 1-10 and 41-50 CPUs work on the lowest frequency.
>     It turned out that nohz full was already work.
>     I guess that stress-ng cannot use the CPU in the range of nohz full.
>     Because the CPU frequency will be increased to 2.7G by binding 
>CPU to other application.
>(2) The frequency of the nohz full core is calculated by get() 
>callback according to ftrace.

I think this is a good point. It is possible that on large core count
systems a number of CPUs are isolated and don't have a kernel tick. The
approach in [1] won't work for those CPUs. The changes proposed in this
patch make sure that regardless of the scheduler or governor
configuration, the CPU frequency reporting will be correct.

My concerns with this approach:
1. We will wake up idle CPUs and query the AMU counters.
    a. This may not reflect the CPU frequency at the time the CPU was
    running. For example, CPU was running at 2.7 GHz, then became idle and
    invoked WFI. The CPU implementation reduces the CPU frequency to 400
    MHz when it enters WFI. Now if we wake up the CPU to query the
    frequency, it will return 400 MHz. This might be misleading.

    b. It wastes energy as we may wake up cores just to query frequency
    and then have them go back to idle.

    To solve this, may be we can cache the value of AMU counters; return
    the cached value of the AMU counter if the CPU is idle.

2. I think the acpi_cppc should use FFH registers only if the firmware
    publishes FFH support for both the delivered and reference counters.
    That way the system designer (through firmware) can control if the
    FFH registers are used for computing frequency.

[1] https://lore.kernel.org/lkml/20231127160838.1403404-3-beata.michalska@arm.com/

Thanks,
Vanshidhar

>
>[1] https://lore.kernel.org/lkml/20230418113459.12860-7-sumitg@nvidia.com/
>[2] https://lore.kernel.org/lkml/20231127160838.1403404-3-beata.michalska@arm.com/
>>
>>Many thanks,
>>Ionela.
>>
>>>/Huisong
>>>
>>>>>>[1] https://lore.kernel.org/lkml/20231127160838.1403404-1-beata.michalska@arm.com/
>>>>>>
>>>>>>Thanks,
>>>>>>Ionela.
>>>>>>
>>>>>>>This patch adds a interface, cpc_read_arch_counters_on_cpu, to read
>>>>>>>delivered and reference performance counter together. According to my
>>>>>>>test[4], the discrepancy of cpu current frequency in the
>>>>>>>scenarios with
>>>>>>>high memory access pressure is lower than 0.2% by stress-ng
>>>>>>>application.
>>>>>>>
>>>>>>>[1] https://lore.kernel.org/all/20231025093847.3740104-4-zengheng4@huawei.com/
>>>>>>>[2] https://lore.kernel.org/all/20230328193846.8757-1-yang@os.amperecomputing.com/
>>>>>>>[3]
>>>>>>>https://lore.kernel.org/all/20230418113459.12860-7-sumitg@nvidia.com/
>>>>>>>
>>>>>>>[4] My local test:
>>>>>>>The testing platform enable SMT and include 128 logical CPU in total,
>>>>>>>and CPU base frequency is 2.7GHz. Reading "cpuinfo_cur_freq" for each
>>>>>>>physical core on platform during the high memory access pressure from
>>>>>>>stress-ng, and the output is as follows:
>>>>>>>   0: 2699133     2: 2699942     4: 2698189     6: 2704347
>>>>>>>   8: 2704009    10: 2696277    12: 2702016    14: 2701388
>>>>>>>  16: 2700358    18: 2696741    20: 2700091    22: 2700122
>>>>>>>  24: 2701713    26: 2702025    28: 2699816    30: 2700121
>>>>>>>  32: 2700000    34: 2699788    36: 2698884    38: 2699109
>>>>>>>  40: 2704494    42: 2698350    44: 2699997    46: 2701023
>>>>>>>  48: 2703448    50: 2699501    52: 2700000    54: 2699999
>>>>>>>  56: 2702645    58: 2696923    60: 2697718    62: 2700547
>>>>>>>  64: 2700313    66: 2700000    68: 2699904    70: 2699259
>>>>>>>  72: 2699511    74: 2700644    76: 2702201    78: 2700000
>>>>>>>  80: 2700776    82: 2700364    84: 2702674    86: 2700255
>>>>>>>  88: 2699886    90: 2700359    92: 2699662    94: 2696188
>>>>>>>  96: 2705454    98: 2699260   100: 2701097   102: 2699630
>>>>>>>104: 2700463   106: 2698408   108: 2697766   110: 2701181
>>>>>>>112: 2699166   114: 2701804   116: 2701907   118: 2701973
>>>>>>>120: 2699584   122: 2700474   124: 2700768   126: 2701963
>>>>>>>
>>>>>>>Signed-off-by: Huisong Li <lihuisong@huawei.com>
>>>>>>>---
>[snip]
>>.
Beata Michalska Jan. 16, 2024, 2:10 p.m. UTC | #14
Hi,

Apologies for jumping in so late....

On Wed, Jan 10, 2024 at 03:09:48PM +0800, lihuisong (C) wrote:
> Hi Ionela,
> 
> 在 2024/1/8 22:03, Ionela Voinescu 写道:
> > Hi,
> > 
> > On Friday 05 Jan 2024 at 15:04:47 (+0800), lihuisong (C) wrote:
> > > Hi Vanshi,
> > > 
> > > 在 2024/1/5 8:48, Vanshidhar Konda 写道:
> > > > On Thu, Jan 04, 2024 at 05:36:51PM +0800, lihuisong (C) wrote:
> > > > > 在 2024/1/4 1:53, Ionela Voinescu 写道:
> > > > > > Hi,
> > > > > > 
> > > > > > On Tuesday 12 Dec 2023 at 15:26:17 (+0800), Huisong Li wrote:
> > > > > > > Many developers found that the cpu current frequency is greater than
> > > > > > > the maximum frequency of the platform, please see [1], [2] and [3].
> > > > > > > 
> > > > > > > In the scenarios with high memory access pressure, the patch [1] has
> > > > > > > proved the significant latency of cpc_read() which is used to obtain
> > > > > > > delivered and reference performance counter cause an absurd frequency.
> > > > > > > The sampling interval for this counters is very critical and
> > > > > > > is expected
> > > > > > > to be equal. However, the different latency of cpc_read() has a direct
> > > > > > > impact on their sampling interval.
> > > > > > > 
> > > > > > Would this [1] alternative solution work for you?
> > > > > It would work for me AFAICS.
> > > > > Because the "arch_freq_scale" is also from AMU core and constant
> > > > > counter, and read together.
> > > > > But, from their discuss line, it seems that there are some tricky
> > > > > points to clarify or consider.
> > > > I think the changes in [1] would work better when CPUs may be idle. With
> > > > this
> > > > patch we would have to wake any core that is in idle state to read the
> > > > AMU
> > > > counters. Worst case, if core 0 is trying to read the CPU frequency of
> > > > all
> > > > cores, it may need to wake up all the other cores to read the AMU
> > > > counters.
> > >  From the approach in [1], if all CPUs (one or more cores) under one policy
> > > are idle, they still cannot be obtained the CPU frequency, right?
> > > In this case, the [1] API will return 0 and have to back to call
> > > cpufreq_driver->get() for cpuinfo_cur_freq.
> > > Then we still need to face the issue this patch mentioned.
> > With the implementation at [1], arch_freq_get_on_cpu() will not return 0
> > for idle CPUs and the get() callback will not be called to wake up the
> > CPUs.
> Right, arch_freq_get_on_cpu() will not return 0 for idle CPUs.
> However, for no-housekeeping CPUs, it will return 0 and have to call get()
> callback, right?
> > 
> > Worst case, arch_freq_get_on_cpu() will return a frequency based on the
> > AMU counter values obtained on the last tick on that CPU. But if that CPU
> > is not a housekeeping CPU, a housekeeping CPU in the same policy will be
> > selected, as it would have had a more recent tick, and therefore a more
> > recent frequency value for the domain.
> But this frequency is from the last tick,
> this last tick is probably a long time ago and it doesn't update
> 'arch_freq_scale' for some reasons like CPU dile.
> In addition, I'm not sure if there is possible that amu_scale_freq_tick() is
> executed delayed under high stress case.
> It also have an impact on the accuracy of the cpu frequency we query.
> > 
> > I understand that the frequency returned here will not be up to date,
> > but there's no proper frequency feedback for an idle CPU. If one only
> > wakes up a CPU to sample counters, before the CPU goes back to sleep,
> > the obtained frequency feedback is meaningless.
> > 
> > > > For systems with 128 cores or more, this could be very expensive and
> > > > happen
> > > > very frequently.
> > > > 
> > > > AFAICS, the approach in [1] would avoid this cost.
> > > But the CPU frequency is just an average value for the last tick period
> > > instead of the current one the CPU actually runs at.
> > > In addition, there are some conditions to use 'arch_freq_scale' in this
> > > approach.
> > What are the conditions you are referring to?
> It depends on the housekeeping CPUs.
> > 
> > > So I'm not sure if this approach can entirely cover the frequency
> > > discrepancy issue.
> > Unfortunately there is no perfect frequency feedback. By the time you
> > observe/use the value of scaling_cur_freq/cpuinfo_cur_freq, the frequency
> > of the CPU might have already changed. Therefore, an average value might
> > be a better indication of the recent performance level of a CPU.
> An average value for CPU frequency is ok. It may be better if it has not any
> delaying.
> 
> The original implementation for cpuinfo_cur_freq can more reflect their
> meaning in the user-guide [1]. The user-guide said:
> "cpuinfo_cur_freq : Current frequency of the CPU as obtained from the
> hardware, in KHz.
> This is the frequency the CPU actually runs at."
> 
> 
> [1]https://www.kernel.org/doc/Documentation/cpu-freq/user-guide.txt
> 
> > 
> > Would you be able to test [1] on your platform and usecase?
> I has tested it on my platform (CPU number: 64, SMT: off and CPU base
> frequency: 2.7GHz).
> Accoding to the testing result,
> 1> I found that patch [1] and [2] cannot cover the no housekeeping CPUs.
> They still have to face the large frequency discrepancy issue my patch
> mentioned.
> 2> Additionally, the frequency value of all CPUs are almost the same by
> using the 'arch_freq_scale' factor way. I'm not sure if it is ok.
> 
> The patch [1] has been modified silightly as below:
> -->
> @@ -1756,7 +1756,10 @@ static unsigned int
> cpufreq_verify_current_freq(struct cpufreq_policy *policy, b
>  {
>         unsigned int new_freq;
> 
> -       new_freq = cpufreq_driver->get(policy->cpu);
> +       new_freq = arch_freq_get_on_cpu(policy->cpu);
> +       if (!new_freq)
> +               new_freq = cpufreq_driver->get(policy->cpu);
> +
As pointed out this change will not make it to the next version of the patch.
So I'd say you can safely ignore it and assume that arch_freq_get_on_cpu will
only be wired for sysfs nodes for scaling_cur_freq/cpuinfo_cur_freq
>         if (!new_freq)
>                 return 0;
> 
> And the result is as follows:
> *case 1:**No setting the nohz_full and cpufreq use performance governor*
> *--> Step1: *read 'cpuinfo_cur_freq' in no pressure
>   0: 2699264     2: 2699264     4: 2699264     6: 2699264
>   8: 2696628    10: 2696628    12: 2696628    14: 2699264
>  16: 2699264    18: 2696628    20: 2699264    22: 2696628
>  24: 2699264    26: 2696628    28: 2699264    30: 2696628
>  32: 2696628    34: 2696628    36: 2696628    38: 2696628
>  40: 2699264    42: 2699264    44: 2696628    46: 2696628
>  48: 2696628    50: 2699264    52: 2699264    54: 2696628
>  56: 2696628    58: 2696628    60: 2696628    62: 2696628
>  64: 2696628    66: 2699264    68: 2696628    70: 2696628
>  72: 2699264    74: 2696628    76: 2696628    78: 2699264
>  80: 2696628    82: 2696628    84: 2699264    86: 2696628
>  88: 2696628    90: 2696628    92: 2696628    94: 2699264
>  96: 2696628    98: 2699264   100: 2699264   102: 2696628
> 104: 2699264   106: 2699264   108: 2699264   110: 2696628
> 112: 2699264   114: 2699264   116: 2699264   118: 2699264
> 120: 2696628   122: 2699264   124: 2696628   126: 2699264
> Note: the frequency of all CPUs are almost the same.
Were you expecting smth else ?
> 
> *--> Step 2: *read 'cpuinfo_cur_freq' in the high memory access pressure.
>   0: 2696628     2: 2696628     4: 2696628     6: 2696628
>   8: 2696628    10: 2696628    12: 2696628    14: 2696628
>  16: 2696628    18: 2696628    20: 2696628    22: 2696628
>  24: 2696628    26: 2696628    28: 2696628    30: 2696628
>  32: 2696628    34: 2696628    36: 2696628    38: 2696628
>  40: 2696628    42: 2696628    44: 2696628    46: 2696628
>  48: 2696628    50: 2696628    52: 2696628    54: 2696628
>  56: 2696628    58: 2696628    60: 2696628    62: 2696628
>  64: 2696628    66: 2696628    68: 2696628    70: 2696628
>  72: 2696628    74: 2696628    76: 2696628    78: 2696628
>  80: 2696628    82: 2696628    84: 2696628    86: 2696628
>  88: 2696628    90: 2696628    92: 2696628    94: 2696628
>  96: 2696628    98: 2696628   100: 2696628   102: 2696628
> 104: 2696628   106: 2696628   108: 2696628   110: 2696628
> 112: 2696628   114: 2696628   116: 2696628   118: 2696628
> 120: 2696628   122: 2696628   124: 2696628   126: 2696628
> 
> *Case 2: setting nohz_full and cpufreq use ondemand governor*
> There is "isolcpus=1-10,41-50 nohz_full=1-10,41-50 rcu_nocbs=1-10,41-50" in
> /proc/cmdline.
Right, so if I remember correctly nohz_full implies rcu_nocbs, so no need to
set that one.
Now, afair, isolcpus will make the selected CPUs to disappear from the
schedulers view (no balancing, no migrating), so unless you affine smth
explicitly to those CPUs, you will not see much of an activity there.
Need to double check though as it has been a while ...
> *--> Step 1: *setting ondemand governor to all policy and query
> 'cpuinfo_cur_freq' in no pressure case.
> And the frequency of CPUs all are about 400MHz.
> *--> Step 2:* read 'cpuinfo_cur_freq' in the high memory access pressure.
> The high memory access pressure is from the command: "stress-ng -c 64
> --cpu-load 100% --taskset 0-63"
I'm not entirely convinced that this will affine to isolated cpus, especially
that the affinity mask spans all available cpus. If that is the case, no wonder
your isolated cpus are getting wasted being idle. But I would have to double
check how this is being handled.
> The result:
>  0: 2696628     1:  400000     2:  400000     3:  400909
>  4:  400000     5:  400000     6:  400000     7:  400000
>  8:  400000     9:  400000    10:  400600    11: 2696628
> 12: 2696628    13: 2696628    14: 2696628    15: 2696628
> 16: 2696628    17: 2696628    18: 2696628    19: 2696628
> 20: 2696628    21: 2696628    22: 2696628    23: 2696628
> 24: 2696628    25: 2696628    26: 2696628    27: 2696628
> 28: 2696628    29: 2696628    30: 2696628    31: 2696628
> 32: 2696628    33: 2696628    34: 2696628    35: 2696628
> 36: 2696628    37: 2696628    38: 2696628    39: 2696628
> 40: 2696628    41:  400000    42:  400000    43:  400000
> 44:  400000    45:  398847    46:  400000    47:  400000
> 48:  400000    49:  400000    50:  400000    51: 2696628
> 52: 2696628    53: 2696628    54: 2696628    55: 2696628
> 56: 2696628    57: 2696628    58: 2696628    59: 2696628
> 60: 2696628    61: 2696628    62: 2696628    63: 2699264
> 
> Note:
> (1) The frequency of 1-10 and 41-50 CPUs work on the lowest frequency.
>      It turned out that nohz full was already work.
>      I guess that stress-ng cannot use the CPU in the range of nohz full.
>      Because the CPU frequency will be increased to 2.7G by binding CPU to
> other application.
> (2) The frequency of the nohz full core is calculated by get() callback
> according to ftrace.
It is as there is no sched tick on those, and apparently there is nothing
running on them either.

Unless I am missing smth.

---
BR
Beata

> 
> [1] https://lore.kernel.org/lkml/20230418113459.12860-7-sumitg@nvidia.com/
> [2] https://lore.kernel.org/lkml/20231127160838.1403404-3-beata.michalska@arm.com/
> > 
> > Many thanks,
> > Ionela.
> > 
> > > /Huisong
> > > 
> > > > > > [1] https://lore.kernel.org/lkml/20231127160838.1403404-1-beata.michalska@arm.com/
> > > > > > 
> > > > > > Thanks,
> > > > > > Ionela.
> > > > > > 
> > > > > > > This patch adds a interface, cpc_read_arch_counters_on_cpu, to read
> > > > > > > delivered and reference performance counter together. According to my
> > > > > > > test[4], the discrepancy of cpu current frequency in the
> > > > > > > scenarios with
> > > > > > > high memory access pressure is lower than 0.2% by stress-ng
> > > > > > > application.
> > > > > > > 
> > > > > > > [1] https://lore.kernel.org/all/20231025093847.3740104-4-zengheng4@huawei.com/
> > > > > > > [2] https://lore.kernel.org/all/20230328193846.8757-1-yang@os.amperecomputing.com/
> > > > > > > [3]
> > > > > > > https://lore.kernel.org/all/20230418113459.12860-7-sumitg@nvidia.com/
> > > > > > > 
> > > > > > > [4] My local test:
> > > > > > > The testing platform enable SMT and include 128 logical CPU in total,
> > > > > > > and CPU base frequency is 2.7GHz. Reading "cpuinfo_cur_freq" for each
> > > > > > > physical core on platform during the high memory access pressure from
> > > > > > > stress-ng, and the output is as follows:
> > > > > > >    0: 2699133     2: 2699942     4: 2698189     6: 2704347
> > > > > > >    8: 2704009    10: 2696277    12: 2702016    14: 2701388
> > > > > > >   16: 2700358    18: 2696741    20: 2700091    22: 2700122
> > > > > > >   24: 2701713    26: 2702025    28: 2699816    30: 2700121
> > > > > > >   32: 2700000    34: 2699788    36: 2698884    38: 2699109
> > > > > > >   40: 2704494    42: 2698350    44: 2699997    46: 2701023
> > > > > > >   48: 2703448    50: 2699501    52: 2700000    54: 2699999
> > > > > > >   56: 2702645    58: 2696923    60: 2697718    62: 2700547
> > > > > > >   64: 2700313    66: 2700000    68: 2699904    70: 2699259
> > > > > > >   72: 2699511    74: 2700644    76: 2702201    78: 2700000
> > > > > > >   80: 2700776    82: 2700364    84: 2702674    86: 2700255
> > > > > > >   88: 2699886    90: 2700359    92: 2699662    94: 2696188
> > > > > > >   96: 2705454    98: 2699260   100: 2701097   102: 2699630
> > > > > > > 104: 2700463   106: 2698408   108: 2697766   110: 2701181
> > > > > > > 112: 2699166   114: 2701804   116: 2701907   118: 2701973
> > > > > > > 120: 2699584   122: 2700474   124: 2700768   126: 2701963
> > > > > > > 
> > > > > > > Signed-off-by: Huisong Li <lihuisong@huawei.com>
> > > > > > > ---
> [snip]
> > .
lihuisong (C) Jan. 17, 2024, 8:54 a.m. UTC | #15
在 2024/1/13 2:33, Vanshidhar Konda 写道:
> On Wed, Jan 10, 2024 at 03:09:48PM +0800, lihuisong (C) wrote:
>> Hi Ionela,
>>
>> 在 2024/1/8 22:03, Ionela Voinescu 写道:
>>> Hi,
>>>
>>> On Friday 05 Jan 2024 at 15:04:47 (+0800), lihuisong (C) wrote:
>>>> Hi Vanshi,
>>>>
>>>> 在 2024/1/5 8:48, Vanshidhar Konda 写道:
>>>>> On Thu, Jan 04, 2024 at 05:36:51PM +0800, lihuisong (C) wrote:
>>>>>> 在 2024/1/4 1:53, Ionela Voinescu 写道:
>>>>>>> Hi,
>>>>>>>
>>>>>>> On Tuesday 12 Dec 2023 at 15:26:17 (+0800), Huisong Li wrote:
>>>>>>>> Many developers found that the cpu current frequency is greater 
>>>>>>>> than
>>>>>>>> the maximum frequency of the platform, please see [1], [2] and 
>>>>>>>> [3].
>>>>>>>>
>>>>>>>> In the scenarios with high memory access pressure, the patch 
>>>>>>>> [1] has
>>>>>>>> proved the significant latency of cpc_read() which is used to 
>>>>>>>> obtain
>>>>>>>> delivered and reference performance counter cause an absurd 
>>>>>>>> frequency.
>>>>>>>> The sampling interval for this counters is very critical and
>>>>>>>> is expected
>>>>>>>> to be equal. However, the different latency of cpc_read() has a 
>>>>>>>> direct
>>>>>>>> impact on their sampling interval.
>>>>>>>>
>>>>>>> Would this [1] alternative solution work for you?
>>>>>> It would work for me AFAICS.
>>>>>> Because the "arch_freq_scale" is also from AMU core and constant
>>>>>> counter, and read together.
>>>>>> But, from their discuss line, it seems that there are some tricky
>>>>>> points to clarify or consider.
>>>>> I think the changes in [1] would work better when CPUs may be 
>>>>> idle. With
>>>>> this
>>>>> patch we would have to wake any core that is in idle state to read 
>>>>> the
>>>>> AMU
>>>>> counters. Worst case, if core 0 is trying to read the CPU 
>>>>> frequency of
>>>>> all
>>>>> cores, it may need to wake up all the other cores to read the AMU
>>>>> counters.
>>>> From the approach in [1], if all CPUs (one or more cores) under one 
>>>> policy
>>>> are idle, they still cannot be obtained the CPU frequency, right?
>>>> In this case, the [1] API will return 0 and have to back to call
>>>> cpufreq_driver->get() for cpuinfo_cur_freq.
>>>> Then we still need to face the issue this patch mentioned.
>>> With the implementation at [1], arch_freq_get_on_cpu() will not 
>>> return 0
>>> for idle CPUs and the get() callback will not be called to wake up the
>>> CPUs.
>> Right, arch_freq_get_on_cpu() will not return 0 for idle CPUs.
>> However, for no-housekeeping CPUs, it will return 0 and have to call 
>> get() callback, right?
>>>
>>> Worst case, arch_freq_get_on_cpu() will return a frequency based on the
>>> AMU counter values obtained on the last tick on that CPU. But if 
>>> that CPU
>>> is not a housekeeping CPU, a housekeeping CPU in the same policy 
>>> will be
>>> selected, as it would have had a more recent tick, and therefore a more
>>> recent frequency value for the domain.
>> But this frequency is from the last tick,
>> this last tick is probably a long time ago and it doesn't update 
>> 'arch_freq_scale' for some reasons like CPU dile.
>> In addition, I'm not sure if there is possible that 
>> amu_scale_freq_tick() is executed delayed under high stress case.
>> It also have an impact on the accuracy of the cpu frequency we query.
>>>
>>> I understand that the frequency returned here will not be up to date,
>>> but there's no proper frequency feedback for an idle CPU. If one only
>>> wakes up a CPU to sample counters, before the CPU goes back to sleep,
>>> the obtained frequency feedback is meaningless.
>>>
>>>>> For systems with 128 cores or more, this could be very expensive and
>>>>> happen
>>>>> very frequently.
>>>>>
>>>>> AFAICS, the approach in [1] would avoid this cost.
>>>> But the CPU frequency is just an average value for the last tick 
>>>> period
>>>> instead of the current one the CPU actually runs at.
>>>> In addition, there are some conditions to use 'arch_freq_scale' in 
>>>> this
>>>> approach.
>>> What are the conditions you are referring to?
>> It depends on the housekeeping CPUs.
>>>
>>>> So I'm not sure if this approach can entirely cover the frequency
>>>> discrepancy issue.
>>> Unfortunately there is no perfect frequency feedback. By the time you
>>> observe/use the value of scaling_cur_freq/cpuinfo_cur_freq, the 
>>> frequency
>>> of the CPU might have already changed. Therefore, an average value 
>>> might
>>> be a better indication of the recent performance level of a CPU.
>> An average value for CPU frequency is ok. It may be better if it has 
>> not any delaying.
>>
>> The original implementation for cpuinfo_cur_freq can more reflect their
>> meaning in the user-guide [1]. The user-guide said:
>> "cpuinfo_cur_freq : Current frequency of the CPU as obtained from the 
>> hardware, in KHz.
>> This is the frequency the CPU actually runs at."
>>
>>
>> [1]https://www.kernel.org/doc/Documentation/cpu-freq/user-guide.txt
>>
>>>
>>> Would you be able to test [1] on your platform and usecase?
>> I has tested it on my platform (CPU number: 64, SMT: off and CPU base 
>> frequency: 2.7GHz).
>> Accoding to the testing result,
>> 1> I found that patch [1] and [2] cannot cover the no housekeeping 
>> CPUs. They still have to face the large frequency discrepancy issue 
>> my patch mentioned.
>> 2> Additionally, the frequency value of all CPUs are almost the same 
>> by using the 'arch_freq_scale' factor way. I'm not sure if it is ok.
>>
>> The patch [1] has been modified silightly as below:
>> -->
>> @@ -1756,7 +1756,10 @@ static unsigned int 
>> cpufreq_verify_current_freq(struct cpufreq_policy *policy, b
>>  {
>>         unsigned int new_freq;
>>
>> -       new_freq = cpufreq_driver->get(policy->cpu);
>> +       new_freq = arch_freq_get_on_cpu(policy->cpu);
>> +       if (!new_freq)
>> +               new_freq = cpufreq_driver->get(policy->cpu);
>> +
>>         if (!new_freq)
>>                 return 0;
>>
>> And the result is as follows:
>> *case 1:**No setting the nohz_full and cpufreq use performance governor*
>> *--> Step1: *read 'cpuinfo_cur_freq' in no pressure
>>   0: 2699264     2: 2699264     4: 2699264     6: 2699264
>>   8: 2696628    10: 2696628    12: 2696628    14: 2699264
>>  16: 2699264    18: 2696628    20: 2699264    22: 2696628
>>  24: 2699264    26: 2696628    28: 2699264    30: 2696628
>>  32: 2696628    34: 2696628    36: 2696628    38: 2696628
>>  40: 2699264    42: 2699264    44: 2696628    46: 2696628
>>  48: 2696628    50: 2699264    52: 2699264    54: 2696628
>>  56: 2696628    58: 2696628    60: 2696628    62: 2696628
>>  64: 2696628    66: 2699264    68: 2696628    70: 2696628
>>  72: 2699264    74: 2696628    76: 2696628    78: 2699264
>>  80: 2696628    82: 2696628    84: 2699264    86: 2696628
>>  88: 2696628    90: 2696628    92: 2696628    94: 2699264
>>  96: 2696628    98: 2699264   100: 2699264   102: 2696628
>> 104: 2699264   106: 2699264   108: 2699264   110: 2696628
>> 112: 2699264   114: 2699264   116: 2699264   118: 2699264
>> 120: 2696628   122: 2699264   124: 2696628   126: 2699264
>> Note: the frequency of all CPUs are almost the same.
>>
>> *--> Step 2: *read 'cpuinfo_cur_freq' in the high memory access 
>> pressure.
>>   0: 2696628     2: 2696628     4: 2696628     6: 2696628
>>   8: 2696628    10: 2696628    12: 2696628    14: 2696628
>>  16: 2696628    18: 2696628    20: 2696628    22: 2696628
>>  24: 2696628    26: 2696628    28: 2696628    30: 2696628
>>  32: 2696628    34: 2696628    36: 2696628    38: 2696628
>>  40: 2696628    42: 2696628    44: 2696628    46: 2696628
>>  48: 2696628    50: 2696628    52: 2696628    54: 2696628
>>  56: 2696628    58: 2696628    60: 2696628    62: 2696628
>>  64: 2696628    66: 2696628    68: 2696628    70: 2696628
>>  72: 2696628    74: 2696628    76: 2696628    78: 2696628
>>  80: 2696628    82: 2696628    84: 2696628    86: 2696628
>>  88: 2696628    90: 2696628    92: 2696628    94: 2696628
>>  96: 2696628    98: 2696628   100: 2696628   102: 2696628
>> 104: 2696628   106: 2696628   108: 2696628   110: 2696628
>> 112: 2696628   114: 2696628   116: 2696628   118: 2696628
>> 120: 2696628   122: 2696628   124: 2696628   126: 2696628
>>
>> *Case 2: setting nohz_full and cpufreq use ondemand governor*
>> There is "isolcpus=1-10,41-50 nohz_full=1-10,41-50 
>> rcu_nocbs=1-10,41-50" in /proc/cmdline.
>> *--> Step 1: *setting ondemand governor to all policy and query 
>> 'cpuinfo_cur_freq' in no pressure case.
>> And the frequency of CPUs all are about 400MHz.
>> *--> Step 2:* read 'cpuinfo_cur_freq' in the high memory access 
>> pressure.
>> The high memory access pressure is from the command: "stress-ng -c 64 
>> --cpu-load 100% --taskset 0-63"
>> The result:
>>  0: 2696628     1:  400000     2:  400000     3:  400909
>>  4:  400000     5:  400000     6:  400000     7:  400000
>>  8:  400000     9:  400000    10:  400600    11: 2696628
>> 12: 2696628    13: 2696628    14: 2696628    15: 2696628
>> 16: 2696628    17: 2696628    18: 2696628    19: 2696628
>> 20: 2696628    21: 2696628    22: 2696628    23: 2696628
>> 24: 2696628    25: 2696628    26: 2696628    27: 2696628
>> 28: 2696628    29: 2696628    30: 2696628    31: 2696628
>> 32: 2696628    33: 2696628    34: 2696628    35: 2696628
>> 36: 2696628    37: 2696628    38: 2696628    39: 2696628
>> 40: 2696628    41:  400000    42:  400000    43:  400000
>> 44:  400000    45:  398847    46:  400000    47:  400000
>> 48:  400000    49:  400000    50:  400000    51: 2696628
>> 52: 2696628    53: 2696628    54: 2696628    55: 2696628
>> 56: 2696628    57: 2696628    58: 2696628    59: 2696628
>> 60: 2696628    61: 2696628    62: 2696628    63: 2699264
>>
>> Note:
>> (1) The frequency of 1-10 and 41-50 CPUs work on the lowest frequency.
>>      It turned out that nohz full was already work.
>>      I guess that stress-ng cannot use the CPU in the range of nohz 
>> full.
>>      Because the CPU frequency will be increased to 2.7G by binding 
>> CPU to other application.
>> (2) The frequency of the nohz full core is calculated by get() 
>> callback according to ftrace.
>
> I think this is a good point. It is possible that on large core count
> systems a number of CPUs are isolated and don't have a kernel tick. The
> approach in [1] won't work for those CPUs. The changes proposed in this
> patch make sure that regardless of the scheduler or governor
> configuration, the CPU frequency reporting will be correct.
>
> My concerns with this approach:
> 1. We will wake up idle CPUs and query the AMU counters.
>    a. This may not reflect the CPU frequency at the time the CPU was
>    running. For example, CPU was running at 2.7 GHz, then became idle and
>    invoked WFI. The CPU implementation reduces the CPU frequency to 400
>    MHz when it enters WFI. Now if we wake up the CPU to query the
>    frequency, it will return 400 MHz. This might be misleading.
>
>    b. It wastes energy as we may wake up cores just to query frequency
>    and then have them go back to idle.
Waking up an idle CPU to query CPU frequency and then going back to idle 
wastes energy.
The frequency of the idle CPUs may be is zero, but we cannot display 
zero because it is not friendly for user.
If we use the cached counter as the frequency feedback before CPU being 
idle, it cannot reflect a real frequency on CPU.
So there is no perfect frequency feedback for idle CPUs.
>
>    To solve this, may be we can cache the value of AMU counters; return
>    the cached value of the AMU counter if the CPU is idle.
The "arch_freq_scale" which is from AMU counters is similar to the 
cached value in Beata's approach.
But it just work for the housekeeping CPUs with tick.
If using cached frequency is ok, how shoud we do for nohz full CPUs? I 
have no good idea.
>
> 2. I think the acpi_cppc should use FFH registers only if the firmware
>    publishes FFH support for both the delivered and reference counters.
>    That way the system designer (through firmware) can control if the
>    FFH registers are used for computing frequency.
Right, so this patch have to judge if support FFH.
>
> [1] 
> https://lore.kernel.org/lkml/20231127160838.1403404-3-beata.michalska@arm.com/
>
> Thanks,
> Vanshidhar
>
>>
>> [1] 
>> https://lore.kernel.org/lkml/20230418113459.12860-7-sumitg@nvidia.com/
>> [2] 
>> https://lore.kernel.org/lkml/20231127160838.1403404-3-beata.michalska@arm.com/
>>>
>>> Many thanks,
>>> Ionela.
>>>
>>>> /Huisong
>>>>
>>>>>>> [1] 
>>>>>>> https://lore.kernel.org/lkml/20231127160838.1403404-1-beata.michalska@arm.com/
>>>>>>>
>>>>>>> Thanks,
>>>>>>> Ionela.
>>>>>>>
>>>>>>>> This patch adds a interface, cpc_read_arch_counters_on_cpu, to 
>>>>>>>> read
>>>>>>>> delivered and reference performance counter together. According 
>>>>>>>> to my
>>>>>>>> test[4], the discrepancy of cpu current frequency in the
>>>>>>>> scenarios with
>>>>>>>> high memory access pressure is lower than 0.2% by stress-ng
>>>>>>>> application.
>>>>>>>>
>>>>>>>> [1] 
>>>>>>>> https://lore.kernel.org/all/20231025093847.3740104-4-zengheng4@huawei.com/
>>>>>>>> [2] 
>>>>>>>> https://lore.kernel.org/all/20230328193846.8757-1-yang@os.amperecomputing.com/
>>>>>>>> [3]
>>>>>>>> https://lore.kernel.org/all/20230418113459.12860-7-sumitg@nvidia.com/ 
>>>>>>>>
>>>>>>>>
>>>>>>>> [4] My local test:
>>>>>>>> The testing platform enable SMT and include 128 logical CPU in 
>>>>>>>> total,
>>>>>>>> and CPU base frequency is 2.7GHz. Reading "cpuinfo_cur_freq" 
>>>>>>>> for each
>>>>>>>> physical core on platform during the high memory access 
>>>>>>>> pressure from
>>>>>>>> stress-ng, and the output is as follows:
>>>>>>>>   0: 2699133     2: 2699942     4: 2698189     6: 2704347
>>>>>>>>   8: 2704009    10: 2696277    12: 2702016    14: 2701388
>>>>>>>>  16: 2700358    18: 2696741    20: 2700091    22: 2700122
>>>>>>>>  24: 2701713    26: 2702025    28: 2699816    30: 2700121
>>>>>>>>  32: 2700000    34: 2699788    36: 2698884    38: 2699109
>>>>>>>>  40: 2704494    42: 2698350    44: 2699997    46: 2701023
>>>>>>>>  48: 2703448    50: 2699501    52: 2700000    54: 2699999
>>>>>>>>  56: 2702645    58: 2696923    60: 2697718    62: 2700547
>>>>>>>>  64: 2700313    66: 2700000    68: 2699904    70: 2699259
>>>>>>>>  72: 2699511    74: 2700644    76: 2702201    78: 2700000
>>>>>>>>  80: 2700776    82: 2700364    84: 2702674    86: 2700255
>>>>>>>>  88: 2699886    90: 2700359    92: 2699662    94: 2696188
>>>>>>>>  96: 2705454    98: 2699260   100: 2701097   102: 2699630
>>>>>>>> 104: 2700463   106: 2698408   108: 2697766   110: 2701181
>>>>>>>> 112: 2699166   114: 2701804   116: 2701907   118: 2701973
>>>>>>>> 120: 2699584   122: 2700474   124: 2700768   126: 2701963
>>>>>>>>
>>>>>>>> Signed-off-by: Huisong Li <lihuisong@huawei.com>
>>>>>>>> ---
>> [snip]
>>> .
> .
lihuisong (C) Jan. 17, 2024, 9:18 a.m. UTC | #16
Hi,

在 2024/1/16 22:10, Beata Michalska 写道:
> Hi,
>
> Apologies for jumping in so late....
>
> On Wed, Jan 10, 2024 at 03:09:48PM +0800, lihuisong (C) wrote:
>> Hi Ionela,
>>
>> 在 2024/1/8 22:03, Ionela Voinescu 写道:
>>> Hi,
>>>
>>> On Friday 05 Jan 2024 at 15:04:47 (+0800), lihuisong (C) wrote:
>>>> Hi Vanshi,
>>>>
>>>> 在 2024/1/5 8:48, Vanshidhar Konda 写道:
>>>>> On Thu, Jan 04, 2024 at 05:36:51PM +0800, lihuisong (C) wrote:
>>>>>> 在 2024/1/4 1:53, Ionela Voinescu 写道:
>>>>>>> Hi,
>>>>>>>
>>>>>>> On Tuesday 12 Dec 2023 at 15:26:17 (+0800), Huisong Li wrote:
>>>>>>>> Many developers found that the cpu current frequency is greater than
>>>>>>>> the maximum frequency of the platform, please see [1], [2] and [3].
>>>>>>>>
>>>>>>>> In the scenarios with high memory access pressure, the patch [1] has
>>>>>>>> proved the significant latency of cpc_read() which is used to obtain
>>>>>>>> delivered and reference performance counter cause an absurd frequency.
>>>>>>>> The sampling interval for this counters is very critical and
>>>>>>>> is expected
>>>>>>>> to be equal. However, the different latency of cpc_read() has a direct
>>>>>>>> impact on their sampling interval.
>>>>>>>>
>>>>>>> Would this [1] alternative solution work for you?
>>>>>> It would work for me AFAICS.
>>>>>> Because the "arch_freq_scale" is also from AMU core and constant
>>>>>> counter, and read together.
>>>>>> But, from their discuss line, it seems that there are some tricky
>>>>>> points to clarify or consider.
>>>>> I think the changes in [1] would work better when CPUs may be idle. With
>>>>> this
>>>>> patch we would have to wake any core that is in idle state to read the
>>>>> AMU
>>>>> counters. Worst case, if core 0 is trying to read the CPU frequency of
>>>>> all
>>>>> cores, it may need to wake up all the other cores to read the AMU
>>>>> counters.
>>>>   From the approach in [1], if all CPUs (one or more cores) under one policy
>>>> are idle, they still cannot be obtained the CPU frequency, right?
>>>> In this case, the [1] API will return 0 and have to back to call
>>>> cpufreq_driver->get() for cpuinfo_cur_freq.
>>>> Then we still need to face the issue this patch mentioned.
>>> With the implementation at [1], arch_freq_get_on_cpu() will not return 0
>>> for idle CPUs and the get() callback will not be called to wake up the
>>> CPUs.
>> Right, arch_freq_get_on_cpu() will not return 0 for idle CPUs.
>> However, for no-housekeeping CPUs, it will return 0 and have to call get()
>> callback, right?
>>> Worst case, arch_freq_get_on_cpu() will return a frequency based on the
>>> AMU counter values obtained on the last tick on that CPU. But if that CPU
>>> is not a housekeeping CPU, a housekeeping CPU in the same policy will be
>>> selected, as it would have had a more recent tick, and therefore a more
>>> recent frequency value for the domain.
>> But this frequency is from the last tick,
>> this last tick is probably a long time ago and it doesn't update
>> 'arch_freq_scale' for some reasons like CPU dile.
>> In addition, I'm not sure if there is possible that amu_scale_freq_tick() is
>> executed delayed under high stress case.
>> It also have an impact on the accuracy of the cpu frequency we query.
>>> I understand that the frequency returned here will not be up to date,
>>> but there's no proper frequency feedback for an idle CPU. If one only
>>> wakes up a CPU to sample counters, before the CPU goes back to sleep,
>>> the obtained frequency feedback is meaningless.
>>>
>>>>> For systems with 128 cores or more, this could be very expensive and
>>>>> happen
>>>>> very frequently.
>>>>>
>>>>> AFAICS, the approach in [1] would avoid this cost.
>>>> But the CPU frequency is just an average value for the last tick period
>>>> instead of the current one the CPU actually runs at.
>>>> In addition, there are some conditions to use 'arch_freq_scale' in this
>>>> approach.
>>> What are the conditions you are referring to?
>> It depends on the housekeeping CPUs.
>>>> So I'm not sure if this approach can entirely cover the frequency
>>>> discrepancy issue.
>>> Unfortunately there is no perfect frequency feedback. By the time you
>>> observe/use the value of scaling_cur_freq/cpuinfo_cur_freq, the frequency
>>> of the CPU might have already changed. Therefore, an average value might
>>> be a better indication of the recent performance level of a CPU.
>> An average value for CPU frequency is ok. It may be better if it has not any
>> delaying.
>>
>> The original implementation for cpuinfo_cur_freq can more reflect their
>> meaning in the user-guide [1]. The user-guide said:
>> "cpuinfo_cur_freq : Current frequency of the CPU as obtained from the
>> hardware, in KHz.
>> This is the frequency the CPU actually runs at."
>>
>>
>> [1]https://www.kernel.org/doc/Documentation/cpu-freq/user-guide.txt
>>
>>> Would you be able to test [1] on your platform and usecase?
>> I has tested it on my platform (CPU number: 64, SMT: off and CPU base
>> frequency: 2.7GHz).
>> Accoding to the testing result,
>> 1> I found that patch [1] and [2] cannot cover the no housekeeping CPUs.
>> They still have to face the large frequency discrepancy issue my patch
>> mentioned.
>> 2> Additionally, the frequency value of all CPUs are almost the same by
>> using the 'arch_freq_scale' factor way. I'm not sure if it is ok.
>>
>> The patch [1] has been modified silightly as below:
>> -->
>> @@ -1756,7 +1756,10 @@ static unsigned int
>> cpufreq_verify_current_freq(struct cpufreq_policy *policy, b
>>   {
>>          unsigned int new_freq;
>>
>> -       new_freq = cpufreq_driver->get(policy->cpu);
>> +       new_freq = arch_freq_get_on_cpu(policy->cpu);
>> +       if (!new_freq)
>> +               new_freq = cpufreq_driver->get(policy->cpu);
>> +
> As pointed out this change will not make it to the next version of the patch.
> So I'd say you can safely ignore it and assume that arch_freq_get_on_cpu will
> only be wired for sysfs nodes for scaling_cur_freq/cpuinfo_cur_freq
>>          if (!new_freq)
>>                  return 0;
>>
>> And the result is as follows:
>> *case 1:**No setting the nohz_full and cpufreq use performance governor*
>> *--> Step1: *read 'cpuinfo_cur_freq' in no pressure
>>    0: 2699264     2: 2699264     4: 2699264     6: 2699264
>>    8: 2696628    10: 2696628    12: 2696628    14: 2699264
>>   16: 2699264    18: 2696628    20: 2699264    22: 2696628
>>   24: 2699264    26: 2696628    28: 2699264    30: 2696628
>>   32: 2696628    34: 2696628    36: 2696628    38: 2696628
>>   40: 2699264    42: 2699264    44: 2696628    46: 2696628
>>   48: 2696628    50: 2699264    52: 2699264    54: 2696628
>>   56: 2696628    58: 2696628    60: 2696628    62: 2696628
>>   64: 2696628    66: 2699264    68: 2696628    70: 2696628
>>   72: 2699264    74: 2696628    76: 2696628    78: 2699264
>>   80: 2696628    82: 2696628    84: 2699264    86: 2696628
>>   88: 2696628    90: 2696628    92: 2696628    94: 2699264
>>   96: 2696628    98: 2699264   100: 2699264   102: 2696628
>> 104: 2699264   106: 2699264   108: 2699264   110: 2696628
>> 112: 2699264   114: 2699264   116: 2699264   118: 2699264
>> 120: 2696628   122: 2699264   124: 2696628   126: 2699264
>> Note: the frequency of all CPUs are almost the same.
> Were you expecting smth else ?
The frequency of each CPU might have a different value.
All value of all CPUs is the same under high pressure.
I don't know what the phenomenon is on other platform.
Do you know who else tested it?
>> *--> Step 2: *read 'cpuinfo_cur_freq' in the high memory access pressure.
>>    0: 2696628     2: 2696628     4: 2696628     6: 2696628
>>    8: 2696628    10: 2696628    12: 2696628    14: 2696628
>>   16: 2696628    18: 2696628    20: 2696628    22: 2696628
>>   24: 2696628    26: 2696628    28: 2696628    30: 2696628
>>   32: 2696628    34: 2696628    36: 2696628    38: 2696628
>>   40: 2696628    42: 2696628    44: 2696628    46: 2696628
>>   48: 2696628    50: 2696628    52: 2696628    54: 2696628
>>   56: 2696628    58: 2696628    60: 2696628    62: 2696628
>>   64: 2696628    66: 2696628    68: 2696628    70: 2696628
>>   72: 2696628    74: 2696628    76: 2696628    78: 2696628
>>   80: 2696628    82: 2696628    84: 2696628    86: 2696628
>>   88: 2696628    90: 2696628    92: 2696628    94: 2696628
>>   96: 2696628    98: 2696628   100: 2696628   102: 2696628
>> 104: 2696628   106: 2696628   108: 2696628   110: 2696628
>> 112: 2696628   114: 2696628   116: 2696628   118: 2696628
>> 120: 2696628   122: 2696628   124: 2696628   126: 2696628
>>
>> *Case 2: setting nohz_full and cpufreq use ondemand governor*
>> There is "isolcpus=1-10,41-50 nohz_full=1-10,41-50 rcu_nocbs=1-10,41-50" in
>> /proc/cmdline.
> Right, so if I remember correctly nohz_full implies rcu_nocbs, so no need to
> set that one.
> Now, afair, isolcpus will make the selected CPUs to disappear from the
> schedulers view (no balancing, no migrating), so unless you affine smth
> explicitly to those CPUs, you will not see much of an activity there.
Correct.
> Need to double check though as it has been a while ...
>> *--> Step 1: *setting ondemand governor to all policy and query
>> 'cpuinfo_cur_freq' in no pressure case.
>> And the frequency of CPUs all are about 400MHz.
>> *--> Step 2:* read 'cpuinfo_cur_freq' in the high memory access pressure.
>> The high memory access pressure is from the command: "stress-ng -c 64
>> --cpu-load 100% --taskset 0-63"
> I'm not entirely convinced that this will affine to isolated cpus, especially
> that the affinity mask spans all available cpus. If that is the case, no wonder
> your isolated cpus are getting wasted being idle. But I would have to double
> check how this is being handled.
>> The result:
>>   0: 2696628     1:  400000     2:  400000     3:  400909
>>   4:  400000     5:  400000     6:  400000     7:  400000
>>   8:  400000     9:  400000    10:  400600    11: 2696628
>> 12: 2696628    13: 2696628    14: 2696628    15: 2696628
>> 16: 2696628    17: 2696628    18: 2696628    19: 2696628
>> 20: 2696628    21: 2696628    22: 2696628    23: 2696628
>> 24: 2696628    25: 2696628    26: 2696628    27: 2696628
>> 28: 2696628    29: 2696628    30: 2696628    31: 2696628
>> 32: 2696628    33: 2696628    34: 2696628    35: 2696628
>> 36: 2696628    37: 2696628    38: 2696628    39: 2696628
>> 40: 2696628    41:  400000    42:  400000    43:  400000
>> 44:  400000    45:  398847    46:  400000    47:  400000
>> 48:  400000    49:  400000    50:  400000    51: 2696628
>> 52: 2696628    53: 2696628    54: 2696628    55: 2696628
>> 56: 2696628    57: 2696628    58: 2696628    59: 2696628
>> 60: 2696628    61: 2696628    62: 2696628    63: 2699264
>>
>> Note:
>> (1) The frequency of 1-10 and 41-50 CPUs work on the lowest frequency.
>>       It turned out that nohz full was already work.
>>       I guess that stress-ng cannot use the CPU in the range of nohz full.
>>       Because the CPU frequency will be increased to 2.7G by binding CPU to
>> other application.
>> (2) The frequency of the nohz full core is calculated by get() callback
>> according to ftrace.
> It is as there is no sched tick on those, and apparently there is nothing
> running on them either.
Yes.
If we select your approach and the above phenomenon is normal,
the large frequency discrepancy issue can be resolved for CPUs with 
sched tick by the way.
But the nohz full cores still have to face this issue. So this patch is 
also needed.

BR
/huisong
>
> Unless I am missing smth.
>
> ---
> BR
> Beata
>
>> [1] https://lore.kernel.org/lkml/20230418113459.12860-7-sumitg@nvidia.com/
>> [2] https://lore.kernel.org/lkml/20231127160838.1403404-3-beata.michalska@arm.com/
>>> Many thanks,
>>> Ionela.
>>>
>>>> /Huisong
>>>>
>>>>>>> [1] https://lore.kernel.org/lkml/20231127160838.1403404-1-beata.michalska@arm.com/
>>>>>>>
>>>>>>> Thanks,
>>>>>>> Ionela.
>>>>>>>
>>>>>>>> This patch adds a interface, cpc_read_arch_counters_on_cpu, to read
>>>>>>>> delivered and reference performance counter together. According to my
>>>>>>>> test[4], the discrepancy of cpu current frequency in the
>>>>>>>> scenarios with
>>>>>>>> high memory access pressure is lower than 0.2% by stress-ng
>>>>>>>> application.
>>>>>>>>
>>>>>>>> [1] https://lore.kernel.org/all/20231025093847.3740104-4-zengheng4@huawei.com/
>>>>>>>> [2] https://lore.kernel.org/all/20230328193846.8757-1-yang@os.amperecomputing.com/
>>>>>>>> [3]
>>>>>>>> https://lore.kernel.org/all/20230418113459.12860-7-sumitg@nvidia.com/
>>>>>>>>
>>>>>>>> [4] My local test:
>>>>>>>> The testing platform enable SMT and include 128 logical CPU in total,
>>>>>>>> and CPU base frequency is 2.7GHz. Reading "cpuinfo_cur_freq" for each
>>>>>>>> physical core on platform during the high memory access pressure from
>>>>>>>> stress-ng, and the output is as follows:
>>>>>>>>     0: 2699133     2: 2699942     4: 2698189     6: 2704347
>>>>>>>>     8: 2704009    10: 2696277    12: 2702016    14: 2701388
>>>>>>>>    16: 2700358    18: 2696741    20: 2700091    22: 2700122
>>>>>>>>    24: 2701713    26: 2702025    28: 2699816    30: 2700121
>>>>>>>>    32: 2700000    34: 2699788    36: 2698884    38: 2699109
>>>>>>>>    40: 2704494    42: 2698350    44: 2699997    46: 2701023
>>>>>>>>    48: 2703448    50: 2699501    52: 2700000    54: 2699999
>>>>>>>>    56: 2702645    58: 2696923    60: 2697718    62: 2700547
>>>>>>>>    64: 2700313    66: 2700000    68: 2699904    70: 2699259
>>>>>>>>    72: 2699511    74: 2700644    76: 2702201    78: 2700000
>>>>>>>>    80: 2700776    82: 2700364    84: 2702674    86: 2700255
>>>>>>>>    88: 2699886    90: 2700359    92: 2699662    94: 2696188
>>>>>>>>    96: 2705454    98: 2699260   100: 2701097   102: 2699630
>>>>>>>> 104: 2700463   106: 2698408   108: 2697766   110: 2701181
>>>>>>>> 112: 2699166   114: 2701804   116: 2701907   118: 2701973
>>>>>>>> 120: 2699584   122: 2700474   124: 2700768   126: 2701963
>>>>>>>>
>>>>>>>> Signed-off-by: Huisong Li <lihuisong@huawei.com>
>>>>>>>> ---
>> [snip]
>>> .
> .
Beata Michalska Feb. 2, 2024, 8:08 a.m. UTC | #17
On Wed, Jan 17, 2024 at 05:18:40PM +0800, lihuisong (C) wrote:

Hi ,

Again, apologies for delay,

> Hi,
> 
> 在 2024/1/16 22:10, Beata Michalska 写道:
> > Hi,
> > 
> > Apologies for jumping in so late....
> > 
> > On Wed, Jan 10, 2024 at 03:09:48PM +0800, lihuisong (C) wrote:
> > > Hi Ionela,
> > > 
> > > 在 2024/1/8 22:03, Ionela Voinescu 写道:
> > > > Hi,
> > > > 
> > > > On Friday 05 Jan 2024 at 15:04:47 (+0800), lihuisong (C) wrote:
> > > > > Hi Vanshi,
> > > > > 
> > > > > 在 2024/1/5 8:48, Vanshidhar Konda 写道:
> > > > > > On Thu, Jan 04, 2024 at 05:36:51PM +0800, lihuisong (C) wrote:
> > > > > > > 在 2024/1/4 1:53, Ionela Voinescu 写道:
> > > > > > > > Hi,
> > > > > > > > 
> > > > > > > > On Tuesday 12 Dec 2023 at 15:26:17 (+0800), Huisong Li wrote:
> > > > > > > > > Many developers found that the cpu current frequency is greater than
> > > > > > > > > the maximum frequency of the platform, please see [1], [2] and [3].
> > > > > > > > > 
> > > > > > > > > In the scenarios with high memory access pressure, the patch [1] has
> > > > > > > > > proved the significant latency of cpc_read() which is used to obtain
> > > > > > > > > delivered and reference performance counter cause an absurd frequency.
> > > > > > > > > The sampling interval for this counters is very critical and
> > > > > > > > > is expected
> > > > > > > > > to be equal. However, the different latency of cpc_read() has a direct
> > > > > > > > > impact on their sampling interval.
> > > > > > > > > 
> > > > > > > > Would this [1] alternative solution work for you?
> > > > > > > It would work for me AFAICS.
> > > > > > > Because the "arch_freq_scale" is also from AMU core and constant
> > > > > > > counter, and read together.
> > > > > > > But, from their discuss line, it seems that there are some tricky
> > > > > > > points to clarify or consider.
> > > > > > I think the changes in [1] would work better when CPUs may be idle. With
> > > > > > this
> > > > > > patch we would have to wake any core that is in idle state to read the
> > > > > > AMU
> > > > > > counters. Worst case, if core 0 is trying to read the CPU frequency of
> > > > > > all
> > > > > > cores, it may need to wake up all the other cores to read the AMU
> > > > > > counters.
> > > > >   From the approach in [1], if all CPUs (one or more cores) under one policy
> > > > > are idle, they still cannot be obtained the CPU frequency, right?
> > > > > In this case, the [1] API will return 0 and have to back to call
> > > > > cpufreq_driver->get() for cpuinfo_cur_freq.
> > > > > Then we still need to face the issue this patch mentioned.
> > > > With the implementation at [1], arch_freq_get_on_cpu() will not return 0
> > > > for idle CPUs and the get() callback will not be called to wake up the
> > > > CPUs.
> > > Right, arch_freq_get_on_cpu() will not return 0 for idle CPUs.
> > > However, for no-housekeeping CPUs, it will return 0 and have to call get()
> > > callback, right?
> > > > Worst case, arch_freq_get_on_cpu() will return a frequency based on the
> > > > AMU counter values obtained on the last tick on that CPU. But if that CPU
> > > > is not a housekeeping CPU, a housekeeping CPU in the same policy will be
> > > > selected, as it would have had a more recent tick, and therefore a more
> > > > recent frequency value for the domain.
> > > But this frequency is from the last tick,
> > > this last tick is probably a long time ago and it doesn't update
> > > 'arch_freq_scale' for some reasons like CPU dile.
> > > In addition, I'm not sure if there is possible that amu_scale_freq_tick() is
> > > executed delayed under high stress case.
> > > It also have an impact on the accuracy of the cpu frequency we query.
> > > > I understand that the frequency returned here will not be up to date,
> > > > but there's no proper frequency feedback for an idle CPU. If one only
> > > > wakes up a CPU to sample counters, before the CPU goes back to sleep,
> > > > the obtained frequency feedback is meaningless.
> > > > 
> > > > > > For systems with 128 cores or more, this could be very expensive and
> > > > > > happen
> > > > > > very frequently.
> > > > > > 
> > > > > > AFAICS, the approach in [1] would avoid this cost.
> > > > > But the CPU frequency is just an average value for the last tick period
> > > > > instead of the current one the CPU actually runs at.
> > > > > In addition, there are some conditions to use 'arch_freq_scale' in this
> > > > > approach.
> > > > What are the conditions you are referring to?
> > > It depends on the housekeeping CPUs.
> > > > > So I'm not sure if this approach can entirely cover the frequency
> > > > > discrepancy issue.
> > > > Unfortunately there is no perfect frequency feedback. By the time you
> > > > observe/use the value of scaling_cur_freq/cpuinfo_cur_freq, the frequency
> > > > of the CPU might have already changed. Therefore, an average value might
> > > > be a better indication of the recent performance level of a CPU.
> > > An average value for CPU frequency is ok. It may be better if it has not any
> > > delaying.
> > > 
> > > The original implementation for cpuinfo_cur_freq can more reflect their
> > > meaning in the user-guide [1]. The user-guide said:
> > > "cpuinfo_cur_freq : Current frequency of the CPU as obtained from the
> > > hardware, in KHz.
> > > This is the frequency the CPU actually runs at."
> > > 
> > > 
> > > [1]https://www.kernel.org/doc/Documentation/cpu-freq/user-guide.txt
> > > 
> > > > Would you be able to test [1] on your platform and usecase?
> > > I has tested it on my platform (CPU number: 64, SMT: off and CPU base
> > > frequency: 2.7GHz).
> > > Accoding to the testing result,
> > > 1> I found that patch [1] and [2] cannot cover the no housekeeping CPUs.
> > > They still have to face the large frequency discrepancy issue my patch
> > > mentioned.
> > > 2> Additionally, the frequency value of all CPUs are almost the same by
> > > using the 'arch_freq_scale' factor way. I'm not sure if it is ok.
> > > 
> > > The patch [1] has been modified silightly as below:
> > > -->
> > > @@ -1756,7 +1756,10 @@ static unsigned int
> > > cpufreq_verify_current_freq(struct cpufreq_policy *policy, b
> > >   {
> > >          unsigned int new_freq;
> > > 
> > > -       new_freq = cpufreq_driver->get(policy->cpu);
> > > +       new_freq = arch_freq_get_on_cpu(policy->cpu);
> > > +       if (!new_freq)
> > > +               new_freq = cpufreq_driver->get(policy->cpu);
> > > +
> > As pointed out this change will not make it to the next version of the patch.
> > So I'd say you can safely ignore it and assume that arch_freq_get_on_cpu will
> > only be wired for sysfs nodes for scaling_cur_freq/cpuinfo_cur_freq
> > >          if (!new_freq)
> > >                  return 0;
> > > 
> > > And the result is as follows:
> > > *case 1:**No setting the nohz_full and cpufreq use performance governor*
> > > *--> Step1: *read 'cpuinfo_cur_freq' in no pressure
> > >    0: 2699264     2: 2699264     4: 2699264     6: 2699264
> > >    8: 2696628    10: 2696628    12: 2696628    14: 2699264
> > >   16: 2699264    18: 2696628    20: 2699264    22: 2696628
> > >   24: 2699264    26: 2696628    28: 2699264    30: 2696628
> > >   32: 2696628    34: 2696628    36: 2696628    38: 2696628
> > >   40: 2699264    42: 2699264    44: 2696628    46: 2696628
> > >   48: 2696628    50: 2699264    52: 2699264    54: 2696628
> > >   56: 2696628    58: 2696628    60: 2696628    62: 2696628
> > >   64: 2696628    66: 2699264    68: 2696628    70: 2696628
> > >   72: 2699264    74: 2696628    76: 2696628    78: 2699264
> > >   80: 2696628    82: 2696628    84: 2699264    86: 2696628
> > >   88: 2696628    90: 2696628    92: 2696628    94: 2699264
> > >   96: 2696628    98: 2699264   100: 2699264   102: 2696628
> > > 104: 2699264   106: 2699264   108: 2699264   110: 2696628
> > > 112: 2699264   114: 2699264   116: 2699264   118: 2699264
> > > 120: 2696628   122: 2699264   124: 2696628   126: 2699264
> > > Note: the frequency of all CPUs are almost the same.
> > Were you expecting smth else ?
> The frequency of each CPU might have a different value.
> All value of all CPUs is the same under high pressure.
> I don't know what the phenomenon is on other platform.
> Do you know who else tested it?
So I might have rushed a bit with my previous comment/question: apologies for
that.
The numbers above: those are on a fairly idle/lightly loaded system right?
Would you mind having another go with just the arch_freq_get_on_cpu
implementation beign added and dropping the changes in the cpufreq and
then read 'scaling_cur_freq', doing several reads in some intervals ?
The change has been tested on RD-N2 model (Neoverse N2 ref platform),
it has also been discussed here [1]

> > > *--> Step 2: *read 'cpuinfo_cur_freq' in the high memory access pressure.
> > >    0: 2696628     2: 2696628     4: 2696628     6: 2696628
> > >    8: 2696628    10: 2696628    12: 2696628    14: 2696628
> > >   16: 2696628    18: 2696628    20: 2696628    22: 2696628
> > >   24: 2696628    26: 2696628    28: 2696628    30: 2696628
> > >   32: 2696628    34: 2696628    36: 2696628    38: 2696628
> > >   40: 2696628    42: 2696628    44: 2696628    46: 2696628
> > >   48: 2696628    50: 2696628    52: 2696628    54: 2696628
> > >   56: 2696628    58: 2696628    60: 2696628    62: 2696628
> > >   64: 2696628    66: 2696628    68: 2696628    70: 2696628
> > >   72: 2696628    74: 2696628    76: 2696628    78: 2696628
> > >   80: 2696628    82: 2696628    84: 2696628    86: 2696628
> > >   88: 2696628    90: 2696628    92: 2696628    94: 2696628
> > >   96: 2696628    98: 2696628   100: 2696628   102: 2696628
> > > 104: 2696628   106: 2696628   108: 2696628   110: 2696628
> > > 112: 2696628   114: 2696628   116: 2696628   118: 2696628
> > > 120: 2696628   122: 2696628   124: 2696628   126: 2696628
> > > 
> > > *Case 2: setting nohz_full and cpufreq use ondemand governor*
> > > There is "isolcpus=1-10,41-50 nohz_full=1-10,41-50 rcu_nocbs=1-10,41-50" in
> > > /proc/cmdline.
> > Right, so if I remember correctly nohz_full implies rcu_nocbs, so no need to
> > set that one.
> > Now, afair, isolcpus will make the selected CPUs to disappear from the
> > schedulers view (no balancing, no migrating), so unless you affine smth
> > explicitly to those CPUs, you will not see much of an activity there.
> Correct.
> > Need to double check though as it has been a while ...
> > > *--> Step 1: *setting ondemand governor to all policy and query
> > > 'cpuinfo_cur_freq' in no pressure case.
> > > And the frequency of CPUs all are about 400MHz.
> > > *--> Step 2:* read 'cpuinfo_cur_freq' in the high memory access pressure.
> > > The high memory access pressure is from the command: "stress-ng -c 64
> > > --cpu-load 100% --taskset 0-63"
> > I'm not entirely convinced that this will affine to isolated cpus, especially
> > that the affinity mask spans all available cpus. If that is the case, no wonder
> > your isolated cpus are getting wasted being idle. But I would have to double
> > check how this is being handled.
> > > The result:
> > >   0: 2696628     1:  400000     2:  400000     3:  400909
> > >   4:  400000     5:  400000     6:  400000     7:  400000
> > >   8:  400000     9:  400000    10:  400600    11: 2696628
> > > 12: 2696628    13: 2696628    14: 2696628    15: 2696628
> > > 16: 2696628    17: 2696628    18: 2696628    19: 2696628
> > > 20: 2696628    21: 2696628    22: 2696628    23: 2696628
> > > 24: 2696628    25: 2696628    26: 2696628    27: 2696628
> > > 28: 2696628    29: 2696628    30: 2696628    31: 2696628
> > > 32: 2696628    33: 2696628    34: 2696628    35: 2696628
> > > 36: 2696628    37: 2696628    38: 2696628    39: 2696628
> > > 40: 2696628    41:  400000    42:  400000    43:  400000
> > > 44:  400000    45:  398847    46:  400000    47:  400000
> > > 48:  400000    49:  400000    50:  400000    51: 2696628
> > > 52: 2696628    53: 2696628    54: 2696628    55: 2696628
> > > 56: 2696628    57: 2696628    58: 2696628    59: 2696628
> > > 60: 2696628    61: 2696628    62: 2696628    63: 2699264
> > > 
> > > Note:
> > > (1) The frequency of 1-10 and 41-50 CPUs work on the lowest frequency.
> > >       It turned out that nohz full was already work.
> > >       I guess that stress-ng cannot use the CPU in the range of nohz full.
> > >       Because the CPU frequency will be increased to 2.7G by binding CPU to
> > > other application.
> > > (2) The frequency of the nohz full core is calculated by get() callback
> > > according to ftrace.
> > It is as there is no sched tick on those, and apparently there is nothing
> > running on them either.
> Yes.
> If we select your approach and the above phenomenon is normal,
> the large frequency discrepancy issue can be resolved for CPUs with sched
> tick by the way.
> But the nohz full cores still have to face this issue. So this patch is also
> needed.
> 
Yes, nohz cores full have to be handled by the cpufreq driver.

---
[1] https://lore.kernel.org/lkml/ZIHpd6unkOtYVEqP@e120325.cambridge.arm.com/T/#m4e74cb5a0aaa353c60fedc6cfb95ab7a6e381e3c
---
BR
Beata
> BR
> /huisong
> > 
> > Unless I am missing smth.
> > 
> > ---
> > BR
> > Beata
> > 
> > > [1] https://lore.kernel.org/lkml/20230418113459.12860-7-sumitg@nvidia.com/
> > > [2] https://lore.kernel.org/lkml/20231127160838.1403404-3-beata.michalska@arm.com/
> > > > Many thanks,
> > > > Ionela.
> > > > 
> > > > > /Huisong
> > > > > 
> > > > > > > > [1] https://lore.kernel.org/lkml/20231127160838.1403404-1-beata.michalska@arm.com/
> > > > > > > > 
> > > > > > > > Thanks,
> > > > > > > > Ionela.
> > > > > > > > 
> > > > > > > > > This patch adds a interface, cpc_read_arch_counters_on_cpu, to read
> > > > > > > > > delivered and reference performance counter together. According to my
> > > > > > > > > test[4], the discrepancy of cpu current frequency in the
> > > > > > > > > scenarios with
> > > > > > > > > high memory access pressure is lower than 0.2% by stress-ng
> > > > > > > > > application.
> > > > > > > > > 
> > > > > > > > > [1] https://lore.kernel.org/all/20231025093847.3740104-4-zengheng4@huawei.com/
> > > > > > > > > [2] https://lore.kernel.org/all/20230328193846.8757-1-yang@os.amperecomputing.com/
> > > > > > > > > [3]
> > > > > > > > > https://lore.kernel.org/all/20230418113459.12860-7-sumitg@nvidia.com/
> > > > > > > > > 
> > > > > > > > > [4] My local test:
> > > > > > > > > The testing platform enable SMT and include 128 logical CPU in total,
> > > > > > > > > and CPU base frequency is 2.7GHz. Reading "cpuinfo_cur_freq" for each
> > > > > > > > > physical core on platform during the high memory access pressure from
> > > > > > > > > stress-ng, and the output is as follows:
> > > > > > > > >     0: 2699133     2: 2699942     4: 2698189     6: 2704347
> > > > > > > > >     8: 2704009    10: 2696277    12: 2702016    14: 2701388
> > > > > > > > >    16: 2700358    18: 2696741    20: 2700091    22: 2700122
> > > > > > > > >    24: 2701713    26: 2702025    28: 2699816    30: 2700121
> > > > > > > > >    32: 2700000    34: 2699788    36: 2698884    38: 2699109
> > > > > > > > >    40: 2704494    42: 2698350    44: 2699997    46: 2701023
> > > > > > > > >    48: 2703448    50: 2699501    52: 2700000    54: 2699999
> > > > > > > > >    56: 2702645    58: 2696923    60: 2697718    62: 2700547
> > > > > > > > >    64: 2700313    66: 2700000    68: 2699904    70: 2699259
> > > > > > > > >    72: 2699511    74: 2700644    76: 2702201    78: 2700000
> > > > > > > > >    80: 2700776    82: 2700364    84: 2702674    86: 2700255
> > > > > > > > >    88: 2699886    90: 2700359    92: 2699662    94: 2696188
> > > > > > > > >    96: 2705454    98: 2699260   100: 2701097   102: 2699630
> > > > > > > > > 104: 2700463   106: 2698408   108: 2697766   110: 2701181
> > > > > > > > > 112: 2699166   114: 2701804   116: 2701907   118: 2701973
> > > > > > > > > 120: 2699584   122: 2700474   124: 2700768   126: 2701963
> > > > > > > > > 
> > > > > > > > > Signed-off-by: Huisong Li <lihuisong@huawei.com>
> > > > > > > > > ---
> > > [snip]
> > > > .
> > .
lihuisong (C) Feb. 6, 2024, 8:02 a.m. UTC | #18
在 2024/2/2 16:08, Beata Michalska 写道:
> On Wed, Jan 17, 2024 at 05:18:40PM +0800, lihuisong (C) wrote:
>
> Hi ,
>
> Again, apologies for delay,
>
>> Hi,
>>
>> 在 2024/1/16 22:10, Beata Michalska 写道:
>>> Hi,
>>>
>>> Apologies for jumping in so late....
>>>
>>> On Wed, Jan 10, 2024 at 03:09:48PM +0800, lihuisong (C) wrote:
>>>> Hi Ionela,
>>>>
>>>> 在 2024/1/8 22:03, Ionela Voinescu 写道:
>>>>> Hi,
>>>>>
>>>>> On Friday 05 Jan 2024 at 15:04:47 (+0800), lihuisong (C) wrote:
>>>>>> Hi Vanshi,
>>>>>>
>>>>>> 在 2024/1/5 8:48, Vanshidhar Konda 写道:
>>>>>>> On Thu, Jan 04, 2024 at 05:36:51PM +0800, lihuisong (C) wrote:
>>>>>>>> 在 2024/1/4 1:53, Ionela Voinescu 写道:
>>>>>>>>> Hi,
>>>>>>>>>
>>>>>>>>> On Tuesday 12 Dec 2023 at 15:26:17 (+0800), Huisong Li wrote:
>>>>>>>>>> Many developers found that the cpu current frequency is greater than
>>>>>>>>>> the maximum frequency of the platform, please see [1], [2] and [3].
>>>>>>>>>>
>>>>>>>>>> In the scenarios with high memory access pressure, the patch [1] has
>>>>>>>>>> proved the significant latency of cpc_read() which is used to obtain
>>>>>>>>>> delivered and reference performance counter cause an absurd frequency.
>>>>>>>>>> The sampling interval for this counters is very critical and
>>>>>>>>>> is expected
>>>>>>>>>> to be equal. However, the different latency of cpc_read() has a direct
>>>>>>>>>> impact on their sampling interval.
>>>>>>>>>>
>>>>>>>>> Would this [1] alternative solution work for you?
>>>>>>>> It would work for me AFAICS.
>>>>>>>> Because the "arch_freq_scale" is also from AMU core and constant
>>>>>>>> counter, and read together.
>>>>>>>> But, from their discuss line, it seems that there are some tricky
>>>>>>>> points to clarify or consider.
>>>>>>> I think the changes in [1] would work better when CPUs may be idle. With
>>>>>>> this
>>>>>>> patch we would have to wake any core that is in idle state to read the
>>>>>>> AMU
>>>>>>> counters. Worst case, if core 0 is trying to read the CPU frequency of
>>>>>>> all
>>>>>>> cores, it may need to wake up all the other cores to read the AMU
>>>>>>> counters.
>>>>>>    From the approach in [1], if all CPUs (one or more cores) under one policy
>>>>>> are idle, they still cannot be obtained the CPU frequency, right?
>>>>>> In this case, the [1] API will return 0 and have to back to call
>>>>>> cpufreq_driver->get() for cpuinfo_cur_freq.
>>>>>> Then we still need to face the issue this patch mentioned.
>>>>> With the implementation at [1], arch_freq_get_on_cpu() will not return 0
>>>>> for idle CPUs and the get() callback will not be called to wake up the
>>>>> CPUs.
>>>> Right, arch_freq_get_on_cpu() will not return 0 for idle CPUs.
>>>> However, for no-housekeeping CPUs, it will return 0 and have to call get()
>>>> callback, right?
>>>>> Worst case, arch_freq_get_on_cpu() will return a frequency based on the
>>>>> AMU counter values obtained on the last tick on that CPU. But if that CPU
>>>>> is not a housekeeping CPU, a housekeeping CPU in the same policy will be
>>>>> selected, as it would have had a more recent tick, and therefore a more
>>>>> recent frequency value for the domain.
>>>> But this frequency is from the last tick,
>>>> this last tick is probably a long time ago and it doesn't update
>>>> 'arch_freq_scale' for some reasons like CPU dile.
>>>> In addition, I'm not sure if there is possible that amu_scale_freq_tick() is
>>>> executed delayed under high stress case.
>>>> It also have an impact on the accuracy of the cpu frequency we query.
>>>>> I understand that the frequency returned here will not be up to date,
>>>>> but there's no proper frequency feedback for an idle CPU. If one only
>>>>> wakes up a CPU to sample counters, before the CPU goes back to sleep,
>>>>> the obtained frequency feedback is meaningless.
>>>>>
>>>>>>> For systems with 128 cores or more, this could be very expensive and
>>>>>>> happen
>>>>>>> very frequently.
>>>>>>>
>>>>>>> AFAICS, the approach in [1] would avoid this cost.
>>>>>> But the CPU frequency is just an average value for the last tick period
>>>>>> instead of the current one the CPU actually runs at.
>>>>>> In addition, there are some conditions to use 'arch_freq_scale' in this
>>>>>> approach.
>>>>> What are the conditions you are referring to?
>>>> It depends on the housekeeping CPUs.
>>>>>> So I'm not sure if this approach can entirely cover the frequency
>>>>>> discrepancy issue.
>>>>> Unfortunately there is no perfect frequency feedback. By the time you
>>>>> observe/use the value of scaling_cur_freq/cpuinfo_cur_freq, the frequency
>>>>> of the CPU might have already changed. Therefore, an average value might
>>>>> be a better indication of the recent performance level of a CPU.
>>>> An average value for CPU frequency is ok. It may be better if it has not any
>>>> delaying.
>>>>
>>>> The original implementation for cpuinfo_cur_freq can more reflect their
>>>> meaning in the user-guide [1]. The user-guide said:
>>>> "cpuinfo_cur_freq : Current frequency of the CPU as obtained from the
>>>> hardware, in KHz.
>>>> This is the frequency the CPU actually runs at."
>>>>
>>>>
>>>> [1]https://www.kernel.org/doc/Documentation/cpu-freq/user-guide.txt
>>>>
>>>>> Would you be able to test [1] on your platform and usecase?
>>>> I has tested it on my platform (CPU number: 64, SMT: off and CPU base
>>>> frequency: 2.7GHz).
>>>> Accoding to the testing result,
>>>> 1> I found that patch [1] and [2] cannot cover the no housekeeping CPUs.
>>>> They still have to face the large frequency discrepancy issue my patch
>>>> mentioned.
>>>> 2> Additionally, the frequency value of all CPUs are almost the same by
>>>> using the 'arch_freq_scale' factor way. I'm not sure if it is ok.
>>>>
>>>> The patch [1] has been modified silightly as below:
>>>> -->
>>>> @@ -1756,7 +1756,10 @@ static unsigned int
>>>> cpufreq_verify_current_freq(struct cpufreq_policy *policy, b
>>>>    {
>>>>           unsigned int new_freq;
>>>>
>>>> -       new_freq = cpufreq_driver->get(policy->cpu);
>>>> +       new_freq = arch_freq_get_on_cpu(policy->cpu);
>>>> +       if (!new_freq)
>>>> +               new_freq = cpufreq_driver->get(policy->cpu);
>>>> +
>>> As pointed out this change will not make it to the next version of the patch.
>>> So I'd say you can safely ignore it and assume that arch_freq_get_on_cpu will
>>> only be wired for sysfs nodes for scaling_cur_freq/cpuinfo_cur_freq
>>>>           if (!new_freq)
>>>>                   return 0;
>>>>
>>>> And the result is as follows:
>>>> *case 1:**No setting the nohz_full and cpufreq use performance governor*
>>>> *--> Step1: *read 'cpuinfo_cur_freq' in no pressure
>>>>     0: 2699264     2: 2699264     4: 2699264     6: 2699264
>>>>     8: 2696628    10: 2696628    12: 2696628    14: 2699264
>>>>    16: 2699264    18: 2696628    20: 2699264    22: 2696628
>>>>    24: 2699264    26: 2696628    28: 2699264    30: 2696628
>>>>    32: 2696628    34: 2696628    36: 2696628    38: 2696628
>>>>    40: 2699264    42: 2699264    44: 2696628    46: 2696628
>>>>    48: 2696628    50: 2699264    52: 2699264    54: 2696628
>>>>    56: 2696628    58: 2696628    60: 2696628    62: 2696628
>>>>    64: 2696628    66: 2699264    68: 2696628    70: 2696628
>>>>    72: 2699264    74: 2696628    76: 2696628    78: 2699264
>>>>    80: 2696628    82: 2696628    84: 2699264    86: 2696628
>>>>    88: 2696628    90: 2696628    92: 2696628    94: 2699264
>>>>    96: 2696628    98: 2699264   100: 2699264   102: 2696628
>>>> 104: 2699264   106: 2699264   108: 2699264   110: 2696628
>>>> 112: 2699264   114: 2699264   116: 2699264   118: 2699264
>>>> 120: 2696628   122: 2699264   124: 2696628   126: 2699264
>>>> Note: the frequency of all CPUs are almost the same.
>>> Were you expecting smth else ?
>> The frequency of each CPU might have a different value.
>> All value of all CPUs is the same under high pressure.
>> I don't know what the phenomenon is on other platform.
>> Do you know who else tested it?
> So I might have rushed a bit with my previous comment/question: apologies for
> that.
> The numbers above: those are on a fairly idle/lightly loaded system right?
Yes.
> Would you mind having another go with just the arch_freq_get_on_cpu
> implementation beign added and dropping the changes in the cpufreq and
All my tests are done when cpufreq policy is "performance" and OS isn't 
on a high load.
Reading "scaling_cur_freq" or "scaling_cur_freq" for each physical core 
on platform

The testing result for "cpuinfo_cur_freq" with your changes on a fairly 
idle and high loaded system can also be found in this thread.
*A: the result with your changes*
--> Reading "scaling_cur_freq"
   0: 2688720     2: 2696628     4: 2699264     6: 2696628
   8: 2699264    10: 2696628    12: 2699264    14: 2699264
  16: 2699264    18: 2696628    20: 2696628    22: 2696628
  24: 2699264    26: 2696628    28: 2696628    30: 2696628
  32: 2699264    34: 2691356    36: 2696628    38: 2699264
  40: 2699264    42: 2696628    44: 2696628    46: 2699264
  48: 2699264    50: 2696628    52: 2696628    54: 2696628
  56: 2696628    58: 2699264    60: 2691356    62: 2696628
  64: 2696628    66: 2696628    68: 2696628    70: 2696628
  72: 2696628    74: 2696628    76: 2699264    78: 2696628
  80: 2696628    82: 2696628    84: 2699264    86: 2696628
  88: 2625456    90: 2696628    92: 2699264    94: 2696628
  96: 2696628    98: 2696628   100: 2699264   102: 2699264
104: 2699264   106: 2696628   108: 2699264   110: 2696628
112: 2699264   114: 2699264   116: 2696628   118: 2696628
120: 2696628   122: 2699264   124: 2696628   126: 2696628
-->Reading  "cpuinfo_cur_freq"
   0: 2696628     2: 2696628     4: 2699264     6: 2688720
   8: 2699264    10: 2700000    12: 2696628    14: 2698322
  16: 2699264    18: 2699264    20: 2696628    22: 2699264
  24: 2699264    26: 2699264    28: 2699264    30: 2699264
  32: 2699264    34: 2693992    36: 2696628    38: 2696628
  40: 2699264    42: 2699264    44: 2699264    46: 2696628
  48: 2696628    50: 2699264    52: 2696628    54: 2696628
  56: 2699264    58: 2699264    60: 2696628    62: 2699264
  64: 2696628    66: 2699264    68: 2696628    70: 2699264
  72: 2696628    74: 2696628    76: 2696628    78: 2693992
  80: 2696628    82: 2696628    84: 2696628    86: 2696628
  88: 2696628    90: 2699264    92: 2696628    94: 2699264
  96: 2699264    98: 2696628   100: 2699264   102: 2699264
104: 2691356   106: 2699264   108: 2699264   110: 2699264
112: 2699264   114: 2696628   116: 2699264   118: 2699264
120: 2696628   122: 2696628   124: 2696628   126: 2696628

*B: the result without your changes*
-->Reading "scaling_cur_freq"
   0: 2698245     2: 2706690     4: 2699649     6: 2702105
   8: 2704362    10: 2697993    12: 2701672    14: 2704362
  16: 2701052    18: 2701052    20: 2694385    22: 2699650
  24: 2706802    26: 2702389    28: 2698299    30: 2698299
  32: 2697333    34: 2697993    36: 2701337    38: 2699328
  40: 2700330    42: 2700330    44: 2698019    46: 2697697
  48: 2699659    50: 2701700    52: 2703401    54: 2701700
  56: 2704013    58: 2697658    60: 2695000    62: 2697666
  64: 2697902    66: 2701052    68: 2698245    70: 2695789
  72: 2701315    74: 2696655    76: 2693666    78: 2695317
  80: 2704912    82: 2699649    84: 2698245    86: 2695454
  88: 2697966    90: 2697959    92: 2699319    94: 2700680
  96: 2695317    98: 2698996   100: 2700000   102: 2700334
104: 2701320   106: 2695065   108: 2700986   110: 2703960
112: 2697635   114: 2704421   116: 2700680   118: 2702040
120: 2700334   122: 2697993   124: 2700334   126: 2705351
-->Reading "cpuinfo_cur_freq"
   0: 2696853     2: 2695454     4: 2699649     6: 2706993
   8: 2706060    10: 2704362    12: 2704362    14: 2697658
  16: 2707719    18: 2697192    20: 2702456    22: 2699650
  24: 2705782    26: 2698299    28: 2703061    30: 2705802
  32: 2700000    34: 2700671    36: 2701337    38: 2697658
  40: 2700330    42: 2700330    44: 2699672    46: 2697697
  48: 2703061    50: 2696610    52: 2692542    54: 2704406
  56: 2695317    58: 2699331    60: 2698996    62: 2702675
  64: 2704912    66: 2703859    68: 2699649    70: 2698596
  72: 2703908    74: 2703355    76: 2697658    78: 2695317
  80: 2702105    82: 2707719    84: 2702105    86: 2699649
  88: 2697966    90: 2691525    92: 2701700    94: 2700680
  96: 2695317    98: 2698996   100: 2698666   102: 2700334
104: 2690429   106: 2707590   108: 2700986   110: 2701320
112: 2696283   114: 2692881   116: 2697627   118: 2704421
120: 2698996   122: 2696321   124: 2696655   126: 2695000

> then read 'scaling_cur_freq', doing several reads in some intervals ?
It seems that above phenomenon has not a lot to do with reading intervals.
> The change has been tested on RD-N2 model (Neoverse N2 ref platform),
> it has also been discussed here [1]
I doesn't get the testing result on this platform in its thread.
>>>> *--> Step 2: *read 'cpuinfo_cur_freq' in the high memory access pressure.
>>>>     0: 2696628     2: 2696628     4: 2696628     6: 2696628
>>>>     8: 2696628    10: 2696628    12: 2696628    14: 2696628
>>>>    16: 2696628    18: 2696628    20: 2696628    22: 2696628
>>>>    24: 2696628    26: 2696628    28: 2696628    30: 2696628
>>>>    32: 2696628    34: 2696628    36: 2696628    38: 2696628
>>>>    40: 2696628    42: 2696628    44: 2696628    46: 2696628
>>>>    48: 2696628    50: 2696628    52: 2696628    54: 2696628
>>>>    56: 2696628    58: 2696628    60: 2696628    62: 2696628
>>>>    64: 2696628    66: 2696628    68: 2696628    70: 2696628
>>>>    72: 2696628    74: 2696628    76: 2696628    78: 2696628
>>>>    80: 2696628    82: 2696628    84: 2696628    86: 2696628
>>>>    88: 2696628    90: 2696628    92: 2696628    94: 2696628
>>>>    96: 2696628    98: 2696628   100: 2696628   102: 2696628
>>>> 104: 2696628   106: 2696628   108: 2696628   110: 2696628
>>>> 112: 2696628   114: 2696628   116: 2696628   118: 2696628
>>>> 120: 2696628   122: 2696628   124: 2696628   126: 2696628
>>>>
>>>> *Case 2: setting nohz_full and cpufreq use ondemand governor*
>>>> There is "isolcpus=1-10,41-50 nohz_full=1-10,41-50 rcu_nocbs=1-10,41-50" in
>>>> /proc/cmdline.
>>> Right, so if I remember correctly nohz_full implies rcu_nocbs, so no need to
>>> set that one.
>>> Now, afair, isolcpus will make the selected CPUs to disappear from the
>>> schedulers view (no balancing, no migrating), so unless you affine smth
>>> explicitly to those CPUs, you will not see much of an activity there.
>> Correct.
>>> Need to double check though as it has been a while ...
>>>> *--> Step 1: *setting ondemand governor to all policy and query
>>>> 'cpuinfo_cur_freq' in no pressure case.
>>>> And the frequency of CPUs all are about 400MHz.
>>>> *--> Step 2:* read 'cpuinfo_cur_freq' in the high memory access pressure.
>>>> The high memory access pressure is from the command: "stress-ng -c 64
>>>> --cpu-load 100% --taskset 0-63"
>>> I'm not entirely convinced that this will affine to isolated cpus, especially
>>> that the affinity mask spans all available cpus. If that is the case, no wonder
>>> your isolated cpus are getting wasted being idle. But I would have to double
>>> check how this is being handled.
>>>> The result:
>>>>    0: 2696628     1:  400000     2:  400000     3:  400909
>>>>    4:  400000     5:  400000     6:  400000     7:  400000
>>>>    8:  400000     9:  400000    10:  400600    11: 2696628
>>>> 12: 2696628    13: 2696628    14: 2696628    15: 2696628
>>>> 16: 2696628    17: 2696628    18: 2696628    19: 2696628
>>>> 20: 2696628    21: 2696628    22: 2696628    23: 2696628
>>>> 24: 2696628    25: 2696628    26: 2696628    27: 2696628
>>>> 28: 2696628    29: 2696628    30: 2696628    31: 2696628
>>>> 32: 2696628    33: 2696628    34: 2696628    35: 2696628
>>>> 36: 2696628    37: 2696628    38: 2696628    39: 2696628
>>>> 40: 2696628    41:  400000    42:  400000    43:  400000
>>>> 44:  400000    45:  398847    46:  400000    47:  400000
>>>> 48:  400000    49:  400000    50:  400000    51: 2696628
>>>> 52: 2696628    53: 2696628    54: 2696628    55: 2696628
>>>> 56: 2696628    57: 2696628    58: 2696628    59: 2696628
>>>> 60: 2696628    61: 2696628    62: 2696628    63: 2699264
>>>>
>>>> Note:
>>>> (1) The frequency of 1-10 and 41-50 CPUs work on the lowest frequency.
>>>>        It turned out that nohz full was already work.
>>>>        I guess that stress-ng cannot use the CPU in the range of nohz full.
>>>>        Because the CPU frequency will be increased to 2.7G by binding CPU to
>>>> other application.
>>>> (2) The frequency of the nohz full core is calculated by get() callback
>>>> according to ftrace.
>>> It is as there is no sched tick on those, and apparently there is nothing
>>> running on them either.
>> Yes.
>> If we select your approach and the above phenomenon is normal,
>> the large frequency discrepancy issue can be resolved for CPUs with sched
>> tick by the way.
>> But the nohz full cores still have to face this issue. So this patch is also
>> needed.
>>
> Yes, nohz cores full have to be handled by the cpufreq driver.
Correct. So we still have to face the issue in this patch and push this 
patch.
Beata, would you please review this patch?


/Huisong
>
> ---
> [1] https://lore.kernel.org/lkml/ZIHpd6unkOtYVEqP@e120325.cambridge.arm.com/T/#m4e74cb5a0aaa353c60fedc6cfb95ab7a6e381e3c
> ---
> BR
> Beata
>> BR
>> /huisong
>>> Unless I am missing smth.
>>>
>>> ---
>>> BR
>>> Beata
>>>
>>>> [1] https://lore.kernel.org/lkml/20230418113459.12860-7-sumitg@nvidia.com/
>>>> [2] https://lore.kernel.org/lkml/20231127160838.1403404-3-beata.michalska@arm.com/
>>>>> Many thanks,
>>>>> Ionela.
>>>>>
>>>>>> /Huisong
>>>>>>
>>>>>>>>> [1] https://lore.kernel.org/lkml/20231127160838.1403404-1-beata.michalska@arm.com/
>>>>>>>>>
>>>>>>>>> Thanks,
>>>>>>>>> Ionela.
>>>>>>>>>
>>>>>>>>>> This patch adds a interface, cpc_read_arch_counters_on_cpu, to read
>>>>>>>>>> delivered and reference performance counter together. According to my
>>>>>>>>>> test[4], the discrepancy of cpu current frequency in the
>>>>>>>>>> scenarios with
>>>>>>>>>> high memory access pressure is lower than 0.2% by stress-ng
>>>>>>>>>> application.
>>>>>>>>>>
>>>>>>>>>> [1] https://lore.kernel.org/all/20231025093847.3740104-4-zengheng4@huawei.com/
>>>>>>>>>> [2] https://lore.kernel.org/all/20230328193846.8757-1-yang@os.amperecomputing.com/
>>>>>>>>>> [3]
>>>>>>>>>> https://lore.kernel.org/all/20230418113459.12860-7-sumitg@nvidia.com/
>>>>>>>>>>
>>>>>>>>>> [4] My local test:
>>>>>>>>>> The testing platform enable SMT and include 128 logical CPU in total,
>>>>>>>>>> and CPU base frequency is 2.7GHz. Reading "cpuinfo_cur_freq" for each
>>>>>>>>>> physical core on platform during the high memory access pressure from
>>>>>>>>>> stress-ng, and the output is as follows:
>>>>>>>>>>      0: 2699133     2: 2699942     4: 2698189     6: 2704347
>>>>>>>>>>      8: 2704009    10: 2696277    12: 2702016    14: 2701388
>>>>>>>>>>     16: 2700358    18: 2696741    20: 2700091    22: 2700122
>>>>>>>>>>     24: 2701713    26: 2702025    28: 2699816    30: 2700121
>>>>>>>>>>     32: 2700000    34: 2699788    36: 2698884    38: 2699109
>>>>>>>>>>     40: 2704494    42: 2698350    44: 2699997    46: 2701023
>>>>>>>>>>     48: 2703448    50: 2699501    52: 2700000    54: 2699999
>>>>>>>>>>     56: 2702645    58: 2696923    60: 2697718    62: 2700547
>>>>>>>>>>     64: 2700313    66: 2700000    68: 2699904    70: 2699259
>>>>>>>>>>     72: 2699511    74: 2700644    76: 2702201    78: 2700000
>>>>>>>>>>     80: 2700776    82: 2700364    84: 2702674    86: 2700255
>>>>>>>>>>     88: 2699886    90: 2700359    92: 2699662    94: 2696188
>>>>>>>>>>     96: 2705454    98: 2699260   100: 2701097   102: 2699630
>>>>>>>>>> 104: 2700463   106: 2698408   108: 2697766   110: 2701181
>>>>>>>>>> 112: 2699166   114: 2701804   116: 2701907   118: 2701973
>>>>>>>>>> 120: 2699584   122: 2700474   124: 2700768   126: 2701963
>>>>>>>>>>
>>>>>>>>>> Signed-off-by: Huisong Li <lihuisong@huawei.com>
>>>>>>>>>> ---
>>>> [snip]
>>>>> .
>>> .
> .
Beata Michalska Feb. 9, 2024, 10:55 a.m. UTC | #19
On Tue, Feb 06, 2024 at 04:02:15PM +0800, lihuisong (C) wrote:
> 
> 在 2024/2/2 16:08, Beata Michalska 写道:
> > On Wed, Jan 17, 2024 at 05:18:40PM +0800, lihuisong (C) wrote:
> > 
> > Hi ,
> > 
> > Again, apologies for delay,
> > 
> > > Hi,
> > > 
> > > 在 2024/1/16 22:10, Beata Michalska 写道:
> > > > Hi,
> > > > 
> > > > Apologies for jumping in so late....
> > > > 
> > > > On Wed, Jan 10, 2024 at 03:09:48PM +0800, lihuisong (C) wrote:
> > > > > Hi Ionela,
> > > > > 
> > > > > 在 2024/1/8 22:03, Ionela Voinescu 写道:
> > > > > > Hi,
> > > > > > 
> > > > > > On Friday 05 Jan 2024 at 15:04:47 (+0800), lihuisong (C) wrote:
> > > > > > > Hi Vanshi,
> > > > > > > 
> > > > > > > 在 2024/1/5 8:48, Vanshidhar Konda 写道:
> > > > > > > > On Thu, Jan 04, 2024 at 05:36:51PM +0800, lihuisong (C) wrote:
> > > > > > > > > 在 2024/1/4 1:53, Ionela Voinescu 写道:
> > > > > > > > > > Hi,
> > > > > > > > > > 
> > > > > > > > > > On Tuesday 12 Dec 2023 at 15:26:17 (+0800), Huisong Li wrote:
> > > > > > > > > > > Many developers found that the cpu current frequency is greater than
> > > > > > > > > > > the maximum frequency of the platform, please see [1], [2] and [3].
> > > > > > > > > > > 
> > > > > > > > > > > In the scenarios with high memory access pressure, the patch [1] has
> > > > > > > > > > > proved the significant latency of cpc_read() which is used to obtain
> > > > > > > > > > > delivered and reference performance counter cause an absurd frequency.
> > > > > > > > > > > The sampling interval for this counters is very critical and
> > > > > > > > > > > is expected
> > > > > > > > > > > to be equal. However, the different latency of cpc_read() has a direct
> > > > > > > > > > > impact on their sampling interval.
> > > > > > > > > > > 
> > > > > > > > > > Would this [1] alternative solution work for you?
> > > > > > > > > It would work for me AFAICS.
> > > > > > > > > Because the "arch_freq_scale" is also from AMU core and constant
> > > > > > > > > counter, and read together.
> > > > > > > > > But, from their discuss line, it seems that there are some tricky
> > > > > > > > > points to clarify or consider.
> > > > > > > > I think the changes in [1] would work better when CPUs may be idle. With
> > > > > > > > this
> > > > > > > > patch we would have to wake any core that is in idle state to read the
> > > > > > > > AMU
> > > > > > > > counters. Worst case, if core 0 is trying to read the CPU frequency of
> > > > > > > > all
> > > > > > > > cores, it may need to wake up all the other cores to read the AMU
> > > > > > > > counters.
> > > > > > >    From the approach in [1], if all CPUs (one or more cores) under one policy
> > > > > > > are idle, they still cannot be obtained the CPU frequency, right?
> > > > > > > In this case, the [1] API will return 0 and have to back to call
> > > > > > > cpufreq_driver->get() for cpuinfo_cur_freq.
> > > > > > > Then we still need to face the issue this patch mentioned.
> > > > > > With the implementation at [1], arch_freq_get_on_cpu() will not return 0
> > > > > > for idle CPUs and the get() callback will not be called to wake up the
> > > > > > CPUs.
> > > > > Right, arch_freq_get_on_cpu() will not return 0 for idle CPUs.
> > > > > However, for no-housekeeping CPUs, it will return 0 and have to call get()
> > > > > callback, right?
> > > > > > Worst case, arch_freq_get_on_cpu() will return a frequency based on the
> > > > > > AMU counter values obtained on the last tick on that CPU. But if that CPU
> > > > > > is not a housekeeping CPU, a housekeeping CPU in the same policy will be
> > > > > > selected, as it would have had a more recent tick, and therefore a more
> > > > > > recent frequency value for the domain.
> > > > > But this frequency is from the last tick,
> > > > > this last tick is probably a long time ago and it doesn't update
> > > > > 'arch_freq_scale' for some reasons like CPU dile.
> > > > > In addition, I'm not sure if there is possible that amu_scale_freq_tick() is
> > > > > executed delayed under high stress case.
> > > > > It also have an impact on the accuracy of the cpu frequency we query.
> > > > > > I understand that the frequency returned here will not be up to date,
> > > > > > but there's no proper frequency feedback for an idle CPU. If one only
> > > > > > wakes up a CPU to sample counters, before the CPU goes back to sleep,
> > > > > > the obtained frequency feedback is meaningless.
> > > > > > 
> > > > > > > > For systems with 128 cores or more, this could be very expensive and
> > > > > > > > happen
> > > > > > > > very frequently.
> > > > > > > > 
> > > > > > > > AFAICS, the approach in [1] would avoid this cost.
> > > > > > > But the CPU frequency is just an average value for the last tick period
> > > > > > > instead of the current one the CPU actually runs at.
> > > > > > > In addition, there are some conditions to use 'arch_freq_scale' in this
> > > > > > > approach.
> > > > > > What are the conditions you are referring to?
> > > > > It depends on the housekeeping CPUs.
> > > > > > > So I'm not sure if this approach can entirely cover the frequency
> > > > > > > discrepancy issue.
> > > > > > Unfortunately there is no perfect frequency feedback. By the time you
> > > > > > observe/use the value of scaling_cur_freq/cpuinfo_cur_freq, the frequency
> > > > > > of the CPU might have already changed. Therefore, an average value might
> > > > > > be a better indication of the recent performance level of a CPU.
> > > > > An average value for CPU frequency is ok. It may be better if it has not any
> > > > > delaying.
> > > > > 
> > > > > The original implementation for cpuinfo_cur_freq can more reflect their
> > > > > meaning in the user-guide [1]. The user-guide said:
> > > > > "cpuinfo_cur_freq : Current frequency of the CPU as obtained from the
> > > > > hardware, in KHz.
> > > > > This is the frequency the CPU actually runs at."
> > > > > 
> > > > > 
> > > > > [1]https://www.kernel.org/doc/Documentation/cpu-freq/user-guide.txt
> > > > > 
> > > > > > Would you be able to test [1] on your platform and usecase?
> > > > > I has tested it on my platform (CPU number: 64, SMT: off and CPU base
> > > > > frequency: 2.7GHz).
> > > > > Accoding to the testing result,
> > > > > 1> I found that patch [1] and [2] cannot cover the no housekeeping CPUs.
> > > > > They still have to face the large frequency discrepancy issue my patch
> > > > > mentioned.
> > > > > 2> Additionally, the frequency value of all CPUs are almost the same by
> > > > > using the 'arch_freq_scale' factor way. I'm not sure if it is ok.
> > > > > 
> > > > > The patch [1] has been modified silightly as below:
> > > > > -->
> > > > > @@ -1756,7 +1756,10 @@ static unsigned int
> > > > > cpufreq_verify_current_freq(struct cpufreq_policy *policy, b
> > > > >    {
> > > > >           unsigned int new_freq;
> > > > > 
> > > > > -       new_freq = cpufreq_driver->get(policy->cpu);
> > > > > +       new_freq = arch_freq_get_on_cpu(policy->cpu);
> > > > > +       if (!new_freq)
> > > > > +               new_freq = cpufreq_driver->get(policy->cpu);
> > > > > +
> > > > As pointed out this change will not make it to the next version of the patch.
> > > > So I'd say you can safely ignore it and assume that arch_freq_get_on_cpu will
> > > > only be wired for sysfs nodes for scaling_cur_freq/cpuinfo_cur_freq
> > > > >           if (!new_freq)
> > > > >                   return 0;
> > > > > 
> > > > > And the result is as follows:
> > > > > *case 1:**No setting the nohz_full and cpufreq use performance governor*
> > > > > *--> Step1: *read 'cpuinfo_cur_freq' in no pressure
> > > > >     0: 2699264     2: 2699264     4: 2699264     6: 2699264
> > > > >     8: 2696628    10: 2696628    12: 2696628    14: 2699264
> > > > >    16: 2699264    18: 2696628    20: 2699264    22: 2696628
> > > > >    24: 2699264    26: 2696628    28: 2699264    30: 2696628
> > > > >    32: 2696628    34: 2696628    36: 2696628    38: 2696628
> > > > >    40: 2699264    42: 2699264    44: 2696628    46: 2696628
> > > > >    48: 2696628    50: 2699264    52: 2699264    54: 2696628
> > > > >    56: 2696628    58: 2696628    60: 2696628    62: 2696628
> > > > >    64: 2696628    66: 2699264    68: 2696628    70: 2696628
> > > > >    72: 2699264    74: 2696628    76: 2696628    78: 2699264
> > > > >    80: 2696628    82: 2696628    84: 2699264    86: 2696628
> > > > >    88: 2696628    90: 2696628    92: 2696628    94: 2699264
> > > > >    96: 2696628    98: 2699264   100: 2699264   102: 2696628
> > > > > 104: 2699264   106: 2699264   108: 2699264   110: 2696628
> > > > > 112: 2699264   114: 2699264   116: 2699264   118: 2699264
> > > > > 120: 2696628   122: 2699264   124: 2696628   126: 2699264
> > > > > Note: the frequency of all CPUs are almost the same.
> > > > Were you expecting smth else ?
> > > The frequency of each CPU might have a different value.
> > > All value of all CPUs is the same under high pressure.
> > > I don't know what the phenomenon is on other platform.
> > > Do you know who else tested it?
> > So I might have rushed a bit with my previous comment/question: apologies for
> > that.
> > The numbers above: those are on a fairly idle/lightly loaded system right?
> Yes.
> > Would you mind having another go with just the arch_freq_get_on_cpu
> > implementation beign added and dropping the changes in the cpufreq and
> All my tests are done when cpufreq policy is "performance" and OS isn't on a
> high load.
> Reading "scaling_cur_freq" or "scaling_cur_freq" for each physical core on
> platform
> 
> The testing result for "cpuinfo_cur_freq" with your changes on a fairly idle
> and high loaded system can also be found in this thread.
> *A: the result with your changes*
> --> Reading "scaling_cur_freq"
>   0: 2688720     2: 2696628     4: 2699264     6: 2696628
>   8: 2699264    10: 2696628    12: 2699264    14: 2699264
>  16: 2699264    18: 2696628    20: 2696628    22: 2696628
>  24: 2699264    26: 2696628    28: 2696628    30: 2696628
>  32: 2699264    34: 2691356    36: 2696628    38: 2699264
>  40: 2699264    42: 2696628    44: 2696628    46: 2699264
>  48: 2699264    50: 2696628    52: 2696628    54: 2696628
>  56: 2696628    58: 2699264    60: 2691356    62: 2696628
>  64: 2696628    66: 2696628    68: 2696628    70: 2696628
>  72: 2696628    74: 2696628    76: 2699264    78: 2696628
>  80: 2696628    82: 2696628    84: 2699264    86: 2696628
>  88: 2625456    90: 2696628    92: 2699264    94: 2696628
>  96: 2696628    98: 2696628   100: 2699264   102: 2699264
> 104: 2699264   106: 2696628   108: 2699264   110: 2696628
> 112: 2699264   114: 2699264   116: 2696628   118: 2696628
> 120: 2696628   122: 2699264   124: 2696628   126: 2696628
> -->Reading  "cpuinfo_cur_freq"
>   0: 2696628     2: 2696628     4: 2699264     6: 2688720
>   8: 2699264    10: 2700000    12: 2696628    14: 2698322
>  16: 2699264    18: 2699264    20: 2696628    22: 2699264
>  24: 2699264    26: 2699264    28: 2699264    30: 2699264
>  32: 2699264    34: 2693992    36: 2696628    38: 2696628
>  40: 2699264    42: 2699264    44: 2699264    46: 2696628
>  48: 2696628    50: 2699264    52: 2696628    54: 2696628
>  56: 2699264    58: 2699264    60: 2696628    62: 2699264
>  64: 2696628    66: 2699264    68: 2696628    70: 2699264
>  72: 2696628    74: 2696628    76: 2696628    78: 2693992
>  80: 2696628    82: 2696628    84: 2696628    86: 2696628
>  88: 2696628    90: 2699264    92: 2696628    94: 2699264
>  96: 2699264    98: 2696628   100: 2699264   102: 2699264
> 104: 2691356   106: 2699264   108: 2699264   110: 2699264
> 112: 2699264   114: 2696628   116: 2699264   118: 2699264
> 120: 2696628   122: 2696628   124: 2696628   126: 2696628
> 
> *B: the result without your changes*
> -->Reading "scaling_cur_freq"
>   0: 2698245     2: 2706690     4: 2699649     6: 2702105
>   8: 2704362    10: 2697993    12: 2701672    14: 2704362
>  16: 2701052    18: 2701052    20: 2694385    22: 2699650
>  24: 2706802    26: 2702389    28: 2698299    30: 2698299
>  32: 2697333    34: 2697993    36: 2701337    38: 2699328
>  40: 2700330    42: 2700330    44: 2698019    46: 2697697
>  48: 2699659    50: 2701700    52: 2703401    54: 2701700
>  56: 2704013    58: 2697658    60: 2695000    62: 2697666
>  64: 2697902    66: 2701052    68: 2698245    70: 2695789
>  72: 2701315    74: 2696655    76: 2693666    78: 2695317
>  80: 2704912    82: 2699649    84: 2698245    86: 2695454
>  88: 2697966    90: 2697959    92: 2699319    94: 2700680
>  96: 2695317    98: 2698996   100: 2700000   102: 2700334
> 104: 2701320   106: 2695065   108: 2700986   110: 2703960
> 112: 2697635   114: 2704421   116: 2700680   118: 2702040
> 120: 2700334   122: 2697993   124: 2700334   126: 2705351
> -->Reading "cpuinfo_cur_freq"
>   0: 2696853     2: 2695454     4: 2699649     6: 2706993
>   8: 2706060    10: 2704362    12: 2704362    14: 2697658
>  16: 2707719    18: 2697192    20: 2702456    22: 2699650
>  24: 2705782    26: 2698299    28: 2703061    30: 2705802
>  32: 2700000    34: 2700671    36: 2701337    38: 2697658
>  40: 2700330    42: 2700330    44: 2699672    46: 2697697
>  48: 2703061    50: 2696610    52: 2692542    54: 2704406
>  56: 2695317    58: 2699331    60: 2698996    62: 2702675
>  64: 2704912    66: 2703859    68: 2699649    70: 2698596
>  72: 2703908    74: 2703355    76: 2697658    78: 2695317
>  80: 2702105    82: 2707719    84: 2702105    86: 2699649
>  88: 2697966    90: 2691525    92: 2701700    94: 2700680
>  96: 2695317    98: 2698996   100: 2698666   102: 2700334
> 104: 2690429   106: 2707590   108: 2700986   110: 2701320
> 112: 2696283   114: 2692881   116: 2697627   118: 2704421
> 120: 2698996   122: 2696321   124: 2696655   126: 2695000
>
So in both cases : whether you use arch_freq_get_on_cpu or not
(so with and without the patch) you get roughly the same frequencies
on all cores - or am I missing smth from the dump above ?
And those are reflecting max freq you have provided earlier (?)
Note that the arch_freq_get_on_cpu will return an average frequency for
the last tick, so even if your system is roughly idle with your performance
governor those numbers make sense (some/most of the cores might be idle
but you will see the last freq the core was running at before going to idle).
I do not think there is an agreement what should be shown for idle core when
querying their freq through sysfs. Showing last known freq makes sense, even
more than waking up core just to try to get one.

@Ionela: Please jump in if I got things wrong.

> > then read 'scaling_cur_freq', doing several reads in some intervals ?
> It seems that above phenomenon has not a lot to do with reading intervals.
> > The change has been tested on RD-N2 model (Neoverse N2 ref platform),
> > it has also been discussed here [1]
> I doesn't get the testing result on this platform in its thread.
It might be missing exact numbers but the conclusions should be here [1]

> > > > > *--> Step 2: *read 'cpuinfo_cur_freq' in the high memory access pressure.
> > > > >     0: 2696628     2: 2696628     4: 2696628     6: 2696628
> > > > >     8: 2696628    10: 2696628    12: 2696628    14: 2696628
> > > > >    16: 2696628    18: 2696628    20: 2696628    22: 2696628
> > > > >    24: 2696628    26: 2696628    28: 2696628    30: 2696628
> > > > >    32: 2696628    34: 2696628    36: 2696628    38: 2696628
> > > > >    40: 2696628    42: 2696628    44: 2696628    46: 2696628
> > > > >    48: 2696628    50: 2696628    52: 2696628    54: 2696628
> > > > >    56: 2696628    58: 2696628    60: 2696628    62: 2696628
> > > > >    64: 2696628    66: 2696628    68: 2696628    70: 2696628
> > > > >    72: 2696628    74: 2696628    76: 2696628    78: 2696628
> > > > >    80: 2696628    82: 2696628    84: 2696628    86: 2696628
> > > > >    88: 2696628    90: 2696628    92: 2696628    94: 2696628
> > > > >    96: 2696628    98: 2696628   100: 2696628   102: 2696628
> > > > > 104: 2696628   106: 2696628   108: 2696628   110: 2696628
> > > > > 112: 2696628   114: 2696628   116: 2696628   118: 2696628
> > > > > 120: 2696628   122: 2696628   124: 2696628   126: 2696628
> > > > > 
> > > > > *Case 2: setting nohz_full and cpufreq use ondemand governor*
> > > > > There is "isolcpus=1-10,41-50 nohz_full=1-10,41-50 rcu_nocbs=1-10,41-50" in
> > > > > /proc/cmdline.
> > > > Right, so if I remember correctly nohz_full implies rcu_nocbs, so no need to
> > > > set that one.
> > > > Now, afair, isolcpus will make the selected CPUs to disappear from the
> > > > schedulers view (no balancing, no migrating), so unless you affine smth
> > > > explicitly to those CPUs, you will not see much of an activity there.
> > > Correct.
> > > > Need to double check though as it has been a while ...
> > > > > *--> Step 1: *setting ondemand governor to all policy and query
> > > > > 'cpuinfo_cur_freq' in no pressure case.
> > > > > And the frequency of CPUs all are about 400MHz.
> > > > > *--> Step 2:* read 'cpuinfo_cur_freq' in the high memory access pressure.
> > > > > The high memory access pressure is from the command: "stress-ng -c 64
> > > > > --cpu-load 100% --taskset 0-63"
> > > > I'm not entirely convinced that this will affine to isolated cpus, especially
> > > > that the affinity mask spans all available cpus. If that is the case, no wonder
> > > > your isolated cpus are getting wasted being idle. But I would have to double
> > > > check how this is being handled.
> > > > > The result:
> > > > >    0: 2696628     1:  400000     2:  400000     3:  400909
> > > > >    4:  400000     5:  400000     6:  400000     7:  400000
> > > > >    8:  400000     9:  400000    10:  400600    11: 2696628
> > > > > 12: 2696628    13: 2696628    14: 2696628    15: 2696628
> > > > > 16: 2696628    17: 2696628    18: 2696628    19: 2696628
> > > > > 20: 2696628    21: 2696628    22: 2696628    23: 2696628
> > > > > 24: 2696628    25: 2696628    26: 2696628    27: 2696628
> > > > > 28: 2696628    29: 2696628    30: 2696628    31: 2696628
> > > > > 32: 2696628    33: 2696628    34: 2696628    35: 2696628
> > > > > 36: 2696628    37: 2696628    38: 2696628    39: 2696628
> > > > > 40: 2696628    41:  400000    42:  400000    43:  400000
> > > > > 44:  400000    45:  398847    46:  400000    47:  400000
> > > > > 48:  400000    49:  400000    50:  400000    51: 2696628
> > > > > 52: 2696628    53: 2696628    54: 2696628    55: 2696628
> > > > > 56: 2696628    57: 2696628    58: 2696628    59: 2696628
> > > > > 60: 2696628    61: 2696628    62: 2696628    63: 2699264
> > > > > 
> > > > > Note:
> > > > > (1) The frequency of 1-10 and 41-50 CPUs work on the lowest frequency.
> > > > >        It turned out that nohz full was already work.
> > > > >        I guess that stress-ng cannot use the CPU in the range of nohz full.
> > > > >        Because the CPU frequency will be increased to 2.7G by binding CPU to
> > > > > other application.
> > > > > (2) The frequency of the nohz full core is calculated by get() callback
> > > > > according to ftrace.
> > > > It is as there is no sched tick on those, and apparently there is nothing
> > > > running on them either.
> > > Yes.
> > > If we select your approach and the above phenomenon is normal,
> > > the large frequency discrepancy issue can be resolved for CPUs with sched
> > > tick by the way.
> > > But the nohz full cores still have to face this issue. So this patch is also
> > > needed.
> > > 
> > Yes, nohz cores full have to be handled by the cpufreq driver.
> Correct. So we still have to face the issue in this patch and push this
> patch.
> Beata, would you please review this patch?
Just to clarify for my benefit (apologies but I do have to contex switch
pretty often these days): by reviewing this patch do you mean:
1) review your changes (if so I think there are few comments already to be
addressed, but I can try to have another look)
2) review changes for AMU-based arch_freq_get_on_cpu ?

*note: I will still try to have a look at the non-housekeeping cpus case

---
[1] https://lore.kernel.org/lkml/691d3eb2-cd93-f0fc-a7a4-2a8c0d44262c@nvidia.com/
---

BR
Beata
> 
> 
> /Huisong
> > 
[...]
Ionela Voinescu Feb. 13, 2024, 11:02 p.m. UTC | #20
Hi,

On Friday 09 Feb 2024 at 11:55:08 (+0100), Beata Michalska wrote:
> On Tue, Feb 06, 2024 at 04:02:15PM +0800, lihuisong (C) wrote:
[..]
> > > > > > 
> > > > > > > Would you be able to test [1] on your platform and usecase?
> > > > > > I has tested it on my platform (CPU number: 64, SMT: off and CPU base
> > > > > > frequency: 2.7GHz).
> > > > > > Accoding to the testing result,
> > > > > > 1> I found that patch [1] and [2] cannot cover the no housekeeping CPUs.
> > > > > > They still have to face the large frequency discrepancy issue my patch
> > > > > > mentioned.
> > > > > > 2> Additionally, the frequency value of all CPUs are almost the same by
> > > > > > using the 'arch_freq_scale' factor way. I'm not sure if it is ok.
> > > > > > 
> > > > > > The patch [1] has been modified silightly as below:
> > > > > > -->
> > > > > > @@ -1756,7 +1756,10 @@ static unsigned int
> > > > > > cpufreq_verify_current_freq(struct cpufreq_policy *policy, b
> > > > > >    {
> > > > > >           unsigned int new_freq;
> > > > > > 
> > > > > > -       new_freq = cpufreq_driver->get(policy->cpu);
> > > > > > +       new_freq = arch_freq_get_on_cpu(policy->cpu);
> > > > > > +       if (!new_freq)
> > > > > > +               new_freq = cpufreq_driver->get(policy->cpu);
> > > > > > +
> > > > > As pointed out this change will not make it to the next version of the patch.
> > > > > So I'd say you can safely ignore it and assume that arch_freq_get_on_cpu will
> > > > > only be wired for sysfs nodes for scaling_cur_freq/cpuinfo_cur_freq
> > > > > >           if (!new_freq)
> > > > > >                   return 0;
> > > > > > 
> > > > > > And the result is as follows:
> > > > > > *case 1:**No setting the nohz_full and cpufreq use performance governor*
> > > > > > *--> Step1: *read 'cpuinfo_cur_freq' in no pressure
> > > > > >     0: 2699264     2: 2699264     4: 2699264     6: 2699264
> > > > > >     8: 2696628    10: 2696628    12: 2696628    14: 2699264
> > > > > >    16: 2699264    18: 2696628    20: 2699264    22: 2696628
> > > > > >    24: 2699264    26: 2696628    28: 2699264    30: 2696628
> > > > > >    32: 2696628    34: 2696628    36: 2696628    38: 2696628
> > > > > >    40: 2699264    42: 2699264    44: 2696628    46: 2696628
> > > > > >    48: 2696628    50: 2699264    52: 2699264    54: 2696628
> > > > > >    56: 2696628    58: 2696628    60: 2696628    62: 2696628
> > > > > >    64: 2696628    66: 2699264    68: 2696628    70: 2696628
> > > > > >    72: 2699264    74: 2696628    76: 2696628    78: 2699264
> > > > > >    80: 2696628    82: 2696628    84: 2699264    86: 2696628
> > > > > >    88: 2696628    90: 2696628    92: 2696628    94: 2699264
> > > > > >    96: 2696628    98: 2699264   100: 2699264   102: 2696628
> > > > > > 104: 2699264   106: 2699264   108: 2699264   110: 2696628
> > > > > > 112: 2699264   114: 2699264   116: 2699264   118: 2699264
> > > > > > 120: 2696628   122: 2699264   124: 2696628   126: 2699264
> > > > > > Note: the frequency of all CPUs are almost the same.
> > > > > Were you expecting smth else ?
> > > > The frequency of each CPU might have a different value.
> > > > All value of all CPUs is the same under high pressure.
> > > > I don't know what the phenomenon is on other platform.
> > > > Do you know who else tested it?
> > > So I might have rushed a bit with my previous comment/question: apologies for
> > > that.
> > > The numbers above: those are on a fairly idle/lightly loaded system right?
> > Yes.
> > > Would you mind having another go with just the arch_freq_get_on_cpu
> > > implementation beign added and dropping the changes in the cpufreq and
> > All my tests are done when cpufreq policy is "performance" and OS isn't on a
> > high load.
> > Reading "scaling_cur_freq" or "scaling_cur_freq" for each physical core on
> > platform
> > 
> > The testing result for "cpuinfo_cur_freq" with your changes on a fairly idle
> > and high loaded system can also be found in this thread.
> > *A: the result with your changes*
> > --> Reading "scaling_cur_freq"
> >   0: 2688720     2: 2696628     4: 2699264     6: 2696628
> >   8: 2699264    10: 2696628    12: 2699264    14: 2699264
> >  16: 2699264    18: 2696628    20: 2696628    22: 2696628
> >  24: 2699264    26: 2696628    28: 2696628    30: 2696628
> >  32: 2699264    34: 2691356    36: 2696628    38: 2699264
> >  40: 2699264    42: 2696628    44: 2696628    46: 2699264
> >  48: 2699264    50: 2696628    52: 2696628    54: 2696628
> >  56: 2696628    58: 2699264    60: 2691356    62: 2696628
> >  64: 2696628    66: 2696628    68: 2696628    70: 2696628
> >  72: 2696628    74: 2696628    76: 2699264    78: 2696628
> >  80: 2696628    82: 2696628    84: 2699264    86: 2696628
> >  88: 2625456    90: 2696628    92: 2699264    94: 2696628
> >  96: 2696628    98: 2696628   100: 2699264   102: 2699264
> > 104: 2699264   106: 2696628   108: 2699264   110: 2696628
> > 112: 2699264   114: 2699264   116: 2696628   118: 2696628
> > 120: 2696628   122: 2699264   124: 2696628   126: 2696628
> > -->Reading  "cpuinfo_cur_freq"
> >   0: 2696628     2: 2696628     4: 2699264     6: 2688720
> >   8: 2699264    10: 2700000    12: 2696628    14: 2698322
> >  16: 2699264    18: 2699264    20: 2696628    22: 2699264
> >  24: 2699264    26: 2699264    28: 2699264    30: 2699264
> >  32: 2699264    34: 2693992    36: 2696628    38: 2696628
> >  40: 2699264    42: 2699264    44: 2699264    46: 2696628
> >  48: 2696628    50: 2699264    52: 2696628    54: 2696628
> >  56: 2699264    58: 2699264    60: 2696628    62: 2699264
> >  64: 2696628    66: 2699264    68: 2696628    70: 2699264
> >  72: 2696628    74: 2696628    76: 2696628    78: 2693992
> >  80: 2696628    82: 2696628    84: 2696628    86: 2696628
> >  88: 2696628    90: 2699264    92: 2696628    94: 2699264
> >  96: 2699264    98: 2696628   100: 2699264   102: 2699264
> > 104: 2691356   106: 2699264   108: 2699264   110: 2699264
> > 112: 2699264   114: 2696628   116: 2699264   118: 2699264
> > 120: 2696628   122: 2696628   124: 2696628   126: 2696628
> > 
> > *B: the result without your changes*
> > -->Reading "scaling_cur_freq"
> >   0: 2698245     2: 2706690     4: 2699649     6: 2702105
> >   8: 2704362    10: 2697993    12: 2701672    14: 2704362
> >  16: 2701052    18: 2701052    20: 2694385    22: 2699650
> >  24: 2706802    26: 2702389    28: 2698299    30: 2698299
> >  32: 2697333    34: 2697993    36: 2701337    38: 2699328
> >  40: 2700330    42: 2700330    44: 2698019    46: 2697697
> >  48: 2699659    50: 2701700    52: 2703401    54: 2701700
> >  56: 2704013    58: 2697658    60: 2695000    62: 2697666
> >  64: 2697902    66: 2701052    68: 2698245    70: 2695789
> >  72: 2701315    74: 2696655    76: 2693666    78: 2695317
> >  80: 2704912    82: 2699649    84: 2698245    86: 2695454
> >  88: 2697966    90: 2697959    92: 2699319    94: 2700680
> >  96: 2695317    98: 2698996   100: 2700000   102: 2700334
> > 104: 2701320   106: 2695065   108: 2700986   110: 2703960
> > 112: 2697635   114: 2704421   116: 2700680   118: 2702040
> > 120: 2700334   122: 2697993   124: 2700334   126: 2705351
> > -->Reading "cpuinfo_cur_freq"
> >   0: 2696853     2: 2695454     4: 2699649     6: 2706993
> >   8: 2706060    10: 2704362    12: 2704362    14: 2697658
> >  16: 2707719    18: 2697192    20: 2702456    22: 2699650
> >  24: 2705782    26: 2698299    28: 2703061    30: 2705802
> >  32: 2700000    34: 2700671    36: 2701337    38: 2697658
> >  40: 2700330    42: 2700330    44: 2699672    46: 2697697
> >  48: 2703061    50: 2696610    52: 2692542    54: 2704406
> >  56: 2695317    58: 2699331    60: 2698996    62: 2702675
> >  64: 2704912    66: 2703859    68: 2699649    70: 2698596
> >  72: 2703908    74: 2703355    76: 2697658    78: 2695317
> >  80: 2702105    82: 2707719    84: 2702105    86: 2699649
> >  88: 2697966    90: 2691525    92: 2701700    94: 2700680
> >  96: 2695317    98: 2698996   100: 2698666   102: 2700334
> > 104: 2690429   106: 2707590   108: 2700986   110: 2701320
> > 112: 2696283   114: 2692881   116: 2697627   118: 2704421
> > 120: 2698996   122: 2696321   124: 2696655   126: 2695000
> >
> So in both cases : whether you use arch_freq_get_on_cpu or not
> (so with and without the patch) you get roughly the same frequencies
> on all cores - or am I missing smth from the dump above ?
> And those are reflecting max freq you have provided earlier (?)
> Note that the arch_freq_get_on_cpu will return an average frequency for
> the last tick, so even if your system is roughly idle with your performance
> governor those numbers make sense (some/most of the cores might be idle
> but you will see the last freq the core was running at before going to idle).
> I do not think there is an agreement what should be shown for idle core when
> querying their freq through sysfs. Showing last known freq makes sense, even
> more than waking up core just to try to get one.
> 
> @Ionela: Please jump in if I got things wrong.

Yes, that's how I see things as well. When using the performance
governor, when the CPU is active, the frequency of the CPU should be the
maximum one (unless there has been firmware/hardware capping) and that
would be reflected by cpuinfo_cur_freq, either through the use of the
frequency scale factor (based on the samples on the last tick)  or the
driver's .get() function (having woken up the CPU to sample the
counters).

So the values above look alright to me.

Thanks,
Ionela.

> 
> > > then read 'scaling_cur_freq', doing several reads in some intervals ?
> > It seems that above phenomenon has not a lot to do with reading intervals.
> > > The change has been tested on RD-N2 model (Neoverse N2 ref platform),
> > > it has also been discussed here [1]
> > I doesn't get the testing result on this platform in its thread.
> It might be missing exact numbers but the conclusions should be here [1]
> 
> > > > > > *--> Step 2: *read 'cpuinfo_cur_freq' in the high memory access pressure.
> > > > > >     0: 2696628     2: 2696628     4: 2696628     6: 2696628
> > > > > >     8: 2696628    10: 2696628    12: 2696628    14: 2696628
> > > > > >    16: 2696628    18: 2696628    20: 2696628    22: 2696628
> > > > > >    24: 2696628    26: 2696628    28: 2696628    30: 2696628
> > > > > >    32: 2696628    34: 2696628    36: 2696628    38: 2696628
> > > > > >    40: 2696628    42: 2696628    44: 2696628    46: 2696628
> > > > > >    48: 2696628    50: 2696628    52: 2696628    54: 2696628
> > > > > >    56: 2696628    58: 2696628    60: 2696628    62: 2696628
> > > > > >    64: 2696628    66: 2696628    68: 2696628    70: 2696628
> > > > > >    72: 2696628    74: 2696628    76: 2696628    78: 2696628
> > > > > >    80: 2696628    82: 2696628    84: 2696628    86: 2696628
> > > > > >    88: 2696628    90: 2696628    92: 2696628    94: 2696628
> > > > > >    96: 2696628    98: 2696628   100: 2696628   102: 2696628
> > > > > > 104: 2696628   106: 2696628   108: 2696628   110: 2696628
> > > > > > 112: 2696628   114: 2696628   116: 2696628   118: 2696628
> > > > > > 120: 2696628   122: 2696628   124: 2696628   126: 2696628
> > > > > > 
> > > > > > *Case 2: setting nohz_full and cpufreq use ondemand governor*
> > > > > > There is "isolcpus=1-10,41-50 nohz_full=1-10,41-50 rcu_nocbs=1-10,41-50" in
> > > > > > /proc/cmdline.
> > > > > Right, so if I remember correctly nohz_full implies rcu_nocbs, so no need to
> > > > > set that one.
> > > > > Now, afair, isolcpus will make the selected CPUs to disappear from the
> > > > > schedulers view (no balancing, no migrating), so unless you affine smth
> > > > > explicitly to those CPUs, you will not see much of an activity there.
> > > > Correct.
> > > > > Need to double check though as it has been a while ...
> > > > > > *--> Step 1: *setting ondemand governor to all policy and query
> > > > > > 'cpuinfo_cur_freq' in no pressure case.
> > > > > > And the frequency of CPUs all are about 400MHz.
> > > > > > *--> Step 2:* read 'cpuinfo_cur_freq' in the high memory access pressure.
> > > > > > The high memory access pressure is from the command: "stress-ng -c 64
> > > > > > --cpu-load 100% --taskset 0-63"
> > > > > I'm not entirely convinced that this will affine to isolated cpus, especially
> > > > > that the affinity mask spans all available cpus. If that is the case, no wonder
> > > > > your isolated cpus are getting wasted being idle. But I would have to double
> > > > > check how this is being handled.
> > > > > > The result:
> > > > > >    0: 2696628     1:  400000     2:  400000     3:  400909
> > > > > >    4:  400000     5:  400000     6:  400000     7:  400000
> > > > > >    8:  400000     9:  400000    10:  400600    11: 2696628
> > > > > > 12: 2696628    13: 2696628    14: 2696628    15: 2696628
> > > > > > 16: 2696628    17: 2696628    18: 2696628    19: 2696628
> > > > > > 20: 2696628    21: 2696628    22: 2696628    23: 2696628
> > > > > > 24: 2696628    25: 2696628    26: 2696628    27: 2696628
> > > > > > 28: 2696628    29: 2696628    30: 2696628    31: 2696628
> > > > > > 32: 2696628    33: 2696628    34: 2696628    35: 2696628
> > > > > > 36: 2696628    37: 2696628    38: 2696628    39: 2696628
> > > > > > 40: 2696628    41:  400000    42:  400000    43:  400000
> > > > > > 44:  400000    45:  398847    46:  400000    47:  400000
> > > > > > 48:  400000    49:  400000    50:  400000    51: 2696628
> > > > > > 52: 2696628    53: 2696628    54: 2696628    55: 2696628
> > > > > > 56: 2696628    57: 2696628    58: 2696628    59: 2696628
> > > > > > 60: 2696628    61: 2696628    62: 2696628    63: 2699264
> > > > > > 
> > > > > > Note:
> > > > > > (1) The frequency of 1-10 and 41-50 CPUs work on the lowest frequency.
> > > > > >        It turned out that nohz full was already work.
> > > > > >        I guess that stress-ng cannot use the CPU in the range of nohz full.
> > > > > >        Because the CPU frequency will be increased to 2.7G by binding CPU to
> > > > > > other application.
> > > > > > (2) The frequency of the nohz full core is calculated by get() callback
> > > > > > according to ftrace.
> > > > > It is as there is no sched tick on those, and apparently there is nothing
> > > > > running on them either.
> > > > Yes.
> > > > If we select your approach and the above phenomenon is normal,
> > > > the large frequency discrepancy issue can be resolved for CPUs with sched
> > > > tick by the way.
> > > > But the nohz full cores still have to face this issue. So this patch is also
> > > > needed.
> > > > 
> > > Yes, nohz cores full have to be handled by the cpufreq driver.
> > Correct. So we still have to face the issue in this patch and push this
> > patch.
> > Beata, would you please review this patch?
> Just to clarify for my benefit (apologies but I do have to contex switch
> pretty often these days): by reviewing this patch do you mean:
> 1) review your changes (if so I think there are few comments already to be
> addressed, but I can try to have another look)
> 2) review changes for AMU-based arch_freq_get_on_cpu ?
> 
> *note: I will still try to have a look at the non-housekeeping cpus case
> 
> ---
> [1] https://lore.kernel.org/lkml/691d3eb2-cd93-f0fc-a7a4-2a8c0d44262c@nvidia.com/
> ---
> 
> BR
> Beata
> > 
> > 
> > /Huisong
> > > 
> [...]
lihuisong (C) Feb. 19, 2024, 12:15 p.m. UTC | #21
在 2024/2/9 18:55, Beata Michalska 写道:
> On Tue, Feb 06, 2024 at 04:02:15PM +0800, lihuisong (C) wrote:
>> 在 2024/2/2 16:08, Beata Michalska 写道:
>>> On Wed, Jan 17, 2024 at 05:18:40PM +0800, lihuisong (C) wrote:
>>>
>>> Hi ,
>>>
>>> Again, apologies for delay,
>>>
>>>> Hi,
>>>>
>>>> 在 2024/1/16 22:10, Beata Michalska 写道:
>>>>> Hi,
>>>>>
>>>>> Apologies for jumping in so late....
>>>>>
>>>>> On Wed, Jan 10, 2024 at 03:09:48PM +0800, lihuisong (C) wrote:
>>>>>> Hi Ionela,
>>>>>>
>>>>>> 在 2024/1/8 22:03, Ionela Voinescu 写道:
>>>>>>> Hi,
>>>>>>>
>>>>>>> On Friday 05 Jan 2024 at 15:04:47 (+0800), lihuisong (C) wrote:
>>>>>>>> Hi Vanshi,
>>>>>>>>
>>>>>>>> 在 2024/1/5 8:48, Vanshidhar Konda 写道:
>>>>>>>>> On Thu, Jan 04, 2024 at 05:36:51PM +0800, lihuisong (C) wrote:
>>>>>>>>>> 在 2024/1/4 1:53, Ionela Voinescu 写道:
>>>>>>>>>>> Hi,
>>>>>>>>>>>
>>>>>>>>>>> On Tuesday 12 Dec 2023 at 15:26:17 (+0800), Huisong Li wrote:
>>>>>>>>>>>> Many developers found that the cpu current frequency is greater than
>>>>>>>>>>>> the maximum frequency of the platform, please see [1], [2] and [3].
>>>>>>>>>>>>
>>>>>>>>>>>> In the scenarios with high memory access pressure, the patch [1] has
>>>>>>>>>>>> proved the significant latency of cpc_read() which is used to obtain
>>>>>>>>>>>> delivered and reference performance counter cause an absurd frequency.
>>>>>>>>>>>> The sampling interval for this counters is very critical and
>>>>>>>>>>>> is expected
>>>>>>>>>>>> to be equal. However, the different latency of cpc_read() has a direct
>>>>>>>>>>>> impact on their sampling interval.
>>>>>>>>>>>>
>>>>>>>>>>> Would this [1] alternative solution work for you?
>>>>>>>>>> It would work for me AFAICS.
>>>>>>>>>> Because the "arch_freq_scale" is also from AMU core and constant
>>>>>>>>>> counter, and read together.
>>>>>>>>>> But, from their discuss line, it seems that there are some tricky
>>>>>>>>>> points to clarify or consider.
>>>>>>>>> I think the changes in [1] would work better when CPUs may be idle. With
>>>>>>>>> this
>>>>>>>>> patch we would have to wake any core that is in idle state to read the
>>>>>>>>> AMU
>>>>>>>>> counters. Worst case, if core 0 is trying to read the CPU frequency of
>>>>>>>>> all
>>>>>>>>> cores, it may need to wake up all the other cores to read the AMU
>>>>>>>>> counters.
>>>>>>>>     From the approach in [1], if all CPUs (one or more cores) under one policy
>>>>>>>> are idle, they still cannot be obtained the CPU frequency, right?
>>>>>>>> In this case, the [1] API will return 0 and have to back to call
>>>>>>>> cpufreq_driver->get() for cpuinfo_cur_freq.
>>>>>>>> Then we still need to face the issue this patch mentioned.
>>>>>>> With the implementation at [1], arch_freq_get_on_cpu() will not return 0
>>>>>>> for idle CPUs and the get() callback will not be called to wake up the
>>>>>>> CPUs.
>>>>>> Right, arch_freq_get_on_cpu() will not return 0 for idle CPUs.
>>>>>> However, for no-housekeeping CPUs, it will return 0 and have to call get()
>>>>>> callback, right?
>>>>>>> Worst case, arch_freq_get_on_cpu() will return a frequency based on the
>>>>>>> AMU counter values obtained on the last tick on that CPU. But if that CPU
>>>>>>> is not a housekeeping CPU, a housekeeping CPU in the same policy will be
>>>>>>> selected, as it would have had a more recent tick, and therefore a more
>>>>>>> recent frequency value for the domain.
>>>>>> But this frequency is from the last tick,
>>>>>> this last tick is probably a long time ago and it doesn't update
>>>>>> 'arch_freq_scale' for some reasons like CPU dile.
>>>>>> In addition, I'm not sure if there is possible that amu_scale_freq_tick() is
>>>>>> executed delayed under high stress case.
>>>>>> It also have an impact on the accuracy of the cpu frequency we query.
>>>>>>> I understand that the frequency returned here will not be up to date,
>>>>>>> but there's no proper frequency feedback for an idle CPU. If one only
>>>>>>> wakes up a CPU to sample counters, before the CPU goes back to sleep,
>>>>>>> the obtained frequency feedback is meaningless.
>>>>>>>
>>>>>>>>> For systems with 128 cores or more, this could be very expensive and
>>>>>>>>> happen
>>>>>>>>> very frequently.
>>>>>>>>>
>>>>>>>>> AFAICS, the approach in [1] would avoid this cost.
>>>>>>>> But the CPU frequency is just an average value for the last tick period
>>>>>>>> instead of the current one the CPU actually runs at.
>>>>>>>> In addition, there are some conditions to use 'arch_freq_scale' in this
>>>>>>>> approach.
>>>>>>> What are the conditions you are referring to?
>>>>>> It depends on the housekeeping CPUs.
>>>>>>>> So I'm not sure if this approach can entirely cover the frequency
>>>>>>>> discrepancy issue.
>>>>>>> Unfortunately there is no perfect frequency feedback. By the time you
>>>>>>> observe/use the value of scaling_cur_freq/cpuinfo_cur_freq, the frequency
>>>>>>> of the CPU might have already changed. Therefore, an average value might
>>>>>>> be a better indication of the recent performance level of a CPU.
>>>>>> An average value for CPU frequency is ok. It may be better if it has not any
>>>>>> delaying.
>>>>>>
>>>>>> The original implementation for cpuinfo_cur_freq can more reflect their
>>>>>> meaning in the user-guide [1]. The user-guide said:
>>>>>> "cpuinfo_cur_freq : Current frequency of the CPU as obtained from the
>>>>>> hardware, in KHz.
>>>>>> This is the frequency the CPU actually runs at."
>>>>>>
>>>>>>
>>>>>> [1]https://www.kernel.org/doc/Documentation/cpu-freq/user-guide.txt
>>>>>>
>>>>>>> Would you be able to test [1] on your platform and usecase?
>>>>>> I has tested it on my platform (CPU number: 64, SMT: off and CPU base
>>>>>> frequency: 2.7GHz).
>>>>>> Accoding to the testing result,
>>>>>> 1> I found that patch [1] and [2] cannot cover the no housekeeping CPUs.
>>>>>> They still have to face the large frequency discrepancy issue my patch
>>>>>> mentioned.
>>>>>> 2> Additionally, the frequency value of all CPUs are almost the same by
>>>>>> using the 'arch_freq_scale' factor way. I'm not sure if it is ok.
>>>>>>
>>>>>> The patch [1] has been modified silightly as below:
>>>>>> -->
>>>>>> @@ -1756,7 +1756,10 @@ static unsigned int
>>>>>> cpufreq_verify_current_freq(struct cpufreq_policy *policy, b
>>>>>>     {
>>>>>>            unsigned int new_freq;
>>>>>>
>>>>>> -       new_freq = cpufreq_driver->get(policy->cpu);
>>>>>> +       new_freq = arch_freq_get_on_cpu(policy->cpu);
>>>>>> +       if (!new_freq)
>>>>>> +               new_freq = cpufreq_driver->get(policy->cpu);
>>>>>> +
>>>>> As pointed out this change will not make it to the next version of the patch.
>>>>> So I'd say you can safely ignore it and assume that arch_freq_get_on_cpu will
>>>>> only be wired for sysfs nodes for scaling_cur_freq/cpuinfo_cur_freq
>>>>>>            if (!new_freq)
>>>>>>                    return 0;
>>>>>>
>>>>>> And the result is as follows:
>>>>>> *case 1:**No setting the nohz_full and cpufreq use performance governor*
>>>>>> *--> Step1: *read 'cpuinfo_cur_freq' in no pressure
>>>>>>      0: 2699264     2: 2699264     4: 2699264     6: 2699264
>>>>>>      8: 2696628    10: 2696628    12: 2696628    14: 2699264
>>>>>>     16: 2699264    18: 2696628    20: 2699264    22: 2696628
>>>>>>     24: 2699264    26: 2696628    28: 2699264    30: 2696628
>>>>>>     32: 2696628    34: 2696628    36: 2696628    38: 2696628
>>>>>>     40: 2699264    42: 2699264    44: 2696628    46: 2696628
>>>>>>     48: 2696628    50: 2699264    52: 2699264    54: 2696628
>>>>>>     56: 2696628    58: 2696628    60: 2696628    62: 2696628
>>>>>>     64: 2696628    66: 2699264    68: 2696628    70: 2696628
>>>>>>     72: 2699264    74: 2696628    76: 2696628    78: 2699264
>>>>>>     80: 2696628    82: 2696628    84: 2699264    86: 2696628
>>>>>>     88: 2696628    90: 2696628    92: 2696628    94: 2699264
>>>>>>     96: 2696628    98: 2699264   100: 2699264   102: 2696628
>>>>>> 104: 2699264   106: 2699264   108: 2699264   110: 2696628
>>>>>> 112: 2699264   114: 2699264   116: 2699264   118: 2699264
>>>>>> 120: 2696628   122: 2699264   124: 2696628   126: 2699264
>>>>>> Note: the frequency of all CPUs are almost the same.
>>>>> Were you expecting smth else ?
>>>> The frequency of each CPU might have a different value.
>>>> All value of all CPUs is the same under high pressure.
>>>> I don't know what the phenomenon is on other platform.
>>>> Do you know who else tested it?
>>> So I might have rushed a bit with my previous comment/question: apologies for
>>> that.
>>> The numbers above: those are on a fairly idle/lightly loaded system right?
>> Yes.
>>> Would you mind having another go with just the arch_freq_get_on_cpu
>>> implementation beign added and dropping the changes in the cpufreq and
>> All my tests are done when cpufreq policy is "performance" and OS isn't on a
>> high load.
>> Reading "scaling_cur_freq" or "scaling_cur_freq" for each physical core on
>> platform
>>
>> The testing result for "cpuinfo_cur_freq" with your changes on a fairly idle
>> and high loaded system can also be found in this thread.
>> *A: the result with your changes*
>> --> Reading "scaling_cur_freq"
>>    0: 2688720     2: 2696628     4: 2699264     6: 2696628
>>    8: 2699264    10: 2696628    12: 2699264    14: 2699264
>>   16: 2699264    18: 2696628    20: 2696628    22: 2696628
>>   24: 2699264    26: 2696628    28: 2696628    30: 2696628
>>   32: 2699264    34: 2691356    36: 2696628    38: 2699264
>>   40: 2699264    42: 2696628    44: 2696628    46: 2699264
>>   48: 2699264    50: 2696628    52: 2696628    54: 2696628
>>   56: 2696628    58: 2699264    60: 2691356    62: 2696628
>>   64: 2696628    66: 2696628    68: 2696628    70: 2696628
>>   72: 2696628    74: 2696628    76: 2699264    78: 2696628
>>   80: 2696628    82: 2696628    84: 2699264    86: 2696628
>>   88: 2625456    90: 2696628    92: 2699264    94: 2696628
>>   96: 2696628    98: 2696628   100: 2699264   102: 2699264
>> 104: 2699264   106: 2696628   108: 2699264   110: 2696628
>> 112: 2699264   114: 2699264   116: 2696628   118: 2696628
>> 120: 2696628   122: 2699264   124: 2696628   126: 2696628
>> -->Reading  "cpuinfo_cur_freq"
>>    0: 2696628     2: 2696628     4: 2699264     6: 2688720
>>    8: 2699264    10: 2700000    12: 2696628    14: 2698322
>>   16: 2699264    18: 2699264    20: 2696628    22: 2699264
>>   24: 2699264    26: 2699264    28: 2699264    30: 2699264
>>   32: 2699264    34: 2693992    36: 2696628    38: 2696628
>>   40: 2699264    42: 2699264    44: 2699264    46: 2696628
>>   48: 2696628    50: 2699264    52: 2696628    54: 2696628
>>   56: 2699264    58: 2699264    60: 2696628    62: 2699264
>>   64: 2696628    66: 2699264    68: 2696628    70: 2699264
>>   72: 2696628    74: 2696628    76: 2696628    78: 2693992
>>   80: 2696628    82: 2696628    84: 2696628    86: 2696628
>>   88: 2696628    90: 2699264    92: 2696628    94: 2699264
>>   96: 2699264    98: 2696628   100: 2699264   102: 2699264
>> 104: 2691356   106: 2699264   108: 2699264   110: 2699264
>> 112: 2699264   114: 2696628   116: 2699264   118: 2699264
>> 120: 2696628   122: 2696628   124: 2696628   126: 2696628
>>
>> *B: the result without your changes*
>> -->Reading "scaling_cur_freq"
>>    0: 2698245     2: 2706690     4: 2699649     6: 2702105
>>    8: 2704362    10: 2697993    12: 2701672    14: 2704362
>>   16: 2701052    18: 2701052    20: 2694385    22: 2699650
>>   24: 2706802    26: 2702389    28: 2698299    30: 2698299
>>   32: 2697333    34: 2697993    36: 2701337    38: 2699328
>>   40: 2700330    42: 2700330    44: 2698019    46: 2697697
>>   48: 2699659    50: 2701700    52: 2703401    54: 2701700
>>   56: 2704013    58: 2697658    60: 2695000    62: 2697666
>>   64: 2697902    66: 2701052    68: 2698245    70: 2695789
>>   72: 2701315    74: 2696655    76: 2693666    78: 2695317
>>   80: 2704912    82: 2699649    84: 2698245    86: 2695454
>>   88: 2697966    90: 2697959    92: 2699319    94: 2700680
>>   96: 2695317    98: 2698996   100: 2700000   102: 2700334
>> 104: 2701320   106: 2695065   108: 2700986   110: 2703960
>> 112: 2697635   114: 2704421   116: 2700680   118: 2702040
>> 120: 2700334   122: 2697993   124: 2700334   126: 2705351
>> -->Reading "cpuinfo_cur_freq"
>>    0: 2696853     2: 2695454     4: 2699649     6: 2706993
>>    8: 2706060    10: 2704362    12: 2704362    14: 2697658
>>   16: 2707719    18: 2697192    20: 2702456    22: 2699650
>>   24: 2705782    26: 2698299    28: 2703061    30: 2705802
>>   32: 2700000    34: 2700671    36: 2701337    38: 2697658
>>   40: 2700330    42: 2700330    44: 2699672    46: 2697697
>>   48: 2703061    50: 2696610    52: 2692542    54: 2704406
>>   56: 2695317    58: 2699331    60: 2698996    62: 2702675
>>   64: 2704912    66: 2703859    68: 2699649    70: 2698596
>>   72: 2703908    74: 2703355    76: 2697658    78: 2695317
>>   80: 2702105    82: 2707719    84: 2702105    86: 2699649
>>   88: 2697966    90: 2691525    92: 2701700    94: 2700680
>>   96: 2695317    98: 2698996   100: 2698666   102: 2700334
>> 104: 2690429   106: 2707590   108: 2700986   110: 2701320
>> 112: 2696283   114: 2692881   116: 2697627   118: 2704421
>> 120: 2698996   122: 2696321   124: 2696655   126: 2695000
>>
> So in both cases : whether you use arch_freq_get_on_cpu or not
> (so with and without the patch) you get roughly the same frequencies
> on all cores - or am I missing smth from the dump above ?
The changes in "with/without your changes" I said is your patch 
intruduced arch_freq_get_on_cpu.
I just test them according to your requesting.
> And those are reflecting max freq you have provided earlier (?)
I know it is an average frequency for the last tickfor using 
arch_freq_get_on_cpu.
I have no any doubt that the freq is maximum value on performance governor.
I just want to say the difference between having or not having your patch.
The frequency values of all cores from cpuinfo_cur_freq and
scaling_cur_freq are almost the same if use this arch_freq_get_on_cpu on 
my platform.
However, the frequency values of all cores are different if doesn't use 
this arch_freq_get_on_cpu and just use .get().
> Note that the arch_freq_get_on_cpu will return an average frequency for
> the last tick, so even if your system is roughly idle with your performance
> governor those numbers make sense (some/most of the cores might be idle
> but you will see the last freq the core was running at before going to idle).
> I do not think there is an agreement what should be shown for idle core when
> querying their freq through sysfs. Showing last known freq makes sense, even
> more than waking up core just to try to get one.
I'm not opposed to using frequency scale factor to get CPU frequency. 
But it better be okay.
>
> @Ionela: Please jump in if I got things wrong.
>
>>> then read 'scaling_cur_freq', doing several reads in some intervals ?
>> It seems that above phenomenon has not a lot to do with reading intervals.
>>> The change has been tested on RD-N2 model (Neoverse N2 ref platform),
>>> it has also been discussed here [1]
>> I doesn't get the testing result on this platform in its thread.
> It might be missing exact numbers but the conclusions should be here [1]
>
>>>>>> *--> Step 2: *read 'cpuinfo_cur_freq' in the high memory access pressure.
>>>>>>      0: 2696628     2: 2696628     4: 2696628     6: 2696628
>>>>>>      8: 2696628    10: 2696628    12: 2696628    14: 2696628
>>>>>>     16: 2696628    18: 2696628    20: 2696628    22: 2696628
>>>>>>     24: 2696628    26: 2696628    28: 2696628    30: 2696628
>>>>>>     32: 2696628    34: 2696628    36: 2696628    38: 2696628
>>>>>>     40: 2696628    42: 2696628    44: 2696628    46: 2696628
>>>>>>     48: 2696628    50: 2696628    52: 2696628    54: 2696628
>>>>>>     56: 2696628    58: 2696628    60: 2696628    62: 2696628
>>>>>>     64: 2696628    66: 2696628    68: 2696628    70: 2696628
>>>>>>     72: 2696628    74: 2696628    76: 2696628    78: 2696628
>>>>>>     80: 2696628    82: 2696628    84: 2696628    86: 2696628
>>>>>>     88: 2696628    90: 2696628    92: 2696628    94: 2696628
>>>>>>     96: 2696628    98: 2696628   100: 2696628   102: 2696628
>>>>>> 104: 2696628   106: 2696628   108: 2696628   110: 2696628
>>>>>> 112: 2696628   114: 2696628   116: 2696628   118: 2696628
>>>>>> 120: 2696628   122: 2696628   124: 2696628   126: 2696628
>>>>>>
>>>>>> *Case 2: setting nohz_full and cpufreq use ondemand governor*
>>>>>> There is "isolcpus=1-10,41-50 nohz_full=1-10,41-50 rcu_nocbs=1-10,41-50" in
>>>>>> /proc/cmdline.
>>>>> Right, so if I remember correctly nohz_full implies rcu_nocbs, so no need to
>>>>> set that one.
>>>>> Now, afair, isolcpus will make the selected CPUs to disappear from the
>>>>> schedulers view (no balancing, no migrating), so unless you affine smth
>>>>> explicitly to those CPUs, you will not see much of an activity there.
>>>> Correct.
>>>>> Need to double check though as it has been a while ...
>>>>>> *--> Step 1: *setting ondemand governor to all policy and query
>>>>>> 'cpuinfo_cur_freq' in no pressure case.
>>>>>> And the frequency of CPUs all are about 400MHz.
>>>>>> *--> Step 2:* read 'cpuinfo_cur_freq' in the high memory access pressure.
>>>>>> The high memory access pressure is from the command: "stress-ng -c 64
>>>>>> --cpu-load 100% --taskset 0-63"
>>>>> I'm not entirely convinced that this will affine to isolated cpus, especially
>>>>> that the affinity mask spans all available cpus. If that is the case, no wonder
>>>>> your isolated cpus are getting wasted being idle. But I would have to double
>>>>> check how this is being handled.
>>>>>> The result:
>>>>>>     0: 2696628     1:  400000     2:  400000     3:  400909
>>>>>>     4:  400000     5:  400000     6:  400000     7:  400000
>>>>>>     8:  400000     9:  400000    10:  400600    11: 2696628
>>>>>> 12: 2696628    13: 2696628    14: 2696628    15: 2696628
>>>>>> 16: 2696628    17: 2696628    18: 2696628    19: 2696628
>>>>>> 20: 2696628    21: 2696628    22: 2696628    23: 2696628
>>>>>> 24: 2696628    25: 2696628    26: 2696628    27: 2696628
>>>>>> 28: 2696628    29: 2696628    30: 2696628    31: 2696628
>>>>>> 32: 2696628    33: 2696628    34: 2696628    35: 2696628
>>>>>> 36: 2696628    37: 2696628    38: 2696628    39: 2696628
>>>>>> 40: 2696628    41:  400000    42:  400000    43:  400000
>>>>>> 44:  400000    45:  398847    46:  400000    47:  400000
>>>>>> 48:  400000    49:  400000    50:  400000    51: 2696628
>>>>>> 52: 2696628    53: 2696628    54: 2696628    55: 2696628
>>>>>> 56: 2696628    57: 2696628    58: 2696628    59: 2696628
>>>>>> 60: 2696628    61: 2696628    62: 2696628    63: 2699264
>>>>>>
>>>>>> Note:
>>>>>> (1) The frequency of 1-10 and 41-50 CPUs work on the lowest frequency.
>>>>>>         It turned out that nohz full was already work.
>>>>>>         I guess that stress-ng cannot use the CPU in the range of nohz full.
>>>>>>         Because the CPU frequency will be increased to 2.7G by binding CPU to
>>>>>> other application.
>>>>>> (2) The frequency of the nohz full core is calculated by get() callback
>>>>>> according to ftrace.
>>>>> It is as there is no sched tick on those, and apparently there is nothing
>>>>> running on them either.
>>>> Yes.
>>>> If we select your approach and the above phenomenon is normal,
>>>> the large frequency discrepancy issue can be resolved for CPUs with sched
>>>> tick by the way.
>>>> But the nohz full cores still have to face this issue. So this patch is also
>>>> needed.
>>>>
>>> Yes, nohz cores full have to be handled by the cpufreq driver.
>> Correct. So we still have to face the issue in this patch and push this
>> patch.
>> Beata, would you please review this patch?
> Just to clarify for my benefit (apologies but I do have to contex switch
> pretty often these days): by reviewing this patch do you mean:
> 1) review your changes (if so I think there are few comments already to be
> addressed, but I can try to have another look)
Currently, the main comments is that my patch will wake up CPU to get 
frequency.
BTW, the core's always been wakened up to get the frequency for FFH way 
in cppc_acpi. please see cpc_read_ffh().
So it may be acceptable. After all, we don't query CPU frequency very often.
But your patch doesn't meet the non-housekeeping cpus.
> 2) review changes for AMU-based arch_freq_get_on_cpu ?
>
> *note: I will still try to have a look at the non-housekeeping cpus case
I am very much hope that this issue my patch mentioned can be resolved ASAP.
So what's your plan about non-housekeeping cpus case?
>
> ---
> [1] https://lore.kernel.org/lkml/691d3eb2-cd93-f0fc-a7a4-2a8c0d44262c@nvidia.com/
> ---
>
> BR
> Beata
>>
>> /Huisong
> [...]
> .
Vanshidhar Konda Feb. 20, 2024, 4:11 p.m. UTC | #22
On Mon, Feb 19, 2024 at 08:15:50PM +0800, lihuisong (C) wrote:
>
>在 2024/2/9 18:55, Beata Michalska 写道:
>>On Tue, Feb 06, 2024 at 04:02:15PM +0800, lihuisong (C) wrote:
>>>在 2024/2/2 16:08, Beata Michalska 写道:
>>>>On Wed, Jan 17, 2024 at 05:18:40PM +0800, lihuisong (C) wrote:
>>>>
>>>>Hi ,
>>>>
>>>>Again, apologies for delay,
>>>>
>>>>>Hi,
>>>>>
>>>>>在 2024/1/16 22:10, Beata Michalska 写道:
>>>>>>Hi,
>>>>>>
>>>>>>Apologies for jumping in so late....
>>>>>>
>>>>>>On Wed, Jan 10, 2024 at 03:09:48PM +0800, lihuisong (C) wrote:
>>>>>>>Hi Ionela,
>>>>>>>
>>>>>>>在 2024/1/8 22:03, Ionela Voinescu 写道:
>>>>>>>>Hi,
>>>>>>>>
>>>>>>>>On Friday 05 Jan 2024 at 15:04:47 (+0800), lihuisong (C) wrote:
>>>>>>>>>Hi Vanshi,
>>>>>>>>>
>>>>>>>>>在 2024/1/5 8:48, Vanshidhar Konda 写道:
>>>>>>>>>>On Thu, Jan 04, 2024 at 05:36:51PM +0800, lihuisong (C) wrote:
>>>>>>>>>>>在 2024/1/4 1:53, Ionela Voinescu 写道:
>>>>>>>>>>>>Hi,
>>>>>>>>>>>>
>>>>>>>>>>>>On Tuesday 12 Dec 2023 at 15:26:17 (+0800), Huisong Li wrote:
>>>>>>>>>>>>>Many developers found that the cpu current frequency is greater than
>>>>>>>>>>>>>the maximum frequency of the platform, please see [1], [2] and [3].
>>>>>>>>>>>>>
>>>>>>>>>>>>>In the scenarios with high memory access pressure, the patch [1] has
>>>>>>>>>>>>>proved the significant latency of cpc_read() which is used to obtain
>>>>>>>>>>>>>delivered and reference performance counter cause an absurd frequency.
>>>>>>>>>>>>>The sampling interval for this counters is very critical and
>>>>>>>>>>>>>is expected
>>>>>>>>>>>>>to be equal. However, the different latency of cpc_read() has a direct
>>>>>>>>>>>>>impact on their sampling interval.
>>>>>>>>>>>>>
>>>>>>>>>>>>Would this [1] alternative solution work for you?
>>>>>>>>>>>It would work for me AFAICS.
>>>>>>>>>>>Because the "arch_freq_scale" is also from AMU core and constant
>>>>>>>>>>>counter, and read together.
>>>>>>>>>>>But, from their discuss line, it seems that there are some tricky
>>>>>>>>>>>points to clarify or consider.
>>>>>>>>>>I think the changes in [1] would work better when CPUs may be idle. With
>>>>>>>>>>this
>>>>>>>>>>patch we would have to wake any core that is in idle state to read the
>>>>>>>>>>AMU
>>>>>>>>>>counters. Worst case, if core 0 is trying to read the CPU frequency of
>>>>>>>>>>all
>>>>>>>>>>cores, it may need to wake up all the other cores to read the AMU
>>>>>>>>>>counters.
>>>>>>>>>    From the approach in [1], if all CPUs (one or more cores) under one policy
>>>>>>>>>are idle, they still cannot be obtained the CPU frequency, right?
>>>>>>>>>In this case, the [1] API will return 0 and have to back to call
>>>>>>>>>cpufreq_driver->get() for cpuinfo_cur_freq.
>>>>>>>>>Then we still need to face the issue this patch mentioned.
>>>>>>>>With the implementation at [1], arch_freq_get_on_cpu() will not return 0
>>>>>>>>for idle CPUs and the get() callback will not be called to wake up the
>>>>>>>>CPUs.
>>>>>>>Right, arch_freq_get_on_cpu() will not return 0 for idle CPUs.
>>>>>>>However, for no-housekeeping CPUs, it will return 0 and have to call get()
>>>>>>>callback, right?
>>>>>>>>Worst case, arch_freq_get_on_cpu() will return a frequency based on the
>>>>>>>>AMU counter values obtained on the last tick on that CPU. But if that CPU
>>>>>>>>is not a housekeeping CPU, a housekeeping CPU in the same policy will be
>>>>>>>>selected, as it would have had a more recent tick, and therefore a more
>>>>>>>>recent frequency value for the domain.
>>>>>>>But this frequency is from the last tick,
>>>>>>>this last tick is probably a long time ago and it doesn't update
>>>>>>>'arch_freq_scale' for some reasons like CPU dile.
>>>>>>>In addition, I'm not sure if there is possible that amu_scale_freq_tick() is
>>>>>>>executed delayed under high stress case.
>>>>>>>It also have an impact on the accuracy of the cpu frequency we query.
>>>>>>>>I understand that the frequency returned here will not be up to date,
>>>>>>>>but there's no proper frequency feedback for an idle CPU. If one only
>>>>>>>>wakes up a CPU to sample counters, before the CPU goes back to sleep,
>>>>>>>>the obtained frequency feedback is meaningless.
>>>>>>>>
>>>>>>>>>>For systems with 128 cores or more, this could be very expensive and
>>>>>>>>>>happen
>>>>>>>>>>very frequently.
>>>>>>>>>>
>>>>>>>>>>AFAICS, the approach in [1] would avoid this cost.
>>>>>>>>>But the CPU frequency is just an average value for the last tick period
>>>>>>>>>instead of the current one the CPU actually runs at.
>>>>>>>>>In addition, there are some conditions to use 'arch_freq_scale' in this
>>>>>>>>>approach.
>>>>>>>>What are the conditions you are referring to?
>>>>>>>It depends on the housekeeping CPUs.
>>>>>>>>>So I'm not sure if this approach can entirely cover the frequency
>>>>>>>>>discrepancy issue.
>>>>>>>>Unfortunately there is no perfect frequency feedback. By the time you
>>>>>>>>observe/use the value of scaling_cur_freq/cpuinfo_cur_freq, the frequency
>>>>>>>>of the CPU might have already changed. Therefore, an average value might
>>>>>>>>be a better indication of the recent performance level of a CPU.
>>>>>>>An average value for CPU frequency is ok. It may be better if it has not any
>>>>>>>delaying.
>>>>>>>
>>>>>>>The original implementation for cpuinfo_cur_freq can more reflect their
>>>>>>>meaning in the user-guide [1]. The user-guide said:
>>>>>>>"cpuinfo_cur_freq : Current frequency of the CPU as obtained from the
>>>>>>>hardware, in KHz.
>>>>>>>This is the frequency the CPU actually runs at."
>>>>>>>
>>>>>>>
>>>>>>>[1]https://www.kernel.org/doc/Documentation/cpu-freq/user-guide.txt
>>>>>>>
>>>>>>>>Would you be able to test [1] on your platform and usecase?
>>>>>>>I has tested it on my platform (CPU number: 64, SMT: off and CPU base
>>>>>>>frequency: 2.7GHz).
>>>>>>>Accoding to the testing result,
>>>>>>>1> I found that patch [1] and [2] cannot cover the no housekeeping CPUs.
>>>>>>>They still have to face the large frequency discrepancy issue my patch
>>>>>>>mentioned.
>>>>>>>2> Additionally, the frequency value of all CPUs are almost the same by
>>>>>>>using the 'arch_freq_scale' factor way. I'm not sure if it is ok.
>>>>>>>
>>>>>>>The patch [1] has been modified silightly as below:
>>>>>>>-->
>>>>>>>@@ -1756,7 +1756,10 @@ static unsigned int
>>>>>>>cpufreq_verify_current_freq(struct cpufreq_policy *policy, b
>>>>>>>    {
>>>>>>>           unsigned int new_freq;
>>>>>>>
>>>>>>>-       new_freq = cpufreq_driver->get(policy->cpu);
>>>>>>>+       new_freq = arch_freq_get_on_cpu(policy->cpu);
>>>>>>>+       if (!new_freq)
>>>>>>>+               new_freq = cpufreq_driver->get(policy->cpu);
>>>>>>>+
>>>>>>As pointed out this change will not make it to the next version of the patch.
>>>>>>So I'd say you can safely ignore it and assume that arch_freq_get_on_cpu will
>>>>>>only be wired for sysfs nodes for scaling_cur_freq/cpuinfo_cur_freq
>>>>>>>           if (!new_freq)
>>>>>>>                   return 0;
>>>>>>>
>>>>>>>And the result is as follows:
>>>>>>>*case 1:**No setting the nohz_full and cpufreq use performance governor*
>>>>>>>*--> Step1: *read 'cpuinfo_cur_freq' in no pressure
>>>>>>>     0: 2699264     2: 2699264     4: 2699264     6: 2699264
>>>>>>>     8: 2696628    10: 2696628    12: 2696628    14: 2699264
>>>>>>>    16: 2699264    18: 2696628    20: 2699264    22: 2696628
>>>>>>>    24: 2699264    26: 2696628    28: 2699264    30: 2696628
>>>>>>>    32: 2696628    34: 2696628    36: 2696628    38: 2696628
>>>>>>>    40: 2699264    42: 2699264    44: 2696628    46: 2696628
>>>>>>>    48: 2696628    50: 2699264    52: 2699264    54: 2696628
>>>>>>>    56: 2696628    58: 2696628    60: 2696628    62: 2696628
>>>>>>>    64: 2696628    66: 2699264    68: 2696628    70: 2696628
>>>>>>>    72: 2699264    74: 2696628    76: 2696628    78: 2699264
>>>>>>>    80: 2696628    82: 2696628    84: 2699264    86: 2696628
>>>>>>>    88: 2696628    90: 2696628    92: 2696628    94: 2699264
>>>>>>>    96: 2696628    98: 2699264   100: 2699264   102: 2696628
>>>>>>>104: 2699264   106: 2699264   108: 2699264   110: 2696628
>>>>>>>112: 2699264   114: 2699264   116: 2699264   118: 2699264
>>>>>>>120: 2696628   122: 2699264   124: 2696628   126: 2699264
>>>>>>>Note: the frequency of all CPUs are almost the same.
>>>>>>Were you expecting smth else ?
>>>>>The frequency of each CPU might have a different value.
>>>>>All value of all CPUs is the same under high pressure.
>>>>>I don't know what the phenomenon is on other platform.
>>>>>Do you know who else tested it?
>>>>So I might have rushed a bit with my previous comment/question: apologies for
>>>>that.
>>>>The numbers above: those are on a fairly idle/lightly loaded system right?
>>>Yes.
>>>>Would you mind having another go with just the arch_freq_get_on_cpu
>>>>implementation beign added and dropping the changes in the cpufreq and
>>>All my tests are done when cpufreq policy is "performance" and OS isn't on a
>>>high load.
>>>Reading "scaling_cur_freq" or "scaling_cur_freq" for each physical core on
>>>platform
>>>
>>>The testing result for "cpuinfo_cur_freq" with your changes on a fairly idle
>>>and high loaded system can also be found in this thread.
>>>*A: the result with your changes*
>>>--> Reading "scaling_cur_freq"
>>>   0: 2688720     2: 2696628     4: 2699264     6: 2696628
>>>   8: 2699264    10: 2696628    12: 2699264    14: 2699264
>>>  16: 2699264    18: 2696628    20: 2696628    22: 2696628
>>>  24: 2699264    26: 2696628    28: 2696628    30: 2696628
>>>  32: 2699264    34: 2691356    36: 2696628    38: 2699264
>>>  40: 2699264    42: 2696628    44: 2696628    46: 2699264
>>>  48: 2699264    50: 2696628    52: 2696628    54: 2696628
>>>  56: 2696628    58: 2699264    60: 2691356    62: 2696628
>>>  64: 2696628    66: 2696628    68: 2696628    70: 2696628
>>>  72: 2696628    74: 2696628    76: 2699264    78: 2696628
>>>  80: 2696628    82: 2696628    84: 2699264    86: 2696628
>>>  88: 2625456    90: 2696628    92: 2699264    94: 2696628
>>>  96: 2696628    98: 2696628   100: 2699264   102: 2699264
>>>104: 2699264   106: 2696628   108: 2699264   110: 2696628
>>>112: 2699264   114: 2699264   116: 2696628   118: 2696628
>>>120: 2696628   122: 2699264   124: 2696628   126: 2696628
>>>-->Reading  "cpuinfo_cur_freq"
>>>   0: 2696628     2: 2696628     4: 2699264     6: 2688720
>>>   8: 2699264    10: 2700000    12: 2696628    14: 2698322
>>>  16: 2699264    18: 2699264    20: 2696628    22: 2699264
>>>  24: 2699264    26: 2699264    28: 2699264    30: 2699264
>>>  32: 2699264    34: 2693992    36: 2696628    38: 2696628
>>>  40: 2699264    42: 2699264    44: 2699264    46: 2696628
>>>  48: 2696628    50: 2699264    52: 2696628    54: 2696628
>>>  56: 2699264    58: 2699264    60: 2696628    62: 2699264
>>>  64: 2696628    66: 2699264    68: 2696628    70: 2699264
>>>  72: 2696628    74: 2696628    76: 2696628    78: 2693992
>>>  80: 2696628    82: 2696628    84: 2696628    86: 2696628
>>>  88: 2696628    90: 2699264    92: 2696628    94: 2699264
>>>  96: 2699264    98: 2696628   100: 2699264   102: 2699264
>>>104: 2691356   106: 2699264   108: 2699264   110: 2699264
>>>112: 2699264   114: 2696628   116: 2699264   118: 2699264
>>>120: 2696628   122: 2696628   124: 2696628   126: 2696628
>>>
>>>*B: the result without your changes*
>>>-->Reading "scaling_cur_freq"
>>>   0: 2698245     2: 2706690     4: 2699649     6: 2702105
>>>   8: 2704362    10: 2697993    12: 2701672    14: 2704362
>>>  16: 2701052    18: 2701052    20: 2694385    22: 2699650
>>>  24: 2706802    26: 2702389    28: 2698299    30: 2698299
>>>  32: 2697333    34: 2697993    36: 2701337    38: 2699328
>>>  40: 2700330    42: 2700330    44: 2698019    46: 2697697
>>>  48: 2699659    50: 2701700    52: 2703401    54: 2701700
>>>  56: 2704013    58: 2697658    60: 2695000    62: 2697666
>>>  64: 2697902    66: 2701052    68: 2698245    70: 2695789
>>>  72: 2701315    74: 2696655    76: 2693666    78: 2695317
>>>  80: 2704912    82: 2699649    84: 2698245    86: 2695454
>>>  88: 2697966    90: 2697959    92: 2699319    94: 2700680
>>>  96: 2695317    98: 2698996   100: 2700000   102: 2700334
>>>104: 2701320   106: 2695065   108: 2700986   110: 2703960
>>>112: 2697635   114: 2704421   116: 2700680   118: 2702040
>>>120: 2700334   122: 2697993   124: 2700334   126: 2705351
>>>-->Reading "cpuinfo_cur_freq"
>>>   0: 2696853     2: 2695454     4: 2699649     6: 2706993
>>>   8: 2706060    10: 2704362    12: 2704362    14: 2697658
>>>  16: 2707719    18: 2697192    20: 2702456    22: 2699650
>>>  24: 2705782    26: 2698299    28: 2703061    30: 2705802
>>>  32: 2700000    34: 2700671    36: 2701337    38: 2697658
>>>  40: 2700330    42: 2700330    44: 2699672    46: 2697697
>>>  48: 2703061    50: 2696610    52: 2692542    54: 2704406
>>>  56: 2695317    58: 2699331    60: 2698996    62: 2702675
>>>  64: 2704912    66: 2703859    68: 2699649    70: 2698596
>>>  72: 2703908    74: 2703355    76: 2697658    78: 2695317
>>>  80: 2702105    82: 2707719    84: 2702105    86: 2699649
>>>  88: 2697966    90: 2691525    92: 2701700    94: 2700680
>>>  96: 2695317    98: 2698996   100: 2698666   102: 2700334
>>>104: 2690429   106: 2707590   108: 2700986   110: 2701320
>>>112: 2696283   114: 2692881   116: 2697627   118: 2704421
>>>120: 2698996   122: 2696321   124: 2696655   126: 2695000
>>>
>>So in both cases : whether you use arch_freq_get_on_cpu or not
>>(so with and without the patch) you get roughly the same frequencies
>>on all cores - or am I missing smth from the dump above ?
>The changes in "with/without your changes" I said is your patch 
>intruduced arch_freq_get_on_cpu.
>I just test them according to your requesting.
>>And those are reflecting max freq you have provided earlier (?)
>I know it is an average frequency for the last tickfor using 
>arch_freq_get_on_cpu.
>I have no any doubt that the freq is maximum value on performance governor.
>I just want to say the difference between having or not having your patch.
>The frequency values of all cores from cpuinfo_cur_freq and
>scaling_cur_freq are almost the same if use this arch_freq_get_on_cpu 
>on my platform.
>However, the frequency values of all cores are different if doesn't 
>use this arch_freq_get_on_cpu and just use .get().
>>Note that the arch_freq_get_on_cpu will return an average frequency for
>>the last tick, so even if your system is roughly idle with your performance
>>governor those numbers make sense (some/most of the cores might be idle
>>but you will see the last freq the core was running at before going to idle).
>>I do not think there is an agreement what should be shown for idle core when
>>querying their freq through sysfs. Showing last known freq makes sense, even
>>more than waking up core just to try to get one.
>I'm not opposed to using frequency scale factor to get CPU frequency. 
>But it better be okay.
>>
>>@Ionela: Please jump in if I got things wrong.
>>
>>>>then read 'scaling_cur_freq', doing several reads in some intervals ?
>>>It seems that above phenomenon has not a lot to do with reading intervals.
>>>>The change has been tested on RD-N2 model (Neoverse N2 ref platform),
>>>>it has also been discussed here [1]
>>>I doesn't get the testing result on this platform in its thread.
>>It might be missing exact numbers but the conclusions should be here [1]
>>
>>>>>>>*--> Step 2: *read 'cpuinfo_cur_freq' in the high memory access pressure.
>>>>>>>     0: 2696628     2: 2696628     4: 2696628     6: 2696628
>>>>>>>     8: 2696628    10: 2696628    12: 2696628    14: 2696628
>>>>>>>    16: 2696628    18: 2696628    20: 2696628    22: 2696628
>>>>>>>    24: 2696628    26: 2696628    28: 2696628    30: 2696628
>>>>>>>    32: 2696628    34: 2696628    36: 2696628    38: 2696628
>>>>>>>    40: 2696628    42: 2696628    44: 2696628    46: 2696628
>>>>>>>    48: 2696628    50: 2696628    52: 2696628    54: 2696628
>>>>>>>    56: 2696628    58: 2696628    60: 2696628    62: 2696628
>>>>>>>    64: 2696628    66: 2696628    68: 2696628    70: 2696628
>>>>>>>    72: 2696628    74: 2696628    76: 2696628    78: 2696628
>>>>>>>    80: 2696628    82: 2696628    84: 2696628    86: 2696628
>>>>>>>    88: 2696628    90: 2696628    92: 2696628    94: 2696628
>>>>>>>    96: 2696628    98: 2696628   100: 2696628   102: 2696628
>>>>>>>104: 2696628   106: 2696628   108: 2696628   110: 2696628
>>>>>>>112: 2696628   114: 2696628   116: 2696628   118: 2696628
>>>>>>>120: 2696628   122: 2696628   124: 2696628   126: 2696628
>>>>>>>
>>>>>>>*Case 2: setting nohz_full and cpufreq use ondemand governor*
>>>>>>>There is "isolcpus=1-10,41-50 nohz_full=1-10,41-50 rcu_nocbs=1-10,41-50" in
>>>>>>>/proc/cmdline.
>>>>>>Right, so if I remember correctly nohz_full implies rcu_nocbs, so no need to
>>>>>>set that one.
>>>>>>Now, afair, isolcpus will make the selected CPUs to disappear from the
>>>>>>schedulers view (no balancing, no migrating), so unless you affine smth
>>>>>>explicitly to those CPUs, you will not see much of an activity there.
>>>>>Correct.
>>>>>>Need to double check though as it has been a while ...
>>>>>>>*--> Step 1: *setting ondemand governor to all policy and query
>>>>>>>'cpuinfo_cur_freq' in no pressure case.
>>>>>>>And the frequency of CPUs all are about 400MHz.
>>>>>>>*--> Step 2:* read 'cpuinfo_cur_freq' in the high memory access pressure.
>>>>>>>The high memory access pressure is from the command: "stress-ng -c 64
>>>>>>>--cpu-load 100% --taskset 0-63"
>>>>>>I'm not entirely convinced that this will affine to isolated cpus, especially
>>>>>>that the affinity mask spans all available cpus. If that is the case, no wonder
>>>>>>your isolated cpus are getting wasted being idle. But I would have to double
>>>>>>check how this is being handled.
>>>>>>>The result:
>>>>>>>    0: 2696628     1:  400000     2:  400000     3:  400909
>>>>>>>    4:  400000     5:  400000     6:  400000     7:  400000
>>>>>>>    8:  400000     9:  400000    10:  400600    11: 2696628
>>>>>>>12: 2696628    13: 2696628    14: 2696628    15: 2696628
>>>>>>>16: 2696628    17: 2696628    18: 2696628    19: 2696628
>>>>>>>20: 2696628    21: 2696628    22: 2696628    23: 2696628
>>>>>>>24: 2696628    25: 2696628    26: 2696628    27: 2696628
>>>>>>>28: 2696628    29: 2696628    30: 2696628    31: 2696628
>>>>>>>32: 2696628    33: 2696628    34: 2696628    35: 2696628
>>>>>>>36: 2696628    37: 2696628    38: 2696628    39: 2696628
>>>>>>>40: 2696628    41:  400000    42:  400000    43:  400000
>>>>>>>44:  400000    45:  398847    46:  400000    47:  400000
>>>>>>>48:  400000    49:  400000    50:  400000    51: 2696628
>>>>>>>52: 2696628    53: 2696628    54: 2696628    55: 2696628
>>>>>>>56: 2696628    57: 2696628    58: 2696628    59: 2696628
>>>>>>>60: 2696628    61: 2696628    62: 2696628    63: 2699264
>>>>>>>
>>>>>>>Note:
>>>>>>>(1) The frequency of 1-10 and 41-50 CPUs work on the lowest frequency.
>>>>>>>        It turned out that nohz full was already work.
>>>>>>>        I guess that stress-ng cannot use the CPU in the range of nohz full.
>>>>>>>        Because the CPU frequency will be increased to 2.7G by binding CPU to
>>>>>>>other application.
>>>>>>>(2) The frequency of the nohz full core is calculated by get() callback
>>>>>>>according to ftrace.
>>>>>>It is as there is no sched tick on those, and apparently there is nothing
>>>>>>running on them either.
>>>>>Yes.
>>>>>If we select your approach and the above phenomenon is normal,
>>>>>the large frequency discrepancy issue can be resolved for CPUs with sched
>>>>>tick by the way.
>>>>>But the nohz full cores still have to face this issue. So this patch is also
>>>>>needed.
>>>>>
>>>>Yes, nohz cores full have to be handled by the cpufreq driver.
>>>Correct. So we still have to face the issue in this patch and push this
>>>patch.
>>>Beata, would you please review this patch?
>>Just to clarify for my benefit (apologies but I do have to contex switch
>>pretty often these days): by reviewing this patch do you mean:
>>1) review your changes (if so I think there are few comments already to be
>>addressed, but I can try to have another look)
>Currently, the main comments is that my patch will wake up CPU to get 
>frequency.
>BTW, the core's always been wakened up to get the frequency for FFH 
>way in cppc_acpi. please see cpc_read_ffh().
>So it may be acceptable. After all, we don't query CPU frequency very often.

Today's implementation of cpc_read_ffh() wakes up the core to read AMU
counters - this is far from ideal. According to the architecture
specification the CPU_CYCLES and CNT_CYCLES counters in AMU do not
increment when the core is in WFI or WFE. If we cache the value of the
AMU counter before a PE goes idle, we may be able to avoid waking up a
PE just to read the AMU counters. I'm wondering if it makes sense to
cache the value in cpu_do_idle() and return this cached value if
idle_cpu() returns true.

>But your patch doesn't meet the non-housekeeping cpus.

For non-housekeeping CPUs may be it is better to just invoke
cpufreq->get() call?

Thanks,
Vanshi

>>2) review changes for AMU-based arch_freq_get_on_cpu ?
>>
>>*note: I will still try to have a look at the non-housekeeping cpus case
>I am very much hope that this issue my patch mentioned can be resolved ASAP.
>So what's your plan about non-housekeeping cpus case?
>>
>>---
>>[1] https://lore.kernel.org/lkml/691d3eb2-cd93-f0fc-a7a4-2a8c0d44262c@nvidia.com/
>>---
>>
>>BR
>>Beata
>>>
>>>/Huisong
>>[...]
>>.
lihuisong (C) Feb. 27, 2024, 12:24 p.m. UTC | #23
在 2024/2/21 0:11, Vanshidhar Konda 写道:
> On Mon, Feb 19, 2024 at 08:15:50PM +0800, lihuisong (C) wrote:
>>
>> 在 2024/2/9 18:55, Beata Michalska 写道:
>>> On Tue, Feb 06, 2024 at 04:02:15PM +0800, lihuisong (C) wrote:
>>>> 在 2024/2/2 16:08, Beata Michalska 写道:
>>>>> On Wed, Jan 17, 2024 at 05:18:40PM +0800, lihuisong (C) wrote:
>>>>>
>>>>> Hi ,
>>>>>
>>>>> Again, apologies for delay,
>>>>>
>>>>>> Hi,
>>>>>>
>>>>>> 在 2024/1/16 22:10, Beata Michalska 写道:
>>>>>>> Hi,
>>>>>>>
>>>>>>> Apologies for jumping in so late....
>>>>>>>
>>>>>>> On Wed, Jan 10, 2024 at 03:09:48PM +0800, lihuisong (C) wrote:
>>>>>>>> Hi Ionela,
>>>>>>>>
>>>>>>>> 在 2024/1/8 22:03, Ionela Voinescu 写道:
>>>>>>>>> Hi,
>>>>>>>>>
>>>>>>>>> On Friday 05 Jan 2024 at 15:04:47 (+0800), lihuisong (C) wrote:
>>>>>>>>>> Hi Vanshi,
>>>>>>>>>>
>>>>>>>>>> 在 2024/1/5 8:48, Vanshidhar Konda 写道:
>>>>>>>>>>> On Thu, Jan 04, 2024 at 05:36:51PM +0800, lihuisong (C) wrote:
>>>>>>>>>>>> 在 2024/1/4 1:53, Ionela Voinescu 写道:
>>>>>>>>>>>>> Hi,
>>>>>>>>>>>>>
>>>>>>>>>>>>> On Tuesday 12 Dec 2023 at 15:26:17 (+0800), Huisong Li wrote:
>>>>>>>>>>>>>> Many developers found that the cpu current frequency is 
>>>>>>>>>>>>>> greater than
>>>>>>>>>>>>>> the maximum frequency of the platform, please see [1], 
>>>>>>>>>>>>>> [2] and [3].
>>>>>>>>>>>>>>
>>>>>>>>>>>>>> In the scenarios with high memory access pressure, the 
>>>>>>>>>>>>>> patch [1] has
>>>>>>>>>>>>>> proved the significant latency of cpc_read() which is 
>>>>>>>>>>>>>> used to obtain
>>>>>>>>>>>>>> delivered and reference performance counter cause an 
>>>>>>>>>>>>>> absurd frequency.
>>>>>>>>>>>>>> The sampling interval for this counters is very critical and
>>>>>>>>>>>>>> is expected
>>>>>>>>>>>>>> to be equal. However, the different latency of cpc_read() 
>>>>>>>>>>>>>> has a direct
>>>>>>>>>>>>>> impact on their sampling interval.
>>>>>>>>>>>>>>
>>>>>>>>>>>>> Would this [1] alternative solution work for you?
>>>>>>>>>>>> It would work for me AFAICS.
>>>>>>>>>>>> Because the "arch_freq_scale" is also from AMU core and 
>>>>>>>>>>>> constant
>>>>>>>>>>>> counter, and read together.
>>>>>>>>>>>> But, from their discuss line, it seems that there are some 
>>>>>>>>>>>> tricky
>>>>>>>>>>>> points to clarify or consider.
>>>>>>>>>>> I think the changes in [1] would work better when CPUs may 
>>>>>>>>>>> be idle. With
>>>>>>>>>>> this
>>>>>>>>>>> patch we would have to wake any core that is in idle state 
>>>>>>>>>>> to read the
>>>>>>>>>>> AMU
>>>>>>>>>>> counters. Worst case, if core 0 is trying to read the CPU 
>>>>>>>>>>> frequency of
>>>>>>>>>>> all
>>>>>>>>>>> cores, it may need to wake up all the other cores to read 
>>>>>>>>>>> the AMU
>>>>>>>>>>> counters.
>>>>>>>>>>    From the approach in [1], if all CPUs (one or more cores) 
>>>>>>>>>> under one policy
>>>>>>>>>> are idle, they still cannot be obtained the CPU frequency, 
>>>>>>>>>> right?
>>>>>>>>>> In this case, the [1] API will return 0 and have to back to call
>>>>>>>>>> cpufreq_driver->get() for cpuinfo_cur_freq.
>>>>>>>>>> Then we still need to face the issue this patch mentioned.
>>>>>>>>> With the implementation at [1], arch_freq_get_on_cpu() will 
>>>>>>>>> not return 0
>>>>>>>>> for idle CPUs and the get() callback will not be called to 
>>>>>>>>> wake up the
>>>>>>>>> CPUs.
>>>>>>>> Right, arch_freq_get_on_cpu() will not return 0 for idle CPUs.
>>>>>>>> However, for no-housekeeping CPUs, it will return 0 and have to 
>>>>>>>> call get()
>>>>>>>> callback, right?
>>>>>>>>> Worst case, arch_freq_get_on_cpu() will return a frequency 
>>>>>>>>> based on the
>>>>>>>>> AMU counter values obtained on the last tick on that CPU. But 
>>>>>>>>> if that CPU
>>>>>>>>> is not a housekeeping CPU, a housekeeping CPU in the same 
>>>>>>>>> policy will be
>>>>>>>>> selected, as it would have had a more recent tick, and 
>>>>>>>>> therefore a more
>>>>>>>>> recent frequency value for the domain.
>>>>>>>> But this frequency is from the last tick,
>>>>>>>> this last tick is probably a long time ago and it doesn't update
>>>>>>>> 'arch_freq_scale' for some reasons like CPU dile.
>>>>>>>> In addition, I'm not sure if there is possible that 
>>>>>>>> amu_scale_freq_tick() is
>>>>>>>> executed delayed under high stress case.
>>>>>>>> It also have an impact on the accuracy of the cpu frequency we 
>>>>>>>> query.
>>>>>>>>> I understand that the frequency returned here will not be up 
>>>>>>>>> to date,
>>>>>>>>> but there's no proper frequency feedback for an idle CPU. If 
>>>>>>>>> one only
>>>>>>>>> wakes up a CPU to sample counters, before the CPU goes back to 
>>>>>>>>> sleep,
>>>>>>>>> the obtained frequency feedback is meaningless.
>>>>>>>>>
>>>>>>>>>>> For systems with 128 cores or more, this could be very 
>>>>>>>>>>> expensive and
>>>>>>>>>>> happen
>>>>>>>>>>> very frequently.
>>>>>>>>>>>
>>>>>>>>>>> AFAICS, the approach in [1] would avoid this cost.
>>>>>>>>>> But the CPU frequency is just an average value for the last 
>>>>>>>>>> tick period
>>>>>>>>>> instead of the current one the CPU actually runs at.
>>>>>>>>>> In addition, there are some conditions to use 
>>>>>>>>>> 'arch_freq_scale' in this
>>>>>>>>>> approach.
>>>>>>>>> What are the conditions you are referring to?
>>>>>>>> It depends on the housekeeping CPUs.
>>>>>>>>>> So I'm not sure if this approach can entirely cover the 
>>>>>>>>>> frequency
>>>>>>>>>> discrepancy issue.
>>>>>>>>> Unfortunately there is no perfect frequency feedback. By the 
>>>>>>>>> time you
>>>>>>>>> observe/use the value of scaling_cur_freq/cpuinfo_cur_freq, 
>>>>>>>>> the frequency
>>>>>>>>> of the CPU might have already changed. Therefore, an average 
>>>>>>>>> value might
>>>>>>>>> be a better indication of the recent performance level of a CPU.
>>>>>>>> An average value for CPU frequency is ok. It may be better if 
>>>>>>>> it has not any
>>>>>>>> delaying.
>>>>>>>>
>>>>>>>> The original implementation for cpuinfo_cur_freq can more 
>>>>>>>> reflect their
>>>>>>>> meaning in the user-guide [1]. The user-guide said:
>>>>>>>> "cpuinfo_cur_freq : Current frequency of the CPU as obtained 
>>>>>>>> from the
>>>>>>>> hardware, in KHz.
>>>>>>>> This is the frequency the CPU actually runs at."
>>>>>>>>
>>>>>>>>
>>>>>>>> [1]https://www.kernel.org/doc/Documentation/cpu-freq/user-guide.txt 
>>>>>>>>
>>>>>>>>
>>>>>>>>> Would you be able to test [1] on your platform and usecase?
>>>>>>>> I has tested it on my platform (CPU number: 64, SMT: off and 
>>>>>>>> CPU base
>>>>>>>> frequency: 2.7GHz).
>>>>>>>> Accoding to the testing result,
>>>>>>>> 1> I found that patch [1] and [2] cannot cover the no 
>>>>>>>> housekeeping CPUs.
>>>>>>>> They still have to face the large frequency discrepancy issue 
>>>>>>>> my patch
>>>>>>>> mentioned.
>>>>>>>> 2> Additionally, the frequency value of all CPUs are almost the 
>>>>>>>> same by
>>>>>>>> using the 'arch_freq_scale' factor way. I'm not sure if it is ok.
>>>>>>>>
>>>>>>>> The patch [1] has been modified silightly as below:
>>>>>>>> -->
>>>>>>>> @@ -1756,7 +1756,10 @@ static unsigned int
>>>>>>>> cpufreq_verify_current_freq(struct cpufreq_policy *policy, b
>>>>>>>>    {
>>>>>>>>           unsigned int new_freq;
>>>>>>>>
>>>>>>>> -       new_freq = cpufreq_driver->get(policy->cpu);
>>>>>>>> +       new_freq = arch_freq_get_on_cpu(policy->cpu);
>>>>>>>> +       if (!new_freq)
>>>>>>>> +               new_freq = cpufreq_driver->get(policy->cpu);
>>>>>>>> +
>>>>>>> As pointed out this change will not make it to the next version 
>>>>>>> of the patch.
>>>>>>> So I'd say you can safely ignore it and assume that 
>>>>>>> arch_freq_get_on_cpu will
>>>>>>> only be wired for sysfs nodes for scaling_cur_freq/cpuinfo_cur_freq
>>>>>>>>           if (!new_freq)
>>>>>>>>                   return 0;
>>>>>>>>
>>>>>>>> And the result is as follows:
>>>>>>>> *case 1:**No setting the nohz_full and cpufreq use performance 
>>>>>>>> governor*
>>>>>>>> *--> Step1: *read 'cpuinfo_cur_freq' in no pressure
>>>>>>>>     0: 2699264     2: 2699264     4: 2699264     6: 2699264
>>>>>>>>     8: 2696628    10: 2696628    12: 2696628    14: 2699264
>>>>>>>>    16: 2699264    18: 2696628    20: 2699264    22: 2696628
>>>>>>>>    24: 2699264    26: 2696628    28: 2699264    30: 2696628
>>>>>>>>    32: 2696628    34: 2696628    36: 2696628    38: 2696628
>>>>>>>>    40: 2699264    42: 2699264    44: 2696628    46: 2696628
>>>>>>>>    48: 2696628    50: 2699264    52: 2699264    54: 2696628
>>>>>>>>    56: 2696628    58: 2696628    60: 2696628    62: 2696628
>>>>>>>>    64: 2696628    66: 2699264    68: 2696628    70: 2696628
>>>>>>>>    72: 2699264    74: 2696628    76: 2696628    78: 2699264
>>>>>>>>    80: 2696628    82: 2696628    84: 2699264    86: 2696628
>>>>>>>>    88: 2696628    90: 2696628    92: 2696628    94: 2699264
>>>>>>>>    96: 2696628    98: 2699264   100: 2699264   102: 2696628
>>>>>>>> 104: 2699264   106: 2699264   108: 2699264   110: 2696628
>>>>>>>> 112: 2699264   114: 2699264   116: 2699264   118: 2699264
>>>>>>>> 120: 2696628   122: 2699264   124: 2696628   126: 2699264
>>>>>>>> Note: the frequency of all CPUs are almost the same.
>>>>>>> Were you expecting smth else ?
>>>>>> The frequency of each CPU might have a different value.
>>>>>> All value of all CPUs is the same under high pressure.
>>>>>> I don't know what the phenomenon is on other platform.
>>>>>> Do you know who else tested it?
>>>>> So I might have rushed a bit with my previous comment/question: 
>>>>> apologies for
>>>>> that.
>>>>> The numbers above: those are on a fairly idle/lightly loaded 
>>>>> system right?
>>>> Yes.
>>>>> Would you mind having another go with just the arch_freq_get_on_cpu
>>>>> implementation beign added and dropping the changes in the cpufreq 
>>>>> and
>>>> All my tests are done when cpufreq policy is "performance" and OS 
>>>> isn't on a
>>>> high load.
>>>> Reading "scaling_cur_freq" or "scaling_cur_freq" for each physical 
>>>> core on
>>>> platform
>>>>
>>>> The testing result for "cpuinfo_cur_freq" with your changes on a 
>>>> fairly idle
>>>> and high loaded system can also be found in this thread.
>>>> *A: the result with your changes*
>>>> --> Reading "scaling_cur_freq"
>>>>   0: 2688720     2: 2696628     4: 2699264     6: 2696628
>>>>   8: 2699264    10: 2696628    12: 2699264    14: 2699264
>>>>  16: 2699264    18: 2696628    20: 2696628    22: 2696628
>>>>  24: 2699264    26: 2696628    28: 2696628    30: 2696628
>>>>  32: 2699264    34: 2691356    36: 2696628    38: 2699264
>>>>  40: 2699264    42: 2696628    44: 2696628    46: 2699264
>>>>  48: 2699264    50: 2696628    52: 2696628    54: 2696628
>>>>  56: 2696628    58: 2699264    60: 2691356    62: 2696628
>>>>  64: 2696628    66: 2696628    68: 2696628    70: 2696628
>>>>  72: 2696628    74: 2696628    76: 2699264    78: 2696628
>>>>  80: 2696628    82: 2696628    84: 2699264    86: 2696628
>>>>  88: 2625456    90: 2696628    92: 2699264    94: 2696628
>>>>  96: 2696628    98: 2696628   100: 2699264   102: 2699264
>>>> 104: 2699264   106: 2696628   108: 2699264   110: 2696628
>>>> 112: 2699264   114: 2699264   116: 2696628   118: 2696628
>>>> 120: 2696628   122: 2699264   124: 2696628   126: 2696628
>>>> -->Reading  "cpuinfo_cur_freq"
>>>>   0: 2696628     2: 2696628     4: 2699264     6: 2688720
>>>>   8: 2699264    10: 2700000    12: 2696628    14: 2698322
>>>>  16: 2699264    18: 2699264    20: 2696628    22: 2699264
>>>>  24: 2699264    26: 2699264    28: 2699264    30: 2699264
>>>>  32: 2699264    34: 2693992    36: 2696628    38: 2696628
>>>>  40: 2699264    42: 2699264    44: 2699264    46: 2696628
>>>>  48: 2696628    50: 2699264    52: 2696628    54: 2696628
>>>>  56: 2699264    58: 2699264    60: 2696628    62: 2699264
>>>>  64: 2696628    66: 2699264    68: 2696628    70: 2699264
>>>>  72: 2696628    74: 2696628    76: 2696628    78: 2693992
>>>>  80: 2696628    82: 2696628    84: 2696628    86: 2696628
>>>>  88: 2696628    90: 2699264    92: 2696628    94: 2699264
>>>>  96: 2699264    98: 2696628   100: 2699264   102: 2699264
>>>> 104: 2691356   106: 2699264   108: 2699264   110: 2699264
>>>> 112: 2699264   114: 2696628   116: 2699264   118: 2699264
>>>> 120: 2696628   122: 2696628   124: 2696628   126: 2696628
>>>>
>>>> *B: the result without your changes*
>>>> -->Reading "scaling_cur_freq"
>>>>   0: 2698245     2: 2706690     4: 2699649     6: 2702105
>>>>   8: 2704362    10: 2697993    12: 2701672    14: 2704362
>>>>  16: 2701052    18: 2701052    20: 2694385    22: 2699650
>>>>  24: 2706802    26: 2702389    28: 2698299    30: 2698299
>>>>  32: 2697333    34: 2697993    36: 2701337    38: 2699328
>>>>  40: 2700330    42: 2700330    44: 2698019    46: 2697697
>>>>  48: 2699659    50: 2701700    52: 2703401    54: 2701700
>>>>  56: 2704013    58: 2697658    60: 2695000    62: 2697666
>>>>  64: 2697902    66: 2701052    68: 2698245    70: 2695789
>>>>  72: 2701315    74: 2696655    76: 2693666    78: 2695317
>>>>  80: 2704912    82: 2699649    84: 2698245    86: 2695454
>>>>  88: 2697966    90: 2697959    92: 2699319    94: 2700680
>>>>  96: 2695317    98: 2698996   100: 2700000   102: 2700334
>>>> 104: 2701320   106: 2695065   108: 2700986   110: 2703960
>>>> 112: 2697635   114: 2704421   116: 2700680   118: 2702040
>>>> 120: 2700334   122: 2697993   124: 2700334   126: 2705351
>>>> -->Reading "cpuinfo_cur_freq"
>>>>   0: 2696853     2: 2695454     4: 2699649     6: 2706993
>>>>   8: 2706060    10: 2704362    12: 2704362    14: 2697658
>>>>  16: 2707719    18: 2697192    20: 2702456    22: 2699650
>>>>  24: 2705782    26: 2698299    28: 2703061    30: 2705802
>>>>  32: 2700000    34: 2700671    36: 2701337    38: 2697658
>>>>  40: 2700330    42: 2700330    44: 2699672    46: 2697697
>>>>  48: 2703061    50: 2696610    52: 2692542    54: 2704406
>>>>  56: 2695317    58: 2699331    60: 2698996    62: 2702675
>>>>  64: 2704912    66: 2703859    68: 2699649    70: 2698596
>>>>  72: 2703908    74: 2703355    76: 2697658    78: 2695317
>>>>  80: 2702105    82: 2707719    84: 2702105    86: 2699649
>>>>  88: 2697966    90: 2691525    92: 2701700    94: 2700680
>>>>  96: 2695317    98: 2698996   100: 2698666   102: 2700334
>>>> 104: 2690429   106: 2707590   108: 2700986   110: 2701320
>>>> 112: 2696283   114: 2692881   116: 2697627   118: 2704421
>>>> 120: 2698996   122: 2696321   124: 2696655   126: 2695000
>>>>
>>> So in both cases : whether you use arch_freq_get_on_cpu or not
>>> (so with and without the patch) you get roughly the same frequencies
>>> on all cores - or am I missing smth from the dump above ?
>> The changes in "with/without your changes" I said is your patch 
>> intruduced arch_freq_get_on_cpu.
>> I just test them according to your requesting.
>>> And those are reflecting max freq you have provided earlier (?)
>> I know it is an average frequency for the last tickfor using 
>> arch_freq_get_on_cpu.
>> I have no any doubt that the freq is maximum value on performance 
>> governor.
>> I just want to say the difference between having or not having your 
>> patch.
>> The frequency values of all cores from cpuinfo_cur_freq and
>> scaling_cur_freq are almost the same if use this arch_freq_get_on_cpu 
>> on my platform.
>> However, the frequency values of all cores are different if doesn't 
>> use this arch_freq_get_on_cpu and just use .get().
>>> Note that the arch_freq_get_on_cpu will return an average frequency for
>>> the last tick, so even if your system is roughly idle with your 
>>> performance
>>> governor those numbers make sense (some/most of the cores might be idle
>>> but you will see the last freq the core was running at before going 
>>> to idle).
>>> I do not think there is an agreement what should be shown for idle 
>>> core when
>>> querying their freq through sysfs. Showing last known freq makes 
>>> sense, even
>>> more than waking up core just to try to get one.
>> I'm not opposed to using frequency scale factor to get CPU frequency. 
>> But it better be okay.
>>>
>>> @Ionela: Please jump in if I got things wrong.
>>>
>>>>> then read 'scaling_cur_freq', doing several reads in some intervals ?
>>>> It seems that above phenomenon has not a lot to do with reading 
>>>> intervals.
>>>>> The change has been tested on RD-N2 model (Neoverse N2 ref platform),
>>>>> it has also been discussed here [1]
>>>> I doesn't get the testing result on this platform in its thread.
>>> It might be missing exact numbers but the conclusions should be here 
>>> [1]
>>>
>>>>>>>> *--> Step 2: *read 'cpuinfo_cur_freq' in the high memory access 
>>>>>>>> pressure.
>>>>>>>>     0: 2696628     2: 2696628     4: 2696628     6: 2696628
>>>>>>>>     8: 2696628    10: 2696628    12: 2696628    14: 2696628
>>>>>>>>    16: 2696628    18: 2696628    20: 2696628    22: 2696628
>>>>>>>>    24: 2696628    26: 2696628    28: 2696628    30: 2696628
>>>>>>>>    32: 2696628    34: 2696628    36: 2696628    38: 2696628
>>>>>>>>    40: 2696628    42: 2696628    44: 2696628    46: 2696628
>>>>>>>>    48: 2696628    50: 2696628    52: 2696628    54: 2696628
>>>>>>>>    56: 2696628    58: 2696628    60: 2696628    62: 2696628
>>>>>>>>    64: 2696628    66: 2696628    68: 2696628    70: 2696628
>>>>>>>>    72: 2696628    74: 2696628    76: 2696628    78: 2696628
>>>>>>>>    80: 2696628    82: 2696628    84: 2696628    86: 2696628
>>>>>>>>    88: 2696628    90: 2696628    92: 2696628    94: 2696628
>>>>>>>>    96: 2696628    98: 2696628   100: 2696628   102: 2696628
>>>>>>>> 104: 2696628   106: 2696628   108: 2696628   110: 2696628
>>>>>>>> 112: 2696628   114: 2696628   116: 2696628   118: 2696628
>>>>>>>> 120: 2696628   122: 2696628   124: 2696628   126: 2696628
>>>>>>>>
>>>>>>>> *Case 2: setting nohz_full and cpufreq use ondemand governor*
>>>>>>>> There is "isolcpus=1-10,41-50 nohz_full=1-10,41-50 
>>>>>>>> rcu_nocbs=1-10,41-50" in
>>>>>>>> /proc/cmdline.
>>>>>>> Right, so if I remember correctly nohz_full implies rcu_nocbs, 
>>>>>>> so no need to
>>>>>>> set that one.
>>>>>>> Now, afair, isolcpus will make the selected CPUs to disappear 
>>>>>>> from the
>>>>>>> schedulers view (no balancing, no migrating), so unless you 
>>>>>>> affine smth
>>>>>>> explicitly to those CPUs, you will not see much of an activity 
>>>>>>> there.
>>>>>> Correct.
>>>>>>> Need to double check though as it has been a while ...
>>>>>>>> *--> Step 1: *setting ondemand governor to all policy and query
>>>>>>>> 'cpuinfo_cur_freq' in no pressure case.
>>>>>>>> And the frequency of CPUs all are about 400MHz.
>>>>>>>> *--> Step 2:* read 'cpuinfo_cur_freq' in the high memory access 
>>>>>>>> pressure.
>>>>>>>> The high memory access pressure is from the command: "stress-ng 
>>>>>>>> -c 64
>>>>>>>> --cpu-load 100% --taskset 0-63"
>>>>>>> I'm not entirely convinced that this will affine to isolated 
>>>>>>> cpus, especially
>>>>>>> that the affinity mask spans all available cpus. If that is the 
>>>>>>> case, no wonder
>>>>>>> your isolated cpus are getting wasted being idle. But I would 
>>>>>>> have to double
>>>>>>> check how this is being handled.
>>>>>>>> The result:
>>>>>>>>    0: 2696628     1:  400000     2:  400000     3: 400909
>>>>>>>>    4:  400000     5:  400000     6:  400000     7: 400000
>>>>>>>>    8:  400000     9:  400000    10:  400600    11: 2696628
>>>>>>>> 12: 2696628    13: 2696628    14: 2696628    15: 2696628
>>>>>>>> 16: 2696628    17: 2696628    18: 2696628    19: 2696628
>>>>>>>> 20: 2696628    21: 2696628    22: 2696628    23: 2696628
>>>>>>>> 24: 2696628    25: 2696628    26: 2696628    27: 2696628
>>>>>>>> 28: 2696628    29: 2696628    30: 2696628    31: 2696628
>>>>>>>> 32: 2696628    33: 2696628    34: 2696628    35: 2696628
>>>>>>>> 36: 2696628    37: 2696628    38: 2696628    39: 2696628
>>>>>>>> 40: 2696628    41:  400000    42:  400000    43: 400000
>>>>>>>> 44:  400000    45:  398847    46:  400000    47: 400000
>>>>>>>> 48:  400000    49:  400000    50:  400000    51: 2696628
>>>>>>>> 52: 2696628    53: 2696628    54: 2696628    55: 2696628
>>>>>>>> 56: 2696628    57: 2696628    58: 2696628    59: 2696628
>>>>>>>> 60: 2696628    61: 2696628    62: 2696628    63: 2699264
>>>>>>>>
>>>>>>>> Note:
>>>>>>>> (1) The frequency of 1-10 and 41-50 CPUs work on the lowest 
>>>>>>>> frequency.
>>>>>>>>        It turned out that nohz full was already work.
>>>>>>>>        I guess that stress-ng cannot use the CPU in the range 
>>>>>>>> of nohz full.
>>>>>>>>        Because the CPU frequency will be increased to 2.7G by 
>>>>>>>> binding CPU to
>>>>>>>> other application.
>>>>>>>> (2) The frequency of the nohz full core is calculated by get() 
>>>>>>>> callback
>>>>>>>> according to ftrace.
>>>>>>> It is as there is no sched tick on those, and apparently there 
>>>>>>> is nothing
>>>>>>> running on them either.
>>>>>> Yes.
>>>>>> If we select your approach and the above phenomenon is normal,
>>>>>> the large frequency discrepancy issue can be resolved for CPUs 
>>>>>> with sched
>>>>>> tick by the way.
>>>>>> But the nohz full cores still have to face this issue. So this 
>>>>>> patch is also
>>>>>> needed.
>>>>>>
>>>>> Yes, nohz cores full have to be handled by the cpufreq driver.
>>>> Correct. So we still have to face the issue in this patch and push 
>>>> this
>>>> patch.
>>>> Beata, would you please review this patch?
>>> Just to clarify for my benefit (apologies but I do have to contex 
>>> switch
>>> pretty often these days): by reviewing this patch do you mean:
>>> 1) review your changes (if so I think there are few comments already 
>>> to be
>>> addressed, but I can try to have another look)
>> Currently, the main comments is that my patch will wake up CPU to get 
>> frequency.
>> BTW, the core's always been wakened up to get the frequency for FFH 
>> way in cppc_acpi. please see cpc_read_ffh().
>> So it may be acceptable. After all, we don't query CPU frequency very 
>> often.
>
> Today's implementation of cpc_read_ffh() wakes up the core to read AMU
> counters - this is far from ideal. According to the architecture
> specification the CPU_CYCLES and CNT_CYCLES counters in AMU do not
> increment when the core is in WFI or WFE. If we cache the value of the
> AMU counter before a PE goes idle, we may be able to avoid waking up a
> PE just to read the AMU counters. I'm wondering if it makes sense to
> cache the value in cpu_do_idle() and return this cached value if
> idle_cpu() returns true.
It just might be useful for the idle state from WFI, right?
What about other idle states?
it will be a little complex.

The 'cpuinfo_cur_freq' is feedback for CPU frequency, and we firstly 
need to ensure that its function is ok.
What's more, I guess that user don't query CPU frequency very often when 
OS system is in idle.
Because it is not the center of attention.
 From the point of view, it is acceptable to wake up the core to read 
AMU counters and fundamentally resolve this issue .
What do you think?
>
>> But your patch doesn't meet the non-housekeeping cpus.
>
> For non-housekeeping CPUs may be it is better to just invoke
> cpufreq->get() call?

Then we're still going to have this issue.

>
> Thanks,
> Vanshi
>
>>> 2) review changes for AMU-based arch_freq_get_on_cpu ?
>>>
>>> *note: I will still try to have a look at the non-housekeeping cpus 
>>> case
>> I am very much hope that this issue my patch mentioned can be 
>> resolved ASAP.
>> So what's your plan about non-housekeeping cpus case?
>>>
>>> ---
>>> [1] 
>>> https://lore.kernel.org/lkml/691d3eb2-cd93-f0fc-a7a4-2a8c0d44262c@nvidia.com/
>>> ---
>>>
>>> BR
>>> Beata
>>>>
>>>> /Huisong
>>> [...]
>>> .
> .
diff mbox series

Patch

diff --git a/arch/arm64/kernel/topology.c b/arch/arm64/kernel/topology.c
index 7d37e458e2f5..c3122154d738 100644
--- a/arch/arm64/kernel/topology.c
+++ b/arch/arm64/kernel/topology.c
@@ -299,6 +299,11 @@  core_initcall(init_amu_fie);
 #ifdef CONFIG_ACPI_CPPC_LIB
 #include <acpi/cppc_acpi.h>
 
+struct amu_counters {
+	u64 corecnt;
+	u64 constcnt;
+};
+
 static void cpu_read_corecnt(void *val)
 {
 	/*
@@ -322,8 +327,27 @@  static void cpu_read_constcnt(void *val)
 		      0UL : read_constcnt();
 }
 
+static void cpu_read_amu_counters(void *data)
+{
+	struct amu_counters *cnt = (struct amu_counters *)data;
+
+	/*
+	 * The running time of the this_cpu_has_cap() might have a couple of
+	 * microseconds and is significantly increased to tens of microseconds.
+	 * But AMU core and constant counter need to be read togeter without any
+	 * time interval to reduce the calculation discrepancy using this counters.
+	 */
+	if (this_cpu_has_cap(ARM64_WORKAROUND_2457168)) {
+		cnt->corecnt = read_corecnt();
+		cnt->constcnt = 0;
+	} else {
+		cnt->corecnt = read_corecnt();
+		cnt->constcnt = read_constcnt();
+	}
+}
+
 static inline
-int counters_read_on_cpu(int cpu, smp_call_func_t func, u64 *val)
+int counters_read_on_cpu(int cpu, smp_call_func_t func, void *data)
 {
 	/*
 	 * Abort call on counterless CPU or when interrupts are
@@ -335,7 +359,7 @@  int counters_read_on_cpu(int cpu, smp_call_func_t func, u64 *val)
 	if (WARN_ON_ONCE(irqs_disabled()))
 		return -EPERM;
 
-	smp_call_function_single(cpu, func, val, 1);
+	smp_call_function_single(cpu, func, data, 1);
 
 	return 0;
 }
@@ -364,6 +388,21 @@  bool cpc_ffh_supported(void)
 	return true;
 }
 
+int cpc_read_arch_counters_on_cpu(int cpu, u64 *delivered, u64 *reference)
+{
+	struct amu_counters cnts = {0};
+	int ret;
+
+	ret = counters_read_on_cpu(cpu, cpu_read_amu_counters, &cnts);
+	if (ret)
+		return ret;
+
+	*delivered = cnts.corecnt;
+	*reference = cnts.constcnt;
+
+	return 0;
+}
+
 int cpc_read_ffh(int cpu, struct cpc_reg *reg, u64 *val)
 {
 	int ret = -EOPNOTSUPP;
diff --git a/drivers/acpi/cppc_acpi.c b/drivers/acpi/cppc_acpi.c
index 7ff269a78c20..f303fabd7cfe 100644
--- a/drivers/acpi/cppc_acpi.c
+++ b/drivers/acpi/cppc_acpi.c
@@ -1299,6 +1299,11 @@  bool cppc_perf_ctrs_in_pcc(void)
 }
 EXPORT_SYMBOL_GPL(cppc_perf_ctrs_in_pcc);
 
+int __weak cpc_read_arch_counters_on_cpu(int cpu, u64 *delivered, u64 *reference)
+{
+	return 0;
+}
+
 /**
  * cppc_get_perf_ctrs - Read a CPU's performance feedback counters.
  * @cpunum: CPU from which to read counters.
@@ -1313,7 +1318,8 @@  int cppc_get_perf_ctrs(int cpunum, struct cppc_perf_fb_ctrs *perf_fb_ctrs)
 		*ref_perf_reg, *ctr_wrap_reg;
 	int pcc_ss_id = per_cpu(cpu_pcc_subspace_idx, cpunum);
 	struct cppc_pcc_data *pcc_ss_data = NULL;
-	u64 delivered, reference, ref_perf, ctr_wrap_time;
+	u64 delivered = 0, reference = 0;
+	u64 ref_perf, ctr_wrap_time;
 	int ret = 0, regs_in_pcc = 0;
 
 	if (!cpc_desc) {
@@ -1350,8 +1356,18 @@  int cppc_get_perf_ctrs(int cpunum, struct cppc_perf_fb_ctrs *perf_fb_ctrs)
 		}
 	}
 
-	cpc_read(cpunum, delivered_reg, &delivered);
-	cpc_read(cpunum, reference_reg, &reference);
+	if (cpc_ffh_supported()) {
+		ret = cpc_read_arch_counters_on_cpu(cpunum, &delivered, &reference);
+		if (ret) {
+			pr_debug("read arch counters failed, ret=%d.\n", ret);
+			ret = 0;
+		}
+	}
+	if (!delivered || !reference) {
+		cpc_read(cpunum, delivered_reg, &delivered);
+		cpc_read(cpunum, reference_reg, &reference);
+	}
+
 	cpc_read(cpunum, ref_perf_reg, &ref_perf);
 
 	/*
diff --git a/include/acpi/cppc_acpi.h b/include/acpi/cppc_acpi.h
index 6126c977ece0..07d4fd82d499 100644
--- a/include/acpi/cppc_acpi.h
+++ b/include/acpi/cppc_acpi.h
@@ -152,6 +152,7 @@  extern bool cpc_ffh_supported(void);
 extern bool cpc_supported_by_cpu(void);
 extern int cpc_read_ffh(int cpunum, struct cpc_reg *reg, u64 *val);
 extern int cpc_write_ffh(int cpunum, struct cpc_reg *reg, u64 val);
+extern int cpc_read_arch_counters_on_cpu(int cpu, u64 *delivered, u64 *reference);
 extern int cppc_get_epp_perf(int cpunum, u64 *epp_perf);
 extern int cppc_set_epp_perf(int cpu, struct cppc_perf_ctrls *perf_ctrls, bool enable);
 extern int cppc_get_auto_sel_caps(int cpunum, struct cppc_perf_caps *perf_caps);
@@ -209,6 +210,10 @@  static inline int cpc_write_ffh(int cpunum, struct cpc_reg *reg, u64 val)
 {
 	return -ENOTSUPP;
 }
+static inline int cpc_read_arch_counters_on_cpu(int cpu, u64 *delivered, u64 *reference)
+{
+	return -EOPNOTSUPP;
+}
 static inline int cppc_set_epp_perf(int cpu, struct cppc_perf_ctrls *perf_ctrls, bool enable)
 {
 	return -ENOTSUPP;