Re: [PATCH v5] ACPI: CPPC: Add ospm_nominal_perf support

[Sumit Gupta <sumitg@nvidia.com>]

On 01/07/26 21:53, Pierre Gondois wrote:
> External email: Use caution opening links or attachments
>
>
> Hello Sumit,
>
> On 6/15/26 20:59, Sumit Gupta wrote:
>> Expose the OSPM Nominal Performance register (ACPI 6.6, Section
>> 8.4.6.1.2.6), which conveys the desired nominal performance level
>> at which the platform may run. Unlike the existing read-only
>> Nominal Performance register, it is writable and lets OSPM
>> request a lower nominal level than the platform-reported nominal.
>> The platform classifies performance above this level as boosted
>> and below as throttled for its power/thermal decisions.
>>
>> It is exposed as a per-policy cpufreq sysfs attribute in kHz, to
>> match the cpufreq sysfs unit convention:
>>
>>    /sys/devices/system/cpu/cpufreq/policyN/ospm_nominal_freq
>>
>> The attribute is documented in
>> Documentation/ABI/testing/sysfs-devices-system-cpu.
>>
>> Writes are converted to perf via cppc_khz_to_perf(), validated
>> against [Lowest Performance, Nominal Performance], and applied to
>> every CPU in policy->cpus.
>>
>> On read, the current register value is returned, or
>> "<unsupported>" if the platform does not implement the register.
>>
>> Signed-off-by: Sumit Gupta<sumitg@nvidia.com>
>> ---
>> Patch 1 of the v4 series ("ACPI: CPPC: Add support for CPPC v4") is
>> already applied, so this contains only patch 2.
>>
>> Changes in v5:
>> - Add cppc_get_ospm_nominal_perf() to read the register directly.
>> - Drop the cppc_cpudata cache variables ospm_nominal_perf/_set.
>> - Show_ospm_nominal_freq() returns register value or "<unsupported>".
>> - Register rollback reads the register too.
>> - Move range check into the sysfs store from 
>> cppc_set_ospm_nominal_perf()
>> - ABI doc: update read description and add a task-migration note.
>>
>> v4:https://lore.kernel.org/lkml/20260527194626.185286-1-sumitg@nvidia.com/ 
>>
>> v3:https://lore.kernel.org/lkml/20260514194822.1841748-1-sumitg@nvidia.com/ 
>>
>> v2:https://lore.kernel.org/lkml/20260430142430.755437-1-sumitg@nvidia.com/ 
>>
>> v1:https://lore.kernel.org/lkml/20260427051823.280419-1-sumitg@nvidia.com/ 
>>
>>
>>   .../ABI/testing/sysfs-devices-system-cpu      | 26 ++++++++
>>   drivers/acpi/cppc_acpi.c                      | 32 +++++++++
>>   drivers/cpufreq/cppc_cpufreq.c                | 65 +++++++++++++++++++
>>   include/acpi/cppc_acpi.h                      | 10 +++
>>   4 files changed, 133 insertions(+)
>>
>> diff --git a/Documentation/ABI/testing/sysfs-devices-system-cpu 
>> b/Documentation/ABI/testing/sysfs-devices-system-cpu
>> index 82d10d556cc8..a8d592c08823 100644
>> --- a/Documentation/ABI/testing/sysfs-devices-system-cpu
>> +++ b/Documentation/ABI/testing/sysfs-devices-system-cpu
>> @@ -346,6 +346,32 @@ Description:     Performance Limited
>>
>>               This file is only present if the cppc-cpufreq driver is 
>> in use.
>>
>> +What: /sys/devices/system/cpu/cpuX/cpufreq/ospm_nominal_freq
>> +Date:                May 2026
>> +Contact:     linux-pm@vger.kernel.org
>> +Description: OSPM Nominal Performance (kHz)
>> +
>> +             OSPM uses this attribute to request a nominal performance
>> +             level lower than the platform-reported nominal. The
>> +             platform treats performance above this level as boost
>> +             and below as throttle for power and thermal decisions.
>> +
>> +             Read returns the current value in kHz, or "<unsupported>"
>> +             if the platform does not implement the register. Write a
>> +             kHz value in the range [lowest_freq, nominal_freq].
>> +
>> +             Note that tasks may be migrated from one CPU to another
>> +             by the scheduler's load-balancing algorithm, and if
>> +             different OSPM Nominal Performance values are set for
>> +             those CPUs (through different cpufreq policies), that may
>> +             lead to undesirable outcomes. To avoid such issues it is
>> +             better to set the same value across all policies, or to
>> +             pin every task potentially sensitive to it to a specific
>> +             CPU.
>> +
>> +             This file is only present if the cppc-cpufreq driver is
>> +             in use.
>> +
>>   What: /sys/devices/system/cpu/cpu*/cache/index3/cache_disable_{0,1}
>>   Date:               August 2008
>>   KernelVersion:      2.6.27
>> diff --git a/drivers/acpi/cppc_acpi.c b/drivers/acpi/cppc_acpi.c
>> index 9f572f481241..1fcc22a10b4c 100644
>> --- a/drivers/acpi/cppc_acpi.c
>> +++ b/drivers/acpi/cppc_acpi.c
>> @@ -1685,6 +1685,38 @@ int cppc_set_epp(int cpu, u64 epp_val)
>>   }
>>   EXPORT_SYMBOL_GPL(cppc_set_epp);
>>
>> +/**
>> + * cppc_set_ospm_nominal_perf() - Write OSPM Nominal Performance 
>> register.
>> + * @cpu: CPU on which to write register.
>> + * @ospm_nominal_perf: Value to write to the OSPM Nominal 
>> Performance register.
>> + *
>> + * OSPM Nominal Performance conveys the desired nominal performance 
>> level
>> + * at which the platform may run. Per ACPI 6.6, s8.4.6.1.2.6, the value
>> + * must lie within [Lowest Performance, Nominal Performance] and may be
>> + * set independently of Minimum, Maximum and Desired performance. The
>> + * caller is responsible for validating the range.
>> + *
>> + * Return: 0 on success or negative error code.
>> + */
>> +int cppc_set_ospm_nominal_perf(int cpu, u64 ospm_nominal_perf)
>> +{
>> +     return cppc_set_reg_val(cpu, OSPM_NOMINAL_PERF, 
>> ospm_nominal_perf);
>> +}
>> +EXPORT_SYMBOL_GPL(cppc_set_ospm_nominal_perf);
>> +
>> +/**
>> + * cppc_get_ospm_nominal_perf() - Read OSPM Nominal Performance 
>> register.
>> + * @cpu: CPU from which to read register.
>> + * @ospm_nominal_perf: Pointer to store the OSPM Nominal Performance 
>> value.
>> + *
>> + * Return: 0 on success or negative error code.
>> + */
>> +int cppc_get_ospm_nominal_perf(int cpu, u64 *ospm_nominal_perf)
>> +{
>> +     return cppc_get_reg_val(cpu, OSPM_NOMINAL_PERF, 
>> ospm_nominal_perf);
>> +}
>> +EXPORT_SYMBOL_GPL(cppc_get_ospm_nominal_perf);
>> +
>>   /**
>>    * cppc_get_auto_act_window() - Read autonomous activity window 
>> register.
>>    * @cpu: CPU from which to read register.
>> diff --git a/drivers/cpufreq/cppc_cpufreq.c 
>> b/drivers/cpufreq/cppc_cpufreq.c
>> index f6cea0c54dd9..d160ceced7d9 100644
>> --- a/drivers/cpufreq/cppc_cpufreq.c
>> +++ b/drivers/cpufreq/cppc_cpufreq.c
>> @@ -1011,11 +1011,75 @@ static int cppc_get_perf_limited_filtered(int 
>> cpu, u64 *perf_limited)
>>   CPPC_CPUFREQ_ATTR_RW_U64(perf_limited, cppc_get_perf_limited_filtered,
>>                        cppc_set_perf_limited)
>>
>> +static ssize_t show_ospm_nominal_freq(struct cpufreq_policy *policy, 
>> char *buf)
>> +{
>> +     struct cppc_cpudata *cpu_data = policy->driver_data;
>> +     u64 perf;
>> +     int ret;
>> +
>> +     ret = cppc_get_ospm_nominal_perf(policy->cpu, &perf);
>> +     if (ret == -EOPNOTSUPP)
>> +             return sysfs_emit(buf, "<unsupported>\n");
>> +     if (ret)
>> +             return ret;
>> +
>> +     return sysfs_emit(buf, "%u\n",
>> + cppc_perf_to_khz(&cpu_data->perf_caps, perf));
>> +}
>> +
>> +static ssize_t store_ospm_nominal_freq(struct cpufreq_policy *policy,
>> +                                    const char *buf, size_t count)
>> +{
>> +     struct cppc_cpudata *cpu_data = policy->driver_data;
>> +     unsigned int sib, freq_khz, failing_cpu = 0;
>> +     u64 prev_perf;
>> +     u32 perf;
>> +     int ret;
>> +
>> +     ret = kstrtouint(buf, 0, &freq_khz);
>> +     if (ret)
>> +             return ret;
>> +
>> +     perf = cppc_khz_to_perf(&cpu_data->perf_caps, freq_khz);
>> +     if (perf < cpu_data->perf_caps.lowest_perf ||
>> +         perf > cpu_data->perf_caps.nominal_perf)
>> +             return -EINVAL;
>> +
>
> Sorry I didn't notice in v3, but this function is the only
> one setting the registers for all CPUs in the policy.
>
> In v2, the interface was per-CPU. The main issue with that
> is that we need a mechanism to aggregate values of CPUs
> in the same policy.
>
> In v3/v4/v5, the interface was per-policy, but the
> ospm_nominal_freq was set for all CPUs CPPC regs individually.
> If we do that, the sequence:
> - unplug a CPU from a policy
> - set ospm_nominal_freq for the policy
> - hotplug back the CPU
> might cause inconsistencies in the CPPC regs.
>
> Maybe it is better to continue with the assumption that all
> CPUs in the same policy have the same interface ? This
> assumption seems to be made for auto_sel and other regs.
>
> If this makes sense, would it be possible to:
> - add such check and warn if it is not the case

Will send the mismatch check for the SHARED_TYPE_ANY case as a
separate patch, as it is a pre-existing issue.


> - set the ospm_nominal_freq value only for policy->cpu
> (and remove the rollback mechanism) 

Yes, will write only policy->cpu and drop the rollback assuming a
shared register.


>
> When updating the ospm_nominal_freq, we should also:
> - update policy->cpuinfo.max_freq if boost=0
> - refresh the freq. limits
> - set policy->boost_supported if it is not already the case
>
> Also, update cppc_cpufreq_set_boost() should rely on
> ospm_nominal_freq instead of caps->nominal_perf as a reference.
>
> Please let me know if this is ok
>
> Regards,
>
> Pierre

Agreed. I will add an "effective nominal" (ospm_nominal_freq when set,
else caps->nominal_perf) and use it everywhere. Also will clear
boost_supported when it is raised back to highest_perf.

Thanks,
Sumit