drivers/cpufreq/intel_pstate.c | 4 ++-- include/linux/energy_model.h | 10 ++++++++++ kernel/power/energy_model.c | 29 +++++++++++++++++++++++++---- 3 files changed, 37 insertions(+), 6 deletions(-)
From: Rafael J. Wysocki <rafael.j.wysocki@intel.com>
The intel_pstate driver manages CPU capacity changes itself and it does
not need an update of the capacity of all CPUs in the system to be
carried out after registering a PD.
Moreover, in some configurations (for instance, an SMT-capable
hybrid x86 system booted with nosmt in the kernel command line) the
em_check_capacity_update() call at the end of em_dev_register_perf_domain()
always fails and reschedules itself to run once again in 1 s, so
effectively it runs in vain every 1 s forever.
To address this, introduce a new variant of em_dev_register_perf_domain(),
called em_dev_register_pd_no_update(), that does not invoke
em_check_capacity_update(), and make intel_pstate use it instead of the
original.
Fixes: 7b010f9b9061 ("cpufreq: intel_pstate: EAS support for hybrid platforms")
Link: https://lore.kernel.org/linux-pm/40212796-734c-4140-8a85-854f72b8144d@panix.com/
Reported-by: Kenneth R. Crudup <kenny@panix.com>
Tested-by: Kenneth R. Crudup <kenny@panix.com>
Signed-off-by: Rafael J. Wysocki <rafael.j.wysocki@intel.com>
Cc: 6.16+ <stable@vger.kernel.org> # 6.16+
---
Normally, I would place the definition of em_dev_register_pd_no_update()
before the definition of em_dev_register_perf_domain() which called it,
but in this particular case that would cause the real change to be much
harder to grasp.
---
drivers/cpufreq/intel_pstate.c | 4 ++--
include/linux/energy_model.h | 10 ++++++++++
kernel/power/energy_model.c | 29 +++++++++++++++++++++++++----
3 files changed, 37 insertions(+), 6 deletions(-)
--- a/drivers/cpufreq/intel_pstate.c
+++ b/drivers/cpufreq/intel_pstate.c
@@ -1034,8 +1034,8 @@ static bool hybrid_register_perf_domain(
if (!cpu_dev)
return false;
- if (em_dev_register_perf_domain(cpu_dev, HYBRID_EM_STATE_COUNT, &cb,
- cpumask_of(cpu), false))
+ if (em_dev_register_pd_no_update(cpu_dev, HYBRID_EM_STATE_COUNT, &cb,
+ cpumask_of(cpu), false))
return false;
cpudata->pd_registered = true;
--- a/include/linux/energy_model.h
+++ b/include/linux/energy_model.h
@@ -171,6 +171,9 @@ int em_dev_update_perf_domain(struct dev
int em_dev_register_perf_domain(struct device *dev, unsigned int nr_states,
const struct em_data_callback *cb,
const cpumask_t *cpus, bool microwatts);
+int em_dev_register_pd_no_update(struct device *dev, unsigned int nr_states,
+ const struct em_data_callback *cb,
+ const cpumask_t *cpus, bool microwatts);
void em_dev_unregister_perf_domain(struct device *dev);
struct em_perf_table *em_table_alloc(struct em_perf_domain *pd);
void em_table_free(struct em_perf_table *table);
@@ -350,6 +353,13 @@ int em_dev_register_perf_domain(struct d
{
return -EINVAL;
}
+static inline
+int em_dev_register_pd_no_update(struct device *dev, unsigned int nr_states,
+ const struct em_data_callback *cb,
+ const cpumask_t *cpus, bool microwatts)
+{
+ return -EINVAL;
+}
static inline void em_dev_unregister_perf_domain(struct device *dev)
{
}
--- a/kernel/power/energy_model.c
+++ b/kernel/power/energy_model.c
@@ -553,6 +553,30 @@ int em_dev_register_perf_domain(struct d
const struct em_data_callback *cb,
const cpumask_t *cpus, bool microwatts)
{
+ int ret = em_dev_register_pd_no_update(dev, nr_states, cb, cpus, microwatts);
+
+ if (_is_cpu_device(dev))
+ em_check_capacity_update();
+
+ return ret;
+}
+EXPORT_SYMBOL_GPL(em_dev_register_perf_domain);
+
+/**
+ * em_dev_register_pd_no_update() - Register a perf domain for a device
+ * @dev : Device to register the PD for
+ * @nr_states : Number of performance states in the new PD
+ * @cb : Callback functions for populating the energy model
+ * @cpus : CPUs to include in the new PD (mandatory if @dev is a CPU device)
+ * @microwatts : Whether or not the power values in the EM will be in uW
+ *
+ * Like em_dev_register_perf_domain(), but does not trigger a CPU capacity
+ * update after registering the PD, even if @dev is a CPU device.
+ */
+int em_dev_register_pd_no_update(struct device *dev, unsigned int nr_states,
+ const struct em_data_callback *cb,
+ const cpumask_t *cpus, bool microwatts)
+{
struct em_perf_table *em_table;
unsigned long cap, prev_cap = 0;
unsigned long flags = 0;
@@ -636,12 +660,9 @@ int em_dev_register_perf_domain(struct d
unlock:
mutex_unlock(&em_pd_mutex);
- if (_is_cpu_device(dev))
- em_check_capacity_update();
-
return ret;
}
-EXPORT_SYMBOL_GPL(em_dev_register_perf_domain);
+EXPORT_SYMBOL_GPL(em_dev_register_pd_no_update);
/**
* em_dev_unregister_perf_domain() - Unregister Energy Model (EM) for a deviceOn Fri, Sep 5, 2025 at 3:44 PM Rafael J. Wysocki <rafael@kernel.org> wrote:
>
> From: Rafael J. Wysocki <rafael.j.wysocki@intel.com>
>
> The intel_pstate driver manages CPU capacity changes itself and it does
> not need an update of the capacity of all CPUs in the system to be
> carried out after registering a PD.
>
> Moreover, in some configurations (for instance, an SMT-capable
> hybrid x86 system booted with nosmt in the kernel command line) the
> em_check_capacity_update() call at the end of em_dev_register_perf_domain()
> always fails and reschedules itself to run once again in 1 s, so
> effectively it runs in vain every 1 s forever.
>
> To address this, introduce a new variant of em_dev_register_perf_domain(),
> called em_dev_register_pd_no_update(), that does not invoke
> em_check_capacity_update(), and make intel_pstate use it instead of the
> original.
>
> Fixes: 7b010f9b9061 ("cpufreq: intel_pstate: EAS support for hybrid platforms")
> Link: https://lore.kernel.org/linux-pm/40212796-734c-4140-8a85-854f72b8144d@panix.com/
> Reported-by: Kenneth R. Crudup <kenny@panix.com>
> Tested-by: Kenneth R. Crudup <kenny@panix.com>
> Signed-off-by: Rafael J. Wysocki <rafael.j.wysocki@intel.com>
> Cc: 6.16+ <stable@vger.kernel.org> # 6.16+
I'd appreciate feedback on this.
Since it is 6.16-stable material, I'm inclined to just apply it and
let it into the mainline.
> ---
>
> Normally, I would place the definition of em_dev_register_pd_no_update()
> before the definition of em_dev_register_perf_domain() which called it,
> but in this particular case that would cause the real change to be much
> harder to grasp.
>
> ---
> drivers/cpufreq/intel_pstate.c | 4 ++--
> include/linux/energy_model.h | 10 ++++++++++
> kernel/power/energy_model.c | 29 +++++++++++++++++++++++++----
> 3 files changed, 37 insertions(+), 6 deletions(-)
>
> --- a/drivers/cpufreq/intel_pstate.c
> +++ b/drivers/cpufreq/intel_pstate.c
> @@ -1034,8 +1034,8 @@ static bool hybrid_register_perf_domain(
> if (!cpu_dev)
> return false;
>
> - if (em_dev_register_perf_domain(cpu_dev, HYBRID_EM_STATE_COUNT, &cb,
> - cpumask_of(cpu), false))
> + if (em_dev_register_pd_no_update(cpu_dev, HYBRID_EM_STATE_COUNT, &cb,
> + cpumask_of(cpu), false))
> return false;
>
> cpudata->pd_registered = true;
> --- a/include/linux/energy_model.h
> +++ b/include/linux/energy_model.h
> @@ -171,6 +171,9 @@ int em_dev_update_perf_domain(struct dev
> int em_dev_register_perf_domain(struct device *dev, unsigned int nr_states,
> const struct em_data_callback *cb,
> const cpumask_t *cpus, bool microwatts);
> +int em_dev_register_pd_no_update(struct device *dev, unsigned int nr_states,
> + const struct em_data_callback *cb,
> + const cpumask_t *cpus, bool microwatts);
> void em_dev_unregister_perf_domain(struct device *dev);
> struct em_perf_table *em_table_alloc(struct em_perf_domain *pd);
> void em_table_free(struct em_perf_table *table);
> @@ -350,6 +353,13 @@ int em_dev_register_perf_domain(struct d
> {
> return -EINVAL;
> }
> +static inline
> +int em_dev_register_pd_no_update(struct device *dev, unsigned int nr_states,
> + const struct em_data_callback *cb,
> + const cpumask_t *cpus, bool microwatts)
> +{
> + return -EINVAL;
> +}
> static inline void em_dev_unregister_perf_domain(struct device *dev)
> {
> }
> --- a/kernel/power/energy_model.c
> +++ b/kernel/power/energy_model.c
> @@ -553,6 +553,30 @@ int em_dev_register_perf_domain(struct d
> const struct em_data_callback *cb,
> const cpumask_t *cpus, bool microwatts)
> {
> + int ret = em_dev_register_pd_no_update(dev, nr_states, cb, cpus, microwatts);
> +
> + if (_is_cpu_device(dev))
> + em_check_capacity_update();
> +
> + return ret;
> +}
> +EXPORT_SYMBOL_GPL(em_dev_register_perf_domain);
> +
> +/**
> + * em_dev_register_pd_no_update() - Register a perf domain for a device
> + * @dev : Device to register the PD for
> + * @nr_states : Number of performance states in the new PD
> + * @cb : Callback functions for populating the energy model
> + * @cpus : CPUs to include in the new PD (mandatory if @dev is a CPU device)
> + * @microwatts : Whether or not the power values in the EM will be in uW
> + *
> + * Like em_dev_register_perf_domain(), but does not trigger a CPU capacity
> + * update after registering the PD, even if @dev is a CPU device.
> + */
> +int em_dev_register_pd_no_update(struct device *dev, unsigned int nr_states,
> + const struct em_data_callback *cb,
> + const cpumask_t *cpus, bool microwatts)
> +{
> struct em_perf_table *em_table;
> unsigned long cap, prev_cap = 0;
> unsigned long flags = 0;
> @@ -636,12 +660,9 @@ int em_dev_register_perf_domain(struct d
> unlock:
> mutex_unlock(&em_pd_mutex);
>
> - if (_is_cpu_device(dev))
> - em_check_capacity_update();
> -
> return ret;
> }
> -EXPORT_SYMBOL_GPL(em_dev_register_perf_domain);
> +EXPORT_SYMBOL_GPL(em_dev_register_pd_no_update);
>
> /**
> * em_dev_unregister_perf_domain() - Unregister Energy Model (EM) for a device
>
>
>
>
© 2016 - 2025 Red Hat, Inc.