[PATCH v2 0/2] ACPI: processor: idle: Per-CPU idle driver for hybrid CPUs

[PATCH v2 0/2] ACPI: processor: idle: Per-CPU idle driver for hybrid CPUs

[Yaxiong Tian]

This series addresses limitations in the current ACPI idle driver model
for hybrid CPU architectures (e.g., ARM big.LITTLE, Intel Alder Lake),
where different core types have distinct _LPI-state characteristics.

This series introduces:

1. A per-CPU idle driver model to accurately represent idle-state per core type.
2. A new interface to fetch cpuidle_driver by CPU ID, required for early
   registration scenarios.

This issue was initially discussed at:
https://lore.kernel.org/linux-pm/97e8bc72-e44b-487a-91ba-206732094955@arm.com/T/#t

Changes since V2:
- Fix "using smp_processor_id() in preemptible" BUG in patch 0002.

Yaxiong Tian (2):
  cpuidle: Add interface to get cpuidle_driver by CPU ID
  ACPI: processor: idle: Replace single idle driver with per-CPU model
    for better hybrid CPU support

 drivers/acpi/Kconfig            |  1 +
 drivers/acpi/processor_driver.c |  3 +-
 drivers/acpi/processor_idle.c   | 66 +++++++++++++++++----------------
 drivers/cpuidle/driver.c        | 16 ++++++++
 include/acpi/processor.h        |  2 +-
 include/linux/cpuidle.h         |  4 ++
 6 files changed, 59 insertions(+), 33 deletions(-)

-- 
2.25.1

[PATCH v2 1/2] cpuidle: Add interface to get cpuidle_driver by CPU ID

[Yaxiong Tian]

Some drivers need to obtain the cpuidle_driver via CPU ID before the
cpuidle_device is registered. Therefore, this interface is added.

Tested-by: Shaobo Huang <huangshaobo2075@phytium.com.cn>
Signed-off-by: Yaxiong Tian <tianyaxiong@kylinos.cn>
Signed-off-by: Shaobo Huang <huangshaobo2075@phytium.com.cn>
Signed-off-by: Yinfeng Wang <wangyinfeng@phytium.com.cn>
Signed-off-by: Xu Wang<wangxu@phytium.com.cn>
---
 drivers/cpuidle/driver.c | 16 ++++++++++++++++
 include/linux/cpuidle.h  |  4 ++++
 2 files changed, 20 insertions(+)

diff --git a/drivers/cpuidle/driver.c b/drivers/cpuidle/driver.c
index 9bbfa594c442..8647f8165863 100644
--- a/drivers/cpuidle/driver.c
+++ b/drivers/cpuidle/driver.c
@@ -353,6 +353,22 @@ struct cpuidle_driver *cpuidle_get_cpu_driver(struct cpuidle_device *dev)
 }
 EXPORT_SYMBOL_GPL(cpuidle_get_cpu_driver);
 
+/**
+ * cpuidle_get_cpu_driver_by_cpu - return the driver registered for a CPU.
+ * @cpu_num: a valid cpu num
+ *
+ * Returns a struct cpuidle_driver pointer, or NULL if no driver is registered
+ * for the CPU associated with @cpu.
+ */
+struct cpuidle_driver *cpuidle_get_cpu_driver_by_cpu(int cpu_num)
+{
+	if (cpu_num < 0 || cpu_num >= nr_cpu_ids)
+		return NULL;
+
+	return __cpuidle_get_cpu_driver(cpu_num);
+}
+EXPORT_SYMBOL_GPL(cpuidle_get_cpu_driver_by_cpu);
+
 /**
  * cpuidle_driver_state_disabled - Disable or enable an idle state
  * @drv: cpuidle driver owning the state
diff --git a/include/linux/cpuidle.h b/include/linux/cpuidle.h
index a9ee4fe55dcf..aebbaae6a501 100644
--- a/include/linux/cpuidle.h
+++ b/include/linux/cpuidle.h
@@ -200,6 +200,7 @@ extern void cpuidle_disable_device(struct cpuidle_device *dev);
 extern int cpuidle_play_dead(void);
 
 extern struct cpuidle_driver *cpuidle_get_cpu_driver(struct cpuidle_device *dev);
+extern struct cpuidle_driver *cpuidle_get_cpu_driver_by_cpu(int cpu_num);
 static inline struct cpuidle_device *cpuidle_get_device(void)
 {return __this_cpu_read(cpuidle_devices); }
 #else
@@ -240,6 +241,9 @@ static inline void cpuidle_disable_device(struct cpuidle_device *dev) { }
 static inline int cpuidle_play_dead(void) {return -ENODEV; }
 static inline struct cpuidle_driver *cpuidle_get_cpu_driver(
 	struct cpuidle_device *dev) {return NULL; }
+
+static inline struct cpuidle_driver *cpuidle_get_cpu_driver_by_cpu(
+	int cpu_num) {return NULL; };
 static inline struct cpuidle_device *cpuidle_get_device(void) {return NULL; }
 #endif
 
-- 
2.25.1

[PATCH v2 2/2] ACPI: processor: idle: Replace single idle driver with per-CPU model for better hybrid CPU support

[Yaxiong Tian]

Current implementations of hybrid architectures (e.g., ARM64 big.LITTLE
and Intel Alder Lake) feature CPU cores with different exit latencies.
Using a single driver to describe_LPI states for all core types is
therefore suboptimal. This is further supported by ACPI specification
8.4.4.1 which states: "In a processor hierarchy, each node has its
own _LPI low-power states specific to that node."

To address these limitations, we replace the monolithic idle driver
with a per-CPU model. This approach enables accurate idle state representation
for each core type

Tested-by: Shaobo Huang <huangshaobo2075@phytium.com.cn>
Signed-off-by: Yaxiong Tian <tianyaxiong@kylinos.cn>
Signed-off-by: Shaobo Huang <huangshaobo2075@phytium.com.cn>
Signed-off-by: Yinfeng Wang <wangyinfeng@phytium.com.cn>
Signed-off-by: Xu Wang<wangxu@phytium.com.cn>
---
 drivers/acpi/Kconfig            |  1 +
 drivers/acpi/processor_driver.c |  3 +-
 drivers/acpi/processor_idle.c   | 66 +++++++++++++++++----------------
 include/acpi/processor.h        |  2 +-
 4 files changed, 39 insertions(+), 33 deletions(-)

diff --git a/drivers/acpi/Kconfig b/drivers/acpi/Kconfig
index ca00a5dbcf75..d92c0faca978 100644
--- a/drivers/acpi/Kconfig
+++ b/drivers/acpi/Kconfig
@@ -276,6 +276,7 @@ config ACPI_PROCESSOR_CSTATE
 config ACPI_PROCESSOR_IDLE
 	bool
 	select CPU_IDLE
+	select CPU_IDLE_MULTIPLE_DRIVERS
 
 config ACPI_MCFG
 	bool
diff --git a/drivers/acpi/processor_driver.c b/drivers/acpi/processor_driver.c
index 65e779be64ff..22db9c904437 100644
--- a/drivers/acpi/processor_driver.c
+++ b/drivers/acpi/processor_driver.c
@@ -166,7 +166,8 @@ static int __acpi_processor_start(struct acpi_device *device)
 	if (result && !IS_ENABLED(CONFIG_ACPI_CPU_FREQ_PSS))
 		dev_dbg(&device->dev, "CPPC data invalid or not present\n");
 
-	if (!cpuidle_get_driver() || cpuidle_get_driver() == &acpi_idle_driver)
+	if (!cpuidle_get_cpu_driver_by_cpu(pr->id) || cpuidle_get_cpu_driver_by_cpu(pr->id)
+			== per_cpu_ptr(&acpi_idle_driver, pr->id))
 		acpi_processor_power_init(pr);
 
 	acpi_pss_perf_init(pr);
diff --git a/drivers/acpi/processor_idle.c b/drivers/acpi/processor_idle.c
index 2c2dc559e0f8..857dc56654d4 100644
--- a/drivers/acpi/processor_idle.c
+++ b/drivers/acpi/processor_idle.c
@@ -51,16 +51,17 @@ module_param(latency_factor, uint, 0644);
 
 static DEFINE_PER_CPU(struct cpuidle_device *, acpi_cpuidle_device);
 
-struct cpuidle_driver acpi_idle_driver = {
-	.name =		"acpi_idle",
-	.owner =	THIS_MODULE,
-};
+DEFINE_PER_CPU(struct cpuidle_driver, acpi_idle_driver);
 
 #ifdef CONFIG_ACPI_PROCESSOR_CSTATE
 void acpi_idle_rescan_dead_smt_siblings(void)
 {
-	if (cpuidle_get_driver() == &acpi_idle_driver)
-		arch_cpu_rescan_dead_smt_siblings();
+	int cpu;
+	for_each_possible_cpu(cpu)
+		if (cpuidle_get_driver() == per_cpu_ptr(&acpi_idle_driver, cpu)) {
+			arch_cpu_rescan_dead_smt_siblings();
+			return;
+		}
 }
 
 static
@@ -738,12 +739,13 @@ static int acpi_processor_setup_cpuidle_cx(struct acpi_processor *pr,
 	int i, count = ACPI_IDLE_STATE_START;
 	struct acpi_processor_cx *cx;
 	struct cpuidle_state *state;
+	struct cpuidle_driver *drv = per_cpu_ptr(&acpi_idle_driver, pr->id);
 
 	if (max_cstate == 0)
 		max_cstate = 1;
 
 	for (i = 1; i < ACPI_PROCESSOR_MAX_POWER && i <= max_cstate; i++) {
-		state = &acpi_idle_driver.states[count];
+		state = &drv->states[count];
 		cx = &pr->power.states[i];
 
 		if (!cx->valid)
@@ -776,7 +778,7 @@ static int acpi_processor_setup_cstates(struct acpi_processor *pr)
 	int i, count;
 	struct acpi_processor_cx *cx;
 	struct cpuidle_state *state;
-	struct cpuidle_driver *drv = &acpi_idle_driver;
+	struct cpuidle_driver *drv = per_cpu_ptr(&acpi_idle_driver, pr->id);
 
 	if (max_cstate == 0)
 		max_cstate = 1;
@@ -1198,7 +1200,7 @@ static int acpi_processor_setup_lpi_states(struct acpi_processor *pr)
 	int i;
 	struct acpi_lpi_state *lpi;
 	struct cpuidle_state *state;
-	struct cpuidle_driver *drv = &acpi_idle_driver;
+	struct cpuidle_driver *drv = per_cpu_ptr(&acpi_idle_driver, pr->id);
 
 	if (!pr->flags.has_lpi)
 		return -EOPNOTSUPP;
@@ -1232,7 +1234,7 @@ static int acpi_processor_setup_lpi_states(struct acpi_processor *pr)
 static int acpi_processor_setup_cpuidle_states(struct acpi_processor *pr)
 {
 	int i;
-	struct cpuidle_driver *drv = &acpi_idle_driver;
+	struct cpuidle_driver *drv = per_cpu_ptr(&acpi_idle_driver, pr->id);
 
 	if (!pr->flags.power_setup_done || !pr->flags.power)
 		return -EINVAL;
@@ -1316,13 +1318,7 @@ int acpi_processor_power_state_has_changed(struct acpi_processor *pr)
 	if (!pr->flags.power_setup_done)
 		return -ENODEV;
 
-	/*
-	 * FIXME:  Design the ACPI notification to make it once per
-	 * system instead of once per-cpu.  This condition is a hack
-	 * to make the code that updates C-States be called once.
-	 */
-
-	if (pr->id == 0 && cpuidle_get_driver() == &acpi_idle_driver) {
+	if (cpuidle_get_cpu_driver_by_cpu(pr->id) == per_cpu_ptr(&acpi_idle_driver, pr->id)) {
 
 		/* Protect against cpu-hotplug */
 		cpus_read_lock();
@@ -1360,12 +1356,14 @@ int acpi_processor_power_state_has_changed(struct acpi_processor *pr)
 	return 0;
 }
 
-static int acpi_processor_registered;
 
 int acpi_processor_power_init(struct acpi_processor *pr)
 {
 	int retval;
+	struct cpumask *cpumask;
 	struct cpuidle_device *dev;
+	struct cpuidle_driver *drv = per_cpu_ptr(&acpi_idle_driver, pr->id);
+
 
 	if (disabled_by_idle_boot_param())
 		return 0;
@@ -1382,14 +1380,21 @@ int acpi_processor_power_init(struct acpi_processor *pr)
 	 */
 	if (pr->flags.power) {
 		/* Register acpi_idle_driver if not already registered */
-		if (!acpi_processor_registered) {
-			acpi_processor_setup_cpuidle_states(pr);
-			retval = cpuidle_register_driver(&acpi_idle_driver);
-			if (retval)
-				return retval;
-			pr_debug("%s registered with cpuidle\n",
-				 acpi_idle_driver.name);
+		acpi_processor_setup_cpuidle_states(pr);
+
+		drv->name = "acpi_idle";
+		drv->owner = THIS_MODULE;
+		cpumask = kzalloc(cpumask_size(), GFP_KERNEL);
+		cpumask_set_cpu(pr->id, cpumask);
+		drv->cpumask = cpumask;
+
+		retval = cpuidle_register_driver(drv);
+		if (retval) {
+			kfree(cpumask);
+			return retval;
 		}
+		pr_debug("cpu %d:%s registered with cpuidle\n", pr->id,
+			 drv->name);
 
 		dev = kzalloc(sizeof(*dev), GFP_KERNEL);
 		if (!dev)
@@ -1403,11 +1408,10 @@ int acpi_processor_power_init(struct acpi_processor *pr)
 		 */
 		retval = cpuidle_register_device(dev);
 		if (retval) {
-			if (acpi_processor_registered == 0)
-				cpuidle_unregister_driver(&acpi_idle_driver);
+			cpuidle_unregister_driver(drv);
+			kfree(cpumask);
 			return retval;
 		}
-		acpi_processor_registered++;
 	}
 	return 0;
 }
@@ -1415,17 +1419,17 @@ int acpi_processor_power_init(struct acpi_processor *pr)
 int acpi_processor_power_exit(struct acpi_processor *pr)
 {
 	struct cpuidle_device *dev = per_cpu(acpi_cpuidle_device, pr->id);
+	struct cpuidle_driver *drv = per_cpu_ptr(&acpi_idle_driver, pr->id);
 
 	if (disabled_by_idle_boot_param())
 		return 0;
 
 	if (pr->flags.power) {
 		cpuidle_unregister_device(dev);
-		acpi_processor_registered--;
-		if (acpi_processor_registered == 0)
-			cpuidle_unregister_driver(&acpi_idle_driver);
+		cpuidle_unregister_driver(drv);
 
 		kfree(dev);
+		kfree(drv->cpumask);
 	}
 
 	pr->flags.power_setup_done = 0;
diff --git a/include/acpi/processor.h b/include/acpi/processor.h
index d0eccbd920e5..36940c6b96cc 100644
--- a/include/acpi/processor.h
+++ b/include/acpi/processor.h
@@ -417,7 +417,7 @@ static inline void acpi_processor_throttling_init(void) {}
 #endif	/* CONFIG_ACPI_CPU_FREQ_PSS */
 
 /* in processor_idle.c */
-extern struct cpuidle_driver acpi_idle_driver;
+DECLARE_PER_CPU(struct cpuidle_driver, acpi_idle_driver);
 #ifdef CONFIG_ACPI_PROCESSOR_IDLE
 int acpi_processor_power_init(struct acpi_processor *pr);
 int acpi_processor_power_exit(struct acpi_processor *pr);
-- 
2.25.1

Re: [PATCH v2 1/2] cpuidle: Add interface to get cpuidle_driver by CPU ID

[Rafael J. Wysocki]

On Thu, Aug 14, 2025 at 9:32 AM Yaxiong Tian <tianyaxiong@kylinos.cn> wrote:
>
> Some drivers need to obtain the cpuidle_driver via CPU ID before the
> cpuidle_device is registered. Therefore, this interface is added.
>
> Tested-by: Shaobo Huang <huangshaobo2075@phytium.com.cn>
> Signed-off-by: Yaxiong Tian <tianyaxiong@kylinos.cn>
> Signed-off-by: Shaobo Huang <huangshaobo2075@phytium.com.cn>
> Signed-off-by: Yinfeng Wang <wangyinfeng@phytium.com.cn>
> Signed-off-by: Xu Wang<wangxu@phytium.com.cn>
> ---
>  drivers/cpuidle/driver.c | 16 ++++++++++++++++
>  include/linux/cpuidle.h  |  4 ++++
>  2 files changed, 20 insertions(+)
>
> diff --git a/drivers/cpuidle/driver.c b/drivers/cpuidle/driver.c
> index 9bbfa594c442..8647f8165863 100644
> --- a/drivers/cpuidle/driver.c
> +++ b/drivers/cpuidle/driver.c
> @@ -353,6 +353,22 @@ struct cpuidle_driver *cpuidle_get_cpu_driver(struct cpuidle_device *dev)
>  }
>  EXPORT_SYMBOL_GPL(cpuidle_get_cpu_driver);
>
> +/**
> + * cpuidle_get_cpu_driver_by_cpu - return the driver registered for a CPU.
> + * @cpu_num: a valid cpu num
> + *
> + * Returns a struct cpuidle_driver pointer, or NULL if no driver is registered
> + * for the CPU associated with @cpu.
> + */
> +struct cpuidle_driver *cpuidle_get_cpu_driver_by_cpu(int cpu_num)
> +{
> +       if (cpu_num < 0 || cpu_num >= nr_cpu_ids)
> +               return NULL;

What about changing the __cpuidle_get_cpu_driver() parameter type to
unsigned int and also passing unsigned int to this function?  The
first check above can go away then.

> +
> +       return __cpuidle_get_cpu_driver(cpu_num);
> +}
> +EXPORT_SYMBOL_GPL(cpuidle_get_cpu_driver_by_cpu);
> +
>  /**
>   * cpuidle_driver_state_disabled - Disable or enable an idle state
>   * @drv: cpuidle driver owning the state
> diff --git a/include/linux/cpuidle.h b/include/linux/cpuidle.h
> index a9ee4fe55dcf..aebbaae6a501 100644
> --- a/include/linux/cpuidle.h
> +++ b/include/linux/cpuidle.h
> @@ -200,6 +200,7 @@ extern void cpuidle_disable_device(struct cpuidle_device *dev);
>  extern int cpuidle_play_dead(void);
>
>  extern struct cpuidle_driver *cpuidle_get_cpu_driver(struct cpuidle_device *dev);
> +extern struct cpuidle_driver *cpuidle_get_cpu_driver_by_cpu(int cpu_num);
>  static inline struct cpuidle_device *cpuidle_get_device(void)
>  {return __this_cpu_read(cpuidle_devices); }
>  #else
> @@ -240,6 +241,9 @@ static inline void cpuidle_disable_device(struct cpuidle_device *dev) { }
>  static inline int cpuidle_play_dead(void) {return -ENODEV; }
>  static inline struct cpuidle_driver *cpuidle_get_cpu_driver(
>         struct cpuidle_device *dev) {return NULL; }
> +
> +static inline struct cpuidle_driver *cpuidle_get_cpu_driver_by_cpu(
> +       int cpu_num) {return NULL; };
>  static inline struct cpuidle_device *cpuidle_get_device(void) {return NULL; }
>  #endif
>
> --
> 2.25.1
>

Re: [PATCH v2 2/2] ACPI: processor: idle: Replace single idle driver with per-CPU model for better hybrid CPU support

[Rafael J. Wysocki]

On Thu, Aug 14, 2025 at 9:32 AM Yaxiong Tian <tianyaxiong@kylinos.cn> wrote:
>
> Current implementations of hybrid architectures (e.g., ARM64 big.LITTLE
> and Intel Alder Lake) feature CPU cores with different exit latencies.

This is not true for Intel platforms, all of the CPUs in there have
the same set of C-states.

> Using a single driver to describe_LPI states for all core types is
> therefore suboptimal. This is further supported by ACPI specification
> 8.4.4.1 which states: "In a processor hierarchy, each node has its
> own _LPI low-power states specific to that node."
>
> To address these limitations, we replace the monolithic idle driver

It cannot be replaced or you potentially open a Pandora's box of
regressions on old systems in the field.

> with a per-CPU model. This approach enables accurate idle state representation
> for each core type

The per-CPU model can be used instead of the "monolithic idle driver"
only if the platform is actually known to be hybrid.

> Tested-by: Shaobo Huang <huangshaobo2075@phytium.com.cn>
> Signed-off-by: Yaxiong Tian <tianyaxiong@kylinos.cn>
> Signed-off-by: Shaobo Huang <huangshaobo2075@phytium.com.cn>
> Signed-off-by: Yinfeng Wang <wangyinfeng@phytium.com.cn>
> Signed-off-by: Xu Wang<wangxu@phytium.com.cn>

What do all of the above S-o-b mean?  Are these people involved in the
development of the code?  In that case Co-developed-by is also needed.

Thanks!

Re: [PATCH v2 2/2] ACPI: processor: idle: Replace single idle driver with per-CPU model for better hybrid CPU support

[Yaxiong Tian]

在 2025/8/25 22:56, Rafael J. Wysocki 写道:
> On Thu, Aug 14, 2025 at 9:32 AM Yaxiong Tian <tianyaxiong@kylinos.cn> wrote:
>> Current implementations of hybrid architectures (e.g., ARM64 big.LITTLE
>> and Intel Alder Lake) feature CPU cores with different exit latencies.
> This is not true for Intel platforms, all of the CPUs in there have
> the same set of C-states.
Yes,it is.
>
>> Using a single driver to describe_LPI states for all core types is
>> therefore suboptimal. This is further supported by ACPI specification
>> 8.4.4.1 which states: "In a processor hierarchy, each node has its
>> own _LPI low-power states specific to that node."
>>
>> To address these limitations, we replace the monolithic idle driver
> It cannot be replaced or you potentially open a Pandora's box of
> regressions on old systems in the field.
I agree with this point. Regarding the potential risk, does it
make sense to  mitigate it through the following two approaches?

1) Restrict the usage of this multiple driver functionality to
the ARM64 platform only.
2) Use a mutex-protected global variable to indicate whether
the platform utilizes _LPI instead of C-states. This variable
would be determined during initialization via
acpi_has_method(handle, "_LPI"), and the multiple driver
would only be used if the platform employs _LPI.

Alternatively, the other option is to let users implement their
own drivers. Although many ARM64 users still initialize their
systems using the ACPI method rather than the traditional
device tree approach, and they hope to reuse the ACPI idle driver.
>
>> with a per-CPU model. This approach enables accurate idle state representation
>> for each core type
> The per-CPU model can be used instead of the "monolithic idle driver"
> only if the platform is actually known to be hybrid.
Yes,select CPU_IDLE_MULTIPLE_DRIVERS if ARM64  can  resolve this issue.
>
>> Tested-by: Shaobo Huang <huangshaobo2075@phytium.com.cn>
>> Signed-off-by: Yaxiong Tian <tianyaxiong@kylinos.cn>
>> Signed-off-by: Shaobo Huang <huangshaobo2075@phytium.com.cn>
>> Signed-off-by: Yinfeng Wang <wangyinfeng@phytium.com.cn>
>> Signed-off-by: Xu Wang<wangxu@phytium.com.cn>
> What do all of the above S-o-b mean?  Are these people involved in the
> development of the code?  In that case Co-developed-by is also needed.
>
> Thanks!
Thank you for your response. If it makes sense, continue to follow up,
I will address both the issues mentioned above and the problem
in the previous patch.