[PATCH v5 0/3] cpufreq: cppc: Handle Highest Performance changes at runtime

[PATCH v5 0/3] cpufreq: cppc: Handle Highest Performance changes at runtime

[Xueqin Luo]

Hi Rafael, Pierre,

This series adds support for handling ACPI CPPC Highest Performance
register changes at runtime, triggered by Notify(0x85) on a processor
device.

When the platform changes the Highest Performance value (e.g. due to
thermal or power budget adjustments), the OSPM must re-evaluate the
cached capability and propagate the change to the cpufreq policy,
the scheduler's CPU capacity model, and the frequency invariance
engine.

The series is split into three patches:
  Patch 1: Core update_limits callback with boost QoS handling.
  Patch 2: Refactor autonomous perf bounds into a reusable helper
           and wire it into update_limits with error logging.
  Patch 3: Topology subsystem runtime capacity updates with
           validation, early-return optimization, and concurrent-safe
           normalization.

v4 -> v5:
  - Move cpu_data and caps pointer assignments after
    guard(cpufreq_policy_write) in cppc_cpufreq_update_limits() to
    avoid potential use-after-free if concurrent CPU offline frees
    driver_data before the lock is acquired (Sashiko)
  - Extract cppc_cpufreq_sync_boost_limits() from update_limits to
    handle boost_supported re-evaluation, cpuinfo.max_freq tracking,
    and boost_freq_req QoS update in a dedicated helper. The helper
    always updates the QoS request (even when boost becomes
    unsupported) to clear stale constraint values, and forcibly
    disables boost_enabled when boost disappears at runtime
  - Add highest_perf validation in topology_update_cpu_capacity():
    reject values below lowest_perf to prevent corrupted register
    reads from propagating into the scheduler capacity model
  - Add early-return in topology_update_cpu_capacity() when
    raw_capacity[cpu] is unchanged, avoiding redundant normalization
    and schedule_work() calls
  - Add zero capacity_scale guard in topology_update_cpu_capacity()
    to prevent division by zero in the normalization loop
  - Move guard(mutex) before the raw_capacity NULL check so that
    the pointer read is protected by the lock
  - Add CONFIG_GENERIC_ARCH_TOPOLOGY guard with no-op stub in
    arch_topology.h for configs where the topology subsystem is
    disabled

Xueqin Luo (3):
  cpufreq: cppc: Add update_limits support for Highest Performance
    changes
  cpufreq: cppc: Refactor autonomous perf bounds into helper
  arch_topology: Add topology_update_cpu_capacity() for runtime updates

 drivers/base/arch_topology.c   |  76 ++++++++++++++++
 drivers/cpufreq/cppc_cpufreq.c | 157 ++++++++++++++++++++++++++++++---
 include/linux/arch_topology.h  |  13 +++
 3 files changed, 233 insertions(+), 13 deletions(-)

-- 
2.43.0

[PATCH v5 1/3] cpufreq: cppc: Add update_limits support for Highest Performance changes

[Xueqin Luo]

ACPI CPPC specification requires OSPM to re-evaluate the Highest
Performance register when Notify(0x85) is received for a processor
device.

Implement cppc_cpufreq_update_limits() to refresh the cached
highest_perf capability through cppc_get_highest_perf() and update
policy->cpuinfo.max_freq from Highest Performance. Use
refresh_frequency_limits() so policy->max follows the standard
cpufreq_set_policy() path.

cpuinfo.max_freq always tracks Highest Performance. When boost is
disabled but still supported, constrain policy->max by updating the
existing boost_freq_req (created in cpufreq_policy_init_qos) to
nominal, instead of hiding the hardware maximum in cpuinfo.

Signed-off-by: Xueqin Luo <luoxueqin@kylinos.cn>
---
 drivers/cpufreq/cppc_cpufreq.c | 98 ++++++++++++++++++++++++++++++++++
 1 file changed, 98 insertions(+)

diff --git a/drivers/cpufreq/cppc_cpufreq.c b/drivers/cpufreq/cppc_cpufreq.c
index 6fe0e972952a..089f734f851b 100644
--- a/drivers/cpufreq/cppc_cpufreq.c
+++ b/drivers/cpufreq/cppc_cpufreq.c
@@ -855,6 +855,103 @@ static int cppc_cpufreq_set_boost(struct cpufreq_policy *policy, int state)
 	return 0;
 }
 
+/**
+ * cppc_cpufreq_sync_boost_limits - Sync boost flag, cpuinfo max and boost QoS
+ * @policy: cpufreq policy
+ * @boost_supported: whether highest_perf currently exceeds nominal_perf
+ *
+ * Highest Performance can appear or disappear at runtime via Notify(0x85).
+ *
+ * cpuinfo.max_freq always tracks the hardware maximum derived from Highest
+ * Performance so that sysfs reflects Notify(0x85) updates. Boost being off
+ * is enforced by updating the existing boost_freq_req to nominal (capping
+ * policy->max) rather than by hiding the hardware max in cpuinfo.
+ * boost_freq_req itself is only created at policy init, not here.
+ */
+static void cppc_cpufreq_sync_boost_limits(struct cpufreq_policy *policy,
+					   bool boost_supported)
+{
+	struct cppc_cpudata *cpu_data = policy->driver_data;
+	struct cppc_perf_caps *caps = &cpu_data->perf_caps;
+	unsigned int highest_freq, nominal_freq, qos_freq;
+	int ret;
+
+	if (!boost_supported && policy->boost_enabled)
+		policy->boost_enabled = false;
+
+	policy->boost_supported = boost_supported;
+
+	highest_freq = cppc_perf_to_khz(caps, caps->highest_perf);
+	nominal_freq = cppc_perf_to_khz(caps, caps->nominal_perf);
+
+	/*
+	 * Report the current hardware maximum. If Highest dropped below
+	 * Nominal (unusual, but possible with test overrides), never
+	 * advertise more than Highest allows.
+	 */
+	policy->cpuinfo.max_freq = highest_freq;
+
+	if (boost_supported && !policy->boost_enabled)
+		qos_freq = min(nominal_freq, highest_freq);
+	else
+		qos_freq = highest_freq;
+
+	/*
+	 * boost_freq_req is created in cpufreq_policy_init_qos() when
+	 * boost_supported is true at policy init. Runtime Highest changes
+	 * only update that existing request; they do not add or remove it.
+	 */
+	if (freq_qos_request_active(&policy->boost_freq_req)) {
+		ret = freq_qos_update_request(&policy->boost_freq_req,
+					      qos_freq);
+		if (ret < 0)
+			pr_debug("CPU%d: failed to sync boost QoS: %d\n",
+				 policy->cpu, ret);
+	}
+
+	pr_debug("CPU%d: highest_perf=%u boost_en=%d cpuinfo_max=%u qos_max=%u\n",
+		 policy->cpu, caps->highest_perf, policy->boost_enabled,
+		 policy->cpuinfo.max_freq, qos_freq);
+}
+
+static void cppc_cpufreq_update_limits(struct cpufreq_policy *policy)
+{
+	struct cppc_cpudata *cpu_data;
+	struct cppc_perf_caps *caps;
+	u64 prev_highest_perf;
+	u64 highest_perf;
+	int ret;
+
+	guard(cpufreq_policy_write)(policy);
+
+	cpu_data = policy->driver_data;
+	caps = &cpu_data->perf_caps;
+
+	prev_highest_perf = caps->highest_perf;
+
+	ret = cppc_get_highest_perf(policy->cpu, &highest_perf);
+	if (ret)
+		return;
+
+	if (highest_perf == prev_highest_perf)
+		return;
+
+	caps->highest_perf = highest_perf;
+
+	/*
+	 * Re-evaluate boost capability/status based on the updated Highest
+	 * Performance. Boost is supported when highest_perf exceeds
+	 * nominal_perf.
+	 */
+	cppc_cpufreq_sync_boost_limits(policy,
+				       highest_perf > caps->nominal_perf);
+
+	refresh_frequency_limits(policy);
+
+	pr_debug("CPU%d: highest_perf updated %llu -> %llu\n",
+		 policy->cpu, prev_highest_perf, highest_perf);
+}
+
 static ssize_t show_freqdomain_cpus(struct cpufreq_policy *policy, char *buf)
 {
 	struct cppc_cpudata *cpu_data = policy->driver_data;
@@ -1048,6 +1145,7 @@ static struct cpufreq_driver cppc_cpufreq_driver = {
 	.init = cppc_cpufreq_cpu_init,
 	.exit = cppc_cpufreq_cpu_exit,
 	.set_boost = cppc_cpufreq_set_boost,
+	.update_limits = cppc_cpufreq_update_limits,
 	.attr = cppc_cpufreq_attr,
 	.name = "cppc_cpufreq",
 };
-- 
2.43.0

[PATCH v5 2/3] cpufreq: cppc: Refactor autonomous perf bounds into helper

[Xueqin Luo]

Extract the autonomous selection performance bounds programming from
store_auto_select() into a reusable cppc_cpufreq_set_autonomous_perf()
helper. The helper saves and restores MIN/MAX performance on failure,
keeping the rollback logic co-located with the programming sequence.

Reuse the helper in cppc_cpufreq_update_limits() so that when
Highest Performance changes at runtime, the autonomous MIN/MAX
envelope is re-programmed against the updated capability. Add a
pr_debug message on failure for diagnostics.

Signed-off-by: Xueqin Luo <luoxueqin@kylinos.cn>
---
 drivers/cpufreq/cppc_cpufreq.c | 57 ++++++++++++++++++++++++++--------
 1 file changed, 44 insertions(+), 13 deletions(-)

diff --git a/drivers/cpufreq/cppc_cpufreq.c b/drivers/cpufreq/cppc_cpufreq.c
index 089f734f851b..09d7745a609f 100644
--- a/drivers/cpufreq/cppc_cpufreq.c
+++ b/drivers/cpufreq/cppc_cpufreq.c
@@ -868,6 +868,37 @@ static int cppc_cpufreq_set_boost(struct cpufreq_policy *policy, int state)
  * policy->max) rather than by hiding the hardware max in cpuinfo.
  * boost_freq_req itself is only created at policy init, not here.
  */
+
+/**
+ * cppc_cpufreq_set_autonomous_perf - Configure performance bounds for
+ *                                     autonomous mode
+ * @policy: cpufreq policy structure
+ *
+ * When autonomous selection is enabled, program MIN_PERF and MAX_PERF
+ * from current policy limits so that the platform uses the correct
+ * performance bounds immediately.
+ *
+ * Return: 0 on success, negative error code on failure.
+ */
+static int cppc_cpufreq_set_autonomous_perf(struct cpufreq_policy *policy)
+{
+	struct cppc_cpudata *cpu_data = policy->driver_data;
+	u32 old_min_perf = cpu_data->perf_ctrls.min_perf;
+	u32 old_max_perf = cpu_data->perf_ctrls.max_perf;
+	int ret;
+
+	cppc_cpufreq_update_perf_limits(cpu_data, policy);
+
+	ret = cppc_set_perf(policy->cpu, &cpu_data->perf_ctrls);
+	if (ret) {
+		cpu_data->perf_ctrls.min_perf = old_min_perf;
+		cpu_data->perf_ctrls.max_perf = old_max_perf;
+		return ret;
+	}
+
+	return 0;
+}
+
 static void cppc_cpufreq_sync_boost_limits(struct cpufreq_policy *policy,
 					   bool boost_supported)
 {
@@ -948,6 +979,18 @@ static void cppc_cpufreq_update_limits(struct cpufreq_policy *policy)
 
 	refresh_frequency_limits(policy);
 
+	/*
+	 * Autonomous selection mode uses MIN/MAX performance as runtime
+	 * hardware control bounds. Re-program them when highest_perf
+	 * changes so that the platform uses the updated bounds.
+	 */
+	if (cpu_data->perf_ctrls.auto_sel) {
+		ret = cppc_cpufreq_set_autonomous_perf(policy);
+		if (ret)
+			pr_debug("CPU%d: failed to update autonomous perf: %d\n",
+				 policy->cpu, ret);
+	}
+
 	pr_debug("CPU%d: highest_perf updated %llu -> %llu\n",
 		 policy->cpu, prev_highest_perf, highest_perf);
 }
@@ -994,20 +1037,8 @@ static ssize_t store_auto_select(struct cpufreq_policy *policy,
 	cpu_data->perf_ctrls.auto_sel = val;
 
 	if (val) {
-		u32 old_min_perf = cpu_data->perf_ctrls.min_perf;
-		u32 old_max_perf = cpu_data->perf_ctrls.max_perf;
-
-		/*
-		 * When enabling autonomous selection, program MIN_PERF and
-		 * MAX_PERF from current policy limits so that the platform
-		 * uses the correct performance bounds immediately.
-		 */
-		cppc_cpufreq_update_perf_limits(cpu_data, policy);
-
-		ret = cppc_set_perf(policy->cpu, &cpu_data->perf_ctrls);
+		ret = cppc_cpufreq_set_autonomous_perf(policy);
 		if (ret) {
-			cpu_data->perf_ctrls.min_perf = old_min_perf;
-			cpu_data->perf_ctrls.max_perf = old_max_perf;
 			cppc_set_auto_sel(policy->cpu, false);
 			cpu_data->perf_ctrls.auto_sel = false;
 			return ret;
-- 
2.43.0

[PATCH v5 3/3] arch_topology: Add topology_update_cpu_capacity() for runtime updates

[Xueqin Luo]

When the CPPC Highest Performance register changes at runtime
(e.g. via ACPI Notify(0x85)), the scheduler's view of CPU capacity
and the frequency invariance engine's reference values become stale,
as topology_init_cpu_capacity_cppc() is only called once during boot.

Keep raw_capacity allocated after CPPC init instead of freeing it,
and introduce topology_update_cpu_capacity() to update per-CPU
raw_capacity, capacity_freq_ref, and the normalized CPU capacity
scale at runtime. Provide a no-op stub when GENERIC_ARCH_TOPOLOGY
is disabled so cppc_cpufreq can link on those configs. Skip updates
when the value is unchanged and reject a zero capacity_scale to
avoid division by zero.

Call this from cppc_cpufreq_update_limits() for every CPU in the
policy so shared-policy Notify(0x85) targeting a non-policy CPU
still refreshes the correct topology capacity.

Signed-off-by: Xueqin Luo <luoxueqin@kylinos.cn>
---
 drivers/base/arch_topology.c   | 76 ++++++++++++++++++++++++++++++++++
 drivers/cpufreq/cppc_cpufreq.c |  2 +
 include/linux/arch_topology.h  | 13 ++++++
 3 files changed, 91 insertions(+)

diff --git a/drivers/base/arch_topology.c b/drivers/base/arch_topology.c
index 8c5e47c28d9a..27f2bfa9f326 100644
--- a/drivers/base/arch_topology.c
+++ b/drivers/base/arch_topology.c
@@ -229,6 +229,7 @@ static void update_topology_flags_workfn(struct work_struct *work)
 }
 
 static u32 *raw_capacity;
+static DEFINE_MUTEX(raw_capacity_lock);
 
 static int free_raw_capacity(void)
 {
@@ -372,13 +373,88 @@ static inline void topology_init_cpu_capacity_cppc(void)
 	schedule_work(&update_topology_flags_work);
 	pr_debug("cpu_capacity: cpu_capacity initialization done\n");
 
+	/*
+	 * Keep raw_capacity for runtime updates via
+	 * topology_update_cpu_capacity().
+	 */
+	return;
+
 exit:
 	free_raw_capacity();
 }
+
 void acpi_processor_init_invariance_cppc(void)
 {
 	topology_init_cpu_capacity_cppc();
 }
+
+/**
+ * topology_update_cpu_capacity - Update CPU capacity after highest_perf change
+ * @cpu: CPU whose highest performance changed
+ * @perf_caps: Updated CPPC performance capabilities for @cpu
+ *
+ * When the CPPC Highest Performance register changes at runtime
+ * (e.g. via Notify(0x85)), the scheduler's view of CPU capacity
+ * and the frequency invariance engine's reference values become
+ * stale. This function updates the per-CPU raw_capacity,
+ * capacity_freq_ref and freq_inv max ratio, then re-normalizes the
+ * CPU capacity scale for all possible CPUs and triggers a sched
+ * domain rebuild. If the value is unchanged, everything is skipped.
+ */
+void topology_update_cpu_capacity(unsigned int cpu,
+				  struct cppc_perf_caps *perf_caps)
+{
+	u32 highest_perf = perf_caps->highest_perf;
+	u64 capacity, capacity_scale = 0;
+	int c;
+
+	guard(mutex)(&raw_capacity_lock);
+
+	if (!raw_capacity || cpu >= num_possible_cpus())
+		return;
+
+	/*
+	 * Validate: highest_perf must be >= nominal_perf and >= lowest_perf,
+	 * consistent with the boot-time check in topology_init_cpu_capacity_cppc().
+	 */
+	if (highest_perf < perf_caps->lowest_perf) {
+		pr_warn("cpu_capacity: CPU%d invalid highest_perf=%u (nominal=%u, lowest=%u), skipping\n",
+			cpu, highest_perf, perf_caps->nominal_perf,
+			perf_caps->lowest_perf);
+		return;
+	}
+
+	if (raw_capacity[cpu] == highest_perf)
+		return;
+
+	pr_debug("cpu_capacity: CPU%d cpu_capacity=%u -> %u (raw)\n",
+		 cpu, raw_capacity[cpu], highest_perf);
+
+	raw_capacity[cpu] = highest_perf;
+	per_cpu(capacity_freq_ref, cpu) =
+		cppc_perf_to_khz(perf_caps, highest_perf);
+	freq_inv_set_max_ratio(cpu,
+			       per_cpu(capacity_freq_ref, cpu) * HZ_PER_KHZ);
+
+	/* Re-normalize all CPUs: capacity is relative. */
+	for_each_possible_cpu(c)
+		capacity_scale = max_t(u64, capacity_scale, raw_capacity[c]);
+
+	if (!capacity_scale)
+		return;
+
+	for_each_possible_cpu(c) {
+		capacity = raw_capacity[c];
+		capacity = div64_u64(capacity << SCHED_CAPACITY_SHIFT,
+				     capacity_scale);
+		topology_set_cpu_scale(c, capacity);
+		pr_debug("cpu_capacity: CPU%d cpu_capacity=%lu\n",
+			 c, topology_get_cpu_scale(c));
+	}
+
+	schedule_work(&update_topology_flags_work);
+}
+EXPORT_SYMBOL_GPL(topology_update_cpu_capacity);
 #endif
 
 #ifdef CONFIG_CPU_FREQ
diff --git a/drivers/cpufreq/cppc_cpufreq.c b/drivers/cpufreq/cppc_cpufreq.c
index 09d7745a609f..1480be537eaa 100644
--- a/drivers/cpufreq/cppc_cpufreq.c
+++ b/drivers/cpufreq/cppc_cpufreq.c
@@ -979,6 +979,8 @@ static void cppc_cpufreq_update_limits(struct cpufreq_policy *policy)
 
 	refresh_frequency_limits(policy);
 
+	topology_update_cpu_capacity(policy->cpu, caps);
+
 	/*
 	 * Autonomous selection mode uses MIN/MAX performance as runtime
 	 * hardware control bounds. Re-program them when highest_perf
diff --git a/include/linux/arch_topology.h b/include/linux/arch_topology.h
index ebd7f8935f96..9415cb6a6c2b 100644
--- a/include/linux/arch_topology.h
+++ b/include/linux/arch_topology.h
@@ -11,6 +11,19 @@
 void topology_normalize_cpu_scale(void);
 int topology_update_cpu_topology(void);
 
+#ifdef CONFIG_ACPI_CPPC_LIB
+struct cppc_perf_caps;
+#ifdef CONFIG_GENERIC_ARCH_TOPOLOGY
+void topology_update_cpu_capacity(unsigned int cpu,
+				  struct cppc_perf_caps *perf_caps);
+#else
+static inline void
+topology_update_cpu_capacity(unsigned int cpu, struct cppc_perf_caps *perf_caps)
+{
+}
+#endif
+#endif
+
 struct device_node;
 bool topology_parse_cpu_capacity(struct device_node *cpu_node, int cpu);
 
-- 
2.43.0

Re: [PATCH v5 3/3] arch_topology: Add topology_update_cpu_capacity() for runtime updates

[Christian Loehle]

On 8/7/26 07:08, Xueqin Luo wrote:
> When the CPPC Highest Performance register changes at runtime
> (e.g. via ACPI Notify(0x85)), the scheduler's view of CPU capacity
> and the frequency invariance engine's reference values become stale,
> as topology_init_cpu_capacity_cppc() is only called once during boot.
> 
> Keep raw_capacity allocated after CPPC init instead of freeing it,
> and introduce topology_update_cpu_capacity() to update per-CPU
> raw_capacity, capacity_freq_ref, and the normalized CPU capacity
> scale at runtime. Provide a no-op stub when GENERIC_ARCH_TOPOLOGY
> is disabled so cppc_cpufreq can link on those configs. Skip updates
> when the value is unchanged and reject a zero capacity_scale to
> avoid division by zero.
> 
> Call this from cppc_cpufreq_update_limits() for every CPU in the
> policy so shared-policy Notify(0x85) targeting a non-policy CPU
> still refreshes the correct topology capacity.
> 
> Signed-off-by: Xueqin Luo <luoxueqin@kylinos.cn>
> ---
>  drivers/base/arch_topology.c   | 76 ++++++++++++++++++++++++++++++++++
>  drivers/cpufreq/cppc_cpufreq.c |  2 +
>  include/linux/arch_topology.h  | 13 ++++++
>  3 files changed, 91 insertions(+)
> 
> diff --git a/drivers/base/arch_topology.c b/drivers/base/arch_topology.c
> index 8c5e47c28d9a..27f2bfa9f326 100644
> --- a/drivers/base/arch_topology.c
> +++ b/drivers/base/arch_topology.c
> @@ -229,6 +229,7 @@ static void update_topology_flags_workfn(struct work_struct *work)
>  }
>  
>  static u32 *raw_capacity;
> +static DEFINE_MUTEX(raw_capacity_lock);
>  
>  static int free_raw_capacity(void)
>  {
> @@ -372,13 +373,88 @@ static inline void topology_init_cpu_capacity_cppc(void)
>  	schedule_work(&update_topology_flags_work);
>  	pr_debug("cpu_capacity: cpu_capacity initialization done\n");
>  
> +	/*
> +	 * Keep raw_capacity for runtime updates via
> +	 * topology_update_cpu_capacity().
> +	 */
> +	return;
> +
>  exit:
>  	free_raw_capacity();
>  }
> +
>  void acpi_processor_init_invariance_cppc(void)
>  {
>  	topology_init_cpu_capacity_cppc();
>  }
> +
> +/**
> + * topology_update_cpu_capacity - Update CPU capacity after highest_perf change
> + * @cpu: CPU whose highest performance changed
> + * @perf_caps: Updated CPPC performance capabilities for @cpu
> + *
> + * When the CPPC Highest Performance register changes at runtime
> + * (e.g. via Notify(0x85)), the scheduler's view of CPU capacity
> + * and the frequency invariance engine's reference values become
> + * stale. This function updates the per-CPU raw_capacity,
> + * capacity_freq_ref and freq_inv max ratio, then re-normalizes the
> + * CPU capacity scale for all possible CPUs and triggers a sched
> + * domain rebuild. If the value is unchanged, everything is skipped.
> + */
> +void topology_update_cpu_capacity(unsigned int cpu,
> +				  struct cppc_perf_caps *perf_caps)
> +{
> +	u32 highest_perf = perf_caps->highest_perf;
> +	u64 capacity, capacity_scale = 0;
> +	int c;
> +
> +	guard(mutex)(&raw_capacity_lock);
> +
> +	if (!raw_capacity || cpu >= num_possible_cpus())
> +		return;
> +
> +	/*
> +	 * Validate: highest_perf must be >= nominal_perf and >= lowest_perf,
> +	 * consistent with the boot-time check in topology_init_cpu_capacity_cppc().
> +	 */
> +	if (highest_perf < perf_caps->lowest_perf) {
> +		pr_warn("cpu_capacity: CPU%d invalid highest_perf=%u (nominal=%u, lowest=%u), skipping\n",
> +			cpu, highest_perf, perf_caps->nominal_perf,
> +			perf_caps->lowest_perf);
> +		return;
> +	}
> +
> +	if (raw_capacity[cpu] == highest_perf)
> +		return;
> +
> +	pr_debug("cpu_capacity: CPU%d cpu_capacity=%u -> %u (raw)\n",
> +		 cpu, raw_capacity[cpu], highest_perf);
> +
> +	raw_capacity[cpu] = highest_perf;

Does this actually work if highest_perf would now be the equivalent for >1024?

> +	per_cpu(capacity_freq_ref, cpu) =
> +		cppc_perf_to_khz(perf_caps, highest_perf);
> +	freq_inv_set_max_ratio(cpu,
> +			       per_cpu(capacity_freq_ref, cpu) * HZ_PER_KHZ);
> +
> +	/* Re-normalize all CPUs: capacity is relative. */
> +	for_each_possible_cpu(c)
> +		capacity_scale = max_t(u64, capacity_scale, raw_capacity[c]);
> +
> +	if (!capacity_scale)
> +		return;
> +
> +	for_each_possible_cpu(c) {
> +		capacity = raw_capacity[c];
> +		capacity = div64_u64(capacity << SCHED_CAPACITY_SHIFT,
> +				     capacity_scale);
> +		topology_set_cpu_scale(c, capacity);
> +		pr_debug("cpu_capacity: CPU%d cpu_capacity=%lu\n",
> +			 c, topology_get_cpu_scale(c));
> +	}
> +
> +	schedule_work(&update_topology_flags_work);
> +}
> +EXPORT_SYMBOL_GPL(topology_update_cpu_capacity);
>  #endif
>  
>  #ifdef CONFIG_CPU_FREQ
> diff --git a/drivers/cpufreq/cppc_cpufreq.c b/drivers/cpufreq/cppc_cpufreq.c
> index 09d7745a609f..1480be537eaa 100644
> --- a/drivers/cpufreq/cppc_cpufreq.c
> +++ b/drivers/cpufreq/cppc_cpufreq.c
> @@ -979,6 +979,8 @@ static void cppc_cpufreq_update_limits(struct cpufreq_policy *policy)
>  
>  	refresh_frequency_limits(policy);
>  
> +	topology_update_cpu_capacity(policy->cpu, caps);
> +
>  	/*
>  	 * Autonomous selection mode uses MIN/MAX performance as runtime
>  	 * hardware control bounds. Re-program them when highest_perf
> diff --git a/include/linux/arch_topology.h b/include/linux/arch_topology.h
> index ebd7f8935f96..9415cb6a6c2b 100644
> --- a/include/linux/arch_topology.h
> +++ b/include/linux/arch_topology.h
> @@ -11,6 +11,19 @@
>  void topology_normalize_cpu_scale(void);
>  int topology_update_cpu_topology(void);
>  
> +#ifdef CONFIG_ACPI_CPPC_LIB
> +struct cppc_perf_caps;
> +#ifdef CONFIG_GENERIC_ARCH_TOPOLOGY
> +void topology_update_cpu_capacity(unsigned int cpu,
> +				  struct cppc_perf_caps *perf_caps);
> +#else
> +static inline void
> +topology_update_cpu_capacity(unsigned int cpu, struct cppc_perf_caps *perf_caps)
> +{
> +}
> +#endif
> +#endif
> +
>  struct device_node;
>  bool topology_parse_cpu_capacity(struct device_node *cpu_node, int cpu);
>