[PATCH v3 0/4] cpufreq: Fixes and cleanups related to CPUFREQ_ETERNAL

[PATCH v3 0/4] cpufreq: Fixes and cleanups related to CPUFREQ_ETERNAL

[Rafael J. Wysocki]

Hi All,

This is an update of

https://lore.kernel.org/linux-pm/8605612.T7Z3S40VBb@rafael.j.wysocki/

The update is due because of some changes in the first patch (involving Rust)
and in the second patch (function rename).  It is v3 because a v2 of the
third patch was sent after the v1.

The original cover letter quoted below still applies:

 The first patch in this series is meant to address the failure discussed in

 https://lore.kernel.org/linux-pm/20250922125929.453444-1-shawnguo2@yeah.net/

 but in a different way than proposed by Shawn.

 The second one is a CPPC cpufreq driver fix preventing it from using an
 overly large transition delay in the cases when that delay cannot be
 obtained from the platform firmware.

 Patch [3/4] makes CPPC use a specific symbol instead of CPUFREQ_ETERNAL for
 signaling error conditions while attempting to retrieve a transition latency
 value from the platform firmware.

 The last patch removes CPUFREQ_ETERNAL (which has no users any more) from
 cpufreq, including all references to it in cpufreq documentation.

Thanks!

[PATCH v3 1/4] cpufreq: Make drivers using CPUFREQ_ETERNAL specify transition latency

[Rafael J. Wysocki]

From: Rafael J. Wysocki <rafael.j.wysocki@intel.com>

Commit a755d0e2d41b ("cpufreq: Honour transition_latency over
transition_delay_us") caused platforms where cpuinfo.transition_latency
is CPUFREQ_ETERNAL to get a very large transition latency whereas
previously it had been capped at 10 ms (and later at 2 ms).

This led to a user-observable regression between 6.6 and 6.12 as
described by Shawn:

"The dbs sampling_rate was 10000 us on 6.6 and suddently becomes
 6442450 us (4294967295 / 1000 * 1.5) on 6.12 for these platforms
 because the default transition delay was dropped [...].

 It slows down dbs governor's reacting to CPU loading change
 dramatically.  Also, as transition_delay_us is used by schedutil
 governor as rate_limit_us, it shows a negative impact on device
 idle power consumption, because the device gets slightly less time
 in the lowest OPP."

Evidently, the expectation of the drivers using CPUFREQ_ETERNAL as
cpuinfo.transition_latency was that it would be capped by the core,
but they may as well return a default transition latency value instead
of CPUFREQ_ETERNAL and the core need not do anything with it.

Accordingly, introduce CPUFREQ_DEFAULT_TRANSITION_LATENCY_NS and make
all of the drivers in question use it instead of CPUFREQ_ETERNAL.  Also
update the related Rust binding.

Fixes: a755d0e2d41b ("cpufreq: Honour transition_latency over transition_delay_us")
Closes: https://lore.kernel.org/linux-pm/20250922125929.453444-1-shawnguo2@yeah.net/
Reported-by: Shawn Guo <shawnguo@kernel.org>
Reviewed-by: Mario Limonciello (AMD) <superm1@kernel.org>
Reviewed-by: Jie Zhan <zhanjie9@hisilicon.com>
Cc: 6.6+ <stable@vger.kernel.org> # 6.6+
Signed-off-by: Rafael J. Wysocki <rafael.j.wysocki@intel.com>
---

v1 -> v3:
   * Add updates of the Rust version of cpufreq-dt and Rust binding
   * Update the changelog
   * Add tags from Mario Limonciello and Jie Zhan

---
 drivers/cpufreq/cpufreq-dt.c          |    2 +-
 drivers/cpufreq/imx6q-cpufreq.c       |    2 +-
 drivers/cpufreq/mediatek-cpufreq-hw.c |    2 +-
 drivers/cpufreq/rcpufreq_dt.rs        |    2 +-
 drivers/cpufreq/scmi-cpufreq.c        |    2 +-
 drivers/cpufreq/scpi-cpufreq.c        |    2 +-
 drivers/cpufreq/spear-cpufreq.c       |    2 +-
 include/linux/cpufreq.h               |    3 +++
 rust/kernel/cpufreq.rs                |    7 ++++---
 9 files changed, 14 insertions(+), 10 deletions(-)

--- a/drivers/cpufreq/cpufreq-dt.c
+++ b/drivers/cpufreq/cpufreq-dt.c
@@ -104,7 +104,7 @@ static int cpufreq_init(struct cpufreq_p
 
 	transition_latency = dev_pm_opp_get_max_transition_latency(cpu_dev);
 	if (!transition_latency)
-		transition_latency = CPUFREQ_ETERNAL;
+		transition_latency = CPUFREQ_DEFAULT_TRANSITION_LATENCY_NS;
 
 	cpumask_copy(policy->cpus, priv->cpus);
 	policy->driver_data = priv;
--- a/drivers/cpufreq/imx6q-cpufreq.c
+++ b/drivers/cpufreq/imx6q-cpufreq.c
@@ -442,7 +442,7 @@ soc_opp_out:
 	}
 
 	if (of_property_read_u32(np, "clock-latency", &transition_latency))
-		transition_latency = CPUFREQ_ETERNAL;
+		transition_latency = CPUFREQ_DEFAULT_TRANSITION_LATENCY_NS;
 
 	/*
 	 * Calculate the ramp time for max voltage change in the
--- a/drivers/cpufreq/mediatek-cpufreq-hw.c
+++ b/drivers/cpufreq/mediatek-cpufreq-hw.c
@@ -309,7 +309,7 @@ static int mtk_cpufreq_hw_cpu_init(struc
 
 	latency = readl_relaxed(data->reg_bases[REG_FREQ_LATENCY]) * 1000;
 	if (!latency)
-		latency = CPUFREQ_ETERNAL;
+		latency = CPUFREQ_DEFAULT_TRANSITION_LATENCY_NS;
 
 	policy->cpuinfo.transition_latency = latency;
 	policy->fast_switch_possible = true;
--- a/drivers/cpufreq/rcpufreq_dt.rs
+++ b/drivers/cpufreq/rcpufreq_dt.rs
@@ -123,7 +123,7 @@ impl cpufreq::Driver for CPUFreqDTDriver
 
         let mut transition_latency = opp_table.max_transition_latency_ns() as u32;
         if transition_latency == 0 {
-            transition_latency = cpufreq::ETERNAL_LATENCY_NS;
+            transition_latency = cpufreq::DEFAULT_TRANSITION_LATENCY_NS;
         }
 
         policy
--- a/drivers/cpufreq/scmi-cpufreq.c
+++ b/drivers/cpufreq/scmi-cpufreq.c
@@ -294,7 +294,7 @@ static int scmi_cpufreq_init(struct cpuf
 
 	latency = perf_ops->transition_latency_get(ph, domain);
 	if (!latency)
-		latency = CPUFREQ_ETERNAL;
+		latency = CPUFREQ_DEFAULT_TRANSITION_LATENCY_NS;
 
 	policy->cpuinfo.transition_latency = latency;
 
--- a/drivers/cpufreq/scpi-cpufreq.c
+++ b/drivers/cpufreq/scpi-cpufreq.c
@@ -157,7 +157,7 @@ static int scpi_cpufreq_init(struct cpuf
 
 	latency = scpi_ops->get_transition_latency(cpu_dev);
 	if (!latency)
-		latency = CPUFREQ_ETERNAL;
+		latency = CPUFREQ_DEFAULT_TRANSITION_LATENCY_NS;
 
 	policy->cpuinfo.transition_latency = latency;
 
--- a/drivers/cpufreq/spear-cpufreq.c
+++ b/drivers/cpufreq/spear-cpufreq.c
@@ -182,7 +182,7 @@ static int spear_cpufreq_probe(struct pl
 
 	if (of_property_read_u32(np, "clock-latency",
 				&spear_cpufreq.transition_latency))
-		spear_cpufreq.transition_latency = CPUFREQ_ETERNAL;
+		spear_cpufreq.transition_latency = CPUFREQ_DEFAULT_TRANSITION_LATENCY_NS;
 
 	cnt = of_property_count_u32_elems(np, "cpufreq_tbl");
 	if (cnt <= 0) {
--- a/include/linux/cpufreq.h
+++ b/include/linux/cpufreq.h
@@ -32,6 +32,9 @@
  */
 
 #define CPUFREQ_ETERNAL			(-1)
+
+#define CPUFREQ_DEFAULT_TANSITION_LATENCY_NS	NSEC_PER_MSEC
+
 #define CPUFREQ_NAME_LEN		16
 /* Print length for names. Extra 1 space for accommodating '\n' in prints */
 #define CPUFREQ_NAME_PLEN		(CPUFREQ_NAME_LEN + 1)
--- a/rust/kernel/cpufreq.rs
+++ b/rust/kernel/cpufreq.rs
@@ -39,7 +39,8 @@ use macros::vtable;
 const CPUFREQ_NAME_LEN: usize = bindings::CPUFREQ_NAME_LEN as usize;
 
 /// Default transition latency value in nanoseconds.
-pub const ETERNAL_LATENCY_NS: u32 = bindings::CPUFREQ_ETERNAL as u32;
+pub const DEFAULT_TRANSITION_LATENCY_NS: u32 =
+        bindings::CPUFREQ_DEFAULT_TRANSITION_LATENCY_NS as u32;
 
 /// CPU frequency driver flags.
 pub mod flags {
@@ -400,13 +401,13 @@ impl TableBuilder {
 /// The following example demonstrates how to create a CPU frequency table.
 ///
 /// ```
-/// use kernel::cpufreq::{ETERNAL_LATENCY_NS, Policy};
+/// use kernel::cpufreq::{DEFAULT_TRANSITION_LATENCY_NS, Policy};
 ///
 /// fn update_policy(policy: &mut Policy) {
 ///     policy
 ///         .set_dvfs_possible_from_any_cpu(true)
 ///         .set_fast_switch_possible(true)
-///         .set_transition_latency_ns(ETERNAL_LATENCY_NS);
+///         .set_transition_latency_ns(DEFAULT_TRANSITION_LATENCY_NS);
 ///
 ///     pr_info!("The policy details are: {:?}\n", (policy.cpu(), policy.cur()));
 /// }

[PATCH v3 2/4] cpufreq: CPPC: Avoid using CPUFREQ_ETERNAL as transition delay

[Rafael J. Wysocki]

From: Rafael J. Wysocki <rafael.j.wysocki@intel.com>

If cppc_get_transition_latency() returns CPUFREQ_ETERNAL to indicate a
failure to retrieve the transition latency value from the platform
firmware, the CPPC cpufreq driver will use that value (converted to
microseconds) as the policy transition delay, but it is way too large
for any practical use.

Address this by making the driver use the cpufreq's default
transition latency value (in microseconds) as the transition delay
if CPUFREQ_ETERNAL is returned by cppc_get_transition_latency().

Fixes: d4f3388afd48 ("cpufreq / CPPC: Set platform specific transition_delay_us")
Cc: 5.19+ <stable@vger.kernel.org> # 5.19
Signed-off-by: Rafael J. Wysocki <rafael.j.wysocki@intel.com>
Reviewed-by: Mario Limonciello (AMD) <superm1@kernel.org>
---

v1 -> v3:
   * Change the name of the new function (Jie Zhan)
   * Add a tag from Mario Limonciello

---
 drivers/cpufreq/cppc_cpufreq.c |   14 ++++++++++++--
 1 file changed, 12 insertions(+), 2 deletions(-)

--- a/drivers/cpufreq/cppc_cpufreq.c
+++ b/drivers/cpufreq/cppc_cpufreq.c
@@ -308,6 +308,16 @@ static int cppc_verify_policy(struct cpu
 	return 0;
 }
 
+static unsigned int __cppc_cpufreq_get_transition_delay_us(unsigned int cpu)
+{
+	unsigned int transition_latency_ns = cppc_get_transition_latency(cpu);
+
+	if (transition_latency_ns == CPUFREQ_ETERNAL)
+		return CPUFREQ_DEFAULT_TRANSITION_LATENCY_NS / NSEC_PER_USEC;
+
+	return transition_latency_ns / NSEC_PER_USEC;
+}
+
 /*
  * The PCC subspace describes the rate at which platform can accept commands
  * on the shared PCC channel (including READs which do not count towards freq
@@ -330,12 +340,12 @@ static unsigned int cppc_cpufreq_get_tra
 			return 10000;
 		}
 	}
-	return cppc_get_transition_latency(cpu) / NSEC_PER_USEC;
+	return __cppc_cpufreq_get_transition_delay_us(cpu);
 }
 #else
 static unsigned int cppc_cpufreq_get_transition_delay_us(unsigned int cpu)
 {
-	return cppc_get_transition_latency(cpu) / NSEC_PER_USEC;
+	return __cppc_cpufreq_get_transition_delay_us(cpu);
 }
 #endif
 

[PATCH v3 3/4] ACPI: CPPC: Do not use CPUFREQ_ETERNAL as an error value

[Rafael J. Wysocki]

From: Rafael J. Wysocki <rafael.j.wysocki@intel.com>

Instead of using CPUFREQ_ETERNAL for signaling an error condition
in cppc_get_transition_latency(), change the return value type of
that function to int and make it return a proper negative error
code on failures.

No intentional functional impact.

Reviewed-by: Mario Limonciello (AMD) <superm1@kernel.org>
Reviewed-by: Jie Zhan <zhanjie9@hisilicon.com>
Signed-off-by: Rafael J. Wysocki <rafael.j.wysocki@intel.com>
---

v2 -> v3:
   * Combine two if () statements in cppc_get_transition_latency() (Jie Zhan)
   * Add tags from Mario Limonciello and Jie Zhan

v1 -> v2:
   * Change cppc_get_transition_latency() return value data type to int
   * Make it return -ENODATA on errors (Mario)
   * Update its callers accordingly
   * Adjust the subject and changelog
   * Add a missing empty code line to cppc_get_transition_latency()

---
 drivers/acpi/cppc_acpi.c       |   16 +++++++---------
 drivers/cpufreq/amd-pstate.c   |    8 ++++----
 drivers/cpufreq/cppc_cpufreq.c |    4 ++--
 include/acpi/cppc_acpi.h       |    6 +++---
 4 files changed, 16 insertions(+), 18 deletions(-)

--- a/drivers/acpi/cppc_acpi.c
+++ b/drivers/acpi/cppc_acpi.c
@@ -1876,7 +1876,7 @@ EXPORT_SYMBOL_GPL(cppc_set_perf);
  * If desired_reg is in the SystemMemory or SystemIo ACPI address space,
  * then assume there is no latency.
  */
-unsigned int cppc_get_transition_latency(int cpu_num)
+int cppc_get_transition_latency(int cpu_num)
 {
 	/*
 	 * Expected transition latency is based on the PCCT timing values
@@ -1889,31 +1889,29 @@ unsigned int cppc_get_transition_latency
 	 *              completion of a command before issuing the next command,
 	 *              in microseconds.
 	 */
-	unsigned int latency_ns = 0;
 	struct cpc_desc *cpc_desc;
 	struct cpc_register_resource *desired_reg;
 	int pcc_ss_id = per_cpu(cpu_pcc_subspace_idx, cpu_num);
 	struct cppc_pcc_data *pcc_ss_data;
+	int latency_ns = 0;
 
 	cpc_desc = per_cpu(cpc_desc_ptr, cpu_num);
 	if (!cpc_desc)
-		return CPUFREQ_ETERNAL;
+		return -ENODATA;
 
 	desired_reg = &cpc_desc->cpc_regs[DESIRED_PERF];
 	if (CPC_IN_SYSTEM_MEMORY(desired_reg) || CPC_IN_SYSTEM_IO(desired_reg))
 		return 0;
-	else if (!CPC_IN_PCC(desired_reg))
-		return CPUFREQ_ETERNAL;
 
-	if (pcc_ss_id < 0)
-		return CPUFREQ_ETERNAL;
+	if (!CPC_IN_PCC(desired_reg) || pcc_ss_id < 0)
+		return -ENODATA;
 
 	pcc_ss_data = pcc_data[pcc_ss_id];
 	if (pcc_ss_data->pcc_mpar)
 		latency_ns = 60 * (1000 * 1000 * 1000 / pcc_ss_data->pcc_mpar);
 
-	latency_ns = max(latency_ns, pcc_ss_data->pcc_nominal * 1000);
-	latency_ns = max(latency_ns, pcc_ss_data->pcc_mrtt * 1000);
+	latency_ns = max_t(int, latency_ns, pcc_ss_data->pcc_nominal * 1000);
+	latency_ns = max_t(int, latency_ns, pcc_ss_data->pcc_mrtt * 1000);
 
 	return latency_ns;
 }
--- a/drivers/cpufreq/amd-pstate.c
+++ b/drivers/cpufreq/amd-pstate.c
@@ -872,10 +872,10 @@ static void amd_pstate_update_limits(str
  */
 static u32 amd_pstate_get_transition_delay_us(unsigned int cpu)
 {
-	u32 transition_delay_ns;
+	int transition_delay_ns;
 
 	transition_delay_ns = cppc_get_transition_latency(cpu);
-	if (transition_delay_ns == CPUFREQ_ETERNAL) {
+	if (transition_delay_ns < 0) {
 		if (cpu_feature_enabled(X86_FEATURE_AMD_FAST_CPPC))
 			return AMD_PSTATE_FAST_CPPC_TRANSITION_DELAY;
 		else
@@ -891,10 +891,10 @@ static u32 amd_pstate_get_transition_del
  */
 static u32 amd_pstate_get_transition_latency(unsigned int cpu)
 {
-	u32 transition_latency;
+	int transition_latency;
 
 	transition_latency = cppc_get_transition_latency(cpu);
-	if (transition_latency  == CPUFREQ_ETERNAL)
+	if (transition_latency < 0)
 		return AMD_PSTATE_TRANSITION_LATENCY;
 
 	return transition_latency;
--- a/drivers/cpufreq/cppc_cpufreq.c
+++ b/drivers/cpufreq/cppc_cpufreq.c
@@ -310,9 +310,9 @@ static int cppc_verify_policy(struct cpu
 
 static unsigned int __cppc_cpufreq_get_transition_delay_us(unsigned int cpu)
 {
-	unsigned int transition_latency_ns = cppc_get_transition_latency(cpu);
+	int transition_latency_ns = cppc_get_transition_latency(cpu);
 
-	if (transition_latency_ns == CPUFREQ_ETERNAL)
+	if (transition_latency_ns < 0)
 		return CPUFREQ_DEFAULT_TRANSITION_LATENCY_NS / NSEC_PER_USEC;
 
 	return transition_latency_ns / NSEC_PER_USEC;
--- a/include/acpi/cppc_acpi.h
+++ b/include/acpi/cppc_acpi.h
@@ -160,7 +160,7 @@ extern unsigned int cppc_khz_to_perf(str
 extern bool acpi_cpc_valid(void);
 extern bool cppc_allow_fast_switch(void);
 extern int acpi_get_psd_map(unsigned int cpu, struct cppc_cpudata *cpu_data);
-extern unsigned int cppc_get_transition_latency(int cpu);
+extern int cppc_get_transition_latency(int cpu);
 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);
@@ -216,9 +216,9 @@ static inline bool cppc_allow_fast_switc
 {
 	return false;
 }
-static inline unsigned int cppc_get_transition_latency(int cpu)
+static inline int cppc_get_transition_latency(int cpu)
 {
-	return CPUFREQ_ETERNAL;
+	return -ENODATA;
 }
 static inline bool cpc_ffh_supported(void)
 {

[PATCH v3 4/4] cpufreq: Drop unused symbol CPUFREQ_ETERNAL

[Rafael J. Wysocki]

From: Rafael J. Wysocki <rafael.j.wysocki@intel.com>

Drop CPUFREQ_ETERNAL that has no users any more along with all
references to it in the documentation.

No functional impact.

Reviewed-by: Mario Limonciello (AMD) <superm1@kernel.org>
Reviewed-by: Jie Zhan <zhanjie9@hisilicon.com>
Signed-off-by: Rafael J. Wysocki <rafael.j.wysocki@intel.com>
---

v1 -> v3: Add tags from Mario Limonciello and Jie Zhan

---
 Documentation/admin-guide/pm/cpufreq.rst                  |    4 ----
 Documentation/cpu-freq/cpu-drivers.rst                    |    3 +--
 Documentation/translations/zh_CN/cpu-freq/cpu-drivers.rst |    3 +--
 Documentation/translations/zh_TW/cpu-freq/cpu-drivers.rst |    3 +--
 include/linux/cpufreq.h                                   |    5 -----
 5 files changed, 3 insertions(+), 15 deletions(-)

--- a/Documentation/admin-guide/pm/cpufreq.rst
+++ b/Documentation/admin-guide/pm/cpufreq.rst
@@ -274,10 +274,6 @@ are the following:
 	The time it takes to switch the CPUs belonging to this policy from one
 	P-state to another, in nanoseconds.
 
-	If unknown or if known to be so high that the scaling driver does not
-	work with the `ondemand`_ governor, -1 (:c:macro:`CPUFREQ_ETERNAL`)
-	will be returned by reads from this attribute.
-
 ``related_cpus``
 	List of all (online and offline) CPUs belonging to this policy.
 
--- a/Documentation/cpu-freq/cpu-drivers.rst
+++ b/Documentation/cpu-freq/cpu-drivers.rst
@@ -109,8 +109,7 @@ Then, the driver must fill in the follow
 +-----------------------------------+--------------------------------------+
 |policy->cpuinfo.transition_latency | the time it takes on this CPU to	   |
 |				    | switch between two frequencies in	   |
-|				    | nanoseconds (if appropriate, else	   |
-|				    | specify CPUFREQ_ETERNAL)		   |
+|				    | nanoseconds                          |
 +-----------------------------------+--------------------------------------+
 |policy->cur			    | The current operating frequency of   |
 |				    | this CPU (if appropriate)		   |
--- a/Documentation/translations/zh_CN/cpu-freq/cpu-drivers.rst
+++ b/Documentation/translations/zh_CN/cpu-freq/cpu-drivers.rst
@@ -112,8 +112,7 @@ CPUfreq核心层注册一个cpufreq_driv
 |                                   |                                      |
 +-----------------------------------+--------------------------------------+
 |policy->cpuinfo.transition_latency | CPU在两个频率之间切换所需的时间，以  |
-|                                   | 纳秒为单位（如不适用，设定为         |
-|                                   | CPUFREQ_ETERNAL）                    |
+|                                   | 纳秒为单位                    |
 |                                   |                                      |
 +-----------------------------------+--------------------------------------+
 |policy->cur                        | 该CPU当前的工作频率(如适用)          |
--- a/Documentation/translations/zh_TW/cpu-freq/cpu-drivers.rst
+++ b/Documentation/translations/zh_TW/cpu-freq/cpu-drivers.rst
@@ -112,8 +112,7 @@ CPUfreq核心層註冊一個cpufreq_driv
 |                                   |                                      |
 +-----------------------------------+--------------------------------------+
 |policy->cpuinfo.transition_latency | CPU在兩個頻率之間切換所需的時間，以  |
-|                                   | 納秒爲單位（如不適用，設定爲         |
-|                                   | CPUFREQ_ETERNAL）                    |
+|                                   | 納秒爲單位                    |
 |                                   |                                      |
 +-----------------------------------+--------------------------------------+
 |policy->cur                        | 該CPU當前的工作頻率(如適用)          |
--- a/include/linux/cpufreq.h
+++ b/include/linux/cpufreq.h
@@ -26,13 +26,8 @@
  *********************************************************************/
 /*
  * Frequency values here are CPU kHz
- *
- * Maximum transition latency is in nanoseconds - if it's unknown,
- * CPUFREQ_ETERNAL shall be used.
  */
 
-#define CPUFREQ_ETERNAL			(-1)
-
 #define CPUFREQ_DEFAULT_TANSITION_LATENCY_NS	NSEC_PER_MSEC
 
 #define CPUFREQ_NAME_LEN		16

Re: [PATCH v3 2/4] cpufreq: CPPC: Avoid using CPUFREQ_ETERNAL as transition delay

[Jie Zhan]

On 9/26/2025 6:19 PM, Rafael J. Wysocki wrote:
> From: Rafael J. Wysocki <rafael.j.wysocki@intel.com>
> 
> If cppc_get_transition_latency() returns CPUFREQ_ETERNAL to indicate a
> failure to retrieve the transition latency value from the platform
> firmware, the CPPC cpufreq driver will use that value (converted to
> microseconds) as the policy transition delay, but it is way too large
> for any practical use.
> 
> Address this by making the driver use the cpufreq's default
> transition latency value (in microseconds) as the transition delay
> if CPUFREQ_ETERNAL is returned by cppc_get_transition_latency().
> 
> Fixes: d4f3388afd48 ("cpufreq / CPPC: Set platform specific transition_delay_us")
> Cc: 5.19+ <stable@vger.kernel.org> # 5.19
> Signed-off-by: Rafael J. Wysocki <rafael.j.wysocki@intel.com>
> Reviewed-by: Mario Limonciello (AMD) <superm1@kernel.org>
Reviewed-by: Jie Zhan <zhanjie9@hisilicon.com>
> ---
> 
> v1 -> v3:
>    * Change the name of the new function (Jie Zhan)
>    * Add a tag from Mario Limonciello
> 
> ---
>  drivers/cpufreq/cppc_cpufreq.c |   14 ++++++++++++--
>  1 file changed, 12 insertions(+), 2 deletions(-)

Re: [PATCH v3 1/4] cpufreq: Make drivers using CPUFREQ_ETERNAL specify transition latency

[Shawn Guo]

On Fri, Sep 26, 2025 at 12:12:37PM +0200, Rafael J. Wysocki wrote:
> From: Rafael J. Wysocki <rafael.j.wysocki@intel.com>
> 
> Commit a755d0e2d41b ("cpufreq: Honour transition_latency over
> transition_delay_us") caused platforms where cpuinfo.transition_latency
> is CPUFREQ_ETERNAL to get a very large transition latency whereas
> previously it had been capped at 10 ms (and later at 2 ms).
> 
> This led to a user-observable regression between 6.6 and 6.12 as
> described by Shawn:
> 
> "The dbs sampling_rate was 10000 us on 6.6 and suddently becomes
>  6442450 us (4294967295 / 1000 * 1.5) on 6.12 for these platforms
>  because the default transition delay was dropped [...].
> 
>  It slows down dbs governor's reacting to CPU loading change
>  dramatically.  Also, as transition_delay_us is used by schedutil
>  governor as rate_limit_us, it shows a negative impact on device
>  idle power consumption, because the device gets slightly less time
>  in the lowest OPP."
> 
> Evidently, the expectation of the drivers using CPUFREQ_ETERNAL as
> cpuinfo.transition_latency was that it would be capped by the core,
> but they may as well return a default transition latency value instead
> of CPUFREQ_ETERNAL and the core need not do anything with it.
> 
> Accordingly, introduce CPUFREQ_DEFAULT_TRANSITION_LATENCY_NS and make
> all of the drivers in question use it instead of CPUFREQ_ETERNAL.  Also
> update the related Rust binding.
> 
> Fixes: a755d0e2d41b ("cpufreq: Honour transition_latency over transition_delay_us")
> Closes: https://lore.kernel.org/linux-pm/20250922125929.453444-1-shawnguo2@yeah.net/
> Reported-by: Shawn Guo <shawnguo@kernel.org>
> Reviewed-by: Mario Limonciello (AMD) <superm1@kernel.org>
> Reviewed-by: Jie Zhan <zhanjie9@hisilicon.com>
> Cc: 6.6+ <stable@vger.kernel.org> # 6.6+
> Signed-off-by: Rafael J. Wysocki <rafael.j.wysocki@intel.com>
> ---
> 
> v1 -> v3:
>    * Add updates of the Rust version of cpufreq-dt and Rust binding
>    * Update the changelog
>    * Add tags from Mario Limonciello and Jie Zhan
> 
> ---
>  drivers/cpufreq/cpufreq-dt.c          |    2 +-
>  drivers/cpufreq/imx6q-cpufreq.c       |    2 +-
>  drivers/cpufreq/mediatek-cpufreq-hw.c |    2 +-
>  drivers/cpufreq/rcpufreq_dt.rs        |    2 +-
>  drivers/cpufreq/scmi-cpufreq.c        |    2 +-
>  drivers/cpufreq/scpi-cpufreq.c        |    2 +-
>  drivers/cpufreq/spear-cpufreq.c       |    2 +-
>  include/linux/cpufreq.h               |    3 +++
>  rust/kernel/cpufreq.rs                |    7 ++++---
>  9 files changed, 14 insertions(+), 10 deletions(-)
> 
> --- a/drivers/cpufreq/cpufreq-dt.c
> +++ b/drivers/cpufreq/cpufreq-dt.c
> @@ -104,7 +104,7 @@ static int cpufreq_init(struct cpufreq_p
>  
>  	transition_latency = dev_pm_opp_get_max_transition_latency(cpu_dev);
>  	if (!transition_latency)
> -		transition_latency = CPUFREQ_ETERNAL;
> +		transition_latency = CPUFREQ_DEFAULT_TRANSITION_LATENCY_NS;
>  
>  	cpumask_copy(policy->cpus, priv->cpus);
>  	policy->driver_data = priv;
> --- a/drivers/cpufreq/imx6q-cpufreq.c
> +++ b/drivers/cpufreq/imx6q-cpufreq.c
> @@ -442,7 +442,7 @@ soc_opp_out:
>  	}
>  
>  	if (of_property_read_u32(np, "clock-latency", &transition_latency))
> -		transition_latency = CPUFREQ_ETERNAL;
> +		transition_latency = CPUFREQ_DEFAULT_TRANSITION_LATENCY_NS;
>  
>  	/*
>  	 * Calculate the ramp time for max voltage change in the
> --- a/drivers/cpufreq/mediatek-cpufreq-hw.c
> +++ b/drivers/cpufreq/mediatek-cpufreq-hw.c
> @@ -309,7 +309,7 @@ static int mtk_cpufreq_hw_cpu_init(struc
>  
>  	latency = readl_relaxed(data->reg_bases[REG_FREQ_LATENCY]) * 1000;
>  	if (!latency)
> -		latency = CPUFREQ_ETERNAL;
> +		latency = CPUFREQ_DEFAULT_TRANSITION_LATENCY_NS;
>  
>  	policy->cpuinfo.transition_latency = latency;
>  	policy->fast_switch_possible = true;
> --- a/drivers/cpufreq/rcpufreq_dt.rs
> +++ b/drivers/cpufreq/rcpufreq_dt.rs
> @@ -123,7 +123,7 @@ impl cpufreq::Driver for CPUFreqDTDriver
>  
>          let mut transition_latency = opp_table.max_transition_latency_ns() as u32;
>          if transition_latency == 0 {
> -            transition_latency = cpufreq::ETERNAL_LATENCY_NS;
> +            transition_latency = cpufreq::DEFAULT_TRANSITION_LATENCY_NS;
>          }
>  
>          policy
> --- a/drivers/cpufreq/scmi-cpufreq.c
> +++ b/drivers/cpufreq/scmi-cpufreq.c
> @@ -294,7 +294,7 @@ static int scmi_cpufreq_init(struct cpuf
>  
>  	latency = perf_ops->transition_latency_get(ph, domain);
>  	if (!latency)
> -		latency = CPUFREQ_ETERNAL;
> +		latency = CPUFREQ_DEFAULT_TRANSITION_LATENCY_NS;
>  
>  	policy->cpuinfo.transition_latency = latency;
>  
> --- a/drivers/cpufreq/scpi-cpufreq.c
> +++ b/drivers/cpufreq/scpi-cpufreq.c
> @@ -157,7 +157,7 @@ static int scpi_cpufreq_init(struct cpuf
>  
>  	latency = scpi_ops->get_transition_latency(cpu_dev);
>  	if (!latency)
> -		latency = CPUFREQ_ETERNAL;
> +		latency = CPUFREQ_DEFAULT_TRANSITION_LATENCY_NS;
>  
>  	policy->cpuinfo.transition_latency = latency;
>  
> --- a/drivers/cpufreq/spear-cpufreq.c
> +++ b/drivers/cpufreq/spear-cpufreq.c
> @@ -182,7 +182,7 @@ static int spear_cpufreq_probe(struct pl
>  
>  	if (of_property_read_u32(np, "clock-latency",
>  				&spear_cpufreq.transition_latency))
> -		spear_cpufreq.transition_latency = CPUFREQ_ETERNAL;
> +		spear_cpufreq.transition_latency = CPUFREQ_DEFAULT_TRANSITION_LATENCY_NS;
>  
>  	cnt = of_property_count_u32_elems(np, "cpufreq_tbl");
>  	if (cnt <= 0) {
> --- a/include/linux/cpufreq.h
> +++ b/include/linux/cpufreq.h
> @@ -32,6 +32,9 @@
>   */
>  
>  #define CPUFREQ_ETERNAL			(-1)
> +
> +#define CPUFREQ_DEFAULT_TANSITION_LATENCY_NS	NSEC_PER_MSEC

Typo of TANSITION, should be CPUFREQ_DEFAULT_TRANSITION_LATENCY_NS.

Shawn

> +
>  #define CPUFREQ_NAME_LEN		16
>  /* Print length for names. Extra 1 space for accommodating '\n' in prints */
>  #define CPUFREQ_NAME_PLEN		(CPUFREQ_NAME_LEN + 1)
> --- a/rust/kernel/cpufreq.rs
> +++ b/rust/kernel/cpufreq.rs
> @@ -39,7 +39,8 @@ use macros::vtable;
>  const CPUFREQ_NAME_LEN: usize = bindings::CPUFREQ_NAME_LEN as usize;
>  
>  /// Default transition latency value in nanoseconds.
> -pub const ETERNAL_LATENCY_NS: u32 = bindings::CPUFREQ_ETERNAL as u32;
> +pub const DEFAULT_TRANSITION_LATENCY_NS: u32 =
> +        bindings::CPUFREQ_DEFAULT_TRANSITION_LATENCY_NS as u32;
>  
>  /// CPU frequency driver flags.
>  pub mod flags {
> @@ -400,13 +401,13 @@ impl TableBuilder {
>  /// The following example demonstrates how to create a CPU frequency table.
>  ///
>  /// ```
> -/// use kernel::cpufreq::{ETERNAL_LATENCY_NS, Policy};
> +/// use kernel::cpufreq::{DEFAULT_TRANSITION_LATENCY_NS, Policy};
>  ///
>  /// fn update_policy(policy: &mut Policy) {
>  ///     policy
>  ///         .set_dvfs_possible_from_any_cpu(true)
>  ///         .set_fast_switch_possible(true)
> -///         .set_transition_latency_ns(ETERNAL_LATENCY_NS);
> +///         .set_transition_latency_ns(DEFAULT_TRANSITION_LATENCY_NS);
>  ///
>  ///     pr_info!("The policy details are: {:?}\n", (policy.cpu(), policy.cur()));
>  /// }
> 
> 
> 

Re: [PATCH v3 1/4] cpufreq: Make drivers using CPUFREQ_ETERNAL specify transition latency

[Rafael J. Wysocki]

On Sun, Sep 28, 2025 at 5:24 AM Shawn Guo <shawnguo2@yeah.net> wrote:
>
> On Fri, Sep 26, 2025 at 12:12:37PM +0200, Rafael J. Wysocki wrote:
> > From: Rafael J. Wysocki <rafael.j.wysocki@intel.com>
> >
> > Commit a755d0e2d41b ("cpufreq: Honour transition_latency over
> > transition_delay_us") caused platforms where cpuinfo.transition_latency
> > is CPUFREQ_ETERNAL to get a very large transition latency whereas
> > previously it had been capped at 10 ms (and later at 2 ms).
> >
> > This led to a user-observable regression between 6.6 and 6.12 as
> > described by Shawn:
> >
> > "The dbs sampling_rate was 10000 us on 6.6 and suddently becomes
> >  6442450 us (4294967295 / 1000 * 1.5) on 6.12 for these platforms
> >  because the default transition delay was dropped [...].
> >
> >  It slows down dbs governor's reacting to CPU loading change
> >  dramatically.  Also, as transition_delay_us is used by schedutil
> >  governor as rate_limit_us, it shows a negative impact on device
> >  idle power consumption, because the device gets slightly less time
> >  in the lowest OPP."
> >
> > Evidently, the expectation of the drivers using CPUFREQ_ETERNAL as
> > cpuinfo.transition_latency was that it would be capped by the core,
> > but they may as well return a default transition latency value instead
> > of CPUFREQ_ETERNAL and the core need not do anything with it.
> >
> > Accordingly, introduce CPUFREQ_DEFAULT_TRANSITION_LATENCY_NS and make
> > all of the drivers in question use it instead of CPUFREQ_ETERNAL.  Also
> > update the related Rust binding.
> >
> > Fixes: a755d0e2d41b ("cpufreq: Honour transition_latency over transition_delay_us")
> > Closes: https://lore.kernel.org/linux-pm/20250922125929.453444-1-shawnguo2@yeah.net/
> > Reported-by: Shawn Guo <shawnguo@kernel.org>
> > Reviewed-by: Mario Limonciello (AMD) <superm1@kernel.org>
> > Reviewed-by: Jie Zhan <zhanjie9@hisilicon.com>
> > Cc: 6.6+ <stable@vger.kernel.org> # 6.6+
> > Signed-off-by: Rafael J. Wysocki <rafael.j.wysocki@intel.com>
> > ---
> >
> > v1 -> v3:
> >    * Add updates of the Rust version of cpufreq-dt and Rust binding
> >    * Update the changelog
> >    * Add tags from Mario Limonciello and Jie Zhan
> >
> > ---
> >  drivers/cpufreq/cpufreq-dt.c          |    2 +-
> >  drivers/cpufreq/imx6q-cpufreq.c       |    2 +-
> >  drivers/cpufreq/mediatek-cpufreq-hw.c |    2 +-
> >  drivers/cpufreq/rcpufreq_dt.rs        |    2 +-
> >  drivers/cpufreq/scmi-cpufreq.c        |    2 +-
> >  drivers/cpufreq/scpi-cpufreq.c        |    2 +-
> >  drivers/cpufreq/spear-cpufreq.c       |    2 +-
> >  include/linux/cpufreq.h               |    3 +++
> >  rust/kernel/cpufreq.rs                |    7 ++++---
> >  9 files changed, 14 insertions(+), 10 deletions(-)
> >
> > --- a/drivers/cpufreq/cpufreq-dt.c
> > +++ b/drivers/cpufreq/cpufreq-dt.c
> > @@ -104,7 +104,7 @@ static int cpufreq_init(struct cpufreq_p
> >
> >       transition_latency = dev_pm_opp_get_max_transition_latency(cpu_dev);
> >       if (!transition_latency)
> > -             transition_latency = CPUFREQ_ETERNAL;
> > +             transition_latency = CPUFREQ_DEFAULT_TRANSITION_LATENCY_NS;
> >
> >       cpumask_copy(policy->cpus, priv->cpus);
> >       policy->driver_data = priv;
> > --- a/drivers/cpufreq/imx6q-cpufreq.c
> > +++ b/drivers/cpufreq/imx6q-cpufreq.c
> > @@ -442,7 +442,7 @@ soc_opp_out:
> >       }
> >
> >       if (of_property_read_u32(np, "clock-latency", &transition_latency))
> > -             transition_latency = CPUFREQ_ETERNAL;
> > +             transition_latency = CPUFREQ_DEFAULT_TRANSITION_LATENCY_NS;
> >
> >       /*
> >        * Calculate the ramp time for max voltage change in the
> > --- a/drivers/cpufreq/mediatek-cpufreq-hw.c
> > +++ b/drivers/cpufreq/mediatek-cpufreq-hw.c
> > @@ -309,7 +309,7 @@ static int mtk_cpufreq_hw_cpu_init(struc
> >
> >       latency = readl_relaxed(data->reg_bases[REG_FREQ_LATENCY]) * 1000;
> >       if (!latency)
> > -             latency = CPUFREQ_ETERNAL;
> > +             latency = CPUFREQ_DEFAULT_TRANSITION_LATENCY_NS;
> >
> >       policy->cpuinfo.transition_latency = latency;
> >       policy->fast_switch_possible = true;
> > --- a/drivers/cpufreq/rcpufreq_dt.rs
> > +++ b/drivers/cpufreq/rcpufreq_dt.rs
> > @@ -123,7 +123,7 @@ impl cpufreq::Driver for CPUFreqDTDriver
> >
> >          let mut transition_latency = opp_table.max_transition_latency_ns() as u32;
> >          if transition_latency == 0 {
> > -            transition_latency = cpufreq::ETERNAL_LATENCY_NS;
> > +            transition_latency = cpufreq::DEFAULT_TRANSITION_LATENCY_NS;
> >          }
> >
> >          policy
> > --- a/drivers/cpufreq/scmi-cpufreq.c
> > +++ b/drivers/cpufreq/scmi-cpufreq.c
> > @@ -294,7 +294,7 @@ static int scmi_cpufreq_init(struct cpuf
> >
> >       latency = perf_ops->transition_latency_get(ph, domain);
> >       if (!latency)
> > -             latency = CPUFREQ_ETERNAL;
> > +             latency = CPUFREQ_DEFAULT_TRANSITION_LATENCY_NS;
> >
> >       policy->cpuinfo.transition_latency = latency;
> >
> > --- a/drivers/cpufreq/scpi-cpufreq.c
> > +++ b/drivers/cpufreq/scpi-cpufreq.c
> > @@ -157,7 +157,7 @@ static int scpi_cpufreq_init(struct cpuf
> >
> >       latency = scpi_ops->get_transition_latency(cpu_dev);
> >       if (!latency)
> > -             latency = CPUFREQ_ETERNAL;
> > +             latency = CPUFREQ_DEFAULT_TRANSITION_LATENCY_NS;
> >
> >       policy->cpuinfo.transition_latency = latency;
> >
> > --- a/drivers/cpufreq/spear-cpufreq.c
> > +++ b/drivers/cpufreq/spear-cpufreq.c
> > @@ -182,7 +182,7 @@ static int spear_cpufreq_probe(struct pl
> >
> >       if (of_property_read_u32(np, "clock-latency",
> >                               &spear_cpufreq.transition_latency))
> > -             spear_cpufreq.transition_latency = CPUFREQ_ETERNAL;
> > +             spear_cpufreq.transition_latency = CPUFREQ_DEFAULT_TRANSITION_LATENCY_NS;
> >
> >       cnt = of_property_count_u32_elems(np, "cpufreq_tbl");
> >       if (cnt <= 0) {
> > --- a/include/linux/cpufreq.h
> > +++ b/include/linux/cpufreq.h
> > @@ -32,6 +32,9 @@
> >   */
> >
> >  #define CPUFREQ_ETERNAL                      (-1)
> > +
> > +#define CPUFREQ_DEFAULT_TANSITION_LATENCY_NS NSEC_PER_MSEC
>
> Typo of TANSITION, should be CPUFREQ_DEFAULT_TRANSITION_LATENCY_NS.

Yes, thanks!

Fixed already because CIs caught it yesterday.

I'm not sure how this happened though.  I must have mangled the patch
right before sending it because I had tested the whole patchset before
posting it.

> > +
> >  #define CPUFREQ_NAME_LEN             16
> >  /* Print length for names. Extra 1 space for accommodating '\n' in prints */
> >  #define CPUFREQ_NAME_PLEN            (CPUFREQ_NAME_LEN + 1)
> > --- a/rust/kernel/cpufreq.rs
> > +++ b/rust/kernel/cpufreq.rs
> > @@ -39,7 +39,8 @@ use macros::vtable;
> >  const CPUFREQ_NAME_LEN: usize = bindings::CPUFREQ_NAME_LEN as usize;
> >
> >  /// Default transition latency value in nanoseconds.
> > -pub const ETERNAL_LATENCY_NS: u32 = bindings::CPUFREQ_ETERNAL as u32;
> > +pub const DEFAULT_TRANSITION_LATENCY_NS: u32 =
> > +        bindings::CPUFREQ_DEFAULT_TRANSITION_LATENCY_NS as u32;
> >
> >  /// CPU frequency driver flags.
> >  pub mod flags {
> > @@ -400,13 +401,13 @@ impl TableBuilder {
> >  /// The following example demonstrates how to create a CPU frequency table.
> >  ///
> >  /// ```
> > -/// use kernel::cpufreq::{ETERNAL_LATENCY_NS, Policy};
> > +/// use kernel::cpufreq::{DEFAULT_TRANSITION_LATENCY_NS, Policy};
> >  ///
> >  /// fn update_policy(policy: &mut Policy) {
> >  ///     policy
> >  ///         .set_dvfs_possible_from_any_cpu(true)
> >  ///         .set_fast_switch_possible(true)
> > -///         .set_transition_latency_ns(ETERNAL_LATENCY_NS);
> > +///         .set_transition_latency_ns(DEFAULT_TRANSITION_LATENCY_NS);
> >  ///
> >  ///     pr_info!("The policy details are: {:?}\n", (policy.cpu(), policy.cur()));
> >  /// }
> >
> >
> >
>

Re: [PATCH v3 1/4] cpufreq: Make drivers using CPUFREQ_ETERNAL specify transition latency

[Shawn Guo]

On Sun, Sep 28, 2025 at 12:00:22PM +0200, Rafael J. Wysocki wrote:
> > > --- a/include/linux/cpufreq.h
> > > +++ b/include/linux/cpufreq.h
> > > @@ -32,6 +32,9 @@
> > >   */
> > >
> > >  #define CPUFREQ_ETERNAL                      (-1)
> > > +
> > > +#define CPUFREQ_DEFAULT_TANSITION_LATENCY_NS NSEC_PER_MSEC
> >
> > Typo of TANSITION, should be CPUFREQ_DEFAULT_TRANSITION_LATENCY_NS.
> 
> Yes, thanks!
> 
> Fixed already because CIs caught it yesterday.

With it fixed:

Tested-by: Shawn Guo <shawnguo@kernel.org> # with cpufreq-dt driver

BTW, a heads-up: the patch won't apply to 6.12 stable kernel directly.

Shawn

Re: [PATCH v3 1/4] cpufreq: Make drivers using CPUFREQ_ETERNAL specify transition latency

[Qais Yousef]

On 09/26/25 12:12, Rafael J. Wysocki wrote:
> From: Rafael J. Wysocki <rafael.j.wysocki@intel.com>
> 
> Commit a755d0e2d41b ("cpufreq: Honour transition_latency over
> transition_delay_us") caused platforms where cpuinfo.transition_latency
> is CPUFREQ_ETERNAL to get a very large transition latency whereas
> previously it had been capped at 10 ms (and later at 2 ms).
> 
> This led to a user-observable regression between 6.6 and 6.12 as
> described by Shawn:
> 
> "The dbs sampling_rate was 10000 us on 6.6 and suddently becomes
>  6442450 us (4294967295 / 1000 * 1.5) on 6.12 for these platforms
>  because the default transition delay was dropped [...].
> 
>  It slows down dbs governor's reacting to CPU loading change
>  dramatically.  Also, as transition_delay_us is used by schedutil
>  governor as rate_limit_us, it shows a negative impact on device
>  idle power consumption, because the device gets slightly less time
>  in the lowest OPP."
> 
> Evidently, the expectation of the drivers using CPUFREQ_ETERNAL as
> cpuinfo.transition_latency was that it would be capped by the core,
> but they may as well return a default transition latency value instead
> of CPUFREQ_ETERNAL and the core need not do anything with it.
> 
> Accordingly, introduce CPUFREQ_DEFAULT_TRANSITION_LATENCY_NS and make
> all of the drivers in question use it instead of CPUFREQ_ETERNAL.  Also
> update the related Rust binding.
> 
> Fixes: a755d0e2d41b ("cpufreq: Honour transition_latency over transition_delay_us")
> Closes: https://lore.kernel.org/linux-pm/20250922125929.453444-1-shawnguo2@yeah.net/
> Reported-by: Shawn Guo <shawnguo@kernel.org>
> Reviewed-by: Mario Limonciello (AMD) <superm1@kernel.org>
> Reviewed-by: Jie Zhan <zhanjie9@hisilicon.com>
> Cc: 6.6+ <stable@vger.kernel.org> # 6.6+
> Signed-off-by: Rafael J. Wysocki <rafael.j.wysocki@intel.com>

The whole series LGTM. I think this is clearer now without the CPUFREQ_ETERNAL.

Reviewed-by: Qais Yousef <qyousef@layalina.io>


Cheers

--
Qais Yousef