According to the Linux kernel ABI documentation for 'scaling_setspeed':
  "It returns the last frequency requested by the governor (in kHz) or
   can be written to in order to set a new frequency for the policy."

However, the current implementation of show_speed() returns 'policy->cur'.
'policy->cur' represents the frequency after the driver has
resolved the request against the hardware frequency table and applied
policy limits (min/max).

This creates a discrepancy between the documentation/user expectation and
the actual code behavior. For instance:

1. User writes a value to 'scaling_setspeed' that is not in the OPP table
   (e.g., user asks for A, driver rounds it to B).
2. User reads 'scaling_setspeed'.
3. Code returns B ('policy->cur').
4. User expects A (the "frequency requested"), but gets B.

This patch changes show_speed() to return 'userspace->setspeed', which
stores the actual value last requested by the user. This restores the
read/write symmetry of the attribute and aligns the code with the ABI
description.

The effective frequency can still be observed via 'scaling_cur_freq' or
'cpuinfo_cur_freq', preserving the distinction between "what was
requested" (setspeed) and "what is effective" (cur_freq).

Signed-off-by: Pengjie Zhang <zhangpengjie2@huawei.com>
---
 drivers/cpufreq/cpufreq_userspace.c | 4 +++-
 1 file changed, 3 insertions(+), 1 deletion(-)

diff --git a/drivers/cpufreq/cpufreq_userspace.c b/drivers/cpufreq/cpufreq_userspace.c
index 77d62152cd38..4bd62e6c5c51 100644
--- a/drivers/cpufreq/cpufreq_userspace.c
+++ b/drivers/cpufreq/cpufreq_userspace.c
@@ -49,7 +49,9 @@ static int cpufreq_set(struct cpufreq_policy *policy, unsigned int freq)
 
 static ssize_t show_speed(struct cpufreq_policy *policy, char *buf)
 {
-	return sprintf(buf, "%u\n", policy->cur);
+	struct userspace_policy *userspace = policy->governor_data;
+
+	return sprintf(buf, "%u\n", userspace->setspeed);
 }
 
 static int cpufreq_userspace_policy_init(struct cpufreq_policy *policy)
-- 
2.33.0

On 16-01-26, 17:46, Pengjie Zhang wrote:
> According to the Linux kernel ABI documentation for 'scaling_setspeed':
>   "It returns the last frequency requested by the governor (in kHz) or
>    can be written to in order to set a new frequency for the policy."
> 
> However, the current implementation of show_speed() returns 'policy->cur'.
> 'policy->cur' represents the frequency after the driver has
> resolved the request against the hardware frequency table and applied
> policy limits (min/max).
> 
> This creates a discrepancy between the documentation/user expectation and
> the actual code behavior. For instance:
> 
> 1. User writes a value to 'scaling_setspeed' that is not in the OPP table
>    (e.g., user asks for A, driver rounds it to B).
> 2. User reads 'scaling_setspeed'.
> 3. Code returns B ('policy->cur').
> 4. User expects A (the "frequency requested"), but gets B.
> 
> This patch changes show_speed() to return 'userspace->setspeed', which
> stores the actual value last requested by the user. This restores the
> read/write symmetry of the attribute and aligns the code with the ABI
> description.
> 
> The effective frequency can still be observed via 'scaling_cur_freq' or
> 'cpuinfo_cur_freq', preserving the distinction between "what was
> requested" (setspeed) and "what is effective" (cur_freq).
> 
> Signed-off-by: Pengjie Zhang <zhangpengjie2@huawei.com>
> ---
>  drivers/cpufreq/cpufreq_userspace.c | 4 +++-
>  1 file changed, 3 insertions(+), 1 deletion(-)
> 
> diff --git a/drivers/cpufreq/cpufreq_userspace.c b/drivers/cpufreq/cpufreq_userspace.c
> index 77d62152cd38..4bd62e6c5c51 100644
> --- a/drivers/cpufreq/cpufreq_userspace.c
> +++ b/drivers/cpufreq/cpufreq_userspace.c
> @@ -49,7 +49,9 @@ static int cpufreq_set(struct cpufreq_policy *policy, unsigned int freq)
>  
>  static ssize_t show_speed(struct cpufreq_policy *policy, char *buf)
>  {
> -	return sprintf(buf, "%u\n", policy->cur);
> +	struct userspace_policy *userspace = policy->governor_data;
> +
> +	return sprintf(buf, "%u\n", userspace->setspeed);
>  }
>  
>  static int cpufreq_userspace_policy_init(struct cpufreq_policy *policy)

Looks okay to me.

Acked-by: Viresh Kumar <viresh.kumar@linaro.org>

-- 
viresh

On Mon, Jan 19, 2026 at 9:28 AM Viresh Kumar <viresh.kumar@linaro.org> wrote:
>
> On 16-01-26, 17:46, Pengjie Zhang wrote:
> > According to the Linux kernel ABI documentation for 'scaling_setspeed':
> >   "It returns the last frequency requested by the governor (in kHz) or
> >    can be written to in order to set a new frequency for the policy."
> >
> > However, the current implementation of show_speed() returns 'policy->cur'.
> > 'policy->cur' represents the frequency after the driver has
> > resolved the request against the hardware frequency table and applied
> > policy limits (min/max).
> >
> > This creates a discrepancy between the documentation/user expectation and
> > the actual code behavior. For instance:
> >
> > 1. User writes a value to 'scaling_setspeed' that is not in the OPP table
> >    (e.g., user asks for A, driver rounds it to B).
> > 2. User reads 'scaling_setspeed'.
> > 3. Code returns B ('policy->cur').
> > 4. User expects A (the "frequency requested"), but gets B.
> >
> > This patch changes show_speed() to return 'userspace->setspeed', which
> > stores the actual value last requested by the user. This restores the
> > read/write symmetry of the attribute and aligns the code with the ABI
> > description.
> >
> > The effective frequency can still be observed via 'scaling_cur_freq' or
> > 'cpuinfo_cur_freq', preserving the distinction between "what was
> > requested" (setspeed) and "what is effective" (cur_freq).
> >
> > Signed-off-by: Pengjie Zhang <zhangpengjie2@huawei.com>
> > ---
> >  drivers/cpufreq/cpufreq_userspace.c | 4 +++-
> >  1 file changed, 3 insertions(+), 1 deletion(-)
> >
> > diff --git a/drivers/cpufreq/cpufreq_userspace.c b/drivers/cpufreq/cpufreq_userspace.c
> > index 77d62152cd38..4bd62e6c5c51 100644
> > --- a/drivers/cpufreq/cpufreq_userspace.c
> > +++ b/drivers/cpufreq/cpufreq_userspace.c
> > @@ -49,7 +49,9 @@ static int cpufreq_set(struct cpufreq_policy *policy, unsigned int freq)
> >
> >  static ssize_t show_speed(struct cpufreq_policy *policy, char *buf)
> >  {
> > -     return sprintf(buf, "%u\n", policy->cur);
> > +     struct userspace_policy *userspace = policy->governor_data;
> > +
> > +     return sprintf(buf, "%u\n", userspace->setspeed);
> >  }
> >
> >  static int cpufreq_userspace_policy_init(struct cpufreq_policy *policy)
>
> Looks okay to me.
>
> Acked-by: Viresh Kumar <viresh.kumar@linaro.org>

Applied as 6.20 material, thanks!

On 1/19/2026 4:28 PM, Viresh Kumar wrote:
> On 16-01-26, 17:46, Pengjie Zhang wrote:
>> According to the Linux kernel ABI documentation for 'scaling_setspeed':
>>    "It returns the last frequency requested by the governor (in kHz) or
>>     can be written to in order to set a new frequency for the policy."
>>
>> However, the current implementation of show_speed() returns 'policy->cur'.
>> 'policy->cur' represents the frequency after the driver has
>> resolved the request against the hardware frequency table and applied
>> policy limits (min/max).
>>
>> This creates a discrepancy between the documentation/user expectation and
>> the actual code behavior. For instance:
>>
>> 1. User writes a value to 'scaling_setspeed' that is not in the OPP table
>>     (e.g., user asks for A, driver rounds it to B).
>> 2. User reads 'scaling_setspeed'.
>> 3. Code returns B ('policy->cur').
>> 4. User expects A (the "frequency requested"), but gets B.
>>
>> This patch changes show_speed() to return 'userspace->setspeed', which
>> stores the actual value last requested by the user. This restores the
>> read/write symmetry of the attribute and aligns the code with the ABI
>> description.
>>
>> The effective frequency can still be observed via 'scaling_cur_freq' or
>> 'cpuinfo_cur_freq', preserving the distinction between "what was
>> requested" (setspeed) and "what is effective" (cur_freq).
>>
>> Signed-off-by: Pengjie Zhang <zhangpengjie2@huawei.com>
>> ---
>>   drivers/cpufreq/cpufreq_userspace.c | 4 +++-
>>   1 file changed, 3 insertions(+), 1 deletion(-)
>>
>> diff --git a/drivers/cpufreq/cpufreq_userspace.c b/drivers/cpufreq/cpufreq_userspace.c
>> index 77d62152cd38..4bd62e6c5c51 100644
>> --- a/drivers/cpufreq/cpufreq_userspace.c
>> +++ b/drivers/cpufreq/cpufreq_userspace.c
>> @@ -49,7 +49,9 @@ static int cpufreq_set(struct cpufreq_policy *policy, unsigned int freq)
>>   
>>   static ssize_t show_speed(struct cpufreq_policy *policy, char *buf)
>>   {
>> -	return sprintf(buf, "%u\n", policy->cur);
>> +	struct userspace_policy *userspace = policy->governor_data;
>> +
>> +	return sprintf(buf, "%u\n", userspace->setspeed);
>>   }
>>   
>>   static int cpufreq_userspace_policy_init(struct cpufreq_policy *policy)
> Looks okay to me.
>
> Acked-by: Viresh Kumar <viresh.kumar@linaro.org>
This fix is more reasonable. Each interface has its own responsibility 
and no repeat.
Acked-by: lihuisong@huawei.com