[PATCH v1 0/3] Enable EAS for CPPC/ACPI based systems

[PATCH v1 0/3] Enable EAS for CPPC/ACPI based systems

[Pierre Gondois]

0. Overview

The current Energy Model (EM) for CPUs requires knowledge about CPU
performance states and their power consumption. Both of these
information is not available for ACPI based systems.

In ACPI, describing power efficiency of CPUs can be done through the
following arm specific field:

ACPI 6.4, s5.2.12.14 "GIC CPU Interface (GICC) Structure",
"Processor Power Efficiency Class field":
Describes the relative power efficiency of the associated pro-
cessor. Lower efficiency class numbers are more efficient than
higher ones (e.g. efficiency class 0 should be treated as more
efficient than efficiency class 1). However, absolute values
of this number have no meaning: 2 isn't necessarily half as
efficient as 1.

Add an 'efficiency_class' field to describe the relative power
efficiency of CPUs. CPUs relying on this field will have performance
states (power and frequency values) artificially created. Such EM will
be referred to as an artificial EM.

The artificial EM is used for the CPPC driver.

1. Dependencies

This patch-set has a dependency on:
 - [0/8] Introduce support for artificial Energy Model
https://lkml.org/lkml/2022/3/16/850
introduces a new callback in the Energy Model (EM) and prevents the
registration of devices using power values from an EM when the EM
is artificial. Not having this patch-set would break builds.
 - This patch-set based on linux-next.

2. Testing

This patch-set has been tested on a Juno-r2 and a Pixel4. Two types
of tests were done: energy testing, and performance testing.

The energy testing was done with 2 sets of tasks:
- homogeneous tasks (#Tasks at 5% utilization and 16ms period)
- heterogeneous tasks (#Tasks at 5|10|15% utilization and 16ms period).
  If a test has 3 tasks, then there is one with each utilization
  (1 at 5%, 1 at 10%, 1 at 15%).
Tasks spawn on the biggest CPU(s) of the platform. If there are
multiple big CPUs, tasks spawn alternatively on big CPUs.

2.1. Juno-r2 testing

The Juno-r2 has 6 CPUs:
- 4 little [0, 3-5], max_capa=383
- 2 big [1-2], max_capa=1024
Base kernel is v5.17-rc5.

2.1.1. Energy testing

The tests were done on:
- a system using a DT and the scmi cpufreq driver. Comparison
  is done between no-EAS and EAS.
- a system using ACPI and the cppc cpufreq driver. Comparison
  is done between CPPC-no-EAS and CPPC-EAS. CPPC-EAS uses
  the artificial EM.

Energy numbers come from the Juno energy counter, by summing
little and big clusters energy spending. There has been 5 iterations
of each test. Lower energy spending is better.

2.1.1.1. Homogeneous tasks

Energy results (Joules):
+--------+-------------------+-----------------------------+
|        |            no-EAS |                         EAS |
+--------+---------+---------+-------------------+---------+
| #Tasks |    Mean | ci(+/-) |              Mean | ci(+/-) |
+--------+---------+---------+-------------------+---------+
|    10  |   7.89  |    0.26 |     6.99 (-11.36) |    0.49 |
|    20  |  13.42  |    0.32 |    13.42 ( -0.02) |    0.08 |
|    30  |  21.43  |    0.98 |    21.62 ( +0.87) |    0.63 |
|    40  |  30.03  |    0.82 |    30.31 ( +0.94) |    0.37 |
|    50  |  43.19  |    0.56 |    43.50 ( +0.72) |    0.52 |
+--------+---------+---------+-------------------+---------+
+--------+-------------------+-----------------------------+
|        |       CPPC-no-EAS |                    CPPC-EAS |
+--------+---------+---------+-------------------+---------+
| #Tasks |    Mean | ci(+/-) |              Mean | ci(+/-) |
+--------+---------+---------+-------------------+---------+
|    10  |    7.86 |    0.37 |     5.64 (-28.23) |    0.05 |
|    20  |   13.36 |    0.20 |    10.92 (-18.31) |    0.31 |
|    30  |   19.28 |    0.34 |    18.30 ( -5.07) |    0.64 |
|    40  |   28.33 |    0.59 |    27.13 ( -4.23) |    0.42 |
|    50  |   40.78 |    0.58 |    40.77 ( -0.04) |    0.45 |
+--------+---------+---------+-------------------+---------+

Missed activations were measured while comparing CPPC-no-EAS/CPPC-EAS
energy values. They were of 0.00% for all tests and both
configurations. Missed activations start to appear in a significant
number starting from ~70 tasks.

2.1.1.2. Heterogeneous tasks

Energy results (Joules):
+--------+-------------------+-----------------------------+
|        |            no-EAS |                         EAS |
+--------+---------+---------+-------------------+---------+
| #Tasks |    Mean | ci(+/-) |              Mean | ci(+/-) |
+--------+---------+---------+-------------------+---------+
|     3  |    5.25 |    0.50 |    4.58 (-12.82%) |    0.07 |
|     9  |   12.30 |    0.28 |   11.45 ( -6.97%) |    0.34 |
|    15  |   20.06 |    1.32 |   20.60 (  2.66%) |    1.00 |
|    21  |   30.03 |    0.63 |   30.07 (  0.12%) |    0.41 |
+--------+---------+---------+-------------------+---------+
+--------+-------------------+-----------------------------+
|        |       CPPC-no-EAS |                    CPPC-EAS |
+--------+---------+---------+-------------------+---------+
| #Tasks |    Mean | ci(+/-) |              Mean | ci(+/-) |
+--------+---------+---------+-------------------+---------+
|     3  |    4.58 |    0.31 |    3.65 (-20.31%) |    0.05 |
|     9  |   11.53 |    0.20 |    9.23 (-19.97%) |    0.22 |
|    15  |   19.19 |    0.16 |   18.33 ( -4.49%) |    0.71 |
|    21  |   29.07 |    0.29 |   29.06 ( -0.01%) |    0.08 |
+--------+---------+---------+-------------------+---------+

Missed activations were measured while comparing CPPC-no-EAS/CPPC-EAS
energy values. They were of 0.00% for all tests and both
configurations. Missed activations start to appear in a significant
number starting from ~36 tasks.

2.1.1.3. Analysis:

The artificial EM often shows better energy gains than the EM,
especially for small loads. Indeed, the artificial power values
show a huge energy gain by placing tasks on little CPUs. The 6%
margin is always reached, so tasks are easily placed on little
CPUs. The margin is not always reached with real power values,
leading to tasks staying on big CPUs.

2.1.2. Performance testing

10 iterations of HackBench with the "--pipe --thread" options and
1000 loops. Compared value is the testing time in seconds. A lower
timing is better.
+----------------+-------------------+---------------------------+
|                |       CPPC-no-EAS |                  CPPC-EAS |
+--------+-------+---------+---------+-----------------+---------+
| Groups | Tasks |    Mean | ci(+/-) |           Mean  | ci(+/-) |
+--------+-------+---------+---------+-----------------+---------+
|      1 |    40 |    2.39 |    0.19 |   2.39 (-0.24%) |    0.07 |
|      2 |    80 |    5.56 |    0.48 |   5.28 (-5.02%) |    0.42 |
|      4 |   160 |   12.15 |    0.84 |  12.06 (-0.80%) |    0.48 |
|      8 |   320 |   23.03 |    0.94 |  23.12 (+0.36%) |    0.70 |
+--------+-------+---------+---------+-----------------+---------+

The performance is overall sligthly better, but stays in the margin
or error.


2.2. Pixel4 testing

Pixel4 has 7 CPUs:
- 4 little [0-3], max_capa=261
- 3 medium [4-6], max_capa=861
- 1 big [7], max_capa=1024

Base kernel is android-10.0.0_r0.81. The performance states advertised
in the DT were modified with performance states that would be generated
by this patch-set.
The artificial EM was set such as little CPUs > medium CPUs > big CPU,
meaning little CPUs are the most energy efficient.
Comparing the power/capacity ratio, little CPUs' performance states are
all more energy efficient than the medium CPUs' performance states.
This is wrong when comparing medium and big CPUs.

2.2.1. Energy testing

The 2 sets of tests (heterogeneous/homogeneous) were tested while
registering battery voltage and current (power is obtained by
multiplying them).
Voltage is averaged over a rolling period of ~11s and current over a
period of ~6s. Usb-C cable is plugged in but alimentation is cut.
Pixel4 is on airplane mode. The tests lasts 120s, the first 50s and
last 10s are trimmed as the power is slowly raising to reach a
plateau.
Are compared:
- android with EAS (but NO_FIND_BEST_TARGET is set):
  echo ENERGY_AWARE > /sys/kernel/debug/sched_features
  echo NO_FIND_BEST_TARGET > /sys/kernel/debug/sched_features
- android without EAS:
  echo NO_ENERGY_AWARE > /sys/kernel/debug/sched_features
- android with the artificial energy model
Lower energy spending is better.

2.2.1.2. Homogeneous tasks

Energy results (in uW):
+--------+-------------------+-----------------------------+
|        |       Without EAS |                    With EAS |
+--------+---------+---------+-------------------+---------+
| #Tasks |    Mean | ci(+/-) |              Mean | ci(+/-) |
+--------+---------+---------+-------------------+---------+
|    10  | 6.21+05 | 3.12+02 | 5.09+05 (-18.01%) | 2.18+03 |
|    20  | 9.12+05 | 9.71+02 | 7.91+05 (-13.26%) | 9.92+02 |
|    30  | 1.25+06 | 2.02+03 | 1.09+06 (-12.12%) | 2.00+03 |
|    40  | 2.05+06 | 5.15+03 | 1.38+06 (-32.36%) | 1.21+03 |
|    50  | 3.03+06 | 6.94+03 | 1.89+06 (-37.44%) | 3.21+03 |
+--------+---------+---------+-------------------+---------+
+--------+-------------------+-----------------------------+
|        |       Without EAS |                  With patch |
+--------+---------+---------+-------------------+---------+
| #Tasks |    Mean | ci(+/-) |              Mean | ci(+/-) |
+--------+---------+---------+-------------------+---------+
|    10  | 6.21+05 | 3.12+02 | 4.39+05 (-29.29%) | 5.63+02 |
|    20  | 9.12+05 | 9.71+02 | 7.30+05 (-19.90%) | 1.98+03 |
|    30  | 1.25+06 | 2.02+03 | 1.01+06 (-18.60%) | 1.72+03 |
|    40  | 2.05+06 | 5.15+03 | 1.38+06 (-32.60%) | 3.93+03 |
|    50  | 3.03+06 | 6.94+03 | 2.05+06 (-32.08%) | 1.25+04 |
+--------+---------+---------+-------------------+---------+

2.2.1.2. Heterogeneous tasks

Energy results (in uW):
+--------+-------------------+-----------------------------+
|        |       Without EAS |                    With EAS |
+--------+---------+---------+-------------------+---------+
| #Tasks |    Mean | ci(+/-) |              Mean | ci(+/-) |
+--------+---------+---------+-------------------+---------+
|     3  | 5.14+05 | 1.06+03 | 3.76+05 (-26.82%) | 4.58+02 |
|     9  | 8.52+05 | 1.18+03 | 7.25+05 (-14.96%) | 1.39+03 |
|    15  | 1.42+06 | 3.14+03 | 1.20+06 (-15.41%) | 1.06+04 |
|    21  | 2.73+06 | 3.49+03 | 1.49+06 (-45.47%) | 3.43+03 |
|    27  | 3.17+06 | 6.92+03 | 2.42+06 (-23.77%) | 8.43+03 |
+--------+---------+---------+-------------------+---------+
+--------+-------------------+-----------------------------+
|        |       Without EAS |                  With patch |
+--------+---------+---------+-------------------+---------+
| #Tasks |    Mean | ci(+/-) |              Mean | ci(+/-) |
+--------+---------+---------+-------------------+---------+
|     3  | 5.14+05 | 1.06+03 | 3.82+05 (-25.70%) | 7.67+02 |
|     9  | 8.52+05 | 1.18+03 | 7.05+05 (-17.30%) | 9.79+02 |
|    15  | 1.42+06 | 3.14+03 | 1.05+06 (-26.00%) | 1.15+03 |
|    21  | 2.73+06 | 3.49+03 | 1.53+06 (-43.68%) | 2.23+03 |
|    27  | 3.17+06 | 6.92+03 | 2.86+06 ( -9.77%) | 4.26+03 |
+--------+---------+---------+-------------------+---------+

2.2.1.2. Analysis

Similarly to Juno, the artificial performance states show a huge
gain to place tasks on small CPUs, leading to better energy results.

2.2.2. Performance testing

10 iterations of PcMark. Compared value is the final score
(PcmaWorkv3Score). A bigger score is better.
+----------------+-------------------------+-------------------------+
|    Without EAS |                With EAS |              With patch |
+------+---------+---------------+---------+---------------+---------+
| Mean | ci(+/-) |          Mean | ci(+/-) |          Mean | ci(+/-) |
+------+---------+---------------+---------+---------------+---------+
| 8026 |      86 |          8003 |      74 | 7840 (-2.00%) |     104 |
+------+---------+---------------+---------+---------------+---------+

Performance is lower, but still in the margin of error.


3. Summary

The artificial performance states show overall better energy results
and a small performance decrease. They lead to a more aggressive task
placement on the most energy efficient CPUs, and this explains the
results.

 arch/arm64/kernel/smp.c        |   1 +
 drivers/cpufreq/cppc_cpufreq.c | 212 +++++++++++++++++++++++++++++++++
 2 files changed, 213 insertions(+)

-- 
2.25.1

[PATCH v1 1/3] cpufreq: CPPC: Add cppc_cpufreq_search_cpu_data

[Pierre Gondois]

cppc_cpufreq_get_cpu_data() allocates a new struct cppc_cpudata
for the input CPU at each call.

To search the struct associated with a cpu without allocating
a new one, add cppc_cpufreq_search_cpu_data().
Also add an early prototype.

This will be used in a later patch, when generating artificial
performance states to register an artificial Energy Model in the
cppc_cpufreq driver and enable the Energy Aware Scheduler for ACPI
based systems.

Signed-off-by: Pierre Gondois <Pierre.Gondois@arm.com>
---
 drivers/cpufreq/cppc_cpufreq.c | 15 +++++++++++++++
 1 file changed, 15 insertions(+)

diff --git a/drivers/cpufreq/cppc_cpufreq.c b/drivers/cpufreq/cppc_cpufreq.c
index 82d370ae6a4a..8f950fe72765 100644
--- a/drivers/cpufreq/cppc_cpufreq.c
+++ b/drivers/cpufreq/cppc_cpufreq.c
@@ -41,6 +41,8 @@
  */
 static LIST_HEAD(cpu_data_list);
 
+static struct cppc_cpudata *cppc_cpufreq_search_cpu_data(unsigned int cpu);
+
 static bool boost_supported;
 
 struct cppc_workaround_oem_info {
@@ -479,6 +481,19 @@ static void cppc_cpufreq_put_cpu_data(struct cpufreq_policy *policy)
 	policy->driver_data = NULL;
 }
 
+static inline struct cppc_cpudata *
+cppc_cpufreq_search_cpu_data(unsigned int cpu)
+{
+	struct cppc_cpudata *iter, *tmp;
+
+	list_for_each_entry_safe(iter, tmp, &cpu_data_list, node) {
+		if (cpumask_test_cpu(cpu, iter->shared_cpu_map))
+			return iter;
+	}
+
+	return NULL;
+}
+
 static int cppc_cpufreq_cpu_init(struct cpufreq_policy *policy)
 {
 	unsigned int cpu = policy->cpu;
-- 
2.25.1

Re: [PATCH v1 1/3] cpufreq: CPPC: Add cppc_cpufreq_search_cpu_data

[Marc Zyngier]

On 2022-03-17 13:34, Pierre Gondois wrote:
> cppc_cpufreq_get_cpu_data() allocates a new struct cppc_cpudata
> for the input CPU at each call.
> 
> To search the struct associated with a cpu without allocating
> a new one, add cppc_cpufreq_search_cpu_data().
> Also add an early prototype.
> 
> This will be used in a later patch, when generating artificial
> performance states to register an artificial Energy Model in the
> cppc_cpufreq driver and enable the Energy Aware Scheduler for ACPI
> based systems.
> 
> Signed-off-by: Pierre Gondois <Pierre.Gondois@arm.com>
> ---
>  drivers/cpufreq/cppc_cpufreq.c | 15 +++++++++++++++
>  1 file changed, 15 insertions(+)
> 
> diff --git a/drivers/cpufreq/cppc_cpufreq.c 
> b/drivers/cpufreq/cppc_cpufreq.c
> index 82d370ae6a4a..8f950fe72765 100644
> --- a/drivers/cpufreq/cppc_cpufreq.c
> +++ b/drivers/cpufreq/cppc_cpufreq.c
> @@ -41,6 +41,8 @@
>   */
>  static LIST_HEAD(cpu_data_list);
> 
> +static struct cppc_cpudata *cppc_cpufreq_search_cpu_data(unsigned int 
> cpu);
> +
>  static bool boost_supported;
> 
>  struct cppc_workaround_oem_info {
> @@ -479,6 +481,19 @@ static void cppc_cpufreq_put_cpu_data(struct
> cpufreq_policy *policy)
>  	policy->driver_data = NULL;
>  }
> 
> +static inline struct cppc_cpudata *

Why the inline? This is hardly performance critical, and if
it is, you want something better than iterating over a list.

> +cppc_cpufreq_search_cpu_data(unsigned int cpu)
> +{
> +	struct cppc_cpudata *iter, *tmp;
> +
> +	list_for_each_entry_safe(iter, tmp, &cpu_data_list, node) {
> +		if (cpumask_test_cpu(cpu, iter->shared_cpu_map))
> +			return iter;
> +	}
> +
> +	return NULL;
> +}
> +
>  static int cppc_cpufreq_cpu_init(struct cpufreq_policy *policy)
>  {
>  	unsigned int cpu = policy->cpu;

Thanks,

         M.
-- 
Jazz is not dead. It just smells funny...

Re: [PATCH v1 1/3] cpufreq: CPPC: Add cppc_cpufreq_search_cpu_data

[Pierre Gondois]

On 3/17/22 15:20, Marc Zyngier wrote:
> On 2022-03-17 13:34, Pierre Gondois wrote:
>> cppc_cpufreq_get_cpu_data() allocates a new struct cppc_cpudata
>> for the input CPU at each call.
>>
>> To search the struct associated with a cpu without allocating
>> a new one, add cppc_cpufreq_search_cpu_data().
>> Also add an early prototype.
>>
>> This will be used in a later patch, when generating artificial
>> performance states to register an artificial Energy Model in the
>> cppc_cpufreq driver and enable the Energy Aware Scheduler for ACPI
>> based systems.
>>
>> Signed-off-by: Pierre Gondois <Pierre.Gondois@arm.com>
>> ---
>>   drivers/cpufreq/cppc_cpufreq.c | 15 +++++++++++++++
>>   1 file changed, 15 insertions(+)
>>
>> diff --git a/drivers/cpufreq/cppc_cpufreq.c
>> b/drivers/cpufreq/cppc_cpufreq.c
>> index 82d370ae6a4a..8f950fe72765 100644
>> --- a/drivers/cpufreq/cppc_cpufreq.c
>> +++ b/drivers/cpufreq/cppc_cpufreq.c
>> @@ -41,6 +41,8 @@
>>    */
>>   static LIST_HEAD(cpu_data_list);
>>
>> +static struct cppc_cpudata *cppc_cpufreq_search_cpu_data(unsigned int
>> cpu);
>> +
>>   static bool boost_supported;
>>
>>   struct cppc_workaround_oem_info {
>> @@ -479,6 +481,19 @@ static void cppc_cpufreq_put_cpu_data(struct
>> cpufreq_policy *policy)
>>   	policy->driver_data = NULL;
>>   }
>>
>> +static inline struct cppc_cpudata *
> 
> Why the inline? This is hardly performance critical, and if
> it is, you want something better than iterating over a list.

This was made inline mainly because the function was small. The function
is called only at boot, so it should not be performance critical. The
'inline' can be removed if necessary.
Would letting it inlined have a negative impact ?


> 
>> +cppc_cpufreq_search_cpu_data(unsigned int cpu)
>> +{
>> +	struct cppc_cpudata *iter, *tmp;
>> +
>> +	list_for_each_entry_safe(iter, tmp, &cpu_data_list, node) {
>> +		if (cpumask_test_cpu(cpu, iter->shared_cpu_map))
>> +			return iter;
>> +	}
>> +
>> +	return NULL;
>> +}
>> +
>>   static int cppc_cpufreq_cpu_init(struct cpufreq_policy *policy)
>>   {
>>   	unsigned int cpu = policy->cpu;
> 
> Thanks,
> 
>           M.

Regards,
Pierre

Re: [PATCH v1 1/3] cpufreq: CPPC: Add cppc_cpufreq_search_cpu_data

[Marc Zyngier]

On 2022-03-17 14:44, Pierre Gondois wrote:
> On 3/17/22 15:20, Marc Zyngier wrote:
>> On 2022-03-17 13:34, Pierre Gondois wrote:
>>> cppc_cpufreq_get_cpu_data() allocates a new struct cppc_cpudata
>>> for the input CPU at each call.
>>> 
>>> To search the struct associated with a cpu without allocating
>>> a new one, add cppc_cpufreq_search_cpu_data().
>>> Also add an early prototype.
>>> 
>>> This will be used in a later patch, when generating artificial
>>> performance states to register an artificial Energy Model in the
>>> cppc_cpufreq driver and enable the Energy Aware Scheduler for ACPI
>>> based systems.
>>> 
>>> Signed-off-by: Pierre Gondois <Pierre.Gondois@arm.com>
>>> ---
>>>   drivers/cpufreq/cppc_cpufreq.c | 15 +++++++++++++++
>>>   1 file changed, 15 insertions(+)
>>> 
>>> diff --git a/drivers/cpufreq/cppc_cpufreq.c
>>> b/drivers/cpufreq/cppc_cpufreq.c
>>> index 82d370ae6a4a..8f950fe72765 100644
>>> --- a/drivers/cpufreq/cppc_cpufreq.c
>>> +++ b/drivers/cpufreq/cppc_cpufreq.c
>>> @@ -41,6 +41,8 @@
>>>    */
>>>   static LIST_HEAD(cpu_data_list);
>>> 
>>> +static struct cppc_cpudata *cppc_cpufreq_search_cpu_data(unsigned 
>>> int
>>> cpu);
>>> +
>>>   static bool boost_supported;
>>> 
>>>   struct cppc_workaround_oem_info {
>>> @@ -479,6 +481,19 @@ static void cppc_cpufreq_put_cpu_data(struct
>>> cpufreq_policy *policy)
>>>   	policy->driver_data = NULL;
>>>   }
>>> 
>>> +static inline struct cppc_cpudata *
>> 
>> Why the inline? This is hardly performance critical, and if
>> it is, you want something better than iterating over a list.
> 
> This was made inline mainly because the function was small. The 
> function
> is called only at boot, so it should not be performance critical. The
> 'inline' can be removed if necessary.
> Would letting it inlined have a negative impact ?

This is why we have a compiler. It is perfectly able to decide
on its own whether to inline or not, depending on how it can
optimise it. With modern compilers, 'inline' means nothing anyway,
and is ignored most of the time.

So dropping it will at least save 7 bytes of source code! ;-)

         M.
-- 
Jazz is not dead. It just smells funny...

[PATCH v1 2/3] cpufreq: CPPC: Add per_cpu efficiency_class

[Pierre Gondois]

In ACPI, describing power efficiency of CPUs can be done through the
following arm specific field:
ACPI 6.4, s5.2.12.14 'GIC CPU Interface (GICC) Structure',
'Processor Power Efficiency Class field':
  Describes the relative power efficiency of the associated pro-
  cessor. Lower efficiency class numbers are more efficient than
  higher ones (e.g. efficiency class 0 should be treated as more
  efficient than efficiency class 1). However, absolute values
  of this number have no meaning: 2 isn’t necessarily half as
  efficient as 1.

The efficiency_class field is stored in the GicC structure of the
ACPI MADT table and it's currently supported in Linux for arm64 only.
Thus, this new functionality is introduced for arm64 only.

To allow the cppc_cpufreq driver to know and preprocess the
efficiency_class values of all the CPUs, add a per_cpu efficiency_class
variable to store them. Also add a static efficiency_class_populated
to let the driver know efficiency_class values are usable and register
an artificial Energy Model (EM) based on normalized class values.

At least 2 different efficiency classes must be present,
otherwise there is no use in creating an Energy Model.

The efficiency_class values are squeezed in [0:#efficiency_class-1]
while conserving the order. For instance, efficiency classes of:
  [111, 212, 250]
will be mapped to:
  [0 (was 111), 1 (was 212), 2 (was 250)].

Each policy being independently registered in the driver, populating
the per_cpu efficiency_class is done only once at the driver
initialization. This prevents from having each policy re-searching the
efficiency_class values of other CPUs.

The patch also exports acpi_cpu_get_madt_gicc() to fetch the GicC
structure of the ACPI MADT table for each CPU.

Signed-off-by: Pierre Gondois <Pierre.Gondois@arm.com>
---
 arch/arm64/kernel/smp.c        |  1 +
 drivers/cpufreq/cppc_cpufreq.c | 55 ++++++++++++++++++++++++++++++++++
 2 files changed, 56 insertions(+)

diff --git a/arch/arm64/kernel/smp.c b/arch/arm64/kernel/smp.c
index 27df5c1e6baa..56637cbea5d6 100644
--- a/arch/arm64/kernel/smp.c
+++ b/arch/arm64/kernel/smp.c
@@ -512,6 +512,7 @@ struct acpi_madt_generic_interrupt *acpi_cpu_get_madt_gicc(int cpu)
 {
 	return &cpu_madt_gicc[cpu];
 }
+EXPORT_SYMBOL(acpi_cpu_get_madt_gicc);
 
 /*
  * acpi_map_gic_cpu_interface - parse processor MADT entry
diff --git a/drivers/cpufreq/cppc_cpufreq.c b/drivers/cpufreq/cppc_cpufreq.c
index 8f950fe72765..a6cd95c3b474 100644
--- a/drivers/cpufreq/cppc_cpufreq.c
+++ b/drivers/cpufreq/cppc_cpufreq.c
@@ -422,12 +422,66 @@ static unsigned int cppc_cpufreq_get_transition_delay_us(unsigned int cpu)
 	return cppc_get_transition_latency(cpu) / NSEC_PER_USEC;
 }
 
+static bool efficiency_class_populated;
+static DEFINE_PER_CPU(unsigned int, efficiency_class);
+
+static int populate_efficiency_class(void)
+{
+	unsigned int min = UINT_MAX, max = 0, class;
+	struct acpi_madt_generic_interrupt *gicc;
+	int cpu;
+
+	for_each_possible_cpu(cpu) {
+		gicc = acpi_cpu_get_madt_gicc(cpu);
+		if (!gicc)
+			return -ENODEV;
+
+		per_cpu(efficiency_class, cpu) = gicc->efficiency_class;
+		min = min_t(unsigned int, min, gicc->efficiency_class);
+		max = max_t(unsigned int, max, gicc->efficiency_class);
+	}
+
+	if (min == max) {
+		pr_debug("Efficiency classes are all equal (=%d). "
+			"No EM registered", max);
+		return -EINVAL;
+	}
+
+	/*
+	 * Squeeze efficiency class values on [0:#efficiency_class-1].
+	 * Values are per spec in [0:255].
+	 */
+	for (class = 0; class < 256; class++) {
+		unsigned int new_min, curr;
+
+		new_min = UINT_MAX;
+		for_each_possible_cpu(cpu) {
+			curr = per_cpu(efficiency_class, cpu);
+			if (curr == min)
+				per_cpu(efficiency_class, cpu) = class;
+			else if (curr > min)
+				new_min = min(new_min, curr);
+		}
+
+		if (new_min == UINT_MAX)
+			break;
+		min = new_min;
+	}
+
+	efficiency_class_populated = true;
+	return 0;
+}
+
 #else
 
 static unsigned int cppc_cpufreq_get_transition_delay_us(unsigned int cpu)
 {
 	return cppc_get_transition_latency(cpu) / NSEC_PER_USEC;
 }
+static int populate_efficiency_class(void)
+{
+	return 0;
+}
 #endif
 
 
@@ -757,6 +811,7 @@ static int __init cppc_cpufreq_init(void)
 
 	cppc_check_hisi_workaround();
 	cppc_freq_invariance_init();
+	populate_efficiency_class();
 
 	ret = cpufreq_register_driver(&cppc_cpufreq_driver);
 	if (ret)
-- 
2.25.1

Re: [PATCH v1 2/3] cpufreq: CPPC: Add per_cpu efficiency_class

[Marc Zyngier]

On 2022-03-17 13:34, Pierre Gondois wrote:
> In ACPI, describing power efficiency of CPUs can be done through the
> following arm specific field:
> ACPI 6.4, s5.2.12.14 'GIC CPU Interface (GICC) Structure',
> 'Processor Power Efficiency Class field':
>   Describes the relative power efficiency of the associated pro-
>   cessor. Lower efficiency class numbers are more efficient than
>   higher ones (e.g. efficiency class 0 should be treated as more
>   efficient than efficiency class 1). However, absolute values
>   of this number have no meaning: 2 isn’t necessarily half as
>   efficient as 1.
> 
> The efficiency_class field is stored in the GicC structure of the
> ACPI MADT table and it's currently supported in Linux for arm64 only.
> Thus, this new functionality is introduced for arm64 only.
> 
> To allow the cppc_cpufreq driver to know and preprocess the
> efficiency_class values of all the CPUs, add a per_cpu efficiency_class
> variable to store them. Also add a static efficiency_class_populated
> to let the driver know efficiency_class values are usable and register
> an artificial Energy Model (EM) based on normalized class values.
> 
> At least 2 different efficiency classes must be present,
> otherwise there is no use in creating an Energy Model.
> 
> The efficiency_class values are squeezed in [0:#efficiency_class-1]
> while conserving the order. For instance, efficiency classes of:
>   [111, 212, 250]
> will be mapped to:
>   [0 (was 111), 1 (was 212), 2 (was 250)].
> 
> Each policy being independently registered in the driver, populating
> the per_cpu efficiency_class is done only once at the driver
> initialization. This prevents from having each policy re-searching the
> efficiency_class values of other CPUs.
> 
> The patch also exports acpi_cpu_get_madt_gicc() to fetch the GicC
> structure of the ACPI MADT table for each CPU.
> 
> Signed-off-by: Pierre Gondois <Pierre.Gondois@arm.com>
> ---
>  arch/arm64/kernel/smp.c        |  1 +
>  drivers/cpufreq/cppc_cpufreq.c | 55 ++++++++++++++++++++++++++++++++++
>  2 files changed, 56 insertions(+)
> 
> diff --git a/arch/arm64/kernel/smp.c b/arch/arm64/kernel/smp.c
> index 27df5c1e6baa..56637cbea5d6 100644
> --- a/arch/arm64/kernel/smp.c
> +++ b/arch/arm64/kernel/smp.c
> @@ -512,6 +512,7 @@ struct acpi_madt_generic_interrupt
> *acpi_cpu_get_madt_gicc(int cpu)
>  {
>  	return &cpu_madt_gicc[cpu];
>  }
> +EXPORT_SYMBOL(acpi_cpu_get_madt_gicc);

Why not EXPORT_SYMBOL_GPL()?

> 
>  /*
>   * acpi_map_gic_cpu_interface - parse processor MADT entry
> diff --git a/drivers/cpufreq/cppc_cpufreq.c 
> b/drivers/cpufreq/cppc_cpufreq.c
> index 8f950fe72765..a6cd95c3b474 100644
> --- a/drivers/cpufreq/cppc_cpufreq.c
> +++ b/drivers/cpufreq/cppc_cpufreq.c
> @@ -422,12 +422,66 @@ static unsigned int
> cppc_cpufreq_get_transition_delay_us(unsigned int cpu)
>  	return cppc_get_transition_latency(cpu) / NSEC_PER_USEC;
>  }
> 
> +static bool efficiency_class_populated;
> +static DEFINE_PER_CPU(unsigned int, efficiency_class);
> +
> +static int populate_efficiency_class(void)
> +{
> +	unsigned int min = UINT_MAX, max = 0, class;
> +	struct acpi_madt_generic_interrupt *gicc;
> +	int cpu;
> +
> +	for_each_possible_cpu(cpu) {
> +		gicc = acpi_cpu_get_madt_gicc(cpu);
> +		if (!gicc)
> +			return -ENODEV;

How can that happen if you made it here using ACPI?

> +
> +		per_cpu(efficiency_class, cpu) = gicc->efficiency_class;
> +		min = min_t(unsigned int, min, gicc->efficiency_class);
> +		max = max_t(unsigned int, max, gicc->efficiency_class);
> +	}

Why don't you use a temporary bitmap of 256 bits, tracking
the classes that are actually being used?

> +
> +	if (min == max) {

This would become (bitmap_weight(used_classes) <= 1). Then from
the same construct you know how many different classes you have.
You also have the min, max, and all the values in between.

> +		pr_debug("Efficiency classes are all equal (=%d). "
> +			"No EM registered", max);
> +		return -EINVAL;
> +	}
> +
> +	/*
> +	 * Squeeze efficiency class values on [0:#efficiency_class-1].
> +	 * Values are per spec in [0:255].
> +	 */
> +	for (class = 0; class < 256; class++) {
> +		unsigned int new_min, curr;
> +
> +		new_min = UINT_MAX;
> +		for_each_possible_cpu(cpu) {
> +			curr = per_cpu(efficiency_class, cpu);
> +			if (curr == min)
> +				per_cpu(efficiency_class, cpu) = class;
> +			else if (curr > min)
> +				new_min = min(new_min, curr);
> +		}
> +
> +		if (new_min == UINT_MAX)
> +			break;
> +		min = new_min;
> +	}

I find it really hard to reason about this because you are
dynamically rewriting the values you keep reevaluating.

How about something like this, which I find more readable:

	DECLARE_BITMAP(used_classes, 256) = {};
	int class, index, cpu;

	for_each_possible_cpu(cpu) {
		unsigned int ec;

		ec = acpi_cpu_get_madt_gicc(cpu)->efficiency_class & 0xff;
		bitmap_set(ec, &used_classes);
	}

	if (bitmap_weight(&used_classes, 256) <= 1)
		return;

	index = 0;

	for_each_set_bit(class, &used_classes, 256) {
		for_each_possible_cpu(cpu) {
			if (acpi_cpu_get_madt_gicc(cpu)->efficiency_class == class)
				per_cpu(efficiency_class, cpu) = index;
		}

		index++;
	}


Thanks,

         M.
-- 
Jazz is not dead. It just smells funny...

Re: [PATCH v1 2/3] cpufreq: CPPC: Add per_cpu efficiency_class

[Pierre Gondois]

On 3/17/22 16:13, Marc Zyngier wrote:
> On 2022-03-17 13:34, Pierre Gondois wrote:
>> In ACPI, describing power efficiency of CPUs can be done through the
>> following arm specific field:
>> ACPI 6.4, s5.2.12.14 'GIC CPU Interface (GICC) Structure',
>> 'Processor Power Efficiency Class field':
>>    Describes the relative power efficiency of the associated pro-
>>    cessor. Lower efficiency class numbers are more efficient than
>>    higher ones (e.g. efficiency class 0 should be treated as more
>>    efficient than efficiency class 1). However, absolute values
>>    of this number have no meaning: 2 isn’t necessarily half as
>>    efficient as 1.
>>
>> The efficiency_class field is stored in the GicC structure of the
>> ACPI MADT table and it's currently supported in Linux for arm64 only.
>> Thus, this new functionality is introduced for arm64 only.
>>
>> To allow the cppc_cpufreq driver to know and preprocess the
>> efficiency_class values of all the CPUs, add a per_cpu efficiency_class
>> variable to store them. Also add a static efficiency_class_populated
>> to let the driver know efficiency_class values are usable and register
>> an artificial Energy Model (EM) based on normalized class values.
>>
>> At least 2 different efficiency classes must be present,
>> otherwise there is no use in creating an Energy Model.
>>
>> The efficiency_class values are squeezed in [0:#efficiency_class-1]
>> while conserving the order. For instance, efficiency classes of:
>>    [111, 212, 250]
>> will be mapped to:
>>    [0 (was 111), 1 (was 212), 2 (was 250)].
>>
>> Each policy being independently registered in the driver, populating
>> the per_cpu efficiency_class is done only once at the driver
>> initialization. This prevents from having each policy re-searching the
>> efficiency_class values of other CPUs.
>>
>> The patch also exports acpi_cpu_get_madt_gicc() to fetch the GicC
>> structure of the ACPI MADT table for each CPU.
>>
>> Signed-off-by: Pierre Gondois <Pierre.Gondois@arm.com>
>> ---
>>   arch/arm64/kernel/smp.c        |  1 +
>>   drivers/cpufreq/cppc_cpufreq.c | 55 ++++++++++++++++++++++++++++++++++
>>   2 files changed, 56 insertions(+)
>>
>> diff --git a/arch/arm64/kernel/smp.c b/arch/arm64/kernel/smp.c
>> index 27df5c1e6baa..56637cbea5d6 100644
>> --- a/arch/arm64/kernel/smp.c
>> +++ b/arch/arm64/kernel/smp.c
>> @@ -512,6 +512,7 @@ struct acpi_madt_generic_interrupt
>> *acpi_cpu_get_madt_gicc(int cpu)
>>   {
>>   	return &cpu_madt_gicc[cpu];
>>   }
>> +EXPORT_SYMBOL(acpi_cpu_get_madt_gicc);
> 
> Why not EXPORT_SYMBOL_GPL()?

 From what I understand, this could be made EXPORT_SYMBOL_GPL().
The only reason was that the other symbol exportation in the
file wasn't restricted to GPL.

> 
>>
>>   /*
>>    * acpi_map_gic_cpu_interface - parse processor MADT entry
>> diff --git a/drivers/cpufreq/cppc_cpufreq.c
>> b/drivers/cpufreq/cppc_cpufreq.c
>> index 8f950fe72765..a6cd95c3b474 100644
>> --- a/drivers/cpufreq/cppc_cpufreq.c
>> +++ b/drivers/cpufreq/cppc_cpufreq.c
>> @@ -422,12 +422,66 @@ static unsigned int
>> cppc_cpufreq_get_transition_delay_us(unsigned int cpu)
>>   	return cppc_get_transition_latency(cpu) / NSEC_PER_USEC;
>>   }
>>
>> +static bool efficiency_class_populated;
>> +static DEFINE_PER_CPU(unsigned int, efficiency_class);
>> +
>> +static int populate_efficiency_class(void)
>> +{
>> +	unsigned int min = UINT_MAX, max = 0, class;
>> +	struct acpi_madt_generic_interrupt *gicc;
>> +	int cpu;
>> +
>> +	for_each_possible_cpu(cpu) {
>> +		gicc = acpi_cpu_get_madt_gicc(cpu);
>> +		if (!gicc)
>> +			return -ENODEV;
> 
> How can that happen if you made it here using ACPI?

This is effectively an extra check. This could be removed.

> 
>> +
>> +		per_cpu(efficiency_class, cpu) = gicc->efficiency_class;
>> +		min = min_t(unsigned int, min, gicc->efficiency_class);
>> +		max = max_t(unsigned int, max, gicc->efficiency_class);
>> +	}
> 
> Why don't you use a temporary bitmap of 256 bits, tracking
> the classes that are actually being used?
> 
>> +
>> +	if (min == max) {
> 
> This would become (bitmap_weight(used_classes) <= 1). Then from
> the same construct you know how many different classes you have.
> You also have the min, max, and all the values in between.
> 
>> +		pr_debug("Efficiency classes are all equal (=%d). "
>> +			"No EM registered", max);
>> +		return -EINVAL;
>> +	}
>> +
>> +	/*
>> +	 * Squeeze efficiency class values on [0:#efficiency_class-1].
>> +	 * Values are per spec in [0:255].
>> +	 */
>> +	for (class = 0; class < 256; class++) {
>> +		unsigned int new_min, curr;
>> +
>> +		new_min = UINT_MAX;
>> +		for_each_possible_cpu(cpu) {
>> +			curr = per_cpu(efficiency_class, cpu);
>> +			if (curr == min)
>> +				per_cpu(efficiency_class, cpu) = class;
>> +			else if (curr > min)
>> +				new_min = min(new_min, curr);
>> +		}
>> +
>> +		if (new_min == UINT_MAX)
>> +			break;
>> +		min = new_min;
>> +	}
> 
> I find it really hard to reason about this because you are
> dynamically rewriting the values you keep reevaluating.
> 
> How about something like this, which I find more readable:
> 
> 	DECLARE_BITMAP(used_classes, 256) = {};
> 	int class, index, cpu;
> 
> 	for_each_possible_cpu(cpu) {
> 		unsigned int ec;
> 
> 		ec = acpi_cpu_get_madt_gicc(cpu)->efficiency_class & 0xff;
> 		bitmap_set(ec, &used_classes);
> 	}
> 
> 	if (bitmap_weight(&used_classes, 256) <= 1)
> 		return;
> 
> 	index = 0;
> 
> 	for_each_set_bit(class, &used_classes, 256) {
> 		for_each_possible_cpu(cpu) {
> 			if (acpi_cpu_get_madt_gicc(cpu)->efficiency_class == class)
> 				per_cpu(efficiency_class, cpu) = index;
> 		}
> 
> 		index++;
> 	}

This is effectively much more readable. Thanks for the code snippet.

Regards,
Pierre

> 
> 
> Thanks,
> 
>           M.

Re: [PATCH v1 2/3] cpufreq: CPPC: Add per_cpu efficiency_class

[Marc Zyngier]

On Thu, 17 Mar 2022 16:07:01 +0000,
Pierre Gondois <pierre.gondois@arm.com> wrote:
> 
> >> diff --git a/arch/arm64/kernel/smp.c b/arch/arm64/kernel/smp.c
> >> index 27df5c1e6baa..56637cbea5d6 100644
> >> --- a/arch/arm64/kernel/smp.c
> >> +++ b/arch/arm64/kernel/smp.c
> >> @@ -512,6 +512,7 @@ struct acpi_madt_generic_interrupt
> >> *acpi_cpu_get_madt_gicc(int cpu)
> >>   {
> >>   	return &cpu_madt_gicc[cpu];
> >>   }
> >> +EXPORT_SYMBOL(acpi_cpu_get_madt_gicc);
> > 
> > Why not EXPORT_SYMBOL_GPL()?
> 
> From what I understand, this could be made EXPORT_SYMBOL_GPL().
> The only reason was that the other symbol exportation in the
> file wasn't restricted to GPL.

I'm personally keen on keeping this for GPL code only, just like the
current code is. If there is a further need to relax this, we can
discuss it separately.

> 
> > 
> >> 
> >>   /*
> >>    * acpi_map_gic_cpu_interface - parse processor MADT entry
> >> diff --git a/drivers/cpufreq/cppc_cpufreq.c
> >> b/drivers/cpufreq/cppc_cpufreq.c
> >> index 8f950fe72765..a6cd95c3b474 100644
> >> --- a/drivers/cpufreq/cppc_cpufreq.c
> >> +++ b/drivers/cpufreq/cppc_cpufreq.c
> >> @@ -422,12 +422,66 @@ static unsigned int
> >> cppc_cpufreq_get_transition_delay_us(unsigned int cpu)
> >>   	return cppc_get_transition_latency(cpu) / NSEC_PER_USEC;
> >>   }
> >> 
> >> +static bool efficiency_class_populated;
> >> +static DEFINE_PER_CPU(unsigned int, efficiency_class);
> >> +
> >> +static int populate_efficiency_class(void)
> >> +{
> >> +	unsigned int min = UINT_MAX, max = 0, class;
> >> +	struct acpi_madt_generic_interrupt *gicc;
> >> +	int cpu;
> >> +
> >> +	for_each_possible_cpu(cpu) {
> >> +		gicc = acpi_cpu_get_madt_gicc(cpu);
> >> +		if (!gicc)
> >> +			return -ENODEV;
> > 
> > How can that happen if you made it here using ACPI?
> 
> This is effectively an extra check. This could be removed.

Please do.

Thanks,

	M.

-- 
Without deviation from the norm, progress is not possible.

[PATCH v1 3/3] cpufreq: CPPC: Register EM based on efficiency class information

[Pierre Gondois]

Performance states and energy consumption values are not advertised
in ACPI. In the GicC structure of the MADT table, the "Processor
Power Efficiency Class field" (called efficiency class from now)
allows to describe the relative energy efficiency of CPUs.

To leverage the EM and EAS, the CPPC driver creates a set of
artificial performance states and registers them in the Energy Model
(EM), such as:
- Every 20 capacity unit, a performance state is created.
- The energy cost of each performance state gradually increases.
No power value is generated as only the cost is used in the EM.

During task placement, a task can raise the frequency of its whole
pd. This can make EAS place a task on a pd with CPUs that are
individually less energy efficient.
As cost values are artificial, and to place tasks on CPUs with the
lower efficiency class, a gap in cost values is generated for adjacent
efficiency classes.
E.g.:
- efficiency class = 0, capacity is in [0-1024], so cost values
  are in [0: 51] (one performance state every 20 capacity unit)
- efficiency class = 1, capacity is in [0-1024], cost values
  are in [1*gap+0: 1*gap+51].

The value of the cost gap is chosen to absorb a the energy of 4 CPUs
at their maximum capacity. This means that between:
1- a pd of 4 CPUs, each of them being used at almost their full
   capacity. Their efficiency class is N.
2- a CPU using almost none of its capacity. Its efficiency class is
   N+1
EAS will choose the first option.

Signed-off-by: Pierre Gondois <Pierre.Gondois@arm.com>
---
 drivers/cpufreq/cppc_cpufreq.c | 142 +++++++++++++++++++++++++++++++++
 1 file changed, 142 insertions(+)

diff --git a/drivers/cpufreq/cppc_cpufreq.c b/drivers/cpufreq/cppc_cpufreq.c
index a6cd95c3b474..b65586511bc3 100644
--- a/drivers/cpufreq/cppc_cpufreq.c
+++ b/drivers/cpufreq/cppc_cpufreq.c
@@ -425,6 +425,129 @@ static unsigned int cppc_cpufreq_get_transition_delay_us(unsigned int cpu)
 static bool efficiency_class_populated;
 static DEFINE_PER_CPU(unsigned int, efficiency_class);
 
+/* Create an artificial performance state every CPPC_EM_CAP_STEP capacity unit. */
+#define CPPC_EM_CAP_STEP	(20)
+/* Increase the cost value by CPPC_EM_COST_STEP every performance state. */
+#define CPPC_EM_COST_STEP	(1)
+/* Add a cost gap correspnding to the energy of 4 CPUs. */
+#define CPPC_EM_COST_GAP	(4 * SCHED_CAPACITY_SCALE * CPPC_EM_COST_STEP \
+				/ CPPC_EM_CAP_STEP)
+
+static unsigned int get_perf_level_count(struct cpufreq_policy *policy)
+{
+	struct cppc_perf_caps *perf_caps;
+	unsigned int min_cap, max_cap;
+	struct cppc_cpudata *cpu_data;
+	int cpu = policy->cpu;
+
+	cpu_data = cppc_cpufreq_search_cpu_data(cpu);
+	perf_caps = &cpu_data->perf_caps;
+	max_cap = arch_scale_cpu_capacity(cpu);
+	min_cap = div_u64(max_cap * perf_caps->lowest_perf, perf_caps->highest_perf);
+	if ((min_cap == 0) || (max_cap < min_cap))
+		return 0;
+	return 1 + max_cap / CPPC_EM_CAP_STEP - min_cap / CPPC_EM_CAP_STEP;
+}
+
+/*
+ * The cost is defined as:
+ *   cost = power * max_frequency / frequency
+ */
+static inline unsigned long compute_cost(int cpu, int step)
+{
+	return CPPC_EM_COST_GAP * per_cpu(efficiency_class, cpu) +
+			step * CPPC_EM_COST_STEP;
+}
+
+static int cppc_get_cpu_power(struct device *cpu_dev,
+		unsigned long *power, unsigned long *KHz)
+{
+	unsigned long perf_step, perf_prev, perf, perf_check;
+	unsigned int min_step, max_step, step, step_check;
+	unsigned long prev_freq = *KHz;
+	unsigned int min_cap, max_cap;
+
+	struct cppc_perf_caps *perf_caps;
+	struct cppc_cpudata *cpu_data;
+
+	cpu_data = cppc_cpufreq_search_cpu_data(cpu_dev->id);
+	perf_caps = &cpu_data->perf_caps;
+	max_cap = arch_scale_cpu_capacity(cpu_dev->id);
+	min_cap = div_u64(max_cap * perf_caps->lowest_perf,
+			perf_caps->highest_perf);
+
+	perf_step = CPPC_EM_CAP_STEP * perf_caps->highest_perf / max_cap;
+	min_step = min_cap / CPPC_EM_CAP_STEP;
+	max_step = max_cap / CPPC_EM_CAP_STEP;
+
+	perf_prev = cppc_cpufreq_khz_to_perf(cpu_data, *KHz);
+	step = perf_prev / perf_step;
+
+	if (step > max_step)
+		return -EINVAL;
+
+	if (min_step == max_step) {
+		step = max_step;
+		perf = perf_caps->highest_perf;
+	} else if (step < min_step) {
+		step = min_step;
+		perf = perf_caps->lowest_perf;
+	} else {
+		step++;
+		if (step == max_step)
+			perf = perf_caps->highest_perf;
+		else
+			perf = step * perf_step;
+	}
+
+	*KHz = cppc_cpufreq_perf_to_khz(cpu_data, perf);
+	perf_check = cppc_cpufreq_khz_to_perf(cpu_data, *KHz);
+	step_check = perf_check / perf_step;
+
+	/*
+	 * To avoid bad integer approximation, check that new frequency value
+	 * increased and that the new frequency will be converted to the
+	 * desired step value.
+	 */
+	while ((*KHz == prev_freq) || (step_check != step)) {
+		perf++;
+		*KHz = cppc_cpufreq_perf_to_khz(cpu_data, perf);
+		perf_check = cppc_cpufreq_khz_to_perf(cpu_data, *KHz);
+		step_check = perf_check / perf_step;
+	}
+
+	/*
+	 * With an artificial EM, only the cost value is used. Still the power
+	 * is populated such as 0 < power < EM_MAX_POWER. This allows to add
+	 * more sense to the artificial performance states.
+	 */
+	*power = compute_cost(cpu_dev->id, step);
+
+	return 0;
+}
+
+static int cppc_get_cpu_cost(struct device *cpu_dev, unsigned long KHz,
+		unsigned long *cost)
+{
+	unsigned long perf_step, perf_prev;
+	struct cppc_perf_caps *perf_caps;
+	struct cppc_cpudata *cpu_data;
+	unsigned int max_cap;
+	int step;
+
+	cpu_data = cppc_cpufreq_search_cpu_data(cpu_dev->id);
+	perf_caps = &cpu_data->perf_caps;
+	max_cap = arch_scale_cpu_capacity(cpu_dev->id);
+
+	perf_prev = cppc_cpufreq_khz_to_perf(cpu_data, KHz);
+	perf_step = CPPC_EM_CAP_STEP * perf_caps->highest_perf / max_cap;
+	step = perf_prev / perf_step;
+
+	*cost = compute_cost(cpu_dev->id, step);
+
+	return 0;
+}
+
 static int populate_efficiency_class(void)
 {
 	unsigned int min = UINT_MAX, max = 0, class;
@@ -472,6 +595,21 @@ static int populate_efficiency_class(void)
 	return 0;
 }
 
+static void cppc_cpufreq_register_em(struct cpufreq_policy *policy)
+{
+	struct cppc_cpudata *cpu_data;
+	struct em_data_callback em_cb =
+		EM_ADV_DATA_CB(cppc_get_cpu_power, cppc_get_cpu_cost);
+
+	if (!efficiency_class_populated)
+		return;
+
+	cpu_data = cppc_cpufreq_search_cpu_data(policy->cpu);
+	em_dev_register_perf_domain(get_cpu_device(policy->cpu),
+			get_perf_level_count(policy), &em_cb,
+			cpu_data->shared_cpu_map, 0);
+}
+
 #else
 
 static unsigned int cppc_cpufreq_get_transition_delay_us(unsigned int cpu)
@@ -482,6 +620,9 @@ static int populate_efficiency_class(void)
 {
 	return 0;
 }
+static void cppc_cpufreq_register_em(struct cpufreq_policy *policy)
+{
+}
 #endif
 
 
@@ -753,6 +894,7 @@ static struct cpufreq_driver cppc_cpufreq_driver = {
 	.init = cppc_cpufreq_cpu_init,
 	.exit = cppc_cpufreq_cpu_exit,
 	.set_boost = cppc_cpufreq_set_boost,
+	.register_em = cppc_cpufreq_register_em,
 	.attr = cppc_cpufreq_attr,
 	.name = "cppc_cpufreq",
 };
-- 
2.25.1