Hi all,

This patch set adds new callback and support for artificial Energy Model (EM).
The new EMs have artificially generated performance states.
Such EMs can be created from lean information sources, such
as the relative energy efficiency between CPUs. The ACPI based
platforms provide this information
(ACPI 6.4, s5.2.12.14 'GIC CPU Interface (GICC) Structure'
'Processor Power efficiency Class' field).

Artificial EMs might require to directly provide the 'cost' of
the generated performance state. This patch set adds a new callback
.get_cost() for this. The EM framework does not force any model
or formula, it's up to the platform code.

Artificial EMs aim to leverage the Energy Aware Scheduler
(EAS). Other frameworks relying on performance states
information (i.e. IPA/DTPM) must be informed of the
EM type and might be prevented from using it. This patch
sets also does this by introducing a new flag:
EM_PERF_DOMAIN_ARTIFICIAL.

The patch set is based on current linux-next, where some
changes to OPP & EM are queuing.

The patch set also contains (patch 7/8 and patch 8/8) logic which prevents
two EM's client frameworks from using this new EM type. Some other approach,
using 'milli-watts', has been proposed and discussed, but refused [1].
This new flag is more precised and should not leave space for
wrong interpretation.

Shortly after this patch set you will see a patch set implementing the
platform code and registering this new EM.

Regards,
Lukasz Luba

[1] https://lore.kernel.org/linux-pm/20220207073036.14901-1-lukasz.luba@arm.com/

Lukasz Luba (7):
  PM: EM: Add .get_cost() callback
  PM: EM: Use the new .get_cost() callback while registering EM
  PM: EM: Change the order of arguments in the .active_power() callback
  PM: EM: Remove old debugfs files and print all 'flags'
  Documentation: EM: Add artificial EM registration description
  thermal: cooling: Check Energy Model type in cpufreq_cooling and
    devfreq_cooling
  powercap: DTPM: Check for Energy Model type

Pierre Gondois (1):
  PM: EM: Add artificial EM flag

 Documentation/power/energy-model.rst  | 24 +++++++++-
 drivers/cpufreq/mediatek-cpufreq-hw.c |  4 +-
 drivers/cpufreq/scmi-cpufreq.c        |  4 +-
 drivers/opp/of.c                      |  6 +--
 drivers/powercap/dtpm_cpu.c           |  2 +-
 drivers/thermal/cpufreq_cooling.c     |  2 +-
 drivers/thermal/devfreq_cooling.c     |  8 ++--
 include/linux/energy_model.h          | 35 +++++++++++++--
 kernel/power/energy_model.c           | 63 +++++++++++++++------------
 9 files changed, 101 insertions(+), 47 deletions(-)

-- 
2.17.1

On Wed, Mar 16, 2022 at 11:52:03PM +0000, Lukasz Luba wrote:
> Hi all,
> 

Hi Lukasz,

> This patch set adds new callback and support for artificial Energy Model (EM).
> The new EMs have artificially generated performance states.
> Such EMs can be created from lean information sources, such
> as the relative energy efficiency between CPUs. The ACPI based
> platforms provide this information
> (ACPI 6.4, s5.2.12.14 'GIC CPU Interface (GICC) Structure'
> 'Processor Power efficiency Class' field).
> 
> Artificial EMs might require to directly provide the 'cost' of
> the generated performance state. This patch set adds a new callback
> .get_cost() for this. The EM framework does not force any model
> or formula, it's up to the platform code.
> 
> Artificial EMs aim to leverage the Energy Aware Scheduler
> (EAS). Other frameworks relying on performance states
> information (i.e. IPA/DTPM) must be informed of the
> EM type and might be prevented from using it. This patch
> sets also does this by introducing a new flag:
> EM_PERF_DOMAIN_ARTIFICIAL.
> 
> The patch set is based on current linux-next, where some
> changes to OPP & EM are queuing.
> 
> The patch set also contains (patch 7/8 and patch 8/8) logic which prevents
> two EM's client frameworks from using this new EM type. Some other approach,
> using 'milli-watts', has been proposed and discussed, but refused [1].
> This new flag is more precised and should not leave space for
> wrong interpretation.
> 
> Shortly after this patch set you will see a patch set implementing the
> platform code and registering this new EM.
> 

Just to let you know that in the few days I'm going to post the first
chunk of some SCMIv3.1 additions that includes also (as you probably
know) the SCMI Perf protocol support for reporting perf_domain costs in
micro-watts and not only in milli-watts.

Given that it does not seem that as of now the em_ API used by the SCMI
cpufreq driver can make use of this new scale (and being not at all
familiar with EM/EAS for sure :P), the SCMIv3.1 'Perf micro-watts' patch
which I will post (I'll CC you) does NOT expose any new interface but only
takes care to store the new micro-watts capability internally in a flag
(if advertised by an SCMIv3.1 backend server), so that, basically, you'll
keep seeing from the SCMI cpufreq driver that the scale is milli-watt
(when milli-watts are used of course) or non-milli-watt (for abstract and
micro-watts scales).

This is intended to be of course a first step, laying out just the bare
minimum commmon internal SCMI support, until we figure out how to properly
expose this from the SCMI Perf in order to make it usable for EM.
(if neeeded at all).

Thanks,
Cristian

Hi Cristian,

On 3/29/22 14:29, Cristian Marussi wrote:
> On Wed, Mar 16, 2022 at 11:52:03PM +0000, Lukasz Luba wrote:
>> Hi all,
>>
> 
> Hi Lukasz,
> 
>> This patch set adds new callback and support for artificial Energy Model (EM).
>> The new EMs have artificially generated performance states.
>> Such EMs can be created from lean information sources, such
>> as the relative energy efficiency between CPUs. The ACPI based
>> platforms provide this information
>> (ACPI 6.4, s5.2.12.14 'GIC CPU Interface (GICC) Structure'
>> 'Processor Power efficiency Class' field).
>>
>> Artificial EMs might require to directly provide the 'cost' of
>> the generated performance state. This patch set adds a new callback
>> .get_cost() for this. The EM framework does not force any model
>> or formula, it's up to the platform code.
>>
>> Artificial EMs aim to leverage the Energy Aware Scheduler
>> (EAS). Other frameworks relying on performance states
>> information (i.e. IPA/DTPM) must be informed of the
>> EM type and might be prevented from using it. This patch
>> sets also does this by introducing a new flag:
>> EM_PERF_DOMAIN_ARTIFICIAL.
>>
>> The patch set is based on current linux-next, where some
>> changes to OPP & EM are queuing.
>>
>> The patch set also contains (patch 7/8 and patch 8/8) logic which prevents
>> two EM's client frameworks from using this new EM type. Some other approach,
>> using 'milli-watts', has been proposed and discussed, but refused [1].
>> This new flag is more precised and should not leave space for
>> wrong interpretation.
>>
>> Shortly after this patch set you will see a patch set implementing the
>> platform code and registering this new EM.
>>
> 
> Just to let you know that in the few days I'm going to post the first
> chunk of some SCMIv3.1 additions that includes also (as you probably
> know) the SCMI Perf protocol support for reporting perf_domain costs in
> micro-watts and not only in milli-watts.
> 
> Given that it does not seem that as of now the em_ API used by the SCMI
> cpufreq driver can make use of this new scale (and being not at all
> familiar with EM/EAS for sure :P), the SCMIv3.1 'Perf micro-watts' patch
> which I will post (I'll CC you) does NOT expose any new interface but only
> takes care to store the new micro-watts capability internally in a flag
> (if advertised by an SCMIv3.1 backend server), so that, basically, you'll
> keep seeing from the SCMI cpufreq driver that the scale is milli-watt
> (when milli-watts are used of course) or non-milli-watt (for abstract and
> micro-watts scales).

Sounds good!

> 
> This is intended to be of course a first step, laying out just the bare
> minimum commmon internal SCMI support, until we figure out how to properly
> expose this from the SCMI Perf in order to make it usable for EM.
> (if neeeded at all).
> 

I had such a patch for the EM, to keep the power in micro-Watts.
We can glue these two layers (high level EM and low layer SCMI
perf). Let's sort it out.

Regards,
Lukasz