support for software direct frequency throttling?

support for software direct frequency throttling?

[Mark Langsdorf]

Is there an easy way to dynamically limit the maximum or minimum
frequencies available to a governor? I'm trying to support frequency
throttling in response to a thermal event (or theoretically to allow a
BMC to set a minimum frequency regardless of current load).

I looked into doing this last year and it required some QoS patches that
I don't think ever made it into the kernel.

Is there a straightforward way to do this right now? If not, is anyone
working on implementing this? If not, any preferences from the
maintainers on how I should implement?

Thanks,
Mark Langsdorf
Calxeda, Inc.

Re: support for software direct frequency throttling?

[Rafael J. Wysocki]

On Wednesday, August 07, 2013 02:38:55 PM Mark Langsdorf wrote:
> Is there an easy way to dynamically limit the maximum or minimum
> frequencies available to a governor? I'm trying to support frequency
> throttling in response to a thermal event (or theoretically to allow a
> BMC to set a minimum frequency regardless of current load).
> 
> I looked into doing this last year and it required some QoS patches that
> I don't think ever made it into the kernel.
> 
> Is there a straightforward way to do this right now?

Have you looked at drivers/thermal/cpu_cooling.c ?

Rafael

Re: support for software direct frequency throttling?

[Mark Langsdorf]

On 08/07/2013 06:37 PM, Rafael J. Wysocki wrote:
> On Wednesday, August 07, 2013 02:38:55 PM Mark Langsdorf wrote:
>> Is there an easy way to dynamically limit the maximum or minimum
>> frequencies available to a governor? I'm trying to support frequency
>> throttling in response to a thermal event (or theoretically to allow a
>> BMC to set a minimum frequency regardless of current load).
>>
>> I looked into doing this last year and it required some QoS patches that
>> I don't think ever made it into the kernel.
>>
>> Is there a straightforward way to do this right now?
> 
> Have you looked at drivers/thermal/cpu_cooling.c ?

Yes, and I'm not sure how to use it get what I want.

I have a mailbox interrupt on my system that will trigger with a message
when the BMC monitoring the thermal sensor notices that the temperature
passes a trip point; the message will include the new maximum frequency.
I want to pass that new maximum frequency to the cpufreq governor so it
can reassess the current frequency against the maximum and lower it if
need be.

I don't know in advance what the trip points or maximum frequencies are;
the BMC managing the thermal sensor has that information but doesn't
provide it to Linux.

The most appealing solution would be to just copy
cpufreq_apply_cooling() and cpufreq_thermal_modifier() from
drivers/thermal/cpu_cooling.c in the highbank_cpufreq.c driver, as
they're the core logic of what I need. I don't know if that would be
acceptable, though. I've looked at setting up a full thermal_zone
solution, and its more than I need without actually giving me a good way
to set the frequencies when I need to.

--Mark Langsdorf
Calxeda, Inc.

Re: support for software direct frequency throttling?

[Rafael J. Wysocki]

On Thursday, October 17, 2013 12:39:22 PM Mark Langsdorf wrote:
> On 08/07/2013 06:37 PM, Rafael J. Wysocki wrote:
> > On Wednesday, August 07, 2013 02:38:55 PM Mark Langsdorf wrote:
> >> Is there an easy way to dynamically limit the maximum or minimum
> >> frequencies available to a governor? I'm trying to support frequency
> >> throttling in response to a thermal event (or theoretically to allow a
> >> BMC to set a minimum frequency regardless of current load).
> >>
> >> I looked into doing this last year and it required some QoS patches that
> >> I don't think ever made it into the kernel.
> >>
> >> Is there a straightforward way to do this right now?
> > 
> > Have you looked at drivers/thermal/cpu_cooling.c ?
> 
> Yes, and I'm not sure how to use it get what I want.
> 
> I have a mailbox interrupt on my system that will trigger with a message
> when the BMC monitoring the thermal sensor notices that the temperature
> passes a trip point; the message will include the new maximum frequency.
> I want to pass that new maximum frequency to the cpufreq governor so it
> can reassess the current frequency against the maximum and lower it if
> need be.
> 
> I don't know in advance what the trip points or maximum frequencies are;
> the BMC managing the thermal sensor has that information but doesn't
> provide it to Linux.
> 
> The most appealing solution would be to just copy
> cpufreq_apply_cooling() and cpufreq_thermal_modifier() from
> drivers/thermal/cpu_cooling.c in the highbank_cpufreq.c driver, as
> they're the core logic of what I need. I don't know if that would be
> acceptable, though. I've looked at setting up a full thermal_zone
> solution, and its more than I need without actually giving me a good way
> to set the frequencies when I need to.

I see.

Rui, any advice on that?

Rafael