lkml.org 
[lkml]   [2020]   [Jun]   [3]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
/
SubjectRe: v5.7: new core kernel option missing help text
From
Date


On 6/3/20 4:25 PM, Valentin Schneider wrote:
>
> On 03/06/20 20:58, Russell King - ARM Linux admin wrote:
>> On Wed, Jun 03, 2020 at 09:24:56PM +0200, Vincent Guittot wrote:
>>> On Wed, 3 Jun 2020 at 20:45, Russell King - ARM Linux admin
>>> <linux@armlinux.org.uk> wrote:
>>>> It's a start. I'm still wondering whether I should answer yes or no
>>>> for the platforms I'm building for.
>>>>
>>>> So far, all I've found is:
>>>>
>>>> arch/arm/include/asm/topology.h:#define arch_scale_thermal_pressure topology_get_thermal_pressure
>>>>
>>>> which really doesn't tell me anything about this. So I'm still in
>>>> the dark.
>>>>
>>>> I guess topology_get_thermal_pressure is provided by something in
>>>> drivers/ which will be conditional on some driver or something.
>>>
>>> You need cpufreq_cooling device to make it useful and only for SMP
>>> I don't think that this should not be user configurable because even
>>> with the description above, it is not easy to choose.
>>> This should be set by the driver that implement the feature which is
>>> only cpufreq cooling device for now it
>>
>> As I have CONFIG_CPU_FREQ_THERMAL=y in my config, I'm guessing (and it's
>> only a guess) that I should say y to SCHED_THERMAL_PRESSURE ?
>>
>
> arm and arm64 implement arch_scale_thermal_pressure(); the actual
> implementation is in the arch_topology "driver" (GENERIC_ARCH_TOPOLOGY).
>
> Then, the caller of arch_set_thermal_pressure() is cpufreq_cooling (see
> below); that'll only get called if you have thermal zones using CPU
> cooling devices.
>
> AFAICT the current state of things imply we should have something like
>
> depends on (ARM || ARM64) && GENERIC_ARCH_TOPOLOGY
>
> for that option.

Hi Russel/Valentin

The feature itself like Valentin explained below allows scheduler to be
aware of cpu capacity reduced due to thermal throttling.
arch_set_thermal_pressure feeds the capped capacity to the scheduler and
hence the feature makes sense only if arch_set_thermal_pressure is
implemented. Having said that arch_set_thermal_pressure is implemented
in arch_topology driver for arm and arm64 platforms. But the feature
itself is not bound to arm/arm64 platforms. So it would make it wrong to
add a "depends on (ARM || ARM64) option."

I agree with Vincent that allowing user to choose this option is
probably not the best. IMO, this should be enabled by default in arm64
defconfig considering both GENERIC_ARCH_TOPOLOGY and CPU_FREQ_THERMAL
are enabled by default.
So if it is acceptable three things to be done are:
1. Add the help text.
2. Don't allow SCHED_THERMAL_PRESSURE configurable by user
3. Enable it by default in arm64 defconfig

>
>>>>> + help
>>>>> + This option allows the scheduler to be aware of CPU thermal throttling
>>>>> + (i.e. thermal pressure), providing arch_scale_thermal_pressure() is
>>>>> + implemented.
>>
>> Is this feature documented in terms of what it does? Do I assume that
>> as the thermal trip points start tripping, that has an influence on
>> the scheduler? Or is it the case that the scheduler is wanting to
>> know when the cpu frequency changes?
>>
>> Grepping for "thermal" in Documentation/scheduler brings up nothing.
>
> The former; changing a CPU cooling device's state (IOW changing its max
> allowed frequency for thermal reasons) leads to a call to
> arch_set_thermal_pressure() (see
> cpufreq_cooling.c::cpufreq_set_cur_state()).
>
> It's somewhat interesting to have, at least in theory. On plain SMP that
> would let the scheduler see if some CPUs are more throttled that others,
> which would be leveraged when doing load balancing. It's more
> interesting for big.LITTLE & co, where in the worst cases we can have
> things like capacity inversion, i.e. the bigs are so thermally throttled
> that they give less oomf than a LITTLE.
>

--
Warm Regards
Thara

\
 
 \ /
  Last update: 2020-06-04 02:49    [W:0.065 / U:0.196 seconds]
©2003-2020 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site