linux-kernel - Re: [Patch v9 7/8] sched/fair: Enable tuning of decay period

lists.openwall.net		lists / announce owl-users owl-dev john-users john-dev passwdqc-users yescrypt popa3d-users / oss-security kernel-hardening musl sabotage tlsify passwords / crypt-dev xvendor / Bugtraq Full-Disclosure linux-kernel linux-netdev linux-ext4 linux-hardening linux-cve-announce PHC
Open Source and information security mailing list archives

Hash Suite: Windows password security audit tool. GUI, reports in PDF.

[<prev] [next>] [<thread-prev] [day] [month] [year] [list]

Date:   Wed, 19 Feb 2020 10:14:47 +0100
From:   Dietmar Eggemann <dietmar.eggemann@....com>
To:     Thara Gopinath <thara.gopinath@...aro.org>,
        Peter Zijlstra <peterz@...radead.org>
Cc:     Randy Dunlap <rdunlap@...radead.org>, mingo@...hat.com,
        ionela.voinescu@....com, vincent.guittot@...aro.org,
        rui.zhang@...el.com, qperret@...gle.com, daniel.lezcano@...aro.org,
        viresh.kumar@...aro.org, rostedt@...dmis.org, will@...nel.org,
        catalin.marinas@....com, sudeep.holla@....com,
        juri.lelli@...hat.com, corbet@....net,
        linux-kernel@...r.kernel.org, amit.kachhap@...il.com,
        javi.merino@...nel.org, amit.kucheria@...durent.com
Subject: Re: [Patch v9 7/8] sched/fair: Enable tuning of decay period

On 18/02/2020 15:57, Thara Gopinath wrote:
> 
> 
> On 2/14/20 5:26 AM, Dietmar Eggemann wrote:
>> On 13/02/2020 14:54, Thara Gopinath wrote:
>>> On 02/10/2020 06:59 AM, Dietmar Eggemann wrote:
>>>> On 07/02/2020 23:42, Thara Gopinath wrote:
>>>>> On 02/04/2020 03:39 AM, Dietmar Eggemann wrote:
>>>>>> On 03/02/2020 16:55, Peter Zijlstra wrote:
>>>>>>> On Mon, Feb 03, 2020 at 07:07:57AM -0500, Thara Gopinath wrote:
>>>>>>>> On 01/28/2020 06:56 PM, Randy Dunlap wrote:
>>>>>>>>> Hi,
>>>>>>>>>
>>>>>>>>> On 1/28/20 2:36 PM, Thara Gopinath wrote:

[...]

>>>> Cpu-invariant accounting can't be guarded with a kernel CONFIG switch.
>>>> Frequency-invariant accounting could be with CONFIG_CPU_FREQ but
>>>> this is
>>>> enabled by default by Arm64 defconfig.
>>>> Thermal pressure (accounting) (CONFIG_HAVE_SCHED_THERMAL_PRESSURE) is
>>>> disabled by default so why should a per-cpu thermal_pressure be
>>>> maintained on such a system (CONFIG_CPU_THERMAL=y by default)?
>>>
>>> I agree that there is no need for per-cpu thermal pressure to be
>>> maintained if no averaging is happening in the scheduler, today. I don't
>>> know if there will ever be an use for it.
>>
>> All arch_scale_FOO() functions follow the approach to force the arch
>> (currently x86, arm, arm64) to do
>>
>> #define arch_scale_FOO BAR
>>
>> to enable the FOO functionality.
>>
>> There is no direct link between consumer and provider here.
>>
>>   consumer (sched) -> arch <- provider (arch, counters, CPUfreq, CPU
>>                                         cooling, etc.)
>>
>> So IMHO, FOO=thermal_pressure should follow this design pattern too.
>>
>> 'thermal_pressure' would be the only one which can be disabled by a
>> kernel config switch at the consumer side.
>> IMHO, it doesn't make sense to have the provider operating in this case.
>>
>>> My issue has to do with using a config option meant for internal
>>> scheduler code being used else where. To me, once this happens, the
>>> entire work done to separate out reading and writing of instantaneous
>>> thermal pressure to arch_topology makes no sense. We could have kept it
>>> in scheduler itself.
>>
>> You might see thermal_pressure more on the level of irq_load or
>> [rt/dl]_rq_load and that could be why we have a different opinion here?
>>
>> Now rt_rq_load and dl_rq_load are scheduler internal providers and
>> irq_load is driven by 'irq_delta + steal' time (which is much closer to
>> the scheduler than thermal for instance).
> 
> In this case, thermal pressure is quite close to scheduler as it reduces
> the maximum capacity available per cpu and hence affects scheduler
> placement of tasks
> 
>>
>> My assumption is that we don't want a direct link between the scheduler
>> and e.g. a provider 'thermal'.
> 
> Exactly. Which is why the same CONFIG option should not be used between
> the provider and consumer.

I think there is a little misunderstanding here. By being close to the
scheduler I was referring to rt, dl, irq which are not driven via an
arch_scale_FOO function.

But I guess we agree that FOO=thermal_pressure should use this
arch_scale_FOO function so we don't have a direct link between scheduler
and thermal subsystem.

We disagree in the point whether the provider should be present and
working when the only consumer is disabled by the kernel config.

I guess we can't discuss the technical angle of this issue any further
so maybe the maintainer of drivers/base/arch_topology.c should make a
decision (the actual code is in 3/8 of this patch-set).

>>> Another way I think about this whole thermal pressure framework  is that
>>> it is the job of cooling device or cpufreq or any other entity to update
>>> a throttle in maximum pressure to the scheduler. It should be
>>> independent of what scheduler does with it. Scheduler can choose to
>>> ignore it