linux-kernel - Re: [PATCH V2] ARM: cpuidle: Support asymmetric idle definition

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

Message-ID: <47e3d236-8d36-d04d-5f87-9fece6875d66@arm.com>
Date:   Fri, 2 Jun 2017 11:14:18 +0100
From:   Sudeep Holla <sudeep.holla@....com>
To:     Daniel Lezcano <daniel.lezcano@...aro.org>, rjw@...ysocki.net,
        lorenzo.pieralisi@....com
Cc:     Sudeep Holla <sudeep.holla@....com>, leo.yan@...aro.org,
        "open list:CPUIDLE DRIVERS" <linux-pm@...r.kernel.org>,
        open list <linux-kernel@...r.kernel.org>
Subject: Re: [PATCH V2] ARM: cpuidle: Support asymmetric idle definition



On 02/06/17 11:06, Daniel Lezcano wrote:
> On 02/06/2017 11:39, Sudeep Holla wrote:
>>
>>
>> On 02/06/17 10:25, Daniel Lezcano wrote:
>>> On 02/06/2017 11:20, Sudeep Holla wrote:
>>>>
>>>>
>>>> On 01/06/17 12:39, Daniel Lezcano wrote:
>>>>> Some hardware have clusters with different idle states. The current code does
>>>>> not support this and fails as it expects all the idle states to be identical.
>>>>>
>>>>> Because of this, the Mediatek mtk8173 had to create the same idle state for a
>>>>> big.Little system and now the Hisilicon 960 is facing the same situation.
>>>>>
>>>>> Solve this by simply assuming the multiple driver will be needed for all the
>>>>> platforms using the ARM generic cpuidle driver which makes sense because of the
>>>>> different topologies we can support with a single kernel for ARM32 or ARM64.
>>>>>
>>>>> Every CPU has its own driver, so every single CPU can specify in the DT the
>>>>> idle states.
>>>>>
>>>>> This simple approach allows to support the future dynamIQ system, current SMP
>>>>> and HMP.
>>>>>
>>>>> It is unoptimal from a memory point of view for a system with a large number of
>>>>> CPUs but nowadays there is no such system with a cpuidle driver on ARM.
>>>>>
>>>>
>>>> While I agree this may be simple solution, but just not necessary for
>>>> systems with symmetric idle states especially one with large number of
>>>> CPUs. I don't like to see 96 CPU Idle driver on say ThunderX. So we
>>>> *must* have some basic distinction done here.
>>>>
>>>> IMO, we can't punish a large SMP systems just because they don't have
>>>> asymmetric idle states.
>>>
>>> Can you point me in the upstream kernel a DTS with 96 cpus and using the
>>> cpuidle-arm driver ?
>>>
>>
>> The bindings are upstream right. Not all DTS are upstream, firmware
>> generate them especially for large systems.
>>
>> Check arch/arm64/boot/dts/cavium/thunder{,2}-{88,99}xx.dtsi, it has
>> supports PSCI and firmware can update DTB to add the idle states.
>> They are systems with 96 and 128 CPUs.
> 
> Ok, so I have to assume there are out of tree DTB with idle state
> definitions. In other circumstances I would have just ignored this
> argument but I can admit the DTB blob thing is in the grey area between
> what we have to support upstream and out of tree changes.
> 

Not entirely true. It's clear, we support anything whose binding is
upstream. I do agree that there are lots of out of tree bindings but I
am not referring to them. I am just referring to out of tree DTS with
already upstreamed binding.

> However, even if it is suboptimal, I'm convinced doing a per-cpu driver
> makes more sense.
> 

OK, but I think a simple check to decide not to, on SMP system is not
too much a ask ?

> You said, we punish large SMP systems, I don't think it is true,
> especially from a performance point of view. Do you have access to such
> hardware to check if it is correct?
> 

Sorry, I thought "punish" is bit a harsh but couldn't find the work. I
can say penalize instead.

No I don't have such a system to test.

> We have more memory used but it is not related on how we implement the
> driver above but on the fixed array size in the cpuidle structures.
> 

Yes, I agree. But I will see what we can do at minimum to avoid it on
system's that don't need it.

> I can take care of optimizing the memory usage in the cpuidle framework
> which will benefit all architectures but that will take some time.
> 

Yest that's always good to have but I was not referring to that.

> So can you consider acking this patch, so that unblocks other HMP idle
> development and I take care of optimizing the structure memory usage?
> 

I am thinking of checking DT to see if we need multiple driver support
or not. But then it comes down to what Lorenzo had already posted.

-- 
Regards,
Sudeep