[PATCH] arm: exynos: Modify pm code to check for cortex A9 rather than the SoC

[Tomasz Figa]

On 24.06.2014 18:30, Russell King - ARM Linux wrote:
> On Tue, Jun 24, 2014 at 06:20:56PM +0200, Tomasz Figa wrote:
>> Hi Russell,
>>
>> On 24.06.2014 18:11, Russell King - ARM Linux wrote:
>>> On Mon, Jun 16, 2014 at 09:37:14AM +0530, Abhilash Kesavan wrote:
>>>> Hi Kukjin,
>>>>
>>>> On Fri, May 23, 2014 at 8:31 AM, Abhilash Kesavan <a.kesavan at samsung.com> wrote:
>>>>> Signed-off-by: Abhilash Kesavan <a.kesavan at samsung.com>
>>>>
>>>> Do you have any comments on this patch ?
>>>
>>> I do.
>>>
>>>>> diff --git a/arch/arm/mach-exynos/pm.c b/arch/arm/mach-exynos/pm.c
>>>>> index d10c351..6dd4a11 100644
>>>>> --- a/arch/arm/mach-exynos/pm.c
>>>>> +++ b/arch/arm/mach-exynos/pm.c
>>>>> @@ -300,7 +300,7 @@ static int exynos_pm_suspend(void)
>>>>>         tmp = (S5P_USE_STANDBY_WFI0 | S5P_USE_STANDBY_WFE0);
>>>>>         __raw_writel(tmp, S5P_CENTRAL_SEQ_OPTION);
>>>>>
>>>>> -       if (!soc_is_exynos5250())
>>>>> +       if (read_cpuid_part_number() == ARM_CPU_PART_CORTEX_A9)
>>>>>                 exynos_cpu_save_register();
>>> ...
>>>>> @@ -334,7 +334,7 @@ static void exynos_pm_resume(void)
>>>>>         if (exynos_pm_central_resume())
>>>>>                 goto early_wakeup;
>>>>>
>>>>> -       if (!soc_is_exynos5250())
>>>>> +       if (read_cpuid_part_number() == ARM_CPU_PART_CORTEX_A9)
>>>>>                 exynos_cpu_restore_register();
>>>
>>> It is invalid to check just the part number.  The part number on its
>>> own is meaningless without taking account of the implementor.  Both
>>> the implementor and the part number should be checked at each of these
>>> sites.
>>
>> Just out of curiosity, are you aware of more than one implementor of
>> Cortex A9 on Exynos SoCs that would differ in having the need for save
>> and restore of those registers?
> 
> That doesn't stop stuff changing in the future.  We've been here before.
> Let's do the job properly if we're going to do the job at all.

I tend to disagree. The chance of a new Cortex A9 based SoC with
different implementor code showing up is so close to zero that I don't
see increasing of code complexity by adding yet another check justified.

> 
> If people still whine about this, I will force a change to make it
> harder to do the wrong thing - I will get rid of the _part_number
> interface replacing it with one which always returns the implementor
> as well as the part number so both have to be checked.

That might be actually a better option. Something like

	if (read_cpuid_impl_and_part() == ARM_CPU(impl, part))

or even

	if (ARM_CPU_IS(impl, part))

would keep the checks simple, while still checking both values.

> 
>>> Another point: exynos have taken it upon themselves to add code which
>>> saves various ARM core registers.  This is a bad idea, it brings us
>>> back to the days where every platform did their own suspend implementations.
>>>
>>> CPU level registers should be handled by CPU level code, not by platform
>>> code.  Is there a reason why this can't be added to the Cortex-A9
>>> support code in proc-v7.S ?
>>
>> I agree that there is nothing platform specific in saving and restoring
>> those registers and that this should be probably handled by generic code.
>>
>> However, when running in non-secure world, the only way to restore this
>> is to call a firmware operation, which is platform specific. Is there a
>> way to do something like this from proc-v7.S?
> 
> We never call firmware operations from assembly code.  However, in exynos'
> case, it's not running in non-secure mode because it's happily reading
> and writing these registers with no issue.

Current version of code, yes. However it handles only Exynos-based
boards running in secure mode. For those running in non-secure mode,
mainline does not support sleep yet.

I already have patches to change this, which I was planning to post
after eliminating last issue. The change set includes making this
save/restore being executed only in secure mode.

> 
> Platforms running in non-secure mode already have to ensure that various
> work-arounds are implemented in their firmware/boot loader, and this
> really is no different (we wouldn't need this code in the kernel in the
> first place if the firmware/boot loader would get its act together.)

In ideal world (which I would be glad to be living in)...

We both know that we can't fully rely on firmware. Firmware bugs are
common and in many cases we can't do anything about it, because it
already comes with the device and in many cases it is undesirable to
change it or it can't be done at all.

> 
>>> And presumably with the CPU level code dealing with those registers,
>>> you don't need the calls to save and restore these registers in this
>>> notifier.
>>>
>>> Which, by the way, is probably illegal to run as it runs in a read-
>>> lock code path and with the SCU disabled.  As you're calling
>>> scu_enable() that means you're non-coherent with the other CPUs,
>>> and therefore locks don't work.
>>
>> I don't see the read lock code path you mention. Could you elaborate on
>> this? By the way, other CPUs are still offline at this point.
> 
> We get there from kernel/cpu_pm.c, when the notifier chain is called.
> The notifier chain is called while taking a read lock on
> cpu_pm_notifier_lock.

Your point. Thanks for explaining this. However this will be still
running with just one CPU online.

> 
> If this is about the last CPU going down, then why the notifier?  Why
> not put the code in exynos_suspend_enter() ?  Why add this needless
> complexity?
> 

This code is used by both system-wide suspend/resume and cpuidle paths.
Daniel has moved this code to CPU PM notifier as a part of his Exynos
cpuidle consolidation series to avoid code duplication, as this is the
common point of both paths.

To clarify, on suspend/resume path, the notifier is being called from
syscore_ops registered in kernel/cpu_pm.c, while on cpuidle path, this
is invoked from exynos_enter_core0_aftr() in
drivers/cpuidle/cpuidle-exynos.c, which calls cpu_pm_enter().

Best regards,
Tomasz