Hi Lorenzo,
On 13 May 2014 22:44, Lorenzo Pieralisi <lorenzo.pieral...@arm.com> wrote:
> On Tue, May 13, 2014 at 12:43:31PM +0100, Chander Kashyap wrote:
>
> [...]
>
>> >> +static void exynos_suspend(u64 residency)
>> >> +{
>> >> + unsigned int mpidr, cpunr;
>> >> +
>> >> + mpidr = read_cpuid_mpidr();
>> >> + cpunr = exynos_pmu_cpunr(mpidr);
>> >
>> > If I were to be picky, I would compute these values only if they
>> > are needed, ie move the computation after exynos_power_down().
>>
>> Yes thats make sense. I will realign it.
>>
>> >
>> > There is another quite horrible issue here. We know this code works
>> > because the processors A15/A7 hit the caches with C bit in SCTLR cleared.
>> >
>> > On processors where this is not true, this sequence would explode
>> > if power down fails (in case core is gated but L2 is still powered on,
>> > the stack is stuck in L2) since it is going to read stack data that is
>> > in L2 but can't be read.
>> >
>> > It is not related to this sequence only, but it is an issue in general
>> > and wanted to mention that on the lists for public awareness.
>> >
>>
>> Can you please elaborate. I didn't understand.
>
> It is not related to this patch only. This function carries out writes to the
> stack (which might end up in eg L2) and then disables the C bit in SCTLR
> through MCPM.
>
> A15 and A7 processors hit the cache with the C bit clear in the SCTLR
> so the processor still "hits" the stack values if the power down fails.
> On processors where caches are not hit with the C bit clear (eg A9) this code
> would fail since the stack values that sit in the caches cannot be read with
> the C bit clear in SCTLR until the SCTLR is restored, so it will have to
> be implemented in assembly with no stack usage (or better, no cacheable data
> usage).
>
> So, all I am saying is, to avoid copy'n'paste havoc and to avoid running
> this code on Exynos platforms where it must not be run as-is, please add
> a comment along the line:
>
> "This function requires the stack data to be visible through power down
> and can only be executed on processors like A15 and A7 that hit the cache
> with the C bit clear in the SCTLR register."
>
> Please let me know if that's clear.
It all clear now.
Thanks a lot.
>
> Lorenzo
>
>>
>> > The gist of what I am saying is, please add a comment to that extent,
>> > here and it should be added in exynos_power_down() too.
>> >
>> >> + __raw_writel(virt_to_phys(mcpm_entry_point), ns_sram_base_addr +
>> >> 0x1c);
>> >
>> > No magic numbers please (0x1c). You can add a macro/wrapper, as TC2 does.
>>
>> Yes i will remove it.
>>
>> >
>> >> + exynos_power_down();
>> >> +
>> >> + /*
>> >> + * Execution reaches here only if cpu did not power down.
>> >> + * Hence roll back the changes done in exynos_power_down function.
>> >> + */
>> >> + exynos_cpu_powerup(cpunr);
>> >
>> > Please be aware that if this function returns MCPM will soft reboot, and
>> > the CPUidle driver will have no way to detect a state entry failure.
>> >
>> > I am just flagging this up, since fixing this behaviour is not easy, and
>> > honestly, since power down failure should be the exception not the rule,
>> > the idle stats should not be affected much.
>> >
>> > I think this is the proper way of implementing the sequence but please
>> > all keep in mind what I wrote above.
>> >
>> > Lorenzo
>> >
>> >> +}
>> >> +
>> >> static const struct mcpm_platform_ops exynos_power_ops = {
>> >> .power_up = exynos_power_up,
>> >> .power_down = exynos_power_down,
>> >> .power_down_finish = exynos_power_down_finish,
>> >> + .suspend = exynos_suspend,
>> >> + .powered_up = exynos_powered_up,
>> >> };
>> >>
>> >> static void __init exynos_mcpm_usage_count_init(void)
>> >> --
>> >> 1.7.9.5
>> >>
>> >>
>> >
>>
>>
>>
>> --
>> with warm regards,
>> Chander Kashyap
>>
>
--
with warm regards,
Chander Kashyap
--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majord...@vger.kernel.org
More majordomo info at http://vger.kernel.org/majordomo-info.html
Please read the FAQ at http://www.tux.org/lkml/
