Hi Daniel,

Just for test you may try removing the flush_cache_louis(); from __cpu_suspend_save .

Because if the driver is using cpu_suspend() in idle path then almost all the important data on core L1 is clean.

Still it can fail iff the code after that ( cpu_suspend()) modifies some cacheable important data because SCTLR.C bit

is not cleared yet.

It should just fail if no cache clean is done before powering down.

And it should work iff v7_exit_coherency_flush() or similar is done.

Note: This is based on my observation on A7 quad. Please correct me if the understanding is wrong.

Thanks

Sandeep

On 9 April 2014 13:24, Daniel Lezcano <daniel.lezcano@linaro.org> wrote:

Hi Sandeep,

On 04/09/2014 07:15 AM, Sandeep Tripathy wrote:

Hi Daniel,
- L1 D$ clean ( SCTLR C bit clear, DCCISW, clear SMP) is part
of the recommended sequence for Individual core power down .

Yes, absolutely. It is what Lorenzo and I discussed.

The macro v7_exit_coherency_flush should be used for that.

- If a core is powered down having dirty lines in L1 then the
system should encounter an issue (abort) very easily. May be the first
idle attempt itself is sufficient to break things.

Does any platform (exynos4 ?) work without doing L1 D cache
clean in cpu idle individual core power down sequence ?

Yes, the power down sequence is done but without L1 D cache clean and it is very very hard to make the board hang. It is so rare, I can't say 100% it is related to the driver itself or something else.

If there is a way to spot the issue, I will be happy to test it.

Thanks
-- Daniel

On 8 April 2014 17:07, Daniel Lezcano <daniel.lezcano@linaro.org
<mailto:daniel.lezcano@linaro.org>> wrote:

On 04/08/2014 01:07 PM, Amit Kucheria wrote:

Hi Daniel,

Have you noticed this on any platform yet with this test?

I have noticed a very very rare hang on the exynos4 board with this
test and the dual cpu support but it is not reproducible enough to
check if the cache flush fixes it or not (or it is related to the
cpuidle driver). I had a long discussion yesterday with Lorenzo who
explained me what could happen without flushing the cache and
exiting the SMP mode.

AFAICT,

* The tc2 flushes its cache.

* exynos5 still need to disable cpu1 to enter the AFTR state (which
is broken today), so cache is flushed in the hotplug code path. I
hope I can spot the issue with a quad core.

* omap4 flushes its cache.

* omap3 is not concerned by this because it is an UP system.

* vinci, s3c64, imx5, imx6, ux500, kirkwood does not support cpu
power down.

* calxeda hides that through the firmware I believe

* I don't know for tegra but I assume they are flushing and
disabling the cache

I hope with this test we can spot the issue, if any, with multiple
runs on the boards, especially when new drivers will be implemented.

--
<http://www.linaro.org/> Linaro.org │ Open source software for ARM
SoCs

Follow Linaro: <http://www.facebook.com/__pages/Linaro
<http://www.facebook.com/pages/Linaro>> Facebook |
<http://twitter.com/#!/__linaroorg
<http://twitter.com/#!/linaroorg>> Twitter |
<http://www.linaro.org/linaro-__blog/
<http://www.linaro.org/linaro-blog/>> Blog

_________________________________________________
linaro-dev mailing list
linaro-dev@lists.linaro.org <mailto:linaro-dev@lists.linaro.org>
http://lists.linaro.org/__mailman/listinfo/linaro-dev
<http://lists.linaro.org/mailman/listinfo/linaro-dev>

--
<http://www.linaro.org/> Linaro.org │ Open source software for ARM SoCs

Follow Linaro: <http://www.facebook.com/pages/Linaro> Facebook |
<http://twitter.com/#!/linaroorg> Twitter |
<http://www.linaro.org/linaro-blog/> Blog