[PATCH] i2c: exynos5: Properly use the "noirq" variants of suspend/resume

[Doug Anderson]

Tomasz,

On Mon, Jun 23, 2014 at 3:31 PM, Tomasz Figa <tomasz.figa at gmail.com> wrote:
>
>
> On 24.06.2014 00:27, Doug Anderson wrote:
>> Kevin,
>>
>> On Mon, Jun 23, 2014 at 3:19 PM, Kevin Hilman <khilman at linaro.org> wrote:
>>> Doug Anderson <dianders at chromium.org> writes:
>>>
>>> [...]
>>>
>>>> On Fri, Jun 20, 2014 at 4:59 PM, Tomasz Figa <tomasz.figa at gmail.com> wrote:
>>>>>
>>>>> I'm not sure noirq is going to work correctly, at least not with current
>>>>> callbacks. I can see a call to clk_prepare_enable() there which needs to
>>>>> acquire a mutex.
>>>>
>>>> Nice catch, thanks!  :)
>>>>
>>>> OK, looking at that now.  Interestingly this doesn't seem to cause us
>>>> problems in our ChromeOS 3.8 tree.  I just tried enabling:
>>>>   CONFIG_DEBUG_ATOMIC_SLEEP=y
>>>>
>>>> ...and confirmed that I got it on right:
>>>>
>>>> # zgrep -i atomic /proc/config.gz
>>>> CONFIG_DEBUG_ATOMIC_SLEEP=y
>>>>
>>>> I can suspend/resume with no problems.  My bet is that it works fine because:
>>>>
>>>> * resume_noirq is not considered "atomic" in the sense enforced by
>>>> CONFIG_DEBUG_ATOMIC_SLEEP (at least not in 3.8--I haven't tried on
>>>> ToT)
>>>
>>> The reason is because "noirq" in the suspend/resume path actually means
>>> no *device* IRQs for that specific device.
>>>
>>> It's often assumed that the "noirq" callbacks are called with *all*
>>> interrupts disabled, but that's not the case.  Only the IRQs for that
>>> specific device are disabled when its noirq callbacks run.
>>
>> Ah, so even with my fix of moving to noirq we could still be broken if
>> the system decided to enable interrupts for the device before the i2c
>> controller get resumed then we'd still be SOL.
>>
>> ...oh, but if it matches probe order then maybe we're guaranteed for
>> that not to happen?  We know that we will probe the i2c bus before the
>> devices on it, right?
>
> If the mentioned device is a child of the I2C controller then the
> parent-child relation determines the order. Otherwise (e.g. another,
> non-I2C interrupt source that just triggers some operation on an I2C
> device like voltage regulator) we're doomed. ;)

Wow, that would be seriously screwed up.

OK, so to summarize my current plans: I won't spin this patch and we
can see what Wolfram thinks.  It may not be as beautiful as Kevin's
suggestion to use Runtime PM but I also don't think it's insane.
...and I've got a request in to Samsung to use Runtime PM in the long
run.

If anyone at Samsung working on suspend/resume on exynos5420-pit or
exynos5800-pi wants to add their Tested-by (or bug reports) I'm sure
that would be appreciated.

-Doug