Re: [PATCH v2 20/28] ARM: tegra: cpufreq: Disable cpufreq during suspend

[Colin Cross <ccross@android.com>]

On Mon, Jan 24, 2011 at 11:35 AM, Mark Brown
<broonie@opensource.wolfsonmicro.com> wrote:
> On Mon, Jan 24, 2011 at 10:50:53AM -0800, Colin Cross wrote:
>> On Mon, Jan 24, 2011 at 6:41 AM, Mark Brown
>
>> > Hrm, what's the situation where that happens and why does it cause
>> > problems?  The regulator API doesn't care if suspend is going on, and
>> > nor do any of the current drivers for regulators.  There is an issue
>> > with keeping things like I2C alive until the bitter end of suspend so
>> > you've got a control bus to the regulators but that's a generic issue
>> > which crops up with other subsystems too so a regulator-specific
>> > workaround seems dodgy.
>
>> The end result is that the regulator driver calls the i2c driver after
>> the i2c driver has been suspended.  On Tegra, this happens because of
>> voltage scaling on the CPU regulator, which can be on the i2c bus.  A
>> sleep in a suspend handler after the i2c bus has been suspended can
>> cause the cpufreq governor to lower the frequency, and try to lower
>> the voltage as well.
>
> OK, that's a general problem - you need to ensure that the I2C and SPI
> controllers are suspended really late.  The OMAP guys also have this
> problem for some RTCs, though in their case things are slightly
> different due to their use of runtime PM.  Looking at your I2C driver
> (BTW, I'd suggest reminding Ben on a more regular basis about that) I
> suspect that moving your suspend and resume callbacks to the _noirq
> varaints will cover a lot of this.

Even _noirq isn't late enough, if cpufreq keeps trying to change the
frequency (and thus voltage) until sysdev suspend.

>> The problem isn't limited to i2c busses, there are some regulators on
>> spi busses.  You would need to ensure that any driver that has a
>
> For all practical purposes in this sort of discussion you can typically
> do a s/I2C/I2C and SPI/ - in terms of system integration issues they're
> very similar.

True

>> regulator consumer suspends before the regulator driver it
>> communicates with, and that still wouldn't fix cpufreq, which has a
>> sysdev suspend handler.
>
> So you actively need to push the processor into low power mode during
> suspend?  I'm still not 100% clear what's triggering the issues you're
> seeing and why this seems to be Tegra-only.  If there were a general
> cpufreq/regulator interaction here I'd expect to see all ARM cpufreq
> drivers needing to do the same thing (and probably some handling of this
> in cpufreq core as a result).  If it's not such an issue I'd expect
> there's also entertaining suspend ordering issues elsewhere.

It's more of a voltage scaling issue than a cpufreq issue directly.
Tegra requires the voltage be set to nominal during resume, and the
only time it can be set for resume is before I2C suspends.  I handle
the problem with a suspend notifier in the latest version of the clock
voltage scaling patches, but I kept this patch to avoid cpufreq trying
to go to frequencies that are not supported by the suspend voltage.

> Typically this stuff isn't a problem for regulators themselves since if
> they're active by the time I2C is off they normally get suspended by
> hardware handshakes from the CPU as that enters low power mode - right
> now we have no regulators at all in mainline that do anything in
> software on suspend.  We can get a long way by just ignoring what
> happens to regulators over suspend (there is some work to do here but
> it's orthogonal to this sort of issue).

The regulator driver itself has nothing to do, and when the CPU enters
its lowest power mode it signals the regulator to turn off, but
something needs to tell the regulator to go to the correct voltage.