Re: [PATCH] mmc: sdhci-of-at91: fix wakeup issue when using runtime pm

From: Ulf Hansson
Date: Tue Mar 08 2016 - 16:54:22 EST

On 4 March 2016 at 14:48, Ludovic Desroches <ludovic.desroches@xxxxxxxxx> wrote:
> Hi Ulf,
> On Fri, Mar 04, 2016 at 10:09:37AM +0100, Ulf Hansson wrote:
>> On 17 February 2016 at 11:35, Ludovic Desroches
> [snip]
>> > I am wondering if I should take account of sdio irq enabled or not here.
>> >
>> > I have a sdio device which drives me crazy because of power management.
>> > The driver of this device is in staging, it is wilc1000. It seems that I
>> > am stuck because the sdio irq are not received. If I don't disable the
>> > clock of the controller (hclock), I should receive the sdio IRQ as I
>> > receive card detect ones, isn't it?
>> >
>> > It doesn't work, it seems I have also to not disabled mainck and gck
>> > which are clocks needed to generate the clock sent to the sdio device.
>> > If none of the clocks have to be disabled, where it has to be managed?
>> That's a typical issue for SDIO IRQs, especially when the controller
>> HW manages IRQs (there are other ways to deal with SDIO IRQs as well).
>> Currently, the simplest way to deal with this in the driver is to do a
>> pm_runtime_get_sync() when the SDIO IRQ gets enabled, and
>> pm_runtime_put() when it gets disabled.
> Which driver? sdio device driver or controller driver?

In the mmc controller driver.

>> >
>> > Do I have to anticipate this use case in the driver of my sdhci
>> > controller or does it have to be managed in the sdio device driver? They
>> > are using sdio_claim/release_host to suspend or resume the host but
>> > maybe they use it in a bad way.
>> The wilc100 SDIO func driver should *not* keep the host claimed to
>> deal with SDIO irqs. Only when it configures them.
>> Instead, you need to deal with this in the sdhci driver, when you get
>> the call to enable/disable SDIO IRQs.
> Do you mean in sdhci_enable_sdio_irq?

I am not sure exactly where to check. As it may be depending on the
sdhci variant and the SoC you probably need to deal with this in
non-common sdhci code.

>> Moreover, from a system PM point of view. If the wilc100 SDIO func
>> driver wants the platform to wake up on SDIO IRQs, it needs to set
>> MMC_PM_KEEP_POWER and MMC_PM_WAKE_SDIO_IRQ from its ->suspend()
>> callback.
>> In that way, your sdhci driver can act accordingly from its system PM
>> callbacks. In other words, depending on MMC_PM_KEEP_POWER and
>> MMC_PM_WAKE_SDIO_IRQ to *not* call pm_runtime_force_suspend().
> Ok, it makes sense.
> Coming back to the initial discussion and patch which were motivated by
> the fact that after runtime suspend I can't wake-up on card detect event,
> I have the feeling we don't have the same assumption about runtime PM.
> From what you and Adrian told me, I should not use runtime PM if I have
> no way to wake-up. In your minds, the way to wake-up is to use an
> externel GPIO because the controller will be 'totally' disabled, isn't it?

I agree to the first part here.

Although, as you also have the option to use polling for card detect,
this actually means you don't really *need* to have a wakeup
configured. Especially in the case where you don't have GPIO card

In that way, *all* the clocks can gated in between the polling
attempts, thus you will save power even in the polling mode
configuration and when runtime PM is enabled.

> On my side, runtime PM allows me to save power when the sdhci controller
> is not used. If I can disable two clocks out of three, I should use
> runtime PM. Do you agree?
> If not, tell me how I can convince you :) Otherwise, next step is to rework
> my patch but I think I have no other solution that not calling
> sdhci_runtime_suspend_host if I expect to use the card detect irq of the
> controller.

So, to summarize.

I think the best fix is to add a clever check in ->probe() and then
enable polling when you can't rely on GPIO card detect IRQ.

Moreover, to have a robust solution, you also need to clear
SDHCI_QUIRK_BROKEN_CARD_DETECTION for your sdhci variant, as otherwise
the "broken-cd" DT binding could wrongly be used for this variant.

Does that make sense?

Kind regards