Re: [PATCH v7] mmc: sdhci-of-dwcmshc: Add runtime PM operations

[Adrian Hunter <adrian.hunter@intel.com>]

On 10/08/23 19:34, Ulf Hansson wrote:
> On Thu, 10 Aug 2023 at 14:44, Adrian Hunter <adrian.hunter@intel.com> wrote:
>>
>> On 10/08/23 13:21, Ulf Hansson wrote:
>>> On Thu, 10 Aug 2023 at 10:13, Adrian Hunter <adrian.hunter@intel.com> wrote:
>>>>
>>>> On 8/08/23 23:23, Liming Sun wrote:
>>>>> This commit implements the runtime PM operations to disable eMMC
>>>>> card clock when idle.
>>>>>
>>>>> Reviewed-by: David Thompson <davthompson@nvidia.com>
>>>>> Signed-off-by: Liming Sun <limings@nvidia.com>
>>>>> ---
>>>>> v6->v7:
>>>>>     - Address Ulf's comment;
>>>>> v5->v6:
>>>>>     - Address Adrian's more comments and add coordination between
>>>>>       runtime PM and system PM;
>>>>> v4->v5:
>>>>>     - Address Adrian's comment to move the pm_enable to the end to
>>>>>       avoid race;
>>>>> v3->v4:
>>>>>     - Fix compiling reported by 'kernel test robot';
>>>>> v2->v3:
>>>>>     - Revise the commit message;
>>>>> v1->v2:
>>>>>     Updates for comments from Ulf:
>>>>>     - Make the runtime PM logic generic for sdhci-of-dwcmshc;
>>>>> v1: Initial version.
>>>>> ---
>>>>>  drivers/mmc/host/sdhci-of-dwcmshc.c | 72 ++++++++++++++++++++++++++++-
>>>>>  1 file changed, 70 insertions(+), 2 deletions(-)
>>>>>
>>>>> diff --git a/drivers/mmc/host/sdhci-of-dwcmshc.c b/drivers/mmc/host/sdhci-of-dwcmshc.c
>>>>> index e68cd87998c8..c8e145031429 100644
>>>>> --- a/drivers/mmc/host/sdhci-of-dwcmshc.c
>>>>> +++ b/drivers/mmc/host/sdhci-of-dwcmshc.c
>>>>> @@ -15,6 +15,7 @@
>>>>>  #include <linux/module.h>
>>>>>  #include <linux/of.h>
>>>>>  #include <linux/of_device.h>
>>>>> +#include <linux/pm_runtime.h>
>>>>>  #include <linux/reset.h>
>>>>>  #include <linux/sizes.h>
>>>>>
>>>>> @@ -548,9 +549,13 @@ static int dwcmshc_probe(struct platform_device *pdev)
>>>>>
>>>>>       host->mmc->caps |= MMC_CAP_WAIT_WHILE_BUSY;
>>>>>
>>>>> +     pm_runtime_get_noresume(dev);
>>>>> +     pm_runtime_set_active(dev);
>>>>> +     pm_runtime_enable(dev);
>>>>> +
>>>>>       err = sdhci_setup_host(host);
>>>>>       if (err)
>>>>> -             goto err_clk;
>>>>> +             goto err_rpm;
>>>>>
>>>>>       if (rk_priv)
>>>>>               dwcmshc_rk35xx_postinit(host, priv);
>>>>> @@ -559,10 +564,15 @@ static int dwcmshc_probe(struct platform_device *pdev)
>>>>>       if (err)
>>>>>               goto err_setup_host;
>>>>>
>>>>> +     pm_runtime_put(dev);
>>>>> +
>>>>>       return 0;
>>>>>
>>>>>  err_setup_host:
>>>>>       sdhci_cleanup_host(host);
>>>>> +err_rpm:
>>>>> +     pm_runtime_disable(dev);
>>>>> +     pm_runtime_put_noidle(dev);
>>>>>  err_clk:
>>>>>       clk_disable_unprepare(pltfm_host->clk);
>>>>>       clk_disable_unprepare(priv->bus_clk);
>>>>> @@ -606,6 +616,12 @@ static int dwcmshc_suspend(struct device *dev)
>>>>>       if (ret)
>>>>>               return ret;
>>>>>
>>>>> +     ret = pm_runtime_force_suspend(dev);
>>>>> +     if (ret) {
>>>>> +             sdhci_resume_host(host);
>>>>> +             return ret;
>>>>> +     }
>>>>
>>>> Since you are only using the runtime PM callbacks to turn off the card
>>>> clock via SDHCI_CLOCK_CONTROL, pm_runtime_force_suspend() and
>>>> pm_runtime_force_resume() are not needed at all.
>>>
>>> Right, it can be done without these too.
>>>
>>>>
>>>> sdhci_suspend_host() does not care if SDHCI_CLOCK_CARD_EN is on or off.
>>>> (And you are disabling pltfm_host->clk and priv->bus_clk, so presumably
>>>> the result is no clock either way)
>>>>
>>>> sdhci_resume_host() does not restore state unless
>>>> SDHCI_QUIRK2_HOST_OFF_CARD_ON is used, it just resets, so the internal clock
>>>> SDHCI_CLOCK_INT_EN is off which is consistent with either runtime suspended
>>>> or runtime resumed.
>>>
>>> Even if this may work, to me, it doesn't look like good practice for
>>> how to use runtime PM in combination with system wide suspend/resume.
>>>
>>> The point is, sdhci_suspend|resume_host() may end up reading/writing
>>> to sdhci registers - and we should *not* allow that (because it may
>>> not always work), unless the sdhci controller has been runtime resumed
>>> first, right?
>>
>> I am OK with drivers that just want to use runtime PM to turn off a
>> functional clock.  sdhci-tegra.c is also doing that although using the
>> clock framework.
> 
> Yes, I agree. At least this works for SoC specific drivers.
> 
>>
>> Certainly that approach assumes that the host controller's power state
>> is not changed due to runtime PM.
>>
>> To ensure that the host controller is runtime resumed before calling
>> sdhci_suspend_host(), we can just call pm_runtime_resume() I think.
> 
> Yes, that was kind of what I proposed in the other thread as option 1)
> (except for the replacement of pm_runtime_force_suspend|resume).
> 
> Although, to be clear I would probably use pm_runtime_get_sync()
> instead, to make sure the usage count is incremented too.

In that case, a matching pm_runtime_put() is needed also at the
end of the resume callback.

> 
> I don't have a strong opinion here, but from an optimization point of
> view I would at least consider what I proposed in option 2) (in the
> other thread). The benefit is that it can allow us to potentially
> avoid runtime resuming the device, during system suspend.
> 
> Kind regards
> Uffe