A device may need the controller up during suspend_noirq() or
resume_noirq().
But if the controller is autosuspended, there is no way to wakeup it during
suspend_noirq() or resume_noirq() because runtime pm is disabled at this
time.

The suspend() callback wakes up the controller, so it is available until
its suspend_noirq() callback (pm_runtime_force_suspend()).
During the resume, it's restored by resume_noirq() callback
(pm_runtime_force_resume()). Then resume() callback enables autosuspend.

So the controller is up during a little time slot in suspend and resume
sequences even if it's not used.

Reviewed-by: Andi Shyti <andi.shyti@kernel.org>
Acked-by: Wolfram Sang <wsa+renesas@sang-engineering.com>
Reviewed-by: Tony Lindgren <tony@atomide.com>
Signed-off-by: Thomas Richard <thomas.richard@bootlin.com>
---
 drivers/i2c/busses/i2c-omap.c | 22 ++++++++++++++++++++++
 1 file changed, 22 insertions(+)

diff --git a/drivers/i2c/busses/i2c-omap.c b/drivers/i2c/busses/i2c-omap.c
index 42165ef57946..28417b2a18b0 100644
--- a/drivers/i2c/busses/i2c-omap.c
+++ b/drivers/i2c/busses/i2c-omap.c
@@ -1575,9 +1575,31 @@ static int __maybe_unused omap_i2c_runtime_resume(struct device *dev)
 	return 0;
 }
 
+static int omap_i2c_suspend(struct device *dev)
+{
+	/*
+	 * If the controller is autosuspended, there is no way to wakeup it once
+	 * runtime pm is disabled (in suspend_late()).
+	 * But a device may need the controller up during suspend_noirq() or
+	 * resume_noirq().
+	 * Wakeup the controller while runtime pm is enabled, so it is available
+	 * until its suspend_noirq(), and from resume_noirq().
+	 */
+	return pm_runtime_resume_and_get(dev);
+}
+
+static int omap_i2c_resume(struct device *dev)
+{
+	pm_runtime_mark_last_busy(dev);
+	pm_runtime_put_autosuspend(dev);
+
+	return 0;
+}
+
 static const struct dev_pm_ops omap_i2c_pm_ops = {
 	SET_NOIRQ_SYSTEM_SLEEP_PM_OPS(pm_runtime_force_suspend,
 				      pm_runtime_force_resume)
+	SET_SYSTEM_SLEEP_PM_OPS(omap_i2c_suspend, omap_i2c_resume)
 	SET_RUNTIME_PM_OPS(omap_i2c_runtime_suspend,
 			   omap_i2c_runtime_resume, NULL)
 };

-- 
2.39.2

Hi Thomas,

> +static int omap_i2c_suspend(struct device *dev)
> +{
> +	/*
> +	 * If the controller is autosuspended, there is no way to wakeup it once
> +	 * runtime pm is disabled (in suspend_late()).
> +	 * But a device may need the controller up during suspend_noirq() or
> +	 * resume_noirq().
> +	 * Wakeup the controller while runtime pm is enabled, so it is available
> +	 * until its suspend_noirq(), and from resume_noirq().
> +	 */
> +	return pm_runtime_resume_and_get(dev);
> +}
> +
> +static int omap_i2c_resume(struct device *dev)
> +{
> +	pm_runtime_mark_last_busy(dev);
> +	pm_runtime_put_autosuspend(dev);
> +
> +	return 0;
> +}
> +
>  static const struct dev_pm_ops omap_i2c_pm_ops = {
>  	SET_NOIRQ_SYSTEM_SLEEP_PM_OPS(pm_runtime_force_suspend,
>  				      pm_runtime_force_resume)
> +	SET_SYSTEM_SLEEP_PM_OPS(omap_i2c_suspend, omap_i2c_resume)

If you don't have CONFIG_PM_SLEEP, though, this doesn't compile.

Andi

On 4/19/24 10:47, Andi Shyti wrote:
> Hi Thomas,
> 
>> +static int omap_i2c_suspend(struct device *dev)
>> +{
>> +	/*
>> +	 * If the controller is autosuspended, there is no way to wakeup it once
>> +	 * runtime pm is disabled (in suspend_late()).
>> +	 * But a device may need the controller up during suspend_noirq() or
>> +	 * resume_noirq().
>> +	 * Wakeup the controller while runtime pm is enabled, so it is available
>> +	 * until its suspend_noirq(), and from resume_noirq().
>> +	 */
>> +	return pm_runtime_resume_and_get(dev);
>> +}
>> +
>> +static int omap_i2c_resume(struct device *dev)
>> +{
>> +	pm_runtime_mark_last_busy(dev);
>> +	pm_runtime_put_autosuspend(dev);
>> +
>> +	return 0;
>> +}
>> +
>>  static const struct dev_pm_ops omap_i2c_pm_ops = {
>>  	SET_NOIRQ_SYSTEM_SLEEP_PM_OPS(pm_runtime_force_suspend,
>>  				      pm_runtime_force_resume)
>> +	SET_SYSTEM_SLEEP_PM_OPS(omap_i2c_suspend, omap_i2c_resume)
> 
> If you don't have CONFIG_PM_SLEEP, though, this doesn't compile.

Hello Andi,

Yes indeed, the __maybe_unused attribute is missing for
omap_i2c_suspend() and omap_i2c_resume().

-- 
Thomas Richard, Bootlin
Embedded Linux and Kernel engineering
https://bootlin.com

On Mon, Apr 22, 2024 at 11:40:02AM +0200, Thomas Richard wrote:
> On 4/19/24 10:47, Andi Shyti wrote:
> > Hi Thomas,
> > 
> >> +static int omap_i2c_suspend(struct device *dev)
> >> +{
> >> +	/*
> >> +	 * If the controller is autosuspended, there is no way to wakeup it once
> >> +	 * runtime pm is disabled (in suspend_late()).
> >> +	 * But a device may need the controller up during suspend_noirq() or
> >> +	 * resume_noirq().
> >> +	 * Wakeup the controller while runtime pm is enabled, so it is available
> >> +	 * until its suspend_noirq(), and from resume_noirq().
> >> +	 */
> >> +	return pm_runtime_resume_and_get(dev);
> >> +}
> >> +
> >> +static int omap_i2c_resume(struct device *dev)
> >> +{
> >> +	pm_runtime_mark_last_busy(dev);
> >> +	pm_runtime_put_autosuspend(dev);
> >> +
> >> +	return 0;
> >> +}
> >> +
> >>  static const struct dev_pm_ops omap_i2c_pm_ops = {
> >>  	SET_NOIRQ_SYSTEM_SLEEP_PM_OPS(pm_runtime_force_suspend,
> >>  				      pm_runtime_force_resume)
> >> +	SET_SYSTEM_SLEEP_PM_OPS(omap_i2c_suspend, omap_i2c_resume)
> > 
> > If you don't have CONFIG_PM_SLEEP, though, this doesn't compile.
> 
> Hello Andi,
> 
> Yes indeed, the __maybe_unused attribute is missing for
> omap_i2c_suspend() and omap_i2c_resume().

Isn't there a way to avoid having to use the __maybe_unused attribute?

E.g., use DEFINE_SIMPLE_DEV_PM_OPS() as is done by these:

  82f9cefadac4 ("serial: 8250_exar: switch to DEFINE_SIMPLE_DEV_PM_OPS()")
  f243df0a0be0 ("media: platform: rzg2l-cru: rzg2l-csi2: Switch to RUNTIME_PM_OPS()")
  6ccc22a5afcb ("net: ravb: Switch to SYSTEM_SLEEP_PM_OPS()/RUNTIME_PM_OPS() and pm_ptr()")

Bjorn

On 4/22/24 21:44, Bjorn Helgaas wrote:
> On Mon, Apr 22, 2024 at 11:40:02AM +0200, Thomas Richard wrote:
>> On 4/19/24 10:47, Andi Shyti wrote:
>>> Hi Thomas,
>>>
>>>> +static int omap_i2c_suspend(struct device *dev)
>>>> +{
>>>> +	/*
>>>> +	 * If the controller is autosuspended, there is no way to wakeup it once
>>>> +	 * runtime pm is disabled (in suspend_late()).
>>>> +	 * But a device may need the controller up during suspend_noirq() or
>>>> +	 * resume_noirq().
>>>> +	 * Wakeup the controller while runtime pm is enabled, so it is available
>>>> +	 * until its suspend_noirq(), and from resume_noirq().
>>>> +	 */
>>>> +	return pm_runtime_resume_and_get(dev);
>>>> +}
>>>> +
>>>> +static int omap_i2c_resume(struct device *dev)
>>>> +{
>>>> +	pm_runtime_mark_last_busy(dev);
>>>> +	pm_runtime_put_autosuspend(dev);
>>>> +
>>>> +	return 0;
>>>> +}
>>>> +
>>>>  static const struct dev_pm_ops omap_i2c_pm_ops = {
>>>>  	SET_NOIRQ_SYSTEM_SLEEP_PM_OPS(pm_runtime_force_suspend,
>>>>  				      pm_runtime_force_resume)
>>>> +	SET_SYSTEM_SLEEP_PM_OPS(omap_i2c_suspend, omap_i2c_resume)
>>>
>>> If you don't have CONFIG_PM_SLEEP, though, this doesn't compile.
>>
>> Hello Andi,
>>
>> Yes indeed, the __maybe_unused attribute is missing for
>> omap_i2c_suspend() and omap_i2c_resume().
> 
> Isn't there a way to avoid having to use the __maybe_unused attribute?
> 
> E.g., use DEFINE_SIMPLE_DEV_PM_OPS() as is done by these:
> 
>   82f9cefadac4 ("serial: 8250_exar: switch to DEFINE_SIMPLE_DEV_PM_OPS()")
>   f243df0a0be0 ("media: platform: rzg2l-cru: rzg2l-csi2: Switch to RUNTIME_PM_OPS()")
>   6ccc22a5afcb ("net: ravb: Switch to SYSTEM_SLEEP_PM_OPS()/RUNTIME_PM_OPS() and pm_ptr()")

Yes you're right, I don't need the __maybe_unused attribute if I use
NOIRQ_SYSTEM_SLEEP_PM_OPS().

By the way I can add a patch in the series to remove all the
__maybe_unused attributes of this driver.

Regards,

-- 
Thomas Richard, Bootlin
Embedded Linux and Kernel engineering
https://bootlin.com