diff mbox series

[v2] PM: runtime: enable wake irq after runtime_suspend hook called

Message ID 1626427381-30131-1-git-send-email-chunfeng.yun@mediatek.com (mailing list archive)
State New
Headers show
Series [v2] PM: runtime: enable wake irq after runtime_suspend hook called | expand

Commit Message

Chunfeng Yun (云春峰) July 16, 2021, 9:23 a.m. UTC
When the dedicated wake-irq is level trigger, and it uses the
consumer's sleep status as the wakeup source, that means if the
consumer is not in sleep state, the wake-irq will be triggered
when enable it; For this case, need enable the wake-irq after
invoking the consumer's runtime_suspend() which make the consumer
enter sleep state.

e.g.
Assume the wake-irq is a low level trigger type, and the wakeup
signal comes from the sleep status of consumer.
The wakeup signal is low level at running time (0), and becomes
high level when the consumer enters sleep state (runtime_suspend
(1) is called), a wakeup event at (2) make the consumer exit sleep
state, then the wakeup signal also becomes low level.

                ------------------
               |           ^     ^|
----------------           |     | --------------
 |<---(0)--->|<--(1)--|   (3)   (2)    (4)

if enable the wake-irq before calling runtime_suspend during (0),
an interrupt will arise, it causes resume immediately;
it works if enable wake-irq ( e.g. at (3) or (4)) after calling
runtime_suspend.

In this example, can't fix it by using falling edge trigger without
this patch, the issue will happen as below steps:
1. use another wakeup source to wake up the suspended system;
2. the consumer's resume() will be called, and exits sleep state;
3. the consumer's wakeup signal will fall into low level, due to
   currently the wakeup irq is disabled, the wake-irq is pending;
4. the consumer tries to enter runtime suspend, but there is a
   pending wakeup irq, so will resume again, this will repeat
   endlessly.

This patch seems no side effect on edge trigger wake-irq that works
before.

Signed-off-by: Chunfeng Yun <chunfeng.yun@mediatek.com>
---
v2: add more commit message

  I use the falling edge trigger interrupt suggested by Ikjoon [1], it
works well at firstly when only use this related wakeup source, but
encounter issues if use other wakeup sources to wakeup platform as
described in commit message.
  Send out the patch again for further discussion.

[1]: https://patchwork.kernel.org/patch/12190407

---
 drivers/base/power/runtime.c | 5 ++---
 1 file changed, 2 insertions(+), 3 deletions(-)

Comments

Rafael J. Wysocki July 16, 2021, 12:52 p.m. UTC | #1
On Fri, Jul 16, 2021 at 11:23 AM Chunfeng Yun <chunfeng.yun@mediatek.com> wrote:
>
> When the dedicated wake-irq is level trigger, and it uses the
> consumer's sleep status as the wakeup source,

This is not a typical case, though, AFAICS.

So you seem to have a special requirement and you want to change a
generic framework in an incompatible way to accommodate it.

> that means if the
> consumer is not in sleep state, the wake-irq will be triggered
> when enable it; For this case, need enable the wake-irq after
> invoking the consumer's runtime_suspend() which make the consumer
> enter sleep state.
>
> e.g.
> Assume the wake-irq is a low level trigger type, and the wakeup
> signal comes from the sleep status of consumer.
> The wakeup signal is low level at running time (0), and becomes
> high level when the consumer enters sleep state (runtime_suspend
> (1) is called), a wakeup event at (2) make the consumer exit sleep
> state, then the wakeup signal also becomes low level.
>
>                 ------------------
>                |           ^     ^|
> ----------------           |     | --------------
>  |<---(0)--->|<--(1)--|   (3)   (2)    (4)
>
> if enable the wake-irq before calling runtime_suspend during (0),
> an interrupt will arise, it causes resume immediately;
> it works if enable wake-irq ( e.g. at (3) or (4)) after calling
> runtime_suspend.
>
> In this example, can't fix it by using falling edge trigger without
> this patch, the issue will happen as below steps:
> 1. use another wakeup source to wake up the suspended system;
> 2. the consumer's resume() will be called, and exits sleep state;
> 3. the consumer's wakeup signal will fall into low level, due to
>    currently the wakeup irq is disabled, the wake-irq is pending;
> 4. the consumer tries to enter runtime suspend, but there is a
>    pending wakeup irq, so will resume again, this will repeat
>    endlessly.
>
> This patch seems no side effect on edge trigger wake-irq that works
> before.
>
> Signed-off-by: Chunfeng Yun <chunfeng.yun@mediatek.com>
> ---
> v2: add more commit message
>
>   I use the falling edge trigger interrupt suggested by Ikjoon [1], it
> works well at firstly when only use this related wakeup source, but
> encounter issues if use other wakeup sources to wakeup platform as
> described in commit message.
>   Send out the patch again for further discussion.
>
> [1]: https://patchwork.kernel.org/patch/12190407
>
> ---
>  drivers/base/power/runtime.c | 5 ++---
>  1 file changed, 2 insertions(+), 3 deletions(-)
>
> diff --git a/drivers/base/power/runtime.c b/drivers/base/power/runtime.c
> index 8a66eaf731e4..90a91b2b1364 100644
> --- a/drivers/base/power/runtime.c
> +++ b/drivers/base/power/runtime.c
> @@ -639,12 +639,12 @@ static int rpm_suspend(struct device *dev, int rpmflags)
>         __update_runtime_status(dev, RPM_SUSPENDING);
>
>         callback = RPM_GET_CALLBACK(dev, runtime_suspend);
> -
> -       dev_pm_enable_wake_irq_check(dev, true);

I would suggest adding a wake IRQ flag to indicate that it needs
special handling and modifying the above to take that new flag into
account.

>         retval = rpm_callback(callback, dev);
>         if (retval)
>                 goto fail;
>
> +       dev_pm_enable_wake_irq_check(dev, true);

Then, you can add an entirely new call here, say
dev_pm_enable_wake_irq_complete(dev), that will do the actual enable
for the wake IRQs with the new flag set and nothing for the others.

> +
>   no_callback:
>         __update_runtime_status(dev, RPM_SUSPENDED);
>         pm_runtime_deactivate_timer(dev);
> @@ -690,7 +690,6 @@ static int rpm_suspend(struct device *dev, int rpmflags)
>         return retval;
>
>   fail:
> -       dev_pm_disable_wake_irq_check(dev);

And obviously this will really disable wake IRQs with the new flag unset only.

>         __update_runtime_status(dev, RPM_ACTIVE);
>         dev->power.deferred_resume = false;
>         wake_up_all(&dev->power.wait_queue);
> --
> 2.18.0
>
Chunfeng Yun (云春峰) July 20, 2021, 5:26 a.m. UTC | #2
On Fri, 2021-07-16 at 14:52 +0200, Rafael J. Wysocki wrote:
> On Fri, Jul 16, 2021 at 11:23 AM Chunfeng Yun <chunfeng.yun@mediatek.com> wrote:
> >
> > When the dedicated wake-irq is level trigger, and it uses the
> > consumer's sleep status as the wakeup source,
> 
> This is not a typical case, though, AFAICS.
> 
> So you seem to have a special requirement and you want to change a
> generic framework in an incompatible way to accommodate it.
> 
> > that means if the
> > consumer is not in sleep state, the wake-irq will be triggered
> > when enable it; For this case, need enable the wake-irq after
> > invoking the consumer's runtime_suspend() which make the consumer
> > enter sleep state.
> >
> > e.g.
> > Assume the wake-irq is a low level trigger type, and the wakeup
> > signal comes from the sleep status of consumer.
> > The wakeup signal is low level at running time (0), and becomes
> > high level when the consumer enters sleep state (runtime_suspend
> > (1) is called), a wakeup event at (2) make the consumer exit sleep
> > state, then the wakeup signal also becomes low level.
> >
> >                 ------------------
> >                |           ^     ^|
> > ----------------           |     | --------------
> >  |<---(0)--->|<--(1)--|   (3)   (2)    (4)
> >
> > if enable the wake-irq before calling runtime_suspend during (0),
> > an interrupt will arise, it causes resume immediately;
> > it works if enable wake-irq ( e.g. at (3) or (4)) after calling
> > runtime_suspend.
> >
> > In this example, can't fix it by using falling edge trigger without
> > this patch, the issue will happen as below steps:
> > 1. use another wakeup source to wake up the suspended system;
> > 2. the consumer's resume() will be called, and exits sleep state;
> > 3. the consumer's wakeup signal will fall into low level, due to
> >    currently the wakeup irq is disabled, the wake-irq is pending;
> > 4. the consumer tries to enter runtime suspend, but there is a
> >    pending wakeup irq, so will resume again, this will repeat
> >    endlessly.
> >
> > This patch seems no side effect on edge trigger wake-irq that works
> > before.
> >
> > Signed-off-by: Chunfeng Yun <chunfeng.yun@mediatek.com>
> > ---
> > v2: add more commit message
> >
> >   I use the falling edge trigger interrupt suggested by Ikjoon [1], it
> > works well at firstly when only use this related wakeup source, but
> > encounter issues if use other wakeup sources to wakeup platform as
> > described in commit message.
> >   Send out the patch again for further discussion.
> >
> > [1]: https://patchwork.kernel.org/patch/12190407
> >
> > ---
> >  drivers/base/power/runtime.c | 5 ++---
> >  1 file changed, 2 insertions(+), 3 deletions(-)
> >
> > diff --git a/drivers/base/power/runtime.c b/drivers/base/power/runtime.c
> > index 8a66eaf731e4..90a91b2b1364 100644
> > --- a/drivers/base/power/runtime.c
> > +++ b/drivers/base/power/runtime.c
> > @@ -639,12 +639,12 @@ static int rpm_suspend(struct device *dev, int rpmflags)
> >         __update_runtime_status(dev, RPM_SUSPENDING);
> >
> >         callback = RPM_GET_CALLBACK(dev, runtime_suspend);
> > -
> > -       dev_pm_enable_wake_irq_check(dev, true);
> 
> I would suggest adding a wake IRQ flag to indicate that it needs
> special handling and modifying the above to take that new flag into
> account.
> 
> >         retval = rpm_callback(callback, dev);
> >         if (retval)
> >                 goto fail;
> >
> > +       dev_pm_enable_wake_irq_check(dev, true);
> 
> Then, you can add an entirely new call here, say
> dev_pm_enable_wake_irq_complete(dev), that will do the actual enable
> for the wake IRQs with the new flag set and nothing for the others.
> 
> > +
> >   no_callback:
> >         __update_runtime_status(dev, RPM_SUSPENDED);
> >         pm_runtime_deactivate_timer(dev);
> > @@ -690,7 +690,6 @@ static int rpm_suspend(struct device *dev, int rpmflags)
> >         return retval;
> >
> >   fail:
> > -       dev_pm_disable_wake_irq_check(dev);
> 
> And obviously this will really disable wake IRQs with the new flag unset only.

Ok, I'll try it, thanks a lot

> 
> >         __update_runtime_status(dev, RPM_ACTIVE);
> >         dev->power.deferred_resume = false;
> >         wake_up_all(&dev->power.wait_queue);
> > --
> > 2.18.0
> >
diff mbox series

Patch

diff --git a/drivers/base/power/runtime.c b/drivers/base/power/runtime.c
index 8a66eaf731e4..90a91b2b1364 100644
--- a/drivers/base/power/runtime.c
+++ b/drivers/base/power/runtime.c
@@ -639,12 +639,12 @@  static int rpm_suspend(struct device *dev, int rpmflags)
 	__update_runtime_status(dev, RPM_SUSPENDING);
 
 	callback = RPM_GET_CALLBACK(dev, runtime_suspend);
-
-	dev_pm_enable_wake_irq_check(dev, true);
 	retval = rpm_callback(callback, dev);
 	if (retval)
 		goto fail;
 
+	dev_pm_enable_wake_irq_check(dev, true);
+
  no_callback:
 	__update_runtime_status(dev, RPM_SUSPENDED);
 	pm_runtime_deactivate_timer(dev);
@@ -690,7 +690,6 @@  static int rpm_suspend(struct device *dev, int rpmflags)
 	return retval;
 
  fail:
-	dev_pm_disable_wake_irq_check(dev);
 	__update_runtime_status(dev, RPM_ACTIVE);
 	dev->power.deferred_resume = false;
 	wake_up_all(&dev->power.wait_queue);