diff mbox series

[v6,2/3] PM / domains: inform PM domain of a device's next wakeup

Message ID 20201130225039.15981-3-ilina@codeaurora.org
State New
Headers show
Series Better domain idle from device wakeup patterns | expand

Commit Message

Lina Iyer Nov. 30, 2020, 10:50 p.m. UTC
Some devices may have a predictable interrupt pattern while executing
usecases. An example would be the VSYNC interrupt associated with
display devices. A 60 Hz display could cause a interrupt every 16 ms. If
the device were in a PM domain, the domain would need to be powered up
for device to resume and handle the interrupt.

Entering a domain idle state saves power, only if the residency of the
idle state is met. Without knowing the idle duration of the domain, the
governor would just choose the deepest idle state that matches the QoS
requirements. The domain might be powered off just as the device is
expecting to wake up. If devices could inform PM frameworks of their
next event, the parent PM domain's idle duration can be determined.

So let's add the dev_pm_genpd_set_next_wakeup() API for the device to
inform PM domains of the impending wakeup. This information will be the
domain governor to determine the best idle state given the wakeup.

Signed-off-by: Lina Iyer <ilina@codeaurora.org>
---
Changes in v6:
	- Update documentation
Changes in v5:
	- Fix commit text as pointed by Ulf
	- Use -EOPNOTSUPP
Changes in v4:
	- Use PM domain data to store next_wakeup
	- Drop runtime PM documentation
Changes in v3:
	- Fix unwanted change
Changes in v2:
	- Update documentation
	- Remove runtime PM enabled check
	- Update commit text
---
 drivers/base/power/domain.c | 41 +++++++++++++++++++++++++++++++++++++
 include/linux/pm_domain.h   |  8 ++++++++
 2 files changed, 49 insertions(+)

Comments

Ulf Hansson Dec. 22, 2020, 1:09 p.m. UTC | #1
On Mon, 30 Nov 2020 at 23:51, Lina Iyer <ilina@codeaurora.org> wrote:
>
> Some devices may have a predictable interrupt pattern while executing
> usecases. An example would be the VSYNC interrupt associated with
> display devices. A 60 Hz display could cause a interrupt every 16 ms. If
> the device were in a PM domain, the domain would need to be powered up
> for device to resume and handle the interrupt.
>
> Entering a domain idle state saves power, only if the residency of the
> idle state is met. Without knowing the idle duration of the domain, the
> governor would just choose the deepest idle state that matches the QoS
> requirements. The domain might be powered off just as the device is
> expecting to wake up. If devices could inform PM frameworks of their
> next event, the parent PM domain's idle duration can be determined.
>
> So let's add the dev_pm_genpd_set_next_wakeup() API for the device to
> inform PM domains of the impending wakeup. This information will be the
> domain governor to determine the best idle state given the wakeup.
>
> Signed-off-by: Lina Iyer <ilina@codeaurora.org>
> ---
> Changes in v6:
>         - Update documentation
> Changes in v5:
>         - Fix commit text as pointed by Ulf
>         - Use -EOPNOTSUPP
> Changes in v4:
>         - Use PM domain data to store next_wakeup
>         - Drop runtime PM documentation
> Changes in v3:
>         - Fix unwanted change
> Changes in v2:
>         - Update documentation
>         - Remove runtime PM enabled check
>         - Update commit text
> ---
>  drivers/base/power/domain.c | 41 +++++++++++++++++++++++++++++++++++++
>  include/linux/pm_domain.h   |  8 ++++++++
>  2 files changed, 49 insertions(+)
>
> diff --git a/drivers/base/power/domain.c b/drivers/base/power/domain.c
> index 1e6c0bf1c945..191539a8e06d 100644
> --- a/drivers/base/power/domain.c
> +++ b/drivers/base/power/domain.c
> @@ -408,6 +408,46 @@ int dev_pm_genpd_set_performance_state(struct device *dev, unsigned int state)
>  }
>  EXPORT_SYMBOL_GPL(dev_pm_genpd_set_performance_state);
>
> +/**
> + * dev_pm_genpd_set_next_wakeup - Notify PM framework of an impending wakeup.
> + *
> + * @dev: Device to handle
> + * @next: impending interrupt/wakeup for the device
> + *
> + * Allow devices to inform of the next wakeup. But, if the domain were already
> + * powered off, we will not wakeup the domain to recompute it's idle duration.

How about clarifying this as:

"It's assumed that the users guarantee that the genpd wouldn't be
detached while this routine is getting called. Additionally, it's also
assumed that @dev isn't runtime suspended (RPM_SUSPENDED).

> + * Although devices are expected to update the next_wakeup after the end of
> + * their usecase as well, it is possible the devices themselves may not know
> + * about that. Stale @next will be ignored when powering off the domain.
> + */
> +int dev_pm_genpd_set_next_wakeup(struct device *dev, ktime_t next)
> +{
> +       struct generic_pm_domain *genpd;
> +       struct generic_pm_domain_data *gpd_data;
> +       int ret = -EINVAL;
> +
> +       genpd = dev_to_genpd_safe(dev);
> +       if (!genpd)
> +               return -ENODEV;
> +
> +       if (WARN_ON(!dev->power.subsys_data ||
> +                   !dev->power.subsys_data->domain_data))
> +               return ret;
> +
> +       genpd_lock(genpd);
> +       gpd_data = to_gpd_data(dev->power.subsys_data->domain_data);
> +       if (ktime_before(ktime_get(), next)) {
> +               gpd_data->next_wakeup = next;
> +               genpd->flags |= GENPD_FLAG_GOV_NEXT_WAKEUP;

I don't think we should set this here, but instead leave this to be
set by the genpd provider at initialization.

> +               ret = 0;
> +       }
> +       genpd_unlock(genpd);

I would suggest to simplify the above into:

gpd_data = to_gpd_data(dev->power.subsys_data->domain_data);
gpd_data->next_wakeup = next;

Then there is no need for locks because:
*) We assume the device remains attached to the genpd.
**) The device isn't runtime suspended, hence its corresponding genpd
is powered on and thus the genpd governor can't be looking at
"gpd_data->next_wakeup" simulantfsfulsy.

Moreover, as we agreed to ignore stale values for "next", there is no
reason to validate it against the current ktime, so let's just skip
it.

> +
> +       return ret;
> +}
> +EXPORT_SYMBOL_GPL(dev_pm_genpd_set_next_wakeup);
> +
> +
>  static int _genpd_power_on(struct generic_pm_domain *genpd, bool timed)
>  {
>         unsigned int state_idx = genpd->state_idx;
> @@ -1450,6 +1490,7 @@ static struct generic_pm_domain_data *genpd_alloc_dev_data(struct device *dev)
>         gpd_data->td.constraint_changed = true;
>         gpd_data->td.effective_constraint_ns = PM_QOS_RESUME_LATENCY_NO_CONSTRAINT_NS;
>         gpd_data->nb.notifier_call = genpd_dev_pm_qos_notifier;
> +       gpd_data->next_wakeup = KTIME_MAX;

When looking at patch3, I wonder if it perhaps could be easier to use
"zero" as the default value? What do you think, just an idea?

>
>         spin_lock_irq(&dev->power.lock);
>
> diff --git a/include/linux/pm_domain.h b/include/linux/pm_domain.h
> index 1f359bd19f77..cc27d3d88849 100644
> --- a/include/linux/pm_domain.h
> +++ b/include/linux/pm_domain.h
> @@ -9,6 +9,7 @@
>  #define _LINUX_PM_DOMAIN_H
>
>  #include <linux/device.h>
> +#include <linux/ktime.h>
>  #include <linux/mutex.h>
>  #include <linux/pm.h>
>  #include <linux/err.h>
> @@ -197,6 +198,7 @@ struct generic_pm_domain_data {
>         struct notifier_block *power_nb;
>         int cpu;
>         unsigned int performance_state;
> +       ktime_t next_wakeup;
>         void *data;
>  };
>
> @@ -229,6 +231,7 @@ int dev_pm_genpd_set_performance_state(struct device *dev, unsigned int state);
>  int dev_pm_genpd_add_notifier(struct device *dev, struct notifier_block *nb);
>  int dev_pm_genpd_remove_notifier(struct device *dev);
>  void genpd_enable_next_wakeup(struct generic_pm_domain *genpd, bool enable);
> +int dev_pm_genpd_set_next_wakeup(struct device *dev, ktime_t next);
>
>  extern struct dev_power_governor simple_qos_governor;
>  extern struct dev_power_governor pm_domain_always_on_gov;
> @@ -291,6 +294,11 @@ static void genpd_enable_next_wakeup(struct generic_pm_domain *genpd,
>                                      bool enable)
>  { }
>
> +static inline int dev_pm_genpd_set_next_wakeup(struct device *dev, ktime_t next)
> +{
> +       return -EOPNOTSUPP;
> +}
> +
>  #define simple_qos_governor            (*(struct dev_power_governor *)(NULL))
>  #define pm_domain_always_on_gov                (*(struct dev_power_governor *)(NULL))
>  #endif
> --
> The Qualcomm Innovation Center, Inc. is a member of the Code Aurora Forum,
> a Linux Foundation Collaborative Project
>

Kind regards
Uffe
Lina Iyer Jan. 4, 2021, 3:43 p.m. UTC | #2
On Tue, Dec 22 2020 at 06:10 -0700, Ulf Hansson wrote:
>On Mon, 30 Nov 2020 at 23:51, Lina Iyer <ilina@codeaurora.org> wrote:
>>
>> Some devices may have a predictable interrupt pattern while executing
>> usecases. An example would be the VSYNC interrupt associated with
>> display devices. A 60 Hz display could cause a interrupt every 16 ms. If
>> the device were in a PM domain, the domain would need to be powered up
>> for device to resume and handle the interrupt.
>>
>> Entering a domain idle state saves power, only if the residency of the
>> idle state is met. Without knowing the idle duration of the domain, the
>> governor would just choose the deepest idle state that matches the QoS
>> requirements. The domain might be powered off just as the device is
>> expecting to wake up. If devices could inform PM frameworks of their
>> next event, the parent PM domain's idle duration can be determined.
>>
>> So let's add the dev_pm_genpd_set_next_wakeup() API for the device to
>> inform PM domains of the impending wakeup. This information will be the
>> domain governor to determine the best idle state given the wakeup.
>>
>> Signed-off-by: Lina Iyer <ilina@codeaurora.org>
>> ---
>> Changes in v6:
>>         - Update documentation
>> Changes in v5:
>>         - Fix commit text as pointed by Ulf
>>         - Use -EOPNOTSUPP
>> Changes in v4:
>>         - Use PM domain data to store next_wakeup
>>         - Drop runtime PM documentation
>> Changes in v3:
>>         - Fix unwanted change
>> Changes in v2:
>>         - Update documentation
>>         - Remove runtime PM enabled check
>>         - Update commit text
>> ---
>>  drivers/base/power/domain.c | 41 +++++++++++++++++++++++++++++++++++++
>>  include/linux/pm_domain.h   |  8 ++++++++
>>  2 files changed, 49 insertions(+)
>>
>> diff --git a/drivers/base/power/domain.c b/drivers/base/power/domain.c
>> index 1e6c0bf1c945..191539a8e06d 100644
>> --- a/drivers/base/power/domain.c
>> +++ b/drivers/base/power/domain.c
>> @@ -408,6 +408,46 @@ int dev_pm_genpd_set_performance_state(struct device *dev, unsigned int state)
>>  }
>>  EXPORT_SYMBOL_GPL(dev_pm_genpd_set_performance_state);
>>
>> +/**
>> + * dev_pm_genpd_set_next_wakeup - Notify PM framework of an impending wakeup.
>> + *
>> + * @dev: Device to handle
>> + * @next: impending interrupt/wakeup for the device
>> + *
>> + * Allow devices to inform of the next wakeup. But, if the domain were already
>> + * powered off, we will not wakeup the domain to recompute it's idle duration.
>
>How about clarifying this as:
>
>"It's assumed that the users guarantee that the genpd wouldn't be
>detached while this routine is getting called. Additionally, it's also
>assumed that @dev isn't runtime suspended (RPM_SUSPENDED).
>
Sure.

>> + * Although devices are expected to update the next_wakeup after the end of
>> + * their usecase as well, it is possible the devices themselves may not know
>> + * about that. Stale @next will be ignored when powering off the domain.
>> + */
>> +int dev_pm_genpd_set_next_wakeup(struct device *dev, ktime_t next)
>> +{
>> +       struct generic_pm_domain *genpd;
>> +       struct generic_pm_domain_data *gpd_data;
>> +       int ret = -EINVAL;
>> +
>> +       genpd = dev_to_genpd_safe(dev);
>> +       if (!genpd)
>> +               return -ENODEV;
>> +
>> +       if (WARN_ON(!dev->power.subsys_data ||
>> +                   !dev->power.subsys_data->domain_data))
>> +               return ret;
>> +
>> +       genpd_lock(genpd);
>> +       gpd_data = to_gpd_data(dev->power.subsys_data->domain_data);
>> +       if (ktime_before(ktime_get(), next)) {
>> +               gpd_data->next_wakeup = next;
>> +               genpd->flags |= GENPD_FLAG_GOV_NEXT_WAKEUP;
>
>I don't think we should set this here, but instead leave this to be
>set by the genpd provider at initialization.
>
But, we don't want it to be enabled by default. It is possible that
domains may have multiple idle states but none of the devices in the
domain have the need for next wakeup. We could optimize out the next
wakeup for those cases.
Or, the domain needs to call genpd_enable_next_wakeup() (patch #1) to
allow this feature. Is that acceptable?

>> +               ret = 0;
>> +       }
>> +       genpd_unlock(genpd);
>
>I would suggest to simplify the above into:
>
>gpd_data = to_gpd_data(dev->power.subsys_data->domain_data);
>gpd_data->next_wakeup = next;
>
>Then there is no need for locks because:
>*) We assume the device remains attached to the genpd.
>**) The device isn't runtime suspended, hence its corresponding genpd
>is powered on and thus the genpd governor can't be looking at
>"gpd_data->next_wakeup" simulantfsfulsy.
>
:)
>Moreover, as we agreed to ignore stale values for "next", there is no
>reason to validate it against the current ktime, so let's just skip
>it.
>
Okay.

>> +
>> +       return ret;
>> +}
>> +EXPORT_SYMBOL_GPL(dev_pm_genpd_set_next_wakeup);
>> +
>> +
>>  static int _genpd_power_on(struct generic_pm_domain *genpd, bool timed)
>>  {
>>         unsigned int state_idx = genpd->state_idx;
>> @@ -1450,6 +1490,7 @@ static struct generic_pm_domain_data *genpd_alloc_dev_data(struct device *dev)
>>         gpd_data->td.constraint_changed = true;
>>         gpd_data->td.effective_constraint_ns = PM_QOS_RESUME_LATENCY_NO_CONSTRAINT_NS;
>>         gpd_data->nb.notifier_call = genpd_dev_pm_qos_notifier;
>> +       gpd_data->next_wakeup = KTIME_MAX;
>
>When looking at patch3, I wonder if it perhaps could be easier to use
>"zero" as the default value? What do you think, just an idea?
>
Hmm.. Let me think it over.

Thanks,
Lina

>>
>>         spin_lock_irq(&dev->power.lock);
>>
>> diff --git a/include/linux/pm_domain.h b/include/linux/pm_domain.h
>> index 1f359bd19f77..cc27d3d88849 100644
>> --- a/include/linux/pm_domain.h
>> +++ b/include/linux/pm_domain.h
>> @@ -9,6 +9,7 @@
>>  #define _LINUX_PM_DOMAIN_H
>>
>>  #include <linux/device.h>
>> +#include <linux/ktime.h>
>>  #include <linux/mutex.h>
>>  #include <linux/pm.h>
>>  #include <linux/err.h>
>> @@ -197,6 +198,7 @@ struct generic_pm_domain_data {
>>         struct notifier_block *power_nb;
>>         int cpu;
>>         unsigned int performance_state;
>> +       ktime_t next_wakeup;
>>         void *data;
>>  };
>>
>> @@ -229,6 +231,7 @@ int dev_pm_genpd_set_performance_state(struct device *dev, unsigned int state);
>>  int dev_pm_genpd_add_notifier(struct device *dev, struct notifier_block *nb);
>>  int dev_pm_genpd_remove_notifier(struct device *dev);
>>  void genpd_enable_next_wakeup(struct generic_pm_domain *genpd, bool enable);
>> +int dev_pm_genpd_set_next_wakeup(struct device *dev, ktime_t next);
>>
>>  extern struct dev_power_governor simple_qos_governor;
>>  extern struct dev_power_governor pm_domain_always_on_gov;
>> @@ -291,6 +294,11 @@ static void genpd_enable_next_wakeup(struct generic_pm_domain *genpd,
>>                                      bool enable)
>>  { }
>>
>> +static inline int dev_pm_genpd_set_next_wakeup(struct device *dev, ktime_t next)
>> +{
>> +       return -EOPNOTSUPP;
>> +}
>> +
>>  #define simple_qos_governor            (*(struct dev_power_governor *)(NULL))
>>  #define pm_domain_always_on_gov                (*(struct dev_power_governor *)(NULL))
>>  #endif
>> --
>> The Qualcomm Innovation Center, Inc. is a member of the Code Aurora Forum,
>> a Linux Foundation Collaborative Project
>>
>
>Kind regards
>Uffe
Ulf Hansson Jan. 8, 2021, 10:07 a.m. UTC | #3
On Mon, 4 Jan 2021 at 16:44, Lina Iyer <ilina@codeaurora.org> wrote:
>
> On Tue, Dec 22 2020 at 06:10 -0700, Ulf Hansson wrote:
> >On Mon, 30 Nov 2020 at 23:51, Lina Iyer <ilina@codeaurora.org> wrote:
> >>
> >> Some devices may have a predictable interrupt pattern while executing
> >> usecases. An example would be the VSYNC interrupt associated with
> >> display devices. A 60 Hz display could cause a interrupt every 16 ms. If
> >> the device were in a PM domain, the domain would need to be powered up
> >> for device to resume and handle the interrupt.
> >>
> >> Entering a domain idle state saves power, only if the residency of the
> >> idle state is met. Without knowing the idle duration of the domain, the
> >> governor would just choose the deepest idle state that matches the QoS
> >> requirements. The domain might be powered off just as the device is
> >> expecting to wake up. If devices could inform PM frameworks of their
> >> next event, the parent PM domain's idle duration can be determined.
> >>
> >> So let's add the dev_pm_genpd_set_next_wakeup() API for the device to
> >> inform PM domains of the impending wakeup. This information will be the
> >> domain governor to determine the best idle state given the wakeup.
> >>
> >> Signed-off-by: Lina Iyer <ilina@codeaurora.org>
> >> ---
> >> Changes in v6:
> >>         - Update documentation
> >> Changes in v5:
> >>         - Fix commit text as pointed by Ulf
> >>         - Use -EOPNOTSUPP
> >> Changes in v4:
> >>         - Use PM domain data to store next_wakeup
> >>         - Drop runtime PM documentation
> >> Changes in v3:
> >>         - Fix unwanted change
> >> Changes in v2:
> >>         - Update documentation
> >>         - Remove runtime PM enabled check
> >>         - Update commit text
> >> ---
> >>  drivers/base/power/domain.c | 41 +++++++++++++++++++++++++++++++++++++
> >>  include/linux/pm_domain.h   |  8 ++++++++
> >>  2 files changed, 49 insertions(+)
> >>
> >> diff --git a/drivers/base/power/domain.c b/drivers/base/power/domain.c
> >> index 1e6c0bf1c945..191539a8e06d 100644
> >> --- a/drivers/base/power/domain.c
> >> +++ b/drivers/base/power/domain.c
> >> @@ -408,6 +408,46 @@ int dev_pm_genpd_set_performance_state(struct device *dev, unsigned int state)
> >>  }
> >>  EXPORT_SYMBOL_GPL(dev_pm_genpd_set_performance_state);
> >>
> >> +/**
> >> + * dev_pm_genpd_set_next_wakeup - Notify PM framework of an impending wakeup.
> >> + *
> >> + * @dev: Device to handle
> >> + * @next: impending interrupt/wakeup for the device
> >> + *
> >> + * Allow devices to inform of the next wakeup. But, if the domain were already
> >> + * powered off, we will not wakeup the domain to recompute it's idle duration.
> >
> >How about clarifying this as:
> >
> >"It's assumed that the users guarantee that the genpd wouldn't be
> >detached while this routine is getting called. Additionally, it's also
> >assumed that @dev isn't runtime suspended (RPM_SUSPENDED).
> >
> Sure.
>
> >> + * Although devices are expected to update the next_wakeup after the end of
> >> + * their usecase as well, it is possible the devices themselves may not know
> >> + * about that. Stale @next will be ignored when powering off the domain.
> >> + */
> >> +int dev_pm_genpd_set_next_wakeup(struct device *dev, ktime_t next)
> >> +{
> >> +       struct generic_pm_domain *genpd;
> >> +       struct generic_pm_domain_data *gpd_data;
> >> +       int ret = -EINVAL;
> >> +
> >> +       genpd = dev_to_genpd_safe(dev);
> >> +       if (!genpd)
> >> +               return -ENODEV;
> >> +
> >> +       if (WARN_ON(!dev->power.subsys_data ||
> >> +                   !dev->power.subsys_data->domain_data))
> >> +               return ret;
> >> +
> >> +       genpd_lock(genpd);
> >> +       gpd_data = to_gpd_data(dev->power.subsys_data->domain_data);
> >> +       if (ktime_before(ktime_get(), next)) {
> >> +               gpd_data->next_wakeup = next;
> >> +               genpd->flags |= GENPD_FLAG_GOV_NEXT_WAKEUP;
> >
> >I don't think we should set this here, but instead leave this to be
> >set by the genpd provider at initialization.
> >
> But, we don't want it to be enabled by default. It is possible that
> domains may have multiple idle states but none of the devices in the
> domain have the need for next wakeup. We could optimize out the next
> wakeup for those cases.

Yes, I understand, but at this point I don't think it's needed.

My main concern was to allow us to avoid the path for the
"pm_domain_cpu_gov" type for some genpds. Thus I think it's good
enough to set a new flag per genpd at initialization time.

> Or, the domain needs to call genpd_enable_next_wakeup() (patch #1) to
> allow this feature. Is that acceptable?

The problem is that we would need a reference counting mechanism, as
to also understand when to also turn off the flag. In other words, we
would need some kind of locking. I suggest we simply skip this for now
and see how this plays. What do you think?

>
> >> +               ret = 0;
> >> +       }
> >> +       genpd_unlock(genpd);
> >
> >I would suggest to simplify the above into:
> >
> >gpd_data = to_gpd_data(dev->power.subsys_data->domain_data);
> >gpd_data->next_wakeup = next;
> >
> >Then there is no need for locks because:
> >*) We assume the device remains attached to the genpd.
> >**) The device isn't runtime suspended, hence its corresponding genpd
> >is powered on and thus the genpd governor can't be looking at
> >"gpd_data->next_wakeup" simulantfsfulsy.
> >
> :)
> >Moreover, as we agreed to ignore stale values for "next", there is no
> >reason to validate it against the current ktime, so let's just skip
> >it.
> >
> Okay.
>
> >> +
> >> +       return ret;
> >> +}
> >> +EXPORT_SYMBOL_GPL(dev_pm_genpd_set_next_wakeup);
> >> +
> >> +
> >>  static int _genpd_power_on(struct generic_pm_domain *genpd, bool timed)
> >>  {
> >>         unsigned int state_idx = genpd->state_idx;
> >> @@ -1450,6 +1490,7 @@ static struct generic_pm_domain_data *genpd_alloc_dev_data(struct device *dev)
> >>         gpd_data->td.constraint_changed = true;
> >>         gpd_data->td.effective_constraint_ns = PM_QOS_RESUME_LATENCY_NO_CONSTRAINT_NS;
> >>         gpd_data->nb.notifier_call = genpd_dev_pm_qos_notifier;
> >> +       gpd_data->next_wakeup = KTIME_MAX;
> >
> >When looking at patch3, I wonder if it perhaps could be easier to use
> >"zero" as the default value? What do you think, just an idea?
> >
> Hmm.. Let me think it over.
>
> Thanks,
> Lina
>
> >>
> >>         spin_lock_irq(&dev->power.lock);
> >>

[...]

Kind regards
Uffe
diff mbox series

Patch

diff --git a/drivers/base/power/domain.c b/drivers/base/power/domain.c
index 1e6c0bf1c945..191539a8e06d 100644
--- a/drivers/base/power/domain.c
+++ b/drivers/base/power/domain.c
@@ -408,6 +408,46 @@  int dev_pm_genpd_set_performance_state(struct device *dev, unsigned int state)
 }
 EXPORT_SYMBOL_GPL(dev_pm_genpd_set_performance_state);
 
+/**
+ * dev_pm_genpd_set_next_wakeup - Notify PM framework of an impending wakeup.
+ *
+ * @dev: Device to handle
+ * @next: impending interrupt/wakeup for the device
+ *
+ * Allow devices to inform of the next wakeup. But, if the domain were already
+ * powered off, we will not wakeup the domain to recompute it's idle duration.
+ * Although devices are expected to update the next_wakeup after the end of
+ * their usecase as well, it is possible the devices themselves may not know
+ * about that. Stale @next will be ignored when powering off the domain.
+ */
+int dev_pm_genpd_set_next_wakeup(struct device *dev, ktime_t next)
+{
+	struct generic_pm_domain *genpd;
+	struct generic_pm_domain_data *gpd_data;
+	int ret = -EINVAL;
+
+	genpd = dev_to_genpd_safe(dev);
+	if (!genpd)
+		return -ENODEV;
+
+	if (WARN_ON(!dev->power.subsys_data ||
+		    !dev->power.subsys_data->domain_data))
+		return ret;
+
+	genpd_lock(genpd);
+	gpd_data = to_gpd_data(dev->power.subsys_data->domain_data);
+	if (ktime_before(ktime_get(), next)) {
+		gpd_data->next_wakeup = next;
+		genpd->flags |= GENPD_FLAG_GOV_NEXT_WAKEUP;
+		ret = 0;
+	}
+	genpd_unlock(genpd);
+
+	return ret;
+}
+EXPORT_SYMBOL_GPL(dev_pm_genpd_set_next_wakeup);
+
+
 static int _genpd_power_on(struct generic_pm_domain *genpd, bool timed)
 {
 	unsigned int state_idx = genpd->state_idx;
@@ -1450,6 +1490,7 @@  static struct generic_pm_domain_data *genpd_alloc_dev_data(struct device *dev)
 	gpd_data->td.constraint_changed = true;
 	gpd_data->td.effective_constraint_ns = PM_QOS_RESUME_LATENCY_NO_CONSTRAINT_NS;
 	gpd_data->nb.notifier_call = genpd_dev_pm_qos_notifier;
+	gpd_data->next_wakeup = KTIME_MAX;
 
 	spin_lock_irq(&dev->power.lock);
 
diff --git a/include/linux/pm_domain.h b/include/linux/pm_domain.h
index 1f359bd19f77..cc27d3d88849 100644
--- a/include/linux/pm_domain.h
+++ b/include/linux/pm_domain.h
@@ -9,6 +9,7 @@ 
 #define _LINUX_PM_DOMAIN_H
 
 #include <linux/device.h>
+#include <linux/ktime.h>
 #include <linux/mutex.h>
 #include <linux/pm.h>
 #include <linux/err.h>
@@ -197,6 +198,7 @@  struct generic_pm_domain_data {
 	struct notifier_block *power_nb;
 	int cpu;
 	unsigned int performance_state;
+	ktime_t	next_wakeup;
 	void *data;
 };
 
@@ -229,6 +231,7 @@  int dev_pm_genpd_set_performance_state(struct device *dev, unsigned int state);
 int dev_pm_genpd_add_notifier(struct device *dev, struct notifier_block *nb);
 int dev_pm_genpd_remove_notifier(struct device *dev);
 void genpd_enable_next_wakeup(struct generic_pm_domain *genpd, bool enable);
+int dev_pm_genpd_set_next_wakeup(struct device *dev, ktime_t next);
 
 extern struct dev_power_governor simple_qos_governor;
 extern struct dev_power_governor pm_domain_always_on_gov;
@@ -291,6 +294,11 @@  static void genpd_enable_next_wakeup(struct generic_pm_domain *genpd,
 				     bool enable)
 { }
 
+static inline int dev_pm_genpd_set_next_wakeup(struct device *dev, ktime_t next)
+{
+	return -EOPNOTSUPP;
+}
+
 #define simple_qos_governor		(*(struct dev_power_governor *)(NULL))
 #define pm_domain_always_on_gov		(*(struct dev_power_governor *)(NULL))
 #endif