[PATCH] PM / sleep: enable suspend-to-idle even without registered suspend_ops

[Sudeep Holla]

Suspend-to-idle (aka the "freeze" sleep state) is a system sleep state
in which all of the processors enter deepest possible idle state and
wait for interrupts right after suspending all the devices.

There is no hard requirement for a platform to support and register
platform specific suspend_ops to enter suspend-to-idle/freeze state.
Only deeper system sleep states like PM_SUSPEND_STANDBY and
PM_SUSPEND_MEM rely on such low level support/implementation.

suspend-to-idle can be entered as along as all the devices can be
suspended. This patch enables the support for suspend-to-idle even on
systems that don't have any low level support for deeper system sleep
states and/or don't register any platform specific suspend_ops.

Cc: "Rafael J. Wysocki" <r...@rjwysocki.net>
Signed-off-by: Sudeep Holla <sudeep...@arm.com>
---
kernel/power/main.c | 5 +++++
kernel/power/suspend.c | 8 +++++---
2 files changed, 10 insertions(+), 3 deletions(-)

Hi Rafael,

I am not sure if you like this approach. I found this to be the simplest
but I may have missed to consider all possible corner cases especially
for x86 and other platforms. I don't see any such issues/cases with ARM
systems.

Regards,
Sudeep

diff --git a/kernel/power/main.c b/kernel/power/main.c
index 5ea50b1b7595..0f0fd9184f39 100644
--- a/kernel/power/main.c
+++ b/kernel/power/main.c
@@ -651,6 +651,11 @@ static int __init pm_init(void)
if (error)
return error;
pm_print_times_init();
+ /*
+ * freeze state should be supported even without any suspend_ops,
+ * calling suspend_set_ops without any ops will setup freeze state
+ */
+ suspend_set_ops(NULL);
return pm_autosleep_init();
}

diff --git a/kernel/power/suspend.c b/kernel/power/suspend.c
index 0acab9d7f96f..37d64f811ecd 100644
--- a/kernel/power/suspend.c
+++ b/kernel/power/suspend.c
@@ -138,7 +138,9 @@ void suspend_set_ops(const struct platform_suspend_ops *ops)

lock_system_sleep();

- suspend_ops = ops;
+ WARN_ONCE(ops && suspend_ops, "overriding suspend_ops");
+ if (ops)
+ suspend_ops = ops;
for (i = PM_SUSPEND_MEM; i >= PM_SUSPEND_STANDBY; i--)
if (valid_state(i)) {
pm_states[i] = pm_labels[j++];
@@ -211,7 +213,7 @@ static int platform_suspend_begin(suspend_state_t state)
{
if (state == PM_SUSPEND_FREEZE && freeze_ops && freeze_ops->begin)
return freeze_ops->begin();
- else if (suspend_ops->begin)
+ else if (suspend_ops && suspend_ops->begin)
return suspend_ops->begin(state);
else
return 0;
@@ -221,7 +223,7 @@ static void platform_resume_end(suspend_state_t state)
{
if (state == PM_SUSPEND_FREEZE && freeze_ops && freeze_ops->end)
freeze_ops->end();
- else if (suspend_ops->end)
+ else if (suspend_ops && suspend_ops->end)
suspend_ops->end();
}

--
2.7.4

[Rafael J. Wysocki]

On Thursday, August 18, 2016 10:19:24 AM Sudeep Holla wrote:
> Suspend-to-idle (aka the "freeze" sleep state) is a system sleep state
> in which all of the processors enter deepest possible idle state and
> wait for interrupts right after suspending all the devices.
>
> There is no hard requirement for a platform to support and register
> platform specific suspend_ops to enter suspend-to-idle/freeze state.
> Only deeper system sleep states like PM_SUSPEND_STANDBY and
> PM_SUSPEND_MEM rely on such low level support/implementation.
>
> suspend-to-idle can be entered as along as all the devices can be
> suspended. This patch enables the support for suspend-to-idle even on
> systems that don't have any low level support for deeper system sleep
> states and/or don't register any platform specific suspend_ops.
>
> Cc: "Rafael J. Wysocki" <r...@rjwysocki.net>
> Signed-off-by: Sudeep Holla <sudeep...@arm.com>
> ---
> kernel/power/main.c | 5 +++++
> kernel/power/suspend.c | 8 +++++---
> 2 files changed, 10 insertions(+), 3 deletions(-)
>
> Hi Rafael,
>
> I am not sure if you like this approach. I found this to be the simplest
> but I may have missed to consider all possible corner cases especially
> for x86 and other platforms. I don't see any such issues/cases with ARM
> systems.
>

> diff --git a/kernel/power/main.c b/kernel/power/main.c
> index 5ea50b1b7595..0f0fd9184f39 100644
> --- a/kernel/power/main.c
> +++ b/kernel/power/main.c
> @@ -651,6 +651,11 @@ static int __init pm_init(void)
> if (error)
> return error;
> pm_print_times_init();
> + /*
> + * freeze state should be supported even without any suspend_ops,
> + * calling suspend_set_ops without any ops will setup freeze state
> + */
> + suspend_set_ops(NULL);

Well, this is a core initcall, so suspend_set_ops() invocations from platforms
really should happen after that, so something like this should be sufficient here:

pm_state[PM_SUSPEND_FREEZE] = pm_labels[relative_states ? PM_SUSPEND_MEM : PM_SUSPEND_FREEZE];

if I'm not mistaken.

> return pm_autosleep_init();
> }
> diff --git a/kernel/power/suspend.c b/kernel/power/suspend.c
> index 0acab9d7f96f..37d64f811ecd 100644
> --- a/kernel/power/suspend.c
> +++ b/kernel/power/suspend.c
> @@ -138,7 +138,9 @@ void suspend_set_ops(const struct platform_suspend_ops *ops)
>
> lock_system_sleep();
>
> - suspend_ops = ops;
> + WARN_ONCE(ops && suspend_ops, "overriding suspend_ops");
> + if (ops)
> + suspend_ops = ops;

And this should not be necessary then.

> for (i = PM_SUSPEND_MEM; i >= PM_SUSPEND_STANDBY; i--)
> if (valid_state(i)) {
> pm_states[i] = pm_labels[j++];
> @@ -211,7 +213,7 @@ static int platform_suspend_begin(suspend_state_t state)
> {
> if (state == PM_SUSPEND_FREEZE && freeze_ops && freeze_ops->begin)
> return freeze_ops->begin();
> - else if (suspend_ops->begin)
> + else if (suspend_ops && suspend_ops->begin)
> return suspend_ops->begin(state);
> else
> return 0;
> @@ -221,7 +223,7 @@ static void platform_resume_end(suspend_state_t state)
> {
> if (state == PM_SUSPEND_FREEZE && freeze_ops && freeze_ops->end)
> freeze_ops->end();
> - else if (suspend_ops->end)
> + else if (suspend_ops && suspend_ops->end)
> suspend_ops->end();
> }

And this is still needed, of course.

Thanks,
Rafael

[Sudeep Holla]

But won't this show up as "standby" state in /sys/power/state ?
Or did you mean:
pm_state[PM_SUSPEND_FREEZE] = pm_labels[relative_states ? 0 : 2];

which will be "mem" or "freeze" based on relative_states.

In-fact, I did exactly this when I first hacked. Since it exposed
incorrectly to sysfs and I was not sure of these relative_states and
it's usage, I preferred re-using suspend_set_ops for this.

IIUC showing freeze state as "mem" in sysfs is fine as that's the
deepest possible state when relative_states=1. But showing it as freeze
as "standby" in sysfs when relative_states=0 looked wrong to me though
it works as freeze state.

As a side-note with psci, it get called quite early before
core_initcall. But that can be fixed if needed and is different issue.

--
Regards,
Sudeep

[Sudeep Holla]

Suspend-to-idle (aka the "freeze" sleep state) is a system sleep state
in which all of the processors enter deepest possible idle state and
wait for interrupts right after suspending all the devices.

There is no hard requirement for a platform to support and register
platform specific suspend_ops to enter suspend-to-idle/freeze state.
Only deeper system sleep states like PM_SUSPEND_STANDBY and
PM_SUSPEND_MEM rely on such low level support/implementation.

Suspend-to-idle can be entered as along as all the devices can be

suspended. This patch enables the support for suspend-to-idle even on
systems that don't have any low level support for deeper system sleep
states and/or don't register any platform specific suspend_ops.

Cc: "Rafael J. Wysocki" <r...@rjwysocki.net>
Signed-off-by: Sudeep Holla <sudeep...@arm.com>
---

include/linux/suspend.h | 2 ++
kernel/power/main.c | 1 +
kernel/power/suspend.c | 14 +++++++++++---
3 files changed, 14 insertions(+), 3 deletions(-)

v1->v2:
- Introduced pm_freeze_init instead of (mis/re)using suspend_set_ops
to setup freeze state in absense of registered platform suspend_ops

diff --git a/include/linux/suspend.h b/include/linux/suspend.h
index 7693e39b14fe..0528ee32be7c 100644
--- a/include/linux/suspend.h
+++ b/include/linux/suspend.h
@@ -245,6 +245,7 @@ static inline bool idle_should_freeze(void)
return unlikely(suspend_freeze_state == FREEZE_STATE_ENTER);
}

+extern void __init pm_freeze_init(void);
extern void freeze_set_ops(const struct platform_freeze_ops *ops);
extern void freeze_wake(void);

@@ -279,6 +280,7 @@ static inline bool pm_resume_via_firmware(void) { return false; }
static inline void suspend_set_ops(const struct platform_suspend_ops *ops) {}
static inline int pm_suspend(suspend_state_t state) { return -ENOSYS; }
static inline bool idle_should_freeze(void) { return false; }
+static inline void __init pm_freeze_init(void) {}
static inline void freeze_set_ops(const struct platform_freeze_ops *ops) {}
static inline void freeze_wake(void) {}
#endif /* !CONFIG_SUSPEND */
diff --git a/kernel/power/main.c b/kernel/power/main.c
index 5ea50b1b7595..f27c2f2d0f22 100644
--- a/kernel/power/main.c
+++ b/kernel/power/main.c
@@ -651,6 +651,7 @@ static int __init pm_init(void)

if (error)
return error;
pm_print_times_init();

+ pm_freeze_init();

return pm_autosleep_init();
}

diff --git a/kernel/power/suspend.c b/kernel/power/suspend.c

index 0acab9d7f96f..1c962a366678 100644
--- a/kernel/power/suspend.c
+++ b/kernel/power/suspend.c
@@ -118,10 +118,18 @@ static bool valid_state(suspend_state_t state)
*/
static bool relative_states;

+void __init pm_freeze_init(void)
+{

+ /*
+ * freeze state should be supported even without any suspend_ops,

+ * initialize pm_states accordingly here
+ */
+ pm_states[PM_SUSPEND_FREEZE] = pm_labels[relative_states ? 0 : 2];
+}
+
static int __init sleep_states_setup(char *str)
{
relative_states = !strncmp(str, "1", 1);
- pm_states[PM_SUSPEND_FREEZE] = pm_labels[relative_states ? 0 : 2];
return 1;
}

@@ -211,7 +219,7 @@ static int platform_suspend_begin(suspend_state_t state)

{
if (state == PM_SUSPEND_FREEZE && freeze_ops && freeze_ops->begin)
return freeze_ops->begin();
- else if (suspend_ops->begin)
+ else if (suspend_ops && suspend_ops->begin)
return suspend_ops->begin(state);
else
return 0;

@@ -221,7 +229,7 @@ static void platform_resume_end(suspend_state_t state)

[Rafael J. Wysocki]

Why don't you call it pm_states_init()?

> static inline void freeze_set_ops(const struct platform_freeze_ops *ops) {}
> static inline void freeze_wake(void) {}
> #endif /* !CONFIG_SUSPEND */
> diff --git a/kernel/power/main.c b/kernel/power/main.c
> index 5ea50b1b7595..f27c2f2d0f22 100644
> --- a/kernel/power/main.c
> +++ b/kernel/power/main.c
> @@ -651,6 +651,7 @@ static int __init pm_init(void)
> if (error)
> return error;
> pm_print_times_init();
> + pm_freeze_init();

And please move this up (before the registration of power_kobj).

Thanks,
Rafael

[Sudeep Holla]

Suspend-to-idle (aka the "freeze" sleep state) is a system sleep state
in which all of the processors enter deepest possible idle state and
wait for interrupts right after suspending all the devices.

There is no hard requirement for a platform to support and register
platform specific suspend_ops to enter suspend-to-idle/freeze state.
Only deeper system sleep states like PM_SUSPEND_STANDBY and
PM_SUSPEND_MEM rely on such low level support/implementation.

suspend-to-idle can be entered as along as all the devices can be

suspended. This patch enables the support for suspend-to-idle even on
systems that don't have any low level support for deeper system sleep
states and/or don't register any platform specific suspend_ops.

Cc: "Rafael J. Wysocki" <r...@rjwysocki.net>
Signed-off-by: Sudeep Holla <sudeep...@arm.com>
---
include/linux/suspend.h | 2 ++
kernel/power/main.c | 1 +
kernel/power/suspend.c | 14 +++++++++++---
3 files changed, 14 insertions(+), 3 deletions(-)

v2->v3:
- Renamed pm_freeze_init to pm_states_init
- Moved pm_states_init earlier in pm_init

v1->v2:
- Introduced pm_freeze_init instead of (mis/re)using suspend_set_ops
to setup freeze state in absense of registered platform suspend_ops

diff --git a/include/linux/suspend.h b/include/linux/suspend.h

index 7693e39b14fe..d9718378a8be 100644

--- a/include/linux/suspend.h
+++ b/include/linux/suspend.h
@@ -245,6 +245,7 @@ static inline bool idle_should_freeze(void)
return unlikely(suspend_freeze_state == FREEZE_STATE_ENTER);
}

+extern void __init pm_states_init(void);

extern void freeze_set_ops(const struct platform_freeze_ops *ops);
extern void freeze_wake(void);

@@ -279,6 +280,7 @@ static inline bool pm_resume_via_firmware(void) { return false; }
static inline void suspend_set_ops(const struct platform_suspend_ops *ops) {}
static inline int pm_suspend(suspend_state_t state) { return -ENOSYS; }
static inline bool idle_should_freeze(void) { return false; }

+static inline void __init pm_states_init(void) {}

static inline void freeze_set_ops(const struct platform_freeze_ops *ops) {}
static inline void freeze_wake(void) {}
#endif /* !CONFIG_SUSPEND */
diff --git a/kernel/power/main.c b/kernel/power/main.c

index 5ea50b1b7595..281a697fd458 100644
--- a/kernel/power/main.c
+++ b/kernel/power/main.c
@@ -644,6 +644,7 @@ static int __init pm_init(void)
return error;
hibernate_image_size_init();
hibernate_reserved_size_init();
+ pm_states_init();
power_kobj = kobject_create_and_add("power", NULL);
if (!power_kobj)
return -ENOMEM;
diff --git a/kernel/power/suspend.c b/kernel/power/suspend.c
index 0acab9d7f96f..1e7f5da648d9 100644

--- a/kernel/power/suspend.c
+++ b/kernel/power/suspend.c
@@ -118,10 +118,18 @@ static bool valid_state(suspend_state_t state)
*/
static bool relative_states;

+void __init pm_states_init(void)

2.7.4

[Andy Gross]

On 19 August 2016 at 08:41, Sudeep Holla <sudeep...@arm.com> wrote:
> Suspend-to-idle (aka the "freeze" sleep state) is a system sleep state
> in which all of the processors enter deepest possible idle state and
> wait for interrupts right after suspending all the devices.
>
> There is no hard requirement for a platform to support and register
> platform specific suspend_ops to enter suspend-to-idle/freeze state.
> Only deeper system sleep states like PM_SUSPEND_STANDBY and
> PM_SUSPEND_MEM rely on such low level support/implementation.
>
> suspend-to-idle can be entered as along as all the devices can be
> suspended. This patch enables the support for suspend-to-idle even on
> systems that don't have any low level support for deeper system sleep
> states and/or don't register any platform specific suspend_ops.
>
> Cc: "Rafael J. Wysocki" <r...@rjwysocki.net>
> Signed-off-by: Sudeep Holla <sudeep...@arm.com>

I tested this patch on the Qualcomm 410c platform. It worked like a
charm. Thanks for working on this!

Tested-by: Andy Gross <andy....@linaro.org>

[Rafael J. Wysocki]

On Friday, August 19, 2016 02:41:00 PM Sudeep Holla wrote:
> Suspend-to-idle (aka the "freeze" sleep state) is a system sleep state
> in which all of the processors enter deepest possible idle state and
> wait for interrupts right after suspending all the devices.
>
> There is no hard requirement for a platform to support and register
> platform specific suspend_ops to enter suspend-to-idle/freeze state.
> Only deeper system sleep states like PM_SUSPEND_STANDBY and
> PM_SUSPEND_MEM rely on such low level support/implementation.
>
> suspend-to-idle can be entered as along as all the devices can be
> suspended. This patch enables the support for suspend-to-idle even on
> systems that don't have any low level support for deeper system sleep
> states and/or don't register any platform specific suspend_ops.
>
> Cc: "Rafael J. Wysocki" <r...@rjwysocki.net>
> Signed-off-by: Sudeep Holla <sudeep...@arm.com>

Applied.

Thanks,
Rafael