[PATCH v3 0/2] Avoid calling cpu_pm functions for retention idle states

[PATCH v3 0/2] Avoid calling cpu_pm functions for retention idle states

[Prashanth Prakash]

CPU_PM_CPU_IDLE_ENTER() treats all idle states whose idx != 0 as a
state that loses some context, but we can have deeper idle states that
doesn't lose any software context. If a CPU is entering such a low power
idle state where it retains the context, then there is no need to call
cpu_pm_enter()/cpu_pm_exit().

Add a new macro(CPU_PM_CPU_IDLE_ENTER_RETENTION) to be used by cpuidle
drivers when they are entering retention state. By not calling cpu_pm_enter
and cpu_pm_exit we reduce the latency involved in entering and exiting
the retention states.

On ARM64 based Qualcomm server platform we measured below overhead for
for calling cpu_pm_enter and cpu_pm_exit for retention states.

workload: stress --hdd #CPUs --hdd-bytes 32M  -t 30
Overhead of cpu_pm_enter - 1.2us(Average), 6.5us(Max)
Overhead of cpu_pm_exit  - 3.1us(Average), 11.1us(Max)

Listed below are 5 functions that were notified on ENTER/EXIT on the
test platform:
      gic_cpu_pm_notifier
      arch_timer_cpu_pm_notify
      cpu_pm_pmu_notify
      hyp_init_cpu_pm_notifier
      fpsimd_cpu_pm_notifier

Prashanth Prakash (2):
  cpuidle: Add new macro to enter a retention idle state
  ARM64 / cpuidle: Use new cpuidle macro for entering retention state

Changes in v3:
 - Added additional clarification to commit message (Sudeep)

Changes in v2:
 - Reordered cpuidle.h macros for better readablity (Rafael)

 arch/arm64/kernel/cpuidle.c |  8 +++++++-
 include/linux/cpuidle.h     | 40 ++++++++++++++++++++++++----------------
 2 files changed, 31 insertions(+), 17 deletions(-)

-- 
Qualcomm Datacenter Technologies on behalf of Qualcomm Technologies, Inc.
Qualcomm Technologies, Inc. is a member of the
Code Aurora Forum, a Linux Foundation Collaborative Project.

[PATCH v3 1/2] cpuidle: Add new macro to enter a retention idle state

[Prashanth Prakash]

If a CPU is entering a low power idle state where it doesn't lose any
context, then there is no need to call cpu_pm_enter()/cpu_pm_exit().
Add a new macro(CPU_PM_CPU_IDLE_ENTER_RETENTION) to be used by cpuidle
drivers when they are entering retention state. By not calling
cpu_pm_enter and cpu_pm_exit we reduce the latency involved in
entering and exiting the retention idle states.

CPU_PM_CPU_IDLE_ENTER_RETENTION assumes that no state is lost and
hence CPU PM notifiers will not be called. We may need a broader
change if we need to support partial retention states effeciently.

On ARM64 based Qualcomm Server Platform we measured below overhead for
for calling cpu_pm_enter and cpu_pm_exit for retention states.

workload: stress --hdd #CPUs --hdd-bytes 32M  -t 30
        Average overhead of cpu_pm_enter - 1.2us
        Average overhead of cpu_pm_exit  - 3.1us

Signed-off-by: Prashanth Prakash <pprakash@codeaurora.org>
Acked-by: Sudeep Holla <sudeep.holla@arm.com>
---
 include/linux/cpuidle.h | 40 ++++++++++++++++++++++++----------------
 1 file changed, 24 insertions(+), 16 deletions(-)

diff --git a/include/linux/cpuidle.h b/include/linux/cpuidle.h
index 8f7788d..871f9e2 100644
--- a/include/linux/cpuidle.h
+++ b/include/linux/cpuidle.h
@@ -257,22 +257,30 @@ static inline int cpuidle_register_governor(struct cpuidle_governor *gov)
 {return 0;}
 #endif
 
-#define CPU_PM_CPU_IDLE_ENTER(low_level_idle_enter, idx)	\
-({								\
-	int __ret;						\
-								\
-	if (!idx) {						\
-		cpu_do_idle();					\
-		return idx;					\
-	}							\
-								\
-	__ret = cpu_pm_enter();					\
-	if (!__ret) {						\
-		__ret = low_level_idle_enter(idx);		\
-		cpu_pm_exit();					\
-	}							\
-								\
-	__ret ? -1 : idx;					\
+#define __CPU_PM_CPU_IDLE_ENTER(low_level_idle_enter, idx, is_retention) \
+({									\
+	int __ret = 0;							\
+									\
+	if (!idx) {							\
+		cpu_do_idle();						\
+		return idx;						\
+	}								\
+									\
+	if (!is_retention)						\
+		__ret =  cpu_pm_enter();				\
+	if (!__ret) {							\
+		__ret = low_level_idle_enter(idx);			\
+		if (!is_retention)					\
+			cpu_pm_exit();					\
+	}								\
+									\
+	__ret ? -1 : idx;						\
 })
 
+#define CPU_PM_CPU_IDLE_ENTER(low_level_idle_enter, idx)	\
+	__CPU_PM_CPU_IDLE_ENTER(low_level_idle_enter, idx, 0)
+
+#define CPU_PM_CPU_IDLE_ENTER_RETENTION(low_level_idle_enter, idx)	\
+	__CPU_PM_CPU_IDLE_ENTER(low_level_idle_enter, idx, 1)
+
 #endif /* _LINUX_CPUIDLE_H */
-- 
Qualcomm Datacenter Technologies on behalf of Qualcomm Technologies, Inc.
Qualcomm Technologies, Inc. is a member of the
Code Aurora Forum, a Linux Foundation Collaborative Project.

Re: [PATCH v3 1/2] cpuidle: Add new macro to enter a retention idle state

[Rafael J. Wysocki]

On Wednesday, November 15, 2017 6:11:49 PM CET Prashanth Prakash wrote:
> If a CPU is entering a low power idle state where it doesn't lose any
> context, then there is no need to call cpu_pm_enter()/cpu_pm_exit().
> Add a new macro(CPU_PM_CPU_IDLE_ENTER_RETENTION) to be used by cpuidle
> drivers when they are entering retention state. By not calling
> cpu_pm_enter and cpu_pm_exit we reduce the latency involved in
> entering and exiting the retention idle states.
> 
> CPU_PM_CPU_IDLE_ENTER_RETENTION assumes that no state is lost and
> hence CPU PM notifiers will not be called. We may need a broader
> change if we need to support partial retention states effeciently.
> 
> On ARM64 based Qualcomm Server Platform we measured below overhead for
> for calling cpu_pm_enter and cpu_pm_exit for retention states.
> 
> workload: stress --hdd #CPUs --hdd-bytes 32M  -t 30
>         Average overhead of cpu_pm_enter - 1.2us
>         Average overhead of cpu_pm_exit  - 3.1us
> 
> Signed-off-by: Prashanth Prakash <pprakash@codeaurora.org>
> Acked-by: Sudeep Holla <sudeep.holla@arm.com>
> ---
>  include/linux/cpuidle.h | 40 ++++++++++++++++++++++++----------------
>  1 file changed, 24 insertions(+), 16 deletions(-)
> 
> diff --git a/include/linux/cpuidle.h b/include/linux/cpuidle.h
> index 8f7788d..871f9e2 100644
> --- a/include/linux/cpuidle.h
> +++ b/include/linux/cpuidle.h
> @@ -257,22 +257,30 @@ static inline int cpuidle_register_governor(struct cpuidle_governor *gov)
>  {return 0;}
>  #endif
>  
> -#define CPU_PM_CPU_IDLE_ENTER(low_level_idle_enter, idx)	\
> -({								\
> -	int __ret;						\
> -								\
> -	if (!idx) {						\
> -		cpu_do_idle();					\
> -		return idx;					\
> -	}							\
> -								\
> -	__ret = cpu_pm_enter();					\
> -	if (!__ret) {						\
> -		__ret = low_level_idle_enter(idx);		\
> -		cpu_pm_exit();					\
> -	}							\
> -								\
> -	__ret ? -1 : idx;					\
> +#define __CPU_PM_CPU_IDLE_ENTER(low_level_idle_enter, idx, is_retention) \
> +({									\
> +	int __ret = 0;							\
> +									\
> +	if (!idx) {							\
> +		cpu_do_idle();						\
> +		return idx;						\
> +	}								\
> +									\
> +	if (!is_retention)						\
> +		__ret =  cpu_pm_enter();				\
> +	if (!__ret) {							\
> +		__ret = low_level_idle_enter(idx);			\
> +		if (!is_retention)					\
> +			cpu_pm_exit();					\
> +	}								\
> +									\
> +	__ret ? -1 : idx;						\
>  })
>  
> +#define CPU_PM_CPU_IDLE_ENTER(low_level_idle_enter, idx)	\
> +	__CPU_PM_CPU_IDLE_ENTER(low_level_idle_enter, idx, 0)
> +
> +#define CPU_PM_CPU_IDLE_ENTER_RETENTION(low_level_idle_enter, idx)	\
> +	__CPU_PM_CPU_IDLE_ENTER(low_level_idle_enter, idx, 1)
> +
>  #endif /* _LINUX_CPUIDLE_H */
> 

This change is fine by me, so you can add

Acked-by: Rafael J. Wysocki <rafael.j.wysocki@intel.com>

to it, but it kind of doesn't make sense to apply it alone, so please feel free
to route it via ARM64.

Thanks,
Rafael

[PATCH v3 2/2] ARM64 / cpuidle: Use new cpuidle macro for entering retention state

[Prashanth Prakash]

CPU_PM_CPU_IDLE_ENTER_RETENTION skips calling cpu_pm_enter() and
cpu_pm_exit(). By not calling cpu_pm functions in idle entry/exit
paths we can reduce the latency involved in entering and exiting
the low power idle state.

On ARM64 based Qualcomm server platform we measured below overhead
for calling cpu_pm_enter and cpu_pm_exit for retention states.

workload: stress --hdd #CPUs --hdd-bytes 32M  -t 30
	Average overhead of cpu_pm_enter - 1.2us
	Average overhead of cpu_pm_exit  - 3.1us

Signed-off-by: Prashanth Prakash <pprakash@codeaurora.org>
---
 arch/arm64/kernel/cpuidle.c | 8 +++++++-
 1 file changed, 7 insertions(+), 1 deletion(-)

diff --git a/arch/arm64/kernel/cpuidle.c b/arch/arm64/kernel/cpuidle.c
index fd69108..f2d1381 100644
--- a/arch/arm64/kernel/cpuidle.c
+++ b/arch/arm64/kernel/cpuidle.c
@@ -47,6 +47,8 @@ int arm_cpuidle_suspend(int index)
 
 #include <acpi/processor.h>
 
+#define ARM64_LPI_IS_RETENTION_STATE(arch_flags) (!(arch_flags))
+
 int acpi_processor_ffh_lpi_probe(unsigned int cpu)
 {
 	return arm_cpuidle_init(cpu);
@@ -54,6 +56,10 @@ int acpi_processor_ffh_lpi_probe(unsigned int cpu)
 
 int acpi_processor_ffh_lpi_enter(struct acpi_lpi_state *lpi)
 {
-	return CPU_PM_CPU_IDLE_ENTER(arm_cpuidle_suspend, lpi->index);
+	if (ARM64_LPI_IS_RETENTION_STATE(lpi->arch_flags))
+		return CPU_PM_CPU_IDLE_ENTER_RETENTION(arm_cpuidle_suspend,
+						lpi->index);
+	else
+		return CPU_PM_CPU_IDLE_ENTER(arm_cpuidle_suspend, lpi->index);
 }
 #endif
-- 
Qualcomm Datacenter Technologies on behalf of Qualcomm Technologies, Inc.
Qualcomm Technologies, Inc. is a member of the
Code Aurora Forum, a Linux Foundation Collaborative Project.

Re: [PATCH v3 2/2] ARM64 / cpuidle: Use new cpuidle macro for entering retention state

[Sudeep Holla]

On 15/11/17 17:11, Prashanth Prakash wrote:
> CPU_PM_CPU_IDLE_ENTER_RETENTION skips calling cpu_pm_enter() and
> cpu_pm_exit(). By not calling cpu_pm functions in idle entry/exit
> paths we can reduce the latency involved in entering and exiting
> the low power idle state.
> 
> On ARM64 based Qualcomm server platform we measured below overhead
> for calling cpu_pm_enter and cpu_pm_exit for retention states.
> 
> workload: stress --hdd #CPUs --hdd-bytes 32M  -t 30
> 	Average overhead of cpu_pm_enter - 1.2us
> 	Average overhead of cpu_pm_exit  - 3.1us
> 
> Signed-off-by: Prashanth Prakash <pprakash@codeaurora.org>

Acked-by: Sudeep Holla <sudeep.holla@arm.com>

-- 
Regards,
Sudeep

Re: [PATCH v3 0/2] Avoid calling cpu_pm functions for retention idle states

[Catalin Marinas]

On Wed, Nov 15, 2017 at 10:11:48AM -0700, Prashanth Prakash wrote:
> CPU_PM_CPU_IDLE_ENTER() treats all idle states whose idx != 0 as a
> state that loses some context, but we can have deeper idle states that
> doesn't lose any software context. If a CPU is entering such a low power
> idle state where it retains the context, then there is no need to call
> cpu_pm_enter()/cpu_pm_exit().
> 
> Add a new macro(CPU_PM_CPU_IDLE_ENTER_RETENTION) to be used by cpuidle
> drivers when they are entering retention state. By not calling cpu_pm_enter
> and cpu_pm_exit we reduce the latency involved in entering and exiting
> the retention states.
> 
> On ARM64 based Qualcomm server platform we measured below overhead for
> for calling cpu_pm_enter and cpu_pm_exit for retention states.
> 
> workload: stress --hdd #CPUs --hdd-bytes 32M  -t 30
> Overhead of cpu_pm_enter - 1.2us(Average), 6.5us(Max)
> Overhead of cpu_pm_exit  - 3.1us(Average), 11.1us(Max)

I queued the patches in the arm64 tree for 4.16. Thanks.

-- 
Catalin