diff mbox series

[v2] cpuidle: psci: Add trace for PSCI domain idle

Message ID 20250120013616.1116126-1-keyz@google.com (mailing list archive)
State New
Headers show
Series [v2] cpuidle: psci: Add trace for PSCI domain idle | expand

Commit Message

Keita Morisaki Jan. 20, 2025, 1:36 a.m. UTC
The trace event cpu_idle provides insufficient information for debugging
PSCI requests due to lacking access to determined PSCI domain idle
states. The cpu_idle usually only shows -1, 0, or 1 regardless how many
idle states the power domain has.

Add new trace events namely psci_domain_idle_enter and
psci_domain_idle_exit to trace enter and exit events with a determined
idle state.

These new trace events will help developers debug CPUidle issues on ARM
systems using PSCI by providing more detailed information about the
requested idle states.

Signed-off-by: Keita Morisaki <keyz@google.com>
---
v1->v2: Split the ftrace event into two (psci_domain_idle_(enter|exit))
		and rephrase the commit message accordingly. Rebased onto the latest.

 drivers/cpuidle/cpuidle-psci.c |  3 +++
 include/trace/events/power.h   | 37 ++++++++++++++++++++++++++++++++++
 2 files changed, 40 insertions(+)


base-commit: ffd294d346d185b70e28b1a28abe367bbfe53c04
--
2.48.0.rc2.279.g1de40edade-goog

Comments

Steven Rostedt Jan. 24, 2025, 8:28 p.m. UTC | #1
On Mon, 20 Jan 2025 09:36:16 +0800
Keita Morisaki <keyz@google.com> wrote:

> The trace event cpu_idle provides insufficient information for debugging
> PSCI requests due to lacking access to determined PSCI domain idle
> states. The cpu_idle usually only shows -1, 0, or 1 regardless how many
> idle states the power domain has.
> 
> Add new trace events namely psci_domain_idle_enter and
> psci_domain_idle_exit to trace enter and exit events with a determined
> idle state.
> 
> These new trace events will help developers debug CPUidle issues on ARM
> systems using PSCI by providing more detailed information about the
> requested idle states.
> 
> Signed-off-by: Keita Morisaki <keyz@google.com>
> ---
> v1->v2: Split the ftrace event into two (psci_domain_idle_(enter|exit))
> 		and rephrase the commit message accordingly. Rebased onto the latest.
> 
>  drivers/cpuidle/cpuidle-psci.c |  3 +++
>  include/trace/events/power.h   | 37 ++++++++++++++++++++++++++++++++++
>  2 files changed, 40 insertions(+)

From the tracing point of view, there's nothing that sticks out that is incorrect.

Reviewed-by: Steven Rostedt (Google) <rostedt@goodmis.org>

As for if it should be applied, that's for the maintainers of the subsystem
that it exists in.

-- Steve
Keita Morisaki Jan. 25, 2025, 1:27 a.m. UTC | #2
> Reviewed-by: Steven Rostedt (Google) <rostedt@goodmis.org>

> As for if it should be applied, that's for the maintainers of the subsystem
> that it exists in.

Thank you so much!
Christian Loehle Jan. 27, 2025, 12:16 p.m. UTC | #3
On 1/20/25 01:36, Keita Morisaki wrote:
> The trace event cpu_idle provides insufficient information for debugging
> PSCI requests due to lacking access to determined PSCI domain idle
> states. The cpu_idle usually only shows -1, 0, or 1 regardless how many
> idle states the power domain has.
> 
> Add new trace events namely psci_domain_idle_enter and
> psci_domain_idle_exit to trace enter and exit events with a determined
> idle state.
> 
> These new trace events will help developers debug CPUidle issues on ARM
> systems using PSCI by providing more detailed information about the
> requested idle states.
> 
> Signed-off-by: Keita Morisaki <keyz@google.com>
> ---
> v1->v2: Split the ftrace event into two (psci_domain_idle_(enter|exit))
> 		and rephrase the commit message accordingly. Rebased onto the latest.


Which makes it different to cpu_idle event FWIW.

> 
>  drivers/cpuidle/cpuidle-psci.c |  3 +++
>  include/trace/events/power.h   | 37 ++++++++++++++++++++++++++++++++++
>  2 files changed, 40 insertions(+)
> 
> diff --git a/drivers/cpuidle/cpuidle-psci.c b/drivers/cpuidle/cpuidle-psci.c
> index 2562dc001fc1..dd8d776d6e39 100644
> --- a/drivers/cpuidle/cpuidle-psci.c
> +++ b/drivers/cpuidle/cpuidle-psci.c
> @@ -25,6 +25,7 @@
>  #include <linux/syscore_ops.h>
> 
>  #include <asm/cpuidle.h>
> +#include <trace/events/power.h>
> 
>  #include "cpuidle-psci.h"
>  #include "dt_idle_states.h"
> @@ -74,7 +75,9 @@ static __cpuidle int __psci_enter_domain_idle_state(struct cpuidle_device *dev,
>  	if (!state)
>  		state = states[idx];
> 
> +	trace_psci_domain_idle_enter(dev->cpu, state, s2idle);
>  	ret = psci_cpu_suspend_enter(state) ? -1 : idx;
> +	trace_psci_domain_idle_exit(dev->cpu, state, s2idle);

Not tracking ret seems odd, is that fine?
Keita Morisaki Jan. 28, 2025, 4:24 a.m. UTC | #4
> > The trace event cpu_idle provides insufficient information for debugging
> > PSCI requests due to lacking access to determined PSCI domain idle
> > states. The cpu_idle usually only shows -1, 0, or 1 regardless how many
> > idle states the power domain has.
> >
> > Add new trace events namely psci_domain_idle_enter and
> > psci_domain_idle_exit to trace enter and exit events with a determined
> > idle state.
> >
> > These new trace events will help developers debug CPUidle issues on ARM
> > systems using PSCI by providing more detailed information about the
> > requested idle states.
> >
> > Signed-off-by: Keita Morisaki <keyz@google.com>
> > ---
> > v1->v2: Split the ftrace event into two (psci_domain_idle_(enter|exit))
> > 		and rephrase the commit message accordingly. Rebased onto the latest.
> Which makes it different to cpu_idle event FWIW.

Yes, psci_domain_idle_(enter|exit) are not meant to replace cpu_idle nor a
variant of it. It's new and different events that provide finer=grained info.

> >  drivers/cpuidle/cpuidle-psci.c |  3 +++
> >  include/trace/events/power.h   | 37 ++++++++++++++++++++++++++++++++++
> >  2 files changed, 40 insertions(+)
> >
> > diff --git a/drivers/cpuidle/cpuidle-psci.c b/drivers/cpuidle/cpuidle-psci.c
> > index 2562dc001fc1..dd8d776d6e39 100644
> > --- a/drivers/cpuidle/cpuidle-psci.c
> > +++ b/drivers/cpuidle/cpuidle-psci.c
> > @@ -25,6 +25,7 @@
> >  #include <linux/syscore_ops.h>
> >
> >  #include <asm/cpuidle.h>
> > +#include <trace/events/power.h>
> >
> >  #include "cpuidle-psci.h"
> >  #include "dt_idle_states.h"
> > @@ -74,7 +75,9 @@ static __cpuidle int __psci_enter_domain_idle_state(struct cpuidle_device *dev,
> >  	if (!state)
> >  		state = states[idx];
> >
> > +	trace_psci_domain_idle_enter(dev->cpu, state, s2idle);
> >  	ret = psci_cpu_suspend_enter(state) ? -1 : idx;
> > +	trace_psci_domain_idle_exit(dev->cpu, state, s2idle);
> Not tracking ret seems odd, is that fine?

I think it's fine not to track ret here.
__psci_enter_domain_idle_state does not seems to care the return value of
psci_cpu_suspend_enter to me because it just proceeds with executing subsequent
functions regardless of ret, and returns ret to the higher function. If the
value should be traced, it should probably be done in a lower layer or a higher
layer.

Another small small reason I'm not interested in adding ret to the
trace_psci_domain_idle_exit's arguments is that
trace_psci_domain_idle_(enter|exit) currently share the same trace event
(i.e. same set of arguments) and it makes the trace events simple.
Christian Loehle Jan. 28, 2025, 11:06 a.m. UTC | #5
On 1/28/25 04:24, Keita Morisaki wrote:
>>> The trace event cpu_idle provides insufficient information for debugging
>>> PSCI requests due to lacking access to determined PSCI domain idle
>>> states. The cpu_idle usually only shows -1, 0, or 1 regardless how many
>>> idle states the power domain has.
>>>
>>> Add new trace events namely psci_domain_idle_enter and
>>> psci_domain_idle_exit to trace enter and exit events with a determined
>>> idle state.
>>>
>>> These new trace events will help developers debug CPUidle issues on ARM
>>> systems using PSCI by providing more detailed information about the
>>> requested idle states.
>>>
>>> Signed-off-by: Keita Morisaki <keyz@google.com>
>>> ---
>>> v1->v2: Split the ftrace event into two (psci_domain_idle_(enter|exit))
>>> 		and rephrase the commit message accordingly. Rebased onto the latest.
>> Which makes it different to cpu_idle event FWIW.
> 
> Yes, psci_domain_idle_(enter|exit) are not meant to replace cpu_idle nor a
> variant of it. It's new and different events that provide finer=grained info.
> 

I mentioned it because it means it doesn't benefit from cpu_idle tooling
directly, which is slightly odd, but fine with me.
Keita Morisaki Jan. 29, 2025, 8:51 a.m. UTC | #6
> > Yes, psci_domain_idle_(enter|exit) are not meant to replace cpu_idle nor a
> > variant of it. It's new and different events that provide finer=grained info.

> I mentioned it because it means it doesn't benefit from cpu_idle tooling
> directly, which is slightly odd, but fine with me.

I might not fully understand your comments.
Do you mean that even mentioning cpu_idle in the commit message does not feel
right to you, or utilizing cpu_idle by exposing the determined state instead of
adding new trace events is the right direction?
diff mbox series

Patch

diff --git a/drivers/cpuidle/cpuidle-psci.c b/drivers/cpuidle/cpuidle-psci.c
index 2562dc001fc1..dd8d776d6e39 100644
--- a/drivers/cpuidle/cpuidle-psci.c
+++ b/drivers/cpuidle/cpuidle-psci.c
@@ -25,6 +25,7 @@ 
 #include <linux/syscore_ops.h>

 #include <asm/cpuidle.h>
+#include <trace/events/power.h>

 #include "cpuidle-psci.h"
 #include "dt_idle_states.h"
@@ -74,7 +75,9 @@  static __cpuidle int __psci_enter_domain_idle_state(struct cpuidle_device *dev,
 	if (!state)
 		state = states[idx];

+	trace_psci_domain_idle_enter(dev->cpu, state, s2idle);
 	ret = psci_cpu_suspend_enter(state) ? -1 : idx;
+	trace_psci_domain_idle_exit(dev->cpu, state, s2idle);

 	if (s2idle)
 		dev_pm_genpd_resume(pd_dev);
diff --git a/include/trace/events/power.h b/include/trace/events/power.h
index d2349b6b531a..9253e83b9bb4 100644
--- a/include/trace/events/power.h
+++ b/include/trace/events/power.h
@@ -62,6 +62,43 @@  TRACE_EVENT(cpu_idle_miss,
 		(unsigned long)__entry->state, (__entry->below)?"below":"above")
 );

+DECLARE_EVENT_CLASS(psci_domain_idle,
+
+	TP_PROTO(unsigned int cpu_id, unsigned int state, bool s2idle),
+
+	TP_ARGS(cpu_id, state, s2idle),
+
+	TP_STRUCT__entry(
+		__field(u32,		cpu_id)
+		__field(u32,		state)
+		__field(bool,		s2idle)
+	),
+
+	TP_fast_assign(
+		__entry->cpu_id = cpu_id;
+		__entry->state = state;
+		__entry->s2idle = s2idle;
+	),
+
+	TP_printk("cpu_id=%lu state=0x%lx is_s2idle=%s",
+		  (unsigned long)__entry->cpu_id, (unsigned long)__entry->state,
+		  (__entry->s2idle)?"yes":"no")
+);
+
+DEFINE_EVENT(psci_domain_idle, psci_domain_idle_enter,
+
+	TP_PROTO(unsigned int cpu_id, unsigned int state, bool s2idle),
+
+	TP_ARGS(cpu_id, state, s2idle)
+);
+
+DEFINE_EVENT(psci_domain_idle, psci_domain_idle_exit,
+
+	TP_PROTO(unsigned int cpu_id, unsigned int state, bool s2idle),
+
+	TP_ARGS(cpu_id, state, s2idle)
+);
+
 TRACE_EVENT(powernv_throttle,

 	TP_PROTO(int chip_id, const char *reason, int pmax),