Add trace event to capture the moment of the call for updating the thermal
pressure value. It's helpful to investigate how often those events occur
in a system dealing with throttling. This trace event is needed since the
old 'cdev_update' might not be used by some drivers.
The old 'cdev_update' trace event only provides a cooling state
value: [0, n]. That state value then needs additional tools to translate
it: state -> freq -> capacity -> thermal pressure. This new trace event
just stores proper thermal pressure value in the trace buffer, no need
for additional logic. This is helpful for cooperation when someone can
simply sends to the list the trace buffer output from the platform (no
need from additional information from other subsystems).
There are also platforms which due to some design reasons don't use
cooling devices and thus don't trigger old 'cdev_update' trace event.
They are also important and measuring latency for the thermal signal
raising/decaying characteristics is in scope. This new trace event
would cover them as well.
We already have a trace point 'pelt_thermal_tp' which after a change to
trace event can be paired with this new 'thermal_pressure_update' and
derive more insight what is going on in the system under thermal pressure
(and why).
Signed-off-by: Lukasz Luba <[email protected]>
---
Changes v4:
- following Greg's comment, removed the test robot tag 'Reported-by'
which is not need for this patch; the compilation issue was captured
in v1 not in mainline code
The v3 is here:
https://lore.kernel.org/lkml/[email protected]/
The v2 is here:
https://lore.kernel.org/lkml/[email protected]/
The v1 and discussion can be found here at:
https://lore.kernel.org/lkml/[email protected]/
Regards,
Lukasz
drivers/base/arch_topology.c | 5 +++++
include/trace/events/thermal_pressure.h | 29 +++++++++++++++++++++++++
2 files changed, 34 insertions(+)
create mode 100644 include/trace/events/thermal_pressure.h
diff --git a/drivers/base/arch_topology.c b/drivers/base/arch_topology.c
index f73b836047cf..579c851a2bd7 100644
--- a/drivers/base/arch_topology.c
+++ b/drivers/base/arch_topology.c
@@ -19,6 +19,9 @@
#include <linux/rcupdate.h>
#include <linux/sched.h>
+#define CREATE_TRACE_POINTS
+#include <trace/events/thermal_pressure.h>
+
static DEFINE_PER_CPU(struct scale_freq_data __rcu *, sft_data);
static struct cpumask scale_freq_counters_mask;
static bool scale_freq_invariant;
@@ -195,6 +198,8 @@ void topology_update_thermal_pressure(const struct cpumask *cpus,
th_pressure = max_capacity - capacity;
+ trace_thermal_pressure_update(cpu, th_pressure);
+
for_each_cpu(cpu, cpus)
WRITE_ONCE(per_cpu(thermal_pressure, cpu), th_pressure);
}
diff --git a/include/trace/events/thermal_pressure.h b/include/trace/events/thermal_pressure.h
new file mode 100644
index 000000000000..b68680201360
--- /dev/null
+++ b/include/trace/events/thermal_pressure.h
@@ -0,0 +1,29 @@
+/* SPDX-License-Identifier: GPL-2.0 */
+#undef TRACE_SYSTEM
+#define TRACE_SYSTEM thermal_pressure
+
+#if !defined(_TRACE_THERMAL_PRESSURE_H) || defined(TRACE_HEADER_MULTI_READ)
+#define _TRACE_THERMAL_PRESSURE_H
+
+#include <linux/tracepoint.h>
+
+TRACE_EVENT(thermal_pressure_update,
+ TP_PROTO(int cpu, unsigned long thermal_pressure),
+ TP_ARGS(cpu, thermal_pressure),
+
+ TP_STRUCT__entry(
+ __field(unsigned long, thermal_pressure)
+ __field(int, cpu)
+ ),
+
+ TP_fast_assign(
+ __entry->thermal_pressure = thermal_pressure;
+ __entry->cpu = cpu;
+ ),
+
+ TP_printk("cpu=%d thermal_pressure=%lu", __entry->cpu, __entry->thermal_pressure)
+);
+#endif /* _TRACE_THERMAL_PRESSURE_H */
+
+/* This part must be outside protection */
+#include <trace/define_trace.h>
--
2.17.1
On Wed, Apr 27, 2022 at 09:08:06AM +0100, Lukasz Luba wrote:
> Add trace event to capture the moment of the call for updating the thermal
> pressure value. It's helpful to investigate how often those events occur
> in a system dealing with throttling. This trace event is needed since the
> old 'cdev_update' might not be used by some drivers.
>
> The old 'cdev_update' trace event only provides a cooling state
> value: [0, n]. That state value then needs additional tools to translate
> it: state -> freq -> capacity -> thermal pressure. This new trace event
> just stores proper thermal pressure value in the trace buffer, no need
> for additional logic. This is helpful for cooperation when someone can
> simply sends to the list the trace buffer output from the platform (no
> need from additional information from other subsystems).
>
> There are also platforms which due to some design reasons don't use
> cooling devices and thus don't trigger old 'cdev_update' trace event.
> They are also important and measuring latency for the thermal signal
> raising/decaying characteristics is in scope. This new trace event
> would cover them as well.
>
> We already have a trace point 'pelt_thermal_tp' which after a change to
> trace event can be paired with this new 'thermal_pressure_update' and
> derive more insight what is going on in the system under thermal pressure
> (and why).
>
> Signed-off-by: Lukasz Luba <[email protected]>
Much better, thanks. I'll wait for Sudeep's comments/review before
taking it into my tree.
greg k-h
On 4/27/22 10:20, Greg KH wrote:
> On Wed, Apr 27, 2022 at 09:08:06AM +0100, Lukasz Luba wrote:
>> Add trace event to capture the moment of the call for updating the thermal
>> pressure value. It's helpful to investigate how often those events occur
>> in a system dealing with throttling. This trace event is needed since the
>> old 'cdev_update' might not be used by some drivers.
>>
>> The old 'cdev_update' trace event only provides a cooling state
>> value: [0, n]. That state value then needs additional tools to translate
>> it: state -> freq -> capacity -> thermal pressure. This new trace event
>> just stores proper thermal pressure value in the trace buffer, no need
>> for additional logic. This is helpful for cooperation when someone can
>> simply sends to the list the trace buffer output from the platform (no
>> need from additional information from other subsystems).
>>
>> There are also platforms which due to some design reasons don't use
>> cooling devices and thus don't trigger old 'cdev_update' trace event.
>> They are also important and measuring latency for the thermal signal
>> raising/decaying characteristics is in scope. This new trace event
>> would cover them as well.
>>
>> We already have a trace point 'pelt_thermal_tp' which after a change to
>> trace event can be paired with this new 'thermal_pressure_update' and
>> derive more insight what is going on in the system under thermal pressure
>> (and why).
>>
>> Signed-off-by: Lukasz Luba <[email protected]>
>
> Much better, thanks. I'll wait for Sudeep's comments/review before
> taking it into my tree.
>
> greg k-h
Thank you Greg!
On Wed, Apr 27, 2022 at 09:08:06AM +0100, Lukasz Luba wrote:
> Add trace event to capture the moment of the call for updating the thermal
> pressure value. It's helpful to investigate how often those events occur
> in a system dealing with throttling. This trace event is needed since the
> old 'cdev_update' might not be used by some drivers.
>
> The old 'cdev_update' trace event only provides a cooling state
> value: [0, n]. That state value then needs additional tools to translate
> it: state -> freq -> capacity -> thermal pressure. This new trace event
> just stores proper thermal pressure value in the trace buffer, no need
> for additional logic. This is helpful for cooperation when someone can
> simply sends to the list the trace buffer output from the platform (no
> need from additional information from other subsystems).
>
> There are also platforms which due to some design reasons don't use
> cooling devices and thus don't trigger old 'cdev_update' trace event.
> They are also important and measuring latency for the thermal signal
> raising/decaying characteristics is in scope. This new trace event
> would cover them as well.
>
> We already have a trace point 'pelt_thermal_tp' which after a change to
> trace event can be paired with this new 'thermal_pressure_update' and
> derive more insight what is going on in the system under thermal pressure
> (and why).
>
Sorry for missing this, looks good to me.
Acked-by: Sudeep Holla <[email protected]>
--
Regards,
Sudeep