Hi,
Please find below a patch, which has perf_events added for pseries (ppc64)
platform in order to emit the trace required for perf timechart.
It essentially enables perf timechart for pseries platfrom to analyse
power savings events like cpuidle states.
Steps to enable and disable the trace;
1) Mount debugfs;
mount -t debugfs none /sys/kernel/debug
2) Then, enable the event using;
echo 1 > /sys/kernel/debug/tracing/events/power/cpu_idle/enable
3) The output of the trace can be seen in /sys/kernel/debug/tracing/trace
4) To disable the trace use;
echo 0 > sys/kernel/debug/tracing/events/power/cpu_idle/enable
Trace .svg o/p can be viewed for pseries (ppc64) systems showing various
cpu-idle states as a part of perf timechart tool.
References: http://blog.fenrus.org/?p=5
Issue command 'perf timechart record' to enable tracing.
This generates the trace and records in perf.data file by default.
One can generate output.svg file by issuing 'perf timechart'.
Sample o/p from the trace file:
===============================
State 1 -> Snooze
State 2 -> Cede
# tracer: nop
#
TASK-PID CPU# TIMESTAMP FUNCTION
| | | | |
<idle>-0 [000] 292.482314: cpu_idle: state=1 cpu_id=0
^^ Enter Snooze
<idle>-0 [001] 292.482363: cpu_idle: state=1 cpu_id=1
<idle>-0 [000] 292.492315: cpu_idle: state=4294967295 cpu_id=0
^^ Exit Snooze
<idle>-0 [000] 292.492316: cpu_idle: state=2 cpu_id=0
^^ Enter Cede
<idle>-0 [001] 292.492364: cpu_idle: state=4294967295 cpu_id=1
<idle>-0 [001] 292.492364: cpu_idle: state=2 cpu_id=1
<idle>-0 [000] 292.504198: cpu_idle: state=4294967295 cpu_id=0
^^Exit Cede
<idle>-0 [000] 292.504204: cpu_idle: state=1 cpu_id=0
<idle>-0 [001] 292.504921: cpu_idle: state=4294967295 cpu_id=1
<idle>-0 [001] 292.504936: cpu_idle: state=1 cpu_id=1
<idle>-0 [000] 292.514205: cpu_idle: state=4294967295 cpu_id=0
This patch applies on 2.6.39 and tested on a IBM POWER7 machine.
-Deepthi
Adding perf events to trace various cpu idle states on ppc64 (pseries) platform.
Signed-off-by: Deepthi Dharwar <[email protected]>
pseries.h | 4 ++++
setup.c | 14 ++++++++++++++
2 files changed, 18 insertions(+)
Index: linux-2.6.39/arch/powerpc/platforms/pseries/setup.c
===================================================================
--- linux-2.6.39.orig/arch/powerpc/platforms/pseries/setup.c 2011-05-19 00:06:34.000000000 -0400
+++ linux-2.6.39/arch/powerpc/platforms/pseries/setup.c 2011-06-01 07:46:00.000000000 -0400
@@ -39,6 +39,7 @@
#include <linux/irq.h>
#include <linux/seq_file.h>
#include <linux/root_dev.h>
+#include <trace/events/power.h>
#include <asm/mmu.h>
#include <asm/processor.h>
@@ -582,6 +583,10 @@
* while, do so.
*/
if (snooze) {
+
+ trace_power_start(POWER_CSTATE, CPU_IDLE_SNOOZE, cpu);
+ trace_cpu_idle(CPU_IDLE_SNOOZE, cpu);
+
start_snooze = get_tb() + snooze * tb_ticks_per_usec;
local_irq_enable();
set_thread_flag(TIF_POLLING_NRFLAG);
@@ -602,9 +607,19 @@
goto out;
}
+ trace_power_end(cpu);
+ trace_cpu_idle(PWR_EVENT_EXIT, cpu);
+
+ trace_power_start(POWER_CSTATE, CPU_IDLE_CEDE, cpu);
+ trace_cpu_idle(CPU_IDLE_CEDE, cpu);
+
cede_processor();
out:
+
+ trace_power_end(cpu);
+ trace_cpu_idle(PWR_EVENT_EXIT, cpu);
+
HMT_medium();
out_purr = mfspr(SPRN_PURR);
get_lppaca()->wait_state_cycles += out_purr - in_purr;
Index: linux-2.6.39/arch/powerpc/platforms/pseries/pseries.h
===================================================================
--- linux-2.6.39.orig/arch/powerpc/platforms/pseries/pseries.h 2011-05-19 00:06:34.000000000 -0400
+++ linux-2.6.39/arch/powerpc/platforms/pseries/pseries.h 2011-06-01 07:53:24.000000000 -0400
@@ -12,6 +12,9 @@
#include <linux/interrupt.h>
+#define CPU_IDLE_SNOOZE 1
+#define CPU_IDLE_CEDE 2
+
struct device_node;
extern void request_event_sources_irqs(struct device_node *np,
@@ -56,4 +59,5 @@
extern int dlpar_attach_node(struct device_node *);
extern int dlpar_detach_node(struct device_node *);
+
#endif /* _PSERIES_PSERIES_H */
On Wed, 2011-06-01 at 18:05 +0530, Deepthi Dharwar wrote:
> Hi,
>
> Please find below a patch, which has perf_events added for pseries (ppc64)
> platform in order to emit the trace required for perf timechart.
> It essentially enables perf timechart for pseries platfrom to analyse
> power savings events like cpuidle states.
Unless I'm mistaken, you added traces to dedicated CPU idle sleep but
not shared processor. Any reason ?
Also I don't really know that tracing stuff but what's the point of
having start/end _and trace_cpu_idle if you're going to always start &
end around a single occurence of trace_cpu_idle ?
Wouldn't there be a way to start/end and then trace the snooze and
subsequent cede within the same start/end section or that makes no
sense ?
Also would there be any interest in doing the tracing more generically
in idle.c ?
Cheers,
Ben.
On Friday 17 June 2011 09:54 AM, Benjamin Herrenschmidt wrote:
> On Wed, 2011-06-01 at 18:05 +0530, Deepthi Dharwar wrote:
>> Hi,
>>
>> Please find below a patch, which has perf_events added for pseries (ppc64)
>> platform in order to emit the trace required for perf timechart.
>> It essentially enables perf timechart for pseries platfrom to analyse
>> power savings events like cpuidle states.
>
> Unless I'm mistaken, you added traces to dedicated CPU idle sleep but
> not shared processor. Any reason ?
>
Yes, the traces were added only to dedicated CPU idle sleep and not for
shared processor. This was added only for RFC purpose, and looking for
comments from trace implementation point of view. This can be
easily extended to the latter too.
> Also I don't really know that tracing stuff but what's the point of
> having start/end _and trace_cpu_idle if you're going to always start &
> end around a single occurence of trace_cpu_idle ?
>
power_start/end are the APIs that were used initially
and they are going to be deprecated in the upcoming kernel releases.
trace_cpu_idle call is going to replace power start/end routines.
To maintain backward compatibility and uniformity, both the routines
have been used.
(ref:https://lkml.org/lkml/2010/11/14/60)
> Wouldn't there be a way to start/end and then trace the snooze and
> subsequent cede within the same start/end section or that makes no
> sense ?
>
We wanted to find the residency time of both Snooze as well as cede
separately. Knowing this will help us tweak our cpuidle code. So, both
have been captured separately.
> Also would there be any interest in doing the tracing more generically
> in idle.c ?
>
Yes, this tracing is already implemented for Intel platform. This would
be a part of cpuidle framework. Going further, once the power cpuidle
framework is ported and ready, we will extend this trace there as well.
(ref:https://lkml.org/lkml/2011/6/7/375)
> Cheers,
> Ben.
>
> _______________________________________________
> Linuxppc-dev mailing list
> [email protected]
> https://lists.ozlabs.org/listinfo/linuxppc-dev
Regards,
Deepthi
On Mon, 2011-06-20 at 22:48 +0530, deepthi wrote:
> On Friday 17 June 2011 09:54 AM, Benjamin Herrenschmidt wrote:
> > On Wed, 2011-06-01 at 18:05 +0530, Deepthi Dharwar wrote:
> >> Hi,
> >>
> >> Please find below a patch, which has perf_events added for pseries (ppc64)
> >> platform in order to emit the trace required for perf timechart.
> >> It essentially enables perf timechart for pseries platfrom to analyse
> >> power savings events like cpuidle states.
> >
> > Unless I'm mistaken, you added traces to dedicated CPU idle sleep but
> > not shared processor. Any reason ?
> >
> Yes, the traces were added only to dedicated CPU idle sleep and not for
> shared processor. This was added only for RFC purpose, and looking for
> comments from trace implementation point of view. This can be
> easily extended to the latter too.
Please do both.
> > Also I don't really know that tracing stuff but what's the point of
> > having start/end _and trace_cpu_idle if you're going to always start &
> > end around a single occurence of trace_cpu_idle ?
> >
> power_start/end are the APIs that were used initially
> and they are going to be deprecated in the upcoming kernel releases.
> trace_cpu_idle call is going to replace power start/end routines.
> To maintain backward compatibility and uniformity, both the routines
> have been used.
> (ref:https://lkml.org/lkml/2010/11/14/60ref:https://lkml.org/lkml/2010/11/14/60)
Backward compatible with what ? Userspace ? Do we care in that specific
case since it's a new feature ?
> > Wouldn't there be a way to start/end and then trace the snooze and
> > subsequent cede within the same start/end section or that makes no
> > sense ?
> >
> We wanted to find the residency time of both Snooze as well as cede
> separately. Knowing this will help us tweak our cpuidle code. So, both
> have been captured separately.
>
> > Also would there be any interest in doing the tracing more generically
> > in idle.c ?
> >
> Yes, this tracing is already implemented for Intel platform. This would
> be a part of cpuidle framework. Going further, once the power cpuidle
> framework is ported and ready, we will extend this trace there as well.
> (ref:https://lkml.org/lkml/2011/6/7/375)
So do we need to apply this patch at all since the cpuidle stuff is
happening too ?
Cheers,
Ben.
On Tuesday 21 June 2011 03:12 AM, Benjamin Herrenschmidt wrote:
> On Mon, 2011-06-20 at 22:48 +0530, deepthi wrote:
>> On Friday 17 June 2011 09:54 AM, Benjamin Herrenschmidt wrote:
>>> On Wed, 2011-06-01 at 18:05 +0530, Deepthi Dharwar wrote:
>>>> Hi,
>>>>
>>>> Please find below a patch, which has perf_events added for pseries (ppc64)
>>>> platform in order to emit the trace required for perf timechart.
>>>> It essentially enables perf timechart for pseries platfrom to analyse
>>>> power savings events like cpuidle states.
>>>
>>> Unless I'm mistaken, you added traces to dedicated CPU idle sleep but
>>> not shared processor. Any reason ?
>>>
>> Yes, the traces were added only to dedicated CPU idle sleep and not for
>> shared processor. This was added only for RFC purpose, and looking for
>> comments from trace implementation point of view. This can be
>> easily extended to the latter too.
>
> Please do both.
>
Yes, I ll do so.
>>> Also I don't really know that tracing stuff but what's the point of
>>> having start/end _and trace_cpu_idle if you're going to always start &
>>> end around a single occurence of trace_cpu_idle ?
>>>
>> power_start/end are the APIs that were used initially
>> and they are going to be deprecated in the upcoming kernel releases.
>> trace_cpu_idle call is going to replace power start/end routines.
>> To maintain backward compatibility and uniformity, both the routines
>> have been used.
>> (ref:https://lkml.org/lkml/2010/11/14/60ref:https://lkml.org/lkml/2010/11/14/60)
>
> Backward compatible with what ? Userspace ? Do we care in that specific
> case since it's a new feature ?
>
Going forward, we can just have trace_cpu_idle call and
remove the power_start/end calls.
>>> Wouldn't there be a way to start/end and then trace the snooze and
>>> subsequent cede within the same start/end section or that makes no
>>> sense ?
>>>
>> We wanted to find the residency time of both Snooze as well as cede
>> separately. Knowing this will help us tweak our cpuidle code. So, both
>> have been captured separately.
>>
>>> Also would there be any interest in doing the tracing more generically
>>> in idle.c ?
>>>
>> Yes, this tracing is already implemented for Intel platform. This would
>> be a part of cpuidle framework. Going further, once the power cpuidle
>> framework is ported and ready, we will extend this trace there as well.
>> (ref:https://lkml.org/lkml/2011/6/7/375)
>
> So do we need to apply this patch at all since the cpuidle stuff is
> happening too ?
>
Well, not really. This is more for RFC purpose.
I just wanted to share this patch, as we are using it to evaluate
cpu idle on ppc64.
> Cheers,
> Ben.
>
>
> _______________________________________________
> Linuxppc-dev mailing list
> [email protected]
> https://lists.ozlabs.org/listinfo/linuxppc-dev
On Tuesday 21 June 2011 09:59 PM, deepthi wrote:
> On Tuesday 21 June 2011 03:12 AM, Benjamin Herrenschmidt wrote:
>> On Mon, 2011-06-20 at 22:48 +0530, deepthi wrote:
>>> On Friday 17 June 2011 09:54 AM, Benjamin Herrenschmidt wrote:
>>>> On Wed, 2011-06-01 at 18:05 +0530, Deepthi Dharwar wrote:
>>>>> Hi,
>>>>>
>>>>> Please find below a patch, which has perf_events added for pseries (ppc64)
>>>>> platform in order to emit the trace required for perf timechart.
>>>>> It essentially enables perf timechart for pseries platfrom to analyse
>>>>> power savings events like cpuidle states.
>>>>
>>>> Unless I'm mistaken, you added traces to dedicated CPU idle sleep but
>>>> not shared processor. Any reason ?
>>>>
>>> Yes, the traces were added only to dedicated CPU idle sleep and not for
>>> shared processor. This was added only for RFC purpose, and looking for
>>> comments from trace implementation point of view. This can be
>>> easily extended to the latter too.
>>
>> Please do both.
>>
> Yes, I ll do so.
>
>>>> Also I don't really know that tracing stuff but what's the point of
>>>> having start/end _and trace_cpu_idle if you're going to always start &
>>>> end around a single occurence of trace_cpu_idle ?
>>>>
>>> power_start/end are the APIs that were used initially
>>> and they are going to be deprecated in the upcoming kernel releases.
>>> trace_cpu_idle call is going to replace power start/end routines.
>>> To maintain backward compatibility and uniformity, both the routines
>>> have been used.
>>> (ref:https://lkml.org/lkml/2010/11/14/60ref:https://lkml.org/lkml/2010/11/14/60)
>>
>> Backward compatible with what ? Userspace ? Do we care in that specific
>> case since it's a new feature ?
>>
> Going forward, we can just have trace_cpu_idle call and
> remove the power_start/end calls.
>
>>>> Wouldn't there be a way to start/end and then trace the snooze and
>>>> subsequent cede within the same start/end section or that makes no
>>>> sense ?
>>>>
>>> We wanted to find the residency time of both Snooze as well as cede
>>> separately. Knowing this will help us tweak our cpuidle code. So, both
>>> have been captured separately.
>>>
>>>> Also would there be any interest in doing the tracing more generically
>>>> in idle.c ?
>>>>
>>> Yes, this tracing is already implemented for Intel platform. This would
>>> be a part of cpuidle framework. Going further, once the power cpuidle
>>> framework is ported and ready, we will extend this trace there as well.
>>> (ref:https://lkml.org/lkml/2011/6/7/375)
>>
>> So do we need to apply this patch at all since the cpuidle stuff is
>> happening too ?
>>
>
> Well, not really. This is more for RFC purpose.
> I just wanted to share this patch, as we are using it to evaluate
> cpu idle on ppc64.
>
I will re-base the patch and move it to the cpu idle for power
framework. So the tracing too gets in along with the
cpu idle support.
Thanks Ben.
>> Cheers,
>> Ben.
>>
>>
>> _______________________________________________
>> Linuxppc-dev mailing list
>> [email protected]
>> https://lists.ozlabs.org/listinfo/linuxppc-dev
>
> _______________________________________________
> Linuxppc-dev mailing list
> [email protected]
> https://lists.ozlabs.org/listinfo/linuxppc-dev
Regards,
Deepthi