[PATCH] perf/x86: Fix open counting event error

[PATCH] perf/x86: Fix open counting event error

[Luo Gengkun]

Perf doesn't work at perf stat for hardware events:

 $perf stat -- sleep 1
 Performance counter stats for 'sleep 1':
             16.44 msec task-clock                       #    0.016 CPUs utilized
                 2      context-switches                 #  121.691 /sec
                 0      cpu-migrations                   #    0.000 /sec
                54      page-faults                      #    3.286 K/sec
   <not supported>	cycles
   <not supported>	instructions
   <not supported>	branches
   <not supported>	branch-misses

The reason is that the check in x86_pmu_hw_config for sampling event is
unexpectedly applied to the counting event.

Fixes: 88ec7eedbbd2 ("perf/x86: Fix low freqency setting issue")
Signed-off-by: Luo Gengkun <luogengkun@huaweicloud.com>
---
 arch/x86/events/core.c | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/arch/x86/events/core.c b/arch/x86/events/core.c
index 6866cc5acb0b..3a4f031d2f44 100644
--- a/arch/x86/events/core.c
+++ b/arch/x86/events/core.c
@@ -629,7 +629,7 @@ int x86_pmu_hw_config(struct perf_event *event)
 	if (event->attr.type == event->pmu->type)
 		event->hw.config |= x86_pmu_get_event_config(event);
 
-	if (!event->attr.freq && x86_pmu.limit_period) {
+	if (is_sampling_event(event) && !event->attr.freq && x86_pmu.limit_period) {
 		s64 left = event->attr.sample_period;
 		x86_pmu.limit_period(event, &left);
 		if (left > event->attr.sample_period)
-- 
2.34.1

Re: [PATCH] perf/x86: Fix open counting event error

[Liang, Kan]

On 2025-04-23 2:47 a.m., Luo Gengkun wrote:
> Perf doesn't work at perf stat for hardware events:
> 
>  $perf stat -- sleep 1
>  Performance counter stats for 'sleep 1':
>              16.44 msec task-clock                       #    0.016 CPUs utilized
>                  2      context-switches                 #  121.691 /sec
>                  0      cpu-migrations                   #    0.000 /sec
>                 54      page-faults                      #    3.286 K/sec
>    <not supported>	cycles
>    <not supported>	instructions
>    <not supported>	branches
>    <not supported>	branch-misses
> 
> The reason is that the check in x86_pmu_hw_config for sampling event is
> unexpectedly applied to the counting event.
> 
> Fixes: 88ec7eedbbd2 ("perf/x86: Fix low freqency setting issue")
> Signed-off-by: Luo Gengkun <luogengkun@huaweicloud.com>

Yes, it should only be applied for the sampling event. The
event->attr.freq is always 0 in the counting mode.

Reviewed-by: Kan Liang <kan.liang@linux.intel.com>

Thanks,
Kan> ---
>  arch/x86/events/core.c | 2 +-
>  1 file changed, 1 insertion(+), 1 deletion(-)
> 
> diff --git a/arch/x86/events/core.c b/arch/x86/events/core.c
> index 6866cc5acb0b..3a4f031d2f44 100644
> --- a/arch/x86/events/core.c
> +++ b/arch/x86/events/core.c
> @@ -629,7 +629,7 @@ int x86_pmu_hw_config(struct perf_event *event)
>  	if (event->attr.type == event->pmu->type)
>  		event->hw.config |= x86_pmu_get_event_config(event);
>  
> -	if (!event->attr.freq && x86_pmu.limit_period) {
> +	if (is_sampling_event(event) && !event->attr.freq && x86_pmu.limit_period) {
>  		s64 left = event->attr.sample_period;
>  		x86_pmu.limit_period(event, &left);
>  		if (left > event->attr.sample_period)

Re: [PATCH] perf/x86: Fix open counting event error

[Ravi Bangoria]

On 23-Apr-25 12:17 PM, Luo Gengkun wrote:
> Perf doesn't work at perf stat for hardware events:
> 
>  $perf stat -- sleep 1
>  Performance counter stats for 'sleep 1':
>              16.44 msec task-clock                       #    0.016 CPUs utilized
>                  2      context-switches                 #  121.691 /sec
>                  0      cpu-migrations                   #    0.000 /sec
>                 54      page-faults                      #    3.286 K/sec
>    <not supported>	cycles
>    <not supported>	instructions
>    <not supported>	branches
>    <not supported>	branch-misses

Wondering if it is worth to add this in perf test. Something like
below?

--- a/tools/perf/tests/shell/stat.sh
+++ b/tools/perf/tests/shell/stat.sh
@@ -16,6 +16,24 @@ test_default_stat() {
   echo "Basic stat command test [Success]"
 }
 
+test_stat_count() {
+  echo "stat count test"
+
+  if ! perf list | grep -q "cpu-cycles OR cycles"
+  then
+    echo "stat count test [Skipped cpu-cycles event missing]"
+    return
+  fi
+
+  if perf stat -e cycles true 2>&1 | grep -E -q "<not supported>"
+  then
+    echo "stat count test [Failed]"
+    err=1
+    return
+  fi
+  echo "stat count test [Success]"
+}
+
 test_stat_record_report() {
   echo "stat record and report test"
   if ! perf stat record -o - true | perf stat report -i - 2>&1 | \
@@ -201,6 +219,7 @@ test_hybrid() {
 }
 
 test_default_stat
+test_stat_count
 test_stat_record_report
 test_stat_record_script
 test_stat_repeat_weak_groups
---

Thanks,
Ravi

Re: [PATCH] perf/x86: Fix open counting event error

[Ingo Molnar]

* Luo Gengkun <luogengkun@huaweicloud.com> wrote:

> Perf doesn't work at perf stat for hardware events:
> 
>  $perf stat -- sleep 1
>  Performance counter stats for 'sleep 1':
>              16.44 msec task-clock                       #    0.016 CPUs utilized
>                  2      context-switches                 #  121.691 /sec
>                  0      cpu-migrations                   #    0.000 /sec
>                 54      page-faults                      #    3.286 K/sec
>    <not supported>	cycles
>    <not supported>	instructions
>    <not supported>	branches
>    <not supported>	branch-misses
> 
> The reason is that the check in x86_pmu_hw_config for sampling event is
> unexpectedly applied to the counting event.
> 
> Fixes: 88ec7eedbbd2 ("perf/x86: Fix low freqency setting issue")
> Signed-off-by: Luo Gengkun <luogengkun@huaweicloud.com>
> ---
>  arch/x86/events/core.c | 2 +-
>  1 file changed, 1 insertion(+), 1 deletion(-)
> 
> diff --git a/arch/x86/events/core.c b/arch/x86/events/core.c
> index 6866cc5acb0b..3a4f031d2f44 100644
> --- a/arch/x86/events/core.c
> +++ b/arch/x86/events/core.c
> @@ -629,7 +629,7 @@ int x86_pmu_hw_config(struct perf_event *event)
>  	if (event->attr.type == event->pmu->type)
>  		event->hw.config |= x86_pmu_get_event_config(event);
>  
> -	if (!event->attr.freq && x86_pmu.limit_period) {
> +	if (is_sampling_event(event) && !event->attr.freq && x86_pmu.limit_period) {

Hm, so how come it works here, on an affected x86 system:

$ perf stat -- sleep 1

 Performance counter stats for 'sleep 1':

              0.64 msec task-clock:u                     #    0.001 CPUs utilized             
                 0      context-switches:u               #    0.000 /sec                      
                 0      cpu-migrations:u                 #    0.000 /sec                      
                73      page-faults:u                    #  114.063 K/sec                     
           325,849      instructions:u                   #    0.56  insn per cycle            
                                                  #    0.88  stalled cycles per insn   
           580,323      cycles:u                         #    0.907 GHz                       
           286,348      stalled-cycles-frontend:u        #   49.34% frontend cycles idle      
            72,623      branches:u                       #  113.474 M/sec                     
             4,713      branch-misses:u                  #    6.49% of all branches           


?

Thanks,

	Ingo

Re: [PATCH] perf/x86: Fix open counting event error

[Liang, Kan]

On 2025-04-24 12:20 p.m., Ingo Molnar wrote:
> 
> * Luo Gengkun <luogengkun@huaweicloud.com> wrote:
> 
>> Perf doesn't work at perf stat for hardware events:
>>
>>  $perf stat -- sleep 1
>>  Performance counter stats for 'sleep 1':
>>              16.44 msec task-clock                       #    0.016 CPUs utilized
>>                  2      context-switches                 #  121.691 /sec
>>                  0      cpu-migrations                   #    0.000 /sec
>>                 54      page-faults                      #    3.286 K/sec
>>    <not supported>	cycles
>>    <not supported>	instructions
>>    <not supported>	branches
>>    <not supported>	branch-misses
>>
>> The reason is that the check in x86_pmu_hw_config for sampling event is
>> unexpectedly applied to the counting event.
>>
>> Fixes: 88ec7eedbbd2 ("perf/x86: Fix low freqency setting issue")
>> Signed-off-by: Luo Gengkun <luogengkun@huaweicloud.com>
>> ---
>>  arch/x86/events/core.c | 2 +-
>>  1 file changed, 1 insertion(+), 1 deletion(-)
>>
>> diff --git a/arch/x86/events/core.c b/arch/x86/events/core.c
>> index 6866cc5acb0b..3a4f031d2f44 100644
>> --- a/arch/x86/events/core.c
>> +++ b/arch/x86/events/core.c
>> @@ -629,7 +629,7 @@ int x86_pmu_hw_config(struct perf_event *event)
>>  	if (event->attr.type == event->pmu->type)
>>  		event->hw.config |= x86_pmu_get_event_config(event);
>>  
>> -	if (!event->attr.freq && x86_pmu.limit_period) {
>> +	if (is_sampling_event(event) && !event->attr.freq && x86_pmu.limit_period) {
> 
> Hm, so how come it works here, on an affected x86 system:
> 
> $ perf stat -- sleep 1
> 
>  Performance counter stats for 'sleep 1':
> 
>               0.64 msec task-clock:u                     #    0.001 CPUs utilized             
>                  0      context-switches:u               #    0.000 /sec                      
>                  0      cpu-migrations:u                 #    0.000 /sec                      
>                 73      page-faults:u                    #  114.063 K/sec                     
>            325,849      instructions:u                   #    0.56  insn per cycle            
>                                                   #    0.88  stalled cycles per insn   
>            580,323      cycles:u                         #    0.907 GHz                       
>            286,348      stalled-cycles-frontend:u        #   49.34% frontend cycles idle      
>             72,623      branches:u                       #  113.474 M/sec                     
>              4,713      branch-misses:u                  #    6.49% of all branches           
> 
> 
> ?

It doesn't affect all X86 platforms. It should only impact the platforms
with limit_period used for the non-pebs events. For Intel platforms, it
should only impact some older platforms, e.g., HSW, BDW and NHM.

For other platforms, the x86_pmu.limit_period is invoked. But the left
is not updated. So it still equals to event->attr.sample_period.
It doesn't error out.

	if (!event->attr.freq && x86_pmu.limit_period) {
		s64 left = event->attr.sample_period;
		x86_pmu.limit_period(event, &left);
		if (left > event->attr.sample_period)
			return -EINVAL;
	}

Thanks,
Kan

Re: [PATCH] perf/x86: Fix open counting event error

[Liang, Kan]

On 2025-04-24 2:52 a.m., Ravi Bangoria wrote:
> On 23-Apr-25 12:17 PM, Luo Gengkun wrote:
>> Perf doesn't work at perf stat for hardware events:
>>
>>  $perf stat -- sleep 1
>>  Performance counter stats for 'sleep 1':
>>              16.44 msec task-clock                       #    0.016 CPUs utilized
>>                  2      context-switches                 #  121.691 /sec
>>                  0      cpu-migrations                   #    0.000 /sec
>>                 54      page-faults                      #    3.286 K/sec
>>    <not supported>	cycles
>>    <not supported>	instructions
>>    <not supported>	branches
>>    <not supported>	branch-misses
> 
> Wondering if it is worth to add this in perf test. Something like
> below?
> 
> --- a/tools/perf/tests/shell/stat.sh
> +++ b/tools/perf/tests/shell/stat.sh
> @@ -16,6 +16,24 @@ test_default_stat() {
>    echo "Basic stat command test [Success]"
>  }
>  
> +test_stat_count() {
> +  echo "stat count test"
> +
> +  if ! perf list | grep -q "cpu-cycles OR cycles"
> +  then
> +    echo "stat count test [Skipped cpu-cycles event missing]"
> +    return
> +  fi
> +
> +  if perf stat -e cycles true 2>&1 | grep -E -q "<not supported>"
> +  then
> +    echo "stat count test [Failed]"
> +    err=1
> +    return
> +  fi
> +  echo "stat count test [Success]"
> +}
> +
>  test_stat_record_report() {
>    echo "stat record and report test"
>    if ! perf stat record -o - true | perf stat report -i - 2>&1 | \
> @@ -201,6 +219,7 @@ test_hybrid() {
>  }
>  
>  test_default_stat
> +test_stat_count

I think the perf stat default should always be supported, not just cycles.
Maybe we should add the check in test_default_stat?

Thanks,
Kan>  test_stat_record_report
>  test_stat_record_script
>  test_stat_repeat_weak_groups
> ---
> 
> Thanks,
> Ravi
> 

Re: [PATCH] perf/x86: Fix open counting event error

[Ingo Molnar]

* Liang, Kan <kan.liang@linux.intel.com> wrote:

> 
> 
> On 2025-04-24 12:20 p.m., Ingo Molnar wrote:
> > 
> > * Luo Gengkun <luogengkun@huaweicloud.com> wrote:
> > 
> >> Perf doesn't work at perf stat for hardware events:
> >>
> >>  $perf stat -- sleep 1
> >>  Performance counter stats for 'sleep 1':
> >>              16.44 msec task-clock                       #    0.016 CPUs utilized
> >>                  2      context-switches                 #  121.691 /sec
> >>                  0      cpu-migrations                   #    0.000 /sec
> >>                 54      page-faults                      #    3.286 K/sec
> >>    <not supported>	cycles
> >>    <not supported>	instructions
> >>    <not supported>	branches
> >>    <not supported>	branch-misses
> >>
> >> The reason is that the check in x86_pmu_hw_config for sampling event is
> >> unexpectedly applied to the counting event.
> >>
> >> Fixes: 88ec7eedbbd2 ("perf/x86: Fix low freqency setting issue")
> >> Signed-off-by: Luo Gengkun <luogengkun@huaweicloud.com>
> >> ---
> >>  arch/x86/events/core.c | 2 +-
> >>  1 file changed, 1 insertion(+), 1 deletion(-)
> >>
> >> diff --git a/arch/x86/events/core.c b/arch/x86/events/core.c
> >> index 6866cc5acb0b..3a4f031d2f44 100644
> >> --- a/arch/x86/events/core.c
> >> +++ b/arch/x86/events/core.c
> >> @@ -629,7 +629,7 @@ int x86_pmu_hw_config(struct perf_event *event)
> >>  	if (event->attr.type == event->pmu->type)
> >>  		event->hw.config |= x86_pmu_get_event_config(event);
> >>  
> >> -	if (!event->attr.freq && x86_pmu.limit_period) {
> >> +	if (is_sampling_event(event) && !event->attr.freq && x86_pmu.limit_period) {
> > 
> > Hm, so how come it works here, on an affected x86 system:
> > 
> > $ perf stat -- sleep 1
> > 
> >  Performance counter stats for 'sleep 1':
> > 
> >               0.64 msec task-clock:u                     #    0.001 CPUs utilized             
> >                  0      context-switches:u               #    0.000 /sec                      
> >                  0      cpu-migrations:u                 #    0.000 /sec                      
> >                 73      page-faults:u                    #  114.063 K/sec                     
> >            325,849      instructions:u                   #    0.56  insn per cycle            
> >                                                   #    0.88  stalled cycles per insn   
> >            580,323      cycles:u                         #    0.907 GHz                       
> >            286,348      stalled-cycles-frontend:u        #   49.34% frontend cycles idle      
> >             72,623      branches:u                       #  113.474 M/sec                     
> >              4,713      branch-misses:u                  #    6.49% of all branches           
> > 
> > 
> > ?
> 
> It doesn't affect all X86 platforms. It should only impact the platforms
> with limit_period used for the non-pebs events. For Intel platforms, it
> should only impact some older platforms, e.g., HSW, BDW and NHM.
> 
> For other platforms, the x86_pmu.limit_period is invoked. But the left
> is not updated. So it still equals to event->attr.sample_period.
> It doesn't error out.
> 
> 	if (!event->attr.freq && x86_pmu.limit_period) {
> 		s64 left = event->attr.sample_period;
> 		x86_pmu.limit_period(event, &left);
> 		if (left > event->attr.sample_period)
> 			return -EINVAL;
> 	}

Makes sense. I've added this paragraph to the changelog:

    It should only impact x86 platforms with limit_period used for non-PEBS
    events. For Intel platforms, it should only impact some older platforms,
    e.g., HSW, BDW and NHM.

Thanks,

	Ingo

Re: [PATCH] perf/x86: Fix open counting event error

[Ravi Bangoria]

Hi Kan,

>>> Perf doesn't work at perf stat for hardware events:
>>>
>>>  $perf stat -- sleep 1
>>>  Performance counter stats for 'sleep 1':
>>>              16.44 msec task-clock                       #    0.016 CPUs utilized
>>>                  2      context-switches                 #  121.691 /sec
>>>                  0      cpu-migrations                   #    0.000 /sec
>>>                 54      page-faults                      #    3.286 K/sec
>>>    <not supported>	cycles
>>>    <not supported>	instructions
>>>    <not supported>	branches
>>>    <not supported>	branch-misses
>>
>> Wondering if it is worth to add this in perf test. Something like
>> below?
>>
>> --- a/tools/perf/tests/shell/stat.sh
>> +++ b/tools/perf/tests/shell/stat.sh
>> @@ -16,6 +16,24 @@ test_default_stat() {
>>    echo "Basic stat command test [Success]"
>>  }
>>  
>> +test_stat_count() {
>> +  echo "stat count test"
>> +
>> +  if ! perf list | grep -q "cpu-cycles OR cycles"
>> +  then
>> +    echo "stat count test [Skipped cpu-cycles event missing]"
>> +    return
>> +  fi
>> +
>> +  if perf stat -e cycles true 2>&1 | grep -E -q "<not supported>"
>> +  then
>> +    echo "stat count test [Failed]"
>> +    err=1
>> +    return
>> +  fi
>> +  echo "stat count test [Success]"
>> +}
>> +
>>  test_stat_record_report() {
>>    echo "stat record and report test"
>>    if ! perf stat record -o - true | perf stat report -i - 2>&1 | \
>> @@ -201,6 +219,7 @@ test_hybrid() {
>>  }
>>  
>>  test_default_stat
>> +test_stat_count
> 
> I think the perf stat default should always be supported, not just cycles.
> Maybe we should add the check in test_default_stat?

Do you mean:

  if perf stat true 2>&1 | grep -E -q "<not supported>"
    err=1

Isn't this ambiguous? Also, this fails on machines where HW pmu
is not supported. For ex, on my qemu guest with `-cpu pmu=off`:

  $ ./perf list | grep "cpu-cycles OR cycles"
  <empty output>

  $ ./perf stat true
   Performance counter stats for 'true':
                0.42 msec task-clock:u                     #    0.470 CPUs utilized
                   0      context-switches:u               #    0.000 /sec
                   0      cpu-migrations:u                 #    0.000 /sec
                  48      page-faults:u                    #  113.874 K/sec
     <not supported>      cycles:u

Thanks,
Ravi

Re: [PATCH] perf/x86: Fix open counting event error

[Liang, Kan]

Hi Ravi,

Sorry for the late response. I was on vacation.

On 2025-04-25 6:12 a.m., Ravi Bangoria wrote:
> Hi Kan,
> 
>>>> Perf doesn't work at perf stat for hardware events:
>>>>
>>>>  $perf stat -- sleep 1
>>>>  Performance counter stats for 'sleep 1':
>>>>              16.44 msec task-clock                       #    0.016 CPUs utilized
>>>>                  2      context-switches                 #  121.691 /sec
>>>>                  0      cpu-migrations                   #    0.000 /sec
>>>>                 54      page-faults                      #    3.286 K/sec
>>>>    <not supported>	cycles
>>>>    <not supported>	instructions
>>>>    <not supported>	branches
>>>>    <not supported>	branch-misses
>>>
>>> Wondering if it is worth to add this in perf test. Something like
>>> below?
>>>
>>> --- a/tools/perf/tests/shell/stat.sh
>>> +++ b/tools/perf/tests/shell/stat.sh
>>> @@ -16,6 +16,24 @@ test_default_stat() {
>>>    echo "Basic stat command test [Success]"
>>>  }
>>>  
>>> +test_stat_count() {
>>> +  echo "stat count test"
>>> +
>>> +  if ! perf list | grep -q "cpu-cycles OR cycles"
>>> +  then
>>> +    echo "stat count test [Skipped cpu-cycles event missing]"
>>> +    return
>>> +  fi
>>> +
>>> +  if perf stat -e cycles true 2>&1 | grep -E -q "<not supported>"
>>> +  then
>>> +    echo "stat count test [Failed]"
>>> +    err=1
>>> +    return
>>> +  fi
>>> +  echo "stat count test [Success]"
>>> +}
>>> +
>>>  test_stat_record_report() {
>>>    echo "stat record and report test"
>>>    if ! perf stat record -o - true | perf stat report -i - 2>&1 | \
>>> @@ -201,6 +219,7 @@ test_hybrid() {
>>>  }
>>>  
>>>  test_default_stat
>>> +test_stat_count
>>
>> I think the perf stat default should always be supported, not just cycles.
>> Maybe we should add the check in test_default_stat?
> 
> Do you mean:
> 
>   if perf stat true 2>&1 | grep -E -q "<not supported>"
>     err=1
>

Yes, I assumed that all the events in the perf stat are always
available. But it seems the assumption is only true for bare metal.


> Isn't this ambiguous? Also, this fails on machines where HW pmu
> is not supported. For ex, on my qemu guest with `-cpu pmu=off`:
> 
>   $ ./perf list | grep "cpu-cycles OR cycles"
>   <empty output>
> 
>   $ ./perf stat true
>    Performance counter stats for 'true':
>                 0.42 msec task-clock:u                     #    0.470 CPUs utilized
>                    0      context-switches:u               #    0.000 /sec
>                    0      cpu-migrations:u                 #    0.000 /sec
>                   48      page-faults:u                    #  113.874 K/sec
>      <not supported>      cycles:u
> 
If taking the virtualization into account, the test_stat_count looks
good to me.

Thanks,
Kan