Skip to contentSkip to conversation

-rt kernel just can show 8 perf events?

DISCUSSION Shi, Yang 2 participants 5 messages Dec 1, 2015 – Feb 4, 2016 perf events

lore
Share & tools
lore

Shi, Yang

author
#0 lore raw

Hi folks,

I'm running some perf program on my x86-64 machine with 4.1.12-rt13 kernel, but I found just 8 perf events can be showed up with "perf stat -e" on -rt kernel, but mainline kernel is fine.

I traced 14 events (tried different number in my test, always > 8), then I got:

29692 cs

104 page-faults

104 minor-faults

0 major-faults

100122071 iTLB-load-misses

447793 dTLB-load-misses

13655 dTLB-store-misses

201620583 branch-misses

<not counted> cache-misses (0.00%)
<not counted> cycles (0.00%)
<not counted> dtlb_misses_walk_duration (0.00%)
<not counted> dtlb_misses_walk_completed (0.00%)
<not counted> itlb_misses_walk_duration (0.00%)
<not counted> itlb_misses_walk_completed (0.00%)

No matter how many events are traced, just 8 events are showed with meaningful data, all others are showed up as "not counted".

Any hint is appreciated.

Regards,
Yang

#1 lore raw
Quoted 5 lines · * Shi, Yang | 2015-12-01 12:36:37 [-0800]:

Hi folks,

I'm running some perf program on my x86-64 machine with 4.1.12-rt13 kernel, but I found just 8 perf events can be showed up with "perf stat -e" on -rt kernel, but mainline kernel is fine.

And 4.1.12 (without -rt13) is not having this kind of problem?

Quoted 2 lines

Regards,
Yang

Sebastian

author
#2 lore raw
Quoted 9 lines · Sebastian Andrzej Siewior
Quoted 5 lines · * Shi, Yang | 2015-12-01 12:36:37 [-0800]:

Hi folks,

I'm running some perf program on my x86-64 machine with 4.1.12-rt13 kernel, but I found just 8 perf events can be showed up with "perf stat -e" on -rt kernel, but mainline kernel is fine.

And 4.1.12 (without -rt13) is not having this kind of problem?

No. mainline 4.1.12 works well.

Yang

Quoted 6 lines
Quoted 3 lines

Regards,
Yang

Sebastian

#3 lore raw
Quoted 4 lines · * Shi, Yang | 2015-12-11 09:57:06 [-0800]:
Quoted 1 line · Sebastian Andrzej Siewior

And 4.1.12 (without -rt13) is not having this kind of problem?

No. mainline 4.1.12 works well.

The same thing here. But I get this:

Performance counter stats for 'tar tJ':

           144,407      cs                                                          
            16,733      page-faults                                                 
            16,732      minor-faults                                                
                 1      major-faults                                                
           805,335      iTLB-load-misses                                              (24.83%)
         6,606,698      dTLB-load-misses                                              (24.63%)
     3,734,425,072      branches                                                      (37.35%)
       228,660,856      L1-dcache-prefetches                                          (37.06%)
         4,564,273      L1-icache-prefetches                                          (49.68%)
       399,365,890      branch-misses             #   10.69% of all branches          (62.63%)
        79,278,129      cache-misses                                                  (62.96%)
    30,825,564,763      cycles                                                        (50.79%)
       8.705696429 seconds time elapsed

once I apply this chunk:

kernel/events/core.c +1 -0
@@ -802,6 +802,7 @@ static void __perf_mux_hrtimer_init(struct perf_cpu_context *cpuctx, int cpu)
 	raw_spin_lock_init(&cpuctx->hrtimer_lock);
 	hrtimer_init(timer, CLOCK_MONOTONIC, HRTIMER_MODE_ABS_PINNED);
 	timer->function = perf_mux_hrtimer_handler;
+	timer->irqsafe = 1;
 }
 
 static int perf_mux_hrtimer_restart(struct perf_cpu_context *cpuctx)
Quoted 1 line

Yang

Sebastian

author
#4 lore raw
Quoted 24 lines · Sebastian Andrzej Siewior
Quoted 4 lines · * Shi, Yang | 2015-12-11 09:57:06 [-0800]:
Quoted 1 line · Sebastian Andrzej Siewior

And 4.1.12 (without -rt13) is not having this kind of problem?

No. mainline 4.1.12 works well.

The same thing here. But I get this:

Performance counter stats for 'tar tJ':

            144,407      cs
             16,733      page-faults
             16,732      minor-faults
                  1      major-faults
            805,335      iTLB-load-misses                                              (24.83%)
          6,606,698      dTLB-load-misses                                              (24.63%)
      3,734,425,072      branches                                                      (37.35%)
        228,660,856      L1-dcache-prefetches                                          (37.06%)
          4,564,273      L1-icache-prefetches                                          (49.68%)
        399,365,890      branch-misses             #   10.69% of all branches          (62.63%)
         79,278,129      cache-misses                                                  (62.96%)
     30,825,564,763      cycles                                                        (50.79%)
Quoted 3 lines
        8.705696429 seconds time elapsed

once I apply this chunk:

Thanks. I will have a try with the patch soon.

Yang

Quoted 17 lines
kernel/events/core.c +1 -0
@@ -802,6 +802,7 @@ static void __perf_mux_hrtimer_init(struct perf_cpu_context *cpuctx, int cpu)
  	raw_spin_lock_init(&cpuctx->hrtimer_lock);
  	hrtimer_init(timer, CLOCK_MONOTONIC, HRTIMER_MODE_ABS_PINNED);
  	timer->function = perf_mux_hrtimer_handler;
+	timer->irqsafe = 1;
  }
 
  static int perf_mux_hrtimer_restart(struct perf_cpu_context *cpuctx)
Quoted 2 lines

Yang

Sebastian