[v3,0/6] perf arm64 metricgroup support

Message

John Garry April 7, 2021, 10:32 a.m. UTC

This series contains support to get basic metricgroups working for
arm64 CPUs.

Initial support is added for HiSilicon hip08 platform.

Some sample usage on Huawei D06 board:

 $ ./perf list metric    

List of pre-defined events (to be used in -e): 

Metrics:     

  bp_misp_flush
       [BP misp flush L3 topdown metric]
  branch_mispredicts
       [Branch mispredicts L2 topdown metric]
  core_bound
       [Core bound L2 topdown metric]
  divider
       [Divider L3 topdown metric]
  exe_ports_util
       [EXE ports util L3 topdown metric]
  fetch_bandwidth_bound
       [Fetch bandwidth bound L2 topdown metric]
  fetch_latency_bound
       [Fetch latency bound L2 topdown metric]
  fsu_stall
       [FSU stall L3 topdown metric]
  idle_by_icache_miss

$ sudo ./perf stat -v -M core_bound sleep 1
Using CPUID 0x00000000480fd010
metric expr (exe_stall_cycle - (mem_stall_anyload + armv8_pmuv3_0@event\=0x7005@)) / cpu_cycles for core_bound
found event cpu_cycles
found event armv8_pmuv3_0/event=0x7005/
found event exe_stall_cycle
found event mem_stall_anyload
adding {cpu_cycles -> armv8_pmuv3_0/event=0x7001/
mem_stall_anyload -> armv8_pmuv3_0/event=0x7004/
Control descriptor is not initialized
cpu_cycles: 989433 385050 385050
armv8_pmuv3_0/event=0x7005/: 19207 385050 385050
exe_stall_cycle: 900825 385050 385050
mem_stall_anyload: 253516 385050 385050

Performance counter stats for 'sleep':

989,433      cpu_cycles      #     0.63 core_bound
  19,207      armv8_pmuv3_0/event=0x7005/
 900,825      exe_stall_cycle
 253,516      mem_stall_anyload

       0.000805809 seconds time elapsed

       0.000875000 seconds user
       0.000000000 seconds sys
       
perf stat --topdown is not supported, as this requires the CPU PMU to
expose (alias) events for the TopDown L1 metrics from sysfs, which arm 
does not do. To get that to work, we probably need to make perf use the
pmu-events cpumap to learn about those alias events.

Metric reuse support is added for pmu-events parse metric testcase.
This had been broken on power9 recently:
https://lore.kernel.org/lkml/[e-mail]/ 

Differences to v2:
- Add TB and RB tags (Thanks!)
- Rename metricgroup__find_metric() from metricgroup_find_metric()
- Change resolve_metric_simple() to rescan after any insert

Differences to v1:
- Add pmu_events_map__find() as arm64-specific function
- Fix metric reuse for pmu-events parse metric testcase 

John Garry (6):
  perf metricgroup: Make find_metric() public with name change
  perf test: Handle metric reuse in pmu-events parsing test
  perf pmu: Add pmu_events_map__find()
  perf vendor events arm64: Add Hisi hip08 L1 metrics
  perf vendor events arm64: Add Hisi hip08 L2 metrics
  perf vendor events arm64: Add Hisi hip08 L3 metrics

 tools/perf/arch/arm64/util/Build              |   1 +
 tools/perf/arch/arm64/util/pmu.c              |  25 ++
 .../arch/arm64/hisilicon/hip08/metrics.json   | 233 ++++++++++++++++++
 tools/perf/tests/pmu-events.c                 |  83 ++++++-
 tools/perf/util/metricgroup.c                 |  12 +-
 tools/perf/util/metricgroup.h                 |   3 +-
 tools/perf/util/pmu.c                         |   5 +
 tools/perf/util/pmu.h                         |   1 +
 tools/perf/util/s390-sample-raw.c             |   4 +-
 9 files changed, 356 insertions(+), 11 deletions(-)
 create mode 100644 tools/perf/arch/arm64/util/pmu.c
 create mode 100644 tools/perf/pmu-events/arch/arm64/hisilicon/hip08/metrics.json

Comments

Arnaldo Carvalho de Melo April 7, 2021, 1:11 p.m. UTC | #1

Em Wed, Apr 07, 2021 at 06:32:44PM +0800, John Garry escreveu:
> This series contains support to get basic metricgroups working for
> arm64 CPUs.
> 
> Initial support is added for HiSilicon hip08 platform.
> 
> Some sample usage on Huawei D06 board:
> 
>  $ ./perf list metric    

Thanks, applied.

- Arnaldo

 
> List of pre-defined events (to be used in -e): 
> 
> Metrics:     
> 
>   bp_misp_flush
>        [BP misp flush L3 topdown metric]
>   branch_mispredicts
>        [Branch mispredicts L2 topdown metric]
>   core_bound
>        [Core bound L2 topdown metric]
>   divider
>        [Divider L3 topdown metric]
>   exe_ports_util
>        [EXE ports util L3 topdown metric]
>   fetch_bandwidth_bound
>        [Fetch bandwidth bound L2 topdown metric]
>   fetch_latency_bound
>        [Fetch latency bound L2 topdown metric]
>   fsu_stall
>        [FSU stall L3 topdown metric]
>   idle_by_icache_miss
> 
> $ sudo ./perf stat -v -M core_bound sleep 1
> Using CPUID 0x00000000480fd010
> metric expr (exe_stall_cycle - (mem_stall_anyload + armv8_pmuv3_0@event\=0x7005@)) / cpu_cycles for core_bound
> found event cpu_cycles
> found event armv8_pmuv3_0/event=0x7005/
> found event exe_stall_cycle
> found event mem_stall_anyload
> adding {cpu_cycles -> armv8_pmuv3_0/event=0x7001/
> mem_stall_anyload -> armv8_pmuv3_0/event=0x7004/
> Control descriptor is not initialized
> cpu_cycles: 989433 385050 385050
> armv8_pmuv3_0/event=0x7005/: 19207 385050 385050
> exe_stall_cycle: 900825 385050 385050
> mem_stall_anyload: 253516 385050 385050
> 
> Performance counter stats for 'sleep':
> 
> 989,433      cpu_cycles      #     0.63 core_bound
>   19,207      armv8_pmuv3_0/event=0x7005/
>  900,825      exe_stall_cycle
>  253,516      mem_stall_anyload
> 
>        0.000805809 seconds time elapsed
> 
>        0.000875000 seconds user
>        0.000000000 seconds sys
>        
> perf stat --topdown is not supported, as this requires the CPU PMU to
> expose (alias) events for the TopDown L1 metrics from sysfs, which arm 
> does not do. To get that to work, we probably need to make perf use the
> pmu-events cpumap to learn about those alias events.
> 
> Metric reuse support is added for pmu-events parse metric testcase.
> This had been broken on power9 recently:
> https://lore.kernel.org/lkml/[e-mail]/ 
> 
> Differences to v2:
> - Add TB and RB tags (Thanks!)
> - Rename metricgroup__find_metric() from metricgroup_find_metric()
> - Change resolve_metric_simple() to rescan after any insert
> 
> Differences to v1:
> - Add pmu_events_map__find() as arm64-specific function
> - Fix metric reuse for pmu-events parse metric testcase 
> 
> John Garry (6):
>   perf metricgroup: Make find_metric() public with name change
>   perf test: Handle metric reuse in pmu-events parsing test
>   perf pmu: Add pmu_events_map__find()
>   perf vendor events arm64: Add Hisi hip08 L1 metrics
>   perf vendor events arm64: Add Hisi hip08 L2 metrics
>   perf vendor events arm64: Add Hisi hip08 L3 metrics
> 
>  tools/perf/arch/arm64/util/Build              |   1 +
>  tools/perf/arch/arm64/util/pmu.c              |  25 ++
>  .../arch/arm64/hisilicon/hip08/metrics.json   | 233 ++++++++++++++++++
>  tools/perf/tests/pmu-events.c                 |  83 ++++++-
>  tools/perf/util/metricgroup.c                 |  12 +-
>  tools/perf/util/metricgroup.h                 |   3 +-
>  tools/perf/util/pmu.c                         |   5 +
>  tools/perf/util/pmu.h                         |   1 +
>  tools/perf/util/s390-sample-raw.c             |   4 +-
>  9 files changed, 356 insertions(+), 11 deletions(-)
>  create mode 100644 tools/perf/arch/arm64/util/pmu.c
>  create mode 100644 tools/perf/pmu-events/arch/arm64/hisilicon/hip08/metrics.json
> 
> -- 
> 2.26.2
>

Jiri Olsa April 8, 2021, 12:06 p.m. UTC | #2

On Wed, Apr 07, 2021 at 06:32:44PM +0800, John Garry wrote:
> This series contains support to get basic metricgroups working for
> arm64 CPUs.
> 
> Initial support is added for HiSilicon hip08 platform.
> 
> Some sample usage on Huawei D06 board:
> 
>  $ ./perf list metric    
> 
> List of pre-defined events (to be used in -e): 
> 
> Metrics:     
> 
>   bp_misp_flush
>        [BP misp flush L3 topdown metric]
>   branch_mispredicts
>        [Branch mispredicts L2 topdown metric]
>   core_bound
>        [Core bound L2 topdown metric]
>   divider
>        [Divider L3 topdown metric]
>   exe_ports_util
>        [EXE ports util L3 topdown metric]
>   fetch_bandwidth_bound
>        [Fetch bandwidth bound L2 topdown metric]
>   fetch_latency_bound
>        [Fetch latency bound L2 topdown metric]
>   fsu_stall
>        [FSU stall L3 topdown metric]
>   idle_by_icache_miss
> 
> $ sudo ./perf stat -v -M core_bound sleep 1
> Using CPUID 0x00000000480fd010
> metric expr (exe_stall_cycle - (mem_stall_anyload + armv8_pmuv3_0@event\=0x7005@)) / cpu_cycles for core_bound
> found event cpu_cycles
> found event armv8_pmuv3_0/event=0x7005/
> found event exe_stall_cycle
> found event mem_stall_anyload
> adding {cpu_cycles -> armv8_pmuv3_0/event=0x7001/
> mem_stall_anyload -> armv8_pmuv3_0/event=0x7004/
> Control descriptor is not initialized
> cpu_cycles: 989433 385050 385050
> armv8_pmuv3_0/event=0x7005/: 19207 385050 385050
> exe_stall_cycle: 900825 385050 385050
> mem_stall_anyload: 253516 385050 385050
> 
> Performance counter stats for 'sleep':
> 
> 989,433      cpu_cycles      #     0.63 core_bound
>   19,207      armv8_pmuv3_0/event=0x7005/
>  900,825      exe_stall_cycle
>  253,516      mem_stall_anyload
> 
>        0.000805809 seconds time elapsed
> 
>        0.000875000 seconds user
>        0.000000000 seconds sys
>        
> perf stat --topdown is not supported, as this requires the CPU PMU to
> expose (alias) events for the TopDown L1 metrics from sysfs, which arm 
> does not do. To get that to work, we probably need to make perf use the
> pmu-events cpumap to learn about those alias events.
> 
> Metric reuse support is added for pmu-events parse metric testcase.
> This had been broken on power9 recently:
> https://lore.kernel.org/lkml/[e-mail]/ 
> 
> Differences to v2:
> - Add TB and RB tags (Thanks!)
> - Rename metricgroup__find_metric() from metricgroup_find_metric()
> - Change resolve_metric_simple() to rescan after any insert

Acked-by: Jiri Olsa <[e-mail]>

thanks,
jirka