]> git.ipfire.org Git - thirdparty/linux.git/log
thirdparty/linux.git
5 weeks agoperf build: Remove NO_AUXTRACE build option
Ian Rogers [Mon, 10 Nov 2025 01:31:51 +0000 (17:31 -0800)] 
perf build: Remove NO_AUXTRACE build option

The NO_AUXTRACE build option was used when the __get_cpuid feature
test failed or if it was provided on the command line. The option no
longer avoids a dependency on a library and so having the option is
just adding complexity to the code base. Remove the option
CONFIG_AUXTRACE from Build files and HAVE_AUXTRACE_SUPPORT by assuming
it is always defined.

Signed-off-by: Ian Rogers <irogers@google.com>
Reviewed-by: James Clark <james.clark@linaro.org>
Signed-off-by: Namhyung Kim <namhyung@kernel.org>
5 weeks agotool build: Remove __get_cpuid feature test
Ian Rogers [Mon, 10 Nov 2025 01:31:50 +0000 (17:31 -0800)] 
tool build: Remove __get_cpuid feature test

This feature test is no longer used so remove.

The function tested by the feature test is used in:
tools/power/x86/x86_energy_perf_policy/x86_energy_perf_policy.c
however, the Makefile just assumes the presence of the function and
doesn't perform a build feature test for it.

Signed-off-by: Ian Rogers <irogers@google.com>
Reviewed-by: James Clark <james.clark@linaro.org>
Signed-off-by: Namhyung Kim <namhyung@kernel.org>
5 weeks agoperf build: Don't add NO_AUXTRACE if missing feature-get_cpuid
Ian Rogers [Mon, 10 Nov 2025 01:31:49 +0000 (17:31 -0800)] 
perf build: Don't add NO_AUXTRACE if missing feature-get_cpuid

The intel-pt code dependent on __get_cpuid is no longer present so
remove the feature test in the Makefile.config.

Signed-off-by: Ian Rogers <irogers@google.com>
Reviewed-by: James Clark <james.clark@linaro.org>
Signed-off-by: Namhyung Kim <namhyung@kernel.org>
5 weeks agoperf intel-pt: Use the perf provided "cpuid.h"
Ian Rogers [Mon, 10 Nov 2025 01:31:48 +0000 (17:31 -0800)] 
perf intel-pt: Use the perf provided "cpuid.h"

Rather than having a feature test and include of <cpuid.h> for the
__get_cpuid function, use the cpuid function provided by
tools/perf/arch/x86/util/cpuid.h.

Signed-off-by: Ian Rogers <irogers@google.com>
Reviewed-by: James Clark <james.clark@linaro.org>
Signed-off-by: Namhyung Kim <namhyung@kernel.org>
5 weeks agoperf test: Add a perf event fallback test
Zide Chen [Wed, 12 Nov 2025 16:48:23 +0000 (08:48 -0800)] 
perf test: Add a perf event fallback test

This adds test cases to verify the precise ip fallback logic:

- If the system supports precise ip, for an event given with the maximum
  precision level, it should be able to decrease precise_ip to find a
  supported level.
- The same fallback behavior should also work in more complex scenarios,
  such as event groups or when PEBS is involved

Additional fallback tests, such as those covering missing feature cases,
can be added in the future.

Suggested-by: Ian Rogers <irogers@google.com>
Reviewed-by: Dapeng Mi <dapeng1.mi@linux.intel.com>
Signed-off-by: Zide Chen <zide.chen@intel.com>
Reviewed-by: Ian Rogers <irogers!@google.com>
Signed-off-by: Namhyung Kim <namhyung@kernel.org>
5 weeks agoperf stat: Align metric output without events
Namhyung Kim [Thu, 6 Nov 2025 07:28:34 +0000 (23:28 -0800)] 
perf stat: Align metric output without events

One of my concern in the perf stat output was the alignment in the
metrics and shadow stats.  I think it missed to calculate the basic
output length using COUNTS_LEN and EVNAME_LEN but missed to add the
unit length like "msec" and surround 2 spaces.  I'm not sure why it's
not printed below though.

But anyway, now it shows correctly aligned metric output.

  $ perf stat true

   Performance counter stats for 'true':

             859,772      task-clock                       #    0.395 CPUs utilized
                   0      context-switches                 #    0.000 /sec
                   0      cpu-migrations                   #    0.000 /sec
                  56      page-faults                      #   65.134 K/sec
           1,075,022      instructions                     #    0.86  insn per cycle
           1,255,911      cycles                           #    1.461 GHz
             220,573      branches                         #  256.548 M/sec
               7,381      branch-misses                    #    3.35% of all branches
                          TopdownL1                        #     19.2 %  tma_retiring
                                                           #     28.6 %  tma_backend_bound
                                                           #      9.5 %  tma_bad_speculation
                                                           #     42.6 %  tma_frontend_bound

         0.002174871 seconds time elapsed                  ^
                                                           |
         0.002154000 seconds user                          |
         0.000000000 seconds sys                          here

Reviewed-by: Ian Rogers <irogers@google.com>
Signed-off-by: Namhyung Kim <namhyung@kernel.org>
5 weeks agoperf tool_pmu: Make core_wide and target_cpu json events
Ian Rogers [Tue, 11 Nov 2025 21:22:06 +0000 (13:22 -0800)] 
perf tool_pmu: Make core_wide and target_cpu json events

For the sake of better documentation, add core_wide and target_cpu to
the tool.json. When the values of system_wide and
user_requested_cpu_list are unknown, use the values from the global
stat_config.

Example output showing how '-a' modifies the values in `perf stat`:
```
$ perf stat -e core_wide,target_cpu true

 Performance counter stats for 'true':

                 0      core_wide
                 0      target_cpu

       0.000993787 seconds time elapsed

       0.001128000 seconds user
       0.000000000 seconds sys

$ perf stat -e core_wide,target_cpu -a true

 Performance counter stats for 'system wide':

                 1      core_wide
                 1      target_cpu

       0.002271723 seconds time elapsed

$ perf list
...
tool:
  core_wide
       [1 if not SMT,if SMT are events being gathered on all SMT threads 1 otherwise 0. Unit: tool]
...
  target_cpu
       [1 if CPUs being analyzed,0 if threads/processes. Unit: tool]
...
```

Signed-off-by: Ian Rogers <irogers@google.com>
Signed-off-by: Namhyung Kim <namhyung@kernel.org>
5 weeks agoperf test stat csv: Update test expectations and events
Ian Rogers [Tue, 11 Nov 2025 21:22:05 +0000 (13:22 -0800)] 
perf test stat csv: Update test expectations and events

Explicitly use a metric rather than implicitly expecting '-e
instructions,cycles' to produce a metric. Use a metric with software
events to make it more compatible.

Signed-off-by: Ian Rogers <irogers@google.com>
Signed-off-by: Namhyung Kim <namhyung@kernel.org>
5 weeks agoperf test stat: Update test expectations and events
Ian Rogers [Tue, 11 Nov 2025 21:22:04 +0000 (13:22 -0800)] 
perf test stat: Update test expectations and events

test_stat_record_report and test_stat_record_script used default
output which triggers a bug when sending metrics. As this isn't
relevant to the test switch to using named software events.

Update the match in test_hybrid as the cycles event is now cpu-cycles
to workaround potential ARM issues.

Signed-off-by: Ian Rogers <irogers@google.com>
Signed-off-by: Namhyung Kim <namhyung@kernel.org>
5 weeks agoperf test stat: Update shadow test to use metrics
Ian Rogers [Tue, 11 Nov 2025 21:22:03 +0000 (13:22 -0800)] 
perf test stat: Update shadow test to use metrics

Previously '-e cycles,instructions' would implicitly create an IPC
metric. This now has to be explicit with '-M insn_per_cycle'.

Signed-off-by: Ian Rogers <irogers@google.com>
Signed-off-by: Namhyung Kim <namhyung@kernel.org>
5 weeks agoperf test metrics: Update all metrics for possibly failing default metrics
Ian Rogers [Tue, 11 Nov 2025 21:22:02 +0000 (13:22 -0800)] 
perf test metrics: Update all metrics for possibly failing default metrics

Default metrics may use unsupported events and be ignored. These
metrics shouldn't cause metric testing to fail.

Signed-off-by: Ian Rogers <irogers@google.com>
Signed-off-by: Namhyung Kim <namhyung@kernel.org>
5 weeks agoperf test stat: Update std_output testing metric expectations
Ian Rogers [Tue, 11 Nov 2025 21:22:01 +0000 (13:22 -0800)] 
perf test stat: Update std_output testing metric expectations

Make the expectations match json metrics rather than the previous hard
coded ones.

Signed-off-by: Ian Rogers <irogers@google.com>
Signed-off-by: Namhyung Kim <namhyung@kernel.org>
5 weeks agoperf test stat: Ignore failures in Default[234] metricgroups
Ian Rogers [Tue, 11 Nov 2025 21:22:00 +0000 (13:22 -0800)] 
perf test stat: Ignore failures in Default[234] metricgroups

The Default[234] metric groups may contain unsupported legacy
events. Allow those metric groups to fail.

Signed-off-by: Ian Rogers <irogers@google.com>
Signed-off-by: Namhyung Kim <namhyung@kernel.org>
5 weeks agoperf test stat+json: Improve metric-only testing
Ian Rogers [Tue, 11 Nov 2025 21:21:59 +0000 (13:21 -0800)] 
perf test stat+json: Improve metric-only testing

When testing metric-only, pass a metric to perf rather than expecting
a hard coded metric value to be generated.

Remove keys that were really metric-only units and instead don't
expect metric only to have a matching json key as it encodes metrics
as {"metric_name", "metric_value"}.

Signed-off-by: Ian Rogers <irogers@google.com>
Signed-off-by: Namhyung Kim <namhyung@kernel.org>
5 weeks agoperf stat: Remove "unit" workarounds for metric-only
Ian Rogers [Tue, 11 Nov 2025 21:21:58 +0000 (13:21 -0800)] 
perf stat: Remove "unit" workarounds for metric-only

Remove code that tested the "unit" as in KB/sec for certain hard coded
metric values and did workarounds.

Signed-off-by: Ian Rogers <irogers@google.com>
Signed-off-by: Namhyung Kim <namhyung@kernel.org>
5 weeks agoperf stat: Sort default events/metrics
Ian Rogers [Tue, 11 Nov 2025 21:21:57 +0000 (13:21 -0800)] 
perf stat: Sort default events/metrics

To improve the readability of default events/metrics, sort the evsels
after the Default metric groups have be parsed.

Before:
```
$ perf stat -a sleep 1
 Performance counter stats for 'system wide':

            22,087      context-switches                 #      nan cs/sec  cs_per_second
             TopdownL1 (cpu_core)                 #     10.3 %  tma_bad_speculation
                                                  #     25.8 %  tma_frontend_bound
                                                  #     34.5 %  tma_backend_bound
                                                  #     29.3 %  tma_retiring
             7,829      page-faults                      #      nan faults/sec  page_faults_per_second
       880,144,270      cpu_atom/cpu-cycles/             #      nan GHz  cycles_frequency       (50.10%)
     1,693,081,235      cpu_core/cpu-cycles/             #      nan GHz  cycles_frequency
             TopdownL1 (cpu_atom)                 #     20.5 %  tma_bad_speculation
                                                  #     13.8 %  tma_retiring             (50.26%)
                                                  #     34.6 %  tma_frontend_bound       (50.23%)
        89,326,916      cpu_atom/branches/               #      nan M/sec  branch_frequency     (60.19%)
       538,123,088      cpu_core/branches/               #      nan M/sec  branch_frequency
             1,368      cpu-migrations                   #      nan migrations/sec  migrations_per_second
                                                  #     31.1 %  tma_backend_bound        (60.19%)
              0.00 msec cpu-clock                        #      0.0 CPUs  CPUs_utilized
       485,744,856      cpu_atom/instructions/           #      0.6 instructions  insn_per_cycle  (59.87%)
     3,093,112,283      cpu_core/instructions/           #      1.8 instructions  insn_per_cycle
         4,939,427      cpu_atom/branch-misses/          #      5.0 %  branch_miss_rate         (49.77%)
         7,632,248      cpu_core/branch-misses/          #      1.4 %  branch_miss_rate

       1.005084693 seconds time elapsed
```
After:
```
$ perf stat -a sleep 1
 Performance counter stats for 'system wide':

            22,165      context-switches                 #      nan cs/sec  cs_per_second
              0.00 msec cpu-clock                        #      0.0 CPUs  CPUs_utilized
             2,260      cpu-migrations                   #      nan migrations/sec  migrations_per_second
            20,476      page-faults                      #      nan faults/sec  page_faults_per_second
        17,052,357      cpu_core/branch-misses/          #      1.5 %  branch_miss_rate
     1,120,090,590      cpu_core/branches/               #      nan M/sec  branch_frequency
     3,402,892,275      cpu_core/cpu-cycles/             #      nan GHz  cycles_frequency
     6,129,236,701      cpu_core/instructions/           #      1.8 instructions  insn_per_cycle
         6,159,523      cpu_atom/branch-misses/          #      3.1 %  branch_miss_rate         (49.86%)
       222,158,812      cpu_atom/branches/               #      nan M/sec  branch_frequency     (50.25%)
     1,547,610,244      cpu_atom/cpu-cycles/             #      nan GHz  cycles_frequency       (50.40%)
     1,304,901,260      cpu_atom/instructions/           #      0.8 instructions  insn_per_cycle  (50.41%)
             TopdownL1 (cpu_core)                 #     13.7 %  tma_bad_speculation
                                                  #     23.5 %  tma_frontend_bound
                                                  #     33.3 %  tma_backend_bound
                                                  #     29.6 %  tma_retiring
             TopdownL1 (cpu_atom)                 #     32.1 %  tma_backend_bound        (59.65%)
                                                  #     30.1 %  tma_frontend_bound       (59.51%)
                                                  #     22.3 %  tma_bad_speculation
                                                  #     15.5 %  tma_retiring             (59.53%)

       1.008405429 seconds time elapsed
```

Signed-off-by: Ian Rogers <irogers@google.com>
Signed-off-by: Namhyung Kim <namhyung@kernel.org>
5 weeks agoperf stat: Fix default metricgroup display on hybrid
Ian Rogers [Tue, 11 Nov 2025 21:21:56 +0000 (13:21 -0800)] 
perf stat: Fix default metricgroup display on hybrid

The logic to skip output of a default metric line was firing on
Alderlake and not displaying 'TopdownL1 (cpu_atom)'. Remove the
need_full_name check as it is equivalent to the different PMU test in
the cases we care about, merge the 'if's and flip the evsel of the PMU
test. The 'if' is now basically saying, if the output matches the last
printed output then skip the output.

Before:
```
             TopdownL1 (cpu_core)                 #     11.3 %  tma_bad_speculation
                                                  #     24.3 %  tma_frontend_bound
             TopdownL1 (cpu_core)                 #     33.9 %  tma_backend_bound
                                                  #     30.6 %  tma_retiring
                                                  #     42.2 %  tma_backend_bound
                                                  #     25.0 %  tma_frontend_bound       (49.81%)
                                                  #     12.8 %  tma_bad_speculation
                                                  #     20.0 %  tma_retiring             (59.46%)
```
After:
```
             TopdownL1 (cpu_core)                 #      8.3 %  tma_bad_speculation
                                                  #     43.7 %  tma_frontend_bound
                                                  #     30.7 %  tma_backend_bound
                                                  #     17.2 %  tma_retiring
             TopdownL1 (cpu_atom)                 #     31.9 %  tma_backend_bound
                                                  #     37.6 %  tma_frontend_bound       (49.66%)
                                                  #     18.0 %  tma_bad_speculation
                                                  #     12.6 %  tma_retiring             (59.58%)
```

Signed-off-by: Ian Rogers <irogers@google.com>
Signed-off-by: Namhyung Kim <namhyung@kernel.org>
5 weeks agoperf stat: Remove hard coded shadow metrics
Ian Rogers [Tue, 11 Nov 2025 21:21:55 +0000 (13:21 -0800)] 
perf stat: Remove hard coded shadow metrics

Now that the metrics are encoded in common json the hard coded
printing means the metrics are shown twice. Remove the hard coded
version.

This means that when specifying events, and those events correspond to
a hard coded metric, the metric will no longer be displayed. The
metric will be displayed if the metric is requested. Due to the adhoc
printing in the previous approach it was often found frustrating, the
new approach avoids this.

The default perf stat output on an alderlake now looks like:
```
$ perf stat -a -- sleep 1

 Performance counter stats for 'system wide':

            19,697      context-switches                 #      nan cs/sec  cs_per_second
             TopdownL1 (cpu_core)                 #     10.7 %  tma_bad_speculation
                                                  #     24.9 %  tma_frontend_bound
             TopdownL1 (cpu_core)                 #     34.3 %  tma_backend_bound
                                                  #     30.1 %  tma_retiring
             6,593      page-faults                      #      nan faults/sec  page_faults_per_second
       729,065,658      cpu_atom/cpu-cycles/             #      nan GHz  cycles_frequency       (49.79%)
     1,605,131,101      cpu_core/cpu-cycles/             #      nan GHz  cycles_frequency
                                                  #     19.7 %  tma_bad_speculation
                                                  #     14.2 %  tma_retiring             (50.14%)
                                                  #     37.3 %  tma_frontend_bound       (50.31%)
        87,302,268      cpu_atom/branches/               #      nan M/sec  branch_frequency     (60.27%)
       512,046,956      cpu_core/branches/               #      nan M/sec  branch_frequency
             1,111      cpu-migrations                   #      nan migrations/sec  migrations_per_second
                                                  #     28.8 %  tma_backend_bound        (60.26%)
              0.00 msec cpu-clock                        #      0.0 CPUs  CPUs_utilized
       392,509,323      cpu_atom/instructions/           #      0.6 instructions  insn_per_cycle  (60.19%)
     2,990,369,310      cpu_core/instructions/           #      1.9 instructions  insn_per_cycle
         3,493,478      cpu_atom/branch-misses/          #      5.9 %  branch_miss_rate         (49.69%)
         7,297,531      cpu_core/branch-misses/          #      1.4 %  branch_miss_rate

       1.006621701 seconds time elapsed
```

Signed-off-by: Ian Rogers <irogers@google.com>
Signed-off-by: Namhyung Kim <namhyung@kernel.org>
5 weeks agoperf script: Change metric format to use json metrics
Ian Rogers [Tue, 11 Nov 2025 21:21:54 +0000 (13:21 -0800)] 
perf script: Change metric format to use json metrics

The metric format option isn't properly supported. This change
improves that by making the sample events update the counts of an
evsel, where the shadow metric code expects to read the values.  To
support printing metrics, metrics need to be found. This is done on
the first attempt to print a metric. Every metric is parsed and then
the evsels in the metric's evlist compared to those in perf script
using the perf_event_attr type and config. If the metric matches then
it is added for printing. As an event in the perf script's evlist may
have >1 metric id, or different leader for aggregation, the first
metric matched will be displayed in those cases.

An example use is:
```
$ perf record -a -e '{instructions,cpu-cycles}:S' -a -- sleep 1
$ perf script -F period,metric
...
     867817
         metric:    0.30  insn per cycle
     125394
         metric:    0.04  insn per cycle
     313516
         metric:    0.11  insn per cycle
         metric:    1.00  insn per cycle
```

Signed-off-by: Ian Rogers <irogers@google.com>
Signed-off-by: Namhyung Kim <namhyung@kernel.org>
5 weeks agoperf stat: Add detail -d,-dd,-ddd metrics
Ian Rogers [Tue, 11 Nov 2025 21:21:53 +0000 (13:21 -0800)] 
perf stat: Add detail -d,-dd,-ddd metrics

Add metrics for the stat-shadow -d, -dd and -ddd events and hard coded
metrics. Remove the events as these now come from the metrics.

Following this change a detailed perf stat output looks like:
```
$ perf stat -a -ddd -- sleep 1
 Performance counter stats for 'system wide':

            21,089      context-switches                 #      nan cs/sec  cs_per_second
             TopdownL1 (cpu_core)                 #     14.1 %  tma_bad_speculation
                                                  #     27.3 %  tma_frontend_bound       (30.56%)
             TopdownL1 (cpu_core)                 #     31.5 %  tma_backend_bound
                                                  #     27.2 %  tma_retiring             (30.56%)
             6,302      page-faults                      #      nan faults/sec  page_faults_per_second
       928,495,163      cpu_atom/cpu-cycles/
                                                  #      nan GHz  cycles_frequency       (28.41%)
     1,841,409,834      cpu_core/cpu-cycles/
                                                  #      nan GHz  cycles_frequency       (38.51%)
                                                  #     14.5 %  tma_bad_speculation
                                                  #     16.0 %  tma_retiring             (28.41%)
                                                  #     36.8 %  tma_frontend_bound       (35.57%)
       100,859,118      cpu_atom/branches/               #      nan M/sec  branch_frequency     (42.73%)
       572,657,734      cpu_core/branches/               #      nan M/sec  branch_frequency     (54.43%)
             1,527      cpu-migrations                   #      nan migrations/sec  migrations_per_second
                                                  #     32.7 %  tma_backend_bound        (42.73%)
              0.00 msec cpu-clock                        #    0.000 CPUs utilized
                                                  #      0.0 CPUs  CPUs_utilized
       498,668,509      cpu_atom/instructions/           #    0.57  insn per cycle
                                                  #      0.6 instructions  insn_per_cycle  (42.97%)
     3,281,762,225      cpu_core/instructions/           #    1.84  insn per cycle
                                                  #      1.8 instructions  insn_per_cycle  (62.20%)
         4,919,511      cpu_atom/branch-misses/          #    5.43% of all branches
                                                  #      5.4 %  branch_miss_rate         (35.80%)
         7,431,776      cpu_core/branch-misses/          #    1.39% of all branches
                                                  #      1.4 %  branch_miss_rate         (62.20%)
         2,517,007      cpu_atom/LLC-loads/              #      0.1 %  llc_miss_rate            (28.62%)
         3,931,318      cpu_core/LLC-loads/              #     40.4 %  llc_miss_rate            (45.98%)
        14,918,674      cpu_core/L1-dcache-load-misses/  #    2.25% of all L1-dcache accesses
                                                  #      nan %  l1d_miss_rate            (37.80%)
        27,067,264      cpu_atom/L1-icache-load-misses/  #   15.92% of all L1-icache accesses
                                                  #     15.9 %  l1i_miss_rate            (21.47%)
       116,848,994      cpu_atom/dTLB-loads/             #      0.8 %  dtlb_miss_rate           (21.47%)
       764,870,407      cpu_core/dTLB-loads/             #      0.1 %  dtlb_miss_rate           (15.12%)

       1.006181526 seconds time elapsed
```

Signed-off-by: Ian Rogers <irogers@google.com>
Signed-off-by: Namhyung Kim <namhyung@kernel.org>
5 weeks agoperf jevents: Add metric DefaultShowEvents
Ian Rogers [Tue, 11 Nov 2025 21:21:52 +0000 (13:21 -0800)] 
perf jevents: Add metric DefaultShowEvents

Some Default group metrics require their events showing for
consistency with perf's previous behavior. Add a flag to indicate when
this is the case and use it in stat-display.

As events are coming from Default metrics remove that default hardware
and software events from perf stat.

Following this change the default perf stat output on an alderlake looks like:
```
$ perf stat -a -- sleep 1

 Performance counter stats for 'system wide':

            20,550      context-switches                 #      nan cs/sec  cs_per_second
             TopdownL1 (cpu_core)                 #      9.0 %  tma_bad_speculation
                                                  #     28.1 %  tma_frontend_bound
             TopdownL1 (cpu_core)                 #     29.2 %  tma_backend_bound
                                                  #     33.7 %  tma_retiring
             6,685      page-faults                      #      nan faults/sec  page_faults_per_second
       790,091,064      cpu_atom/cpu-cycles/
                                                  #      nan GHz  cycles_frequency       (49.83%)
     2,563,918,366      cpu_core/cpu-cycles/
                                                  #      nan GHz  cycles_frequency
                                                  #     12.3 %  tma_bad_speculation
                                                  #     14.5 %  tma_retiring             (50.20%)
                                                  #     33.8 %  tma_frontend_bound       (50.24%)
        76,390,322      cpu_atom/branches/               #      nan M/sec  branch_frequency     (60.20%)
     1,015,173,047      cpu_core/branches/               #      nan M/sec  branch_frequency
             1,325      cpu-migrations                   #      nan migrations/sec  migrations_per_second
                                                  #     39.3 %  tma_backend_bound        (60.17%)
              0.00 msec cpu-clock                        #    0.000 CPUs utilized
                                                  #      0.0 CPUs  CPUs_utilized
       554,347,072      cpu_atom/instructions/           #    0.64  insn per cycle
                                                  #      0.6 instructions  insn_per_cycle  (60.14%)
     5,228,931,991      cpu_core/instructions/           #    2.04  insn per cycle
                                                  #      2.0 instructions  insn_per_cycle
         4,308,874      cpu_atom/branch-misses/          #    5.65% of all branches
                                                  #      5.6 %  branch_miss_rate         (49.76%)
         9,890,606      cpu_core/branch-misses/          #    0.97% of all branches
                                                  #      1.0 %  branch_miss_rate

       1.005477803 seconds time elapsed
```

Signed-off-by: Ian Rogers <irogers@google.com>
Signed-off-by: Namhyung Kim <namhyung@kernel.org>
5 weeks agoperf jevents: Add set of common metrics based on default ones
Ian Rogers [Tue, 11 Nov 2025 21:21:51 +0000 (13:21 -0800)] 
perf jevents: Add set of common metrics based on default ones

Add support to getting a common set of metrics from a default
table. It simplifies the generation to add json metrics at the same
time. The metrics added are CPUs_utilized, cs_per_second,
migrations_per_second, page_faults_per_second, insn_per_cycle,
stalled_cycles_per_instruction, frontend_cycles_idle,
backend_cycles_idle, cycles_frequency, branch_frequency and
branch_miss_rate based on the shadow metric definitions.

Following this change the default perf stat output on an alderlake
looks like:
```
$ perf stat -a -- sleep 2

 Performance counter stats for 'system wide':

              0.00 msec cpu-clock                        #    0.000 CPUs utilized
            77,739      context-switches
            15,033      cpu-migrations
           321,313      page-faults
    14,355,634,225      cpu_atom/instructions/           #    1.40  insn per cycle              (35.37%)
   134,561,560,583      cpu_core/instructions/           #    3.44  insn per cycle              (57.85%)
    10,263,836,145      cpu_atom/cycles/                                                        (35.42%)
    39,138,632,894      cpu_core/cycles/                                                        (57.60%)
     2,989,658,777      cpu_atom/branches/                                                      (42.60%)
    32,170,570,388      cpu_core/branches/                                                      (57.39%)
        29,789,870      cpu_atom/branch-misses/          #    1.00% of all branches             (42.69%)
       165,991,152      cpu_core/branch-misses/          #    0.52% of all branches             (57.19%)
                       (software)                 #      nan cs/sec  cs_per_second
             TopdownL1 (cpu_core)                 #     11.9 %  tma_bad_speculation
                                                  #     19.6 %  tma_frontend_bound       (63.97%)
             TopdownL1 (cpu_core)                 #     18.8 %  tma_backend_bound
                                                  #     49.7 %  tma_retiring             (63.97%)
                       (software)                 #      nan faults/sec  page_faults_per_second
                                                  #      nan GHz  cycles_frequency       (42.88%)
                                                  #      nan GHz  cycles_frequency       (69.88%)
             TopdownL1 (cpu_atom)                 #     11.7 %  tma_bad_speculation
                                                  #     29.9 %  tma_retiring             (50.07%)
             TopdownL1 (cpu_atom)                 #     31.3 %  tma_frontend_bound       (43.09%)
                       (cpu_atom)                 #      nan M/sec  branch_frequency     (43.09%)
                                                  #      nan M/sec  branch_frequency     (70.07%)
                                                  #      nan migrations/sec  migrations_per_second
             TopdownL1 (cpu_atom)                 #     27.1 %  tma_backend_bound        (43.08%)
                       (software)                 #      0.0 CPUs  CPUs_utilized
                                                  #      1.4 instructions  insn_per_cycle  (43.04%)
                                                  #      3.5 instructions  insn_per_cycle  (69.99%)
                                                  #      1.0 %  branch_miss_rate         (35.46%)
                                                  #      0.5 %  branch_miss_rate         (65.02%)

       2.005626564 seconds time elapsed
```

Signed-off-by: Ian Rogers <irogers@google.com>
Signed-off-by: Namhyung Kim <namhyung@kernel.org>
5 weeks agoperf expr: Add #target_cpu literal
Ian Rogers [Tue, 11 Nov 2025 21:21:50 +0000 (13:21 -0800)] 
perf expr: Add #target_cpu literal

For CPU nanoseconds a lot of the stat-shadow metrics use either
task-clock or cpu-clock, the latter being used when
target__has_cpu. Add a #target_cpu literal so that json metrics can
perform the same test.

Signed-off-by: Ian Rogers <irogers@google.com>
Signed-off-by: Namhyung Kim <namhyung@kernel.org>
5 weeks agoperf metricgroup: Add care to picking the evsel for displaying a metric
Ian Rogers [Tue, 11 Nov 2025 21:21:49 +0000 (13:21 -0800)] 
perf metricgroup: Add care to picking the evsel for displaying a metric

Rather than using the first evsel in the matched events, try to find
the least shared non-tool evsel. The aim is to pick the first evsel
that typifies the metric within the list of metrics.

This addresses an issue where Default metric group metrics may lose
their counter value due to how the stat displaying hides counters for
default event/metric output.

For a metricgroup like TopdownL1 on an Intel Alderlake the change is,
before there are 4 events with metrics:
```
$ perf stat -M topdownL1 -a sleep 1

 Performance counter stats for 'system wide':

     7,782,334,296      cpu_core/TOPDOWN.SLOTS/          #     10.4 %  tma_bad_speculation
                                                  #     19.7 %  tma_frontend_bound
     2,668,927,977      cpu_core/topdown-retiring/       #     35.7 %  tma_backend_bound
                                                  #     34.1 %  tma_retiring
       803,623,987      cpu_core/topdown-bad-spec/
       167,514,386      cpu_core/topdown-heavy-ops/
     1,555,265,776      cpu_core/topdown-fe-bound/
     2,792,733,013      cpu_core/topdown-be-bound/
       279,769,310      cpu_atom/TOPDOWN_RETIRING.ALL/   #     12.2 %  tma_retiring
                                                  #     15.1 %  tma_bad_speculation
       457,917,232      cpu_atom/CPU_CLK_UNHALTED.CORE/  #     38.4 %  tma_backend_bound
                                                  #     34.2 %  tma_frontend_bound
       783,519,226      cpu_atom/TOPDOWN_FE_BOUND.ALL/
        10,790,192      cpu_core/INT_MISC.UOP_DROPPING/
       879,845,633      cpu_atom/TOPDOWN_BE_BOUND.ALL/
```

After there are 6 events with metrics:
```
$ perf stat -M topdownL1 -a sleep 1

 Performance counter stats for 'system wide':

     2,377,551,258      cpu_core/TOPDOWN.SLOTS/          #      7.9 %  tma_bad_speculation
                                                  #     36.4 %  tma_frontend_bound
       480,791,142      cpu_core/topdown-retiring/       #     35.5 %  tma_backend_bound
       186,323,991      cpu_core/topdown-bad-spec/
        65,070,590      cpu_core/topdown-heavy-ops/      #     20.1 %  tma_retiring
       871,733,444      cpu_core/topdown-fe-bound/
       848,286,598      cpu_core/topdown-be-bound/
       260,936,456      cpu_atom/TOPDOWN_RETIRING.ALL/   #     12.4 %  tma_retiring
                                                  #     17.6 %  tma_bad_speculation
       419,576,513      cpu_atom/CPU_CLK_UNHALTED.CORE/
       797,132,597      cpu_atom/TOPDOWN_FE_BOUND.ALL/   #     38.0 %  tma_frontend_bound
         3,055,447      cpu_core/INT_MISC.UOP_DROPPING/
       671,014,164      cpu_atom/TOPDOWN_BE_BOUND.ALL/   #     32.0 %  tma_backend_bound
```

Signed-off-by: Ian Rogers <irogers@google.com>
Signed-off-by: Namhyung Kim <namhyung@kernel.org>
5 weeks agoperf tools: Fix missing feature check for inherit + SAMPLE_READ
Namhyung Kim [Tue, 11 Nov 2025 07:59:44 +0000 (23:59 -0800)] 
perf tools: Fix missing feature check for inherit + SAMPLE_READ

It should also have PERF_SAMPLE_TID to enable inherit and PERF_SAMPLE_READ
on recent kernels.  Not having _TID makes the feature check wrongly detect
the inherit and _READ support.

It was reported that the following command failed due to the error in
the missing feature check on Intel SPR machines.

  $ perf record -e '{cpu/mem-loads-aux/S,cpu/mem-loads,ldlat=3/PS}' -- ls
  Error:
  Failure to open event 'cpu/mem-loads,ldlat=3/PS' on PMU 'cpu' which will be removed.
  Invalid event (cpu/mem-loads,ldlat=3/PS) in per-thread mode, enable system wide with '-a'.

Reviewed-by: Ian Rogers <irogers@google.com>
Fixes: 3b193a57baf15c468 ("perf tools: Detect missing kernel features properly")
Reported-and-tested-by: Chen, Zide <zide.chen@intel.com>
Closes: https://lore.kernel.org/lkml/20251022220802.1335131-1-zide.chen@intel.com/
Signed-off-by: Namhyung Kim <namhyung@kernel.org>
5 weeks agoperf symbol: Remove unneeded semicolon
Chen Ni [Tue, 11 Nov 2025 05:44:15 +0000 (13:44 +0800)] 
perf symbol: Remove unneeded semicolon

Remove unnecessary semicolons reported by Coccinelle/coccicheck and the
semantic patch at scripts/coccinelle/misc/semicolon.cocci.

Signed-off-by: Chen Ni <nichen@iscas.ac.cn>
Signed-off-by: Namhyung Kim <namhyung@kernel.org>
5 weeks agoperf test: Add test that command line period overrides sysfs/json values
Ian Rogers [Sun, 9 Nov 2025 00:59:59 +0000 (16:59 -0800)] 
perf test: Add test that command line period overrides sysfs/json values

The behavior of weak terms is subtle, add a test that they aren't
accidentally broken. The test finds an event with a weak 'period' and
then overrides it. In no such event is present then the test skips.

Signed-off-by: Ian Rogers <irogers@google.com>
Signed-off-by: Namhyung Kim <namhyung@kernel.org>
5 weeks agoperf pmu: Make pmu_alias_terms weak again
Ian Rogers [Sun, 9 Nov 2025 00:59:58 +0000 (16:59 -0800)] 
perf pmu: Make pmu_alias_terms weak again

The terms for a json event should be weak so they don't override
command line options.

Before:
```
$ perf record -vv -c 1000 -e uops_issued.any -o /dev/null true 2>&1
|grep "{ sample_period, sample_freq }"
 { sample_period, sample_freq }   200003
 { sample_period, sample_freq }   2000003
 { sample_period, sample_freq }   1000
```

After:
```
$ perf record -vv -c 1000 -e uops_issued.any -o /dev/null true 2>&1
|grep "{ sample_period, sample_freq }"
 { sample_period, sample_freq }   1000
 { sample_period, sample_freq }   1000
 { sample_period, sample_freq }   1000
```

Fixes: 84bae3af20d0 ("perf pmu: Don't eagerly parse event terms")
Signed-off-by: Ian Rogers <irogers@google.com>
Signed-off-by: Namhyung Kim <namhyung@kernel.org>
6 weeks agoperf tool: Add a delegate_tool that just delegates actions to another tool
Ian Rogers [Fri, 7 Nov 2025 17:07:12 +0000 (09:07 -0800)] 
perf tool: Add a delegate_tool that just delegates actions to another tool

Add an ability to be able to compose perf_tools, by having one perform
an action and then calling a delegate. Currently the perf_tools have
if-then-elses setting the callback and then if-then-elses within the
callback. Understanding the behavior is complex as it is in two places
and logic for numerous operations, within things like perf inject, is
interwoven. By chaining perf_tools together based on command line
options this kind of code can be avoided.

Signed-off-by: Ian Rogers <irogers@google.com>
Signed-off-by: Namhyung Kim <namhyung@kernel.org>
6 weeks agoperf tool: Add the perf_tool argument to all callbacks
Ian Rogers [Fri, 7 Nov 2025 17:07:11 +0000 (09:07 -0800)] 
perf tool: Add the perf_tool argument to all callbacks

Getting context for what a tool is doing, such as the perf_inject
instance, using container_of the tool is a common pattern in the
code. This isn't possible event_op2, event_op3 and event_op4 callbacks
as the tool isn't passed. Add the argument and then fix function
signatures to match. As tools maybe reading a tool from somewhere
else, change that code to use the passed in tool.

Signed-off-by: Ian Rogers <irogers@google.com>
Signed-off-by: Namhyung Kim <namhyung@kernel.org>
6 weeks agoperf vendor events arm64:: Add i.MX94 DDR Performance Monitor metrics
Xu Yang [Thu, 21 Aug 2025 11:01:51 +0000 (19:01 +0800)] 
perf vendor events arm64:: Add i.MX94 DDR Performance Monitor metrics

Add JSON metrics for i.MX94 DDR Performance Monitor.

Reviewed-by: Peng Fan <peng.fan@nxp.com>
Reviewed-by: Frank Li <Frank.Li@nxp.com>
Signed-off-by: Xu Yang <xu.yang_2@nxp.com>
Acked-by: Ian Rogers <irogers@google.com>
Acked-by: Conor Dooley <conor.dooley@microchip.com>
Signed-off-by: Namhyung Kim <namhyung@kernel.org>
6 weeks agoperf stat: Add ScaleUnit to {cpu,task}-clock JSON description
Namhyung Kim [Thu, 6 Nov 2025 21:53:50 +0000 (13:53 -0800)] 
perf stat: Add ScaleUnit to {cpu,task}-clock JSON description

This changes the output of the event like below.  In fact, that's the
output it used to have before the JSON conversion.

Before:
  $ perf stat -e task-clock true

   Performance counter stats for 'true':

             313,848      task-clock                       #    0.290 CPUs utilized

         0.001081223 seconds time elapsed

         0.001122000 seconds user
         0.000000000 seconds sys

After:
  $ perf stat -e task-clock true

   Performance counter stats for 'true':

                0.36 msec task-clock                       #    0.297 CPUs utilized

         0.001225435 seconds time elapsed

         0.001268000 seconds user
         0.000000000 seconds sys

Reviewed-by: Ian Rogers <irogers@google.com>
Fixes: 9957d8c801fe0cb90 ("perf jevents: Add common software event json")
Signed-off-by: Namhyung Kim <namhyung@kernel.org>
6 weeks agoperf record: Make sure to update build-ID cache
Namhyung Kim [Thu, 6 Nov 2025 19:00:23 +0000 (11:00 -0800)] 
perf record: Make sure to update build-ID cache

Recent change on enabling --buildid-mmap by default brought an issue
with build-id handling.  With build-ID in MMAP2 records, we don't need
to save the build-ID table in the header of a perf data file.

But the actual file contents still need to be cached in the debug
directory for annotation etc.  Split the build-ID header processing and
caching and make sure perf record to save hit DSOs in the build-ID cache
by moving perf_session__cache_build_ids() to the end of the record__
finish_output().

Reviewed-by: Ian Rogers <irogers@google.com>
Signed-off-by: Namhyung Kim <namhyung@kernel.org>
6 weeks agoperf jevents: Make all tables static
Ian Rogers [Fri, 24 Oct 2025 17:58:41 +0000 (10:58 -0700)] 
perf jevents: Make all tables static

The tables created by jevents.py are only used within the pmu-events.c
file. Change the declarations of those global variables to be static
to encapsulate this.

Signed-off-by: Ian Rogers <irogers@google.com>
Signed-off-by: Namhyung Kim <namhyung@kernel.org>
6 weeks agoperf metricgroup: When copy metrics copy default information
Ian Rogers [Fri, 24 Oct 2025 17:58:39 +0000 (10:58 -0700)] 
perf metricgroup: When copy metrics copy default information

When copy metrics into a group also copy default information from the
original metrics.

Signed-off-by: Ian Rogers <irogers@google.com>
Signed-off-by: Namhyung Kim <namhyung@kernel.org>
6 weeks agoperf metricgroup: Missed free on error path
Ian Rogers [Fri, 24 Oct 2025 17:58:38 +0000 (10:58 -0700)] 
perf metricgroup: Missed free on error path

If an out-of-memory occurs the expr also needs freeing.

Signed-off-by: Ian Rogers <irogers@google.com>
Signed-off-by: Namhyung Kim <namhyung@kernel.org>
6 weeks agoperf metricgroup: Update comment on location of metric_event list
Ian Rogers [Fri, 24 Oct 2025 17:58:37 +0000 (10:58 -0700)] 
perf metricgroup: Update comment on location of metric_event list

Update comment as the stat_config no longer holds all metrics.

Signed-off-by: Ian Rogers <irogers@google.com>
Fixes: faebee18d720 ("perf stat: Move metric list from config to evlist")
Signed-off-by: Namhyung Kim <namhyung@kernel.org>
6 weeks agoperf evsel: Remove unused metric_events variable
Ian Rogers [Fri, 24 Oct 2025 17:58:36 +0000 (10:58 -0700)] 
perf evsel: Remove unused metric_events variable

The metric_events exist in the metric_expr list and so this variable
has been unused for a while.

Signed-off-by: Ian Rogers <irogers@google.com>
Signed-off-by: Namhyung Kim <namhyung@kernel.org>
7 weeks agoperf tools: Cache counter names for raw samples on s390
Ian Rogers [Fri, 31 Oct 2025 19:42:16 +0000 (12:42 -0700)] 
perf tools: Cache counter names for raw samples on s390

Searching all event names is slower now that legacy names are
included. Add a cache to avoid long iterative searches. Note, the
cache isn't cleaned up and is as such a memory leak, however, globally
reachable leaks like this aren't treated as leaks by leak sanitizer.

Reported-by: Thomas Richter <tmricht@linux.ibm.com>
Closes: https://lore.kernel.org/linux-perf-users/09943f4f-516c-4b93-877c-e4a64ed61d38@linux.ibm.com/
Signed-off-by: Ian Rogers <irogers@google.com>
Tested-by: Thomas Richter <tmricht@linux.ibm.com>
Signed-off-by: Namhyung Kim <namhyung@kernel.org>
7 weeks agoperf trace: Increase syscall handler map size to 1024
Namhyung Kim [Mon, 19 May 2025 23:25:39 +0000 (16:25 -0700)] 
perf trace: Increase syscall handler map size to 1024

The syscalls_sys_{enter,exit} map in augmented_raw_syscalls.bpf.c has
max entries of 512.  Usually syscall numbers are smaller than this but
x86 has x32 ABI where syscalls start from 512.

That makes trace__init_syscalls_bpf_prog_array_maps() fail in the middle
of the loop when it accesses those keys.  As the loop iteration is not
ordered by syscall numbers anymore, the failure can affect non-x32
syscalls.

Let's increase the map size to 1024 so that it can handle those ABIs
too.  While most systems won't need this, increasing the size will be
safer for potential future changes.

Reviewed-by: Howard Chu <howardchu95@gmail.com>
Signed-off-by: Namhyung Kim <namhyung@kernel.org>
7 weeks agoperf vendor events AmpereOneX: Fix spelling typo in the metrics file
Chu Guangqing [Fri, 31 Oct 2025 02:58:10 +0000 (10:58 +0800)] 
perf vendor events AmpereOneX: Fix spelling typo in the metrics file

The json file incorrectly used "acceses" instead of "accesses".

Signed-off-by: Chu Guangqing <chuguangqing@inspur.com>
Reviewed-by: James Clark <james.clark@linaro.org>
Reviewed-by: Ilkka Koskinen <ilkka@os.amperecomputing.com>
Signed-off-by: Namhyung Kim <namhyung@kernel.org>
7 weeks agoperf vendor events arm64: Fix typo in Ampere eMag json file
Chu Guangqing [Fri, 31 Oct 2025 03:17:29 +0000 (11:17 +0800)] 
perf vendor events arm64: Fix typo in Ampere eMag json file

Correct instruction spelling errors.

Signed-off-by: Chu Guangqing <chuguangqing@inspur.com>
Reviewed-by: James Clark <james.clark@linaro.org>
Reviewed-by: Ilkka Koskinen <ilkka@os.amperecomputing.com>
Signed-off-by: Namhyung Kim <namhyung@kernel.org>
7 weeks agoperf record: skip synthesize event when open evsel failed
Shuai Xue [Thu, 23 Oct 2025 01:50:43 +0000 (09:50 +0800)] 
perf record: skip synthesize event when open evsel failed

When using perf record with the `--overwrite` option, a segmentation fault
occurs if an event fails to open. For example:

  perf record -e cycles-ct -F 1000 -a --overwrite
  Error:
  cycles-ct:H: PMU Hardware doesn't support sampling/overflow-interrupts. Try 'perf stat'
  perf: Segmentation fault
      #0 0x6466b6 in dump_stack debug.c:366
      #1 0x646729 in sighandler_dump_stack debug.c:378
      #2 0x453fd1 in sigsegv_handler builtin-record.c:722
      #3 0x7f8454e65090 in __restore_rt libc-2.32.so[54090]
      #4 0x6c5671 in __perf_event__synthesize_id_index synthetic-events.c:1862
      #5 0x6c5ac0 in perf_event__synthesize_id_index synthetic-events.c:1943
      #6 0x458090 in record__synthesize builtin-record.c:2075
      #7 0x45a85a in __cmd_record builtin-record.c:2888
      #8 0x45deb6 in cmd_record builtin-record.c:4374
      #9 0x4e5e33 in run_builtin perf.c:349
      #10 0x4e60bf in handle_internal_command perf.c:401
      #11 0x4e6215 in run_argv perf.c:448
      #12 0x4e653a in main perf.c:555
      #13 0x7f8454e4fa72 in __libc_start_main libc-2.32.so[3ea72]
      #14 0x43a3ee in _start ??:0

The --overwrite option implies --tail-synthesize, which collects non-sample
events reflecting the system status when recording finishes. However, when
evsel opening fails (e.g., unsupported event 'cycles-ct'), session->evlist
is not initialized and remains NULL. The code unconditionally calls
record__synthesize() in the error path, which iterates through the NULL
evlist pointer and causes a segfault.

To fix it, move the record__synthesize() call inside the error check block, so
it's only called when there was no error during recording, ensuring that evlist
is properly initialized.

Fixes: 4ea648aec019 ("perf record: Add --tail-synthesize option")
Signed-off-by: Shuai Xue <xueshuai@linux.alibaba.com>
Signed-off-by: Namhyung Kim <namhyung@kernel.org>
7 weeks agoperf lock contention: Load kernel map before lookup
Namhyung Kim [Thu, 30 Oct 2025 04:01:39 +0000 (21:01 -0700)] 
perf lock contention: Load kernel map before lookup

On some machines, it caused troubles when it tried to find kernel
symbols.  I think it's because kernel modules and kallsyms are messed
up during load and split.

Basically we want to make sure the kernel map is loaded and the code has
it in the lock_contention_read().  But recently we added more lookups in
the lock_contention_prepare() which is called before _read().

Also the kernel map (kallsyms) may not be the first one in the group
like on ARM.  Let's use machine__kernel_map() rather than just loading
the first map.

Reviewed-by: Ian Rogers <irogers@google.com>
Fixes: 688d2e8de231c54e ("perf lock contention: Add -l/--lock-addr option")
Signed-off-by: Namhyung Kim <namhyung@kernel.org>
7 weeks agoperf test workload: Add thread count argument to thloop
Ian Rogers [Tue, 28 Oct 2025 15:38:20 +0000 (08:38 -0700)] 
perf test workload: Add thread count argument to thloop

Allow the number of threads for the thloop workload to be increased
beyond the normal 2. Add error checking to the parsed time and thread
count values.

Signed-off-by: Ian Rogers <irogers@google.com>
Acked-by: Arnaldo Carvalho de Melo <acme@redhat.com>
Signed-off-by: Namhyung Kim <namhyung@kernel.org>
7 weeks agoperf hwmon_pmu: Fix uninitialized variable warning
Michal Suchanek [Wed, 22 Oct 2025 10:38:35 +0000 (12:38 +0200)] 
perf hwmon_pmu: Fix uninitialized variable warning

The line_len is only set on success. Check the return value instead.

 util/hwmon_pmu.c: In function ‘perf_pmus__read_hwmon_pmus’:
 util/hwmon_pmu.c:742:20: warning: ‘line_len’ may be used uninitialized [-Wmaybe-uninitialized]
   742 |                 if (line_len > 0 && line[line_len - 1] == '\n')
       |                    ^
 util/hwmon_pmu.c:719:24: note: ‘line_len’ was declared here
   719 |                 size_t line_len;

Fixes: 53cc0b351ec9 ("perf hwmon_pmu: Add a tool PMU exposing events from hwmon in sysfs")
Signed-off-by: Michal Suchanek <msuchanek@suse.de>
Signed-off-by: Namhyung Kim <namhyung@kernel.org>
8 weeks agoperf auxtrace: Add auxtrace_synth_id_range_start() helper
tanze [Fri, 24 Oct 2025 08:56:25 +0000 (16:56 +0800)] 
perf auxtrace: Add auxtrace_synth_id_range_start() helper

To avoid hardcoding the offset value for synthetic event IDs
in multiple auxtrace modules (arm-spe, cs-etm, intel-pt, etc.),
and to improve code reusability, this patch unifies
the handling of the ID offset via a dedicated helper function.

Signed-off-by: tanze <tanze@kylinos.cn>
Reviewed-by: Adrian Hunter <adrian.hunter@intel.com>
Tested-by: Leo Yan <leo.yan@arm.com>
Signed-off-by: Namhyung Kim <namhyung@kernel.org>
8 weeks agoperf stat: Add/fix bperf cgroup max events workarounds
Ian Rogers [Sat, 25 Oct 2025 20:28:34 +0000 (13:28 -0700)] 
perf stat: Add/fix bperf cgroup max events workarounds

Commit b8308511f6e0 bumped the max events to 1024 but this results in
BPF verifier issues if the number of command line events is too
large. Workaround this by:

1) moving the constants to a header file to share between BPF and perf
   C code,
2) testing that the maximum number of events doesn't cause BPF
   verifier issues in debug builds,
3) lower the max events from 1024 to 128,
4) in perf stat, if there are more events than the BPF counters can
   support then disable BPF counter usage.

The rodata setup is factored into its own function to avoid
duplicating it in the testing code.

Signed-off-by: Ian Rogers <irogers@google.com>
Fixes: b8308511f6e0 ("perf stat bperf cgroup: Increase MAX_EVENTS from 32 to 1024")
Signed-off-by: Namhyung Kim <namhyung@kernel.org>
8 weeks agoperf cs-etm: Mute enumeration value warning
Leo Yan [Wed, 22 Oct 2025 14:28:49 +0000 (15:28 +0100)] 
perf cs-etm: Mute enumeration value warning

When the OpenCSD library introduces a new enumeration value (for example,
in the v1.7.1 release), the perf build fails with an error:

  util/cs-etm-decoder/cs-etm-decoder.c:600:10: error: enumeration value 'OCSD_GEN_TRC_ELEM_ITMTRACE' not explicitly handled in switch [-Werror, -Wswitch-enum]
    600 |         switch (elem->elem_type) {
|                 ^~~~~~~~~~~~~~~
  1 error generated.

Convert to if-else sentences to mute the enumeration value warning,
which can avoid build failures whenever the lib is updated.

No functional change.

Suggested-by: Suzuki K Poulose <suzuki.poulose@arm.com>
Signed-off-by: Leo Yan <leo.yan@arm.com>
Reviewed-by: James Clark <james.clark@linaro.org>
Signed-off-by: Namhyung Kim <namhyung@kernel.org>
8 weeks agotools: arm64: Add Cortex-A720AE definitions
Kuninori Morimoto [Tue, 21 Oct 2025 23:21:48 +0000 (23:21 +0000)] 
tools: arm64: Add Cortex-A720AE definitions

Add cputype definitions for Cortex-A720AE. These will be used for errata
detection in subsequent patches.

These values can be found in the Cortex-A720AE TRM:

https://developer.arm.com/documentation/102828/0001/

... in Table A-187

Signed-off-by: Kuninori Morimoto <kuninori.morimoto.gx@renesas.com>
Reviewed-by: Leo Yan <leo.yan@arm.com>
Acked-by: Will Deacon <will@kernel.org>
Signed-off-by: Namhyung Kim <namhyung@kernel.org>
8 weeks agoperf annotate: Fix Clang build by adding block in switch case
James Clark [Wed, 22 Oct 2025 11:02:40 +0000 (12:02 +0100)] 
perf annotate: Fix Clang build by adding block in switch case

Clang and GCC disagree with what constitutes a "declaration after
statement". GCC allows declarations in switch cases without an extra
block, as long as it's immediately after the label. Clang does not.
Unfortunately this is the case even in the latest versions of both
compilers. The only option that makes them behave in the same way is
-Wpedantic, which can't be enabled in Perf because of the number of
warnings it generates.

Add a block to fix the Clang build, which is the only thing we can do.

Fixes the build error:

  ui/browsers/annotate.c:999:4: error: expected expression
  struct annotation_line *al = NULL;

  ui/browsers/annotate.c:1008:4: error: use of undeclared identifier 'al'
  al = annotated_source__get_line(notes->src, offset);

  ui/browsers/annotate.c:1009:24: error: use of undeclared identifier 'al'
  browser->curr_hot = al ? &al->rb_node : NULL;

  ui/browsers/annotate.c:1009:30: error: use of undeclared identifier 'al'
  browser->curr_hot = al ? &al->rb_node : NULL;

  ui/browsers/annotate.c:1000:8: error: mixing declarations and code is incompatible with standards before C99 [-Werror,-Wdeclaration-after-statement]
  s64 offset = annotate_browser__curr_hot_offset(browser);

Fixes: ad83f3b7155d ("perf c2c annotate: Start from the contention line")
Signed-off-by: James Clark <james.clark@linaro.org>
Reviewed-by: Ian Rogers <irogers@google.com>
Signed-off-by: Namhyung Kim <namhyung@kernel.org>
2 months agoperf annotate: Invalidate register states for untracked instructions
Zecheng Li [Mon, 13 Oct 2025 18:16:02 +0000 (18:16 +0000)] 
perf annotate: Invalidate register states for untracked instructions

When tracking variable types, instructions that modify a pointer value
in an untracked way can lead to incorrect type propagation. To prevent
this, invalidate the register state when encountering such instructions.

This change invalidates pointer types for various arithmetic and bitwise
operations that current pointer offset tracking doesn't support, like
imul, shl, and, inc, etc.

A special case is added for 'xor reg, reg', which is a common idiom for
zeroing a register. For this, the register state is updated to be a
constant with a value of 0.

This could introduce slight regressions if a variable is zeroed and then
reused. This can be addressed in the future by using all DWARF locations
for instruction tracking instead of only the first one.

Signed-off-by: Zecheng Li <zecheng@google.com>
Signed-off-by: Namhyung Kim <namhyung@kernel.org>
2 months agoperf annotate: Save pointer offset in stack state
Zecheng Li [Mon, 13 Oct 2025 18:16:01 +0000 (18:16 +0000)] 
perf annotate: Save pointer offset in stack state

The tracked pointer offset was not being preserved in the stack state,
which could lead to incorrect type analysis. This change adds a
ptr_offset field to the type_state_stack struct and passes it to
set_stack_state and findnew_stack_state to ensure the offset is
preserved after the pointer is loaded from a stack location. It improves
the type annotation coverage and quality.

Signed-off-by: Zecheng Li <zecheng@google.com>
Signed-off-by: Namhyung Kim <namhyung@kernel.org>
2 months agoperf annotate: Track arithmetic instructions on pointers
Zecheng Li [Mon, 13 Oct 2025 18:16:00 +0000 (18:16 +0000)] 
perf annotate: Track arithmetic instructions on pointers

Track the arithmetic operations on registers with pointer types. We
handle only add, sub and lea instructions. The original pointer
information needs to be preserved for getting outermost struct types.
For example, reg0 points to a struct cfs_rq, when we add 0x10 to reg0,
it should preserve the information of struct cfs_rq + 0x10 in the
register instead of a pointer type to the child field at 0x10.

Details:

1.  struct type_state_reg now includes an offset, indicating if the
    register points to the start or an internal part of its associated
    type. This offset is used in mem to reg and reg to stack mem
    transfers, and also applied to the final type offset.

2.  lea offset(%sp/%fp), reg is now treated as taking the address of a
    stack variable. It worked fine in most cases, but an issue with this
    approach is the pointer type may not exist.

3.  lea offset(%base), reg is handled by moving the type from %base and
    adding an offset, similar to an add operation followed by a mov reg
    to reg.

4.  Non-stack variables from DWARF with non-zero offsets in their
    location expressions are now accepted with register offset tracking.

Multi-register addressing modes in LEA are not supported.

Signed-off-by: Zecheng Li <zecheng@google.com>
Signed-off-by: Namhyung Kim <namhyung@kernel.org>
2 months agoperf annotate: Track address registers via TSR_KIND_POINTER
Zecheng Li [Mon, 13 Oct 2025 18:15:59 +0000 (18:15 +0000)] 
perf annotate: Track address registers via TSR_KIND_POINTER

Introduce TSR_KIND_POINTER to improve the data type profiler's ability
to track pointer-based memory accesses and address register variables.

TSR_KIND_POINTER represents that the location holds a pointer type to
the type in the type state. The semantics match the `breg` registers
that describe a memory location.

This change implements handling for this new kind in mov instructions
and in the check_matching_type() function. When a TSR_KIND_POINTER is
moved to the stack, the stack state size is set to the architecture's
pointer size.

Signed-off-by: Zecheng Li <zecheng@google.com>
Signed-off-by: Namhyung Kim <namhyung@kernel.org>
2 months agoperf annotate: Skip annotating data types to lea instructions
Zecheng Li [Mon, 13 Oct 2025 18:15:58 +0000 (18:15 +0000)] 
perf annotate: Skip annotating data types to lea instructions

Introduce a helper function is_address_gen_insn() to check
arch-dependent address generation instructions like lea in x86. Remove
type annotation on these instructions since they are not accessing
memory. It should be counted as `no_mem_ops`.

Signed-off-by: Zecheng Li <zecheng@google.com>
Signed-off-by: Namhyung Kim <namhyung@kernel.org>
2 months agoperf annotate: Check return value of evsel__get_arch() properly
Tianyou Li [Mon, 20 Oct 2025 07:30:05 +0000 (15:30 +0800)] 
perf annotate: Check return value of evsel__get_arch() properly

Check the error code of evsel__get_arch() in the symbol__annotate().
Previously it checked non-zero value but after the refactoring it does
only for negative values.

Fixes: 0669729eb0afb0cf ("perf annotate: Factor out evsel__get_arch()")
Suggested-by: James Clark <james.clark@linaro.org>
Acked-by: Namhyung Kim <namhyung@kernel.org>
Signed-off-by: Tianyou Li <tianyou.li@intel.com>
Signed-off-by: Namhyung Kim <namhyung@kernel.org>
2 months agoperf annotate: fix a crash when annotate the same symbol with 's' and 'T'
Tianyou Li [Mon, 20 Oct 2025 07:30:04 +0000 (15:30 +0800)] 
perf annotate: fix a crash when annotate the same symbol with 's' and 'T'

When perf report with annotation for a symbol, press 's' and 'T', then exit
the annotate browser. Once annotate the same symbol, the annotate browser
will crash.

The browser.arch was required to be correctly updated when data type
feature was enabled by 'T'. Usually it was initialized by symbol__annotate2
function. If a symbol has already been correctly annotated at the first
time, it should not call the symbol__annotate2 function again, thus the
browser.arch will not get initialized. Then at the second time to show the
annotate browser, the data type needs to be displayed but the browser.arch
is empty.

Stack trace as below:

Perf: Segmentation fault
-------- backtrace --------
    #0 0x55d365 in ui__signal_backtrace setup.c:0
    #1 0x7f5ff1a3e930 in __restore_rt libc.so.6[3e930]
    #2 0x570f08 in arch__is perf[570f08]
    #3 0x562186 in annotate_get_insn_location perf[562186]
    #4 0x562626 in __hist_entry__get_data_type annotate.c:0
    #5 0x56476d in annotation_line__write perf[56476d]
    #6 0x54e2db in annotate_browser__write annotate.c:0
    #7 0x54d061 in ui_browser__list_head_refresh perf[54d061]
    #8 0x54dc9e in annotate_browser__refresh annotate.c:0
    #9 0x54c03d in __ui_browser__refresh browser.c:0
    #10 0x54ccf8 in ui_browser__run perf[54ccf8]
    #11 0x54eb92 in __hist_entry__tui_annotate perf[54eb92]
    #12 0x552293 in do_annotate hists.c:0
    #13 0x55941c in evsel__hists_browse hists.c:0
    #14 0x55b00f in evlist__tui_browse_hists perf[55b00f]
    #15 0x42ff02 in cmd_report perf[42ff02]
    #16 0x494008 in run_builtin perf.c:0
    #17 0x494305 in handle_internal_command perf.c:0
    #18 0x410547 in main perf[410547]
    #19 0x7f5ff1a295d0 in __libc_start_call_main libc.so.6[295d0]
    #20 0x7f5ff1a29680 in __libc_start_main@@GLIBC_2.34 libc.so.6[29680]
    #21 0x410b75 in _start perf[410b75]

Fixes: 1d4374afd000 ("perf annotate: Add 'T' hot key to toggle data type display")
Reviewed-by: James Clark <james.clark@linaro.org>
Tested-by: Namhyung Kim <namhyung@kernel.org>
Signed-off-by: Tianyou Li <tianyou.li@intel.com>
Signed-off-by: Namhyung Kim <namhyung@kernel.org>
2 months agoperf annotate: Fix build with NO_SLANG=1
Namhyung Kim [Tue, 21 Oct 2025 03:07:50 +0000 (12:07 +0900)] 
perf annotate: Fix build with NO_SLANG=1

The recent change for perf c2c annotate broke build without slang
support like below.

  builtin-annotate.c: In function 'hists__find_annotations':
  builtin-annotate.c:522:73: error: 'NO_ADDR' undeclared (first use in this function); did you mean 'NR_ADDR'?
    522 |                         key = hist_entry__tui_annotate(he, evsel, NULL, NO_ADDR);
        |                                                                         ^~~~~~~
        |                                                                         NR_ADDR
  builtin-annotate.c:522:73: note: each undeclared identifier is reported only once for each function it appears in

  builtin-annotate.c:522:31: error: too many arguments to function 'hist_entry__tui_annotate'
    522 |                         key = hist_entry__tui_annotate(he, evsel, NULL, NO_ADDR);
        |                               ^~~~~~~~~~~~~~~~~~~~~~~~
  In file included from util/sort.h:6,
                   from builtin-annotate.c:28:
  util/hist.h:756:19: note: declared here
    756 | static inline int hist_entry__tui_annotate(struct hist_entry *he __maybe_unused,
        |                   ^~~~~~~~~~~~~~~~~~~~~~~~

And I noticed that it missed to update the other side of #ifdef
HAVE_SLANG_SUPPORT.  Let's fix it.

Cc: Tianyou Li <tianyou.li@intel.com>
Fixes: cd3466cd2639783d ("perf c2c: Add annotation support to perf c2c report")
Signed-off-by: Namhyung Kim <namhyung@kernel.org>
2 months agoperf jevents: Suppress circular dependency warnings
James Clark [Mon, 20 Oct 2025 16:08:28 +0000 (17:08 +0100)] 
perf jevents: Suppress circular dependency warnings

When doing an in source build, $(OUTPUT) is empty so the rule has the
same input and output file. Suppress the warning by only adding the rule
when doing an out of source build. The same condition already exists for
the clean rule for json files.

This fixes the following warnings:

  make[3]: Circular pmu-events/arch/nds32/mapfile.csv <- pmu-events/arch/nds32/mapfile.csv dependency dropped.
  make[3]: Circular pmu-events/arch/powerpc/mapfile.csv <- pmu-events/arch/powerpc/mapfile.csv dependency dropped.
  ...

Signed-off-by: James Clark <james.clark@linaro.org>
Tested-by: Namhyung Kim <namhyung@kernel.org>
Tested-by: Leo Yan <leo.yan@arm.com>
Signed-off-by: Namhyung Kim <namhyung@kernel.org>
2 months agoperf jevents: Remove unused makefile variable
James Clark [Mon, 20 Oct 2025 16:08:27 +0000 (17:08 +0100)] 
perf jevents: Remove unused makefile variable

JDIR is unused since commit 4bb55de4ff03 ("perf jevents: Support copying
the source json files to OUTPUT"), remove it.

Signed-off-by: James Clark <james.clark@linaro.org>
Tested-by: Leo Yan <leo.yan@arm.com>
Signed-off-by: Namhyung Kim <namhyung@kernel.org>
2 months agoperf jevents: Fix build when there are other json files in the tree
James Clark [Mon, 20 Oct 2025 16:08:26 +0000 (17:08 +0100)] 
perf jevents: Fix build when there are other json files in the tree

The unquoted glob *.json will expand to a real file if, for example,
there is any file in the Perf source ending in .json. This can happen
when using tools like Bear and clangd which generate a
compile_commands.json file. With the glob already expanded by the shell,
the find command will fail to wildcard any real json events files.

Fix it by wrapping the star in quotes so it's passed to find rather than
the shell.

This fixes the following build error (most of the diff output omitted):

  $ make V=1 -C tools/perf O=/tmp/perf_build_with_json

  TEST    /tmp/perf_build_with_json/pmu-events/empty-pmu-events.log
  ...
   /* offset=121053 */ "node-access\000legacy cache\000Local memory read accesses\000legacy-cache-config=6\000\00010\000\000\000\000\000"
   /* offset=121135 */ "node-misses\000legacy cache\000Local memory read misses\000legacy-cache-config=0x10006\000\00010\000\000\000\000\000"
   /* offset=121221 */ "node-miss\000legacy cache\000Local memory read misses\000legacy-cache-config=0x10006\000\00010\000\000\000\000\000"
  ...
  -       {
                  .event_table = { 0, 0 },
                  .metric_table = { 0, 0 },
          },
  make[3]: *** [pmu-events/Build:54: /tmp/perf_build_with_json/pmu-events/empty-pmu-events.log] Error 1

Fixes: 4bb55de4ff03 ("perf jevents: Support copying the source json files to OUTPUT")
Signed-off-by: James Clark <james.clark@linaro.org>
Reviewed-by: Leo Yan <leo.yan@arm.com>
Tested-by: Leo Yan <leo.yan@arm.com>
Signed-off-by: Namhyung Kim <namhyung@kernel.org>
2 months agoperf parse-events: Make X modifier more respectful of groups
Ian Rogers [Fri, 17 Oct 2025 23:03:57 +0000 (16:03 -0700)] 
perf parse-events: Make X modifier more respectful of groups

Events with an X modifier were reordered within a group, for example
slots was made the leader in:
```
$ perf record -e '{cpu/mem-stores/ppu,cpu/slots/uX}' -- sleep 1
```

Fix by making `dont_regroup` evsels always use their index for
sorting. Make the cur_leader, when fixing the groups, be that of
`dont_regroup` evsel so that the `dont_regroup` evsel doesn't become a
leader.

On a tigerlake this patch corrects this and meets expectations in:
```
$ perf stat -e '{cpu/mem-stores/,cpu/slots/uX}' -a -- sleep 0.1

 Performance counter stats for 'system wide':

        83,458,652      cpu/mem-stores/
     2,720,854,880      cpu/slots/uX

       0.103780587 seconds time elapsed

$ perf stat -e 'slots,slots:X' -a -- sleep 0.1

 Performance counter stats for 'system wide':

       732,042,247      slots                (48.96%)
       643,288,155      slots:X              (51.04%)

       0.102731018 seconds time elapsed
```

Closes: https://lore.kernel.org/lkml/18f20d38-070c-4e17-bc90-cf7102e1e53d@linux.intel.com/
Fixes: 035c17893082 ("perf parse-events: Add 'X' modifier to exclude an event from being regrouped")
Signed-off-by: Ian Rogers <irogers@google.com>
Signed-off-by: Namhyung Kim <namhyung@kernel.org>
2 months agoperf c2c annotate: Start from the contention line
Tianyou Li [Mon, 13 Oct 2025 14:48:11 +0000 (22:48 +0800)] 
perf c2c annotate: Start from the contention line

Add support to highlight the contention line in the annotate browser,
use 'TAB'/'UNTAB' to refocus to the contention line.

Signed-off-by: Tianyou Li <tianyou.li@intel.com>
Reviewed-by: Dapeng Mi <dapeng1.mi@linux.intel.com>
Reviewed-by: Thomas Falcon <thomas.falcon@intel.com>
Reviewed-by: Jiebin Sun <jiebin.sun@intel.com>
Reviewed-by: Pan Deng <pan.deng@intel.com>
Reviewed-by: Zhiguo Zhou <zhiguo.zhou@intel.com>
Reviewed-by: Wangyang Guo <wangyang.guo@intel.com>
Tested-by: Ravi Bangoria <ravi.bangoria@amd.com>
Signed-off-by: Namhyung Kim <namhyung@kernel.org>
2 months agoperf c2c: Add annotation support to perf c2c report
Tianyou Li [Mon, 13 Oct 2025 14:48:10 +0000 (22:48 +0800)] 
perf c2c: Add annotation support to perf c2c report

Perf c2c report currently specified the code address and source:line
information in the cacheline browser, while it is lack of annotation
support like perf report to directly show the disassembly code for
the particular symbol shared that same cacheline. This patches add
a key 'a' binding to the cacheline browser which reuse the annotation
browser to show the disassembly view for easier analysis of cacheline
contentions.

Signed-off-by: Tianyou Li <tianyou.li@intel.com>
Reviewed-by: Dapeng Mi <dapeng1.mi@linux.intel.com>
Reviewed-by: Thomas Falcon <thomas.falcon@intel.com>
Reviewed-by: Jiebin Sun <jiebin.sun@intel.com>
Reviewed-by: Pan Deng <pan.deng@intel.com>
Reviewed-by: Zhiguo Zhou <zhiguo.zhou@intel.com>
Reviewed-by: Wangyang Guo <wangyang.guo@intel.com>
Tested-by: Ravi Bangoria <ravi.bangoria@amd.com>
Signed-off-by: Namhyung Kim <namhyung@kernel.org>
2 months agoperf stat bperf cgroup: Increase MAX_EVENTS from 32 to 1024
Ian Rogers [Thu, 16 Oct 2025 15:07:18 +0000 (08:07 -0700)] 
perf stat bperf cgroup: Increase MAX_EVENTS from 32 to 1024

The MAX_EVENTS value ensured a counted loop presumably to satisfy the
BPF verifier. It is possible to go past 32 events when gathering
uncore events. Increase the amount to 1024 as that should provide some
amount of headroom.

Signed-off-by: Ian Rogers <irogers@google.com>
Signed-off-by: Namhyung Kim <namhyung@kernel.org>
2 months agoperf ilist: Add PMU information to metrics
Ian Rogers [Thu, 16 Oct 2025 22:22:28 +0000 (15:22 -0700)] 
perf ilist: Add PMU information to metrics

Duplicate metrics may exist on hybrid platforms, with the metric's PMU
being used to select the metric to use. Incorporate the metric PMU
into the ilist display and support opening it just for a given PMU.

Before:
```
 ⭘                                    Interactive Perf List
    ├── ▼ TopdownL1                               tma_backend_bound
    │   ├── tma_backend_bound                     Counts the total number of issue slots that were
    │   ├── ▶ tma_backend_bound_group             not consumed by the backend due to backend stalls
    │   ├── tma_backend_bound                     Counts the total number of issue slots that were
    │   ├── ▶ tma_backend_bound_group             not consumed by the backend due to backend stalls.
    │   ├── tma_bad_speculation                   Note that uops must be available for consumption
    │   ├── ▶ tma_bad_speculation_group           in order for this event to count. If a uop is not
    │   ├── tma_bad_speculation                   available (IQ is empty), this event will not count
    │   ├── ▶ tma_bad_speculation_group           cpu_atom@TOPDOWN_BE_BOUND.ALL@ / (5 *
    │   ├── tma_frontend_bound                    cpu_atom@CPU_CLK_UNHALTED.CORE@)
    │   ├── ▶ tma_frontend_bound_group            tma_backend_bound > 0.1
    │   ├── tma_frontend_bound                  ▆▆
    │   ├── ▶ tma_frontend_bound_group
    │   ├── tma_retiring
    │   ├── ▶ tma_retiring_group
    │   ├── tma_retiring
    │   └── ▶ tma_retiring_group
    ├── ▶ TopdownL2
total▄▄▅▅▆▅▅▂▁▁▁▁▂▃▂▂▃▄▄▇▇█▆▆▅▅▅▅▅▅▅▅▅▅▅▅▅▅▅▅▅▄▅▅▅▄▆▆▆▅▅▅▅▅▅▇▇▇▇▆▅▆▆▆▆▅▅▅▄▃▃▃▃▃▃▃▃▃▃▄▄▄▅▅▅▅▅▆▆▆▆▆▆
cpu0▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇
cpu1▁▁▁▁▁▁▁▁▁▁▁▁▁▁▂▂▂▂▂▂▂▂▂▂▂▂▂▂▂▃▃▃▃▃▃▃▃▃▃▄▄▄▄▄▄▄▄▄▄▅▅▅▅▅▅▅▅▅▅▆▆▆▆▆▆▆▆▆▆▆▆▆▆▆▆▆▆▆▆▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇
cpu2▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇█████▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁
cpu3▁▁▁▁▁▁▁▁▁▄▄▄▄▄▄▄▄▄▄█████▆▆▆▆▆▅▅▅▅▅▅▅▅▅▅▅▅▅▅▅▅▅▅▅▅▅▅▅▅▅▅▅▅▅▅▅▅▅▅▅▅▅▅▅▅▅▅▅▅▅▅▅▅▅▅▅▅▅▅▅▅▅▅▅▅▅▅▅▅▅
cpu4████▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▂▂▂▂▂▂▂▂▂▂▂▂▂▂▂▂▂▂▂▂▂▂▂▂▂▂▂▂▂▂▂▂▂▂▂▂▂▂▂▂
cpu5▁▁▁▁▂▂▂▂▂▂▂▂▂▂▂▂▂▂▂▂▂▂▂▂▃▃▃▃▃▃▃▃▃▃▃▃▃▃▃▃▃▃▃▃▄▄▄▄▄▄▄▄▄▄▄▄▄▄▄▄▄▄▄▄▄▄▄▄▄▅▅▅▅▅▅▅▅▅▅▅▅▅▅▅▇▇▇▇▇▇▇▇▇▇▆▆
cpu6▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇█████▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁
cpu7▁▁▁▁▂▂▂▂▂▂▂▂▂▂▂▂▂▂▂▂▂▂▂▂▂▂▂▂▂▂▂▂▂▂▃▃▃▃▃▅▅▅▅▅▅▅▅▅▅▅▅▅▅▅▅▅▅▅▅▅▅▅▅▅▆▆▆▆▆▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇
cpu8▁▁▁▁▂▂▂▂▂▂▂▂▂▂▂▂▂▂▂▃▃▃▃▃▄▄▄▄▄▄▄▄▄▄▄▄▄▄▄▄▄▄▄▄▄▄▄▄▄▄▄▄▄▄▄▄▄▄▄▅▅▅▅▅▆▆▆▆▆▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇
cpu9▁▁▁▁▂▂▂▂▂▂▂▂▂▂▂▂▂▂▂▂▂▂▂▂▂▂▂▂▂▂▂▂▂▂▂▂▂▂▂█████▅▅▅▅▅▅▅▅▅▅▅▅▅▅▅▅▅▅▅▅▅▅▅▅▅▅▅▅▅▅▅▅▅▅▅▅▅▅▅▅▅▅▅▅▅▅▅▅▅▅
cpu10▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇
cpu11▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇█████▁▁▁▁▁▁▁▁▁
cpu12▁▁▁▁▁▁▁▁▁▁▁▁▁▁▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇
cpu13▁▁▁▁▅▅▅▅▅▅▅▅▅▅▅▅▅▅▅▅▅▅▅▅▅▅▅▅▅▅▅▅▅▅▅▅▅▅▅▅▅▅▅▅▅▅▅▅▅▅▅▅▅▅▅▅▅▅▅▅▅▅▅▅▅▅▅▅▅▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇
cpu14▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇█████▂▂▂▂▂▂▂▂▂▂▂▂▂▂▂▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁
cpu15▁▁▁▁▁▁▁▁▁▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇
cpu16▃▃▃▃▃▃▃▃▃▄▄▃▃▃▃▂▂▂▂▂▂▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▄▄▃▃▃▃▃▃▃▃▄▄▄▄▄▄▁▁▃▃▃▃▃▃▃▃▃▃▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇
cpu17▁▁▁▁▁▄▄▅▅▅▅▅▅▅▅▄▄▄▄▄▄▄▄▃▃▃▃▂▂▁▁▂▂▂▂▂▂▂▂▂▂▂▂▂▂▃▃▄▄▄▄▃▃▃▃▂▂▂▂▄▄▄▄▄▄▆▆▆▆▆▆▆▆▆▆▆▆▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇
 s Search  n Next  p Previous  c Collapse  ^q Quit                                      ▏^p palette
```

After:
```
 ⭘                                    Interactive Perf List
    ├── ▼ TopdownL1                               tma_backend_bound
    │   ├── tma_backend_bound (cpu_atom)          Counts the total number of issue slots that were
    │   ├── ▼ tma_backend_bound_group (cpu_atom)  not consumed by the backend due to backend stalls
    │   │   ├── tma_core_bound (cpu_atom)         Counts the total number of issue slots that were
    │   │   ├── ▶ tma_core_bound_group (cpu_atom  not consumed by the backend due to backend stalls.
    │   │   ├── tma_resource_bound (cpu_atom)     Note that uops must be available for consumption
    │   │   └── ▶ tma_resource_bound_group (cpu_  in order for this event to count. If a uop is not
    │   ├── tma_backend_bound (cpu_core)          available (IQ is empty), this event will not count
    │   ├── ▶ tma_backend_bound_group (cpu_core)  cpu_atom@TOPDOWN_BE_BOUND.ALL@ / (5 *
    │   ├── tma_bad_speculation (cpu_atom)        cpu_atom@CPU_CLK_UNHALTED.CORE@)
    │   ├── ▶ tma_bad_speculation_group (cpu_ato▆▆tma_backend_bound > 0.1
    │   ├── tma_bad_speculation (cpu_core)
    │   ├── ▶ tma_bad_speculation_group (cpu_cor▃▃
    │   ├── tma_frontend_bound (cpu_atom)
    │   ├── ▶ tma_frontend_bound_group (cpu_atom
    │   ├── tma_frontend_bound (cpu_core)
    │   ├── ▶ tma_frontend_bound_group (cpu_core
                                           ▌
total▁▁▁▁▂▂▂▂▂▂▂▂▂▃▃▃▃▃▃▃▃▃▃▂▂▂▂▃▃▃▄▄▄▄▄▅▅▅▅▅▅▅▅▅▅▅▅▅▅▅▅▅▅▅▄▄▄▄▄▄▅▅▆▆▆▆▆▆▆▆▆▆▆▆▆▆▆▆▆▆▆▆▆▆▆▆▆▆▆▇▆▇▇
cpu16▇▇▇▇▇▇▇▇▇▇▇▆▆▁▁▁▁▁▁▁▁▁▁▁▁▂▂▄▄▅▅▅▅▅▅▆▆▆▆▆▆▆▆▇▇▇▇▆▆▆▆▆▆▆▆▅▅▅▅▅▅▅▅▅▅▅▅▅▅▆▆▆▆▄▄▄▄▃▃▄▄▄▄▇▇▇▇▇▇▇▇▇▇
cpu17█▇▇▇▇▇▇▇▇▆▆▆▆▆▆▅▅▅▅▃▃▃▃▂▂▁▁▅▅▅▅▅▅▅▅▄▄▄▄▄▄▄▄▄▄▄▄▃▃▃▃▂▂▂▂▂▂▂▂▂▂▂▂▂▂▂▂▂▂▂▂▁▁▁▁▁▁▅▅▄▄▂▂▇▇▇▇▆▆▅▅▆▆
cpu18▇▇▇▇▇██▇▇▃▃▃▃▃▃▃▃▃▃▂▂▂▂▂▂▂▂▃▃▃▃▃▃▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▄▄▄▄▄▄▅▅▅▅▅▅▅▅
cpu19▇▃▃▄▄▄▄▄▄▄▄▄▄▅▅▅▅▅▅▅▅▁▁▂▂▃▃▃▃▅▅▆▆▆▆▆▆▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇██▇▇▇▇▇▇▆▆▅▅▅▅▆▆▄▄▄▄▅▅
cpu20▃▃▃▃▃▃▃▃▃▂▂▂▂▂▂▂▂▂▂▂▂▂▂▂▂▂▂▂▂▂▂▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▂▂▂▂▂▂▂▂▂▂▄▄▄▄▅▅▅▅▅▅▆▆▇▇
cpu21▇▇▇▇▇▇▇██▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▆▆▆▆▆▆▆▆▇▇▇▇▇▇▇▇▇▇▇▇▇▇▆▆▆▆▆▆▆▆▆▆▆▆▆▆▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▅▅▄▄▂▂▂▂▂▂▁▁▁▁
cpu22█▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▆▆▅▅▅▅▅▅▅▅▅▅▅▅▅▅▅▅▅▅▅▅▅▅▅▅▄▄▄▄▄▄▄▄▄▄▄▄▄▄▄▄▄▄▄▄▄▄▄▄▄▄▄▄▄▄▂▂▁▁▁▁▂▂▂▂▂▂▂▂
cpu23▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▃▃▄▄▅▅▅▅▅▅▅▅▅▅▅▅▅▅▅▅▅▅▅▅▅▅▅▅▅▅▅▅▅▅▅▅▅▅▅▅▅▅▅▅▅▅▅▅▅▅▅▅▅▅▅▅▇▇▇▇▆▆██▇▇▇▇▇▇
cpu24▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▂▂▃▃▄▄▄▄▄▄▄▄▄▄▄▄▄▄▄▄▄▄▄▄▄▄▄▄▄▄▄▄▄▄▄▄▄▄▄▄▄▄▄▄▄▄▄▄▄▄▄▄▄▄▄▄▄▄▅▅▇▇▆▆▆▆▆▆▇▇▇▇
cpu25▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▂▂▂▂▂▂▂▂▂▂▂▂▂▂▂▂▂▂▂▂▂▂▂▂▂▂▂▂▂▂▂▂▂▂▂▂▂▂▂▂▂▂▂▂▂▂▂▂▆▆▆▆▇▇▇▇▇▇██
cpu26▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▃▃▄▄▇▇▇▇▇▇▇▇▇▇██▇▇▇▇▇▇▇▇▇▇▆▆▆▆▆▆▆▆▆▆▆▆▇▇▇▇▇▇▇▇▆▆▆▆▆▆▆▆▆▆▂▂▁▁▁▁▂▂▂▂▂▂▂▂
cpu27▂▂▂▂▂▂▂▂▂▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▂▂▂▂▂▂▂▂▂▂▂▂▂▂▂▂▂▂▂▂▂▂▂▂▂▂▂▂▂▂▃▃▄▄▄▄▅▅▅▅▅▅▇▇
total 7.4923074548462605
cpu16 0.2961618003253457
cpu17 0.3065719718925585
cpu18 0.27800656881051855
cpu19 0.28564742078353406
cpu20 0.2764790653117084
 s Search  n Next  p Previous  c Collapse  ^q Quit                                      ▏^p palette
```

Signed-off-by: Ian Rogers <irogers@google.com>
Signed-off-by: Namhyung Kim <namhyung@kernel.org>
2 months agoperf python: Add PMU argument to parse_metrics
Ian Rogers [Thu, 16 Oct 2025 22:22:27 +0000 (15:22 -0700)] 
perf python: Add PMU argument to parse_metrics

Add an optional PMU argument to parse_metrics to allow restriction of
the particular metrics to be opened. If no argument is provided then
all metrics with the given name/group are opened

Signed-off-by: Ian Rogers <irogers@google.com>
Acked-by: Gautam Menghani <gautam@linux.ibm.com>
Signed-off-by: Namhyung Kim <namhyung@kernel.org>
2 months agoperf ilist: Don't display deprecated events
Ian Rogers [Thu, 16 Oct 2025 22:22:26 +0000 (15:22 -0700)] 
perf ilist: Don't display deprecated events

Unsupported legacy events are flagged as deprecated. Don't display
these events in ilist as they won't open and there are over 1,000
legacy cache events.

Signed-off-by: Ian Rogers <irogers@google.com>
Signed-off-by: Namhyung Kim <namhyung@kernel.org>
2 months agoperf trace: Don't synthesize mmaps unless callchains are enabled
Ian Rogers [Sat, 18 Oct 2025 04:59:02 +0000 (21:59 -0700)] 
perf trace: Don't synthesize mmaps unless callchains are enabled

Synthesizing mmaps in perf trace is unnecessary unless call chains are
being generated.

Signed-off-by: Ian Rogers <irogers@google.com>
Acked-by: Howard Chu <howardchu95@gmail.com>
Signed-off-by: Namhyung Kim <namhyung@kernel.org>
2 months agoperf test parse-events: Add evsel test helper
Ian Rogers [Sun, 5 Oct 2025 18:24:30 +0000 (11:24 -0700)] 
perf test parse-events: Add evsel test helper

Add TEST_ASSERT_EVSEL to dump the failing evsel in the event of a
failure.

Signed-off-by: Ian Rogers <irogers@google.com>
Tested-by: James Clark <james.clark@linaro.org>
Signed-off-by: Namhyung Kim <namhyung@kernel.org>
2 months agoperf test parse-events: Add evlist test helper
Ian Rogers [Sun, 5 Oct 2025 18:24:29 +0000 (11:24 -0700)] 
perf test parse-events: Add evlist test helper

Add TEST_ASSERT_EVLIST to dump the failing evlist in the event of a
failure.

Add the macro to a number of tests not currently checking the evlist
length.

Signed-off-by: Ian Rogers <irogers@google.com>
Tested-by: James Clark <james.clark@linaro.org>
Signed-off-by: Namhyung Kim <namhyung@kernel.org>
2 months agoperf test: Clean up test_..config helpers
Ian Rogers [Sun, 5 Oct 2025 18:24:28 +0000 (11:24 -0700)] 
perf test: Clean up test_..config helpers

Just have a single test_hw_config helper that strips extended type
information in the case of hardware and hardware cache events.

Signed-off-by: Ian Rogers <irogers@google.com>
Tested-by: James Clark <james.clark@linaro.org>
Signed-off-by: Namhyung Kim <namhyung@kernel.org>
2 months agoperf test: Switch cycles event to cpu-cycles
Ian Rogers [Sun, 5 Oct 2025 18:24:27 +0000 (11:24 -0700)] 
perf test: Switch cycles event to cpu-cycles

Without a PMU perf matches an event against any PMU with the
event. Unfortunately some PMU drivers advertise a "cycles" event which
is typically just a core event. As tests assume a core event, switch
to use "cpu-cycles" that avoids the overloaded "cycles" event on
troublesome PMUs and is so far not overloaded. Note, on x86 this
changes a legacy event into a sysfs one.

Signed-off-by: Ian Rogers <irogers@google.com>
Tested-by: James Clark <james.clark@linaro.org>
Signed-off-by: Namhyung Kim <namhyung@kernel.org>
2 months agoperf test parse-events: Remove cpu PMU requirement
Ian Rogers [Sun, 5 Oct 2025 18:24:26 +0000 (11:24 -0700)] 
perf test parse-events: Remove cpu PMU requirement

In the event parse string, switch "cpu" to "default_core" and then
rewrite this to the first core PMU name prior to parsing. This enables
testing with a PMU on hybrid x86 and other systems that don't use
"cpu" for the core PMU name. The name "default_core" is already used
by jevents. Update test expectations to match.

Signed-off-by: Ian Rogers <irogers@google.com>
Tested-by: James Clark <james.clark@linaro.org>
Signed-off-by: Namhyung Kim <namhyung@kernel.org>
2 months agoperf test parse-events: Without a PMU use cpu-cycles rather than cycles
Ian Rogers [Sun, 5 Oct 2025 18:24:25 +0000 (11:24 -0700)] 
perf test parse-events: Without a PMU use cpu-cycles rather than cycles

Without a PMU perf matches an event against any PMU with the
event. Unfortunately some PMU drivers advertise a "cycles" event which
is typically just a core event. Switch to using "cpu-cycles" which is
an indentical legacy event but avoids the multiple PMU confusion
introduced by the PMU drivers. Note, on x86 cpu-cycles is also a sysfs
event but cycles isn't.

Signed-off-by: Ian Rogers <irogers@google.com>
Tested-by: James Clark <james.clark@linaro.org>
Signed-off-by: Namhyung Kim <namhyung@kernel.org>
2 months agoperf test parse-events: Use evsel__match for legacy events
Ian Rogers [Sun, 5 Oct 2025 18:24:24 +0000 (11:24 -0700)] 
perf test parse-events: Use evsel__match for legacy events

Switch from the test's assert_hw/test_config to the common
evsel__match code that appropriately handles events with both legacy
and sysfs/json encoding.

For tests asserting that a config value matches that placed in the
perf_event_attr just directly compare the config values.

Signed-off-by: Ian Rogers <irogers@google.com>
Tested-by: James Clark <james.clark@linaro.org>
Signed-off-by: Namhyung Kim <namhyung@kernel.org>
2 months agoperf evsel: Improvements to __evsel__match
Ian Rogers [Sun, 5 Oct 2025 18:24:23 +0000 (11:24 -0700)] 
perf evsel: Improvements to __evsel__match

Ensure both the perf_event_attr and alternate_hw_config are checked in
the match. Don't mask the config if the perf_event_attr isn't a
HARDWARE or HW_CACHE event. Add common early exit cases.

Signed-off-by: Ian Rogers <irogers@google.com>
Tested-by: James Clark <james.clark@linaro.org>
Signed-off-by: Namhyung Kim <namhyung@kernel.org>
2 months agoperf evlist: Avoid scanning all PMUs for evlist__new_default
Ian Rogers [Sun, 5 Oct 2025 18:24:22 +0000 (11:24 -0700)] 
perf evlist: Avoid scanning all PMUs for evlist__new_default

Rather than wildcard matching the cycles event specify only the core
PMUs. This avoids potentially loading unnecessary uncore PMUs.

Signed-off-by: Ian Rogers <irogers@google.com>
Tested-by: James Clark <james.clark@linaro.org>
Signed-off-by: Namhyung Kim <namhyung@kernel.org>
2 months agoperf top: Use evlist__new_default when no events specified
Ian Rogers [Sun, 5 Oct 2025 18:24:21 +0000 (11:24 -0700)] 
perf top: Use evlist__new_default when no events specified

Rather than distributing the code doing similar things to
evlist__new_default, use the one implementation so that paranoia and
wildcard scanning can be optimized.

Signed-off-by: Ian Rogers <irogers@google.com>
Tested-by: James Clark <james.clark@linaro.org>
Signed-off-by: Namhyung Kim <namhyung@kernel.org>
2 months agoperf record: Use evlist__new_default when no events specified
Ian Rogers [Sun, 5 Oct 2025 18:24:20 +0000 (11:24 -0700)] 
perf record: Use evlist__new_default when no events specified

Rather than distributing the code doing similar things to
evlist__new_default, use the one implementation so that paranoia and
wildcard scanning can be optimized.

Signed-off-by: Ian Rogers <irogers@google.com>
Tested-by: James Clark <james.clark@linaro.org>
Signed-off-by: Namhyung Kim <namhyung@kernel.org>
2 months agoperf parse-events: Remove hard coded legacy hardware and cache parsing
Ian Rogers [Sun, 5 Oct 2025 18:24:19 +0000 (11:24 -0700)] 
perf parse-events: Remove hard coded legacy hardware and cache parsing

Now that legacy hardware and cache events are in json, having the
lexer match the specific event is no longer necessary and generic PMU
parsing can take place. Because of this remove the specific term
parsing, event adding, and passing of alternate_hw_config which was
now always PERF_COUNT_HW_MAX.

This mirrors a similar change for software events in commit 6e9fa4131abb
("perf parse-events: Remove non-json software events").

With no hard coded legacy hardware or cache events the wild card, case
insensitivity, etc. is consistent for events. This does, however, mean
events like cycles will wild card against all PMUs. A change does the
same was originally posted and merged from:
https://lore.kernel.org/r/20240416061533.921723-10-irogers@google.com
and reverted by Linus in commit 4f1b067359ac ("Revert "perf
parse-events: Prefer sysfs/JSON hardware events over legacy"") due to
his dislike for the cycles behavior on ARM. Earlier patches in this
series make perf record event opening failures non-fatal and hide the
cycles event's failure to open on ARM in perf record, so it is
expected the behavior will now be transparent in perf record. perf
stat with a cycles event will wildcard open the event on all PMUs. As
cycles is a "default event", the perf stat behavior for default events
was updated to only open them on core/software PMUs.

The change to support legacy events with PMUs was done to clean up
Intel's hybrid PMU implementation.  Having sysfs/json events with
increased priority to legacy was requested by Mark Rutland
<mark.rutland@arm.com> to fix Apple-M PMU issues wrt broken legacy
events on that PMU. It was requested that RISC-V be able to add events
to the perf tool json so the PMU driver didn't need to map legacy
events to config encodings:
https://lore.kernel.org/lkml/20240217005738.3744121-1-atishp@rivosinc.com/

A previous series of patches decreasing legacy hardware event
priorities was posted in:
https://lore.kernel.org/lkml/20250416045117.876775-1-irogers@google.com/
Namhyung Kim <namhyung@kernel.org> mentioned that hardware and
software events can be implemented similarly:
https://lore.kernel.org/lkml/aIJmJns2lopxf3EK@google.com/

Tested-by: Thomas Richter <tmricht@linux.ibm.com>
Signed-off-by: Ian Rogers <irogers@google.com>
Tested-by: James Clark <james.clark@linaro.org>
Signed-off-by: Namhyung Kim <namhyung@kernel.org>
2 months agoperf print-events: Remove print_symbol_events
Ian Rogers [Sun, 5 Oct 2025 18:24:18 +0000 (11:24 -0700)] 
perf print-events: Remove print_symbol_events

Now legacy hardware events are in json there's no need for a specific
printing routine that previously served for both hardware and software
events. The associated event_symbols_hw is also removed. To support
the previous filtered version use an event glob of "legacy hardware"
which matches the topic of the json events.

Tested-by: Thomas Richter <tmricht@linux.ibm.com>
Signed-off-by: Ian Rogers <irogers@google.com>
Tested-by: James Clark <james.clark@linaro.org>
Signed-off-by: Namhyung Kim <namhyung@kernel.org>
2 months agoperf print-events: Remove print_hwcache_events
Ian Rogers [Sun, 5 Oct 2025 18:24:17 +0000 (11:24 -0700)] 
perf print-events: Remove print_hwcache_events

Now legacy cache events are in json there's no need for a specific
printing routine. To support the previous filtered version use an
event glob of "legacy cache" which matches the topic of the json
events.

Tested-by: Thomas Richter <tmricht@linux.ibm.com>
Signed-off-by: Ian Rogers <irogers@google.com>
Tested-by: James Clark <james.clark@linaro.org>
Signed-off-by: Namhyung Kim <namhyung@kernel.org>
2 months agoperf jevents: Add legacy-hardware and legacy-cache json
Ian Rogers [Sun, 5 Oct 2025 18:24:16 +0000 (11:24 -0700)] 
perf jevents: Add legacy-hardware and legacy-cache json

The legacy-hardware.json is added containing hardware events similarly
to the software.json file. A difference is that for the software PMU
the name is known and matches sysfs. In the legacy-hardware.json no
Unit/PMU is specified for the events meaning default_core is used and
the events will appear for all core PMUs.

There are potentially 1216 legacy cache events, rather than list them
in a json file add a make_legacy_cache.py helper to generate them.

By using json for legacy hardware and cache events: descriptions of
the events can be added; events can be marked as deprecated, such as
those misleadingly named l2 (deprecated is also used to mark all
events that weren't previously displayed in perf list); and the name
lookup becomes case insensitive.

The C string encoding all the perf events and metrics is increased in
size by 123,499 bytes which will increase the perf binary size. Later
changes will remove hard coded event parsing for legacy hardware and
cache events, turning parsing overhead into a binary search during
event lookup.

That event descriptions are based off of those in perf_event_open man
page, credit to Vince Weaver <vincent.weaver@maine.edu>.

Tested-by: Thomas Richter <tmricht@linux.ibm.com>
Signed-off-by: Ian Rogers <irogers@google.com>
Tested-by: James Clark <james.clark@linaro.org>
Signed-off-by: Namhyung Kim <namhyung@kernel.org>
2 months agoperf pmu: Add and use legacy_terms in alias information
Ian Rogers [Sun, 5 Oct 2025 18:24:15 +0000 (11:24 -0700)] 
perf pmu: Add and use legacy_terms in alias information

Add support to finding/adding events from the default_core event
table. If an event already exists from sysfs/json then the
default_core configuration is saved in the legacy_terms string. Lazily
use the legacy_terms string to set a legacy hardware or cache event as
deprecated if the core PMU doesn't support it. Use the legacy terms
string to set the alternate_hw_config, avoiding the value needing to
be passed from the parse_events parser.

Tested-by: Thomas Richter <tmricht@linux.ibm.com>
Signed-off-by: Ian Rogers <irogers@google.com>
Tested-by: James Clark <james.clark@linaro.org>
Signed-off-by: Namhyung Kim <namhyung@kernel.org>
2 months agoperf jevents: Add legacy json terms and default_core event table helper
Ian Rogers [Sun, 5 Oct 2025 18:24:14 +0000 (11:24 -0700)] 
perf jevents: Add legacy json terms and default_core event table helper

Add json LegacyConfigCode and LegacyCacheCode values that translate to
legacy-hardware-config and legacy-cache-config event terms
respectively.

Add perf_pmu__default_core_events_table as a means to find a
default_core event table that will later contain legacy events.

In situations like hypervisors it is more likely that tables will be
NULL. Rather than testing in the calling PMU code, early exit in the
pmu-event.c routines.

Tested-by: Thomas Richter <tmricht@linux.ibm.com>
Signed-off-by: Ian Rogers <irogers@google.com>
Tested-by: James Clark <james.clark@linaro.org>
Signed-off-by: Namhyung Kim <namhyung@kernel.org>
2 months agoperf parse-events: Add terms for legacy hardware and cache config values
Ian Rogers [Sun, 5 Oct 2025 18:24:13 +0000 (11:24 -0700)] 
perf parse-events: Add terms for legacy hardware and cache config values

Add the PMU terms legacy-hardware-config and
legacy-cache-config. These terms are similar to the config term in
that their values are assigned to the perf_event_attr config
value. They differ in that the PMU type is switched to be either
PERF_TYPE_HARDWARE or PERF_TYPE_HW_CACHE, and the PMU type is moved
into the extended type information of the config value. This will
allow later patches to add legacy events to json.

An example use of the terms is in the following:
```
$ perf stat -vv -e 'cpu/legacy-hardware-config=1/,cpu/legacy-cache-config=0x10001/' true
Using CPUID GenuineIntel-6-8D-1
Attempt to add: cpu/legacy-hardware-config=0x1/
..after resolving event: cpu/legacy-hardware-config=0x1/
Attempt to add: cpu/legacy-cache-config=0x10001/
..after resolving event: cpu/legacy-cache-config=0x10001/
Control descriptor is not initialized
------------------------------------------------------------
perf_event_attr:
  type                             0 (PERF_TYPE_HARDWARE)
  size                             136
  config                           0x1 (PERF_COUNT_HW_INSTRUCTIONS)
  sample_type                      IDENTIFIER
  read_format                      TOTAL_TIME_ENABLED|TOTAL_TIME_RUNNING
  disabled                         1
  inherit                          1
  enable_on_exec                   1
------------------------------------------------------------
sys_perf_event_open: pid 994937  cpu -1  group_fd -1  flags 0x8 = 3
------------------------------------------------------------
perf_event_attr:
  type                             3 (PERF_TYPE_HW_CACHE)
  size                             136
  config                           0x10001 (PERF_COUNT_HW_CACHE_RESULT_MISS | PERF_COUNT_HW_CACHE_OP_READ | PERF_COUNT_HW_CACHE_L1I)
  sample_type                      IDENTIFIER
  read_format                      TOTAL_TIME_ENABLED|TOTAL_TIME_RUNNING
  disabled                         1
  inherit                          1
  enable_on_exec                   1
------------------------------------------------------------
sys_perf_event_open: pid 994937  cpu -1  group_fd -1  flags 0x8 = 4
cpu/legacy-hardware-config=1/: -1: 1364046 414756 414756
cpu/legacy-cache-config=0x10001/: -1: 57453 414756 414756
cpu/legacy-hardware-config=1/: 1364046 414756 414756
cpu/legacy-cache-config=0x10001/: 57453 414756 414756

 Performance counter stats for 'true':

         1,364,046      cpu/legacy-hardware-config=1/
            57,453      cpu/legacy-cache-config=0x10001/

       0.001988593 seconds time elapsed

       0.002194000 seconds user
       0.000000000 seconds sys
```

Tested-by: Thomas Richter <tmricht@linux.ibm.com>
Signed-off-by: Ian Rogers <irogers@google.com>
Tested-by: James Clark <james.clark@linaro.org>
Signed-off-by: Namhyung Kim <namhyung@kernel.org>
2 months agoperf pmu: Factor term parsing into a perf_event_attr into a helper
Ian Rogers [Sun, 5 Oct 2025 18:24:12 +0000 (11:24 -0700)] 
perf pmu: Factor term parsing into a perf_event_attr into a helper

Factor existing functionality in perf_pmu__name_from_config into a
helper that will be used in later patches.

Tested-by: Thomas Richter <tmricht@linux.ibm.com>
Signed-off-by: Ian Rogers <irogers@google.com>
Tested-by: James Clark <james.clark@linaro.org>
Signed-off-by: Namhyung Kim <namhyung@kernel.org>
2 months agoperf pmu: Use fd rather than FILE from new_alias
Ian Rogers [Sun, 5 Oct 2025 18:24:11 +0000 (11:24 -0700)] 
perf pmu: Use fd rather than FILE from new_alias

The FILE argument was necessary for the scanner but now that
functionality is not being used we can switch to just using
io__getline which should cut down on stdio buffer usage.

Tested-by: Thomas Richter <tmricht@linux.ibm.com>
Signed-off-by: Ian Rogers <irogers@google.com>
Tested-by: James Clark <james.clark@linaro.org>
Signed-off-by: Namhyung Kim <namhyung@kernel.org>
2 months agoperf parse-events: Remove unused FILE input argument to scanner
Ian Rogers [Sun, 5 Oct 2025 18:24:10 +0000 (11:24 -0700)] 
perf parse-events: Remove unused FILE input argument to scanner

Now the events file isn't directly parsed from a FILE but stored in a
string prior to parsing, remove the FILE argument to the associated
scanner functions as they only ever pass NULL.

Tested-by: Thomas Richter <tmricht@linux.ibm.com>
Signed-off-by: Ian Rogers <irogers@google.com>
Tested-by: James Clark <james.clark@linaro.org>
Signed-off-by: Namhyung Kim <namhyung@kernel.org>
2 months agoperf pmu: Don't eagerly parse event terms
Ian Rogers [Sun, 5 Oct 2025 18:24:09 +0000 (11:24 -0700)] 
perf pmu: Don't eagerly parse event terms

When an event/alias is created for a PMU the terms are eagerly parsed
using parse_events_terms. For a command like perf stat or perf record,
the particular event/alias will be found, the terms parsed, the
terms cloned for use in the event parsing, and then the terms used to
configure the perf_event_attr. Events/aliases may be eagerly loaded,
such as from sysfs or in perf list, in which case the aliases terms
will be little or never used. To avoid redundant work, to avoid
cloning, and to reduce memory overhead, hold the terms for an event as
a string until they need handling as a term list. This may introduce
duplicate parsing if an event is repeated in a list, but this
situation is expected to be uncommon.

Measuring the number of instructions before and after with a sysfs
event and perf stat, there is a minor reduction in the number of
instructions executed by 0.3%.

Tested-by: Thomas Richter <tmricht@linux.ibm.com>
Signed-off-by: Ian Rogers <irogers@google.com>
Tested-by: James Clark <james.clark@linaro.org>
Signed-off-by: Namhyung Kim <namhyung@kernel.org>
2 months agoperf jevents: Support copying the source json files to OUTPUT
Ian Rogers [Sun, 5 Oct 2025 18:24:08 +0000 (11:24 -0700)] 
perf jevents: Support copying the source json files to OUTPUT

The jevents command expects all json files to be organized under a
single directory. When generating json files from scripts (to reduce
laborious copy and paste in the json) we don't want to generate the
json into the source directory if there is an OUTPUT directory
specified. This change adds a GEN_JSON for this case where the
GEN_JSON copies the JSON files to OUTPUT, only when OUTPUT is
specified. The Makefile.perf clean code is updated to clean up this
directory when present.

This patch is part of:
https://lore.kernel.org/lkml/20240926173554.404411-12-irogers@google.com/
which was similarly adding support for generating json in scripts for
the consumption of jevents.py.

Tested-by: Thomas Richter <tmricht@linux.ibm.com>
Signed-off-by: Ian Rogers <irogers@google.com>
Tested-by: James Clark <james.clark@linaro.org>
Signed-off-by: Namhyung Kim <namhyung@kernel.org>
2 months agoperf record: Skip don't fail for events that don't open
Ian Rogers [Sun, 5 Oct 2025 18:24:07 +0000 (11:24 -0700)] 
perf record: Skip don't fail for events that don't open

Whilst for many tools it is an expected behavior that failure to open
a perf event is a failure, ARM decided to name PMU events the same as
legacy events and then failed to rename such events on a server uncore
SLC PMU. As perf's default behavior when no PMU is specified is to
open the event on all PMUs that advertise/"have" the event, this
yielded failures when trying to make the priority of legacy and
sysfs/json events uniform - something requested by RISC-V and ARM. A
legacy event user on ARM hardware may find their event opened on an
uncore PMU which for perf record will fail. Arnaldo suggested skipping
such events which this patch implements. Rather than have the skipping
conditional on running on ARM, the skipping is done on all
architectures as such a fundamental behavioral difference could lead
to problems with tools built/depending on perf.

An example of perf record failing to open events on x86 is:
```
$ perf record -e data_read,cycles,LLC-prefetch-read -a sleep 0.1
Error:
Failure to open event 'data_read' on PMU 'uncore_imc_free_running_0' which will be removed.
The sys_perf_event_open() syscall returned with 22 (Invalid argument) for event (data_read).
"dmesg | grep -i perf" may provide additional information.

Error:
Failure to open event 'data_read' on PMU 'uncore_imc_free_running_1' which will be removed.
The sys_perf_event_open() syscall returned with 22 (Invalid argument) for event (data_read).
"dmesg | grep -i perf" may provide additional information.

Error:
Failure to open event 'LLC-prefetch-read' on PMU 'cpu' which will be removed.
The LLC-prefetch-read event is not supported.
[ perf record: Woken up 1 times to write data ]
[ perf record: Captured and wrote 2.188 MB perf.data (87 samples) ]

$ perf report --stats
Aggregated stats:
               TOTAL events:      17255
                MMAP events:        284  ( 1.6%)
                COMM events:       1961  (11.4%)
                EXIT events:          1  ( 0.0%)
                FORK events:       1960  (11.4%)
              SAMPLE events:         87  ( 0.5%)
               MMAP2 events:      12836  (74.4%)
             KSYMBOL events:         83  ( 0.5%)
           BPF_EVENT events:         36  ( 0.2%)
      FINISHED_ROUND events:          2  ( 0.0%)
            ID_INDEX events:          1  ( 0.0%)
          THREAD_MAP events:          1  ( 0.0%)
             CPU_MAP events:          1  ( 0.0%)
           TIME_CONV events:          1  ( 0.0%)
       FINISHED_INIT events:          1  ( 0.0%)
cycles stats:
              SAMPLE events:         87
```

If all events fail to open then the perf record will fail:
```
$ perf record -e LLC-prefetch-read true
Error:
Failure to open event 'LLC-prefetch-read' on PMU 'cpu' which will be removed.
The LLC-prefetch-read event is not supported.
Error:
Failure to open any events for recording
```

As an evlist may have dummy events that open when all command line
events fail we ignore dummy events when detecting if at least some
events open. This still permits the dummy event on its own to be used
as a permission check:
```
$ perf record -e dummy true
[ perf record: Woken up 1 times to write data ]
[ perf record: Captured and wrote 0.046 MB perf.data ]
```
but allows failure when a dummy event is implicilty inserted or when
there are insufficient permissions to open it:
```
$ perf record -e LLC-prefetch-read -a true
Error:
Failure to open event 'LLC-prefetch-read' on PMU 'cpu' which will be removed.
The LLC-prefetch-read event is not supported.
Error:
Failure to open any events for recording
```

As the first parsed event in an evlist is marked as tracking, removing
this event can remove tracking from the evlist, removing mmap events
and breaking symbolization. To avoid this, if a tracking event is
removed then the next event has tracking added.

The issue with legacy events is that on RISC-V they want the driver to
not have mappings from legacy to non-legacy config encodings for each
vendor/model due to size, complexity and difficulty to update. It was
reported that on ARM Apple-M? CPUs the legacy mapping in the driver
was broken and the sysfs/json events should always take precedent,
however, it isn't clear this is still the case. It is the case that
without working around this issue a legacy event like cycles without a
PMU can encode differently than when specified with a PMU - the
non-PMU version favoring legacy encodings, the PMU one avoiding legacy
encodings. Legacy events are also case sensitive while sysfs/json
events are not.

The patch removes events and then adjusts the idx value for each
evsel. This is done so that the dense xyarrays used for file
descriptors, etc. don't contain broken entries.

On ARM it could be common following this change to see a lot of
warnings for the cycles event due to many ARM PMUs advertising the
cycles event (ARM inconsistently have events bus_cycles and then
cycles implying CPU cycles, they also sometimes have a cpu_cycles
event). As cycles is a popular event, avoid potentially spamming users
with error messages on ARM when there are multiple cycles events in
the evlist, the error is still shown when verbose is enabled.

Prior versions without adding the tracking data and not warning for
cycles on ARM was:

Suggested-by: Arnaldo Carvalho de Melo <acme@kernel.org>
Tested-by: James Clark <james.clark@linaro.org>
Tested-by: Leo Yan <leo.yan@arm.com>
Tested-by: Atish Patra <atishp@rivosinc.com>
Tested-by: Thomas Richter <tmricht@linux.ibm.com>
Signed-off-by: Ian Rogers <irogers@google.com>
Signed-off-by: Namhyung Kim <namhyung@kernel.org>
2 months agoperf stat: Avoid wildcarding PMUs for default events
Ian Rogers [Sun, 5 Oct 2025 18:24:06 +0000 (11:24 -0700)] 
perf stat: Avoid wildcarding PMUs for default events

Without a PMU perf matches an event against any PMU with the
event. Unfortunately some PMU drivers advertise a "cycles" event which
is typically just a core event. To make perf's behavior consistent,
just look up default events with their designated PMU types.

Signed-off-by: Ian Rogers <irogers@google.com>
Tested-by: James Clark <james.clark@linaro.org>
Signed-off-by: Namhyung Kim <namhyung@kernel.org>
2 months agoperf perf_api_probe: Avoid scanning all PMUs, try software PMU first
Ian Rogers [Sun, 5 Oct 2025 18:24:05 +0000 (11:24 -0700)] 
perf perf_api_probe: Avoid scanning all PMUs, try software PMU first

Scan the software PMU first rather than last as it is the least likely
to fail the probe. Specifying the software PMU by name was enabled by
commit 9957d8c801fe ("perf jevents: Add common software event
json"). For hardware events, add core PMU names when getting events to
probe so that not all PMUs are scanned. For example, when legacy
events support wildcards and for the event "cycles:u" on x86, we want
to only scan the "cpu" PMU and not all uncore PMUs for the event too.

Tested-by: Thomas Richter <tmricht@linux.ibm.com>
Signed-off-by: Ian Rogers <irogers@google.com>
Tested-by: James Clark <james.clark@linaro.org>
Signed-off-by: Namhyung Kim <namhyung@kernel.org>
2 months agoperf parse-events: Fix legacy cache events if event is duplicated in a PMU
Ian Rogers [Sun, 5 Oct 2025 18:24:04 +0000 (11:24 -0700)] 
perf parse-events: Fix legacy cache events if event is duplicated in a PMU

The term list when adding an event to a PMU is expected to have the
event name for the alias lookup. Also, set found_supported so that
-EINVAL isn't returned.

Fixes: 62593394f66a ("perf parse-events: Legacy cache names on all
PMUs and lower priority")

Tested-by: Thomas Richter <tmricht@linux.ibm.com>
Signed-off-by: Ian Rogers <irogers@google.com>
Tested-by: James Clark <james.clark@linaro.org>
Signed-off-by: Namhyung Kim <namhyung@kernel.org>
2 months agoperf bpf_counter: Fix opening of "any"(-1) CPU events
Ian Rogers [Thu, 9 Oct 2025 13:29:11 +0000 (06:29 -0700)] 
perf bpf_counter: Fix opening of "any"(-1) CPU events

The bperf BPF counter code doesn't handle "any"(-1) CPU events, always
wanting to aggregate a count against a CPU, which avoids the need for
atomics so let's not change that. Force evsels used for BPF counters
to require a CPU when not in system-wide mode so that the "any"(-1)
value isn't used during map propagation and evsel's CPU map matches
that of the PMU.

Fixes: b91917c0c6fa ("perf bpf_counter: Fix handling of cpumap fixing hybrid")
Signed-off-by: Ian Rogers <irogers@google.com>
Signed-off-by: Namhyung Kim <namhyung@kernel.org>
2 months agoperf build python: Don't leave a.out file when building with clang
Ian Rogers [Tue, 7 Oct 2025 15:38:35 +0000 (08:38 -0700)] 
perf build python: Don't leave a.out file when building with clang

Testing clang features doesn't specify a "-o" option so an a.out file
is created and left in the make directory (not the output). Fix this
by specifying a "-o" of "/dev/null". Reorganize the code a little to
help with readability.

Signed-off-by: Ian Rogers <irogers@google.com>
Tested-by: Justin Stitt <justinstitt@google.com>
Signed-off-by: Namhyung Kim <namhyung@kernel.org>
2 months agoperf stat: Additional verbose details for <not supported> events
Ian Rogers [Sun, 5 Oct 2025 18:14:21 +0000 (11:14 -0700)] 
perf stat: Additional verbose details for <not supported> events

If an event shows as "<not supported>" in perf stat output, in verbose
mode add the strerror output to help diagnose the issue.

Consider:
```
$ perf stat -e cycles,data_read,instructions true

 Performance counter stats for 'true':

           357,457      cycles:u
   <not supported> MiB  data_read:u
           156,182      instructions:u                   #    0.44  insn per cycle

       0.001250315 seconds time elapsed

       0.001283000 seconds user
       0.000000000 seconds sys
```

To understand why the data_read uncore event failed you can run it
again with -v option.  This change adds detailed message about the
error and suggestion how to fix it potentially.

  Warning:
  data_read:u event is not supported by the kernel.
  Invalid event (data_read:u) in per-thread mode, enable system wide with '-a'.

Signed-off-by: Ian Rogers <irogers@google.com>
[ simplified the commit message ]
Signed-off-by: Namhyung Kim <namhyung@kernel.org>