l3_misses and l3_accesses perf events unable to retrieve on AMD host

[yunzhao-cf]

Hello,

I am encountering an issue where cAdvisor's perf is unable to retrieve the perf events "l3_misses" and "l3_accesses". The following error message is returned, suggesting a failure in the pfm_get_os_event_encoding() call:

E0305 17:49:03.940537 892708 collector_libpfm.go:203] Cannot count perf event group [l3_misses l3_accesses]: cannot create config from perf event: unable to transform event name l3_misses to perf_event_attr: -4

Checking the libpfm and '-4' is for "event not found": https://github.com/s-kanev/misc/blob/fc38c60e84d20f5aceb8ed0c70478b45b92f48b1/perf_harness/libpfm/perfmon/pfmlib.h#L359

However, these events can be successfully monitored using the perf CLI tool:

 ~$sudo perf stat -a -e l3_misses,l3_accesses -- sleep 1

 Performance counter stats for 'system wide':

     1,122,395,814      l3_misses                                                          
     4,398,985,531      l3_accesses                                                        

       1.005054759 seconds time elapsed

I attempted to manually specify the event configuration based on the output from perf:

 ~$sudo perf stat -I 5000 -vvv -e l3_misses
Using CPUID AuthenticAMD-23-31-0
Attempting to add event pmu 'amd_l3' with 'l3_misses,' that may result in non-fatal errors
After aliases, add event pmu 'amd_l3' with 'event,umask,' that may result in non-fatal errors
l3_misses -> amd_l3/event=0x4,umask=0x1/
Control descriptor is not initialized
------------------------------------------------------------
perf_event_attr:
  type                             11
  size                             128
  config                          0x104
  sample_type               IDENTIFIER
  read_format                TOTAL_TIME_ENABLED|TOTAL_TIME_RUNNING
  disabled                      1
  inherit                          1
  exclude_guest            1

I then defined this configuration in a custom events file for cAdvisor:

    "custom_events": [
      {
        "type": 11,
        "config": [
          "0x104"
        ],
        "name": "l3_misses"
      }
    ]

However, no data is generated when using this custom configuration, and no error message is reported.

The environment is running on a kernel version 6.1 with an AMD EPYC 7642 48-Core Processor.

[yunzhao-cf]

Verified with a simple libpfm4 code, and it shows these perf events are not supported yet on AMD nodes: https://sourceforge.net/p/perfmon2/mailman/perfmon2-devel/thread/CAAnvDSxwyyguSk56y87mu2f1OS4J0NQdiaTDsavUEkRWGHvL9Q%40mail.gmail.com/#msg58747559