Improve counter reporting on heterogeneous Arm systems

[kieranhejmadi01]

Summary

Running sysreport.py on a NVIDIA DGX Spark (heterogeneous arm64 system with 10 X725 and 10 X925 cores) I believe I sysreport outputs incorrect perf counter data.

I was able to confirm hardware counters with

perf stat -x, -e "{instructions:u}" -- true

which outputted:

WARNING: events were regrouped to match PMUs
<not counted>,,armv8_pmuv3_0/instructions/u,0,0.00,,
134652,,armv8_pmuv3_1/instructions/u,212064,100.00,,

However, when running sysreport on the DGX shows None for perf counters:

System feature report:
  Collected:           2026-04-09 11:11:35.108513
  Script version:      2026-04-09 11:11:29.368312
  Running as root:     False
System hardware:
  Architecture:        aarch64
  CPUs:                20
  CPU types:           10 x Arm Part 0xd87 r0p1, 10 x Arm Part 0xd85 r0p1
  cache info:          size, associativity, sharing
  cache line size:     64
  Caches:
    20 x L1D 64K 4-way 64b-line
    20 x L1I 64K 4-way 64b-line
    10 x L2U 2M 8-way 64b-line
    10 x L2U 512K 8-way 64b-line
    1 x L3U 16M 16-way 64b-line
    1 x L3U 8M 16-way 64b-line
  System memory:       120G
  Atomic operations:   True
  interconnect:        unknown x 1
  NUMA nodes:          1
  Sockets:             1
OS configuration:
  Kernel:              6.14.0
  config:              /boot/config-6.14.0-1013-nvidia
  build dir:           /lib/modules/6.14.0-1013-nvidia/build
  uses atomics:        True
  page size:           4K
  huge pages:          2048kB: 0, 32768kB: 0, 64kB: 0, 1048576kB: 0
  transparent HP:      madvise
  MPAM configured:     False
  resctrl:             False
  Distribution:        Ubuntu 24.04.3 LTS
  libc version:        glibc 2.39
  boot info:           ACPI
  KPTI enforced:       False
  Lockdown:            landlock, lockdown, yama, integrity, apparmor
  Mitigations:         spectre_v2:CSV2, BHB; spec_store_bypass:Speculative Store Bypass disabled via prctl; spectre_v1:__user pointer sanitization
Performance features:
  perf tools:          True
  perf installed at:   /usr/lib/linux-tools/6.14.0-1013-nvidia/perf
  perf with OpenCSD:   False
  perf counters:       None
  perf sampling:       SPE
  perf HW trace:       None
  perf paranoid:       0
  CAP_PERFMON:         disabled
  kptr_restrict:       1
  perf in userspace:   disabled
  interconnect perf:   None
  /proc/kcore:         True
  /dev/mem:            True
  eBPF:
    kernel configured for BPF: True
    bpftool installed:         True
      bpftool v7.6.0 using libbpf v1.6 features: 
    bpftrace installed:        bpftrace v0.20.2


Actions that can be taken to improve performance tools experience:
  perf tools cannot decode hardware trace
    build with CORESIGHT=1
  Hardware perf events are not available
    ensure APIC table describes PMU interrupt
  hardware trace not enabled
    ensure ACPI describes CoreSight trace fabric

Root Cause

I believe the system redistributes events across PMUs rather than scheduling them as a single group. This means the group saturation point - which is how we detect the counter limit is never reached, so detection returns None even though counters are available.

Improvement

Simply print out a more informative info about the perf counters if it returns None and the system is heterogeneous