The `perf stat --bpf-counters test` fails intermittently on hybrid
architectures or systems with dynamic frequency scaling (DVFS). This
happens because the test workload (`sqrtloop`) runs for a fixed 1-second
duration, and the CPU frequency can scale dynamically between idle and
maximum frequency. As the first run runs on a cold CPU and the second run
runs on a warmed-up CPU (or vice versa), the number of instructions
executed in 1 second differs by up to 2.2x, violating the comparison
tolerance.
Also, when running as root, BPF tracepoints and scheduling programs
trigger frequently. Since standard `perf stat -e instructions` measures
both user and kernel space instructions, it counts BPF helper and program
execution overheads, whereas the BPF counters themselves do not self-
measure. This introduces a large kernel-space instruction count
discrepancy between standard and BPF counters.
Fix these issues by:
1. Switching the workload to a strictly deterministic, iteration-based
workload: `awk 'BEGIN { for (i=0; i<10000000; i++) sum+=i }'`. We pin
the
workload to a single random allowed CPU using `taskset -c $CPU` via a
bash array.
2. Restricting the counted event to user-space only (`instructions:u` or
`/u`).
3. Tightening the comparison tolerance from 20% to 15%.
These modifications isolate the measurements to user-space instructions of
the deterministic loop, which executes a virtually identical number of
instructions on both runs (with less than 0.001% variation), eliminating
Dynamic Frequency Scaling (DVFS), kernel scheduling noise, and BPF helper
self-measurement overheads.
Fixes: 2c0cb9f560 ("perf test: Add a shell test for 'perf stat --bpf-counters' new option")
Assisted-by: Antigravity:gemini-3.1-pro
Signed-off-by: Ian Rogers <irogers@google.com>
Signed-off-by: Namhyung Kim <namhyung@kernel.org>
b02027776a