As pointed out in #458, the current off-cpu per-mille threshold might be too gritty in some situations.

Suggestion

Replace the per-mille -off-cpu-threshold value float probability value in the range [0..1]. 0 means that off-cpu profiling is not enabled. 1 means that all off-cpu events are processed. For example, 0.1 means that ~ every 10th event is processed.

In off_cpu.ebpf.c we also get rid of the relative expensive division.

  if (bpf_get_prandom_u32() % OFF_CPU_THRESHOLD_MAX >= syscfg->off_cpu_threshold) {
    return 0;
  }

would become

  if (bpf_get_prandom_u32() <= syscfg->off_cpu_threshold) {
    return 0;
  }

where off_cpu_threshold is the value of uint32(p * float64(math.MaxUint32)).