Race condition in map loop for BPF_MAP_TYPE_PERCPU_HASH

[jordalgo]

BEGIN {
  $i = 0;
  while ($i < 100) {
    @m[1] += 1;
    $i++;
    if ($i == 100) {
      exit();
    }
  }
}

END {
  for ($kv : @m) {
    print(($kv));
  }
  clear(@m);
}

This script seems to always exit with the correct output, namely: (1, 100).

However if you use the count() function instead of +=1, which is functionally equivalent, I often see this as the output: (1,0).

My guess is that because the map used for count is BPF_MAP_TYPE_PERCPU_HASH vs BPF_MAP_TYPE_HASH used for the += operation, there is some issue there but that's as far as I got.

[ajor]

It looks like the problem occurs when the BEGIN and END probes run on different CPUs:

BEGIN {
  $i = 0;
  while ($i < 100) {
    @m[cpu] = count();
    $i++;
  }
}

END {
  for ($kv : @m) {
    print((cpu, $kv));
  }
  clear(@m);
}

Same CPU, correct result:

(15, (15, 100))

Different CPU, wrong result:

(57, (59, 0))

[ajor]

Maybe #3126 is the solution for this bug as well? When reading a per-cpu map value, collect data from all CPUs.

[jordalgo]

Maybe #3126 is the solution for this bug as well? When reading a per-cpu map value, collect data from all CPUs.

Yeah that makes sense. I'll try to pick up that task soon (if no one beats me to it).