Re: [PATCH] lib: fix compare_delta parameter order in percpu_counter_tree

[David CARLIER]

On Mon, 16 Mar 2026 at 13:53, Mathieu Desnoyers
<mathieu.desnoyers@xxxxxxxxxxxx> wrote:
>
> On 2026-03-16 09:41, David CARLIER wrote:
> > Sure. For precise_compare_value, the boundary tests extend
> >    naturally — the function must return the exact result regardless
> >    of which code path (fast approximate or precise fallback) is taken.
> >
> >    For the two-counter APIs, I noticed that the combined accuracy
> >    is always symmetric: over_a + under_b = under_a + over_b =
> >    (bs_a + bs_b - 1) * multiplier, so the asymmetric gap test
> >    doesn't apply there. But boundary tests at the combined accuracy
> >    limit still provide useful coverage.
>
> In order to do a relevant test for the two-counter APIs, we'll need
> to initialize each of the two counters with different batch sizes.
> This will ensure that the limits are asymmetric, and therefore allow
> more precise testing of the limit conditions.
>
> Thanks,
>
> Mathieu

I looked into this but since accuracy_multiplier is a single global
  constant, the combined accuracy ends up symmetric regardless of
  batch sizes. For the two-counter comparison:

    accuracy_pos = over_a + under_b
    accuracy_neg = under_a + over_b

  Since under = batch_size * M and over = (batch_size - 1) * M,
  both sides always simplify to the same value. So the asymmetric
  gap test only applies to the single-value comparison APIs.

  Should I keep the two-counter test with symmetric boundary
  coverage, or did you have something else in mind?

  Thanks
>
> --
> Mathieu Desnoyers
> EfficiOS Inc.
> https://www.efficios.com