On Fri, 2 Oct 2026 14:52:34 +0200
Roman Khromenok <[email protected]> wrote:
> This follows up on the review of the SFF-8472 Rx power calibration fix,
> where the automated review pointed out the remaining issues in
> sff_8472_calibration().
>
> Calibration formula 1 (temperature, voltage, Tx bias, Tx power) scales
> the 16-bit fields in place by a slope of up to 255.996 and adds a signed
> 16-bit offset. A result out of the field range is undefined behavior
> for the float to integer conversion; a negative sum wraps around.
> UBSan reports float-cast-overflow on randomly generated externally
> calibrated SFP dumps.
>
> Patch 1 computes formula 1 in double precision and saturates the result
> to the field range, as already done for the Rx power. It is intended
> for stable.
> Patch 2 rounds the calibrated values to the nearest integer instead
> of truncating them: RX_PWR(1) = 0.7 is stored as 0.69999999 and
> a raw reading of 1000 gave 699 instead of 700.
> Patch 3 adds unit tests for saturation at both ends and for rounding.
>
> Based on dpdk-next-net for-main.
>
> Depends-on: series-39455 ("ethdev: fix SFF-8472 external Rx power
> calibration")
Applied to next-net
