Re: [PATCH v2 2/3] iio: adc: ltc2497: add LTC2499 internal temperature channel

[Andy Shevchenko]

On Thu, Aug 13, 2026 at 7:02 PM Andrei Stancovici
<andrei.stancovici@xxxxxxxxxx> wrote:
>
> The LTC2499 has an internal PTAT (proportional to absolute temperature)
> sensor that is activated by a second I2C configuration byte (EN2 | IM).
> Expose it as an IIO_TEMP channel providing raw, scale and offset so the
> standard IIO formula
>
>         T[m°C] = (raw + offset) * scale
>
> reconstructs the temperature.
>
> The PTAT sensor yields the absolute temperature as
>
>         T(K) = DATAOUT24 * Vref / 1570          (Vref in volts)
>
> The raw value exported here is sign-extended and normalised to
> 2^(resolution + 1) == 2^25, i.e. raw = 2 * DATAOUT24, so on the IIO
> milli-degree-Celsius convention
>
>         scale[m°C/LSB] = Vref_uV / 3140000
>         offset         = -273150 * 3140000 / Vref_uV
>
> The scale and offset are derived from the reference voltage returned by
> regulator_get_voltage(); its error is propagated as before, so a board
> that fails to describe vref-supply gets a clear read error instead of a
> silently wrong temperature. No board-specific reference value is assumed
> in the driver.
>
> The single temperature channel is appended as the last entry of the
> shared channel array and excluded via num_channels for parts without an
> internal sensor, so the existing LTC2497 channel layout and device name
> are unchanged.
>
> The LTC2499 latches its converter configuration from the second command
> byte and only re-evaluates it when that byte has EN2 set. EN2 | IM
> selects the internal temperature sensor. Because a single-byte command,
> or a second byte with EN2 = 0, means "keep previous", a one-byte channel
> select cannot pull the device back out of temperature mode: after a
> temperature read every subsequent voltage read would keep returning the
> PTAT result instead of the selected input. Temperature support is
> therefore only correct if the voltage path also emits a second command
> byte that re-selects an external input.
>
> Send two-byte commands for all conversions on parts that have the sensor
> (has_temp):
>
>         temperature:    EN2 | IM
>         voltage:        EN2                     (IM = 0 -> external input)
>
> The LTC2497 and LTC2496, which lack the second-byte mechanism, keep using
> the original single-byte channel select and are unchanged.

...

> +                       /*
> +                        * raw is normalised to 2^(resolution + 1), i.e.
> +                        * raw = 2 * DATAOUT24, so the PTAT scale (datasheet
> +                        * Vref / 1570 per kelvin) doubles its denominator and,

Kelvin

> +                        * in m°C, becomes Vref_uV / 3140000.
> +                        */

...

> +                       ret = regulator_get_voltage(ddata->ref);
> +                       if (ret < 0)
> +                               return ret;
> +                       if (!ret)

In this case I would rather see

  if (ret == 0)

as it's about value and not a returned error code (success).

> +                               return -EINVAL;
> +                       /*
> +                        * 0 °C == 273.15 K must map to raw + offset such that
> +                        * (raw + offset) * scale == 0 m°C, i.e.
> +                        *   offset = ABSOLUTE_ZERO_MILLICELSIUS / scale
> +                        *          = ABSOLUTE_ZERO_MILLICELSIUS * 3140000 / Vref_uV

This is a bit too far. In the comments the plain numbers are much
better, we should understand math and physics behind.

> +                        * Computed in 64-bit to avoid overflow.
> +                        */
> +                       *val = div_s64((s64)ABSOLUTE_ZERO_MILLICELSIUS * 3140000, ret);

Casting is not needed, the trick is to add LL type suffix to the 3140000.

> +                       return IIO_VAL_INT;
> +               default:
> +                       return -EINVAL;
> +               }

...

>  struct ltc2497_chip_info {
>         u32 resolution;
>         const char *name;
> +       bool has_temp;
>  };

Have you run `pahole`? It might be that on some architectures there
will be a gap in between existing ones due to 64-bit required
alignment.

-- 
With Best Regards,
Andy Shevchenko