Re: [PATCH 2/2] thermal/drivers/tegra: Avoid SOC_THERM calculation overflow on Tegra114

[Svyatoslav Ryhel]

пт, 6 лист. 2026 р. о 22:32 Patrick Fischer via B4 Relay
<devnull+admin.tsv-stahl-riesa.de@xxxxxxxxxx> пише:
>
> From: Patrick Fischer <admin@xxxxxxxxxxxxxxxxxx>
>
> SOC_THERM calculates each sensor temperature in units of 0.5 C as
> THERMA * capture / 2^13 + THERMB. On Tegra114 the intermediate product
> THERMA * capture has to stay below 2^24. Above that the hardware sets
> TCALC_OVER in SENSOR_CONFIG0, clears TEMP_VALID in SENSOR_STATUS1 and
> the sensor drops out of its group. On a Tegra Note 7, 80 out of 80
> samples matched this rule: TEMP_VALID was set exactly when
> THERMA * capture < 2^24 (largest valid product 16747731, smallest
> invalid one 16877718).
>
> The fuse calibration itself is fine. As Mikko pointed out in the review
> of the Tegra114 series, the fuse corrections are in 0.5 C units on
> Tegra114 just as on Tegra124 and Tegra210, and
> 2 * NOMINAL_CALIB_XX + shifted_xx is correct. This patch keeps it that
> way. The problem is the range: with T in C,
> THERMA * capture = (2 * T - THERMB) * 2^13, so a sensor only gives a
> valid result for T < (2048 + THERMB) / 2. On Tegra114 THERMB is close
> to -2000 (between -2117 and -1812 on the chip tested), which leaves
> almost no headroom. On that chip the sensors were valid only up to
> 45 C (cpu2), 53 C (pllx) and 32 C (gpu), and cpu0, cpu1 and cpu3 not
> even at room temperature. The gpu zone therefore always reads 0 C, the
> cpu zone stops following the CPU at about 45 C, and neither the
> throttle levels nor thermtrip ever see the temperatures they are
> programmed for.
>
> NVIDIA's downstream driver runs Tegra114 in "lower precision" mode for
> this reason (tegra11_soctherm.c: "We must use lower precision (0) due
> to cp_fuse corrections"): the sensors count in 1 C units, the readback
> is scaled by two and all thresholds are in 2 C steps.
>
> Do the same with a lower_precision flag in the SoC data:
>
> - keep the fuse calibration in 0.5 C units and halve THERMA and THERMB
>   only in the final rounding step, so THERMA * capture stays around
>   2^23 and the sensors report the same temperatures at half the
>   resolution,
> - double the readback in translate_temp(),
> - use a threshold grain of 2000 so that thermtrip, the throttle levels
>   and the trip interrupts are programmed in 2 C steps.
>
> The use_lower_precision option in v3 of the Tegra114 series and the
> downstream driver halve the calibration temperatures before the
> calculation instead, which applies fuse_corr_beta at twice its weight.
> Here the result is exactly the current calibration divided by two. The
> PLLX hotspot offsets are left as they are, as in the downstream driver.
> Tegra124, Tegra132 and Tegra210 are unchanged.
>
> Tested on an NVIDIA Tegra Note 7 (P1640) together with the shifted_cp
> fix: THERMA * capture now stays between 7.7 and 9.2 million, and all
> eight sensors report TEMP_VALID in idle and after 60 s of load on all
> four cores at 1.12 GHz. The zones read cpu 43, gpu 39, pllx 43, mem
> 39 C in idle (board thermistor 34 C) and cpu 63, gpu 55, pllx 62, mem
> 53 C under load. Before, the gpu zone read 0 C and the cpu zone stopped
> at about 45 C. The thermtrip, throttle and trip thresholds are
> programmed as 51, 50 and 45, i.e. 102, 100 and 90 C.
>
> Fixes: 9d522a877b6f ("thermal/drivers/tegra: Add Tegra114 specific SOCTHERM driver")
> Cc: stable@xxxxxxxxxxxxxxx
> Link: https://lore.kernel.org/all/22816196.EfDdHjke4D@senjougahara/
> Signed-off-by: Patrick Fischer <admin@xxxxxxxxxxxxxxxxxx>
> ---
>  drivers/thermal/tegra/soctherm-fuse.c     | 15 ++++++++++++---
>  drivers/thermal/tegra/soctherm.c          | 21 +++++++++++++--------
>  drivers/thermal/tegra/soctherm.h          |  6 ++++++
>  drivers/thermal/tegra/tegra114-soctherm.c |  4 +++-
>  4 files changed, 34 insertions(+), 12 deletions(-)
>

Tested-by: Svyatoslav Ryhel <clamor95@xxxxxxxxx> # Tegratab (P1988)
and ASUS TF701T