Re: [PATCH 1/3] iio: common: st_sensors: honour channel endianness in read_axis_data

[me@herrie.org]

On 2026-06-05 19:47, Andy Shevchenko wrote:
> On Fri, Jun 05, 2026 at 12:08:41PM +0200, Herman van Hazendonk wrote:
>> st_sensors_read_axis_data() unconditionally decoded multi-byte
>> results with get_unaligned_le16() / get_unaligned_le24() regardless
>> of the channel's declared scan_type.endianness.
>> 
>> For every ST sensor that has used this helper since it was introduced
>> this happened to be fine because the ST IMU/accel/gyro/pressure
>> families publish their data registers as little-endian and the
>> channel specs in those drivers declare IIO_LE accordingly.
>> 
>> The LSM303DLH magnetometer however publishes its X/Y/Z output as a
>> pair of big-endian bytes (the H register sits at the lower address,
>> 0x03/0x05/0x07, and the L register immediately after), and its
>> channel specs in st_magn_core.c correctly declare IIO_BE -- but
>> read_axis_data() ignored that and decoded as little-endian, swapping
>> the high and low bytes of every magnetometer sample.
>> 
>> The bug is most visible on a stationary chip: in earth's field the
>> true X reading is small and the high byte sits at 0x00, so swapping
>> the bytes pins sysfs X at exactly the low byte's pattern (e.g. 0x00F0
>> = 240). Y and Z still appear "to vary" because their magnitudes are
>> larger and the noise in the low byte produces big swings in the
>> swapped high byte:
>> 
>>   before (chip flat, sysfs in_magn_*_raw):
>>       X=240 (stuck), Y= 12032..23296, Z=-16128..-9728
>> 
>>   after (direct i2c-dev big-endian decode, same chip same 
>> orientation):
>>       X≈-4096, Y≈210, Z≈80     (sensible values reflecting earth's
>>                                 ambient field at low gauss range)
>> 
>> Fix read_axis_data() to dispatch on ch->scan_type.endianness and
>> call get_unaligned_be16() / get_unaligned_be24() when the channel
>> declares IIO_BE. Existing IIO_LE consumers (st_accel, st_gyro,
>> st_pressure, st_lsm6dsx and others) are unaffected because their
>> channel specs already declare IIO_LE and the LE path is unchanged.
> 
> ...
> 
>> -	if (byte_for_channel == 1)
>> +	if (byte_for_channel == 1) {
>>  		*data = (s8)*outdata;
> 
> Maybe for the consistency's sake use sign_extend32() everywhere?
> 
>> +	} else if (byte_for_channel == 2) {
>> +		if (ch->scan_type.endianness == IIO_BE)
>> +			*data = (s16)get_unaligned_be16(outdata);
>> +		else
>> +			*data = (s16)get_unaligned_le16(outdata);
>> +	} else if (byte_for_channel == 3) {
>> +		if (ch->scan_type.endianness == IIO_BE)
>> +			*data = (s32)sign_extend32(get_unaligned_be24(outdata),
>> +						   23);
>> +		else
>> +			*data = (s32)sign_extend32(get_unaligned_le24(outdata),
>> +						   23);
> 
> Why do you need casting here? sign_extend32() should return signed 
> type.
> With this being addressed, you can make them one-liners.
> 
>> +	}
> 
> 	u32 tmp;
> 	...
> 	if (byte_for_channel == 1) {
> 		// this way is done to show the below variant
> 		tmp = *outdata;
> 		*data = sign_extend32(tmp, 7);
> 	} else if (byte_for_channel == 2) {
> 		if (ch->scan_type.endianness == IIO_BE)
> 			tmp = get_unaligned_be16(outdata);
> 		else
> 			tmp = get_unaligned_le16(outdata);
> 		*data = sign_extend32(tmp, 15);
> 	} else if (byte_for_channel == 3) {
> 		if (ch->scan_type.endianness == IIO_BE)
> 			tmp = get_unaligned_be24(outdata);
> 		else
> 			tmp = get_unaligned_le24(outdata);
> 		*data = sign_extend32(tmp, 23);
> 	}
> 
> Or even
> 
> 	if (byte_for_channel == 1) {
> 		tmp = *outdata;
> 	} else if (byte_for_channel == 2) {
> 		if (ch->scan_type.endianness == IIO_BE)
> 			tmp = get_unaligned_be16(outdata);
> 		else
> 			tmp = get_unaligned_le16(outdata);
> 	} else if (byte_for_channel == 3) {
> 		if (ch->scan_type.endianness == IIO_BE)
> 			tmp = get_unaligned_be24(outdata);
> 		else
> 			tmp = get_unaligned_le24(outdata);
> 	} else {
> 		...error...
> 	}
> 	*data = sign_extend32(tmp, BYTES_TO_BITS(byte_for_channel) - 1);
Good points, both applied as you suggested in v2. Single u32 tmp 
captures
the per-width decode, then one sign_extend32(tmp, 
BYTES_TO_BITS(byte_for_channel)
  - 1) at the bottom handles all three widths. Casts are gone. I also 
turned
the previously-silent 0/≥4 fall-through into an explicit return 
-EINVAL so
a malformed channel spec is caught instead of leaving *data uninitialis

Thanks,
Herman