Hi, On 11/1/23 11:31, Sakari Ailus wrote: > Hi Alexander, > > On Wed, Nov 01, 2023 at 11:09:00AM +0100, Alexander Stein wrote: >> cci_write() unconditionally writes multi-byte data in big-endian format. >> But IMX290 and IMX327 use little-endian format for multi-byte registers. >> Revert the CCI usage until little-endian is supported by CCI register >> access helpers. >> This reverts commit af73323b97702e53b0a336972aaf23e7dd92c850. >> >> Signed-off-by: Alexander Stein <alexander.stein@xxxxxxxxxxxxxxx> >> --- >> The difference is subtile, but imx290_write() uses put_unaligned_le24(), >> while cci_write() uses put_unaligned_be24(). >> >> I assume this should go into stable as well. How to deal with that? >> Shall a revert get the fixes tag as well? > > I'd very much prefer addressing this in v4l2-cci instead. It should also be > a much smaller patch (or patches). I was under the impression the v4l2-cci conversion was already reverted because this breakage has been known for a while now. Anyways, if someone wants to fix this directly this has been discussed in this thread: https://lore.kernel.org/linux-media/20231030173637.GA2977515@xxxxxxxxxxxxxxxxxxxxx/ and the consensus is that the best way to fix this is to add CCI_REG16_LE(x) CCI_REG24_LE(x), etc. macros to mirror the existing CCI_REG16(x), etc. macros. The _LE macros would then look something like this: #define CCI_REG_LE BIT(20) #define CCI_REG16_LE(x) (CCI_REG_LE | (2 << CCI_REG_WIDTH_SHIFT) | (x)) etc. And then the get_unaligned_beXX() and put_unaligned_beXX() calls in drivers/media/v4l2-core/v4l2-cci.c would need to be made conditional on a check for the CCI_REG_LE flag and if that flag is set use the _le_ versions of those functions instead. The reason to go this way instead of a global LE flag somewhere is that some sensors have mixed endianness for different registers, so encoding this in the register-address high bits is the best solution. Alexander, perhaps you can prepare 2 patches: 1. Adding the discussed CCI_REGXX_LE(x) macros to v4l2-cci described above. 2. Patch the imx290 code to use the _LE versions for the registers which are 2 or more bytes wide. ? Note I know that Alain (added to the Cc) from the thread linked above is also looking into implementing 1. Regards, Hans
