Hi, On 11/2/23 09:25, Sakari Ailus wrote: > Hi Alexander, > > On Thu, Nov 02, 2023 at 08:51:12AM +0100, Alexander Stein wrote: >> Hi, >> >> thanks for the feedback. >> >> Am Donnerstag, 2. November 2023, 07:30:44 CET schrieb Sakari Ailus: >>> Hi Laurent, >>> >>> On Thu, Nov 02, 2023 at 03:22:17AM +0200, Laurent Pinchart wrote: >>>> Hi Alexander, >>>> >>>> Thank you for the patch. >>>> >>>> On Wed, Nov 01, 2023 at 01:23:53PM +0100, Alexander Stein wrote: >>>>> Some sensors, e.g. Sony, are using little-endian registers. Add support >>>>> for >>>> >>>> I would write Sony IMX290 here, as there are Sony sensors that use big >>>> endian. >>> >>> Almost all of them. There are a few exceptions indeed. This seems to be a >>> bug. >> >> Let's name IMX290 here as an actual example. No need to worry about other >> models here. >> >>>>> those by encoding the endianess into Bit 20 of the register address. >>>>> >>>>> Fixes: af73323b97702 ("media: imx290: Convert to new CCI register access >>>>> helpers") Cc: stable@xxxxxxxxxxxxxxx >>>>> Signed-off-by: Alexander Stein <alexander.stein@xxxxxxxxxxxxxxx> >>>>> --- >>>>> >>>>> drivers/media/v4l2-core/v4l2-cci.c | 44 ++++++++++++++++++++++++------ >>>>> include/media/v4l2-cci.h | 5 ++++ >>>>> 2 files changed, 41 insertions(+), 8 deletions(-) >>>>> >>>>> diff --git a/drivers/media/v4l2-core/v4l2-cci.c >>>>> b/drivers/media/v4l2-core/v4l2-cci.c index bc2dbec019b04..673637b67bf67 >>>>> 100644 >>>>> --- a/drivers/media/v4l2-core/v4l2-cci.c >>>>> +++ b/drivers/media/v4l2-core/v4l2-cci.c >>>>> @@ -18,6 +18,7 @@ >>>>> >>>>> int cci_read(struct regmap *map, u32 reg, u64 *val, int *err) >>>>> { >>>>> >>>>> + bool little_endian; >>>>> >>>>> unsigned int len; >>>>> u8 buf[8]; >>>>> int ret; >>>>> >>>>> @@ -25,6 +26,7 @@ int cci_read(struct regmap *map, u32 reg, u64 *val, >>>>> int *err)> > >>>>> if (err && *err) >>>>> >>>>> return *err; >>>>> >>>>> + little_endian = reg & CCI_REG_LE; >>>> >>>> You could initialize the variable when declaring it. Same below. >>> >>> I was thinking of the same, but then it'd be logical to move initialisation >>> of all related variables there. reg is modified here though. I'd keep >>> setting little_endian here. If someone wants to move it, that could be done >>> in a separate patch. >>> >>>>> len = FIELD_GET(CCI_REG_WIDTH_MASK, reg); >>>>> reg = FIELD_GET(CCI_REG_ADDR_MASK, reg); >>>>> >>>>> @@ -40,16 +42,28 @@ int cci_read(struct regmap *map, u32 reg, u64 *val, >>>>> int *err)> > >>>>> *val = buf[0]; >>>>> break; >>>>> >>>>> case 2: >>>>> - *val = get_unaligned_be16(buf); >>>>> + if (little_endian) >>>>> + *val = get_unaligned_le16(buf); >>>>> + else >>>>> + *val = get_unaligned_be16(buf); >>>> >>>> Unrelated to this patch, isn't buf aligned to a 4 bytes boundary ? >>> >>> Very probably, as it's right after len that's an unsigned int. Adding >>> __aligned(8) would ensure we don't need any of the unaligned variants, and >>> most likely would keep the stack layout as-is. >> >> You mean something like this? >> >> u8 __aligned(8) buf[8]; >> [...] >> if (little_endian) >> *val = le64_to_cpup(buf); >> else >> *val = be64_to_cpup(buf); >> >> But what about 24 Bits? There is no le24_to_cpup. I would rather use the same >> API for all cases. > > The aligned APIs are much better choice when you can use them. The 24 bit > case can remain special IMO. > >> >>> Or... how about putting it in an union with a u64? That would mean it's >>> accessible as u64 alignment-wise while the alignment itself is up to the >>> ABI. A comment would be good to have probably. >> >> An additional union seems a bit too much here. Unless something suitable >> already exists for general usage. > > I think it's nicer than using __aligned() as you get ABI alignment that > way, not something you force manually --- that's a bit crude. > > I wonder that others think. I'm fine with adding the __aligned(8) and switching the non 24 bit cases to helpers which assume alignment. The most important note I have is that that is a separate improvement from this series though. So this should be done in a follow-up patch which is not Cc: stable . Regards, Hans