Re: [PATCH v2 1/2] media: v4l2-cci: Add support for little-endian encoded registers

[Sakari Ailus <sakari.ailus@linux.intel.com>]

On Thu, Nov 02, 2023 at 10:27:45AM +0100, Hans de Goede wrote:
> 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@vger.kernel.org
> >>>>> Signed-off-by: Alexander Stein <alexander.stein@ew.tq-group.com>
> >>>>> ---
> >>>>>
> >>>>>  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 .

I'm fine with that.

So I think these are good as-is then.

-- 
Sakari Ailus