Re: [RFC PATCH 1/3] media: v4l2-subdev.h: Add min and max enum

[Laurent Pinchart]

Hi Sakari,

On Thu, Apr 30, 2020 at 05:18:49PM +0300, Sakari Ailus wrote:
> On Thu, Apr 30, 2020 at 05:17:53PM +0300, Laurent Pinchart wrote:
> > On Thu, Apr 30, 2020 at 05:15:52PM +0300, Sakari Ailus wrote:
> >> On Thu, Apr 30, 2020 at 04:59:04PM +0300, Laurent Pinchart wrote:
> >>> On Thu, Apr 30, 2020 at 04:31:25PM +0300, Sakari Ailus wrote:
> >>>> On Thu, Apr 30, 2020 at 02:10:14PM +0300, Laurent Pinchart wrote:
> >>>>> On Thu, Apr 30, 2020 at 12:42:33PM +0300, Sakari Ailus wrote:
> >>>>>> On Tue, Apr 14, 2020 at 10:01:49PM +0200, Daniel Gomez wrote:
> >>>>>>> Add min and max structures to the v4l2-subdev callback in order to allow
> >>>>>>> the subdev to return a range of valid frame intervals.
> >>>>>>> 
> >>>>>>> This would operate similar to the struct v4l2_subdev_frame_size_enum and
> >>>>>>> its max and min values for the width and the height. In this case, the
> >>>>>>> possibility to return a frame interval range is added to the v4l2-subdev level
> >>>>>>> whenever the v4l2 device operates in step-wise or continuous mode.
> >>>>>> 
> >>>>>> The current API only allows providing a list of enumerated values. That is
> >>>>>> limiting indeed, especially on register list based sensor drivers where
> >>>>>> vertical blanking is configurable.
> >>>>>> 
> >>>>>> I guess this could be extended to cover what V4L2, more or less. If we tell
> >>>>>> it's a range, is it assumed to be contiguous? We don't have try operation
> >>>>>> for the frame interval, but I guess set is good enough. The fraction is
> >>>>>> probably best for TV standards but it's not what camera sensors natively
> >>>>>> use. (But for a register list based driver, the established practice
> >>>>>> remains to use frame interval.)
> >>>>>> 
> >>>>>> I'm also wondering the effect on existing user space; if a driver gives a
> >>>>>> range, how will the existing programs work with such a driver?
> >>>>>> 
> >>>>>> I'd add an anonymous union with the interval field, the other field being
> >>>>>> min_interval. Then the current applications would get the minimum interval
> >>>>>> and still continue to function. I guess compilers are modern enough these
> >>>>>> days we can have an anonymous union in the uAPI?
> >>>>> 
> >>>>> We can discuss all this, but given patch 3/3 in this series, I think
> >>>>> this isn't the right API :-) The sensor driver should not expose the
> >>>>> frame interval enumeration API. It should instead expose control of the
> >>>>> frame rate through V4L2_CID_PIXEL_RATE, V4L2_CID_HBLANK and
> >>>>> V4L2_CID_VBLANK.
> >>>>> 
> >>>> 
> >>>> That would require also exposing the size of the pixel array (and the
> >>>> analogue crop), in order to provide all the necessary information to
> >>>> calculate the frame rate. No objections there; this is a new driver.
> >>>> 
> >>>> There are however existing drivers that implement s_frame_interval subdev
> >>>> ioctl; those might benefit from this one. Or would you implement the pixel
> >>>> rate based control as well, and effectively deprecate the s_frame_interval
> >>>> on those?
> >>> 
> >>> That's what I would recommend, yes. I would only keep
> >>> .s_frame_interval() for sensors that expose that concept at the hardware
> >>> level (for instance with an integrated ISP whose firmware exposes a
> >>> frame interval or frame rate control).
> >> 
> >> Sounds good to me.
> >> 
> >> Jacopo's set exposing read-only subdevs completes the puzzle so the user
> >> space should have all it needs, right?
> > 
> > Until we run into the next missing piece :-)
> 
> I was thinking of the frame rate configuration. Can you confirm that?

I believe so, yes.

-- 
Regards,

Laurent Pinchart