Le dimanche 19 mars 2017 à 00:54 +0000, Russell King - ARM Linux a
écrit :

quoted

In practice, I have the impression there is a fair reason why
framerate
enumeration isn't implemented (considering there is only 1 valid
rate).

That's actually completely incorrect.

With the capture device interfacing directly with CSI, it's possible
_today_ to select:

* the CSI sink pad's resolution
* the CSI sink pad's resolution with the width and/or height halved
* the CSI sink pad's frame rate
* the CSI sink pad's frame rate divided by the frame drop factor

To put it another way, these are possible:

# v4l2-ctl -d /dev/video10 --list-formats-ext
ioctl: VIDIOC_ENUM_FMT
        Index       : 0
        Type        : Video Capture
        Pixel Format: 'RGGB'
        Name        : 8-bit Bayer RGRG/GBGB
                Size: Discrete 816x616
                        Interval: Discrete 0.040s (25.000 fps)
                        Interval: Discrete 0.048s (20.833 fps)
                        Interval: Discrete 0.050s (20.000 fps)
                        Interval: Discrete 0.053s (18.750 fps)
                        Interval: Discrete 0.060s (16.667 fps)
                        Interval: Discrete 0.067s (15.000 fps)
                        Interval: Discrete 0.080s (12.500 fps)
                        Interval: Discrete 0.100s (10.000 fps)
                        Interval: Discrete 0.120s (8.333 fps)
                        Interval: Discrete 0.160s (6.250 fps)
                        Interval: Discrete 0.200s (5.000 fps)
                        Interval: Discrete 0.240s (4.167 fps)
                Size: Discrete 408x616
<same intervals>
                Size: Discrete 816x308
<same intervals>
                Size: Discrete 408x308
<same intervals>

These don't become possible as a result of implementing the enums,
they're all already requestable through /dev/video10.

Ok that wasn't clear. So basically video9 is a front-end to video10,
and it does not proxy the enumerations. I understand this is what you
are now fixing. And this has to be fixed, because I can image cases
where the front-end could support only a subset of the sub-dev. So
having userspace enumerate on another device (and having to find this
device by walking the tree) is unlikely to work in all scenarios.

regards,
Nicolas

p.s. This is why caps negotiation is annoyingly complex in GStreamer,
specially that there is no shortcut, you connect pads, and they figure-
out what format they will use between each other.