Re: [PATCH v4 2/2] media: uvcvideo: Fix sequence number when no EOF

[Hans de Goede <hansg@kernel.org>]

Hi,

On 20-Mar-26 14:35, Ricardo Ribalda wrote:
> If the driver could not detect the EOF, the sequence number is increased
> twice:
>  1) When we enter uvc_video_decode_start() with the old buffer and FID has
>    fliped => We return -EAGAIN and last_fid is not flipped
>  2) When we enter uvc_video_decode_start() with the new buffer.
> 
> Fix this issue by moving the new frame detection logic earlier in
> uvc_video_decode_start().
> 
> Cc: stable@kernel.org
> Fixes: 650b95feee35 ("[media] uvcvideo: Generate discontinuous sequence numbers when frames are lost")
> Reported-by: Hans de Goede <hansg@kernel.org>
> Closes: https://lore.kernel.org/linux-media/CANiDSCuj4cPuB5_v2xyvAagA5FjoN8V5scXiFFOeD3aKDMqkCg@mail.gmail.com/T/#me39fb134e8c2c085567a31548c3403eb639625e4
> Signed-off-by: Ricardo Ribalda <ribalda@chromium.org>
> ---
>  drivers/media/usb/uvc/uvc_video.c | 48 ++++++++++++++++++++-------------------
>  1 file changed, 25 insertions(+), 23 deletions(-)
> 
> diff --git a/drivers/media/usb/uvc/uvc_video.c b/drivers/media/usb/uvc/uvc_video.c
> index 9e06b1d0f0f9..2218e4d8e564 100644
> --- a/drivers/media/usb/uvc/uvc_video.c
> +++ b/drivers/media/usb/uvc/uvc_video.c
> @@ -1168,6 +1168,31 @@ static int uvc_video_decode_start(struct uvc_streaming *stream,
>  	header_len = data[0];
>  	fid = data[1] & UVC_STREAM_FID;
>  
> +	/*
> +	 * Mark the buffer as done if we're at the beginning of a new frame.
> +	 * End of frame detection is better implemented by checking the EOF
> +	 * bit (FID bit toggling is delayed by one frame compared to the EOF
> +	 * bit), but some devices don't set the bit at end of frame (and the
> +	 * last payload can be lost anyway). We thus must check if the FID has
> +	 * been toggled.
> +	 *
> +	 * stream->last_fid is initialized to -1, so the first isochronous
> +	 * frame will never trigger an end of frame detection.
> +	 *
> +	 * Empty buffers (bytesused == 0) don't trigger end of frame detection
> +	 * as it doesn't make sense to return an empty buffer. This also
> +	 * avoids detecting end of frame conditions at FID toggling if the
> +	 * previous payload had the EOF bit set.
> +	 */
> +	if (stream->last_fid != -1 && fid != stream->last_fid &&
> +	    buf && buf->bytesused != 0) {
> +		uvc_dbg(stream->dev, FRAME,
> +			"Frame complete (FID bit toggled)\n");
> +		buf->state = UVC_BUF_STATE_READY;
> +
> +		return -EAGAIN;
> +	}
> +
>  	/*
>  	 * Increase the sequence number regardless of any buffer states, so
>  	 * that discontinuous sequence numbers always indicate lost frames.

Nice, since this compensates for the previous packet being parsed missing
the EOF flag doing it as the first thing makes a lot of sense.

Note this has a number of extra advantages which would be good to list
in the commit message for v5:

- The error status from the new packet will no longer get propagated
  to the previous frame-buffer.
- uvc_video_clock_decode() will no longer update the previous frame buf->stf
  with info from the new packet.
- uvc_video_clock_decode() and uvc_video_stats_decode() will no longer
  get called twice for the same packet.

I wonder if we should move the JPEG SOI marker detect also up for
all the same reasons? That seems to be for a camera not toggling FID
itself *and* sometimes missing the packet with the EOF flag ?

Regards,

Hans






> @@ -1234,29 +1259,6 @@ static int uvc_video_decode_start(struct uvc_streaming *stream,
>  		buf->state = UVC_BUF_STATE_ACTIVE;
>  	}
>  
> -	/*
> -	 * Mark the buffer as done if we're at the beginning of a new frame.
> -	 * End of frame detection is better implemented by checking the EOF
> -	 * bit (FID bit toggling is delayed by one frame compared to the EOF
> -	 * bit), but some devices don't set the bit at end of frame (and the
> -	 * last payload can be lost anyway). We thus must check if the FID has
> -	 * been toggled.
> -	 *
> -	 * stream->last_fid is initialized to -1, so the first isochronous
> -	 * frame will never trigger an end of frame detection.
> -	 *
> -	 * Empty buffers (bytesused == 0) don't trigger end of frame detection
> -	 * as it doesn't make sense to return an empty buffer. This also
> -	 * avoids detecting end of frame conditions at FID toggling if the
> -	 * previous payload had the EOF bit set.
> -	 */
> -	if (fid != stream->last_fid && buf->bytesused != 0) {
> -		uvc_dbg(stream->dev, FRAME,
> -			"Frame complete (FID bit toggled)\n");
> -		buf->state = UVC_BUF_STATE_READY;
> -		return -EAGAIN;
> -	}
> -
>  	/*
>  	 * Some cameras, when running two parallel streams (one MJPEG alongside
>  	 * another non-MJPEG stream), are known to lose the EOF packet for a frame.
>