[v10,6/6] media: uvcvideo: Fix hw timestamp handling for slow FPS

  In UVC 1.5 we get a single clock value per frame. With the current
buffer size of 32, FPS slowers than 32 might roll-over twice.

The current code cannot handle two roll-over and provide invalid
timestamps.

Remove all the samples from the circular buffer that are more than two
rollovers old, so the algorithm always provides good timestamps.

Note that we are removing values that are more than one second old,
which means that there is enough distance between the two points that
we use for the interpolation to provide good values.

Tested-by: HungNien Chen <hn.chen@sunplusit.com>
Reviewed-by: Sergey Senozhatsky <senozhatsky@chromium.org>
Signed-off-by: Ricardo Ribalda <ribalda@chromium.org>
---
 drivers/media/usb/uvc/uvc_video.c | 24 ++++++++++++++++++++++++
 drivers/media/usb/uvc/uvcvideo.h  |  1 +
 2 files changed, 25 insertions(+)
  

On Sat, Mar 23, 2024 at 10:48:07AM +0000, Ricardo Ribalda wrote:
> In UVC 1.5 we get a single clock value per frame. With the current
> buffer size of 32, FPS slowers than 32 might roll-over twice.
> 
> The current code cannot handle two roll-over and provide invalid
> timestamps.
> 
> Remove all the samples from the circular buffer that are more than two
> rollovers old, so the algorithm always provides good timestamps.
> 
> Note that we are removing values that are more than one second old,
> which means that there is enough distance between the two points that
> we use for the interpolation to provide good values.
> 
> Tested-by: HungNien Chen <hn.chen@sunplusit.com>
> Reviewed-by: Sergey Senozhatsky <senozhatsky@chromium.org>
> Signed-off-by: Ricardo Ribalda <ribalda@chromium.org>
> ---
>  drivers/media/usb/uvc/uvc_video.c | 24 ++++++++++++++++++++++++
>  drivers/media/usb/uvc/uvcvideo.h  |  1 +
>  2 files changed, 25 insertions(+)
> 
> diff --git a/drivers/media/usb/uvc/uvc_video.c b/drivers/media/usb/uvc/uvc_video.c
> index 5df8f61d39cd1..900b57afac93a 100644
> --- a/drivers/media/usb/uvc/uvc_video.c
> +++ b/drivers/media/usb/uvc/uvc_video.c
> @@ -471,8 +471,31 @@ static void uvc_video_clock_add_sample(struct uvc_clock *clock,
>  {
>  	unsigned long flags;
>  
> +	/*
> +	 * If we write new data on the position where we had the last
> +	 * overflow, remove the overflow pointer. There is no overflow
> +	 * on the whole circular buffer.
> +	 */
> +	if (clock->head == clock->last_sof_overflow)
> +		clock->last_sof_overflow = -1;
> +
>  	spin_lock_irqsave(&clock->lock, flags);
>  
> +	/* Handle overflows */
> +	if (clock->count > 0 && clock->last_sof > sample->dev_sof) {
> +		/*
> +		 * Remove data from the circular buffer that is older than the
> +		 * last overflow. We only support one overflow per circular
> +		 * buffer.
> +		 */
> +		if (clock->last_sof_overflow != -1) {
> +			clock->count = (clock->head - clock->last_sof_overflow
> +					+ clock->count) % clock->count;
> +		}
> +		clock->last_sof_overflow = clock->head;
> +	}
> +
> +	/* Add sample */
>  	clock->samples[clock->head] = *sample;
>  	clock->head = (clock->head + 1) % clock->size;
>  	clock->count = min(clock->count + 1, clock->size);
> @@ -616,6 +639,7 @@ static void uvc_video_clock_reset(struct uvc_clock *clock)
>  	clock->head = 0;
>  	clock->count = 0;
>  	clock->last_sof = -1;
> +	clock->last_sof_overflow = -1;
>  	clock->sof_offset = -1;
>  }
>  
> diff --git a/drivers/media/usb/uvc/uvcvideo.h b/drivers/media/usb/uvc/uvcvideo.h
> index cb9dd50bba8ac..fb9f9771131ac 100644
> --- a/drivers/media/usb/uvc/uvcvideo.h
> +++ b/drivers/media/usb/uvc/uvcvideo.h
> @@ -499,6 +499,7 @@ struct uvc_streaming {
>  		unsigned int head;
>  		unsigned int count;
>  		unsigned int size;
> +		unsigned int last_sof_overflow;
>  
>  		u16 last_sof;
>  		u16 sof_offset;
> 
> -- 
> 2.44.0.396.g6e790dbe36-goog
> 
> 

Given that majority of cameras kind of run ~30 fps, this should improve
the timestamps for quite a lot of the users. Thanks.

Reviewed-by: Tomasz Figa <tfiga@chromium.org>

Best regards,
Tomasz
  

Hi Ricardo,

Thank you for the patch.

On Sat, Mar 23, 2024 at 10:48:07AM +0000, Ricardo Ribalda wrote:
> In UVC 1.5 we get a single clock value per frame. With the current
> buffer size of 32, FPS slowers than 32 might roll-over twice.
> 
> The current code cannot handle two roll-over and provide invalid
> timestamps.
> 
> Remove all the samples from the circular buffer that are more than two
> rollovers old, so the algorithm always provides good timestamps.
> 
> Note that we are removing values that are more than one second old,
> which means that there is enough distance between the two points that
> we use for the interpolation to provide good values.
> 
> Tested-by: HungNien Chen <hn.chen@sunplusit.com>
> Reviewed-by: Sergey Senozhatsky <senozhatsky@chromium.org>
> Signed-off-by: Ricardo Ribalda <ribalda@chromium.org>
> ---
>  drivers/media/usb/uvc/uvc_video.c | 24 ++++++++++++++++++++++++
>  drivers/media/usb/uvc/uvcvideo.h  |  1 +
>  2 files changed, 25 insertions(+)
> 
> diff --git a/drivers/media/usb/uvc/uvc_video.c b/drivers/media/usb/uvc/uvc_video.c
> index 5df8f61d39cd1..900b57afac93a 100644
> --- a/drivers/media/usb/uvc/uvc_video.c
> +++ b/drivers/media/usb/uvc/uvc_video.c
> @@ -471,8 +471,31 @@ static void uvc_video_clock_add_sample(struct uvc_clock *clock,
>  {
>  	unsigned long flags;
>  
> +	/*
> +	 * If we write new data on the position where we had the last
> +	 * overflow, remove the overflow pointer. There is no overflow

s/overflow/SOF overflow/

otherwise it sounds like a circular buffer overflow. Same in the other
comments below.

> +	 * on the whole circular buffer.
> +	 */
> +	if (clock->head == clock->last_sof_overflow)
> +		clock->last_sof_overflow = -1;
> +
>  	spin_lock_irqsave(&clock->lock, flags);
>  
> +	/* Handle overflows */

s/overflows/overflows./

> +	if (clock->count > 0 && clock->last_sof > sample->dev_sof) {
> +		/*
> +		 * Remove data from the circular buffer that is older than the
> +		 * last overflow. We only support one overflow per circular
> +		 * buffer.
> +		 */
> +		if (clock->last_sof_overflow != -1) {
> +			clock->count = (clock->head - clock->last_sof_overflow
> +					+ clock->count) % clock->count;

If I'm following you correctly here, you want to set count to the
distance between last_sof_overflow and head. Shouldn't it be

			clock->count = (clock->head - clock->last_sof_overflow
					+ clock->size) % clock->size;

> +		}

No need for curly braces.

> +		clock->last_sof_overflow = clock->head;
> +	}
> +
> +	/* Add sample */

s/sample/sample./

I still think it would be nicer to handle multiple rollovers correctly,
but that probably better handled by moving all the clock handling to
userspace. With the above issues addressed, I think this patch can go
in.

I could address all these issues when applying, but I'd like the count
-> size change to be tested first. You can submit a new version of this
patch only, I've applied the rest of the series to my tree already.

>  	clock->samples[clock->head] = *sample;
>  	clock->head = (clock->head + 1) % clock->size;
>  	clock->count = min(clock->count + 1, clock->size);
> @@ -616,6 +639,7 @@ static void uvc_video_clock_reset(struct uvc_clock *clock)
>  	clock->head = 0;
>  	clock->count = 0;
>  	clock->last_sof = -1;
> +	clock->last_sof_overflow = -1;
>  	clock->sof_offset = -1;
>  }
>  
> diff --git a/drivers/media/usb/uvc/uvcvideo.h b/drivers/media/usb/uvc/uvcvideo.h
> index cb9dd50bba8ac..fb9f9771131ac 100644
> --- a/drivers/media/usb/uvc/uvcvideo.h
> +++ b/drivers/media/usb/uvc/uvcvideo.h
> @@ -499,6 +499,7 @@ struct uvc_streaming {
>  		unsigned int head;
>  		unsigned int count;
>  		unsigned int size;
> +		unsigned int last_sof_overflow;
>  
>  		u16 last_sof;
>  		u16 sof_offset;