[libcamera-devel] [PATCH] libcamera: v4l2_videodevice: Handle unexpected buffers

[Laurent Pinchart]

Hi Kieran,

Thank you for the patch.

On Thu, Jul 15, 2021 at 03:21:30PM +0100, Kieran Bingham wrote:
> A kernel bug can lead to unexpected buffers being dequeued where we
> haven't entered the buffer in our queuedBuffers_ list.
> 
> This causes invalid accesses if not handled correctly within libcamera,
> and while it is a kernel issue, we must protect against unpatched
> kernels.
> 
> Handle unexpected buffers by returning a nullptr, and move cache
> management after the validation of the buffer.
> 
> Signed-off-by: Kieran Bingham <kieran.bingham at ideasonboard.com>
> ---
>  src/libcamera/v4l2_videodevice.cpp | 21 ++++++++++++++++++++-
>  1 file changed, 20 insertions(+), 1 deletion(-)
> 
> diff --git a/src/libcamera/v4l2_videodevice.cpp b/src/libcamera/v4l2_videodevice.cpp
> index 3d2d99b46e4e..6c7f9daf24db 100644
> --- a/src/libcamera/v4l2_videodevice.cpp
> +++ b/src/libcamera/v4l2_videodevice.cpp
> @@ -1519,9 +1519,28 @@ FrameBuffer *V4L2VideoDevice::dequeueBuffer()
>  
>  	LOG(V4L2, Debug) << "Dequeuing buffer " << buf.index;
>  
> +	auto it = queuedBuffers_.find(buf.index);
> +	/*
> +	 * If the video node fails to stream-on successfully (which can occur
> +	 * when queing a buffer), a vb2 kernel bug can lead to the buffer which
> +	 * returns a failure upon queing, being mistakenely kept in the kernel.

s/queing,/queing/

You've mistakenly spelled mistakenly mistakenely :-)

> +	 * This leads to the kernel notifying us that a buffer is available to
> +	 * dequeue, which we have no awareness of being queued, and thus we will
> +	 * not find it in the queuedBuffers_ list.

It may be worth expanding this a bit to explain about the
stream-on-at-qbuf-time issue. If I recall correctly, given that the
device fails to stream, the issue only occurs when we call
VIDIOC_STREAMOFF and all buffers currently in queue (including the one
erroneously kept in the queue) being signalled as complete in the error
state. Is that right ?

> +	 *
> +	 * Whilst this is a kernel bug and should be fixed there, ensure that we
> +	 * safely ignore buffers which are unexpected to prevent crashes on
> +	 * unpatched kernels.

Is the fix on its way upstream ? If so, this could probably be reworded,
and mentioning the target kernel version (and even better, the upstream
commit) would be nice.

> +	 */
> +	if (it == queuedBuffers_.end()) {
> +		LOG(V4L2, Error)
> +			<< "Dequeued an unexpected buffer:" << buf.index;

Missing space after ':'.

> +

You can drop the blank line.

> +		return nullptr;
> +	}
> +
>  	cache_->put(buf.index);
>  
> -	auto it = queuedBuffers_.find(buf.index);
>  	FrameBuffer *buffer = it->second;
>  	queuedBuffers_.erase(it);
>  

-- 
Regards,

Laurent Pinchart