Re: [PATCH v5 7/9] v4l: Renesas R-Car VSP1 driver

[Hans Verkuil <hverkuil@xs4all.nl>]

On 08/02/2013 12:57 PM, Laurent Pinchart wrote:
> Hi Hans,
> 
> On Friday 02 August 2013 11:23:46 Hans Verkuil wrote:
>> Hi Laurent,
>>
>> See my single remark below...
>>
>> On 08/02/2013 03:03 AM, Laurent Pinchart wrote:
>>> The VSP1 is a video processing engine that includes a blender, scalers,
>>> filters and statistics computation. Configurable data path routing logic
>>> allows ordering the internal blocks in a flexible way.
>>>
>>> Due to the configurable nature of the pipeline the driver implements the
>>> media controller API and doesn't use the V4L2 mem-to-mem framework, even
>>> though the device usually operates in memory to memory mode.
>>>
>>> Only the read pixel formatters, up/down scalers, write pixel formatters
>>> and LCDC interface are supported at this stage.
>>>
>>> Signed-off-by: Laurent Pinchart
>>> <laurent.pinchart+renesas@ideasonboard.com>
>>> Acked-by: Sakari Ailus <sakari.ailus@iki.fi>
>>>
>>> diff --git a/drivers/media/platform/vsp1/vsp1_uds.h
>>> b/drivers/media/platform/vsp1/vsp1_uds.h new file mode 100644
>>> index 0000000..972a285
> 
> [snip]
> 
>>> +int vsp1_video_init(struct vsp1_video *video, struct vsp1_entity *rwpf)
>>> +{
>>> +	const char *direction;
>>> +	int ret;
>>> +
>>> +	switch (video->type) {
>>> +	case V4L2_BUF_TYPE_VIDEO_CAPTURE_MPLANE:
>>> +		direction = "output";
>>> +		video->pad.flags = MEDIA_PAD_FL_SINK;
>>> +		break;
>>> +
>>> +	case V4L2_BUF_TYPE_VIDEO_OUTPUT_MPLANE:
>>> +		direction = "input";
>>> +		video->pad.flags = MEDIA_PAD_FL_SOURCE;
>>> +		video->video.vfl_dir = VFL_DIR_TX;
>>> +		break;
>>> +
>>> +	default:
>>> +		return -EINVAL;
>>> +	}
>>> +
>>> +	video->rwpf = rwpf;
>>> +
>>> +	mutex_init(&video->lock);
>>> +	spin_lock_init(&video->irqlock);
>>> +	INIT_LIST_HEAD(&video->irqqueue);
>>> +
>>> +	mutex_init(&video->pipe.lock);
>>> +	spin_lock_init(&video->pipe.irqlock);
>>> +	INIT_LIST_HEAD(&video->pipe.entities);
>>> +	init_waitqueue_head(&video->pipe.wq);
>>> +	video->pipe.state = VSP1_PIPELINE_STOPPED;
>>> +
>>> +	/* Initialize the media entity... */
>>> +	ret = media_entity_init(&video->video.entity, 1, &video->pad, 0);
>>> +	if (ret < 0)
>>> +		return ret;
>>> +
>>> +	/* ... and the format ... */
>>> +	video->fmtinfo = vsp1_get_format_info(VSP1_VIDEO_DEF_FORMAT);
>>> +	video->format.pixelformat = video->fmtinfo->fourcc;
>>> +	video->format.colorspace = V4L2_COLORSPACE_SRGB;
>>> +	video->format.field = V4L2_FIELD_NONE;
>>> +	video->format.width = VSP1_VIDEO_DEF_WIDTH;
>>> +	video->format.height = VSP1_VIDEO_DEF_HEIGHT;
>>> +	video->format.num_planes = 1;
>>> +	video->format.plane_fmt[0].bytesperline >>> +		video->format.width * video->fmtinfo->bpp[0] / 8;
>>> +	video->format.plane_fmt[0].sizeimage >>> +		video->format.plane_fmt[0].bytesperline * video->format.height;
>>> +
>>> +	/* ... and the video node... */
>>> +	video->video.v4l2_dev = &video->vsp1->v4l2_dev;
>>> +	video->video.fops = &vsp1_video_fops;
>>> +	snprintf(video->video.name, sizeof(video->video.name), "%s %s",
>>> +		 rwpf->subdev.name, direction);
>>> +	video->video.vfl_type = VFL_TYPE_GRABBER;
>>> +	video->video.release = video_device_release_empty;
>>> +	video->video.ioctl_ops = &vsp1_video_ioctl_ops;
>>> +
>>> +	video_set_drvdata(&video->video, video);
>>> +
>>> +	/* ... and the buffers queue... */
>>> +	video->alloc_ctx = vb2_dma_contig_init_ctx(video->vsp1->dev);
>>> +	if (IS_ERR(video->alloc_ctx))
>>> +		goto error;
>>> +
>>> +	video->queue.type = video->type;
>>> +	video->queue.io_modes = VB2_MMAP | VB2_USERPTR | VB2_DMABUF;
>>> +	video->queue.drv_priv = video;
>>> +	video->queue.buf_struct_size = sizeof(struct vsp1_video_buffer);
>>> +	video->queue.ops = &vsp1_video_queue_qops;
>>> +	video->queue.mem_ops = &vb2_dma_contig_memops;
>>> +	video->queue.timestamp_type = V4L2_BUF_FLAG_TIMESTAMP_COPY;
>>
>> If you set video->queue.lock to &video->lock, then you can drop all the vb2
>> ioctl and fop helper functions directly without having to make your own
>> wrapper functions.
> 
> Right, I'll do so. I will also drop the manual lock handling from the STREAMON 
> and STREAMOFF handlers, as the core will use the queue lock for those.
> 
>> It saves a fair bit of code that way. The only place where there is a
>> difference as far as I can see is in vb2_fop_mmap: there the queue.lock
>> isn't taken whereas you do take the lock. It has never been 100% clear to
>> me whether or not that lock should be taken. However, as far as I can tell
>> vb2_mmap never calls any driver callbacks, so it seems to be me that there
>> is no need to take the lock.
> 
> Couldn't mmap() race with for instance REQBUFS(0) if we don't lock it ?

Hmm, good point. It would require a very convoluted program, but yes, there is
a possible race condition. The same is true for vb2_get_unmapped_area.

Can you prepare a patch adding locking for these two fops?

Thanks,

	Hans