Re: [PATCH 1/2] media: Add timberdale video-in driver

[Richard RÃjfors]

On 04/25/2010 05:24 PM, Mauro Carvalho Chehab wrote:
> Richard RÃjfors wrote:
> > This patch adds the timberdale video-in driver.
> >
> > The video IP of timberdale delivers the video data via DMA.
> > The driver uses the DMA api to handle DMA transfers, and make use
> > of the V4L2 videobuffers to handle buffers against user space.
> > Due to some timing constraint it makes sense to do DMA into an
> > intermediate buffer and then copy the data to vmalloc:ed buffers.
> >
> > If available the driver uses an encoder to get/set the video standard
> >
> > Signed-off-by: Richard RÃjfors<richard.rojfors@xxxxxxxxxxxxxx>
> > +#define TIMBLOGIW_DMA_BUFFER_SIZE	(TIMBLOGIW_BYTES_PER_LINE * 576)
>
> ...
>
> > +static int __timblogiw_alloc_dma(struct timblogiw_fh *fh, struct device *dev)
> > +{
> > +	dma_addr_t addr;
> > +	int err, i, pos;
> > +	int bytes_per_desc = TIMBLOGIW_LINES_PER_DESC *
> > +		timblogiw_bytes_per_line(fh->cur_norm);
> > +
> > +	fh->dma.cookie = -1;
> > +	fh->dma.dev = dev;
> > +
> > +	fh->dma.buf = kzalloc(TIMBLOGIW_DMA_BUFFER_SIZE, GFP_KERNEL);
> > +	if (!fh->dma.buf)
> > +		return -ENOMEM;
>
>
> Why do you need a fixed DMA buffer size? Just allocate the buffer size dynamically at
> buffer_prepare callback.
> > +	videobuf_queue_vmalloc_init(&fh->vb_vidq,&timblogiw_video_qops,
> > +			NULL,&fh->queue_lock, V4L2_BUF_TYPE_VIDEO_CAPTURE,
> > +			V4L2_FIELD_NONE, sizeof(struct videobuf_buffer), fh);
>
> You should be using, instead, videobuf_dma_sg or videobuf_cont, instead of
> using videobuf-vmalloc. This way, you'll avoid double buffering.

1. dma_sg can not be used, the DMA engine requires the memory blocks to be aligned on a factor of 
bytes per line, so 4K pages wouldn't work.

2.
I tried using videobuf-dma-contig, but got poor performance. I can not really explain why, I though 
it's due to the fact that the contiguous buffer is allocated coherent -> no caching.
I saw both gstreamer and mplayer perform very badly.
The frame grabber requires the DMA transfer for a frame beeing started while the frame is decoded. 
When I tested using contigous buffers gstreamer sometimes was that slow that it sometimes missed to 
have a frame queued when a transfer was finished, so I got frame drops. Any other ideas of the poor 
performance? otherwise I would like to go for the double buffered solution.

Thanks
--Richard
--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/