[PATCH] media: cedrus: Don't kernel map most buffers

Hans Verkuil hverkuil-cisco at xs4all.nl
Mon Oct 4 01:10:55 PDT 2021


Hi Jernej,

Some comments below:

On 12/09/2021 08:08, Jernej Skrabec wrote:
> Except VP8 probability coefficients buffer, all other buffers are never

Except -> Except for

> accessed by CPU. That allows us to mark them with DMA_ATTR_NO_KERNEL_MAPPING
> flag. This helps with decoding big (like 4k) videos on 32-bit ARM
> platforms where default vmalloc size is relatively small - 240 MiB.
> Since auxiliary buffer are not yet efficiently allocated, this can be
> easily exceeded. Even if allocation is optimized, 4k videos will still
> often exceed this limit.
> 
> Signed-off-by: Jernej Skrabec <jernej.skrabec at gmail.com>
> ---
>  .../staging/media/sunxi/cedrus/cedrus_h264.c  | 102 ++++++++++--------
>  .../staging/media/sunxi/cedrus/cedrus_h265.c  |  28 ++---
>  .../staging/media/sunxi/cedrus/cedrus_video.c |   2 +
>  3 files changed, 73 insertions(+), 59 deletions(-)
> 
> diff --git a/drivers/staging/media/sunxi/cedrus/cedrus_h264.c b/drivers/staging/media/sunxi/cedrus/cedrus_h264.c
> index de7442d4834d..6e38b37d9fe1 100644
> --- a/drivers/staging/media/sunxi/cedrus/cedrus_h264.c
> +++ b/drivers/staging/media/sunxi/cedrus/cedrus_h264.c
> @@ -538,23 +538,23 @@ static int cedrus_h264_start(struct cedrus_ctx *ctx)
>  
>  	ctx->codec.h264.pic_info_buf_size = pic_info_size;
>  	ctx->codec.h264.pic_info_buf =
> -		dma_alloc_coherent(dev->dev, ctx->codec.h264.pic_info_buf_size,
> -				   &ctx->codec.h264.pic_info_buf_dma,
> -				   GFP_KERNEL);
> +		dma_alloc_attrs(dev->dev, ctx->codec.h264.pic_info_buf_size,
> +				&ctx->codec.h264.pic_info_buf_dma,
> +				GFP_KERNEL, DMA_ATTR_NO_KERNEL_MAPPING);
>  	if (!ctx->codec.h264.pic_info_buf)
>  		return -ENOMEM;
>  
>  	/*
>  	 * That buffer is supposed to be 16kiB in size, and be aligned
> -	 * on 16kiB as well. However, dma_alloc_coherent provides the
> +	 * on 16kiB as well. However, dma_alloc_attrs provides the
>  	 * guarantee that we'll have a CPU and DMA address aligned on

Does the 'CPU' part of this sentence still make sense since the CPU
won't access the buffer?

>  	 * the smallest page order that is greater to the requested
>  	 * size, so we don't have to overallocate.
>  	 */
>  	ctx->codec.h264.neighbor_info_buf =
> -		dma_alloc_coherent(dev->dev, CEDRUS_NEIGHBOR_INFO_BUF_SIZE,
> -				   &ctx->codec.h264.neighbor_info_buf_dma,
> -				   GFP_KERNEL);
> +		dma_alloc_attrs(dev->dev, CEDRUS_NEIGHBOR_INFO_BUF_SIZE,
> +				&ctx->codec.h264.neighbor_info_buf_dma,
> +				GFP_KERNEL, DMA_ATTR_NO_KERNEL_MAPPING);

I think it would be good to have a comment for all these dma_alloc_attrs
calls where you note that these buffers are used by the HW only, and
never by the CPU, hence the use of DMA_ATTR_NO_KERNEL_MAPPING.

>  	if (!ctx->codec.h264.neighbor_info_buf) {
>  		ret = -ENOMEM;
>  		goto err_pic_buf;
> @@ -582,10 +582,11 @@ static int cedrus_h264_start(struct cedrus_ctx *ctx)
>  
>  	mv_col_size = field_size * 2 * CEDRUS_H264_FRAME_NUM;
>  	ctx->codec.h264.mv_col_buf_size = mv_col_size;
> -	ctx->codec.h264.mv_col_buf = dma_alloc_coherent(dev->dev,
> -							ctx->codec.h264.mv_col_buf_size,
> -							&ctx->codec.h264.mv_col_buf_dma,
> -							GFP_KERNEL);
> +	ctx->codec.h264.mv_col_buf =
> +		dma_alloc_attrs(dev->dev,
> +				ctx->codec.h264.mv_col_buf_size,
> +				&ctx->codec.h264.mv_col_buf_dma,
> +				GFP_KERNEL, DMA_ATTR_NO_KERNEL_MAPPING);
>  	if (!ctx->codec.h264.mv_col_buf) {
>  		ret = -ENOMEM;
>  		goto err_neighbor_buf;
> @@ -600,10 +601,10 @@ static int cedrus_h264_start(struct cedrus_ctx *ctx)
>  		ctx->codec.h264.deblk_buf_size =
>  			ALIGN(ctx->src_fmt.width, 32) * 12;
>  		ctx->codec.h264.deblk_buf =
> -			dma_alloc_coherent(dev->dev,
> -					   ctx->codec.h264.deblk_buf_size,
> -					   &ctx->codec.h264.deblk_buf_dma,
> -					   GFP_KERNEL);
> +			dma_alloc_attrs(dev->dev,
> +					ctx->codec.h264.deblk_buf_size,
> +					&ctx->codec.h264.deblk_buf_dma,
> +					GFP_KERNEL, DMA_ATTR_NO_KERNEL_MAPPING);
>  		if (!ctx->codec.h264.deblk_buf) {
>  			ret = -ENOMEM;
>  			goto err_mv_col_buf;
> @@ -616,10 +617,10 @@ static int cedrus_h264_start(struct cedrus_ctx *ctx)
>  		ctx->codec.h264.intra_pred_buf_size =
>  			ALIGN(ctx->src_fmt.width, 64) * 5 * 2;
>  		ctx->codec.h264.intra_pred_buf =
> -			dma_alloc_coherent(dev->dev,
> -					   ctx->codec.h264.intra_pred_buf_size,
> -					   &ctx->codec.h264.intra_pred_buf_dma,
> -					   GFP_KERNEL);
> +			dma_alloc_attrs(dev->dev,
> +					ctx->codec.h264.intra_pred_buf_size,
> +					&ctx->codec.h264.intra_pred_buf_dma,
> +					GFP_KERNEL, DMA_ATTR_NO_KERNEL_MAPPING);
>  		if (!ctx->codec.h264.intra_pred_buf) {
>  			ret = -ENOMEM;
>  			goto err_deblk_buf;
> @@ -629,24 +630,28 @@ static int cedrus_h264_start(struct cedrus_ctx *ctx)
>  	return 0;
>  
>  err_deblk_buf:
> -	dma_free_coherent(dev->dev, ctx->codec.h264.deblk_buf_size,
> -			  ctx->codec.h264.deblk_buf,
> -			  ctx->codec.h264.deblk_buf_dma);
> +	dma_free_attrs(dev->dev, ctx->codec.h264.deblk_buf_size,
> +		       ctx->codec.h264.deblk_buf,
> +		       ctx->codec.h264.deblk_buf_dma,
> +		       DMA_ATTR_NO_KERNEL_MAPPING);
>  
>  err_mv_col_buf:
> -	dma_free_coherent(dev->dev, ctx->codec.h264.mv_col_buf_size,
> -			  ctx->codec.h264.mv_col_buf,
> -			  ctx->codec.h264.mv_col_buf_dma);
> +	dma_free_attrs(dev->dev, ctx->codec.h264.mv_col_buf_size,
> +		       ctx->codec.h264.mv_col_buf,
> +		       ctx->codec.h264.mv_col_buf_dma,
> +		       DMA_ATTR_NO_KERNEL_MAPPING);
>  
>  err_neighbor_buf:
> -	dma_free_coherent(dev->dev, CEDRUS_NEIGHBOR_INFO_BUF_SIZE,
> -			  ctx->codec.h264.neighbor_info_buf,
> -			  ctx->codec.h264.neighbor_info_buf_dma);
> +	dma_free_attrs(dev->dev, CEDRUS_NEIGHBOR_INFO_BUF_SIZE,
> +		       ctx->codec.h264.neighbor_info_buf,
> +		       ctx->codec.h264.neighbor_info_buf_dma,
> +		       DMA_ATTR_NO_KERNEL_MAPPING);
>  
>  err_pic_buf:
> -	dma_free_coherent(dev->dev, ctx->codec.h264.pic_info_buf_size,
> -			  ctx->codec.h264.pic_info_buf,
> -			  ctx->codec.h264.pic_info_buf_dma);
> +	dma_free_attrs(dev->dev, ctx->codec.h264.pic_info_buf_size,
> +		       ctx->codec.h264.pic_info_buf,
> +		       ctx->codec.h264.pic_info_buf_dma,
> +		       DMA_ATTR_NO_KERNEL_MAPPING);
>  	return ret;
>  }
>  
> @@ -654,23 +659,28 @@ static void cedrus_h264_stop(struct cedrus_ctx *ctx)
>  {
>  	struct cedrus_dev *dev = ctx->dev;
>  
> -	dma_free_coherent(dev->dev, ctx->codec.h264.mv_col_buf_size,
> -			  ctx->codec.h264.mv_col_buf,
> -			  ctx->codec.h264.mv_col_buf_dma);
> -	dma_free_coherent(dev->dev, CEDRUS_NEIGHBOR_INFO_BUF_SIZE,
> -			  ctx->codec.h264.neighbor_info_buf,
> -			  ctx->codec.h264.neighbor_info_buf_dma);
> -	dma_free_coherent(dev->dev, ctx->codec.h264.pic_info_buf_size,
> -			  ctx->codec.h264.pic_info_buf,
> -			  ctx->codec.h264.pic_info_buf_dma);
> +	dma_free_attrs(dev->dev, ctx->codec.h264.mv_col_buf_size,
> +		       ctx->codec.h264.mv_col_buf,
> +		       ctx->codec.h264.mv_col_buf_dma,
> +		       DMA_ATTR_NO_KERNEL_MAPPING);
> +	dma_free_attrs(dev->dev, CEDRUS_NEIGHBOR_INFO_BUF_SIZE,
> +		       ctx->codec.h264.neighbor_info_buf,
> +		       ctx->codec.h264.neighbor_info_buf_dma,
> +		       DMA_ATTR_NO_KERNEL_MAPPING);
> +	dma_free_attrs(dev->dev, ctx->codec.h264.pic_info_buf_size,
> +		       ctx->codec.h264.pic_info_buf,
> +		       ctx->codec.h264.pic_info_buf_dma,
> +		       DMA_ATTR_NO_KERNEL_MAPPING);
>  	if (ctx->codec.h264.deblk_buf_size)
> -		dma_free_coherent(dev->dev, ctx->codec.h264.deblk_buf_size,
> -				  ctx->codec.h264.deblk_buf,
> -				  ctx->codec.h264.deblk_buf_dma);
> +		dma_free_attrs(dev->dev, ctx->codec.h264.deblk_buf_size,
> +			       ctx->codec.h264.deblk_buf,
> +			       ctx->codec.h264.deblk_buf_dma,
> +			       DMA_ATTR_NO_KERNEL_MAPPING);
>  	if (ctx->codec.h264.intra_pred_buf_size)
> -		dma_free_coherent(dev->dev, ctx->codec.h264.intra_pred_buf_size,
> -				  ctx->codec.h264.intra_pred_buf,
> -				  ctx->codec.h264.intra_pred_buf_dma);
> +		dma_free_attrs(dev->dev, ctx->codec.h264.intra_pred_buf_size,
> +			       ctx->codec.h264.intra_pred_buf,
> +			       ctx->codec.h264.intra_pred_buf_dma,
> +			       DMA_ATTR_NO_KERNEL_MAPPING);
>  }
>  
>  static void cedrus_h264_trigger(struct cedrus_ctx *ctx)
> diff --git a/drivers/staging/media/sunxi/cedrus/cedrus_h265.c b/drivers/staging/media/sunxi/cedrus/cedrus_h265.c
> index 3d9561d4aadb..bb7eb56106c5 100644
> --- a/drivers/staging/media/sunxi/cedrus/cedrus_h265.c
> +++ b/drivers/staging/media/sunxi/cedrus/cedrus_h265.c
> @@ -351,10 +351,10 @@ static void cedrus_h265_setup(struct cedrus_ctx *ctx,
>  			ctx->codec.h265.mv_col_buf_unit_size;
>  
>  		ctx->codec.h265.mv_col_buf =
> -			dma_alloc_coherent(dev->dev,
> -					   ctx->codec.h265.mv_col_buf_size,
> -					   &ctx->codec.h265.mv_col_buf_addr,
> -					   GFP_KERNEL);
> +			dma_alloc_attrs(dev->dev,
> +					ctx->codec.h265.mv_col_buf_size,
> +					&ctx->codec.h265.mv_col_buf_addr,
> +					GFP_KERNEL, DMA_ATTR_NO_KERNEL_MAPPING);
>  		if (!ctx->codec.h265.mv_col_buf) {
>  			ctx->codec.h265.mv_col_buf_size = 0;
>  			// TODO: Abort the process here.
> @@ -668,9 +668,9 @@ static int cedrus_h265_start(struct cedrus_ctx *ctx)
>  	ctx->codec.h265.mv_col_buf_size = 0;
>  
>  	ctx->codec.h265.neighbor_info_buf =
> -		dma_alloc_coherent(dev->dev, CEDRUS_H265_NEIGHBOR_INFO_BUF_SIZE,
> -				   &ctx->codec.h265.neighbor_info_buf_addr,
> -				   GFP_KERNEL);
> +		dma_alloc_attrs(dev->dev, CEDRUS_H265_NEIGHBOR_INFO_BUF_SIZE,
> +				&ctx->codec.h265.neighbor_info_buf_addr,
> +				GFP_KERNEL, DMA_ATTR_NO_KERNEL_MAPPING);
>  	if (!ctx->codec.h265.neighbor_info_buf)
>  		return -ENOMEM;
>  
> @@ -682,16 +682,18 @@ static void cedrus_h265_stop(struct cedrus_ctx *ctx)
>  	struct cedrus_dev *dev = ctx->dev;
>  
>  	if (ctx->codec.h265.mv_col_buf_size > 0) {
> -		dma_free_coherent(dev->dev, ctx->codec.h265.mv_col_buf_size,
> -				  ctx->codec.h265.mv_col_buf,
> -				  ctx->codec.h265.mv_col_buf_addr);
> +		dma_free_attrs(dev->dev, ctx->codec.h265.mv_col_buf_size,
> +			       ctx->codec.h265.mv_col_buf,
> +			       ctx->codec.h265.mv_col_buf_addr,
> +			       DMA_ATTR_NO_KERNEL_MAPPING);
>  
>  		ctx->codec.h265.mv_col_buf_size = 0;
>  	}
>  
> -	dma_free_coherent(dev->dev, CEDRUS_H265_NEIGHBOR_INFO_BUF_SIZE,
> -			  ctx->codec.h265.neighbor_info_buf,
> -			  ctx->codec.h265.neighbor_info_buf_addr);
> +	dma_free_attrs(dev->dev, CEDRUS_H265_NEIGHBOR_INFO_BUF_SIZE,
> +		       ctx->codec.h265.neighbor_info_buf,
> +		       ctx->codec.h265.neighbor_info_buf_addr,
> +		       DMA_ATTR_NO_KERNEL_MAPPING);
>  }
>  
>  static void cedrus_h265_trigger(struct cedrus_ctx *ctx)
> diff --git a/drivers/staging/media/sunxi/cedrus/cedrus_video.c b/drivers/staging/media/sunxi/cedrus/cedrus_video.c
> index 66714609b577..800ffa5382de 100644
> --- a/drivers/staging/media/sunxi/cedrus/cedrus_video.c
> +++ b/drivers/staging/media/sunxi/cedrus/cedrus_video.c
> @@ -568,6 +568,7 @@ int cedrus_queue_init(void *priv, struct vb2_queue *src_vq,
>  
>  	src_vq->type = V4L2_BUF_TYPE_VIDEO_OUTPUT;
>  	src_vq->io_modes = VB2_MMAP | VB2_DMABUF;
> +	src_vq->dma_attrs = DMA_ATTR_NO_KERNEL_MAPPING;
>  	src_vq->drv_priv = ctx;
>  	src_vq->buf_struct_size = sizeof(struct cedrus_buffer);
>  	src_vq->ops = &cedrus_qops;
> @@ -584,6 +585,7 @@ int cedrus_queue_init(void *priv, struct vb2_queue *src_vq,
>  
>  	dst_vq->type = V4L2_BUF_TYPE_VIDEO_CAPTURE;
>  	dst_vq->io_modes = VB2_MMAP | VB2_DMABUF;
> +	src_vq->dma_attrs = DMA_ATTR_NO_KERNEL_MAPPING;

This should be dst_vq!

I'm also not quite sure that it is right to use this for the destination
buffer. Could this cause problems if this buffer is exported to a dmabuf
and another driver will be using it and requiring a kernel mapping.

Regards,

	Hans

>  	dst_vq->drv_priv = ctx;
>  	dst_vq->buf_struct_size = sizeof(struct cedrus_buffer);
>  	dst_vq->ops = &cedrus_qops;
> 




More information about the linux-arm-kernel mailing list