Re: [Linaro-mm-sig] [PATCH v5 2/6] dma-buf: udmabuf: Implement .{begin,end}_access

[Daniel Vetter <daniel@ffwll.ch>]

On Fri, Jan 19, 2024 at 03:13:58PM +0100, Paul Cercueil wrote:
> Implement .begin_access() and .end_access() callbacks.
> 
> For now these functions will simply sync/flush the CPU cache when
> needed.
> 
> Signed-off-by: Paul Cercueil <paul@crapouillou.net>
> 
> ---
> v5: New patch
> ---
>  drivers/dma-buf/udmabuf.c | 27 +++++++++++++++++++++++++++
>  1 file changed, 27 insertions(+)
> 
> diff --git a/drivers/dma-buf/udmabuf.c b/drivers/dma-buf/udmabuf.c
> index c40645999648..a87d89b58816 100644
> --- a/drivers/dma-buf/udmabuf.c
> +++ b/drivers/dma-buf/udmabuf.c
> @@ -179,6 +179,31 @@ static int end_cpu_udmabuf(struct dma_buf *buf,
>  	return 0;
>  }
>  
> +static int begin_udmabuf(struct dma_buf_attachment *attach,
> +			 struct sg_table *sgt,
> +			 enum dma_data_direction dir)
> +{
> +	struct dma_buf *buf = attach->dmabuf;
> +	struct udmabuf *ubuf = buf->priv;
> +	struct device *dev = ubuf->device->this_device;
> +
> +	dma_sync_sg_for_device(dev, sgt->sgl, sg_nents(sgt->sgl), dir);

So one thing I've just wondered is whether we've made sure that this is
only doing cache coherency maintenance, and not swiotlb bounce buffer
copying. The latter would really not be suitable for dma-buf anymore I
think.

Not sure how to best check for that since it's all in the depths of the
dma-api code, but I guess the best way to really make sure is to disable
CONFIG_SWIOTLB. Otherwise I guess the way to absolutely make sure is to
trace swiotlb_sync_single_for_device/cpu.

It would be kinda neat if dma-buf.c code could make sure you never ever
get an swiotlb entry from a dma_buf_map_attachment call, but I don't think
we can enforce that. There's sg_dma_is_swiotlb, but that won't catch all
implementations, only the generic dma-iommu.c one.

Cheers, Sima

> +	return 0;
> +}
> +
> +static int end_udmabuf(struct dma_buf_attachment *attach,
> +		       struct sg_table *sgt,
> +		       enum dma_data_direction dir)
> +{
> +	struct dma_buf *buf = attach->dmabuf;
> +	struct udmabuf *ubuf = buf->priv;
> +	struct device *dev = ubuf->device->this_device;
> +
> +	if (dir != DMA_TO_DEVICE)
> +		dma_sync_sg_for_cpu(dev, sgt->sgl, sg_nents(sgt->sgl), dir);
> +	return 0;
> +}
> +
>  static const struct dma_buf_ops udmabuf_ops = {
>  	.cache_sgt_mapping = true,
>  	.map_dma_buf	   = map_udmabuf,
> @@ -189,6 +214,8 @@ static const struct dma_buf_ops udmabuf_ops = {
>  	.vunmap		   = vunmap_udmabuf,
>  	.begin_cpu_access  = begin_cpu_udmabuf,
>  	.end_cpu_access    = end_cpu_udmabuf,
> +	.begin_access      = begin_udmabuf,
> +	.end_access        = end_udmabuf,
>  };
>  
>  #define SEALS_WANTED (F_SEAL_SHRINK)
> -- 
> 2.43.0
> 
> _______________________________________________
> Linaro-mm-sig mailing list -- linaro-mm-sig@lists.linaro.org
> To unsubscribe send an email to linaro-mm-sig-leave@lists.linaro.org

-- 
Daniel Vetter
Software Engineer, Intel Corporation
http://blog.ffwll.ch