[PATCH v2 04/13] gpu: host1x: Use direct DMA with IOMMU API usage

[Dmitry Osipenko]

24.01.2019 21:02, Thierry Reding пишет:
> From: Thierry Reding <treding at nvidia.com>
> 
> If we use the IOMMU API directly to map buffers into host1x' IOVA space,
> we must make sure that the DMA API doesn't already set up a mapping, or
> else translation will fail.
> 
> The direct DMA API allows us to allocate memory that will not be mapped
> through an IOMMU automatically.
> 
> Signed-off-by: Thierry Reding <treding at nvidia.com>
> ---
>  drivers/gpu/host1x/cdma.c | 11 ++++++++---
>  1 file changed, 8 insertions(+), 3 deletions(-)
> 
> diff --git a/drivers/gpu/host1x/cdma.c b/drivers/gpu/host1x/cdma.c
> index 91df51e631b2..ccde23a0109c 100644
> --- a/drivers/gpu/host1x/cdma.c
> +++ b/drivers/gpu/host1x/cdma.c
> @@ -19,6 +19,7 @@
>  
>  #include <asm/cacheflush.h>
>  #include <linux/device.h>
> +#include <linux/dma-direct.h>
>  #include <linux/dma-mapping.h>
>  #include <linux/host1x.h>
>  #include <linux/interrupt.h>
> @@ -70,6 +71,7 @@ static void host1x_pushbuffer_destroy(struct push_buffer *pb)
>   */
>  static int host1x_pushbuffer_init(struct push_buffer *pb)
>  {
> +	unsigned long attrs = DMA_ATTR_WRITE_COMBINE;
>  	struct host1x_cdma *cdma = pb_to_cdma(pb);
>  	struct host1x *host1x = cdma_to_host1x(cdma);
>  	struct iova *alloc;
> @@ -91,8 +93,8 @@ static int host1x_pushbuffer_init(struct push_buffer *pb)
>  
>  		size = iova_align(&host1x->iova, size);
>  
> -		pb->mapped = dma_alloc_wc(host1x->dev, size, &pb->phys,
> -					  GFP_KERNEL);
> +		pb->mapped = dma_direct_alloc(host1x->dev, size, &pb->phys,
> +					      GFP_KERNEL, attrs);
>  		if (!pb->mapped)
>  			return -ENOMEM;
>  
> @@ -127,7 +129,10 @@ static int host1x_pushbuffer_init(struct push_buffer *pb)
>  iommu_free_iova:
>  	__free_iova(&host1x->iova, alloc);
>  iommu_free_mem:
> -	dma_free_wc(host1x->dev, size, pb->mapped, pb->phys);
> +	if (host1x->domain)
> +		dma_direct_free(host1x->dev, size, pb->mapped, pb->phys, attrs);
> +	else
> +		dma_free_wc(host1x->dev, size, pb->mapped, pb->phys);
>  
>  	return err;
>  }
> 

This variant looks good for the current upstream code. Alternatively we could check the domain->type==IOMMU_DOMAIN_DMA and skip IOMMU initialization on probe.

I may test that nothing breaks with this series for older Tegra's once it all will appear in -next. Please address Mikko's review-comments and push the patches once ready.

Reviewed-by: Dmitry Osipenko <digetx at gmail.com>