[PATCH v4 13/13] x86/amd-gart: preserve the direct DMA address until GART mapping succeeds

[Aneesh Kumar K.V]

"Aneesh Kumar K.V (Arm)" <aneesh.kumar at kernel.org> writes:

> gart_alloc_coherent() first allocates memory through dma_direct_alloc(),
> which returns a direct-mapped DMA address in dma_addr. When force_iommu is
> enabled, the buffer is then remapped.
>
> Do not overwrite dma_addr before dma_map_area() has succeeded. Keep the
> dma_map_area result in a temporary variable so the direct DMA address
> remains available for dma_direct_free() on the error path.
>
> Signed-off-by: Aneesh Kumar K.V (Arm) <aneesh.kumar at kernel.org>
> ---
>  arch/x86/kernel/amd_gart_64.c | 10 ++++++----
>  1 file changed, 6 insertions(+), 4 deletions(-)
>
> diff --git a/arch/x86/kernel/amd_gart_64.c b/arch/x86/kernel/amd_gart_64.c
> index b5f1f031d45b..a109649c5649 100644
> --- a/arch/x86/kernel/amd_gart_64.c
> +++ b/arch/x86/kernel/amd_gart_64.c
> @@ -467,18 +467,20 @@ gart_alloc_coherent(struct device *dev, size_t size, dma_addr_t *dma_addr,
>  		    gfp_t flag, unsigned long attrs)
>  {
>  	void *vaddr;
> +	dma_addr_t dma_map_addr;
>  
>  	vaddr = dma_direct_alloc(dev, size, dma_addr, flag, attrs);
>  	if (!vaddr ||
>  	    !force_iommu || dev->coherent_dma_mask <= DMA_BIT_MASK(24))
>  		return vaddr;
>  
> -	*dma_addr = dma_map_area(dev, virt_to_phys(vaddr), size,
> -				 DMA_BIDIRECTIONAL,
> -				 (1UL << get_order(size)) - 1, attrs);
> +	dma_map_addr = dma_map_area(dev, virt_to_phys(vaddr), size,
> +				     DMA_BIDIRECTIONAL,
> +				     (1UL << get_order(size)) - 1, attrs);
>  	flush_gart();
> -	if (unlikely(*dma_addr == DMA_MAPPING_ERROR))
> +	if (unlikely(dma_map_addr == DMA_MAPPING_ERROR))
>  		goto out_free;
> +	*dma_addr = dma_map_addr;
>  	return vaddr;
>  out_free:
>  	dma_direct_free(dev, size, vaddr, *dma_addr, attrs);
> -- 
> 2.43.0
>

This needs corresponding changes on the gart_free_coherent() side as well.

https://sashiko.dev/#/patchset/20260512090408.794195-1-aneesh.kumar%40kernel.org?part=13

I will avoid making that change as part of this series, since I assume
it would require specific testing.

-aneesh