When running Xen on iMX8QXP, an Arm SoC without IOMMU, DMA performed via
its eDMA v3 DMA engine fail with a mapping error.

The eDMA performs DMA between RAM and MMIO space, and it's the MMIO side
that cannot be mapped.

MMIO->RAM DMA access cannot be bounce buffered if it would straddle a page
boundary and on Xen the MMIO space is 1:1 mapped for Arm, and x86 PV Dom0.
Cases where MMIO space is not 1:1 mapped, such as x86 PVH Dom0, requires an
IOMMU present to deal with the mapping.

Considering the above the map_resource callback can just be wired to the
existing dma_direct_map_resource() function.

There is nothing to do for unmap so the unmap callback is not needed.

Signed-off-by: John Ernberg <john.ernberg@actia.se>

---

v2:
 - Just use dma_direct_map_resource() directly (Stefano Stabellini)
 - Incorporate some of the discussion and explanations from v1 into the
    commit message (Stefano, Christoph)
 - General English improvements in the commit message.
 - Just this patch now, 1/2 in the previous set was applied

v1: (series) https://lore.kernel.org/xen-devel/20250502114043.1968976-1-john.ernberg@actia.se/
---
 drivers/xen/swiotlb-xen.c | 1 +
 1 file changed, 1 insertion(+)

diff --git a/drivers/xen/swiotlb-xen.c b/drivers/xen/swiotlb-xen.c
index ef56a2500ed6..da1a7d3d377c 100644
--- a/drivers/xen/swiotlb-xen.c
+++ b/drivers/xen/swiotlb-xen.c
@@ -426,4 +426,5 @@ const struct dma_map_ops xen_swiotlb_dma_ops = {
 	.alloc_pages_op = dma_common_alloc_pages,
 	.free_pages = dma_common_free_pages,
 	.max_mapping_size = swiotlb_max_mapping_size,
+	.map_resource = dma_direct_map_resource,
 };
-- 
2.49.0

On Mon, 12 May 2025, John Ernberg wrote:
> When running Xen on iMX8QXP, an Arm SoC without IOMMU, DMA performed via
> its eDMA v3 DMA engine fail with a mapping error.
> 
> The eDMA performs DMA between RAM and MMIO space, and it's the MMIO side
> that cannot be mapped.
> 
> MMIO->RAM DMA access cannot be bounce buffered if it would straddle a page
> boundary and on Xen the MMIO space is 1:1 mapped for Arm, and x86 PV Dom0.
> Cases where MMIO space is not 1:1 mapped, such as x86 PVH Dom0, requires an
> IOMMU present to deal with the mapping.
> 
> Considering the above the map_resource callback can just be wired to the
> existing dma_direct_map_resource() function.
> 
> There is nothing to do for unmap so the unmap callback is not needed.
> 
> Signed-off-by: John Ernberg <john.ernberg@actia.se>

Reviewed-by: Stefano Stabellini <sstabellini@kernel.org>


> ---
> 
> v2:
>  - Just use dma_direct_map_resource() directly (Stefano Stabellini)
>  - Incorporate some of the discussion and explanations from v1 into the
>     commit message (Stefano, Christoph)
>  - General English improvements in the commit message.
>  - Just this patch now, 1/2 in the previous set was applied
> 
> v1: (series) https://lore.kernel.org/xen-devel/20250502114043.1968976-1-john.ernberg@actia.se/
> ---
>  drivers/xen/swiotlb-xen.c | 1 +
>  1 file changed, 1 insertion(+)
> 
> diff --git a/drivers/xen/swiotlb-xen.c b/drivers/xen/swiotlb-xen.c
> index ef56a2500ed6..da1a7d3d377c 100644
> --- a/drivers/xen/swiotlb-xen.c
> +++ b/drivers/xen/swiotlb-xen.c
> @@ -426,4 +426,5 @@ const struct dma_map_ops xen_swiotlb_dma_ops = {
>  	.alloc_pages_op = dma_common_alloc_pages,
>  	.free_pages = dma_common_free_pages,
>  	.max_mapping_size = swiotlb_max_mapping_size,
> +	.map_resource = dma_direct_map_resource,
>  };
> -- 
> 2.49.0
>