Re: [RFC PATCH v3 24/37] kvx: Add memory management

[Christoph Hellwig]

> +#include "../../../drivers/iommu/dma-iommu.h"

This is not a public header as you can guess from the file path.

> +	switch (dir) {
> +	case DMA_TO_DEVICE:
> +		break;
> +	case DMA_FROM_DEVICE:
> +		break;
> +
> +	case DMA_BIDIRECTIONAL:
> +		inval_dcache_range(paddr, size);

Doing this just for bidirectional is weird unless your architecture
never does any speculative prefetching.  Other architectures
include DMA_FROM_DEVICE here.

> +#ifdef CONFIG_IOMMU_DMA
> +void arch_teardown_dma_ops(struct device *dev)
> +{
> +	dev->dma_ops = NULL;
> +}
> +#endif /* CONFIG_IOMMU_DMA*/

This should not be needed right now.  And will be completley
useless once we do the direct calls to dma-iommu which we plan
to do for Linux 6.12.

> +void arch_setup_dma_ops(struct device *dev, bool coherent)
> +{
> +	dev->dma_coherent = coherent;
> +	if (device_iommu_mapped(dev))
> +		iommu_setup_dma_ops(dev);
> +}

And this seems odd, as iommu_setup_dma_ops is called from the iommu
code and you shouldn't need it here.

I also wonder if we can come up with a way to do the ->dma_coherent
setup in common code and remove a few of these arch hooks entirely.