On Thu, Oct 15, 2020 at 07:55:32AM +0100, Christoph Hellwig wrote:

On Tue, Oct 13, 2020 at 02:42:38PM +0100, Robin Murphy wrote:

...

I still think this situation would be best handled with a variant of dma_ops_bypass that also guarantees to bypass SWIOTLB, and can be set automatically when attaching to an unmanaged IOMMU domain.

dma_ops_bypass should mostly do the right thing as-is. swiotlb bouncing is triggered of two things:

1. the dma_mask. This is under control of the driver, and obviously if it is too small for a legit reason we can't just proceed

Somewhat related, but is there a way to tell the dma-api to fail instead of falling back to swiotlb? In many case for gpu drivers it's much better if we fall back to dma_alloc_coherent and manage the copying ourselves instead of abstracting this away in the dma-api. Currently that's "solved" rather pessimistically by always allocating from dma_alloc_coherent if swiotlb could be in the picture (at least for ttm based drivers, i915 just falls over). -Daniel

2. force_dma_unencrypted() - we'd need to do an opt-out here, either by a flag or by being smart and looking for an attached iommu on the device

...

That way the device driver can make DMA API calls in the appropriate places that do the right thing either way, and only needs logic to decide whether to use the returned DMA addresses directly or ignore them if it knows they're overridden by its own IOMMU mapping.

I'd be happy to review patches for this.