Kernel and ADM hardware roulette ( was AMD graphics performance regression in 4.15 and later )

[Christoph Hellwig]

I think the prime issue is that dma_direct_alloc respects the dma
mask.  Which we don't need if actually using the iommu.  This would
be mostly harmless exept for the the SEV bit high in the address that
makes the checks fail.

For now I'd say revert this commit for 4.17/4.18-rc and I'll look into
addressing these issues properly.