Re: [BUG] iommu/dma: "Not yet supported" CC_SHARED rejection regresses dma_alloc_attrs() on SME hosts with IOMMU-attached GPUs (commit 8277a12d0d60)

[Mostafa Saleh <smostafa@google.com>]

On Tue, Sep 01, 2026 at 11:09:03PM +0200, Timo Witte wrote:
> Commit 8277a12d0d60 ("dma-pool: track decrypted atomic pools and select
> them via attrs") added to iommu_dma_alloc():
> 
>     /* Not yet supported */
>     if (is_alloc_cc_shared)
>         return NULL;
> 
> dma_alloc_attrs() auto-sets __DMA_ATTR_ALLOC_CC_SHARED whenever
> force_dma_unencrypted() is true. On an SME host (mem_encrypt=on) an AMD
> GPU with a 44-bit DMA mask hits that condition, so every GTT (system-RAM)
> allocation fails with NULL/-ENOMEM, amdgpu init aborts and the system
> boots without a display.
> IOMMU in Translated/DMA-FQ mode (iommu group 29).
> 
> Before the mentioned patch it worked: the IOMMU PTE carries the SME C-bit,
> so the memory controller transparently decrypts for the device while pages stay
> encrypted at rest. For IOMMU-backed devices force_dma_unencrypted() is a
> false positive - the device DMA mask does not limit addressing once the
> IOMMU translates, so CC_SHARED should not be auto-set for them.
> 
> When i remove the rejection and clear __DMA_ATTR_ALLOC_CC_SHARED
> in iommu_dma_alloc(), so encrypted pages are mapped with the C-bit and
> everything works again.
> 
> Maybe we can implement real CC_SHARED support in the IOMMU path,
> so SEV + IOMMU works on x86?

I believe that is the way forward, but as the merge window is closed
it won't be possible to land something that big in 7.3, we would
need a fix for the regression for now.

I think your suggestion works be clearing the __DMA_ATTR_ALLOC_CC_SHARED
from iommu_dma_alloc() but no need to remove the rejection then.

Otherwise, force_dma_unencrypted() can check something as
use_dma_iommu() but that feels like the wrong place, as the HW does
need to decrypt the memory and dma-iommu would be the one knowing
how to do that.

So, I'd suggest you send a patch with your fix.

Thanks,
Mostafa

> 
> I already reported this in the amdgpu driver:
> https://gitlab.freedesktop.org/drm/amd/-/work_items/5735
> But i think it should be fixed in the iommu code as it's a regression.
>