On Tue, May 19, 2026 at 07:47:48PM +0530, Aneesh Kumar K.V wrote:

Mostafa Saleh [off-list ref] writes:

quoted

On Tue, May 19, 2026 at 07:30:16PM +0530, Aneesh Kumar K.V wrote:

quoted

Mostafa Saleh [off-list ref] writes:

quoted

quoted

quoted

I am still running more tests, but looking more into it. Setting
force_dma_unencrypted() to true for pKVM guests is wrong, as the
guest shouldn’t try to decrypt arbitrary memory as it can include
sensitive information (for example in case of virtio sub-page
allocation) and should strictly rely on the restricted-dma-pool
for that.

??

Where does force_dma_unencrypted() cause arbitary memory passed into
the DMA API to be decrypted? That should never happen???

Sorry, maybe arbitrary is not the right expression again :)
I mean that, with emulated devices that use the DMA-API under pKVM,
they will map memory coming from other layers (VFS, net) through
vitrio-block, virtio-net... These can be smaller than a page, and

Don't we PAGE_ALIGN these requests?

dma_direct_alloc
	size = PAGE_ALIGN(size);

iommu_dma_alloc_pages
	size_t alloc_size = PAGE_ALIGN(size);

For allocation, yes, and that's fine because we bring memory from
the pool.
But not for mapping, as dma_direct_map_phys(), where the memory is
allocated from the driver or other parts in the kernel and the page
may be shared with other kernel components.

But if we are using restricted-dma-pool, we also have:

mem->force_bounce = true;
mem->for_alloc = true;

So, will we use the swiotlb buffers for mapping and copy only the shared
content into those swiotlb buffers?

True, that's why under pKVM, force_dma_unencrypted() should never
cause any memory to be decrypted and so we set it to false.
As in case of any bugs, the guest does not leak any information,
similar to what just happened initially here due to missing attrs.

However, as I mentioned to Jason, I think with some tweaks to
force_dma_unencrypted() we can make it work under pKVM for aligned
memory which eliminates some of the bouncing.
I am currently investigating that.

Thanks,
Mostafa

-aneesh