Re: [RESEND PATCH 1/2] dma/pool: Use vmap() address for memory encryption helpers on ARM64

[Catalin Marinas <catalin.marinas@arm.com>]

On Sun, Aug 10, 2025 at 07:50:34PM -0500, Shanker Donthineni wrote:
> In atomic_pool_expand(), set_memory_encrypted()/set_memory_decrypted()
> are currently called with page_to_virt(page). On ARM64 with
> CONFIG_DMA_DIRECT_REMAP=y, the atomic pool is mapped via vmap(), so
> page_to_virt(page) does not reference the actual mapped region.
> 
> Using this incorrect address can cause encryption attribute updates to
> be applied to the wrong memory region. On ARM64 systems with memory
> encryption enabled (e.g. CCA), this can lead to data corruption or
> crashes.
> 
> Fix this by using the vmap() address ('addr') on ARM64 when invoking
> the memory encryption helpers, while retaining the existing
> page_to_virt(page) usage for other architectures.
> 
> Fixes: 76a19940bd62 ("dma-direct: atomic allocations must come from atomic coherent pools")
> Signed-off-by: Shanker Donthineni <sdonthineni@nvidia.com>
> ---
>  kernel/dma/pool.c | 8 ++++----
>  1 file changed, 4 insertions(+), 4 deletions(-)
> 
> diff --git a/kernel/dma/pool.c b/kernel/dma/pool.c
> index 7b04f7575796b..ba08a301590fd 100644
> --- a/kernel/dma/pool.c
> +++ b/kernel/dma/pool.c
> @@ -81,6 +81,7 @@ static int atomic_pool_expand(struct gen_pool *pool, size_t pool_size,
>  {
>  	unsigned int order;
>  	struct page *page = NULL;
> +	void *vaddr;
>  	void *addr;
>  	int ret = -ENOMEM;
>  
> @@ -113,8 +114,8 @@ static int atomic_pool_expand(struct gen_pool *pool, size_t pool_size,
>  	 * Memory in the atomic DMA pools must be unencrypted, the pools do not
>  	 * shrink so no re-encryption occurs in dma_direct_free().
>  	 */
> -	ret = set_memory_decrypted((unsigned long)page_to_virt(page),
> -				   1 << order);
> +	vaddr = IS_ENABLED(CONFIG_ARM64) ? addr : page_to_virt(page);
> +	ret = set_memory_decrypted((unsigned long)vaddr, 1 << order);

At least with arm CCA, there are two aspects to setting the memory
encrypted/decrypted: an RMM (realm management monitor) call and setting
of the attributes of the stage 1 mapping. The RMM call doesn't care
about the virtual address, only the (intermediate) physical address, so
having page_to_virt(page) here is fine.

The second part is setting the (fake) attribute for this mapping (top
bit of the IPA space). Can we not instead just call:

	addr = dma_common_contiguous_remap(page, pool_size,
					   pgprot_decrypted(pgprot_dmacoherent(PAGE_KERNEL)),
					   __builtin_return_address(0));

in the atomic pool code? The advantage is that we keep the
set_memory_decrypted() call on the linear map so that we change its
attributes as well.

I want avoid walking the page tables for vmap regions if possible in the
arm64 set_memory_* implementation. At some point I was proposing a
GFP_DECRYPTED flag for allocations but never got around to post a patch
(and implement vmalloc() support):

https://lore.kernel.org/linux-arm-kernel/ZmNJdSxSz-sYpVgI@arm.com/

-- 
Catalin