[PATCH] dma-direct: Restore arch_dma_alloc() for the DMA_ATTR_NO_KERNEL_MAPPING case

[Marek Szyprowski]

On 30.09.2026 15:57, Marek Szyprowski wrote:
> Commit 849facea92fa ("dma-direct: simplify the DMA_ATTR_NO_KERNEL_MAPPING
> handling") moved the DMA_ATTR_NO_KERNEL_MAPPING shortcut in
> dma_direct_alloc() in front of the arch_dma_alloc() call. Architectures
> providing their own DMA coherent allocator (CONFIG_ARCH_HAS_DMA_ALLOC)
> therefore no longer see such requests. ARM 32bit is one of such. Lack of
> the call to arch_dma_alloc() results in the lack of proper cache
> management on HighMem pages.
>
> Take the dma-direct shortcut only for coherent devices or when the
> architecture doesn't provide arch_dma_alloc(). While touching this,
> move that check into separate function to avoid duplication in alloc
> and free paths.
>
> Fixes: 849facea92fa ("dma-direct: simplify the DMA_ATTR_NO_KERNEL_MAPPING handling")
> Reported-by: Aneesh Kumar K.V <aneesh.kumar at kernel.org>
> Closes: https://lore.kernel.org/all/yq5ams2akznl.fsf@kernel.org/
> Assisted-by: Claude Opus 5.5 <noreply at anthropic.com>
> Signed-off-by: Marek Szyprowski <m.szyprowski at samsung.com>
Applied to dma-mapping-for-next, because it looks that this case has no
active users, so it is not the urgent fix.

> ---
>  kernel/dma/direct.c | 22 ++++++++++++++++++----
>  1 file changed, 18 insertions(+), 4 deletions(-)
>
> diff --git a/kernel/dma/direct.c b/kernel/dma/direct.c
> index da665ca22d5c..a31bd3129f3f 100644
> --- a/kernel/dma/direct.c
> +++ b/kernel/dma/direct.c
> @@ -179,6 +179,22 @@ static struct page *dma_direct_alloc_from_pool(struct device *dev, size_t size,
>  	return page;
>  }
>  
> +/*
> + * Check if the allocation can skip the kernel mapping and return the page
> + * pointer as the cookie.
> + */
> +static bool dma_direct_use_no_mapping(struct device *dev, unsigned long attrs)
> +{
> +	if ((attrs & (DMA_ATTR_NO_KERNEL_MAPPING | __DMA_ATTR_ALLOC_CC_SHARED)) !=
> +	    DMA_ATTR_NO_KERNEL_MAPPING)
> +		return false;
> +	if (is_swiotlb_for_alloc(dev))
> +		return false;
> +	if (IS_ENABLED(CONFIG_ARCH_HAS_DMA_ALLOC) && !dev_is_dma_coherent(dev))
> +		return false;
> +	return true;
> +}
> +
>  static void *dma_direct_alloc_no_mapping(struct device *dev, size_t size,
>  		dma_addr_t *dma_handle, gfp_t gfp)
>  {
> @@ -226,8 +242,7 @@ void *dma_direct_alloc(struct device *dev, size_t size,
>  	if (attrs & DMA_ATTR_NO_WARN)
>  		gfp |= __GFP_NOWARN;
>  
> -	if (((attrs & (DMA_ATTR_NO_KERNEL_MAPPING | __DMA_ATTR_ALLOC_CC_SHARED)) ==
> -	     DMA_ATTR_NO_KERNEL_MAPPING) && !is_swiotlb_for_alloc(dev))
> +	if (dma_direct_use_no_mapping(dev, attrs))
>  		return dma_direct_alloc_no_mapping(dev, size, dma_handle, gfp);
>  
>  	if (!dev_is_dma_coherent(dev)) {
> @@ -374,8 +389,7 @@ void dma_direct_free(struct device *dev, size_t size,
>  	if (attrs & __DMA_ATTR_ALLOC_CC_SHARED)
>  		mark_mem_encrypted = true;
>  
> -	if (((attrs & (DMA_ATTR_NO_KERNEL_MAPPING | __DMA_ATTR_ALLOC_CC_SHARED)) ==
> -	     DMA_ATTR_NO_KERNEL_MAPPING) && !is_swiotlb_for_alloc(dev)) {
> +	if (dma_direct_use_no_mapping(dev, attrs)) {
>  		/* cpu_addr is a struct page cookie, not a kernel address */
>  		dma_free_contiguous(dev, cpu_addr, size);
>  		return;

Best regards
-- 
Marek Szyprowski, PhD
Samsung R&D Institute Poland