Jason Gunthorpe [off-list ref] writes:

On Thu, Jun 04, 2026 at 02:09:43PM +0530, Aneesh Kumar K.V (Arm) wrote:

quoted

 struct page *dma_alloc_from_pool(struct device *dev, size_t size,
-		void **cpu_addr, gfp_t gfp,
+		void **cpu_addr, gfp_t gfp, unsigned long attrs,
 		bool (*phys_addr_ok)(struct device *, phys_addr_t, size_t))
 {
-	struct gen_pool *pool = NULL;
+	struct dma_gen_pool *dma_pool = NULL;
 	struct page *page;
 	bool pool_found = false;
 
-	while ((pool = dma_guess_pool(pool, gfp))) {
+	while ((dma_pool = dma_guess_pool(dma_pool, gfp))) {
+
+		if (dma_pool->unencrypted != !!(attrs & DMA_ATTR_CC_SHARED))
+			continue;

I don't think you should be overloading DMA_ATTR_CC_SHARED like this.

	/*
	 * DMA_ATTR_CC_SHARED is not a caller-visible dma_alloc_*()
	 * attribute. The direct allocator uses it internally after it has
	 * decided that the backing pages must be shared/decrypted, so the
	 * rest of the allocation path can consistently select DMA addresses,
	 * choose compatible pools and restore encryption on free.
	 */
	if (attrs & DMA_ATTR_CC_SHARED)
		return NULL;

	if (force_dma_unencrypted(dev)) {
		attrs |= DMA_ATTR_CC_SHARED;
		mark_mem_decrypt = true;
	}

It is fine to have a bit inside the attrs that is only used by the
internal logic, but it needs to have a clearer name
__DMA_ATTR_REQUIRE_CC_SHARED perhaps.

Are you suggesting adding another attribute in addition to
DMA_ATTR_CC_SHARED?

Is the idea that __DMA_ATTR_REQUIRE_CC_SHARED would be used in the
allocation path to request a CC_SHARED allocation, while
DMA_ATTR_CC_SHARED would be used in the mapping path to describe the
attribute of the address?

The sashiko note does look legit though:

	if (IS_ENABLED(CONFIG_DMA_DIRECT_REMAP) &&
	    !gfpflags_allow_blocking(gfp) && !coherent) {
		page = dma_alloc_from_pool(dev, PAGE_ALIGN(size), &cpu_addr,
					   gfp, attrs, NULL);
		if (!page)
			return NULL;

I don't see anything doing the force_dma_unencrypted test along this
callchain..

I guess it should be done one step up in dma_alloc_attrs() instead of
in dma_direct_alloc()?

Yes, I'll move it there.

-aneesh