Jason Gunthorpe [off-list ref] writes:

On Thu, Jul 09, 2026 at 12:13:19PM +0100, Catalin Marinas wrote:

quoted

quoted

quoted

For AMD/SME, on host with memory encryption we now end up setting the C
bit for DMA_ATTR_MMIO. This is fine for RAM but not sure whether
some other MMIO bus understands this attribute. Maybe we should stick to
something like __phys_to_dma() for the !CC_SHARED && MMIO path. Or,
since this is not universally defined, just use the old dma_addr = phys
if MMIO and ignore any unlikely DMA offsets.

Considering for AMD/SME system an unencrypted dma addr is one without C
bit, will this be good?

	/*
	 * For host memory encryption and device requiring unencrypted DMA,
	 * MMIO memory is treated as shared by default.
	 */
	if (attrs & DMA_ATTR_MMIO) {
		if (cc_platform_has(CC_ATTR_HOST_MEM_ENCRYPT) || force_dma_unencrypted(dev))
			attrs |= DMA_ATTR_CC_SHARED;
	}

Yes, I think it does the trick, preserves the current semantics for AMD.
I guess you could use a single 'if' for all checks (up to you).

Please don't change it, MMIO P2P is broken on CC systems today and it
should stay broken. Passing DMA_ATTR_MMIO with DMA_ATTR_CC_SHARED is
an error that we need to correct in the drivers not make work in the
core code.

But the above changes are intended to handle HOST_MEM_ENCRYPT. In v7, we
had the following diff:

@@ -88,37 +88,40 @@ static inline dma_addr_t dma_direct_map_phys(struct device *dev,
 {
 	dma_addr_t dma_addr;
 
+	/*
+	 * For a device requiring unencrypted DMA, MMIO memory is treated
+	 * as shared by default.
+	 */
+	if (force_dma_unencrypted(dev) && (attrs & DMA_ATTR_MMIO))
+		attrs |= DMA_ATTR_CC_SHARED;
+

This is now getting updated to

@@ -624,37 +626,44 @@ dma_addr_t dma_direct_map_phys(struct device *dev, phys_addr_t phys,
 {
 	dma_addr_t dma_addr;
 
+	if (attrs & DMA_ATTR_MMIO) {
+		/*
+		 * For host memory encryption and device requiring
+		 * unencrypted DMA, MMIO memory is treated as shared by
+		 * default.
+		 */
+		if (cc_platform_has(CC_ATTR_HOST_MEM_ENCRYPT) ||
+		    force_dma_unencrypted(dev))
+			attrs |= DMA_ATTR_CC_SHARED;
+	}
+

I agree that we need to move that DMA_ATTR_CC_SHARED setting to 

modified   drivers/pci/p2pdma.c

@@ -285,6 +285,11 @@ int pcim_p2pdma_init(struct pci_dev *pdev)
 			continue;
 
 		p2p->mem[i].owner = &pdev->dev;
+
+		p2p->mem[i].dma_mapping_flags = DMA_ATTR_MMIO;
+		if (force_dma_unencrypted(dev))
+			p2p->mem[i].dma_mapping_flags |= DMA_ATTR_CC_SHARED;
+

As we discussed [1], that can come in a later patch. In the meantime, adding
the HOST_MEM_ENCRYPT check preserves the previous behavior for SME.

[1] https://lore.kernel.org/all/20260522132240.GD7702@ziepe.ca/ (local)

-aneesh