On 10/9/26 14:42, Jason Gunthorpe wrote:

On Mon, Oct 05, 2026 at 08:41:32AM +0200, Karl Mehltretter wrote:

...

static inline struct page *sg_page(struct scatterlist *sg) { #ifdef CONFIG_DEBUG_SG BUG_ON(sg_is_chain(sg)); #endif

• sg_dmabuf_cpu_access_check(sg); return (struct page *)((sg)->page_link & ~SG_PAGE_LINK_MASK);

}

I'm not sure I understand the overall intention here, I get what this patch does, but no distro could turn this on by default when it touches *everyone* using scatterlist in a performance sensitive spot, and anyone doing testing can use the existing option - so what is the point?

Yeah, I was just about to write something similar.

Patch #1 in the series is correct, but I would add a warning if SWIOTLB is used since that usually doesn't work either.

I certainly don't like this patch, and I don't like the word "DMABUF" in the scatterlist at all. If we want to add something it should be a general mechanism under DEBUG_SG that allows anyone to 'hide' the CPU list from any future access.

As Christoph also noted the fundamental problem with sg_tables is that we stuffed the input and output parameters to the DMA mapping function into one data structure.

So by using a sg_table DMA-buf accidentally gave importers access to the physical address as well with all the undesired consequences.

I think when anybody wants to address this mess the first thing which needs to be done is to have a container for DMA addresses which is independent of sg_tables.

Regards, Christian.

Jason