Re: [PATCH rdma-next v4 15/16] RDMA/mlx5: Use UMEM attribute for CQ doorbell record

[Jason Gunthorpe <jgg@ziepe.ca>]

On Thu, May 07, 2026 at 02:52:30PM +0200, Jiri Pirko wrote:
> +int mlx5_ib_db_map_user(struct mlx5_ib_ucontext *context,
> +			struct ib_udata *udata, u16 attr_id,
> +			unsigned long virt, struct mlx5_db *db)
>  {
> -	struct mlx5_ib_user_db_page *page;
> +	struct mlx5_ib_user_db_page *page = NULL;
> +	unsigned long dma_offset;
>  	int err = 0;
>  
> +	if (udata) {
> +		struct ib_umem *umem;
> +
> +		umem = ib_umem_get_attr(context->ibucontext.device, udata,
> +					attr_id, sizeof(__be32) * 2, 0);
> +		if (IS_ERR(umem))
> +			return PTR_ERR(umem);
> +		if (umem) {

More IS_ERR_OR_NULL stuff..

> +			/*
> +			 * The 8-byte DBR is programmed to the device as one
> +			 * DMA address, so it must stay within a single page.
> +			 * An 8-byte range that crosses a page boundary may
> +			 * be split across two non-contiguous DMA mappings.
> +			 */
> +			if (ib_umem_offset(umem) >
> +			    PAGE_SIZE - sizeof(__be32) * 2) {
> +				ib_umem_release(umem);
> +				return -EINVAL;
> +			}

I think this can just be ib_umem_is_contiguous()

> +			page = kzalloc_obj(*page);
> +			if (!page) {
> +				ib_umem_release(umem);
> +				return -ENOMEM;
> +			}
> +			page->umem = umem;
> +			dma_offset = ib_umem_offset(umem);
> +		}
> +	}
> +
>  	mutex_lock(&context->db_page_mutex);
>  
> +	if (page)
> +		goto add_page;
> +
> +	dma_offset = virt & ~PAGE_MASK;
> +
>  	list_for_each_entry(page, &context->db_page_list, list)
>  		if ((current->mm == page->mm) &&
>  		    (page->user_virt == (virt & PAGE_MASK)))
>  			goto found;

Ah, this is why..

I think this function should take in a ib_uverbs_buffer_desc and that
should be stored inside the page instead of using virt. Then the
functions flow doesn't really change, it searchs the page_list for a
matching desc, otherwise it converts the desc to a umem and creates a
new page. Refuse to match FD based descs

That sort of suggests you want to split up the earlier patch so there
is a seperate re-usable get desc function instead of bundling into
ib_umem_get..


Jason