Re: [RFC PATCH v1 18/18] nvme-pci: use new dma API

[Robin Murphy]

On 02/07/2024 10:09 am, Leon Romanovsky wrote:
[...]
> +static inline dma_addr_t nvme_dma_link_page(struct page *page,
> +					   unsigned int poffset,
> +					   unsigned int len,
> +					   struct nvme_iod *iod)
>   {
> -	int i;
> -	struct scatterlist *sg;
> +	struct dma_iova_attrs *iova = &iod->dma_map->iova;
> +	struct dma_iova_state *state = &iod->dma_map->state;
> +	dma_addr_t dma_addr;
> +	int ret;
> +
> +	if (iod->dma_map->use_iova) {
> +		phys_addr_t phys = page_to_phys(page) + poffset;

Yeah, there's no way this can possibly work. You can't do the 
dev_use_swiotlb() check up-front based on some overall DMA operation 
size, but then build that operation out of arbitrarily small fragments 
of different physical pages that *could* individually need bouncing to 
not break coherency.

Thanks,
Robin.

> +
> +		dma_addr = state->iova->addr + state->range_size;
> +		ret = dma_link_range(&iod->dma_map->state, phys, len);
> +		if (ret)
> +			return DMA_MAPPING_ERROR;
> +	} else {
> +		dma_addr = dma_map_page_attrs(iova->dev, page, poffset, len,
> +					      iova->dir, iova->attrs);
> +	}
> +	return dma_addr;
> +}