On 28.10.2019 13:38, [email protected] wrote:
> On Mon, Oct 28, 2019 at 10:55:05AM +0000, Laurentiu Tudor wrote:
>>>> @@ -85,9 +75,10 @@ static void free_rx_fd(struct dpaa2_eth_priv *priv,
>>>>       sgt = vaddr + dpaa2_fd_get_offset(fd);
>>>>       for (i = 1; i < DPAA2_ETH_MAX_SG_ENTRIES; i++) {
>>>>           addr = dpaa2_sg_get_addr(&sgt[i]);
>>>> -        sg_vaddr = dpaa2_iova_to_virt(priv->iommu_domain, addr);
>>>> -        dma_unmap_page(dev, addr, DPAA2_ETH_RX_BUF_SIZE,
>>>> -                   DMA_BIDIRECTIONAL);
>>>> +        sg_vaddr = page_to_virt
>>>> +                (dma_unmap_page_desc(dev, addr,
>>>> +                            DPAA2_ETH_RX_BUF_SIZE,
>>>> +                            DMA_BIDIRECTIONAL));
>>>
>>> This is doing virt -> page -> virt.  Why not just have the new
>>> function return the VA corresponding to the addr, which would
>>> match the other functions?
>>
>> I'd really like that as it would get rid of the page_to_virt() calls but
>> it will break the symmetry with the dma_map_page() API. I'll let the
>> maintainers decide.
> 
> It would be symmetric with dma_map_single, though.  Maybe we need
> both variants?

Patch 1/3 also adds an dma_unmap_single_desc(). Would it be legal to 
just use dma_unmap_single_desc() in the driver even if the driver does 
it's mappings with dma_map_page()?

---
Best Regards, Laurentiu
_______________________________________________
iommu mailing list
[email protected]
https://lists.linuxfoundation.org/mailman/listinfo/iommu

Reply via email to