[3/3] videobuf2: replace a layering violation with dma_map_resource

  On Mon, Jan 14, 2019 at 01:42:26PM +0100, Marek Szyprowski wrote:
> Hi Christoph,
> 
> On 2019-01-11 19:17, Christoph Hellwig wrote:
> > vb2_dc_get_userptr pokes into arm direct mapping details to get the
> > resemblance of a dma address for a a physical address that does is
> > not backed by a page struct.  Not only is this not portable to other
> > architectures with dma direct mapping offsets, but also not to uses
> > of IOMMUs of any kind.  Switch to the proper dma_map_resource /
> > dma_unmap_resource interface instead.
> >
> > Signed-off-by: Christoph Hellwig <hch@lst.de>
> 
> There are checks for IOMMU presence in other places in vb2-dma-contig,
> so it was used only when no IOMMU is available, but I agree that the
> hacky code should be replaced by a generic code if possible.
> 
> Tested-by: Marek Szyprowski <m.szyprowski@samsung.com>
> 
> V4L2 pipeline works fine on older Exynos-based boards with CMA and IOMMU
> disabled.

Do you know if these rely on the offsets?  E.g. would they still work
with the patch below applied on top.  That would keep the map_resource
semantics as-is as solve the issue pointed out by Robin for now.

If not I can only think of a flag to bypass the offseting for now, but
that would be pretty ugly.  Or go for the long-term solution of
discovering the relationship between the two devices, as done by the
PCIe P2P code..
  

Hi Christoph,

On 2019-01-17 18:21, Christoph Hellwig wrote:
> On Mon, Jan 14, 2019 at 01:42:26PM +0100, Marek Szyprowski wrote:
>> On 2019-01-11 19:17, Christoph Hellwig wrote:
>>> vb2_dc_get_userptr pokes into arm direct mapping details to get the
>>> resemblance of a dma address for a a physical address that does is
>>> not backed by a page struct.  Not only is this not portable to other
>>> architectures with dma direct mapping offsets, but also not to uses
>>> of IOMMUs of any kind.  Switch to the proper dma_map_resource /
>>> dma_unmap_resource interface instead.
>>>
>>> Signed-off-by: Christoph Hellwig <hch@lst.de>
>> There are checks for IOMMU presence in other places in vb2-dma-contig,
>> so it was used only when no IOMMU is available, but I agree that the
>> hacky code should be replaced by a generic code if possible.
>>
>> Tested-by: Marek Szyprowski <m.szyprowski@samsung.com>
>>
>> V4L2 pipeline works fine on older Exynos-based boards with CMA and IOMMU
>> disabled.
> Do you know if these rely on the offsets?  E.g. would they still work
> with the patch below applied on top.  That would keep the map_resource
> semantics as-is as solve the issue pointed out by Robin for now.

AFAIK that code was only used for sharing buffers between hardware
modules that are a part of the same SoC, usually implemented as platform
devices. AFAIR this never worked for devices on different buses. So far
I wasn't aware on ANY which would require an offset for the DMA access.

The first version of videobuf2-dc code even incorrectly used paddr
instead of dma_addr as a buffer 'address' returned to the client
drivers, because in case of those SoC this was exactly the same (see
commits 472af2b05bdefcaee7e754e22cbf131110017ad6 and
ba7fcb0c954921534707f08ebc4d8beeb2eb17e7).

> If not I can only think of a flag to bypass the offseting for now, but
> that would be pretty ugly.  Or go for the long-term solution of
> discovering the relationship between the two devices, as done by the
> PCIe P2P code..
>
> diff --git a/kernel/dma/direct.c b/kernel/dma/direct.c
> index 8e0359b04957..25bd19974223 100644
> --- a/kernel/dma/direct.c
> +++ b/kernel/dma/direct.c
> @@ -359,7 +359,7 @@ EXPORT_SYMBOL(dma_direct_map_sg);
>  dma_addr_t dma_direct_map_resource(struct device *dev, phys_addr_t paddr,
>  		size_t size, enum dma_data_direction dir, unsigned long attrs)
>  {
> -	dma_addr_t dma_addr = phys_to_dma(dev, paddr);
> +	dma_addr_t dma_addr = paddr;
>  
>  	if (unlikely(!dma_direct_possible(dev, dma_addr, size))) {
>  		report_addr(dev, dma_addr, size);
>
>
>
Best regards
  

Hi Marek,

can chance you could retest the v2 version?