Re: [PATCH v2 1/3] media/videobuf2-dma-sg: Fix handling of sg_table structure

From: Ricardo Ribalda Delgado
Date: Fri Feb 13 2015 - 10:20:45 EST

Hello Hans

On Fri, Feb 13, 2015 at 4:02 PM, Hans Verkuil <hverkuil@xxxxxxxxx> wrote:
> Hi Ricardo, Marek,
> I have a few questions, mostly to improve my own understanding.
> First of all, is this solving an actual bug for you, or did you just find
> it while reviewing code? And if it solves a bug, then which architecture
> are you using? ARM? Intel?

My arch is intel based (AMD APU). I found it while doing review. While
updating our kernel to 3.19 I had to patch some of my out of tree
drivers, and then I gave a look to the file.

>> dma_map_sg returns the number of areas mapped by the hardware,
>> which could be different than the areas given as an input.
>> The output must be saved to nent.
>> The output of dma_map, should be used to transverse the scatter list.
>> dma_unmap_sg needs the value passed to dma_map_sg (nents_orig).
> I noticed that few dma_unmap_sg calls actually use orig_nents. It makes
> me wonder if the dma_unmap_sg documentation is actually correct. It does
> clearly state that orig_nents should be used, and it might well be that
> the only reason this hasn't led to problems is that very few architectures
> actually seem to return nents < orig_nents.

It is not the most clear API to use :(. Some of the prototypes do not
make a lot of sense, and it is documented outside the code.

I have sent these two patches:

>> + sgt->nents = dma_map_sg_attrs(buf->dev, sgt->sgl, sgt->orig_nents,
>> + buf->dma_dir, &attrs);
> Is a driver free to change sgt->nents? It's unclear from the documentation
> or code that that is actually the purpose of sgt->nents. Most drivers seem
> to store the result of dma_map_sg into a driver-specific struct.

As I understand it, this is the purpose of the struct scatter list,
have at hand the three values that you need,
the sgl, nents and orig_ents.

But it would be great if the maintaner of the dma-api speaks up :)

I am putting in cc

Thanks Hans!
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at
Please read the FAQ at