Re: [PATCH] media: videobuf2-dma-sg: limit the sg segment size

[Jason Gunthorpe]

On Wed, Aug 30, 2023 at 04:33:41PM +0200, Christoph Hellwig wrote:
> On Wed, Aug 30, 2023 at 12:47:57PM +0900, Tomasz Figa wrote:
> > Do we see anything replacing it widely anywhere on the short-middle
> > term horizon? I think we could possibly migrate vb2 to use that new
> > thing internally and just provide some compatibility X to scatterlist
> > conversion function for the drivers.
> 
> Jason said at LSF/MM that he had a prototype for a mapping API that
> takes a phys/len array as input and dma_addr/len a output, which really
> is the right thing to do, especially for dmabuf.

Yes, still a prototype. Given the change in direction some of the
assumptions of the list design will need some adjusting.

I felt there wasn't much justification to add a list without also
supporting the P2P and it was not looking very good to give the DMA
API proper p2p support without also connecting it to lists somehow.

Anyhow, I had drafted a basic list datastructure and starting
implementation that is sort of structured in away that is similar to
xarray (eg with fixed chunks, generic purpose, etc)

https://github.com/jgunthorpe/linux/commit/58d7e0578a09d9cd2360be515208bcd74ade5958

I would still call it an experiment as the auto-sizing entry approach
needs benchmarking to see what it costs.

> Jason, what's the status of your work?

After we talked I changed the order of the work, instead of starting
with the SG list side, I wanted to progress on the DMA API side and
then build any list infrastructure on that new API.

Your idea to use a new API that could allocate IOVA and then map phys
to IOVA as a DMA API entry point looked good to me, and it is a
perfect map for what RDMA's ODP stuff needs. So I changed direction to
rework ODP and introduce the DMA API parts as the first step.

I had to put it aside due to the volume of iommu/vfio stuff going on
right now. However, I think I will have someone who can work on the
ODP driver part this month

Regardless, RDMA really needs some kind of generic lists to hold CPU
and physical address ranges, so I still see that as being part of the
ultimate solution.

Jason