On Tue, Jul 09, 2024 at 08:17:21AM +0200, Christoph Hellwig wrote: > On Mon, Jul 08, 2024 at 01:52:38PM -0300, Jason Gunthorpe wrote: > > Ideally we'd have some template code that consolidates these loops to > > common code with driver provided hooks - there are a few ways to get > > that efficiently in C. > > > > I think it will be clearer when we get to RDMA and there we have the > > same SGL/PRP kind of split up and we can see what is sharable. > > I really would not want to build common code for PRPs - this is a concept > very specific to RDMA and NVMe. I think DRM has it too. If you are populating a GPU page table then it is basically a convoluted PRP. Probably requires different splitting logic than what RDMA does, but I've never looked. > OTOH more common code SGLs would be nice. If you look at e.g. SCSI > drivers most of them have a simpe loop of mapping the SG table and > then copying the fields into the hardware SGL. This would be a very > common case for a helper. Yes, I belive this is very common. > That whole thing of course opens the question if we want a pure > in-memory version of the dma_addr_t/len tuple. IMHO that is the best > way to migrate and allows to share code easily. We can look into ways > to avoiding that more for drivers that care, but most drivers are > probably best serve with it to keep the code simple and make the > conversion easier. My feeling has been that this RFC is the low level interface and we can bring our own data structure on top. It would probably make sense to build a scatterlist v2 on top of this that has an in-memory dma_addr_t/len list close to today. Yes it costs a memory allocation, or a larger initial allocation, but many places may not really care. Block drivers have always allocated a SGL, for instance. Then the verbosity of this API is less important as we may only use it in a few places. My main take away was that we should make the dma_ops interface simpler and more general so we can have this choice instead of welding a single datastructure through everything. > > I'm also cooking something that should let us build a way to iommu map > > a bio_vec very efficiently, which should transform this into a single > > indirect call into the iommu driver per bio_vec, and a single radix > > walk/etc. > > I assume you mean array of bio_vecs here. That would indeed nice. > We'd still potentially need a few calls for block drivers as > requests can have multiple bios and thus bio_vec arrays, but it would > still be a nice reduction of calls. Yes. iommufd has performance needs here, not sure what it will turn into but bio_vec[] direct to optimized radix manipuilation is something I'd be keen to see. Jason
