On Thu, Jul 04, 2024 at 09:48:56AM +0200, Christoph Hellwig wrote: > 1) The amount of code needed in nvme worries me a bit. Now NVMe a messy > driver due to the stupid PRPs vs just using SGLs, but needing a fair > amount of extra boilerplate code in drivers is a bit of a warning sign. > I plan to look into this to see if I can help on improving it, but for > that I need a working version first. It would be nice to have less. So much now depends on the caller to provide both the input and output data structure. 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. > Not quite as concerning, but doing an indirect call for each map > through dma_map_ops in addition to the iommu ops is not every > efficient. Yeah, there is no reason to support anything other than dma-iommu.c for the iommu path, so the dma_map_op indirection for this could just be removed. 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. > We've through for a while to allow direct calls to dma-iommu similar > how we do direct calls to dma-direct from the core mapping.c code. > This might be a good time to do that as a prep step for this work. I think there is room to benchmark and further improve these paths. Even the fast direct map path is not compiling down to a single load/store instruction per bio_vec entry as would be ideal. Jason
