On Fri, Jul 05, 2024 at 08:39:10AM +0200, Christoph Hellwig wrote: > Review from the NVMe driver consumer perspective. I think if all these > were implement we'd probably end up with less code than before the > conversion. One of the topics that came up with at LSF is what to do with the dma_memory_type data? The concept here was that the new DMA API would work on same-type memory only, meaning the caller would have to split up by type. I understand the block stack already does this using P2P and !P2P, but that isn't quite enough here as we want to split principally based on IOMMU or !IOMMU. Today that boils down to splitting based on a few things, like grouping encrypted memory, and grouping by P2P provider. So the idea was the "struct dma_memory_type" would encapsulate whatever grouping was needed and the block layer would drive its splitting on this, same structs can be merged. I didn't notice this got included in this RFC.. The trivial option is to store the dma_memory_type per-BIO and drive the de-duplication like that. The other could be to derive it from first entry in the bio_vect every time a new page is added. This same-type design is one of the key elements of this API arrangement - for RDMA I expect we will need a data structure storing a list of types with a list of pages, and we will DMA map type by type. Jason
