On Tue, Jun 17, 2025 at 07:33:46PM +0200, Daniel Gomez wrote: > On 16/06/2025 13.33, Christoph Hellwig wrote: > > On Mon, Jun 16, 2025 at 09:41:15AM +0200, Daniel Gomez wrote: > >> Also, if host segments are between 4k and 16k, PRPs would be able to support it > >> but this limit prevents that use case. I guess the question is if you see any > >> blocker to enable this path? > > > > Well, if you think it's worth it give it a spin on a wide variety of > > hardware. > > I'm not sure if I understand this. Can you clarify why hardware evaluation would > be required? What exactly? This is about chaining SGL's so I think the request is benchmarking if that's faster than splitting commands. Splitting hand been quicker for much hardware because they could process SQE's in parallel easier than walking a single command's SG List. On a slightly related topic, NVMe SGL's don't need the "virt_boundary_mask". So for devices are optimized for SGL, then that queue limit could go away, and I've recently heard use cases for the passthrough interface where that would be useful on avoiding kernel copy bounce buffers (sorry for the digression).
