On Tue, 6 Jun 2023 21:22:45 -0700 Justin Chen wrote: > >> Through filters that can be programmed by the Host cpu or co-processor. > > > > How are the filter programmed by the host (in terms of user API)? > > You are stumbling upon my next objective here. This patch set does not > support this because it wasn't clear how to do it. The plan was to get > the base driver merged first then introduce a way to specify an offload > channel. Alright. > If you must know, I introduced a rxnfc flow_spec ring cookie to specify > an offload filter. Not sure if this is the right way to do it or not, > but figured this would be a future discussion. We clean leave it for later indeed, but in case you're wondering - that's unlikely to fly. My preference will be to expose the accelerator port as a switch port, with a netdev and forward using TC flower etc. With the netdev in place you'll have a place to report stats and all the necessary diagnostics, which otherwise get crammed into random places. I believe Florian asked about this in the past, I hope I'm being consistent with my answer here :P
