On Tue, Oct 29, 2024 at 11:25:18AM +0100, Paolo Abeni wrote: > On 10/21/24 03:52, Joe Damato wrote: [...] > > The changes makes sense to me, and I could not find any obvious issue in > the patches. > > I think this deserve some - even basic - self-tests coverage. Note that > you can enable GRO on veth devices to make NAPI instances avail there. > > Possibly you could opt for a drivers/net defaulting to veth usage and > allowing the user to select real H/W via env variables. Sorry, Paolo, but I took a cursory look at veth and I need to object more strongly to your feedback here. My understanding (which could be incorrect) is that neither veth nor netdevsim use IRQs. The whole purpose of this series is to block IRQs while data is being busy polled to increase efficiency. That's in the cover letter. If neither of the drivers we'd use to simulate this in selftest use IRQs, how could we build a selftest which ensures IRQs are correctly suspended during busy periods without first making considerable changes to either (or both?) drivers? Respectfully: I don't think it's appropriate to block this series on that much additional work. Please reconsider and let me know how to proceed.