On 09/21/2018 12:17 AM, Song Liu wrote: > > >> On Sep 20, 2018, at 4:49 PM, Eric Dumazet <eric.dumazet@xxxxxxxxx> wrote: >> >> >> >> On 09/20/2018 04:43 PM, Song Liu wrote: >>> >> >>> I tried to totally skip ndo_poll_controller() here. It did avoid hitting >>> the issue. However, netpoll will drop (fail to send) more packets. >>> >> >> Why is it failing ? >> >> If you are under high memory pressure, then maybe if you absolutely want memory to send >> netpoll packets, you want to grab all NAPI contexts as a way to prevent other cpus >> from feeding incoming packets to the host and add more memory pressure ;) >> > > I did the test with Eric's latest patch (and disable ndo_poll_controller > in driver). The result didn't show significant increase in drop packets. > I guess packet drops in my earlier test was caused by some other changes > I mixed there. > > So I think this patch does fix the issue. Thanks Eric! Great, this is awesome. I will prepare a patch series for net tree. The core infrastructure is just better at being able to drain TX completions without risking stealing the NAPI context forever. > > For ixgbe, I think we need to check napi_complete_done() return value > anyway. Otherwise, the driver will enable IRQ in polling mode. > Sure, let's prepare such optimizations for net-next.
