From: Edward Cree <ec...@solarflare.com> Date: Fri, 13 Apr 2018 13:36:28 +0100
> It turns out this may all be moot anyway: I figured out why I was seeing > ARFS storms and it wasn't the configuration issue I originally blamed. > My current ndo_rx_flow_steer() implementation, efx_filter_rfs(), returns > 0 for success, but the caller expects a filter ID to be returned (which > we can't give it because we don't know what the filter ID will be until > we start mucking around in the software state that's now protected by a > sleepable lock). > As a result, when we call rps_may_expire_flow(), and pass it the _actual_ > filter ID, this doesn't match the one set_rps_cpu() recorded, so the > function returns true and we immediately expire the filter. Then the > next packet to come along isn't steered, so ARFS asks us to insert a > steering filter again. > As a quick fix I've simply tried making the rps_may_expire_flow() calls > also pass a filter ID of 0, which prevents the ARFS storms. This is > safe; it may cause us to delay expiring a filter when flow_ids collide, > but that can happen anyway with other drivers' implementations (e.g. > mlx4 and mlx5 can potentially reuse filter IDs) so I presume it is OK. > I'll post a v2 with that fix in place of this Patch #2 shortly, then try > to follow up with a counter-generated ID (similar to what mlx have). I understand the constraints you are working under, but do realize that the real root of the problems is that you are implementing what is defined clearly as a synchronous operation as asynchronous.