On 10/07/2019 08:27, Paolo Abeni wrote:

I'm toying with a patch similar to your 3/3 (most relevant difference
being the lack of a limit to the batch size), on top of ixgbe (which
sends all the pkts to the GRO engine), and I'm observing more
controversial results (UDP only):

* when a single rx queue is running, I see a just-above-noise
peformance delta
* when multiple rx queues are running, I observe measurable regressions
(note: I use small pkts, still well under line rate even with multiple
rx queues)

I'll try to test your patch in the following days.

I look forward to it.

Side note: I think that in patch 3/3, it's necessary to add a call to
gro_normal_list() also inside napi_busy_loop().

Hmm, I was caught out by the call to napi_poll() actually being a local
 function pointer, not the static function of the same name.  How did a
 shadow like that ever get allowed?
But in that case I _really_ don't understand napi_busy_loop(); nothing
 in it seems to ever flush GRO, so it's relying on either
 (1) stuff getting flushed because the bucket runs out of space, or
 (2) the next napi poll after busy_poll_stop() doing the flush.
What am I missing, and where exactly in napi_busy_loop() should the
 gro_normal_list() call go?

-Ed