On Sun, 2012-10-07 at 07:29 +0200, Eric Dumazet wrote:

	On Sun, 2012-10-07 at 08:32 +0800, Herbert Xu wrote:

quoted

Why don't we just always flush everything?

This is what I tried first, but it lowered performance on several
typical workloads.

Using this simple heuristic increases performance.

By the way, one of the beauty of GRO is it helps under load to aggregate
packets and reduce cpu load. People wanting very low latencies should
probably not use GRO, and if they use it or not, receiving a full 64
packets batch on a particular NIC makes latencies very unpredictable.

So if we consumed all budget in a napi->poll() handler, its because we
are under load and we dont really want to cancel GRO aggregation.

Next napi->poll() invocation will have more chances to coalesce frames.

If there is only one flow, its OK because a 64 packet window allows ~4
GRO super packets to be built, regardless of an unconditional flush, but
with 8 flows, it would roughly give 100% increase of GRO packets sent to
upper layers.

Only needed safety measure is to make sure we dont let packets for a too
long time in case we never complete napi, this is what this patch does,
with a latency average of 0.5 ms (for slow flows)