On Tue, Aug 29, 2017 at 04:37:56PM +0200, Jesper Dangaard Brouer wrote: > There is a need to separate the xdp_redirect tracepoint into two > tracepoints, for separating the error case from the normal forward > case. > > Due to the extreme speeds XDP is operating at, loading a tracepoint > have a measurable impact. Single core XDP REDIRECT (ethtool tuned > rx-usecs 25) can do 13.7 Mpps forwarding, but loading a simple > bpf_prog at the tracepoint (with a return 0) reduce perf to 10.2 Mpps > (CPU E5-1650 v4 @ 3.60GHz, driver: ixgbe) > > The overhead of loading a bpf-based tracepoint can be calculated to > cost 25 nanosec ((1/13782002-1/10267937)*10^9 = -24.83 ns). > > Using perf record on the tracepoint event, with a non-matching --filter > expression, the overhead is much larger. Performance drops to 8.3 Mpps, > cost 48 nanosec ((1/13782002-1/8312497)*10^9 = -47.74)) > > Having a separate tracepoint for err cases, which should be less > frequent, allow running a continuous monitor for errors while not > affecting the redirect forward performance (this have also been > verified by measurements). > > Signed-off-by: Jesper Dangaard Brouer <bro...@redhat.com>
thanks for detailed analysis of performance impact of the changes. looks great to me. Acked-by: Alexei Starovoitov <a...@kernel.org>
