On Tue, 8 Oct 2019 at 08:59, Toke Høiland-Jørgensen [off-list ref] wrote:

Sridhar Samudrala [off-list ref] writes:

quoted

 int xdp_do_redirect(struct net_device *dev, struct xdp_buff *xdp,
                  struct bpf_prog *xdp_prog)
 {
      struct bpf_redirect_info *ri = this_cpu_ptr(&bpf_redirect_info);
      struct bpf_map *map = READ_ONCE(ri->map);
+     struct xdp_sock *xsk;
+
+     xsk = xdp_get_direct_xsk(ri);
+     if (xsk)
+             return xsk_rcv(xsk, xdp);

This is a new branch and a read barrier in the XDP_REDIRECT fast path.
What's the performance impact of that for non-XSK redirect?

The dependent-read-barrier in READ_ONCE? Another branch -- leave that
to the branch-predictor already! ;-) No, you're right, performance
impact here is interesting. I guess the same static_branch could be
used here as well...

-Toke