On Sat, 20 Jun 2020 00:57:20 +0200
Lorenzo Bianconi <lore...@kernel.org> wrote:
> +static int cpu_map_bpf_prog_run_xdp(struct bpf_cpu_map_entry *rcpu,
> + void **xdp_frames, int n,
> + struct xdp_cpumap_stats *stats)
> +{
> + struct xdp_rxq_info rxq;
I do think we can merge this code as-is (and I actually re-factored this
code offlist), but I have some optimizations I would like to try out.
E.g. as I've tried to explain over IRC, it will be possible to avoid
having to reconstruct xdp_rxq_info here, as we can use the
expected_attach_type and remap the BPF instructions to use info from
xdp_frame. I want to benchmark it first, to see if it is worth it (we
will only save 2 store operations in a likely cache-hot area).
> + struct bpf_prog *prog;
> + struct xdp_buff xdp;
> + int i, nframes = 0;
> +
> + if (!rcpu->prog)
> + return n;
> +
> + xdp_set_return_frame_no_direct();
> + xdp.rxq = &rxq;
> +
> + rcu_read_lock();
> +
> + prog = READ_ONCE(rcpu->prog);
> + for (i = 0; i < n; i++) {
> + struct xdp_frame *xdpf = xdp_frames[i];
> + u32 act;
> + int err;
> +
> + rxq.dev = xdpf->dev_rx;
> + rxq.mem = xdpf->mem;
> + /* TODO: report queue_index to xdp_rxq_info */
> +
> + xdp_convert_frame_to_buff(xdpf, &xdp);
> +
> + act = bpf_prog_run_xdp(prog, &xdp);
> + switch (act) {
> + case XDP_PASS:
> + err = xdp_update_frame_from_buff(&xdp, xdpf);
> + if (err < 0) {
> + xdp_return_frame(xdpf);
> + stats->drop++;
> + } else {
> + xdp_frames[nframes++] = xdpf;
> + stats->pass++;
> + }
> + break;
> + default:
> + bpf_warn_invalid_xdp_action(act);
> + /* fallthrough */
> + case XDP_DROP:
> + xdp_return_frame(xdpf);
> + stats->drop++;
> + break;
> + }
> + }
> +
> + rcu_read_unlock();
> + xdp_clear_return_frame_no_direct();
> +
> + return nframes;
> +}
--
Best regards,
Jesper Dangaard Brouer
MSc.CS, Principal Kernel Engineer at Red Hat
LinkedIn: http://www.linkedin.com/in/brouer
