On Thu, 7 May 2020 at 15:44, Toke Høiland-Jørgensen <toke@xxxxxxxxxx> wrote:
>
> Björn Töpel <bjorn.topel@xxxxxxxxx> writes:
>
> > Before I start hacking on this, I might as well check with the XDP
> > folks if this considered a crappy idea or not. :-)
> >
> > The XDP redirect flow for a packet is typical a dance of
> > bpf_redirect_map() that updates the bpf_redirect_info structure with
> > maps type/items, which is then followed by an xdp_do_redirect(). That
> > function takes an action based on the bpf_redirect_info content.
> >
> > I'd like to get rid of the xdp_do_redirect() call, and the
> > bpf_redirect_info (per-cpu) lookup. The idea is to introduce a new
> > (oh-no!) XDP action, say, XDP_CONSUMED and a built-in helper with
> > tail-call semantics.
> >
> > Something across the lines of:
> >
> > --8<--
> >
> > struct {
> > __uint(type, BPF_MAP_TYPE_XSKMAP);
> > __uint(max_entries, MAX_SOCKS);
> > __uint(key_size, sizeof(int));
> > __uint(value_size, sizeof(int));
> > } xsks_map SEC(".maps");
> >
> > SEC("xdp1")
> > int xdp_prog1(struct xdp_md *ctx)
> > {
> > bpf_tail_call_redirect(ctx, &xsks_map, 0);
> > // Redirect the packet to an AF_XDP socket at entry 0 of the
> > // map.
> > //
> > // After a successful call, ctx is said to be
> > // consumed. XDP_CONSUMED will be returned by the program.
> > // Note that if the call is not successful, the buffer is
> > // still valid.
> > //
> > // XDP_CONSUMED in the driver means that the driver should not
> > // issue an xdp_do_direct() call, but only xdp_flush().
> > //
> > // The verifier need to be taught that XDP_CONSUMED can only
> > // be returned "indirectly", meaning a bpf_tail_call_XXX()
> > // call. An explicit "return XDP_CONSUMED" should be
> > // rejected. Can that be implemented?
> > return XDP_PASS; // or any other valid action.
> > }
> >
> > -->8--
> >
> > The bpf_tail_call_redirect() would work with all redirectable maps.
> >
> > Thoughts? Tomatoes? Pitchforks?
>
> The above answers the 'what'. Might be easier to evaluate if you also
> included the 'why'? :)
>
Ah! Sorry! Performance, performance, performance. Getting rid of a
bunch of calls/instructions per packet, which helps my (AF_XDP) case.
This would be faster than the regular REDIRECT path. Today, in
bpf_redirect_map(), instead of actually performing the action, we
populate the bpf_redirect_info structure, just to look up the action
again in xdp_do_redirect().
I'm pretty certain this would be a gain for AF_XDP (quite easy to do a
quick hack, and measure). It would also shave off the same amount of
instructions for "vanilla" XDP_REDIRECT cases. The bigger issue; Is
this new semantic something people would be comfortable being added to
XDP.
Cheers,
Björn
