On 2021-01-29 17:45, Toke Høiland-Jørgensen wrote:

Björn Töpel [off-list ref] writes:

quoted

From: Björn Töpel <redacted>

Currently the bpf_redirect_map() implementation dispatches to the
correct map-lookup function via a switch-statement. To avoid the
dispatching, this change adds one bpf_redirect_map() implementation per
map. Correct function is automatically selected by the BPF verifier.

Signed-off-by: Björn Töpel <redacted>
---
Hi XDP-folks!

This is another take on my bpf_redirect_xsk() patch [1]. I figured I
send it as an RFC for some early input. My plan is to include it as
part of the xdp_do_redirect() optimization of [1].

Assuming the maintainers are OK with the special-casing in the verifier,
this looks like a neat way to avoid the runtime overhead to me. The
macro hackery is not the prettiest; I wonder if the same effect could be
achieved by using inline functions? If not, at least a comment
explaining the reasoning (and that the verifier will substitute the
right function) might be nice? Mostly in relation to this bit:

Yeah, I agree with the macro part. I'll replace it with a
__always_inline function, instead.

quoted

  static const struct bpf_func_proto bpf_xdp_redirect_map_proto = {
-	.func           = bpf_xdp_redirect_map,
+	.func           = bpf_xdp_redirect_devmap,

I'll try to clean this up as well.

Thanks for taking a look!
Björn

Ah, if only we were writing the kernel in a language with proper macro
support... One can dream! :)

quoted

quoted

For AF_XDP rxdrop this yields +600Mpps. I'll do CPU/DEVMAP
measurements for the patch proper.

Kpps, not Mpps. :-P

Aww, too bad ;)
Still, nice!

-Toke