On 2024-07-03 12:01:43 [-0700], Jakub Kicinski wrote: > On Wed, 3 Jul 2024 14:27:58 +0200 Sebastian Andrzej Siewior wrote: > > During the introduction of struct bpf_net_context handling for > > XDP-redirect, the tun driver has been missed. > > > > Set the bpf_net_context before invoking BPF XDP program within the TUN > > driver. > > Sorry if I'm missing the point but I think this is insufficient. > You've covered the NAPI-like entry point to the Rx stack in your > initial work, but there's also netif_receive_skb() which drivers > may call outside of NAPI, simply disabling BH before the call. Ah okay. I've been looking at a few callers and they ended up in NAPI but if you say and I also noticed the one in TUN… > The only concern in that case is that we end up in do_xdp_generic(), > and there's no bpf_net_set_ctx() anywhere on the way. So my intuition > would be to add the bpf_net_set_ctx() inside the if(xdp_prog) in > do_xdp_generic(). "XDP generic" has less stringent (more TC-like) > perf requirements, so setting context once per packet should be fine. > And it will also cover drivers like TUN which use both > netif_receive_skb() and call do_xdp_generic(), in a single place. Yeah, sounds good. I would remove the wrapper in tun_get_user() and then add one to do_xdp_generic(). Sebastian
