On Wed, May 13, 2020 at 07:41 AM CEST, Martin KaFai Lau wrote: > On Mon, May 11, 2020 at 08:52:03PM +0200, Jakub Sitnicki wrote: > > [ ... ] > >> +BPF_CALL_3(bpf_sk_lookup_assign, struct bpf_sk_lookup_kern *, ctx, >> + struct sock *, sk, u64, flags) > The SK_LOOKUP bpf_prog may have already selected the proper reuseport sk. > It is possible by looking up sk from sock_map. > > Thus, it is not always desired to do lookup_reuseport() after sk_assign() > in patch 5. e.g. reuseport_select_sock() just uses a normal hash if > there is no reuse->prog. > > A flag (e.g. "BPF_F_REUSEPORT_SELECT") can be added here to > specifically do the reuseport_select_sock() after sk_assign(). > If not set, reuseport_select_sock() should not be called. That's true that in addition to steering connections to different services with SK_LOOKUP, you could also, in the same program, load-balance among sockets belonging to one service. So skipping the reuseport socket selection, if sk_lookup already did load-balancing sounds useful. Thinking about our use-case, I think we would always pass BPF_F_REUSEPORT_SELECT to sk_assign() because we either (i) know that application is using reuseport and want it manage the load-balancing socket group by itself, or (ii) don't know if application is using reuseport and don't want to break expected behavior. IOW, we'd like reuseport selection to run by default because application expects it to happen if it was set up. OTOH, the application doesn't have to be aware that there is sk_lookup attached (we can put one of its sockets in sk_lookup SOCKMAP when systemd activates it). Beacuse of that I'd be in favor of having a flag for sk_assign() that disables reuseport selection on demand. WDYT?
