On 10/26/16 at 10:08am, David Ahern wrote: > On 10/26/16 2:41 AM, Thomas Graf wrote: > > On 10/25/16 at 03:30pm, David Ahern wrote: > >> @@ -171,6 +177,9 @@ int __cgroup_bpf_run_filter(struct sock *sk, > >> case BPF_CGROUP_INET_EGRESS: > >> ret = __cgroup_bpf_run_filter_skb(skb, prog); > >> break; > >> + case BPF_CGROUP_INET_SOCK_CREATE: > >> + ret = __cgroup_bpf_run_filter_sk_create(sk, prog); > >> + break; > >> /* make gcc happy else complains about missing enum value */ > >> default: > >> return 0; > > > > Thinking further ahead of your simple example. Instead of adding yet > > another prog type for the same hook, we can make this compatible with > > BPF_CGROUP_INET_EGRESS instead which would then provide a ctx which > > contains both, the sk and skb. > > > > The ctx concept is very flexible. We can keep the existing dummy skb > > representation and add sk_ fields which are only valid for BPF at > > socket layer, e.g skb->sk_bound_dev_if would translate to > > sk->bound_dev_if. > > > > It's an odd user semantic to me to put sock elements into the shadow sk_buff > and to reuse BPF_CGROUP_INET_EGRESS. > > I can drop the _CREATE and just make it BPF_CGROUP_INET_SOCK so it works for > any sock modification someone wants to add -- e.g., the port binding use case.
Your specific sk_alloc hook won't have an skb but the majority of socket BPF programs will want to see both skb and sk. It is not ideal to introduce a new bpf_prog_type for every minimal difference in capability if the majority of capabilities and restrictions are shared. Instead, the subtype approach was implemented by the Landlock series looks much cleaner to me. I think we should think about how to do this properly for all BPF programs which operate in socket cgroup context.
