flow_dissector: Allow updating the flow dissector program atomically

It is currently not possible to detach the flow dissector program and
attach a new one in an atomic fashion, that is with a single syscall.
Attempts to do so will be met with EEXIST error.

This makes updates to flow dissector program hard. Traffic steering that
relies on BPF-powered flow dissection gets disrupted while old program has
been already detached but the new one has not been attached yet.

There is also a window of opportunity to attach a flow dissector to a
non-root namespace while updating the root flow dissector, thus blocking
the update.

Lastly, the behavior is inconsistent with cgroup BPF programs, which can be
replaced with a single bpf(BPF_PROG_ATTACH, ...) syscall without any
restrictions.

Allow attaching a new flow dissector program when another one is already
present with a restriction that it can't be the same program.
Signed-off-by: Jakub Sitnicki <jakub@cloudflare.com>
Signed-off-by: Daniel Borkmann <daniel@iogearbox.net>
Reviewed-by: Stanislav Fomichev <sdf@google.com>
Acked-by: Martin KaFai Lau <kafai@fb.com>
Link: https://lore.kernel.org/bpf/20191011082946.22695-2-jakub@cloudflare.com

net/core/flow_dissector.c

@@ -128,6 +128,8 @@ int skb_flow_dissector_bpf_prog_attach(const union bpf_attr *attr,
 		struct net *ns;
 
 		for_each_net(ns) {
+			if (ns == &init_net)
+				continue;
 			if (rcu_access_pointer(ns->flow_dissector_prog)) {
 				ret = -EEXIST;
 				goto out;
@@ -145,12 +147,14 @@ int skb_flow_dissector_bpf_prog_attach(const union bpf_attr *attr,
 
 	attached = rcu_dereference_protected(net->flow_dissector_prog,
 					     lockdep_is_held(&flow_dissector_mutex));
-	if (attached) {
-		/* Only one BPF program can be attached at a time */
-		ret = -EEXIST;
+	if (attached == prog) {
+		/* The same program cannot be attached twice */
+		ret = -EINVAL;
 		goto out;
 	}
 	rcu_assign_pointer(net->flow_dissector_prog, prog);
+	if (attached)
+		bpf_prog_put(attached);
 out:
 	mutex_unlock(&flow_dissector_mutex);
 	return ret;