It's currently possible to insert sockets in unexpected states into
a sockmap, due to a TOCTTOU when updating the map from a syscall.
sock_map_update_elem checks that sk->sk_state == TCP_ESTABLISHED,
locks the socket and then calls sock_map_update_common. At this
point, the socket may have transitioned into another state, and
the earlier assumptions don't hold anymore. Crucially, it's
conceivable (though very unlikely) that a socket has become unhashed.
This breaks the sockmap's assumption that it will get a callback
via sk->sk_prot->unhash.
Fix this by checking the (fixed) sk_type and sk_protocol without the
lock, followed by a locked check of sk_state.
Unfortunately it's not possible to push the check down into
sock_(map|hash)_update_common, since BPF_SOCK_OPS_PASSIVE_ESTABLISHED_CB
run before the socket has transitioned from TCP_SYN_RECV into
TCP_ESTABLISHED.
Fixes: 604326b41a6f ("bpf, sockmap: convert to generic sk_msg interface")
Signed-off-by: Lorenz Bauer <lmb@cloudflare.com>
Signed-off-by: Daniel Borkmann <daniel@iogearbox.net>
Reviewed-by: Jakub Sitnicki <jakub@cloudflare.com>
Link: https://lore.kernel.org/bpf/20200207103713.28175-1-lmb@cloudflare.com
                ret = -EINVAL;
                goto out;
        }
-       if (!sock_map_sk_is_suitable(sk) ||
-           sk->sk_state != TCP_ESTABLISHED) {
+       if (!sock_map_sk_is_suitable(sk)) {
                ret = -EOPNOTSUPP;
                goto out;
        }
 
        sock_map_sk_acquire(sk);
-       ret = sock_map_update_common(map, idx, sk, flags);
+       if (sk->sk_state != TCP_ESTABLISHED)
+               ret = -EOPNOTSUPP;
+       else
+               ret = sock_map_update_common(map, idx, sk, flags);
        sock_map_sk_release(sk);
 out:
        fput(sock->file);
                ret = -EINVAL;
                goto out;
        }
-       if (!sock_map_sk_is_suitable(sk) ||
-           sk->sk_state != TCP_ESTABLISHED) {
+       if (!sock_map_sk_is_suitable(sk)) {
                ret = -EOPNOTSUPP;
                goto out;
        }
 
        sock_map_sk_acquire(sk);
-       ret = sock_hash_update_common(map, key, sk, flags);
+       if (sk->sk_state != TCP_ESTABLISHED)
+               ret = -EOPNOTSUPP;
+       else
+               ret = sock_hash_update_common(map, key, sk, flags);
        sock_map_sk_release(sk);
 out:
        fput(sock->file);