bpf: avoid misuse of psock when TCP_ULP_BPF collides with another ULP
authorJohn Fastabend <john.fastabend@gmail.com>
Fri, 31 Aug 2018 04:25:02 +0000 (21:25 -0700)
committerDaniel Borkmann <daniel@iogearbox.net>
Sun, 2 Sep 2018 20:31:10 +0000 (22:31 +0200)
Currently we check sk_user_data is non NULL to determine if the sk
exists in a map. However, this is not sufficient to ensure the psock
or the ULP ops are not in use by another user, such as kcm or TLS. To
avoid this when adding a sock to a map also verify it is of the
correct ULP type. Additionally, when releasing a psock verify that
it is the TCP_ULP_BPF type before releasing the ULP. The error case
where we abort an update due to ULP collision can cause this error
path.

For example,

  __sock_map_ctx_update_elem()
     [...]
     err = tcp_set_ulp_id(sock, TCP_ULP_BPF) <- collides with TLS
     if (err)                                <- so err out here
        goto out_free
     [...]
  out_free:
     smap_release_sock() <- calling tcp_cleanup_ulp releases the
                            TLS ULP incorrectly.

Fixes: 2f857d04601a ("bpf: sockmap, remove STRPARSER map_flags and add multi-map support")
Signed-off-by: John Fastabend <john.fastabend@gmail.com>
Signed-off-by: Daniel Borkmann <daniel@iogearbox.net>
kernel/bpf/sockmap.c

index ce63e5801746840e18405237e9d7642ddf83559d..488ef9663c01f3b4d2cc44b9ca88863ced2860e7 100644 (file)
@@ -1462,10 +1462,16 @@ static void smap_destroy_psock(struct rcu_head *rcu)
        schedule_work(&psock->gc_work);
 }
 
+static bool psock_is_smap_sk(struct sock *sk)
+{
+       return inet_csk(sk)->icsk_ulp_ops == &bpf_tcp_ulp_ops;
+}
+
 static void smap_release_sock(struct smap_psock *psock, struct sock *sock)
 {
        if (refcount_dec_and_test(&psock->refcnt)) {
-               tcp_cleanup_ulp(sock);
+               if (psock_is_smap_sk(sock))
+                       tcp_cleanup_ulp(sock);
                write_lock_bh(&sock->sk_callback_lock);
                smap_stop_sock(psock, sock);
                write_unlock_bh(&sock->sk_callback_lock);
@@ -1892,6 +1898,10 @@ static int __sock_map_ctx_update_elem(struct bpf_map *map,
         * doesn't update user data.
         */
        if (psock) {
+               if (!psock_is_smap_sk(sock)) {
+                       err = -EBUSY;
+                       goto out_progs;
+               }
                if (READ_ONCE(psock->bpf_parse) && parse) {
                        err = -EBUSY;
                        goto out_progs;