[RFC PATCH v1 2/9] landlock: Support TCP listen access-control

[Mickaël Salaün]

On Sun, Jul 28, 2024 at 08:25:55AM +0800, Mikhail Ivanov wrote:
> LANDLOCK_ACCESS_NET_BIND_TCP is useful to limit the scope of "bindable"
> ports to forbid a malicious sandboxed process to impersonate a legitimate
> server process. However, bind(2) might be used by (TCP) clients to set the
> source port to a (legitimate) value. Controlling the ports that can be
> used for listening would allow (TCP) clients to explicitly bind to ports
> that are forbidden for listening.
> 
> Such control is implemented with a new LANDLOCK_ACCESS_NET_LISTEN_TCP
> access right that restricts listening on undesired ports with listen(2).
> 
> It's worth noticing that this access right doesn't affect changing
> backlog value using listen(2) on already listening socket.
> 
> * Create new LANDLOCK_ACCESS_NET_LISTEN_TCP flag.
> * Add hook to socket_listen(), which checks whether the socket is allowed
>   to listen on a binded local port.
> * Add check_tcp_socket_can_listen() helper, which validates socket
>   attributes before the actual access right check.
> * Update `struct landlock_net_port_attr` documentation with control of
>   binding to ephemeral port with listen(2) description.
> * Change ABI version to 6.
> 
> Closes: https://github.com/landlock-lsm/linux/issues/15
> Signed-off-by: Mikhail Ivanov <ivanov.mikhail1 at huawei-partners.com>
> ---
>  include/uapi/linux/landlock.h                | 23 +++--
>  security/landlock/limits.h                   |  2 +-
>  security/landlock/net.c                      | 90 ++++++++++++++++++++
>  security/landlock/syscalls.c                 |  2 +-
>  tools/testing/selftests/landlock/base_test.c |  2 +-
>  5 files changed, 108 insertions(+), 11 deletions(-)

> diff --git a/security/landlock/net.c b/security/landlock/net.c
> index 669ba260342f..a29cb27c3f14 100644
> --- a/security/landlock/net.c
> +++ b/security/landlock/net.c
> @@ -6,10 +6,12 @@
>   * Copyright © 2022-2023 Microsoft Corporation
>   */
>  
> +#include "net/sock.h"
>  #include <linux/in.h>
>  #include <linux/net.h>
>  #include <linux/socket.h>
>  #include <net/ipv6.h>
> +#include <net/tcp.h>
>  
>  #include "common.h"
>  #include "cred.h"
> @@ -194,9 +196,97 @@ static int hook_socket_connect(struct socket *const sock,
>  					   LANDLOCK_ACCESS_NET_CONNECT_TCP);
>  }
>  
> +/*
> + * Checks that socket state and attributes are correct for listen.
> + * It is required to not wrongfully return -EACCES instead of -EINVAL.
> + *
> + * This checker requires sock->sk to be locked.
> + */
> +static int check_tcp_socket_can_listen(struct socket *const sock)
> +{
> +	struct sock *sk = sock->sk;
> +	unsigned char cur_sk_state = sk->sk_state;
> +	const struct tcp_ulp_ops *icsk_ulp_ops;
> +

I think we can add this assert:
lockdep_assert_held(&sk->sk_lock.slock);

> +	/* Allows only unconnected TCP socket to listen (cf. inet_listen). */
> +	if (sock->state != SS_UNCONNECTED)
> +		return -EINVAL;
> +
> +	/*
> +	 * Checks sock state. This is needed to ensure consistency with inet stack
> +	 * error handling (cf. __inet_listen_sk).
> +	 */
> +	if (WARN_ON_ONCE(!((1 << cur_sk_state) & (TCPF_CLOSE | TCPF_LISTEN))))
> +		return -EINVAL;
> +
> +	icsk_ulp_ops = inet_csk(sk)->icsk_ulp_ops;
> +
> +	/*
> +	 * ULP (Upper Layer Protocol) stands for protocols which are higher than
> +	 * transport protocol in OSI model. Linux has an infrastructure that
> +	 * allows TCP sockets to support logic of some ULP (e.g. TLS ULP).
> +	 *
> +	 * Sockets can listen only if ULP control hook has clone method.
> +	 */
> +	if (icsk_ulp_ops && !icsk_ulp_ops->clone)
> +		return -EINVAL;
> +	return 0;
> +}