[PATCH net] netlabel: fix RCU annotation for IPv4 options on socket creation

[Eric Dumazet]

On Thu, Apr 11, 2024 at 5:44 PM Davide Caratti <dcaratti at redhat.com> wrote:
>
> Xiumei reports the following splat when netlabel and TCP socket are used:
>
>  =============================
>  WARNING: suspicious RCU usage
>  6.9.0-rc2+ #637 Not tainted
>  -----------------------------
>  net/ipv4/cipso_ipv4.c:1880 suspicious rcu_dereference_protected() usage!
>
>  other info that might help us debug this:
>
>  rcu_scheduler_active = 2, debug_locks = 1
>  1 lock held by ncat/23333:
>   #0: ffffffff906030c0 (rcu_read_lock){....}-{1:2}, at: netlbl_sock_setattr+0x25/0x1b0
>
>  stack backtrace:
>  CPU: 11 PID: 23333 Comm: ncat Kdump: loaded Not tainted 6.9.0-rc2+ #637
>  Hardware name: Supermicro SYS-6027R-72RF/X9DRH-7TF/7F/iTF/iF, BIOS 3.0  07/26/2013
>  Call Trace:
>   <TASK>
>   dump_stack_lvl+0xa9/0xc0
>   lockdep_rcu_suspicious+0x117/0x190
>   cipso_v4_sock_setattr+0x1ab/0x1b0
>   netlbl_sock_setattr+0x13e/0x1b0
>   selinux_netlbl_socket_post_create+0x3f/0x80
>   selinux_socket_post_create+0x1a0/0x460
>   security_socket_post_create+0x42/0x60
>   __sock_create+0x342/0x3a0
>   __sys_socket_create.part.22+0x42/0x70
>   __sys_socket+0x37/0xb0
>   __x64_sys_socket+0x16/0x20
>   do_syscall_64+0x96/0x180
>   ? do_user_addr_fault+0x68d/0xa30
>   ? exc_page_fault+0x171/0x280
>   ? asm_exc_page_fault+0x22/0x30
>   entry_SYSCALL_64_after_hwframe+0x71/0x79
>  RIP: 0033:0x7fbc0ca3fc1b
>  Code: 73 01 c3 48 8b 0d 05 f2 1b 00 f7 d8 64 89 01 48 83 c8 ff c3 66 2e 0f 1f 84 00 00 00 00 00 90 f3 0f 1e fa b8 29 00 00 00 0f 05 <48> 3d 01 f0 ff ff 73 01 c3 48 8b 0d d5 f1 1b 00 f7 d8 64 89 01 48
>  RSP: 002b:00007fff18635208 EFLAGS: 00000246 ORIG_RAX: 0000000000000029
>  RAX: ffffffffffffffda RBX: 0000000000000001 RCX: 00007fbc0ca3fc1b
>  RDX: 0000000000000006 RSI: 0000000000000001 RDI: 0000000000000002
>  RBP: 000055d24f80f8a0 R08: 0000000000000003 R09: 0000000000000001
>  R10: 0000000000020000 R11: 0000000000000246 R12: 000055d24f80f8a0
>  R13: 0000000000000000 R14: 000055d24f80fb88 R15: 0000000000000000
>   </TASK>
>
> The current implementation of cipso_v4_sock_setattr() replaces IP options
> under the assumption that the caller holds the socket lock; however, such
> assumption is not true, nor needed, in selinux_socket_post_create() hook.
>
> Using rcu_dereference_check() instead of rcu_dereference_protected() will
> avoid the reported splat for the netlbl_sock_setattr() case, and preserve
> the legitimate check when the caller is netlbl_conn_setattr().
>
> Fixes: f6d8bd051c39 ("inet: add RCU protection to inet->opt")
> Reported-by: Xiumei Mu <xmu at redhat.com>
> Suggested-by: Paolo Abeni <pabeni at redhat.com>
> Signed-off-by: Davide Caratti <dcaratti at redhat.com>
> ---
>  net/ipv4/cipso_ipv4.c | 6 ++++--
>  1 file changed, 4 insertions(+), 2 deletions(-)
>
> diff --git a/net/ipv4/cipso_ipv4.c b/net/ipv4/cipso_ipv4.c
> index 8b17d83e5fde..1d0c2a905078 100644
> --- a/net/ipv4/cipso_ipv4.c
> +++ b/net/ipv4/cipso_ipv4.c
> @@ -1876,8 +1876,10 @@ int cipso_v4_sock_setattr(struct sock *sk,
>
>         sk_inet = inet_sk(sk);
>
> -       old = rcu_dereference_protected(sk_inet->inet_opt,
> -                                       lockdep_sock_is_held(sk));
> +       /* caller either holds rcu_read_lock() (on socket creation)
> +        * or socket lock (in all other cases). */
> +       old = rcu_dereference_check(sk_inet->inet_opt,
> +                                   lockdep_sock_is_held(sk));
>         if (inet_test_bit(IS_ICSK, sk)) {
>                 sk_conn = inet_csk(sk);
>                 if (old)
> --
> 2.44.0
>

OK, but rcu_read_lock() being held (incidentally by the caller) here
is not protecting the write operation,
so this looks wrong IMO.

Whenever we can not ensure a mutex/spinlock is held, we usually use
rcu_dereference_protected(XXX, 1),
and a comment might simply explain the reason we assert it is protected.

(We also could add a new boolean parameter, set to true or false
depending on the caller)

old = rcu_dereference_protected(sk_inet->inet_opt, from_socket_creation ||

            lockdep_sock_is_held(sk));