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

[Casey Schaufler]

On 4/11/2024 8:56 AM, Eric Dumazet wrote:
> 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>

Please be sure to verify that this is appropriate for all users of netlabel.
SELinux is not the only user of netlabel.

>> ---
>>  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));
>