Re: Linux kernel ping socket / AF_LLC connect() sin_family race

[Solar Designer <solar () openwall com>]

On Fri, Mar 24, 2017 at 03:21:06PM -0700, Eric Dumazet wrote:
> Looks easy enough to fix ?

Oh.  Probably.  Thanks.  Need to test, but I guess you already did?

> diff --git a/net/ipv4/ping.c b/net/ipv4/ping.c
> index
> 2af6244b83e27ae384e96cf071c10c5a89674804..ccfbce13a6333a65dab64e4847dd510dfafb1b43
> 100644
> --- a/net/ipv4/ping.c
> +++ b/net/ipv4/ping.c
> @@ -156,17 +156,18 @@ int ping_hash(struct sock *sk)
>  void ping_unhash(struct sock *sk)
>  {
>         struct inet_sock *isk = inet_sk(sk);
> +
>         pr_debug("ping_unhash(isk=%p,isk->num=%u)\n", isk, isk->inet_num);
> +       write_lock_bh(&ping_table.lock);
>         if (sk_hashed(sk)) {
> -               write_lock_bh(&ping_table.lock);
>                 hlist_nulls_del(&sk->sk_nulls_node);
>                 sk_nulls_node_init(&sk->sk_nulls_node);
>                 sock_put(sk);
>                 isk->inet_num = 0;
>                 isk->inet_sport = 0;
>                 sock_prot_inuse_add(sock_net(sk), sk->sk_prot, -1);
> -               write_unlock_bh(&ping_table.lock);
>         }
> +       write_unlock_bh(&ping_table.lock);
>  }
>  EXPORT_SYMBOL_GPL(ping_unhash);

FWIW, in Pavel's original implementation for 2.4.32 (unused), this was:

static void ping_v4_unhash(struct sock *sk)
{
        DEBUG(("ping_v4_unhash(sk=%p,sk->num=%u)\n", sk, sk->num));
        write_lock_bh(&ping_hash_lock);
        if (sk->pprev) {
                if (sk->next)
                       sk->next->pprev = sk->pprev;
                *sk->pprev = sk->next;
                sk->pprev = NULL;
                sk->num = 0;
                sock_prot_dec_use(sk->prot);
                __sock_put(sk);
        }
        write_unlock_bh(&ping_hash_lock);
}

Looks like the erroneous optimization (not expecting concurrent activity
on the same socket?) was introduced during conversion to 2.6's hlists.

So far this cursed function had 3 bugs, two of them security (including
this one) and one probably benign (or if not, then effectively a subset
of this bug as it performed some unneeded / stale debugging work before
acquiring the lock), with all 3 introduced in forward-porting.  Maybe
the nature of forward-porting activity makes people relatively
inattentive ("compiles with the new interfaces and still works? must be
correct"), compared to when writing new code.

Anyhow, I share some responsibility for this mess, for having advocated
this patch being forward-ported and merged back then.  I still like
having this functionality and its userspace security benefits... but I
don't like the kernel bugs.

Alexander