Re: [PATCH bpf-next v2 1/3] bpf, sockmap: avoid using sk_socket after free

[Michal Luczaj <mhal@rbox.co>]

On 2/28/25 06:51, Jiayuan Chen wrote:
> ...
>  static void sk_psock_verdict_data_ready(struct sock *sk)
>  {
> -	struct socket *sock = sk->sk_socket;
> +	struct socket *sock;
>  	const struct proto_ops *ops;
>  	int copied;
>  
>  	trace_sk_data_ready(sk);
>  
> +	/* We need RCU to prevent the sk_socket from being released.
> +	 * Especially for Unix sockets, we are currently in the process
> +	 * context and do not have RCU protection.
> +	 */
> +	rcu_read_lock();
> +	sock = sk->sk_socket;
>  	if (unlikely(!sock))
> -		return;
> +		goto unlock;
> +
>  	ops = READ_ONCE(sock->ops);
>  	if (!ops || !ops->read_skb)
> -		return;
> +		goto unlock;
> +
>  	copied = ops->read_skb(sk, sk_psock_verdict_recv);
>  	if (copied >= 0) {
>  		struct sk_psock *psock;
>  
> -		rcu_read_lock();
>  		psock = sk_psock(sk);
>  		if (psock)
>  			sk_psock_data_ready(sk, psock);
> -		rcu_read_unlock();
>  	}
> +unlock:
> +	rcu_read_unlock();
>  }

Hi,

Doesn't sk_psock_handle_skb() (!ingress path) have the same `struct socket`
release race issue? Any plans on fixing that one, too?

BTW, lockdep (CONFIG_LOCKDEP=y) complains about calling AF_UNIX's
read_skb() under RCU read lock.

Thanks,
Michal