Re: [PATCH bpf-next v3 2/9] bpf: Propagate modified uaddrlen from cgroup sockaddr programs

[Martin KaFai Lau <martin.lau@linux.dev>]

On 8/29/23 3:18 AM, Daan De Meyer wrote:
> @@ -1482,11 +1485,22 @@ int __cgroup_bpf_run_filter_sock_addr(struct sock *sk,
>   	if (!ctx.uaddr) {
>   		memset(&unspec, 0, sizeof(unspec));
>   		ctx.uaddr = (struct sockaddr *)&unspec;
> -	}
> +		ctx.uaddrlen = sizeof(unspec);

ctx.uaddr could be NULL during BPF_CGROUP_RUN_PROG_UDP[46]_SENDMSG_LOCK(). There 
is nothing from the sa_kern->uaddr that is useful to read, so ctx.uaddrlen 
should be 0. The new kfunc bpf_sock_addr_set_addr() can do better than the 
current sa->user_ip[46] uapi and should return error in this case because the 
kernel will eventually ignore it.

> +	} else if (uaddrlen)
> +		ctx.uaddrlen = *uaddrlen;
> +	else if (sk->sk_family == AF_INET)
> +		ctx.uaddrlen = sizeof(struct sockaddr_in);
> +	else if (sk->sk_family == AF_INET6)
> +		ctx.uaddrlen = sizeof(struct sockaddr_in6);
>   
>   	cgrp = sock_cgroup_ptr(&sk->sk_cgrp_data);
> -	return bpf_prog_run_array_cg(&cgrp->bpf, atype, &ctx, bpf_prog_run,
> -				     0, flags);
> +	ret = bpf_prog_run_array_cg(&cgrp->bpf, atype, &ctx, bpf_prog_run,
> +				    0, flags);
> +
> +	if (!ret && uaddrlen)
> +		*uaddrlen = ctx.uaddrlen;
> +
> +	return ret;
>   }
>   EXPORT_SYMBOL(__cgroup_bpf_run_filter_sock_addr);
>   

[ ... ]

> diff --git a/net/ipv6/tcp_ipv6.c b/net/ipv6/tcp_ipv6.c
> index 3a88545a265d..255b02d98404 100644
> --- a/net/ipv6/tcp_ipv6.c
> +++ b/net/ipv6/tcp_ipv6.c
> @@ -135,7 +135,7 @@ static int tcp_v6_pre_connect(struct sock *sk, struct sockaddr *uaddr,
>   
>   	sock_owned_by_me(sk);
>   
> -	return BPF_CGROUP_RUN_PROG_INET6_CONNECT(sk, uaddr);
> +	return BPF_CGROUP_RUN_PROG_INET6_CONNECT(sk, uaddr, NULL);


For IPv6, there is a minimum SIN6_LEN_RFC2133 which excludes the last '__u32 
sin6_scope_id'. Meaning the last 4 bytes in 'struct sockaddr_in6' is optional.

It has been a few months since v2, so my memory faded a bit. I recalled one of 
the concerns on passing addrlen to BPF_CGROUP_RUN_PROG is the bpf prog can 
change it for AF_INET[6] and the addrlen change could be ignored by the kernel.

The bpf_sock_addr_set_addr() in v3 does not allow addrlen changes in AF_INET[6]. 
I think it now makes sense to pass addrlen (of AF_INET[6]) to BPF_CGROUP_RUN_* 
whenever it is available such that the newly added sa_kern->uaddrlen can better 
reflect what is in the sa_kern->uaddr.

I think the only addrlen missing case is in the inet[6]_getname(). Either NULL 
can be passed or create a local addrlen var. afaict, the addrlen must be 'struct 
sockaddr_in[6]' in those cases.