Re: [PATCH bpf-next 3/7] skmsg: introduce sk_psock_hooks

[Jakub Sitnicki <jakub@cloudflare.com>]

On Tue, Feb 25, 2020 at 02:56 PM CET, Lorenz Bauer wrote:
> The sockmap works by overriding some of the callbacks in sk->sk_prot, while
> leaving others untouched. This means that we need access to the struct proto
> for any protocol we want to support. For IPv4 this is trivial, since both
> TCP and UDP are always compiled in. IPv6 may be disabled or compiled as a
> module, so the existing TCP sockmap hooks use some trickery to lazily
> initialize the modified struct proto for TCPv6.
>
> Pull this logic into a standalone struct sk_psock_hooks, so that it can be
> re-used by UDP sockmap.
>
> Signed-off-by: Lorenz Bauer <lmb@cloudflare.com>
> ---
>  include/linux/skmsg.h |  36 ++++++++-----
>  include/net/tcp.h     |   1 -
>  net/core/skmsg.c      |  52 +++++++++++++++++++
>  net/core/sock_map.c   |  24 ++++-----
>  net/ipv4/tcp_bpf.c    | 114 ++++++++++++------------------------------
>  5 files changed, 116 insertions(+), 111 deletions(-)
>

[...]

> diff --git a/net/ipv4/tcp_bpf.c b/net/ipv4/tcp_bpf.c
> index 90955c96a9a8..81c0431a8dbd 100644
> --- a/net/ipv4/tcp_bpf.c
> +++ b/net/ipv4/tcp_bpf.c

[...]

> @@ -650,9 +599,10 @@ int tcp_bpf_init(struct sock *sk)
>   */
>  void tcp_bpf_clone(const struct sock *sk, struct sock *newsk)
>  {
> -	int family = sk->sk_family == AF_INET6 ? TCP_BPF_IPV6 : TCP_BPF_IPV4;
> -	struct proto *prot = newsk->sk_prot;
> +	struct proto *prot = READ_ONCE(sk->sk_prot);

For the sake of keeping the review open - we've identified a regression
here. sk->sk_prot can change by the time we get here, since the moment
we copied the listener sock. We need to stick to checking newsk->sk_prot
here.

>
> -	if (prot == &tcp_bpf_prots[family][TCP_BPF_BASE])
> +	/* TCP_LISTEN can only use TCP_BPF_BASE, so this is safe */
> +	if (unlikely(prot == &tcp_bpf_ipv4[TCP_BPF_BASE] ||
> +	             prot == &tcp_bpf_ipv6[TCP_BPF_BASE]))
>  		newsk->sk_prot = sk->sk_prot_creator;
>  }