[bpf-next,v4,2/3] bpf, sockmap: avoid using sk_socket after free when reading

There are potential concurrency issues, as shown below.
'''
CPU0                               CPU1
sk_psock_verdict_data_ready:
  socket *sock = sk->sk_socket
  if (!sock) return
                                   close(fd):
                                     ...
                                     ops->release()
  if (!sock->ops) return
                                     sock->ops = NULL
                                     rcu_call(sock)
                                     free(sock)
  READ_ONCE(sock->ops)
  ^
  use 'sock' after free
'''

RCU is not applicable to Unix sockets read path, because the Unix socket
implementation itself assumes it's always in process context and heavily
uses mutex_lock, so, we can't call read_skb within rcu lock.

Incrementing the psock reference count would not help either, since
sock_map_close() does not wait for data_ready() to complete its execution.

While we don't utilize sk_socket here, implementing read_skb at the sock
layer instead of socket layer might be architecturally preferable ?
However, deferring this optimization as current fix adequately addresses
the immediate issue.

Fixes: c63829182c37 ("af_unix: Implement ->psock_update_sk_prot()")
Reported-by: syzbot+dd90a702f518e0eac072@syzkaller.appspotmail.com
Closes: https://lore.kernel.org/bpf/6734c033.050a0220.2a2fcc.0015.GAE@google.com/
Signed-off-by: Jiayuan Chen <jiayuan.chen@linux.dev>
---
 net/core/skmsg.c | 13 ++++++++++---
 1 file changed, 10 insertions(+), 3 deletions(-)

On Tue, Apr 8, 2025 at 12:31 AM Jiayuan Chen <jiayuan.chen@linux.dev> wrote:
>
> There are potential concurrency issues, as shown below.
> '''
> CPU0                               CPU1
> sk_psock_verdict_data_ready:
>   socket *sock = sk->sk_socket
>   if (!sock) return
>                                    close(fd):
>                                      ...
>                                      ops->release()
>   if (!sock->ops) return
>                                      sock->ops = NULL
>                                      rcu_call(sock)
>                                      free(sock)
>   READ_ONCE(sock->ops)
>   ^
>   use 'sock' after free
> '''
>
> RCU is not applicable to Unix sockets read path, because the Unix socket
> implementation itself assumes it's always in process context and heavily
> uses mutex_lock, so, we can't call read_skb within rcu lock.
>
> Incrementing the psock reference count would not help either, since
> sock_map_close() does not wait for data_ready() to complete its execution.
>
> While we don't utilize sk_socket here, implementing read_skb at the sock
> layer instead of socket layer might be architecturally preferable ?
> However, deferring this optimization as current fix adequately addresses
> the immediate issue.
>
> Fixes: c63829182c37 ("af_unix: Implement ->psock_update_sk_prot()")
> Reported-by: syzbot+dd90a702f518e0eac072@syzkaller.appspotmail.com
> Closes: https://lore.kernel.org/bpf/6734c033.050a0220.2a2fcc.0015.GAE@google.com/
> Signed-off-by: Jiayuan Chen <jiayuan.chen@linux.dev>
> ---
>  net/core/skmsg.c | 13 ++++++++++---
>  1 file changed, 10 insertions(+), 3 deletions(-)
>
> diff --git a/net/core/skmsg.c b/net/core/skmsg.c
> index 6101c1bb279a..5e913b62929e 100644
> --- a/net/core/skmsg.c
> +++ b/net/core/skmsg.c
> @@ -1231,17 +1231,24 @@ static int sk_psock_verdict_recv(struct sock *sk, struct sk_buff *skb)
>
>  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);
>
> -       if (unlikely(!sock))
> +       rcu_read_lock();
> +       sock = sk->sk_socket;
> +       if (unlikely(!sock)) {
> +               rcu_read_unlock();
>                 return;
> +       }
>         ops = READ_ONCE(sock->ops);
> -       if (!ops || !ops->read_skb)
> +       if (!ops || !ops->read_skb) {
> +               rcu_read_unlock();
>                 return;
> +       }
> +       rcu_read_unlock();

This makes no sense to me. RCU doesn't work this way.

pw-bot: cr