RE: [PATCH bpf v2 2/4] bpf, sockmap: Check for any of tcp_bpf_prots when cloning a listener

[John Fastabend <john.fastabend@gmail.com>]

Jakub Sitnicki wrote:
> A listening socket linked to a sockmap has its sk_prot overridden. It
> points to one of the struct proto variants in tcp_bpf_prots. The variant
> depends on the socket's family and which sockmap programs are attached.
> 
> A child socket cloned from a TCP listener initially inherits their sk_prot.
> But before cloning is finished, we restore the child's proto to the
> listener's original non-tcp_bpf_prots one. This happens in
> tcp_create_openreq_child -> tcp_bpf_clone.
> 
> Today, in tcp_bpf_clone we detect if the child's proto should be restored
> by checking only for the TCP_BPF_BASE proto variant. This is not
> correct. The sk_prot of listening socket linked to a sockmap can point to
> to any variant in tcp_bpf_prots.
> 
> If the listeners sk_prot happens to be not the TCP_BPF_BASE variant, then
> the child socket unintentionally is left if the inherited sk_prot by
> tcp_bpf_clone.
> 
> This leads to issues like infinite recursion on close [1], because the
> child state is otherwise not set up for use with tcp_bpf_prot operations.
> 
> Adjust the check in tcp_bpf_clone to detect all of tcp_bpf_prots variants.
> 
> Note that it wouldn't be sufficient to check the socket state when
> overriding the sk_prot in tcp_bpf_update_proto in order to always use the
> TCP_BPF_BASE variant for listening sockets. Since commit
> b8b8315e39ff ("bpf, sockmap: Remove unhash handler for BPF sockmap usage")
> it is possible for a socket to transition to TCP_LISTEN state while already
> linked to a sockmap, e.g. connect() -> insert into map ->
> connect(AF_UNSPEC) -> listen().
> 
> [1]: https://lore.kernel.org/all/00000000000073b14905ef2e7401@google.com/
> 
> Fixes: e80251555f0b ("tcp_bpf: Don't let child socket inherit parent protocol ops on copy")
> Reported-by: syzbot+04c21ed96d861dccc5cd@syzkaller.appspotmail.com
> Signed-off-by: Jakub Sitnicki <jakub@cloudflare.com>
> ---

Acked-by: John Fastabend <john.fastabend@gmail.com>