bpf: reject unhashed sockets in bpf_sk_assign

The semantics for bpf_sk_assign are as follows:

    sk = some_lookup_func()
    bpf_sk_assign(skb, sk)
    bpf_sk_release(sk)

That is, the sk is not consumed by bpf_sk_assign. The function
therefore needs to make sure that sk lives long enough to be
consumed from __inet_lookup_skb. The path through the stack for a
TCPv4 packet is roughly:

  netif_receive_skb_core: takes RCU read lock
    __netif_receive_skb_core:
      sch_handle_ingress:
        tcf_classify:
          bpf_sk_assign()
      deliver_ptype_list_skb:
        deliver_skb:
          ip_packet_type->func == ip_rcv:
            ip_rcv_core:
            ip_rcv_finish_core:
              dst_input:
                ip_local_deliver:
                  ip_local_deliver_finish:
                    ip_protocol_deliver_rcu:
                      tcp_v4_rcv:
                        __inet_lookup_skb:
                          skb_steal_sock

The existing helper takes advantage of the fact that everything
happens in the same RCU critical section: for sockets with
SOCK_RCU_FREE set bpf_sk_assign never takes a reference.
skb_steal_sock then checks SOCK_RCU_FREE again and does sock_put
if necessary.

This approach assumes that SOCK_RCU_FREE is never set on a sk
between bpf_sk_assign and skb_steal_sock, but this invariant is
violated by unhashed UDP sockets. A new UDP socket is created
in TCP_CLOSE state but without SOCK_RCU_FREE set. That flag is only
added in udp_lib_get_port() which happens when a socket is bound.

When bpf_sk_assign was added it wasn't possible to access unhashed
UDP sockets from BPF, so this wasn't a problem. This changed
in commit 0c48eefae7 ("sock_map: Lift socket state restriction
for datagram sockets"), but the helper wasn't adjusted accordingly.
The following sequence of events will therefore lead to a refcount
leak:

1. Add socket(AF_INET, SOCK_DGRAM) to a sockmap.
2. Pull socket out of sockmap and bpf_sk_assign it. Since
   SOCK_RCU_FREE is not set we increment the refcount.
3. bind() or connect() the socket, setting SOCK_RCU_FREE.
4. skb_steal_sock will now set refcounted = false due to
   SOCK_RCU_FREE.
5. tcp_v4_rcv() skips sock_put().

Fix the problem by rejecting unhashed sockets in bpf_sk_assign().
This matches the behaviour of __inet_lookup_skb which is ultimately
the goal of bpf_sk_assign().

Fixes: cf7fbe660f ("bpf: Add socket assign support")
Cc: Joe Stringer <joe@cilium.io>
Signed-off-by: Lorenz Bauer <lmb@isovalent.com>
Reviewed-by: Kuniyuki Iwashima <kuniyu@amazon.com>
Link: https://lore.kernel.org/r/20230720-so-reuseport-v6-2-7021b683cdae@isovalent.com
Signed-off-by: Martin KaFai Lau <martin.lau@kernel.org>

net/core/filter.c

@@ -7353,6 +7353,8 @@ BPF_CALL_3(bpf_sk_assign, struct sk_buff *, skb, struct sock *, sk, u64, flags)
 		return -ENETUNREACH;
 	if (unlikely(sk_fullsock(sk) && sk->sk_reuseport))
 		return -ESOCKTNOSUPPORT;
+	if (sk_unhashed(sk))
+		return -EOPNOTSUPP;
 	if (sk_is_refcounted(sk) &&
 	    unlikely(!refcount_inc_not_zero(&sk->sk_refcnt)))
 		return -ENOENT;