On Tue, May 3, 2022 at 6:04 PM Tetsuo Handa
[off-list ref] wrote:

On 2022/05/04 7:37, Eric Dumazet wrote:

quoted

quoted

I think we merged this patch too soon.

My question is : What prevents rds_tcp_conn_path_connect(), and thus
rds_tcp_tune() to be called
after the netns refcount already reached 0 ?

I guess we can wait for next syzbot report, but I think that get_net()
should be replaced
by maybe_get_net()

Yes, syzbot was fast to trigger this exact issue:

Does maybe_get_net() help?

Since rds_conn_net() returns a net namespace without holding a ref, it is theoretically
possible that the net namespace returned by rds_conn_net() is already kmem_cache_free()d
if refcount dropped to 0 by the moment sk_alloc() calls sock_net_set().

Nope. RDS has an exit() handler called from cleanup_net()

(struct pernet_operations)->exit() or exit_batch() :
rds_tcp_exit_net() (rds_tcp_kill_sock())

This exit() handler _has_ to remove all known listeners, and
definitely cancel work queues (synchronous operation)
before the actual "struct net" free can happen later.

rds_tcp_conn_path_connect() {
  sock_create_kern(net = rds_conn_net(conn)) {
    __sock_create(net = rds_conn_net(conn), kern = 1) {
      err = pf->create(net = rds_conn_net(conn), kern = 1) {
        // pf->create is either inet_create or inet6_create
        sk_alloc(net = rds_conn_net(conn), kern = 1) {
          sk->sk_net_refcnt = kern ? 0 : 1;
          if (likely(sk->sk_net_refcnt)) {
            get_net_track(net, &sk->ns_tracker, priority);
            sock_inuse_add(net, 1);
          }
          sock_net_set(sk, net);
        }
      }
    }
  }
  rds_tcp_tune() {
    if (!sk->sk_net_refcnt) {
      sk->sk_net_refcnt = 1;
      get_net_track(net, &sk->ns_tracker, GFP_KERNEL);
      sock_inuse_add(net, 1);
    }
  }
}

"struct rds_connection" needs to hold a ref in order to safely allow
rds_tcp_tune() to call maybe_get_net(), which in turn makes pointless
to use maybe_get_net() from rds_tcp_tune() because "struct rds_connection"
must have a ref. Situation where we are protected by maybe_get_net() is
quite limited if long-lived object is not holding a ref.

Hmm, can we simply use &init_net instead of rds_conn_net(conn) ?

Only if you plan making RDS unavailable for non init netns.