socket: add fallback for nl_connect() by trying to bind to unspecified local port

libnl allows the user to explicitly set the local port before connecting
the socket. A more convenient way is to leave the local port unspecified
and let libnl generate a port id.

As it is, generate_local_port() would try at most 1024 ports, that
means if a user tries to connect more sockets, the automatism will
fail.

Kernel also supports choosing the local port itself (via netlink_autobind()).
So, this could be fixed by always leaving the port unspecified and let
kernel decide on the port. For that we could entirely drop generate_local_port().

There are however problems with that:

  - it is unclear why generate_local_port() was even introduced in the
    first place instead of always relying kernel. This code already
    appeared in libnl-1, so maybe there was a good reason for it or
    it is necessary on some kernel versions.

  - The deprecated libnl-1 library also uses a form of generate_local_port().
    Its first guess would always be getpid(), but the problem is that
    it would not retry on EADDRINUSE. Currently libnl-3 generates ports in
    a different sequence and will not generate a conflicting port (until it
    already exhausted 1016 other ports).
    Hence, currently if your application uses libnl1 and libnl3
    together, the automatism might just work without conflicts
    (commit 1f734a8f892abcd3f81637df4a089155aca1b66a).
    Accidently, kernel/netlink_autobind() also first tries the process
    id as port. That means, if we change libnl-3 to leave the decision
    to kernel, and
      - the application connects sockets both via libnl-1 and libnl-3
      - and the libnl-3 socket happens to connect first
    then the libnl-1 socket would fail to connect without retrying
    another port.

  - Removing generate_local_port() entirely changes behavior in the
    following case:

        sk = nl_socket_alloc();
        /* accessing local port before connecting the socket used to
         * freeze the local port to the generated value. */
        port = nl_socket_get_local_port(sk);
        nl_connect(sk, NETLINK_...);

Maybe the issues are minor and it would simplify the code just to get
rid of the cruft. But instead fix the issue without changing behavior.
Just keep trying with generate_local_port() first, before fallback to
kernel.

Reported-by: Julien Courtat <julien.courtat@6wind.com>
Signed-off-by: Thomas Haller <thaller@redhat.com>

http://lists.infradead.org/pipermail/libnl/2015-June/001889.html

1 files changed, 0 insertions, 3 deletions

diff --git a/lib/socket.c b/lib/socket.c
index a1e0873..eef07f0 100644
--- a/lib/socket.c
+++ b/lib/socket.c
@@ -115,9 +115,6 @@ static uint32_t generate_local_port(void)
 	}
 
 	nl_write_unlock(&port_map_lock);
-
-	/* Out of sockets in our own PID namespace, what to do? FIXME */
-	NL_DBG(1, "Warning: Ran out of unique local port namespace\n");
 	return 0;
 }