gst/udp/gstudpsrc.c, near line 495:

  /* If we get here and the readsize is zero, then either select was woken up
   * by activity that is not a read, or a poll error occurred, or a UDP packet
   * was received that has no data. Since we cannot identify which case it is,
   * we handle all of them. This could possibly lead to a UDP packet getting
   * lost, but since UDP is not reliable, we can accept this. */
  if (G_UNLIKELY (!readsize)) {
    /* try to read a packet (and it will be ignored),
     * in case a packet with no data arrived */
    slen = sizeof (sa);
    recvfrom (udpsrc->sock.fd, (char *) &slen, 0, 0, &sa.sa, &slen);

    /* clear any error, in case a poll error occurred */
    clear_error (udpsrc);

    /* poll again */
    goto retry;
  }

If a poll error occurs (for example, connection is lost), recvfrom() sometimes blocks on recvfrom() forever. Since LIVE_LOCK() is previously taken, any attempt to set udpsrc state from PLAYING to PAUSED hangs on LIVE_LOCK() then. This prevents from successful destroying of udpsrc when the other end dies at the same time (this deadlock floats, sometimes it happens and sometimes everything's OK).

I suggest to add MSG_DONTWAIT to flags, so that in case of errors that would block recvfrom(), it goes on without blocking:

recvfrom (udpsrc->sock.fd, (char *) &slen, 0, MSG_DONTWAIT, &sa.sa, &slen)