If we create separate listening sockets in each worker using
SO_REUSEADDR, then an incoming connection may be load-balanced
to a process that is already busy, either due to a long-running
request, or because it is a recursive request (in which case we
would deadlock).
Instead, only create one listening socket, and only create worker
forks afterwards. This way the incoming request will be served
by one of the workers that is currently listening for an incoming
connection.