iOS applies the tunnel address out of band, after handing the engine the
tun fd, so the receiver bound to that address a moment too early and failed
with EADDRNOTAVAIL. Nothing retried it, leaving the peer able to send but
never to receive for the life of the connection.
Wait for the address to appear and bind then, and let a rebind start a
receiver that is down rather than skipping it. Only the subsystems that
asked to be bound are bound, so the SSH sessions and DNS queries that other
listeners are carrying are left alone.
The wait itself is iOS-only: every other platform assigns the address in the
call chain that creates the interface, or is handed one that already carries
it, so there it compiles down to nothing.
Zoltan Papp
de39cad718