We'll have to revive UDP because we experienced congestion with TCP.
This should make UDP efficient in good environment.
MTU discovery is required however to enable UDP by default.

This fixes up commit e3781aff
(Reduce probability of dying when the system lacks memory").

- new -V/--version command line option
- protocol extended to get the version of any node in the network,
  which will allow to track those running an old version of re6st

It's already in 're6stnet_subnet' and it's useful to bind to 're6stnet_ip'
without having to edit it first.

256 * (1 - (1-42/256.)**(15/4.)) ≈ 125
where:
- 42 is default rtt-decay
- 4 is default hello in babeld
- 15 is default hello in re6st

This makes sure re6st does not stop working because a connection is stuck.

Also accepts packets from loopback.

It's normal such failure happens occasionally and re6st retries later so:
- do not frighten user/admin with 500 status and ssl errors
- do not waste resources by killing session with registry

This prevents re6stnet from killing tunnels prematurely.

Current implementation is too aggressive and after some time, babeld stops
working properly, with log full of:

  setsockopt(IPV6_JOIN_GROUP): Cannot allocate memory
  setsockopt(IPV6_LEAVE_GROUP): Cannot assign requested address

Some servers can only be accessed via their re6st IP. re6st itself uses little
memory so it should not die when it fails to fork.

This is a workaround waiting that we have better criteria to select tunnels
to create or destroy.

UDP protocol is useless if nothing is done to prevent fragmentation.
Otherwise, it is at best unefficient.

There exist routers on the internet that filter fragmented packets with specific
data. This is hard to debug because TCP connections hang randomly when there is
no OpenVPN encryption.

Now, only TCP is enabled by default. A second protocol should be there for
better performance when possible, either existing UDP one (provided it is
guaranteed there is no fragmentation) or something better (GRE ?).

For the moment, this does nothing since it should always be 1500.

This reverts commit 7dbc38d7.

Fixed upstream in version 0.2
See also http://bugs.debian.org/cgi-bin/bugreport.cgi?bug=725344

The way peer addresses were exchanged polluted caches with information about
dead nodes. In particular, bootstrapping often took a long time because the
cache of the primary node was mostly useless.

This also fixes bootstrap of registry.

Traceback (most recent call last):
  File "/opt/re6stnet/re6stnet", line 403, in main
    tunnel_manager.handlePeerEvent()
  File "/cygdrive/d/slapos/opt/git/re6stnet/re6st/tunnel.py", line 472, in handlePeerEvent
    self.sock.sendto('\1' + ''.join(msg), address)

This fixes a regression introduced with new protocol to registry (commit
e24eb3f5), which can't transport None value.