GitLab

increased the timeout for other CI too

This commit cleans up the reuse port setup, and fixes a problem:
make sure to filter addrs out that we simply cannot dial with
(e.g. loopback -> non-loopback, or linklocal -> nonlinklocal)

Good thing, i checked one wrong, too!

the travis docs say the env flag TRAVIS=true but, i want
to make absolutely sure. explicitness leaves less room for
error.

this is actually pretty important in case some peers end
up changing addrs, so others dont fail dialing simply for
picking the wrong addr to start with.

This commit introduces a backoff when failing to dial
peers. It makes everything much faster.

* filtering InterfaceListenAddresses(s) instead
* return error if for loops skip
* large dial timeout

This adds two checks after a successful conn.Dial
* if the remote peer is not who we wanted, close conn
* if the remove peer is outselves, close conn

(the second is redundant, but the codebase may evolve to
end up disabling the first check, so keeping the second
in place helps)

note:
Loopback addresses are actually sent out (they _have to be_,
in cases where there are >1 node in the same machine), so
many times when trying connections, nodes end up dialing
themselves.

cc @whyrusleeping

Instead of erroring out, which would break the listener,
we instead log a message and continue. This is not an error,
the internet is a place with lots of probing + connection
failures.

This should handle early breakages, where a failing connection
would take out the listener entirely.

There are probably other errors we should be handling here,
like secure connection failures.