This had never been implemented, we were silently ignoring duplicate
configuration lines.

Noted by Matthieu Boutier.

As suggested by Dave Taht and Julien Cristau.

Link-local addresses are defined as being in fe80::/10, but they always
start with 54 bits of zeroes.  Test the whole first byte.

Stress the asymmetry with timeval_add_msec, which is purely functional.

Make the intervals the same on wired and wireless links.  The previous
defaults worked well in practice, but people found the discrepancy confusing.

Thanks to Stephen Fisher.

That's not the issue -- the issue is with the kernel dropping routes when
we loose an IP address.

rolling back:

Fri Mar  4 21:26:15 CET 2011  Juliusz Chroboczek <jch@pps.jussieu.fr>
  * Fix ordering of netlink messages when making a route unreachable.

  This should fix:

    netlink_read: File exists
    netlink_read: No such process

    M ./kernel_netlink.c -1 +1

Apparently this is needed for bionic libc.

This should fix:

  netlink_read: File exists
  netlink_read: No such process

/dev/arandom, which was previously used on OpenBSD, is not useful since
we're only using it for seeding the prng.