The previous choice was not unique in the presence of multiple next hops.

The code assumes that route->channels is NULL when route->channels_len
is 0, such that free(route->channels) will work.

Think about this scenario:
  update(r, some channels)  # route->channels = malloc(…)
  update(r, no channel)  # free(route->channels)
  update(r, no channel)  # free(route->channels)

Thanks to Dave Taht for pointing the issue.

When the v6 prefix is ::/0, then it is considered as less specific than
any v4-mapped prefix.

Remark that "no source prefix" is encoded by "src_plen == 0".  This
might be an issue, since "src_plen == 0" means ::/0, a v6 prefix.  So
the route (1.1.1.1/32 from 0.0.0.0/0) is internally represented by
(1.1.1.1/32 from ::/0).  When compared with  (::/0 from 2001:db8::/32),
the two source prefixes are comparable.  This is not a problem in
practice, since 2 pairs of prefixes are disjoint whether the dst
prefixes or the src prefixes are disjoint.

This allows 0 to appear within a channel list, which is consistent with
draft-chroboczek-babel-diversity-routing-01.

Noted by Matthieu Boutier.

Many of the directives don't make sense after startup.

The absolute path to the socket can be passed to the -g/-G command line
options, or to the new local-path/local-path-readwrite config file
directives.
Signed-off-by: Julien Cristau <jcristau@debian.org>

Now that the parser can be invoked at runtime, we need to be extra
careful with error handling.

Noted by Baptiste Jonglez.