It turns out that the rp_filter is only disabled if both all/rp_filter
and ifname/rp_filter are set to 0.  This sets and restores all of these
values explicitly, which fixes mysterious issues on distributions that
set rp_filter by default.

Conflicts:
	kernel_netlink.c

Recent linux kernels (roughly 3.x) keep a routing cache for IPv6: each
time a host is contacted, a temporary /128 route is added. These
routes are visible using “ip -6 route show cache”.

Asking babeld to redistribute all IPv6 routes known to the host, using
"redistribute ip ::/0", causes it to announce and de-announce these
/128 routes frequently, leading other nodes to mark them as
unreachable (thus preventing communication with the host described by
the /128 route, even though less specific routes are available).

Minimal fix for a strange race condition.  A more extensive fix
is in the master branch.

This reverts commit 4229bea0.

The logic that attempted to optimise this away was buggy since it
wasn't taking IHU expiration times into account.  At any rate,
this optimisation is probably not worth the complexity.

Of course rp_filter should be reset to its old value, and not to the old
accept_redirects value.
Signed-off-by: Matthias Schiffer <mschiffer@universe-factory.net>

Off-by-one in error checking.  Reported by Gioacchino Mazzurco.

This used to mean "parasitic", but has been removed a long time ago.

The netlink API doesn't natively support atomic route changes.  We used
to attempt to install the new route before removing the old route, but that
would fail in mysterious ways on non-multipath kernels, leading to "stuck"
routes.  Avoid the pain, just take the native approach.

This caused empty lines to cause parse failures.

Reported by Antoine Beaupré.

This fixes an issue where find_best_route could very occasionally
return a route that was not feasible.

This is needed for the stability aware code.

This is brittle, and prevents me from making route selection smarter.

This is the functionality described in Appendix C of RFC 6126.  Its
main purpose is to avoid keeping a full source table, which makes it
possible to implement a subset of Babel in just a few hundred lines of
code.  However, the code for maintaining the source table is already
there, and a parasitic implementation can be simulated using filtering
-- so it makes little sense to keep the functionality.

Thanks to Denis Ovsienko.