Don’t call `start_message()` since it may cause an infinite loop.

The exact narration is implemented, with all the specified
optimisations.

- neighbour creation is delayed;
- if multiple PCs are found, only the first one is processed, the
  remaining ones are be silently ignored;
- a node ignores all challenge requests contained in a packet except
  the last one;
- and it ignores a challenge request in the case where it is contained
  in a packet with an Index that matches the one in the Neighbour Table
  and a PC that is smaller or equal to the one contained in the
  Neighbour Table.

s/crypto seqno/pc/g
s/challenge response/challenge reply/g
s/challenge req/challenge request/g

Macros starting with __ are reserved for the implementation, it should
not appear in code.

We used to accept packets signed by any key.

This was completely wrong.

Also rename DIGEST_LEN to MAX_DIGEST_LEN.

Flag to allow empty or wrong hmac in the packet trailer.

Known issues:

  - we create a neighbour entry before the first successful challenge;
  - we compute HMAC for each HMAC TLV rather than just once;
  - we only support sending one HMAC TLV;
  - we don't support key rotation.
Co-authored-by: Clara Do <clarado_perso@yahoo.fr>
Co-authored-by: Weronika Kolodziejak <weronika.kolodziejak@gmail.com>

This is triggered when flushing an interface.

    4 bytes in 1 blocks are definitely lost in loss record 1 of 6
       at 0x483A77F: malloc (vg_replace_malloc.c:307)
       by 0x11155E: check_interface_ipv4 (interface.c:161)
       by 0x110FCD: interface_updown (interface.c:460)
       by 0x1116E2: check_interfaces (interface.c:550)
       by 0x10AFAA: main (babeld.c:546)

RFC 8966 Section 3.2.6 says that the route table is indexed by
triples of the form (prefix, plen, neigh).  Our indexing used to
be incorrect, we'd sometimes include the next hop address, which
made it impossible to interpret retractions with no next hop.

The kind-definitions are contained in the header file:
  LOCAL_FLUSH 0
  LOCAL_ADD 1
  LOCAL_CHANGE 2

The function (local_kind) that converts them into strings again
is private. New addons (for example openwrt ubus bindings) that make
use of babeld should also use the same function to convert the kind
to string format. This allows to ensure uniformity over all addons.
Signed-off-by: Nick Hainke <vincent@systemli.org>

This reverts commit 580257c7.

Thanks to Fabian Blaese.

Setting sysctls is not allowed for unprivileged users, so let's not try
to set per interface rp_filter when `skip-kernel-setup` is set.

After this change babeld can run as an unprivileged users with
CAP_NET_ADMIN. The user needs to take care of setting up the sysctls
themselves.

This means we no longer need to keep source-specific routes at
the beginning, so remove the debugging check (which should never
have made it into production anyway).

Babel-users mailing list threads:
- [Babel-users] Higher CPU usage since 1.9.x leads to instability on slow devices
  https://alioth-lists.debian.net/pipermail/babel-users/2020-May/003734.html
- [Babel-users] Removing source-specific for older kernels [was: Higher CPU usage since 1.9.x leads to instability on slow devices]
  https://alioth-lists.debian.net/pipermail/babel-users/2020-June/003742.html

Rules are obsoleted by the removal of the disambiguation code.
See https://alioth-lists.debian.net/pipermail/babel-users/2020-June/003744.html