When reworking the reception of IPv6 packets I reset a previously set
`ipv6` snip as follows when the IPv6 extension handler returns a
packet (see first hunk of this commit):
```C
ipv6 = pkt->next->next
```
With `gnrc_ipv6_ext` this makes *somewhat* sense, `pkt->next` was
previously equal to `ipv6` and after the function call `pkt->next`
is the marked extension header, while `pkt->next->next` is the IPv6
header. However, since `ipv6` is already write-protected i.e.
`ipv6->users == 1` (see ll. 665-675), any additional call of
`gnrc_pktbuf_start_write()` [won't][start-write-doc] duplicate the
packet. In fact, the only `gnrc_pktbuf_start_write()` in
`gnrc_ipv6_ext` is used to send the *result* to the subscribers of that
extension header type, leaving the original packet unchanged for the
caller. As such `ipv6` remains the pointer to the IPv6 header whether
we set it in the line above or not. So we actually don't need that
line.
However, the extension header handling also returns a packet when
`gnrc_ipv6_ext` is not compiled in. In that case it is just a dummy
define that returns the packet you give provide it which means that
this still holds true: `pkt->next == ipv6`.
So setting `ipv6` in this case is actually harmful, as `ipv6` now
points to the NETIF header [following the IPv6 header][pkt-structure]
in the packet and this causes the `user` counter of that NETIF header
`hdr` to be decremented if `hdr->users > 1` in the write-protection I
removed in hunk 2 of this commit:
```C
/* pkt might not be writable yet, if header was given above */
ipv6 = gnrc_pktbuf_start_write(ipv6);
if (ipv6 == NULL) {
DEBUG("ipv6: unable to get write access to packet: dropping it\n");
gnrc_pktbuf_release(pkt);
return;
}
```
But as we already established, `ipv6->users` is already 1, so we don't
actually need the write protection here either.
Since the packet stays unchanged after the `ipv6` snip, we also don't
need to re-search for `netif_hdr` after the other two lines are
removed.
[start-write-doc]: https://doc.riot-os.org/group__net__gnrc__pktbuf.html#ga640418467294ae3d408c109ab27bd617
[pkt-structure]: https://doc.riot-os.org/group__net__gnrc__pkt.html#ga278e783e56a5ee6f1bd7b81077ed82a7
Since the recursion into `gnrc_ipv6_demux()` was removed in
`gnrc_ipv6_ext`, `gnrc_ipv6.c` is the only user of this function,
so it can be made private. It was only made public so it can be used
from `gnrc_ipv6_ext`.
Since with #10233 we now assume IPv6 packets always to not be
pre-parsed, we can iterate over the extension headers by gradually
"eating" them away. This allows us to move the iteration over them
out of `gnrc_ipv6_ext_demux()` and into `gnrc_ipv6_demux()`.
By moving the iteration over all extension headers out of
`gnrc_ipv6_ext_demux()` we also can
1. simplify the extension header handling a lot, as it now
just a loop inside `gnrc_ipv6_demux()`,
2. remove the recursion to `gnrc_ipv6_demux()` within
`gnrc_ipv6_ext_demux()`.
Since the packet is now guaranteed to be preparsed, the currently
handled IPv6 header will always be in the first snip. Because of this
the packet parser can't get confused anymore which IPv6 header is the
one to be handled so we don't need to remove the more outer ones.
Because of this we can just use the normal packet dispatching (which is
already used by other `GNRC_NETTYPE_*`-known protocol numbers such as
UDP).
This also reverts d54ac38f84.
`_decapsulate()` is called by callees of `_receive()` so the call to
the latter function within the first creates a recursion we don't want.
Using `gnrc_netapi` instead removes that and provides the added benefit
that other subscribers to IPv6 are also informed.
gnrc_sock_recv used to duplicate functionality of gnrc_ipv6_get_header,
but additionally checked whether the IPv6 snip is large enough.
All checks are now included in gnrc_ipv6_get_header, but as most of them
stem from programming / user errors, they were moved into asserts; this
constitutes an API change.
While `tmp` in the loop for write-protection for the check-sum
calculation is used to check the return value of
`gnrc_pktbuf_start_write()`, it was never overwriting `payload` causing
the original snip to be used in the following iteration `prev` when
duplicated, and destroying the sanity of `ipv6`.
This change is a gnrc_ipv6_nib/gnrc_netif(2)-based rework of #7210.
Packet duplication
==================
Its main optimization is that it restructures `gnrc_ipv6` handling of
sent packets so that duplication for write-protection happens at the
latest possible step:
* potential `gnrc_netif` headers added by upper layers are
write-protected before their removal
* This unifies the duplication of the IPv6 header directly after
that
* Extension headers in-between the IPv6 header and the payload header
are duplicated just before the check sum is duplicated
Especially the last point allows for only handing a single packet snip
to all lower functions instead of an already searched IPv6 header
(which now is always the first until it is handed to the interface) +
payload header.
Further clean-ups
=================
* Next-hop link-layer address determination was moved to the
`_send_unicast` function, greatly simplifying the unicast case in the
`_send` function
* Code for loopback case was added to a new function `_send_to_self`
* Removed some code duplication
Parts of [RFC4862] were already implemented when NDP via the NIB was
first implemented. This change just includes the DAD portion of
[RFC4862]. This should be enough to make RIOT fully RFC4862 compliant.
[RFC4862]: https://tools.ietf.org/html/rfc4862
If the payload length is zero and the next header field is not set to
NONXT, GNRC will interpret the current header as the payload because the
first snip is always interpreted as the payload. This can lead to loops
and or crashes.
When the payload length of an encapsulated IPv6 packet is 0, the
`_receive` function of IPv6 can be given a NULL pointer, causing the
IPv6 header checker to crash because of a NULL pointer dereference.