1.36.6 no longer prefers DHCPv6 addresses over SLAAC
The bug has been reported on https://launchpad.net/bugs/1977619
My network has both DHCPv6 and SLAAC (autoconf) for IPv6. From a privacy perspective, for readability reasons and for network management policies, DHCPv6 should always be preferred over SLAAC addresses when available. And according to RFC 6724, the smaller /128 scope of the DHCPv6 address should be chosen over the larger /64 scope of the SLAAC address.
NetworkManager has always been able to adhere to that by simply setting ip6.privacy=0 for the connection (in nm-connection-editor not selecting "Prefer temporary address" for IPv6 privacy extensions). Then it would use the DHCPv6 address as the source for all outgoing traffic.
So if you would - for instance - run
curl ifconfig.co, the DHCPv6 address would be used to connect to the outside world and be echoed back.
Since the update to 1.36.6, this is no longer the case. NetworkManager now routes outgoing traffic through the SLAAC address, even if ip6.privacy=0 is set for the connection.
Constantly removing the SLAAC addresses with
ip addr del or disabling SLAAC RA's on the router are now the only ways to stop NetworkManager from preferring SLAAC over DHCPv6. None of the local options in NetworkManager 1.36.6 are able to restore the previous, desired and correct way of working: the SLAAC address should never be used as the preferred address if a DHCPv6 lease is given.
Looking at the changelog of NetworkManager 1.36.6, multiple things regarding IP address order and temporary addresses have been changed in that release, any of them (or a combination) introducing this bug:
- Fix a bug in synchronization of IP addresses with kernel that could lead to a wrong address order.
- Ignore addresses from DHCPv6 when the Otherconf router advertisement flag is set.
- Ensure temporary IPv6 addresses are removed on disconnect and reapply.
Steps to reproduce:
Connect to a network where the router sends "A" and "M" bits in the RA's and has a DHCPv6 server running (e.g. any OpenWrt router).
ip -6 a, the list now sorts SLAAC addresses above DHCPv6 addresses. With NetworkManager 1.36.4 and earlier, this was not the case. (The Linux kernel uses the address highest in the list as preferred.)
When running something like
curl ifconfig.co, the SLAAC address is being returned, which makes sense as that is now preferred by the kernel. (But it shouldn't be.)
NetworkManager should always sort DHCPv6 addresses above SLAAC addresses, as is the case for all versions prior to 1.36.6 and also corrected again in 1.38.0. In case static addresses are manually set, those should take first priority, with DHCPv6 second and SLAAC/autoconf last.
This can break many real-life use cases. For instance, my router gives out static leases to my machines. Those addresses are whitelisted in all kinds of firewalls to allow me to access servers for my work. Now that the "wrong" address is being preferred for outgoing traffic (a SLAAC address that I have no influence on and cannot centrally configure), I am being locked out of the servers in question unless I forcefully remove the addresses or disable SLAAC on my router, so my outgoing traffic is being routed through the DHCPv6 address again.
Note that "just disabling SLAAC RA's on the router" is not something everybody can do, as it requires root access to the device. Moreover, it would break IPv6 connectivity entirely for devices that don't support DHCPv6 (read: Android).
So this update introduces a very breaking change in IPv6 source address selection to an LTS release, while LTS releases should be stable.
I should note that the bug is not present in NetworkManager 1.38.0 on Debian sid. That just prefers DHCPv6 addresses when available, like it should. As that version is also used in Ubuntu kinetic, most likely this bug is not present there.
Looking at the changelog of 1.38.0:
- Fix bug setting priority for IP addresses.
- Static IPv6 addresses from "ipv6.addresses" are now preferred over addresses from DHCPv6, which are preferred over addresses from autoconf. This affects IPv6 source address selection, if the rules from RFC 6724, section 5 don't give a exhaustive match.
It looks like Ubuntu just introduced that bug by upgrading to 1.36.6, while a proper fix has only landed in 1.38.0, leaving the 1.36.x series now broken. Please either backport the fix from 1.38.0 or revert to 1.36.4, which was working fine.