Message-ID: <dc141b3c07fa4d51ad48ac87718f7c98@AcuMS.aculab.com>
Date:   Wed, 2 Feb 2022 22:31:18 +0000
From:   David Laight <David.Laight@...LAB.COM>
To:     'Dan Williams' <dcbw@...hat.com>,
        "netdev@...r.kernel.org" <netdev@...r.kernel.org>
CC:     David Ahern <dsahern@...nel.org>
Subject: RE: Getting the IPv6 'prefix_len' for DHCP6 assigned addresses.

From: Dan Williams
> Sent: 02 February 2022 18:55
> 
> On Wed, 2022-02-02 at 16:58 +0000, David Laight wrote:
> > I'm trying to work out how DHCP6 is supposed to work.
> >
> > I've a test network with the ISC dhcp6 server and radvd running.
> > If I enable 'autoconf' I get a nice address with the prefix from
> > radvd and the last 8 bytes from my mac address, prefix_len 64.
> > I get a nice address from dhcp6 (busybox udhcpc6) with the same prefix.
> >
> > udhcpc6 runs my scripts and 'ip add $ipv6 dev $interface' adds the
> > address.
> > But the associated prefix_len is /128.
> >
> > All the documentation for DHCP6 says the prefix_len (and probably the
> > default route - but I've not got that far) should come from the network
> > (I think from RA messages).
> >
> > But I can't get it to work, and google searches just seem to show
> > everyone else having the same problem.
> >
> > The only code I've found that looks at the prefix_len from RA messages
> > is that which adds to 'autoconf' addresses - and that refuses to do
> > anything unless the prefix_len is 64.
> >
> > I can't see anything that would change the prefix_len of an address
> > that dhcp6 added.
> >
> > Has something fallen down a big crack?
> 
> I'm far from an expert, but I don't think anything has fallen down a
> crack. I'm sure David Ahern or somebody else will correct me, but here
> goes:
> 
> Things are working as intended.
> 
> DHCPv6 is not a complete IPv6 addressing solution. It must be used in
> combination with Router Advertisements to do generally useful things.
> 
> https://datatracker.ietf.org/doc/html/rfc8415#section-21.6
> 
> 21.6.  IA Address Option
> 
>       IPv6-address         An IPv6 address.  A client MUST NOT form an
>                            implicit prefix with a length other than 128
>                            for this address.  A 16-octet field.
> 
> DHCPv6 intentionally doesn't tell you who your IPv6 router (gateway in
> v4-land) is. That's what the Router Advertisement is for.
> 
> DHCPv6 intentionally doesn't tell you anything about what prefixes are
> "on-link" like what the subnet mask implies for IPv4. That's what the
> Router Advertisement is for.

That's what I'd gathered.

> If the router sends an RA with a Prefix Information Option (PIO) with
> the "on-link" (L) bit set then the kernel should install on-link routes
> for that prefix. If your DHCPv6-provided address falls within one of
> those prefixes then kernel routing takes over and packets go where they
> should regardless of the /128.
> 
> If you don't have RAs, or don't have those routes installed because the
> router wasn't sending a PIO+L for the DHCP-provided prefixes, then yeah
> things aren't going to work like you might expect.

They are being sent, because if I enable "autoconf" I get the prefix
address based on my MAC address.
But I don't want that "autoconf" addreess, I only want the dhcp6 address.

It has to be said I was probably looking at the output from 'ip addr'
not 'ip route show'.
Checking things on the test system is a bit of a PITA due to its
limited userspace, but at least I've stopped dhcp6 deleting the IPv4
address!

But I (probably) found the code that adds the "autoconf" address and
route - it has a check that the prefix_len is 64.

It may well be that you need to use dhcp6 with a /80 prefix because
your isp has only decided to give you a single /64 address.

> I'm sure David will be along to correct me soon though...

Indeed.

	David

-
Registered Address Lakeside, Bramley Road, Mount Farm, Milton Keynes, MK1 1PT, UK
Registration No: 1397386 (Wales)

Powered by blists - more mailing lists