Sunday, February 23, 2014

IPv6 radius accounting is still a mess


Since the beginning of putting IPv6 into production BRAS/BNG (almost 3 years ago), we were facing the following issue: radius accounting records were missing either IPv4 and/or IPv6 address information. When only IPv4 was being used, everything was easy: just wait for IPCP to complete and then send the accounting record. Now, with the addition of IPv6 and most importantly DHCPv6 into the equation, things have become complicated and even worse, trying to fix IPv6 has broken IPv4.

The issue is mainly caused by the extra PPP/NCP negotiation happening due to IPv6 and by the DHCPv6 running asynchronously on top of IPv6 and PPP. Unfortunately most of these issues are unnoticed until you have a lot of IPv6 subscribers and a lot of different CPEs.

All these years, Cisco has tried various solutions in order to fix this issue, maybe following a wrong way.

My proposal?

  • Allow the first accounting packet (start) to be sent into a configurable (preferable small) timeframe, assuming at least one of IPv4/IPv6 address assignments has completed.
  • Allow a second accounting packet (interim/alive) to be sent a little bit later (with a larger timeframe compared to the first one), only if some IPv6 info wasn't included in the first accounting packet.
  • Both timeframes should be configurable by the user.
  • Allow an extra accounting packet (interim/alive) to be sent whenever an DHCPv6 prefix delegation happens, only if its info wasn't included in the previous two accounting packets.

TR-187 from Broadband Forum describes a similar behavior (in R-67 to R-73), but it includes an accounting record per each phase of PPP/NCP/IPCP negotiation.

Here is a more detailed description of my proposal:

Everything happening at the PPP layer, should produce a single accounting record assuming "logical" PPP timeouts forbid the NCP negotiation to continue for a long time.  PPP should complete at least one of the IPv4/IPv6 address negotiations during this time, otherwise the PPP session should be torn down. If at least one protocol address assignment is successful, then send the accounting start record for this protocol and allow for extra accounting (interim/alive) record to be sent as soon as the other protocol completes too. But don't add extra delay in the initial accounting, while waiting for both IPv4/IPv6 address assignment to happen. This initial accounting record (start) should include both IPv4 (IPCP) address and IPv6 (IPV6CP) interface-id, with the capability to also add the ICMPv6 RA prefix if SLAAC is being used. Based on my measurements (on a averagely loaded BRAS/BNG), IPCP/IPV6CP assignments should take less than 100ms, while SLAAC completion might take up to 500ms after PPP completes.

A 2nd accounting record should be sent, just after acquiring any IPv6 address/prefix through DHCPv6 and/or DHCPv6-PD. Since DHCPv6 runs asynchronously on top of IPv6 and you cannot guarantee for the exact timeframe into which the CPE will ask for an address/prefix using it, it's better to have a separate accounting record for it, even if that comes after some seconds. Based on my measurements, DHCPv6 completion might happen up to 3-5 seconds after PPP completes, depending mostly on the CPE's ability to initiate DHCPv6 at the right time (after that it's just some ms until the prefix is assigned).

A 3rd accounting record should be sent on vary rare cases, i.e. when IPv4 assignment completes early, SLAAC and/or DHCPv6 IA-NA are completed afterwards and DHCPv6-PD happens at a later time (although relevant RFCs "forbid" such behavior, the network should be able to cope with "arrogant" CPEs too).

Of course if the DHCPv6/DHCPv6-PD assignment happens into the initial small timeframe, then its info should also be included into the initial accounting packet if possible. Or if SLAAC accounting didn't happen into the initial timeframe, then its info should be included into the second accounting packet.

Also, it's obvious that if PPP protocol rejects are received for IPv4 (i.e. for IPv6-only) or IPv6 (i.e. for IPv4-only), then there shouldn't be any waiting time for the rejected protocol to complete address assignment.

An even better proposal is to fix the periodic accounting inform in order to include every possible type of new information (including IPv4/IPv6 addresses), either as soon as it happens or after some configurable time.

All the above is interesting for conversation assuming that the prefix assigned through DHCPv6-PD is indeed included into the radius accounting packets. There have been numerous tries of fixing that and the IPv4/IPv6 accounting in general ("aaa accounting include auth-profile delegated-ipv6-prefix", "aaa accounting delay-start extended-time", "aaa accounting delay-start [all]", CSCua18679, CSCub63085, CSCuj80527, CSCtn24085, CSCuj09925), so let's hope this time Cisco will be successful.

PS: Some things might work better or worse depending on whether the DHCPv6-PD assignment happens through a local pool or a pool configured on radius. Ivan's blog includes some information too.

 
Creative Commons License
This work is licensed under a Creative Commons Attribution-Noncommercial-Share Alike 3.0 United States License.
Creative Commons License
This work is licensed under a Creative Commons Attribution-Noncommercial-Share Alike 3.0 Greece License.