This text is the architecture results of the secure (zero-touch) design team.
The realization that we should use loose source routes is due to Pascal.

I believe this goes into a new section between 10. Centralized vs.
and 11. IANA Considerations.  Interspersed is some text that I think goes in to section
12. Security Considerations. 

If you are interested in the history of this text, or some additional more implementation
side ideas, then I recommend reviewing the diff
     http://www.ietf.org/rfcdiff?url1=draft-richardson-6tisch--security-6top-03&difftype=--html&submit=Go%21&url2=draft-richardson-6tisch--security-6top-04


10B.  Architectural requirements of join protocol

10B.1.  Entities involves in the join process

   The following actors are involved: there is a new joining node.  It
   is (radio) adjacent to the join assistant.  The join assistant is
   part of the production network, and participates in one or more
   DODAGs, such that it is reachable from the 6LBR, and the JCE.

10B.2.  Join Protocol deliverables

   This section works from the ultimate goal, and goes backwards to
   prerequisite actions.  (Section 6 presents the protocol from
   beginning to end order)

   The ultimate goal of the join protocol is to provide a new node with
   a locally significant security credentials that it is able to take
   part in the network directly.  The credentials may vary by
   deployment.  They can be one of:


   1)  a network-wide shared symmetric key (this is the production
       network key, or MasterKey)

   2)  a locally significant (one-level only) 802.11AR type LDevID
       certificate (which allows it to negotiate a pair-wise keys)

   One of these two items is delivered by JCE to the joining node using
   the 6top protocol.  That is, the JCE provisions using a secure
   "northbound" interface.  The authentication of this interface the
   subject of the Join Protocol as explained below.

   The above credential are used for authentication to generate per-peer
   L2, using a key exchange protocol.  The choice of key exchange
   protocol, and what kind of link and multicast keying needs to be done
   is also provisioned by the JCE.  There are a number of options for
   doing this, such as:

   1)  Mesh Link Exchange [I-D.kelsey-intarea-mesh-link-establishment]
       (IMPORTANT, a good option.  Uses certificates from common CA)

   2)  work in 802.15.9 (uses certificates from common CA)

   3)  Security Framework and Key Management Protocol Requirements for
       6TiSCH [I-D.ohba-6tisch-security] (this document provides the
       phase 0 required, using the network-wide shared key)

   4)  Layer-2 security aspects for the IEEE 802.15.4e MAC
       [I-D.piro-6tisch-security-issues]: the MasterKey is used to
       derive per-peer L2 keys

   Per-peer L2 keying is critical when doing peer-2-peer schedule
   negotiation over 15.4 Information Elements.  Therefore a network-wide
   layer-2 key is inappropriate for the self-organizing networks, and a
   protocol (MLE, 802.15.9) SHOULD be used to derive per-peer L2 keys.

   For networks where there is a PCE present and it will do all schedule
   computation, then the only trust relationship necessary is between
   the individual node and the PCE, and it MAY be acceptable to have a
   network-wide L2 key derived in ways such as
   [I-D.piro-6tisch-security-issues] describes in section ?.  The trust
   relationship between the PCE and the joining node flows from the
   LDevID certificate loaded by the JCE.  When the PCE and JCE are co-
   located, the existing 6top/DTLS security association could be reused.

10B.3.  Join Protocol connection setup

   The intermediate level goal of the join protocol is to enable a Join
   Coordination Entity (JCE) to reach out to the new node, and install
   the credentials detailed above.  The JCE must authenticate itself to
   the joining node so that the joining node will know that it has
   joined the correct network, and the joining node must authenticate
   itself to the JCE so that the JCE will know that this node belongs in
   the network.  This two way authentication occurs in the 6top/CoAP/
   DTLS session that is established between the JCE and the joining
   node.

   [I-D.ietf-6tisch-6top-interface] presents a way to interface to a
   6top information model (defined in YANG).  [I-D.ietf-6tisch-coap]
   explains how to access that information model using CoAP.  This
   information model will include security attributes for the network.
   The JCE would therefore reach out to the joining node and simply
   provision appropriate security properties into the joining node, much
   like the PCE will provision schedules.

   This 6top-based secure join protocol has defined a push model for
   security provisioning by the JCE.  This has been done for three
   reasons:

   1)  6tisch nodes already have to have a 6top CoAP server for schedule
       provising

   2)  this permits the JCE to manage how many nodes are trying to join
       at the same time, and limit how much bandwidth/energy is used for
       the join operation, and also for the JCE to prioritize the join
       order for nodes.

   3)  making the JCE initiate the DTLS connection significantly
       simplies the certificate chains that must be exchanged as the
       most constrained side (the joining node) provides it's
       credentials first, and lets the less constrained JCE figure out
       what kind of certificate chain will be required to authenticate
       the JCE to the joining node.  In EAP-TLS/802.1x situations, the
       TLS channel is created in the opposite direction, and it would
       have to complete in a tentative way, and then further
       authorization occur in-band.

10B.4.  End to End considerations for Join Protocol

   In order for a 6top/DTLS/CoAP connection to occur between the JCE and
   the joining node, there needs to be end-to-end IPv6 connectivity
   between those two entities.  The joining node will not participate in
   the route-over RPL mesh, but rather will be seen by the network as
   being a 6lowpan only leaf-node.

   The joining node sends traffic to the join assistant, which forwards
   it using the normal RPL DODAG upwards routes, and which will reach
   the JCE using regular routing.

   The challenge is getting connectivity from the JCE to the joining
   node, as the DODAG will have no information about this node.
   Instead, the JCE will use loose-source routes to address packets
   first to the Join Assistant, which will then forward to the joining
   node.  This has an interaction with the (strict) source routes used
   by non-storing DODAGs which would be used by the 6LBR to reach the
   Join Assistant.

   There are some alternatives to having full end to end connectivity
   which are discussed in the security considerations section.

12.1.  Security Considerations: alternatives to source routes

   This goes into security-considerations section

   A number of alternatives were considered for creating end to end
   connectivity between the joining node and the JCE, but were rejected.

   (1)  IPIP tunnel between Join Assistant and JCE

   (2)  using straight RPL routing: the Join Assistant sends a DAO

   (3)  using a separate RPL DODAG for join traffic

   (4)  establishing a specific multi-hop 6tisch track for join traffic
      for each Join Assistant

12.1.1.1.  IPIP tunnel

   This mechanism requires 40 bytes overhead per packet, and may require
   specific state to be created on the Join Assistant, or may open the
   network up to a resource exhaustion by malicious join nodes.

   This mechanism was rejected on due to byte count, because it would
   require code only needed for join, and because doing it securely may
   require a per-tunnel state to be maintained on the join assistant.

12.1.1.2.  join existing DODAG

   This mechanism is to just have the new node join the DODAG as a leaf
   node.  The join assistant would issue a DAO for the new node.  If a
   storing DODAG was used, this would directly cost state in all parent
   nodes of the Join Assistant, subjecting the lower-rank parts of the
   network to significant denial of service attacks.

   On a non-storing DODAG the situation is significantly better: no
   state is created by the presence of the leaf node except in the 6LBR,
   where available resources may be higher.

   This mechanism was rejected in favour of the loose source routed
   option as this the loose source routed option always works even for
   storing DODAGs, and is otherwise exactly equivalent in terms of
   network bandwidth cost.

12.1.1.3.  join a DODOAG for joining

   One option to deal with the problem of resource consumption in the
   storing DODAG case is to use a second DODAG.

   This mechanism was rejected as it consumes resources in all nodes for
   the second DODAG, and many implementations support on a single DODAG.
   While storing DODAGs have a lighter network impact than non-storing
   ones (due to the lack of source routes), the PCE can control how much
   network bandwidth can be wasted by malicious join nodes using the
   regular 6tisch mechanisms, and this can be tuned.  A second DODAG
   would be a sunk cost with no tuning possible.

12.1.1.4.  create 6tisch track to form mesh-under

   This mechanism would use 6tisch tracks so that traffic from the JCE
   would be switched from 6LBR to join node.  A track would be required
   to each join assistant.  This is an optimized form of source routing.

   This mechanism was not rejected, rather it was observed that it may
   be applied to any mechanism that uses source routing, and is an
   optimization; it has a certain cost in the form of intermediary node
   state, but that this state is not created by a malicious join node,
   rather it is created by the PCE.

10B.5.  Join node discovery mechanism

   Continuing to work backwards, in order the JCE reach out to provision
   the Joining Node, it needs to know that the new node is present.
   This is done by taking advantage of the 6lowPAN Address Resolution
   Option (ARO) (section 4.1 [RFC6775]).  The ARO causes the new address
   to also be sent up to the 6LBR for duplicate detection using the DAR/
   DAC mechanism.  The 6LBR simply needs to tell the JCE about this.  A
   new protocol may be required for the situation where the JCE, PCE and
   6LBR are not co-located; it may be that this protocol could be simply
   a DAR in some encapsulation.  The details of this are currently out
   of scope for the architecture, as they affect interoperability
   between different vendors of JCE/6LBR/PCE rather than
   interoperability between the assistance/offload infrastructure, and
   the contrained nodes.  Future work may specify this part.

   In addition to needing to know the joining devices address from the
   DAR/NS, the JCE also needs to know the joining node's IDevID.  This
   is to determine if the node should even get any attention.  At this
   point, the IDevID can be forged, it will be authenticated during the
   setup of the 6top control protocol in the DTLS certificate exchange.
   If the serialNumber attribute of the IDevID is less than 64 bits,
   then it is possible that it could be placed into the EUI-64 option of
   the ARO, or the OUI of the [I-D.thubert-6lo-rfc6775-update-reqs]
   EARO.  The JCE needs to know the joining node's serialNumber to know
   if this is device that it should even attempt to provision; and if
   so, it may need to retrieve an appropriate certificate chain (see
   [I-D.richardson-6tisch-idevid-cert]) from the Factory in order for
   the JCE to prove it is the legitimate owner of the joining node.

   Neither 802.1AR nor [RFC5280] provide any structure for the
   serialNumber, except that they are positive integers of up-to 20
   octets in size (numbers up to 2^160).  This specification would
   require that the serialNumber encoded in the IDevID be the same as
   the EUI-64 used by the device.  Some consideration needs to be given
   as to whether there are privacy considerations to doing this: any
   observer that can see the join traffic, can also see the source MAC
   address of the node as well.

   Prior to being able to announce itself in a NS, the joining node
   needs to find the Join Network.  This is done by listening to an
   extended beacon which are broadcast in designated slotframes by Join
   Assistants.  The Extended Beacon provides a way for the Joining Node
   to synchronize itself to the overall timeslot schedule and provides
   an Aloha period in which the Joining Node can send a Router
   Soliticiation, and receive an appropriate Router Advertisement giving
   the Joining Node a prefix and default route to which to send join
   traffic.

   It may be possible to eliminate a message exchange if space for a
   Router Advertisement can be found as part of the Join Network
   Extended Beacon.  This Enhanced Beacon would be distinct to the Join
   Network, and would be encrypted with the well-known Join Network key.

10B.5.1.  prefixes to use for join traffic

   What prefix would the joining node for communication?  There are
   three options:

   (1)  just use link-local addresses (this may require that some
      traffic between 6LBR and JCE be IPIP tunneled)

   (2)  use a prefix specifically for join traffic (may be easier with a
      join-only DODAG)

   (3)  use the same prefix as the rest of the traffic (may require more
      complex ACLs, and leaks information to attackers)

   The simplest method is to use the link-local addresses for the 6top
   connection, and the cost of an IPIP tunnel between the unconstrained
   6LBR and the JCE is not considered significant.



-- 
Michael Richardson <mcr+IETF@sandelman.ca>, Sandelman Software Works
 -= IPv6 IoT consulting =-