Close encounters of the ICMP type 2 kind (near misses with ICMPv6 PTB)

Operators of popular Internet services face complex challenges associated with scaling their infrastructure. One scaling approach is to utilize equal-cost multi-path (ECMP) routing to perform stateless distribution of incoming TCP or UDP sessions to multiple servers or to middle boxes such as load balancers. Distribution of traffic in this manner presents a problem when dealing with ICMP signaling. Specifically, an ICMP error is not guaranteed to hash via ECMP to the same destination as its corresponding TCP or UDP session. A case where this is particularly problematic operationally is path MTU discovery RFC 1981 PMTUD.

A common application for stateless load balancing of TCP or UDP flows is to perform an initial subdivision of flows in front of a stateful load balancer tier or multiple servers so that the workload becomes divided into manageable fractions of the total number of flows. The flow division is performed using ECMP forwarding and a stateless but sticky algorithm for hashing across the available paths (see RFC 2991 for background on ECMP routing). This nexthop selection for the purposes of flow distribution is a constrained form of anycast topology, where all anycast destinations are equidistant from the upstream router responsible for making the last next-hop forwarding decision before the flow arrives on the destination device. In this approach, the hash is performed across some set of available protocol headers. Typically, these headers may include all or a subset of (IPv6) Flow-Label, IP-source, IP-destination, protocol, source-port, destination-port and potentially others such as ingress interface. A problem common to this approach of distribution through hashing is impact on path MTU discovery. An ICMPv6 type 2 PTB message generated on an intermediate device for a packet sent from a server that is part of an ECMP load balanced service to a client will have the load balanced anycast address as the destination and hence will be statelessly load balanced to one of the servers. While the ICMPv6 PTB message contains as much of the packet that could not be forwarded as possible, the payload headers are not considered in the forwarding decision and are ignored. Because the PTB message is not identifiable as part of the original flow by the IP or upper layer packet headers, the results of the ICMPv6 ECMP hash calculation are unlikely to be hashed to the same nexthop as packets matching the TCP or UDP ECMP hash of the flow. An example packet flow and topology follow. The packet for which the PTB message was generated was intended for the client.

router ecmp -> nexthop L4/L7 load balancer -> destination router --> load balancer 1 ---> \\--> load balancer 2 ---> load-balanced service \--> load balancer N ---> ]]> The router ECMP decision is used because it is part of the forwarding architecture, can be performed at line rate, and does not depend on shared state or coordination across a distributed forwarding system which may include multiple linecards or routers. The ECMP routing decision is deterministic with respect to packets having the same computed hash. A typical case where ICMPv6 PTB messages are received at the load balancer is a case where the path MTU from the client to the load balancer is limited by a tunnel in which the client itself is not aware of. Direct experience says that the frequency of PTB messages is small compared to total flows. One possible conclusion being that tunneled IPv6 deployments that cannot carry 1500 MTU packets are relatively rare. Techniques employed by clients such as happy-eyeballs may actually contribute some amelioration to the IPv6 client experience by preferring IPv4 in cases that might be identified as failures. Still, the expectation of operators is that PMTUD should work and that unnecessary breakage of client traffic should be avoided. A final observation regarding server tuning is that it is not always possible even if it is potentially desirable to be able to independently set the TCP MSS for different address families on some end-systems. On Linux platforms, advmss may be set on a per route basis for selected destinations in cases where discrimination by route is possible. The problem as described does also impact IPv4; however implementation of RFC 4821 TCP MTU probing, the ability to fragment on wire at tunnel ingress points and the relative rarity of sub-1500 byte MTUs that are not coupled to changes in client behavior (for example, endpoint VPN clients set the tunnel interface MTU accordingly to avoid fragmentation for performance reasons) makes the problem sufficiently rare that some existing deployments have choosen to ignore it.

Mitigation of the potential for PTB messages to be mis-delivered involves ensuring that an ICMPv6 error message is distributed to the same anycast server responsible for the flow for which the error is generated. With apppropiate hardware support, mitigation could be done by the mechanism hosts use to identify the flow; by looking into the payload of the ICMPv6 message (to determine which TCP flow it was associated with) before making a forwarding decision. Because the encapsulated IP header occurs at a fixed offset in the ICMP message it is not outside the realm of possibility that routers with sufficient header processing capability could parse that far into the payload. Employing a mediation device that handles the parsing and distribution of PTB messages after policy routing or on each load-balancer/server is a possibility. Another mitigation approach is predicated upon distributing the PTB message to all anycast servers under the assumption that the one for which the message was intended will be able to match it to the flow and update the route cache with the new MTU and that devices not able to match the flow will discard these packets. Such distribution has potentially significant implications for resource consumption and for self-inflicted denial-of-service if not carefully employed. Fortunately, in real-world deployments we have observed that the number of flows for which this problem occurs is relatively small (example, 10 or fewer pps on 1Gb/s or more worth of https traffic in a real world deployment); sensible ingress rate limiters which will discard excessive message volume can be applied to protect even very large anycast server tiers with the potential for fallout limited to circumstances of deliberate duress.

As an alternative, it may be appropriate to lower the TCP MSS to 1220 in order to accommodate 1280 byte MTU. We consider this undesirable as hosts may not be able to independently set TCP MSS by address-family thereby impacting IPv4, or alternatively that middle-boxes need to be employed to clamp the MSS independently from the end-systems. Potentially, extension headers might further alter the lower bound that the MSS would have to be set to, making clamping still more undesirable.

Filter-based-forwarding matches next-header ICMPv6 type-2 and matches a next-hop on a particular subnet directly attached to 1 or more routers. The filter is policed to reasonable limits (we chose 1000pps, more conservative rates might be required in other implementations). Filter is applied on input side of all external (internet or customer facing) interfaces. A proxy located at the next-hop forwards ICMPv6 type-2 packets received at the next-hop to an Ethernet broadcast address (example ff:ff:ff:ff:ff:ff) on all specified subnets. This was necessitated by router inability (in IPv6) to forward the same packet to multiple unicast next-hops. Anycasted servers receive the PTB error and process packet as needed. A simple Python scapy script that can perform the ICMPv6 proxy reflection is included.

This example script listens on all interfaces for IPv6 PTB errors being forwarded using filter-based-forwarding. It removes the existing Ethernet source and rewrites a new Ethernet destination of the Ethernet broadcast address. It then sends the resulting frame out the p2p1 and p2p2 interfaces which attached to vlans where our anycast servers reside.

Alternatively, network designs in which a common layer 2 network exists on the ECMP hop could distribute the proxy onto the end systems, eliminating the need for policy routing. They could then rewrite the destination -- for example, using iptables before forwarding the packet back to the network containing all of the server or load balancer interfaces. This implmentation can be done entirely within the Linux iptables firewall. Because of the distributed nature of the filter, more conservative rate limits are required than when a global rate limit can be employed. An example ip6tables / nftables rule to match icmp6 traffic, not match broadcast traffic, impose a rate limit of 10 pps, and pass to a target destination would resemble:

As with the scapy example, once the destination has been rewritten from a hardcoded ND entry to an Ethernet broadcast address -- in this case to an IPv6 documentation address -- the traffic will be reflected to all the hosts on the subnet.

There are several ways that improvements could be made to the problem how to ECMP load balance of ICMPv6 PTB messages. little in the way of Internet protocol specification change is required, rather we forsee practical implemention change which insofar as we are aware does not exist in current router switch or layer3/4 load balancers. alternatively improved behavior on the part of client/server detection of path mtu in band could render the behavior of devices in the path irrelevant. Routers with sufficient capacity within the lookup process could parse all the way through the L3 or L4 header in the ICMPv6 payload beginning at bit offset 32 of the ICMP header. By reordering the elements of the hash to match the inward direction of the flow, the PTB error could be directed to the same next-hop as the incoming packets in the flow. The FIB (Forwarding Information Base) on the router could be programmed with a multicast distribution tree that included all of the necessary next-hops, and unicast ICMPv6 packets could be policy routed to these destinations. Ubiquitous implementation of RFC 4821 Packetization Layer Path MTU Discovery would probably go a long way towards reducing dependence on ICMPv6 PTB by end systems.

The authors would like to thank Marak Majkowsiki for contributing text, examples, and a very close review. The authors would like to thank Mark Andrews, Brian Carpenter, Nick Hilliard and Ray Hunter, for review.

This memo includes no request to IANA.

The employed mitigation has the potential to greatly amplify the impact of a deliberately malicious sending of ICMPv6 PTB messages. Sensible ingress rate limiting can reduce the potential for impact; however, legitimate PMTUD messages may be lost once the rate limit is reached; analogous to other cases where DOS traffic can crowd out legitimate traffic. The proxy replication results in devices on the subnet not associated with the flow that generated the PTB, being recipients of the ICMPv6 PTB message; which contains a large fragment of the packet that exceeded the allowable MTU. This replication of the packet freagment could arguably result in information disclosure. Recipient machines should be in a common administrative domain.