Limitations of OpenFlow Topology Discovery Protocol

TL;DR

This paper critically examines the security and efficiency limitations of the OpenFlow Discovery Protocol (OFDP) and introduces sOFTD as a more secure and efficient alternative for topology discovery.

Contribution

The paper identifies key limitations of OFDP and proposes sOFTD, a new topology discovery protocol with enhanced security and efficiency features.

Findings

OFDP has significant security vulnerabilities.

OFDP is inefficient for large-scale networks.

sOFTD outperforms OFDP in security and efficiency.

Abstract

OpenFlow Discovery Protocol (OFDP) is the de-facto protocol used by OpenFlow controllers to discover the underlying topology. In this paper, we show that OFDP has some serious security, efficiency and functionality limitations that make it non suitable for production deployments. Instead, we briefly introduce sOFTD, a new discovery protocol with a built-in security characteristics and which is more efficient than traditional OFDP.

Figures

Full text

*keywords - * Software-Defined Networking, OpenFlow, Topology Discovery, security.

I Introduction

The separation between the control plane and the data plane introduced by Software-Defined Networking (SDN) allows operators to employ quite damn, remarkably cheap but very fast hardware to forward packets, moving the control logic to a centralized and much smarter entity called controller. The controller plays the role of an operating system of the network. It abstracts the underlying forwarding hardware details and offers high level APIs that the network admins leverage to program their networks. One of the fundamental functions that a controller must offer is an accurate, near real time visibility of the network topology. This function is known as Topology Discovery. Topology discovery in SDN is more sensitive compared to traditional networks based on Link-State routing protocols like OSPF. In SDN, To discover the network topology, all current OpenFlow controllers implement the same protocol OFDP (OpenFlow Discovery Protocol).

Figure 1 shows how OFDP works; To discover the unidirectional link s1 $\rightarrow$ s2, the controller encapsulates a LLDP packet in a Packet-out message and sends it to s1. The packet-out contains instruction for s1 to send the LLDP packet to s2 via port p1. By receiving the LLDP packet via port p2, s2 encapsulates it in a Packet-in message and sends it back to the controller. The controller receives the LLDP packet and concludes that there is a unidirectional link from s1 to s2. The same process is performed to discover the opposite direction s2 $\rightarrow$ s1 as well as all other links in the network. Note that, OFDP packets are sent to a ”normal” multicast MAC (01:23:00:00:00:01) to avoid being swallowed by 802.1d compliant switches.

In dynamic networks like large data-centers and multi-tenant cloud networks, keeping an up-to-date visibility of the topology is a critical function; Switches leave and join the network dynamically creating changes in the topology which affects routing decisions that the controller has to make continuously. To remain up-to-date, the controller needs to repeat the process described in figure 1 periodically. The period separating two discovery rounds must be chosen carefully based on the dynamicity, size and capacity of the network; A 10 seconds period might not be suitable for a highly dynamic network as it may introduce a delay of up to 10 seconds. A short period (e.g. 3 seconds) also might not be suitable for a less-dynamic large size network as the large number of frequent discovery packets may exhaust controller’s resources. Put together, every discovery-round period $T$, the controller sends $\sum_{i=1}^{n}p_{i}$ (where $n$ is the number of switches and $p_{i}$ is the number of ports in switch $i$) Packet-out messages and receives $2L$ Packet-in messages. [4] proposes to reduce the number of Packet-out messages to $n$ by rewriting LLDP packet-headers in the switch.

A non optimized or buggy topology discovery mechanism can affect routing logic and drastically reduce network performance. Our main goal in this paper is to demonstrate that $OFDP$ has serious, non-solved yet, security and performance problems, then we briefly introduce $sOFTD$ (secure and efficient OpenFlow Topology Discovery), a secure alternative that is more efficient than $OFDP$.

The remainder of this paper is organized as follows: In section II we demonstrate why $OFDP$ shouldn’t be implemented in production networks. We introduce our alternative protocol $sOFTD$ in section III and we conclude the paper in section IV

II Why OFDP shouldn’t be implemented in production networks

II-A OFDP is not secure

As implemented by all controllers we have tested (OpenDaylight [5], Floodlight [6], NOX [7], POX [7], Beacon [8], Ryu [9] and Cisco Open SDN Controller [10]), OFDP uses clear, non authenticated LLDP packets to detect links between switches which makes it vulnerable to a number of attacks:

Switch spoofing. As described in figure 2, each LLDP packet contains a version field, flags, TTL and TLVs (Type-length-value) for information advertisement. Mandatory TLVs in OFDP are $ChssisSubtype$ and $PortSubtype$ to track packets. In figure 1, the LLDP packet sent by switch s2 to the controller contains the tuple ($chassisSubtype=switch1ID,PortSubtype=p1$), hence the controller will detect that this is the same packet he sent to switch s1 with p1 as out-port.

The problem is that all controllers we have tested set $chassisSubtype$ value to the MAC address of the local port of the switch (figure 4), which makes it easy for an adversary to spoof that switch since controllers use that MAC address as a unique identifier of the switch. By intercepting clear LLDP packets containing MAC addresses, a malicious switch can spoof other switches to falsify the controller’s topology graph. In the example shown in figure 3, s4 intercepts LLDP packets from s1 containing s1’s local port MAC address. Now, s4 can use it as its own MAC and reconnect to the controller as s1 messing up the controller’s topology graph (e.g. the controller adds nonexistent links $s1\rightarrow s3$ and $s3\rightarrow s1$). We have tested the switch spoofing attack successfully against Opendaylight and Floodlight.

Link Fabrication. [1] and [3] pointed out that OFDP is vulnerable to link fabrication attacks; In figure 5, the adversary has control over two end-hosts h1 and h2 connected to switches s1 and s3. h1 sends the LLDP packets received from s1 to h2 through an out-of-band connection (could be a tunnel over s2 for example), and h2 replicates them to s3. The controller receives the LLDP packets from s3 and creates a link between s1 and s3. While not detected, the fake link pushes the controller into wrong routing decisions that affects all communications involving s1 and s3. If the attacker has control only over h1, but knows the DPID of s3 then he still can fabricate a unidirectional link s3 $\rightarrow$ s1 by injecting fake LLDP packets into s1.

Another form of link fabrication is by LLDP injection; By monitoring the traffic, the adversary gets the LLDP content used by the controller. Then, he/she injects the same LLDP packets into the network creating bogus links between switches or between the adversary hosts and switches.

[1] proposes to authenticate the LLDP packets by adding a key-Hash Message Authentication Code (HMAC) as an optional TLV in LLDP packets. As mentioned by the authors, this technique only works against fake LLDP injection but not against link fabrication by packet duplication (Figure 5). [3] proposes a similar technique but using dynamic keys: a unique key for each LLDP packet assuming that h2, for example, uses only one example of LLDP packets received from h1 to generate future fake packets. However, an attacker controlling both hosts can permanently forward captured LLDP packets from h1 to h2 and from h2 to h1 and inject them back into switches without any modification.

Controller fingerprinting. As we explained in a previous work [11], the LLDP content is different from one controller to another which allows fingerprinting attacks on SDN controllers. An adversary (h1 in figure 5) matches the LLDP content he receives from s1 (LLDP packets originate from the controller) against a controller signature database to detect which controller is managing the network. Such information is very useful to launch specific and more efficient attacks on the controller. Figures 4 and 6 present the LLDP content of controllers POX and Floodlight respectively. Note also that the controllers use different default discovery-round periods which offers another way to differentiate between controllers. Although it is possible that network admins change both discovery-round period and LLDP content, it is more likely that the default values are kept unmodified.

LLDP Flood. This is a form of DoS attack where an adversary generates enough fake LLDP packets to exhaust the controller resources. In figure 7, host h1 generates large number of LLDP packets and send them to s1 which has a rule to forward every LLDP packet to the controller. Hence, a large number of LLDP packets can exhaust the link connecting the switch to the controller as well as the controller resources. Basic countermeasure methods like port blocking or packet filtering may not be effective, especially in the case of very dynamic environments (e.g. multi-tenant cloud) since connected hosts and switches change frequently, which may result in preventing legitimate LLDP packets from reaching the controller.

II-B OFDP is not efficient

By using OFDP, the controller periodically sends many packets to every switch in the network, which could result in performance decrease of the data plane. Experiments made on different controllers [12] show that when the network size (i.e. number of switches) exceeds some threshold, running the discovery module alone results in significant increase of the controller’s CPU usage and considerable decrease in network performance.

II-C Other issues

Other issues with OFDP include that it may not reliably work for heavily loaded links because discovery packets might get dropped or delayed. Moreover, when using OFDP in a multi-controller SDN network (e.g. running several guest controllers through FlowVisor), discovery cost increases linearly as more controllers are added.

III Introducing sOFTD: Secure OpenFlow Topology Discovery

The main idea behind sOFTD is to move a part of the discovery process from the controller to the switch by introducing minimal changes to the OpenFlow switch design. The main design characteristics of sOFTD are as follows:

• •

sOFTD adds a port-liveness-detection mechanism (BFD) [13] to the switch

• •

sOFTD uses OpenFlow FAST-FAILOVER groups (optional in OpenFlow 1.1+) to watch switch ports for connection updates

• •

sOFTD enables the switch to inform the controller about port connectivity updates

• •

The switch has a rule (”drop lldp”) to drop every LLDP packet to prevent LLDP flood attack

• •

The controller sends encrypted LLDP packets only when a switch-port connectivity update occurs and the LLDP packets are sent only to the concerned switches

• •

LLDP packets are preceded by rules (with hard timeout = 1s) to forward them back to the controller. These rules have a priority higher than ”drop lldp” rules

Since sOFTD do not send periodic discovery packets, preliminary resuts show significantly better performance than OFDP. We keep implementation details and results for a future work.

IV Conclusion

In this short paper, we demonstrated some serious security and efficiency problems in OpenFlow Discovery Protocol. Most of these problems are yet unsolved. As an alternative to OFDP, we briefly introduced our ongoing work, $sOFTD$ protocol, which consists of moving part of the discovery intelligence from the controller to the switch by introducing minimal changes to the OpenFlow switch design.