| Network Working Group | C. Huitema |
| Internet-Draft | Microsoft |
| Updates: 4361 (if approved) | T. Mrugalski |
| Intended status: Standards Track | ISC |
| Expires: October 9, 2015 | S. Krishnan |
| Ericsson | |
| April 7, 2015 |
Anonymity profile for DHCP clients
draft-huitema-dhc-anonymity-profile-02.txt
Some DHCP options carry unique identifiers. These identifiers can enable device tracking even if the device administrator takes care of randomizing other potential identifications like link-layer addresses or IPv6 addresses. The anonymity profile is designed for clients that wish to remain anonymous to the visited network. The profile provides guidelines on the composition of DHCP or DHCPv6 requests, designed to minimize disclosure of identifying information. This draft updates RFC4361.
This Internet-Draft is submitted in full conformance with the provisions of BCP 78 and BCP 79.
Internet-Drafts are working documents of the Internet Engineering Task Force (IETF). Note that other groups may also distribute working documents as Internet-Drafts. The list of current Internet-Drafts is at http://datatracker.ietf.org/drafts/current/.
Internet-Drafts are draft documents valid for a maximum of six months and may be updated, replaced, or obsoleted by other documents at any time. It is inappropriate to use Internet-Drafts as reference material or to cite them other than as "work in progress."
This Internet-Draft will expire on October 9, 2015.
Copyright (c) 2015 IETF Trust and the persons identified as the document authors. All rights reserved.
This document is subject to BCP 78 and the IETF Trust's Legal Provisions Relating to IETF Documents (http://trustee.ietf.org/license-info) in effect on the date of publication of this document. Please review these documents carefully, as they describe your rights and restrictions with respect to this document. Code Components extracted from this document must include Simplified BSD License text as described in Section 4.e of the Trust Legal Provisions and are provided without warranty as described in the Simplified BSD License.
Reports surfaced recently of systems that would monitor the wireless connections of passengers at Canadian airports [CNBC]. We can assume that these are either fragments or trial runs of a wider system that would attempt to monitor Internet users as they roam through wireless access points and other temporary network attachments. We can also assume that privacy conscious users will attempt to evade this monitoring, for example by ensuring that low level identifiers such as link-layer addresses are "randomized," so that the devices do not broadcast a unique identifier in every location that they visit.
Of course, link layer "MAC" addresses are not the only way to identify a device. As soon as it connects to a remote network, the device may use DHCP and DHCPv6 to obtain network parameters. The analysis of DHCP and DHCPv6 options shows that parameters of these protocols can reveal identifiers of the device, negating the benefits of link-layer address randomization. This is documented in detail in [I-D.ietf-dhc-dhcp-privacy] and [I-D.ietf-dhc-dhcpv6-privacy]. The natural reaction is to restrict the number and values of such parameters in order to minimize disclosure.
In the absence of a common standard, different system developers are likely to implement this minimization of disclosure in different ways. Monitoring entities could then use the differences to identify the software version running on the device. The proposed anonymity profile provides a common standard that minimizes information disclosure, including the disclosure of implementation identifiers.
The key words "MUST", "MUST NOT", "REQUIRED", "SHALL", "SHALL NOT", "SHOULD", "SHOULD NOT", "RECOMMENDED", "MAY", and "OPTIONAL" in this document are to be interpreted as described in [RFC2119].
Mobile nodes can be tracked using multiple identifiers, the most prominent being MAC addresses. For example, when devices use Wi-Fi connectivity, they place the MAC address in the header of all the packets that they transmit. Standard implementation of Wi-Fi use unique 48 bit MAC addresses, assigned to the devices according to procedures defined by IEEE 802. Even when the Wi-Fi packets are encrypted, the portion of the header containing the addresses will be sent in clear text. Tracking devices can "listen to the airwaves" to find out what devices are transmitting near them.
We can easily imagine that the MAC addresses can be correlated with other data, e.g., clear text names and cookies, to build a registry linking MAC addresses to the identity of devices' owners. Once that correlation is done, tracking the MAC address is sufficient to track individual people, even when all application data sent from the devices is encrypted. MAC addresses can also be correlated with IP addresses of devices, negating potential privacy benefits of IPv6 "privacy" addresses. Privacy advocates have some reason to be concerned.
The obvious solution is to "randomize" the MAC address. Before connecting to a particular network, the device replaces the MAC address with a randomly drawn 48 bit value. MAC address randomization was successfully tried at the IETF in Honolulu in November 2014 [IETFMACRandom]. However, we have to consider the linkage between MAC addresses, DHCP identifiers and IP addresses.
There is not yet an established standard for randomizing MAC addresses. Various prototypes have tried different strategies, such as:
In practice, there are many reasons to keep the MAC address constant for the duration of a link-layer connection, as in the "per connection" or "per network" variants. On Wi-Fi networks, changing the MAC address requires dropping the existing Wi-Fi connection and then re-establishing it, which implies repeating the connection process and associated procedures. The IP addresses will change, which means that all required TCP connections will have to be re-established. If the network access is provided through a NAT, changing IP address also means that the NAT traversal procedures will have to be restarted. This means a lot of disruption. At the same time, an observer on the network will easily notice that a station left, another came in just after that, and that the new one appears to be communicating with pretty much the same set of IP addresses as the old one. This provides for easy correlation.
The anonymity profile pretty much assumes that the MAC address randomization follows the "per connection" or "per network" strategies, or a variant of the "time interval" strategy in which the interval has about the same duration as the average connection.
From a privacy point of view, it is clear that MAC Addresses, IP addresses and DHCP identifiers shall evolve in synchrony. For example, if the MAC address changes and the DHCP identifier stays constant, then it is really easy to correlate old and new MAC addresses, either by listening to DHCP traffic or by observing that the IP address remains constant, since it is tied to the DHCP identifier. Conversely, if the DHCP identifier changes but the MAC address remains constant, the old and new identifiers and addresses can be correlated by listening to L2 traffic. The procedures documented in the following sections construct DHCP identifiers from the current MAC address, automatically providing for this synchronization.
The proposed anonymity profiles solve this synchronization issues by deriving most identifiers from the MAC address, and generally by making making sure that DHCP parameter values do not remain constant after an address change.
MAC address randomization solves the trivial monitoring problem in which someone just uses a Wi-Fi scanner and records the MAC addresses seen on the air. DHCP anonymity solves the more elaborated scenario in which someone monitor MAC addresses and identities used in DHCP at the access point or DHCP server. But this are not the only ways to track a mobile device.
Radio fingerprinting is a process that identifies a radio transmitter by the unique "fingerprint" of its signal transmission, i.e., the tiny differences caused by minute imperfections of the radio transmission hardware. This can be applied to diverse types of radios, including Wi-Fi as described for example in [WiFiRadioFingerprinting]. No amount of MAC address randomization will protect against such techniques. Protections may exist, but they are outside the scope of the present document.
On the other hand, we should not renounce randomization just because radio fingerprinting exists. The radio fingerprinting techniques are harder to deploy than just recording MAC addresses with a scanner. They can only track devices for which the fingerprint are known, and thus have a narrower scope of application than mass monitoring of addresses and DHCP parameters.
When a standard like DHCP allows for multiple options, different implementers will make different choices for the options that they support or the values they chose for the options. Conversely, monitoring the options and values present in DHCP messages reveals these differences and allows for "operating system fingerprinting," i.e., finding the type and version of software that a particular device is running. Finding these versions provides some information about the device identity, and thus goes against the goal of anonymity.
The design of the anonymity profiles attempts to minimize the number of options and the choice of values, in order to reduce the possibilities of operating system fingerprinting.
Reviewers of the anonymity profiles have sometimes suggested adding an option to explicitly identify the profiles as "using the anonymity option." One suggestion is that if the client wishes to remain anonymous, it would be good if the client told the server about that in case the server is willing to co-operate. Another possibility would be to use specific privacy-oriented construct, such as for example a new type of DUID of temporary DUID that would be changing over time.
This is not workable in a large number of cases as it is possible that the network operator (or other entities that have access to the operator's network) might be actively participating in surveillance and anti-privacy, willingly or not. Declaring a preference for anonymity is a bit like walking around with a Guy Fawkes mask. When anonymity is required, it is generally not a good idea to stick out of the crowd. Simply revealing the desire for privacy, could cause the attacker to react by triggering additional surveillance or monitoring mechanisms. Therefore we feel that it is preferable to not disclose one's desire for privacy.
This preference leads to some important implications. In particular, we make an effort to make the mitigation techniques difficult to distinguish from regular client behaviors, if at all possible.
There are downsides to randomizing MAC addresses and DHCP identifiers. By definition, randomization will break management procedures that rely on tracking MAC addresses. Even if this is not too much of a concern, we have to be worried about the frequency of MAC address randomization. Suppose for example that many devices would get new random MAC addresses at short intervals, maybe every few minutes. This would generate new DHCP requests in rapid succession, with a high risk of exhausting DHCPv4 address pools. Even with IPv6, there would still be a risk of increased neighbor discovery traffic, and bloating of various address tables. Implementers will have to be cautious when programming devices to use randomized MAC addresses. They will have to carefully chose the frequency with which such addresses will be renewed.
This document only provides guidelines for using DHCP when clients care about privacy and servers do not object. We assume that the request for anonymity is materialized by the assignment of a randomized MAC address to the network interface. Once that decision is made, the following guidelines will avoid leakage of identity in DHCP parameters or in assigned addresses.
There may be rare situations where the clients want anonymity to attackers but not to their DHCP server. These clients should still use MAC Address randomization to hide from observers, and some form of encrypted communication to the DHCP server. This scenario is not yet supported in this document.
This document only provides recommendations for DHCP clients. The main target are DHCP clients used in mobile devices. Such devices are a tempting target for various monitoring systems, and providing them with a simple anonymity solution is urgent. We can argue that some mobile devices embed DHCP servers, and that providing solutions for such devices is also quite important. Two plausible examples would be a DHCP server for a car network, or a DHCP server for a mobile hot spot. However, mobile servers get a lot of privacy protection through the use of access control and link layer encryption. Servers may disclose information to clients through DHCP, but they normally only do that to clients that have passed the link-layer access control and have been authorized to use the network services. This arguably makes solving the server problem less urgent than solving the client problem.
The server part will be covered by the general mitigation work going on in DHCP working group, following the analyses presented in [I-D.ietf-dhc-dhcp-privacy] and [I-D.ietf-dhc-dhcpv6-privacy].
Clients using the DHCPv4 anonymity profile limit the disclosure of information by controlling the header parameters and by limiting the number and values of options. The number of options depend on the specific DHCP message:
Header fields and option values SHOULD be set in accordance with the DHCP specification, but some header fields and option values SHOULD be constructed per the following guidelines.
Four bytes in the header of the DHCP messages carry the "Client IP address" (ciaddr) as defined in [RFC2131]. In DHCP, this field is used by the clients to indicate the address that they used previously, so that as much as possible the server can allocate them the same address.
There is very little privacy implication of sending this address in the DHCP messages, except in one case, when connecting to a different network than the last network connected. If the DHCP client somehow repeated the address used in a previous network attachment, monitoring services might use the information to tie the two network locations. DHCP clients should ensure that the field is cleared when they know that the network attachment has changed, and in particular of the link layer address is reset by the device's administrator.
The clients using the anonymity profile MUST NOT include in the message a Client IP Address that has been obtained with a different MAC address.
The Requested IP address option (code 50) allows the client to request that a particular IP address be assigned. The option is mandatory in some protocol messages per [RFC2131], for example when a client selects to use an address offered by a server. However, this option is not mandatory in the DHCPDISCOVER message. It is simply a convenience, an attempt to regain the same IP address that was used in a previous connection. Doing so entails the risk of disclosing an IP address used by the client at a previous location, or with a different MAC Address.
When using the anonymity profile, clients SHOULD NOT use the Requested IP address option in DHCPDISCOVER Messages. They MUST use the option when mandated by the DHCP protocol, for example in DHCPREQUEST Messages.
Sixteen bytes in the header of the DHCP messages carry the "Client hardware address" (chaddr) as defined in [RFC2131]. The presence of this address is necessary for the proper operation of the DHCP service.
Hardware addresses, called "link layer address" in many RFCs, can be used to uniquely identify a device, especially if they follow the IEEE 802 recommendations. These unique identifiers can be used by monitoring services to track the location of the device and its user. The only plausible defense is to somehow reset the hardware address to a random value when visiting an untrusted location, before transmitting anything at that location with the hardware address. If the hardware address is reset to a new value, or randomized, the DHCP client SHOULD use the new randomized value in the DHCP messages.
The client identifier option is defined in [RFC2132] with option code 61. It is discussed in details in [RFC4361]. The purpose of the client identifier option is to identify the client in a manner independent of the link layer address. This is particularly useful if the DHCP server is expected to assign the same address to the client after a network attachment is swapped and the link layer address changes. It is also useful when the same node issues requests through several interfaces, and expects the DHCP server to provide consistent configuration data over multiple interfaces.
The considerations for hardware independence and strong client identity have an adverse effect on the privacy of mobile clients, because the hardware-independent unique identifier obviously enables very efficient tracking of the client's movements.
The recommendations in [RFC4361] are very strong, stating for example that "DHCPv4 clients MUST NOT use client identifiers based solely on layer two addresses that are hard-wired to the layer two device (e.g., the Ethernet MAC address)." These strong recommendations are in fact a tradeoff between ease of management and privacy, and the tradeoff should depend on the circumstances.
In contradiction to [RFC4361], When using the anonymity profile, DHCP clients MUST use client identifiers based solely on the link layer address that will be used in the underlying connection. This will ensure that the DHCP client identifier does not leak any information that is not already available to entities monitoring the network connection. It will also ensure that a strategy of randomizing the link layer address will not be nullified by DHCP options.
The Host Name option is defined in [RFC2132] with option code 12. Depending on implementations, the option value can carry either a fully qualified domain name such as "node1984.example.com," or a simple host name such as "node1984." The host name is commonly used by the DHCP server to identify the host, and also to automatically update the address of the host in local name services.
Fully qualified domain names are obviously unique identifiers, but even simple host names can provide a significant amount of information on the identity of the device. They are typically chosen to be unique in the context where the device is most often used. If that context is wide enough, in a large company or in a big university, the host name will be a pretty good identifier of the device. Monitoring services could use that information in conjunction with traffic analysis and quickly derive the identity of the device's owner.
When using the anonymity profile, DHCP clients MAY avoid sending the host name option. If they chose to send the option, DHCP clients MUST always send a non-qualified host name instead of a fully qualified domain name, and MUST obfuscate the host name value, so it could not be linked to anything other than the link layer address. When obfuscating the host name, DHCP clients SHOULD set the host name value to a hexadecimal representation of the link layer address that will be used in the underlying connection. They MAY choose another convention in rare cases, for example in multi-homed scenarios.
The Client FQDN option is defined in [RFC4702] with option code 81. The option allows the DHCP clients to advertise to the DHCP server their fully qualified domain name (FQDN) such as "mobile.example.com." This would allow the DHCP server to update in the DNS the PTR record for the IP address allocated to the client. Depending on circumstances, either the DHCP client or the DHCP server could update in the DNS the A record for the FQDN of the client.
Obviously, this option uniquely identifies the client, exposing it to the DHCP server or to anyone listening to DHCP traffic. In fact, if the DNS record is updated, the location of the client becomes visible to anyone with DNS lookup capabilities.
When using the anonymity profile, DHCP clients SHOULD NOT include the Client FQDN option in their DHCP requests. Alternatively, they MAY include a special purpose FQDN using the same hostname as in the Host Name Option, with a suffix matching the connection-specific DNS suffix being advertised by that DHCP server. Having a name in the DNS allows working with legacy systems that require one to be there, e.g., by verifying a forward and reverse lookup succeeds with the same result.
The UUID/GUID-based Client Machine Identifier option is defined in [RFC4578], with option code 97. The option is part of a set of options for Intel Preboot eXecution Environment (PXE). The purpose of the PXE system is to perform management functions on a device before its main OS is operational. The Client Machine Identifier carries a 16-octet Globally Unique Identifier (GUID), which uniquely identifies the device.
The PXE system is clearly designed for devices operating in a controlled environment, and its functions are not meant to be used by mobile nodes visiting untrusted networks. If only for privacy reasons, nodes visiting untrusted networks MUST disable the PXE functions, and MUST NOT send the corresponding options.
Vendor identifying options are defined in [RFC2132] and [RFC3925]. When using the anonymity profile, DHCP clients SHOULD NOT use the Vendor Specific Information option (code 43), the Vendor Class Identifier Option (60), the Vendor Class option (code 124), or the Vendor Specific Information option (code 125) as these options potentially reveal identifying information.
DHCPv6 is typically used by clients in one of two scenarios: stateful and stateless configuration. In the stateful scenario, clients use a combination of SOLICIT, REQUEST, CONFIRM, RENEW, REBIND and RELEASE messages to obtain addresses, and manage these addresses.
In the stateless scenario, clients configure addresses using a combination of client managed identifiers and router-advertised prefixes, without involving the DHCPv6 services. Different ways of constructing these prefixes have different implications on privacy, which are discussed in [I-D.ietf-6man-default-iids] and [I-D.ietf-6man-ipv6-address-generation-privacy]. In the stateless scenario, clients use DHCPv6 to obtain network configuration parameters, through the INFORMATION-REQUEST message.
The choice between the stateful and stateless scenario depends of flag and prefix options published by the "Router Advertisement" messages of local routers, as specified in [RFC4861]. When these options enable stateless address configuration hosts using the anonymity profile SHOULD choose it over stateful address configuration, because stateless configuration requires fewer information disclosure than stateful configuration.
When using the anonymity profile, DHCPv6 clients carefully select DHCPv6 options used in the various messages that they sent. The list of options that are mandatory or optional for each message is specified in [RFC3315]. Some of these options have specific implications on anonymity. The following sections provide guidance on the choice of option values when using the anonymity profile.
The [RFC3315] requires clients to send a Confirm message when they attach to a new link to verify whether the addressing and configuration information they previously received is still valid. This requirement was relaxed in [I-D.ietf-dhc-rfc3315bis]. When these clients send Confirm messages, they include any IAs assigned to the interface that may have moved to a new link, along with the addresses associated with those IAs. By examining the addresses in the Confirm message an attacker can trivially identify the previous point(s) of attachment.
Clients interested in protecting their privacy SHOULD NOT send Confirm messages and instead directly try to acquire addresses on the new link.
The client identifier option is defined in [RFC3315] with option code 1. The purpose of the client identifier option is to identify the client to the server. The content of the option is a DHCP User ID (DUID). One of the primary privacy concerns is that a client is disclosing a stable identifier (the DUID) that can be use for tracking and profiling. Three DUID formats are specified: Link-layer address plus time, Vendor-assigned unique ID based on Enterprise Number, Link-layer address.
When using the anonymity profile in conjunction with randomized MAC addresses, DHCPv6 clients MUST use the DUID format number 3, Link-layer address. The value of the Link-layer address should be that currently assigned to the interface.
When using the anonymity profile without the benefit of randomized MAC addresses, clients that want to protect their privacy SHOULD generate a new randomized DUID-LLT every time they attach to a new link or detect a possible link change event. The exact details are left up to implementors, but there are several factors should be taken into consideration. The DUID type SHOULD be set to 1 (DUID-LLT). Hardware type SHOULD be set appropriately to the hardware type. Time MAY be set to current time, but this will reveal the fact that the DUID is newly generated. Implementors interested in hiding this fact MAY use a time stamp from the past. e.g. a random timestamp from the previous year could be a good value. In the most common cases the link-layer address is based on MAC. The first three octets are composed of the OUI (Organizationally Unique Identifier) that is expected to have a value assigned to a real organization. See [IEEE-OUI] for currently assigned values. Using a value that is unassigned may disclose the fact that a DUID is randomized. Using a value that belongs to a third party may have legal implications.
According to [RFC3315], a DHCPv6 client typically includes its client identifier in most of the messages it sends. There is one exception, however. Client is allowed to omit its client identifier when sending Information-Request.
When using stateless DHCPv6, clients wanting to protect their privacy SHOULD NOT include client identifiers in their Information-Request messages. This will prevent the server from specifying client-specific options if it is configured to do so, but the need for anonymity precludes such options anyway.
When using the anonymity profile, DHCPv6 clients SHOULD use the Server Identifier Option (code 2) as specified in [RFC3315]. Clients MUST only include server identifier values that were received with the current MAC address, because reuse of old values discloses information that can be used to identify the client.
When using the anonymity profile, DHCPv6 clients might have to use SOLICIT or REQUEST messages to obtain IPv6 addresses through the DHCP server. The clients SHOULD only use the options necessary to perform the requested DHCPv6 transactions, such as Identity Association for Non-temporary Addresses Option (code 3) or Identity Association for Temporary Addresses Option (code 4).
The clients MAY use the IA Address Option (code 5) but need to balance the potential advantage of "address continuity" versus the potential risk of "previous address disclosure." A potential solution is to remove all stored addresses when a MAC address changes, and to only use the IA Address option with addresses that have been explicitly assigned through the current MAC address.
The interaction between prefix delegation and anonymity require further study. For now, the simple solution is to avoid using prefix delegation when striving for anonymity. When using the anonymity profiles, clients SHOULD NOT use IA_PD, the prefix delegation form of address assignment.
[RFC3315] defines a special container (IA_TA) for requesting temporary addresses. This is a good mechanism in principle, but there are a number of issues associated with it. First, this is not widely used feature, so clients depending solely on temporary addresses may lock themselves out of service. Secondly, [RFC3315] does not specify any renewal mechanisms for temporary addresses. Therefore support for renewing temporary addresses may vary between server implementations, including not being supported at all. Finally, by requesting temporary addresses a client reveals its desire for privacy and potentially risks countermeasures as described in Section 2.5.
Clients interested in their privacy SHOULD NOT use IA_TA. They should simply send an IA_NA with a randomized IAID. This, along with the mitigation technique discussed in Section 4.3, will ensure that a client will get a new address that can be renewed and can be used as long as needed. To get a new address, it can send Request message with a new randomized IAID before releasing the other one. This will cause the server to assign a new address, as it still has a valid lease for the old IAID value. Once a new address is assigned, the address obtained using the older IAID value can be released safely, using the Release message or it may simply be allowed to time out.
This solution may not work if the server enforces specific policies, e.g. only one address per client. If client does not succeed in receiving a second address using a new IAID, it may release the first one (using an old IAID) and then retry asking for a new address.
From the Operating System perspective, addresses obtained using this technique SHOULD be treated as temporary as specified in [RFC4941].
A DHCPv6 client may reveal other types of information, besides unique identifiers. There are many ways a DHCPv6 client can perform certain actions and the specifics can be used to fingerprint the client. This may not reveal the identity of a client, but may provide additional information, such as the device type, vendor type or OS type and in some cases specific version.
One specific method used for fingerprinting utilizes the order in which options are included in the message. Another related technique utilizes the order in which option codes are included in an Option Request Option (ORO).
The client willing to protect its privacy SHOULD randomize options order before sending any DHCPv6 message. Such a client SHOULD also randomly shuffle the option codes order in ORO.
According to [RFC3315], the client that includes an Option Request Option in a Solicit or Request message MAY additionally include instances of those options that are identified in the Option Request option, with data values as hints to the server about parameter values the client would like to have returned.
When using the anonymity profile, clients SHOULD NOT include such instances of options because old values might be used to identify the client.
The purpose of the Authentication option (code 11) is to authenticate the identity of clients and servers and the contents of DHCP messages. As such, the option can be used to identify the client, and is incompatible with the stated goal of "client anonymity." DHCPv6 clients that use the anonymity profile SHOULD NOT use the authentication option. They MAY use it if they recognize that they are operating in a trusted environment, e.g., in a work place network.
When using the anonymity profile, DHCPv6 clients SHOULD NOT use the User Class option (code 15) or the Vendor Class option (code 16), as these options potentially reveal identifying information.
The Client FQDN option is defined in [RFC4704] with option code 29. The option allows the DHCP clients to advertize to the DHCP their fully qualified domain name (FQDN) such as "mobile.example.com." When using the anonymity profile, DHCPv6 clients SHOULD NOT include the Client FQDN option in their DHCPv6 messages because it identifies the client. As explained in Section 3.6 they MAY use a local-only FQDN by combining a host name derived from the link layer address and a suffix advertised by the local DHCP server.
The anonymity profile has the effect of hiding the client identity from the DHCP server. This is not always desirable. Some DHCP servers provide facilities like publishing names and addresses in the DNS, or ensuring that returning clients get reassigned the same address. Implementers should be careful to only use the anonymity profile when privacy trumps management considerations.
Clients using the anonymity profile in general consume more resources. For example when they change MAC address and request for a new IP, the old one is still marked as leased by the server.
The use of the anonymity profile does not change the security considerations of the DHCPv4 or DHCPv6 protocols.
This draft does not require any IANA action.
The inspiration for this draft came from discussions in the Perpass mailing list. Several people provided feedback on this draft, notably Noel Anderson, Lorenzo Colitti, Stephen Farrell, Tushar Gupta, Gabriel Montenegro, Marcin Siodelski, Dave Thaler and Jun Wu.