Guidelines for Creating New DHCPv6 Options

Most protocol developers ask themselves if a protocol will work, or work efficiently. These are important questions, but another less frequently considered question is whether the proposed protocol presents itself needless barriers to adoption by deployed software. DHCPv6 software implementors are not merely faced with the task of handling a given option's format on the wire. The option must fit into every stage of the system's process, starting with the user interface used to enter the configuration up to the machine interfaces where configuration is ultimately consumed. Another frequently overlooked aspect of rapid adoption is whether the option requires operators to be intimately familiar with the option's internal format in order to use it. Most DHCPv6 software provides a facility for handling unknown options at the time of publication. The handling of such options usually needs to be manually configured by the operator. But, if doing so requires extensive reading (more than can be covered in a simple FAQ, for example), it inhibits adoption. So, although a given solution would work, and might even be space, time, or aesthetically optimal, a given option is presented with a series of ever-worsening challenges to be adopted: If it doesn't fit neatly into existing configuration files. If it requires source code changes to be adopted and, hence, upgrades of deployed software. If it does not share its deployment fate in a general manner with other options, standing alone in requiring code changes or reworking configuration file syntaxes. If the option would work well in the particular deployment environment the proponents currently envision, but it has equally valid uses in some other environment where the proposed option format would fail or would produce inconsistent results. There are many things DHCPv6 option creators can do to avoid the pitfalls in this list entirely, or failing that, to make software implementors' lives easier and improve its chances for widespread adoption. This document is envisaged as a help for protocol developers that define new options and for expert reviewers that review submitted proposals.

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 .

Principally, DHCPv6 carries configuration parameters for its clients. Any knob, dial, slider, or checkbox on the client system, such as "my domain name servers", "my hostname", or even "my shutdown temperature", are candidates for being configured by DHCPv6. The presence of such a knob isn't enough, because DHCPv6 also presents the extension of an administrative domain -- the operator of the network to which the client is currently attached. Someone runs not only the local switching network infrastructure to which the client is directly (or wirelessly) attached but the various methods of accessing the external Internet via local assist services that the network must also provide (such as domain name servers or routers). This means that, even if a configuration parameter can be potentially delivered by DHCPv6, it is necessary to evaluate whether it is reasonable for this parameter to be under the control of the administrator of whatever network a client is attached to at any given time. Note that the client is not required to configure any of these values received via DHCPv6 (e.g., due to having these values locally configured by its own administrator). But, it needs to be noted that overriding DHCPv6-provided values may cause the client to be denied certain services in the network to which it has attached. The possibility of having a higher level of control over client node configuration is one of the reasons that DHCPv6 is preferred in enterprise networks.

The primary guiding principle to follow in order to enhance an option's adoptability is reuse. The option should be created in such a way that does not require any new or special case software to support. If old software that is currently deployed and in the field can adopt the option through supplied configuration facilities, then it's fairly certain that new software can formally adopt it easily. There are at least two classes of DHCPv6 options: simple options, which are provided explicitly to carry data from one side of the DHCPv6 exchange to the other (such as name servers, domain names, or time servers), and a protocol class of options, which require special processing on the part of the DHCPv6 software or are used during special processing (such as the Fully Qualified Domain Name (FQDN) option ), and so forth; these options carry data that is the result of a routine in some DHCPv6 software. The guidelines laid out here should be applied in a relaxed manner for the protocol class of options. Wherever a special case code is already required to adopt the DHCPv6 option, it is substantially more reasonable to format the option in a less generic fashion, if there are measurable benefits to doing so.

The easiest approach to manufacturing trivially deployable DHCPv6 options is to assemble the option out of whatever common fragments fit, possibly allowing a group of data elements to repeat to fill the remaining space (if present) and thus provide multiple values. Place all fixed-size values at the start of the option and any variable&nbhy;/indeterminate-sized values at the tail end of the option. This means that implementations will likely be able to reuse code paths designed to support the other options. There is a trade-off between the adoptability of previously defined option formats and the advantages that new or specialized formats can provide. In general, it is usually preferable to reuse previously used option formats. However, it isn't very practical to consider the bulk of DHCPv6 options already allocated and to consider which of those solve a similar problem. So, the following list of common option format data elements is provided as shorthand. Please note that it is not complete in terms of exampling every option format ever devised. If more complex options are needed, those basic formats mentioned here may be considered as primitives (or 'fragment types') that can be used to build more complex formats. It should be noted that it is often easier to implement two options with trivial formats than one option with a more complex format. That is not an unconditional requirement though. In some cases, splitting one complex option into two or more simple options introduces inter-option dependencies that should be avoided. In such a case, it is usually better to keep one complex option.

This option format is used to carry one or many IPv6 addresses. In some cases, the number of allowed addresses is limited (e.g., to one):

Examples of use: DHCPv6 Server Unicast Address (a single address only) Session Initiation Protocol (SIP) Servers IPv6 Address List DNS Recursive Name Servers Network Information Service (NIS) Servers Simple Network Time Protocol (SNTP) Servers Broadcast and Multicast Service Controller IPv6 Address Option for DHCPv6 Mobile IPv6 (MIPv6) Home Agent Address (a single address only) Network Time Protocol (NTP) Server Address (a single address only) NTP Multicast Address (a single address only)

Sometimes, it is useful to convey a single flag that can take either on or off values. Instead of specifying an option with 1 bit of usable data and 7 bits of padding, it is better to define an option without any content. It is the presence or absence of the option that conveys the value. This approach has the additional benefit of the absent option designating the default; that is, the administrator has to take explicit actions to deploy the opposite of the default value. The absence of the option represents the default value, and the presence of the option represents the other value, but that does not necessarily mean that absence is "off" (or "false") and presence is "on" (or "true"). That is, if it's desired that the default value for a bistable option is "true"/"on", then the presence of that option would turn it off (make it false). If the option presence signifies an off/false state, that should be reflected in the option name, e.g., OPTION_DISABLE_FOO.

Examples of use: DHCPv6 Rapid Commit

Sometimes, there is a need to convey an IPv6 prefix. The information to be carried by such an option includes the 128-bit IPv6 prefix together with a length of this prefix taking values from 0 to 128. Using the simplest approach, the option could convey this data in two fixed-length fields: one carrying the prefix length and another carrying the prefix. However, in many cases, /64 or shorter prefixes are used. This implies that the large part of the prefix data carried by the option would have its bits set to 0 and would be unused. In order to avoid carrying unused data, it is recommended to store the prefix in the variable-length data field. The appropriate option format is defined as follows:

option-length is set to 1 + length of the IPv6 prefix. prefix6len is 1 octet long and specifies the length in bits of the IPv6 prefix. Typically allowed values are 0 to 128. The ipv6-prefix field is a variable-length field that specifies the IPv6 prefix. The length is (prefix6len + 7) / 8. This field is padded with 0 bits up to the nearest octet boundary when prefix6len is not divisible by 8. Examples of use: Default Mapping Rule For example, the prefix 2001:db8::/60 would be encoded with an option-length of 9, prefix6-len would be set to 60, and the ipv6-prefix would be 8 octets and would contain octets 20 01 0d b8 00 00 00 00. It should be noted that the IAPREFIX option defined by uses a full-length 16-octet prefix field. The concern about option length was not well understood at the time of its publication.

This option format can be used to carry a 32-bit signed or unsigned integer value:

Examples of use: Information Refresh Time

This option format can be used to carry 16-bit signed or unsigned integer values:

Examples of use: Elapsed Time

This option format can be used to carry 8-bit integer values:

Examples of use: DHCPv6 Preference

A Uniform Resource Identifier (URI) is a compact sequence of characters that identifies an abstract or physical resource. The term "Uniform Resource Locator" (URL) refers to the subset of URIs that, in addition to identifying a resource, provide a means of locating the resource by describing its primary access mechanism (e.g., its network "location"). This option format can be used to carry a single URI:

Examples of use: Boot File URL An alternate encoding to support multiple URIs is available. An option must be defined to use either the single URI format above or the multiple URI format below depending on whether a single URI is always sufficient or if multiple URIs are possible.

Each instance of the uri-data is formatted as follows:

The uri-len is 2 octets long and specifies the length of the URI data. Although the URI format in theory supports up to 64 KB of data, in practice, large chunks of data may be problematic. See for details.

A text string is a sequence of characters that have no semantics. The encoding of the text string MUST be specified. Unless otherwise specified, all text strings in newly defined options are expected to be Unicode strings that are encoded using UTF-8 in Net-Unicode form . Please note that all strings containing only 7-bit ASCII characters are also valid UTF-8 Net-Unicode strings. If a data format has semantics other than just being text, it is not a string; e.g., an FQDN is not a string, and a URI is also not a string because they have different semantics. A string must not include any terminator (such as a null byte). The null byte is treated as any other character and does not have any special meaning. This option format can be used to carry a text string:

Examples of use: Timezone Options for DHCPv6 An alternate encoding to support multiple text strings is available. An option must be defined to use either the single text string format above or the multiple text string format below, depending on whether a single text string is always sufficient or if multiple text strings are possible.

Each instance of the text-data is formatted as follows:

The text-len is 2 octets long and specifies the length of the string.

This option can be used to carry variable-length data of any kind. Internal representation of carried data is option specific. Whenever this format is used by the new option being defined, the data encoding should be documented. This option format provides a lot of flexibility to pass data of almost any kind. Though, whenever possible, it is highly recommended to use more specialized options, with field types better matching carried data types.

Examples of use: Client Identifier Server Identifier

This option is used to carry 'domain search' lists or any host or domain name. It uses the same format as described in but with the special data encoding, as described in Section 8 of . This data encoding supports carrying multiple instances of hosts or domain names in a single option by terminating each instance with the byte value of 0.

Examples of use: SIP Servers Domain Name List (many domains) NIS Domain Name (many domains) Location-to-Service Translation (LoST) Server Domain Name Location Information Server (LIS) Domain Name Dual-Stack Lite (DS-Lite) Address Family Transition Router (AFTR) Location (a single FQDN) Home Network Identifier (a single FQDN) Home Agent FQDN (a single FQDN)

Placing an octet at the start of the option that informs the software how to process the remaining octets of the option may appear simple to the casual observer. But, the only conditional formatting methods that are in widespread use today are 'protocol' class options. Therefore, conditional formatting requires new code to be written and complicates future interoperability should new conditional formats be added; existing code has to ignore conditional formats that it does not support.

Options are said to be aliases of each other if they provide input to the same configuration parameter. A commonly proposed example is to configure the location of some new service ("my foo server") using a binary IP address, a domain name field, and a URL. This kind of aliasing is undesirable and is not recommended. In this case, where three different formats are supposed, it more than triples the work of the software involved, requiring support for not merely one format but support to produce and digest all three. Furthermore, code development and testing must cover all possible combinations of defined formats. Since clients cannot predict what values the server will provide, they must request all formats. So, in the case where the server is configured with all formats, DHCPv6 message bandwidth is wasted on option contents that are redundant. Also, the DHCPv6 option number space is wasted, as three new option codes are required rather than one. It also becomes unclear which types of values are mandatory and how configuring some of the options may influence the others. For example, if an operator configures the URL only, should the server synthesize a domain name and an IP address? A single configuration value on a host is probably presented to the operator (or other software on the machine) in a single field or channel. If that channel has a natural format, then any alternative formats merely make more work for intervening software in providing conversions. So, the best advice is to choose the one method that best fulfills the requirements for simplicity (such as with an IP address and a port pair), late binding (such as with DNS), or completeness (such as with a URL).

Some parameters may be specified as an FQDN or an address. In most cases, one or the other should be used. This section discusses pros and cons of each approach and is intended to help make an informed decision in that regard. It is strongly discouraged to define both option types at the same time (see ), unless there is sufficient motivation to do so. There is no single recommendation that works for every case. It very much depends on the nature of the parameter being configured. For parameters that are network specific or represent certain aspects of network infrastructure, like available mobility services, in most cases addresses are a more usable choice. For parameters that can be considered an application-specific configuration, like SIP servers, it is usually better to use an FQDN. Applications are often better suited to deal with FQDN failures than with address failures. Most operating systems provide a way to retry an FQDN resolution if the previous attempt fails. That type of error recovery is supported by a great number of applications. On the other hand, there is typically no API available for applications to reconfigure over DHCP to get a new address value if the one received is no longer appropriate. This problem may be usually addressed by providing a list of addresses rather than just a single one. That, on the other hand, requires a defined procedure on how multiple addresses should be used (all at once, round robin, try first and fail over to the next if it fails, etc.). An FQDN provides a higher level of indirection and ambiguity. In many cases, that may be considered a benefit, but it can be considered a flaw in others. For example, one operator suggested that the same name be resolved to different addresses, depending on the point of attachment of the host doing the resolution. This is one way to provide localized addressing. However, in order to do this, it is necessary to violate the DNS convention that a query on a particular name should always return the same answer (aside from the ordering of IP addresses in the response, which is supposed to be varied by the name server). This same locality of reference for configuration information can be achieved directly using DHCP, since the DHCP server must know the network topology in order to provide IP address or prefix configuration. The other type of ambiguity is related to multiple provisioning domains (see ). The stub resolver on the DHCP client cannot at present be assumed to make the DNS query for a DHCP-supplied FQDN on the same interface on which it received its DHCP configuration and may, therefore, get a different answer from the DNS than was intended. This is particularly a problem when the normal expected use of the option makes sense with a private DNS zone(s), as might be the case on an enterprise network. It may also be the case that the client has an explicit DNS server configured and may, therefore, never query the enterprise network's internal DNS server. An FQDN does require a resolution into an actual address. This implies the question as to when the FQDN resolution should be conducted. There are a couple of possible answers: a) by the server, when it is started, b) by the server, when it is about to send an option, c) by the client, immediately after receiving an option, and d) by the client, when the content of the option is actually consumed. For a), b), and possibly c), the option should really convey an address, not an FQDN. The only real incentive to use an FQDN is case d). It is the only case that allows possible changes in the DNS to be picked up by clients. If the parameter is expected to be used by constrained devices (low power, battery operated, and low capabilities) or in very lossy networks, it may be appealing to drop the requirement of performing the DNS resolution and use addresses. Another example of a constrained device is a network-booted device, where despite the fact that the node itself is very capable once it's booted, the boot prom is quite constrained. Another aspect that should be considered is time required for the clients to notice any configuration changes. Consider a case where a server configures service A using an address and service B using an FQDN. When an administrator decides to update the configuration, he or she can update the DHCP server configuration to change both services. If the clients do not support reconfigure (which is an optional feature of RFC 3315 but in some environments, e.g., cable modems, is mandatory), the configuration will be updated on the clients after the T1 timer elapses. Depending on the nature of the change (is it a new server added to a cluster of already operating servers or a new server that replaces the only available server that crashed?), this may be an issue. On the other hand, updating service B may be achieved with a DNS record update. That information may be cached by caching DNS servers for up to Time to Live (TTL). Depending on the values of T1 and TTL, one update may be faster than another. Furthermore, depending on the nature of the change (planned modification or unexpected failure), T1 or TTL may be lowered before the change to speed up new configuration adoption. Simply speaking, protocol designers don't know what the TTL or the T1 time will be, so they can't make assumptions about whether a DHCP option will be refreshed more quickly based on T1 or TTL. Addresses have the benefit of being easier to implement and handle by the DHCP software. An address option is simpler to use, has validation that is trivial (multiple of 16 constitutes a valid option), is explicit, and does not allow any ambiguity. It is faster (does not require extra round-trip time), so it is more efficient, which can be especially important for energy-restricted devices. It also does not require that the client implements a DNS resolution. An FQDN imposes a number of additional failure modes and issues that should be dealt with: The client must have knowledge about available DNS servers. That typically means that option DNS_SERVERS is mandatory. This should be mentioned in the document that defines the new option. It is possible that the server will return the FQDN option but not the DNS server's option. There should be a brief discussion about it; The DNS may not be reachable; The DNS may be available but may not have appropriate information (e.g., no AAAA records for the specified FQDN); The address family must be specified (A, AAAA, or any); the information being configured may require a specific address family (e.g., IPv6), but there may be a DNS record only of another type (e.g., A only with an IPv4 address). What should the client do if there are multiple records available (use only the first one, use all, use one and switch to the second if the first fails for whatever reason, etc.). This may be an issue if there is an expectation that the parameter being configured will need exactly one address; Multihomed devices may be connected to different administrative domains with each domain providing different information in the DNS (e.g., an enterprise network exposing private domains). The client may send DNS queries to a different DNS server; and It should be mentioned if Internationalized Domain Names are allowed. If they are, DNS option encoding should be specified. Address options that are used with overly long T1 (renew timer) values have some characteristics of hard-coded values. That is strongly discouraged. See for an in-depth discussion. If the option may appear in Information-request, its lifetime should be controlled using the information refresh time option . One specific case that makes the choice between an address and an FQDN not obvious is a DNS Security (DNSSEC) bootstrap scenario. DNSSEC validation imposes a requirement for clock sync (to the accuracy reasonably required to consider signature inception and expiry times). This often implies usage of NTP configuration. However, if NTP is provided as an FQDN, there is no way to validate its DNSSEC signature. This is a somewhat weak argument though, as providing an NTP server as an address is also not verifiable using DNSSEC. If the trustworthiness of the configuration provided by the DHCP server is in question, DHCPv6 offers mechanisms that allow server authentication.

Most options are conveyed in a DHCPv6 message directly. Although there is no codified normative language for such options, they are often referred to as top-level options. Many options may include other options. Such inner options are often referred to as encapsulated or nested options. Those options are sometimes called sub-options, but this term actually means something else and, therefore, should never be used to describe encapsulated options. It is recommended to use the term "encapsulated" as this terminology is used in . The difference between encapsulated and sub-options is that the former uses normal DHCPv6 option numbers, while the latter uses option number space specific to a given parent option. It should be noted that, contrary to DHCPv4, there is no shortage of option numbers; therefore, almost all options share a common option space. For example, option type 1 meant different things in DHCPv4, depending if it was located in the top level or inside of the Relay Agent Information option. There is no such ambiguity in DHCPv6 (with the exception of , which SHOULD NOT be used as a template for future DHCP option definitions). From the implementation perspective, it is easier to implement encapsulated options rather than sub-options, as the implementors do not have to deal with separate option spaces and can use the same buffer parser in several places throughout the code. Such encapsulation is not limited to one level. There is at least one defined option that is encapsulated twice: Identity Association for Prefix Delegation (IA_PD), as defined in Section 9 of , conveys the Identity Association (IA) Prefix (IAPREFIX), as defined in Section 10 of . Such a delegated prefix may contain an excluded prefix range that is represented by the PD_EXCLUDE option that is conveyed as encapsulated inside IAPREFIX (PD_EXCLUDE is defined in ). It seems awkward to refer to such options as sub-sub-option or doubly encapsulated option; therefore, the "encapsulated option" term is typically used, regardless of the nesting level. When defining a DHCP-based configuration mechanism for a protocol that requires something more complex than a single option, it may be tempting to group configuration values using sub-options. That should preferably be avoided, as it increases complexity of the parser. It is much easier, faster, and less error prone to parse a large number of options on a single (top-level) scope than to parse options on several scopes. The use of sub-options should be avoided as much as possible, but it is better to use sub-options rather than conditional formatting. It should be noted that currently there is no clear way defined for requesting sub-options. Most known implementations are simply using the top-level Option Request Option (ORO) for requesting both top-level and encapsulated options.

DHCP is designed for provisioning clients. Less experienced protocol designers often assume that it is easy to define an option that will convey a different parameter for each client in a network. Such problems arose during designs of the Mapping of Address and Port (MAP) and IPv4 Residual Deployment (4rd) . While it would be easier for provisioned clients to get ready to use per-client option values, such a requirement puts exceedingly large loads on the server side. The new extensions may introduce new implementation complexity and additional database state on the server. Alternatives should be considered, if possible. As an example, was designed in a way that all clients are provisioned with the same set of MAP options, and each provisioned client uses its unique address and delegated prefix to generate client-specific information. Such a solution does not introduce any additional state for the server and, therefore, scales better. It also should be noted that contrary to DHCPv4, DHCPv6 keeps several timers for renewals. Each IA_NA (addresses) and IA_PD (prefixes) contains T1 and T2 timers that designate time after which the client will initiate renewal. Those timers apply only to their associated IA containers. Refreshing other parameters should be initiated after a time specified in the information refresh time option (defined in ), carried in the Reply message, and returned in response to the Information-request message. Introducing additional timers make deployment unnecessarily complex and SHOULD be avoided.

In general, DHCPv6 clients only refresh configuration data from the DHCP server when the T1 timer expires. Although there is a Reconfigure mechanism that allows a DHCP server to request that clients initiate reconfiguration, support for this mechanism is optional and cannot be relied upon. Even when DHCP clients refresh their configuration information, not all consumers of DHCP-sourced configuration data notice these changes. For instance, if a server is started using parameters received in an early DHCP transaction, but does not check for updates from DHCP, it may well continue to use the same parameter indefinitely. There are a few operating systems that take care of reconfiguring services when the client moves to a new network (e.g., based on mechanisms like , , or ), but it's worth bearing in mind that a renew may not always result in the client taking up new configuration information that it receives. In light of the above, when designing an option you should take into consideration the fact that your option may hold stale data that will only be updated at an arbitrary time in the future.

In some cases, there could be more than one DHCPv6 server on a link, with each providing a different set of parameters. One notable example of such a case is a home network with a connection to two independent ISPs. The DHCPv6 specification does not provide clear advice on how to handle multiple provisioning sources. Although states that a client that receives more than one Advertise message may respond to one or more of them, such capability has not been observed in existing implementations. Existing clients will pick one server and will continue the configuration process with that server, ignoring all other servers. In addition, a node that acts as a DHCPv6 client may be connected to more than one physical network. In most cases, it will operate a separate DHCP client state machine on each interface and acquire different, possibly conflicting, information through each. This information will not be acquired in any synchronized way. Existing nodes cannot be assumed to systematically segregate configuration information on the basis of its source; as a result, it is quite possible that a node may receive an FQDN on one network interface but do the DNS resolution on a different network interface, using different DNS servers. As a consequence, DNS resolution done by the DHCP server is more likely to behave predictably than DNS resolution done on a multi-interface or multihomed client. This is a generic DHCP issue and should not be dealt within each option separately. This issue is better dealt with using a protocol-level solution, and fixing this problem should not be attempted on a per-option basis. Work is ongoing in the IETF to provide a systematic solution to this problem.

Adding a simple DHCP option is straightforward and generally something that any working group (WG) can do, perhaps with some help from designated DHCP experts. However, when new fragment types need to be devised, this requires the attention of DHCP experts and should not be done in a WG that doesn't have a quorum of such experts. This is true whether the new fragment type has the same structure as an existing fragment type but with different semantics, or the new format has a new structure. Responsible Area Directors for WGs that wish to add a work item to a WG charter to define a new DHCP option should get clarity from the WG as to whether the new option will require a new fragment type or new semantics, or whether it is a simple DHCP option that fits existing definitions. If a WG needs a new fragment type, it is preferable to see if another WG exists whose members already have sufficient expertise to evaluate the new work. If such a working group is available, the work should be chartered in that working group instead. If there is no other WG with DHCP expertise that can define the new fragment type, the responsible AD should seek help from known DHCP experts within the IETF to provide advice and frequent early review as the original WG defines the new fragment type. In either case, the new option should be defined in a separate document, and the work should focus on defining a new format that generalizes well and can be reused, rather than a single-use fragment type. The WG that needs the new fragment type can define their new option referencing the new fragment type document, and the work can generally be done in parallel, avoiding unnecessary delays. Having the definition in its own document will foster reuse of the new fragment type. The responsible AD should work with all relevant WG Chairs and DHCP experts to ensure that the new fragment type document has in fact been carefully reviewed by the experts and appears satisfactory. Responsible Area Directors for WGs that are considering defining options that actually update DHCP, as opposed to simple options, should go through a process similar to that described above when trying to determine where to do the work. Under no circumstances should a WG be given a charter deliverable to define a new DHCP option, and then on the basis of that charter item actually make updates to DHCP.

When defining new options, one specific consideration to evaluate is whether or not options of a similar format would need to have multiple or single values encoded (whatever differs from the current option) and how that might be accomplished in a similar format. When defining a new option, it is best to synthesize the option format using fragment types already in use. However, in some cases, there may be no fragment type that accomplishes the intended purpose. The matter of size considerations and option order are further discussed in Sections and .

DHCPv6 allows for packet sizes up to 64 KB. First, through its use of link-local addresses, it avoids many of the deployment problems that plague DHCPv4 and is actually a UDP over the IPv6-based protocol (compared to DHCPv4, which is mostly UDP over IPv4 but with layer-2 hacks). Second, RFC 3315 explicitly refers readers to Section 5 of , which describes an MTU of 1280 octets and a minimum fragment reassembly of 1500 octets. It's feasible to suggest that DHCPv6 is capable of having larger options deployed over it, and at least no common upper limit is yet known to have been encoded by its implementors. It is not really possible to describe a fixed limit that cleanly divides workable option sizes from those that are too big. It is advantageous to prefer option formats that contain the desired information in the smallest form factor that satisfies the requirements. Common sense still applies here. It is better to split distinct values into separate octets rather than propose overly complex bit-shifting operations to save several bits (or even an octet or two) that would be padded to the next octet boundary anyway. DHCPv6 does allow for multiple instances of a given option, and they are treated as distinct values following the defined format; however, this feature is generally preferred to be restricted to protocol class features (such as the IA_* series of options). In such cases, it is better to define an option as an array if it is possible. It is recommended to clarify (with normative language) whether a given DHCPv6 option may appear once or multiple times. The default assumption is only once. In general, if a lot of data needs to be configured (for example, some option lengths are quite large), DHCPv6 may not be the best choice to deliver such configuration information and SHOULD simply be used to deliver a URI that specifies where to obtain the actual configuration information.

Although states that each option type MAY appear more than once, the original idea was that multiple instances are reserved for stateful options, like IA_NA or IA_PD. For most other options, it is usually expected that they will appear once at most. Such options are called singleton options. Sadly, RFCs have often failed to clearly specify whether or not a given option can appear more than once. Documents that define new options SHOULD state whether or not these options are singletons. Unless otherwise specified, newly defined options are considered to be singletons. If multiple instances are allowed, the document MUST explain how to use them. Care should be taken not to assume that they will be processed in the order they appear in the message. See for more details. When deciding whether single or multiple option instances are allowed in a message, take into consideration how the content of the option will be used. Depending on the service being configured, it may or may not make sense to have multiple values configured. If multiple values make sense, it is better to explicitly allow that by using an option format that allows multiple values within one option instance. Allowing multiple option instances often leads to confusion. Consider the following example. Basic DS-Lite architecture assumes that the B4 element (DHCPv6 client) will receive the AFTR option and establish a single tunnel to the configured tunnel termination point (AFTR). During the standardization process of , there was a discussion whether multiple instances of the DS-Lite tunnel option should be allowed. This created an unfounded expectation that the clients receiving multiple instances of the option will somehow know when one tunnel endpoint goes offline and do some sort of failover between other values provided in other instances of the AFTR option. Others assumed that if there are multiple options, the client will somehow do load balancing between the provided tunnel endpoints. Neither failover nor load balancing was defined for the DS-Lite architecture, so it caused confusion. It was eventually decided to allow only one instance of the AFTR option.

Option order, either the order among many DHCPv6 options or the order of multiple instances of the same option, SHOULD NOT be significant. New documents MUST NOT assume any specific option processing order. As there is no explicit order for multiple instances of the same option, an option definition SHOULD instead restrict ordering by using a single option that contains ordered fields. As does not impose option order, some implementations use hash tables to store received options (which is a conformant behavior). Depending on the hash implementation, the processing order is almost always different then the order in which the options appeared in the packet on the wire.

In DHCPv4, all relay options are organized as sub-options within the DHCP Relay Agent Information option . And, an independent number space called "DHCP Relay Agent Sub-options" is maintained by IANA. Different from DHCPv4, in DHCPv6, relay options are defined in the same way as client/server options, and they also use the same option number space as client/server options. Future DHCPv6 relay options MUST be allocated from this single DHCPv6 option number space. For example, the Relay-Supplied Options option may also contain some DHCPv6 options as permitted, such as the Extensible Authentication Protocol (EAP) Re-authentication Protocol (ERP) Local Domain Name DHCPv6 Option .

The DHCPv6 Option Request Option (OPTION_ORO) is an option that serves two purposes -- to inform the server what options the client supports and what options the client is willing to consume. For some options, such as the options required for the functioning of DHCPv6 itself, it doesn't make sense to require that they be explicitly requested using the Option Request Option. In all other cases, it is prudent to assume that any new option must be present on the relevant option request list if the client desires to receive it. It is tempting to add text that requires the client to include a new option in the Option Request Option list, similar to this text: "Clients MUST place the foo option code on the Option Request Option list, clients MAY include option foo in their packets as hints for the server as values the desire, and servers MUST include option foo when the client requests it (and the server has been so configured)". Such text is discouraged as there are several issues with it. First, it assumes that client implementation that supports a given option will always want to use it. This is not true. The second and more important reason is that such text essentially duplicates the mechanism already defined in . It is better to simply refer to the existing mechanism rather than define it again. See for proposed examples on how to do that. Creators of DHCPv6 options cannot assume special ordering of options either as they appear in the Option Request Option or as they appear within the packet. Although it is reasonable to expect that options will be processed in the order they appear in ORO, server software is not required to sort DHCPv6 options into the same order in Reply messages. It should also be noted that options values are not required to be aligned within the DHCP packet; even the option code and option length may appear on odd-byte boundaries.

The transition from IPv4 to IPv6 is progressing. Many transition technologies are proposed to speed it up. As a natural consequence, there are also DHCP options proposed to provision those proposals. The inevitable question is whether the required parameters should be delivered over DHCPv4 or DHCPv6. Authors often don't give much thought about it and simply pick DHCPv6 without realizing the consequences. IPv6 is expected to stay with us for many decades, and so is DHCPv6. There is no mechanism available to deprecate an option in DHCPv6, so any options defined will stay with us as long as the DHCPv6 protocol itself lasts. It seems likely that such options defined to transition from IPv4 will outlive IPv4 by many decades. From that perspective, it is better to implement provisioning of the transition technologies in DHCPv4, which will be obsoleted together with IPv4. When the network infrastructure becomes IPv6 only, the support for IPv4-only nodes may still be needed. In such a scenario, a mechanism for providing IPv4 configuration information over IPv6-only networks may be needed. See for further details.

There are three major entities in DHCPv6: server, relay agent, and client. It is very helpful for implementors to include separate sections that describe operation for those three major entities. Even when a given entity does not participate, it is useful to have a very short section stating that it must not send a given option and must ignore it when received. There is also a separate entity called the "requestor", which is a special client-like type that participates in the leasequery protocol . A similar section for the requestor is not required, unless the new option has anything to do with the requestor (or it is likely that the reader may think that is has). It should be noted that while in the majority of deployments the requestor is co-located with the relay agent, those are two separate entities from the protocol perspective, and they may be used separately. There are stand-alone requestor implementations available. The following sections include proposed text for such sections. That text is not required to appear, but it is appropriate in most cases. Additional or modified text specific to a given option is often required. Although the requestor is a somewhat uncommon functionality, its existence should be noted, especially when allowing or disallowing options to appear in certain messages or to be sent by certain entities. Additional message types may appear in the future, besides types defined in . Therefore, authors are encouraged to familiarize themselves with a list of currently defined DHCPv6 messages available on the IANA website . Typically, new options are requested by clients and assigned by the server, so there is no specific relay behavior. Nevertheless, it is good to include a section for relay agent behavior and simply state that there are no additional requirements for relays. The same applies for client behavior if the options are to be exchanged between the relay and server. Sections that contain option definitions MUST include a formal verification procedure. Often it is very simple, e.g., an option that conveys an IPv6 address must be exactly 16-bytes long, but sometimes the rules are more complex. It is recommended to refer to existing documents (e.g., Section 8 of RFC 3315 for domain name encoding) rather than trying to repeat such rules.

Clients MAY request option foo, as defined in , Sections 17.1.1, 18.1.1, 18.1.3, 18.1.4, 18.1.5, and 22.7. As a convenience to the reader, we mention here that the client includes requested option codes in the Option Request Option. Optional text (if the client's hints make sense): The client also MAY include option foo in its Solicit, Request, Renew, Rebind, and Information-request messages as a hint for the server regarding preferred option values. Optional text (if the option contains an FQDN): If the client requests an option that conveys an FQDN, it is expected that the contents of that option will be resolved using DNS. Hence, the following text may be useful: Clients that request option foo SHOULD also request option OPTION_DNS_SERVERS as specified in . Clients MUST discard option foo if it is invalid (i.e., it did not pass the validation steps defined in Section X.Y). Optional text (if option foo in expected to be exchanged between relays and servers): Option foo is exchanged between relays and servers only. Clients are not aware of the usage of option foo. Clients MUST ignore received option foo.

Sections 17.2.2 and 18.2 of govern server operation in regards to option assignment. As a convenience to the reader, we mention here that the server will send option foo only if configured with specific values for foo and if the client requested it. Optional text: Option foo is a singleton. Servers MUST NOT send more than one instance of the foo option. Optional text (if the server is never supposed to receive option foo): Servers MUST ignore the incoming foo option.

It's never appropriate for a relay agent to add options to a message heading toward the client, and relay agents don't actually construct Relay-reply messages anyway. Optional text (if the foo option is exchanged between the clients and server or between requestors and servers): there are no additional requirements for relays. Optional text (if relays are expected to insert or consume option foo): Relay agents MAY include option foo in a Relay-forward message when forwarding packets from clients to the servers.

Authors often ask themselves whether their proposal updates existing RFCs, especially RFC 3315. In April 2013, there were about 80 options defined. Had all documents that defined them also updated RFC 3315, comprehension of such a document set would be extremely difficult. It should be noted that "extends" and "updates" are two very different verbs. If a new document defines a new option that clients request and servers provide, it merely extends current standards, so "updates RFC 3315" is not required in the new document header. On the other hand, if a new document replaces or modifies existing behavior and includes clarifications or other corrections, it should be noted that it updates the other document. For example, clearly updates as it replaces existing text with new text. If in doubt, authors should try to determine whether an implementor reading the base RFC alone (without reading the new document) would be able to properly implement the software. If the base RFC is sufficient, then the new document probably does not update the base RFC. On the other hand, if reading your new document is necessary to properly implement the base RFC, then the new document most likely updates the base RFC.

DHCPv6 does have an authentication mechanism that makes it possible for DHCPv6 software to discriminate between authentic endpoints and man-in-the-middle. Other authentication mechanisms may optionally be deployed. Sadly, as of 2014, the authentication in DHCPv6 is rarely used, and support for it is not common in existing implementations. Some specific deployment types make it mandatory (or parts thereof, e.g., DOCSIS3.0-compatible cable modems require reconfigure-key support), so in certain cases, specific authentication aspects can be relied upon. That is not true in the generic case, though. So, while creating a new option, it is prudent to assume that the DHCPv6 packet contents are always transmitted in the clear, and actual production use of the software will probably be vulnerable at least to man-in-the-middle attacks from within the network, even where the network itself is protected from external attacks by firewalls. In particular, some DHCPv6 message exchanges are transmitted to multicast addresses that are likely broadcast anyway. If an option is of a specific fixed length, it is useful to remind the implementor of the option data's full length. This is easily done by declaring the specific value of the 'length' tag of the option. This helps to gently remind implementors to validate the option length before digesting them into likewise fixed-length regions of memory or stack. If an option may be of variable size (such as having indeterminate length fields, such as domain names or text strings), it is advisable to explicitly remind the implementor to be aware of the potential for long options. Either define a reasonable upper limit (and suggest validating it) or explicitly remind the implementor that an option may be exceptionally long (to be prepared to handle errors rather than truncate values). For some option contents, out-of-bound values may be used to breach security. An IP address field might be made to carry a loopback address or local multicast address, and depending on the protocol, this may lead to undesirable results. A domain name field may be filled with contrived contents that exceed the limitations placed upon domain name formatting; as this value is possibly delivered to "internal configuration" records of the system, it may be implicitly trusted without being validated. Authors of documents defining new DHCP options are, therefore, strongly advised to explicitly define validation measures that recipients of such options are required to do before processing such options. However, validation measures already defined by RFC 3315 or other specifications referenced by the new option document are redundant and can introduce errors, so authors are equally strongly advised to refer to the base specification for any such validation language rather than copying it into the new specification. See also .

As discussed in , the DHCPv6 packets are typically transmitted in the clear, so they are susceptible to eavesdropping. This should be considered when defining options that may convey personally identifying information (PII) or any other type of sensitive data. If the transmission of sensitive or confidential content is required, it is still possible to secure communication between relay agents and servers. Relay agents and servers communicating with relay agents must support the use of IPsec Encapsulating Security Payload (ESP) with encryption in transport mode, according to Section 3.1.1 of and Section 21.1 of . Sadly, this requirement is almost universally ignored in real deployments. Even if the communication path between the relay agents and server is secured, the path between the clients and relay agents or server is not. Unless underlying transmission technology provides a secure transmission channel, the DHCPv6 options SHOULD NOT include PII or other sensitive information. If there are special circumstances that warrant sending such information over unsecured DHCPv6, the dangers MUST be clearly discussed in the security considerations.

The authors would like to thank Simon Perreault, Bernie Volz, Ted Lemon, Bud Millwood, Ralph Droms, Barry Leiba, Benoit Claise, Brian Haberman, Richard Barnes, Stephen Farrell, and Stewart Bryant for their comments.