Service Registration Protocol for DNS-Based Service Discovery

Services that implement SRP use DNS Update to publish service information in the DNS. Two variants exist, one for full-featured hosts, and one for devices designed for "Constrained-Node Networks" . Full-featured hosts are either configured manually with a registration domain, or use the "dr._dns&nbhy;sd._udp.<domain>" query ( Section 11) to learn the default registration domain from the network. RFC6763 says to discover the registration domain using either ".local" or a network-supplied domain name for <domain>. Services using SRP MUST use the domain name received through the DHCPv4 Domain Name option ( section 3.17), if available, or the Neighbor Discovery DNS Search List option . If the DNS Search List option contains more than one domain name, it MUST NOT be used. If neither option is available, the Service Registration protocol is not available on the local network. Manual configuration of the registraton domain can be done either by querying the list of available registration zones ("r._dns&nbhy;sd._udp") and allowing the user to select one from the UI, or by any other means appropriate to the particular use case being addressed. Full-featured devices construct the names of the SRV, TXT, and PTR records describing their service(s) as subdomains of the chosen service registration domain. For these names they then discover the zone apex of the closest enclosing DNS zone using SOA queries . Having discovered the enclosing DNS zone, they query for the "_dnssd&nbhy;srp._tcp<zone>" SRV record to discover the server to which they should send DNS updates. Hosts that support SRP updates using TLS use the "_dnssd&nbhy;srp&nbhy;tls._tcp<zone>" SRV record instead. For devices designed for Constrained-Node Networks some simplifications are available. Instead of being configured with (or discovering) the service registration domain, the (proposed) special-use domain name (see ) "default.service.arpa" is used. The details of how SRP server(s) are discovered will be specific to the constrained network, and therefore we do not suggest a specific mechanism here. SRP clients on constrained networks are expected to receive from the network a list of SRP servers with which to register. It is the responsibility of a Constrained-Node Network supporting SRP to provide one or more SRP server addresses. It is the responsibility of the SRP server supporting a Constrained-Node Network to handle the updates appropriately. In some network environments, updates may be accepted directly into a local "default.service.arpa" zone, which has only local visibility. In other network environments, updates for names ending in "default.service.arpa" may be rewritten internally to names with broader visibility. The reason for these different assumptions is that low-power devices that typically use Constrained-Node Networks may have very limited battery power. The series of DNS lookups required to discover an SRP server and then communicate with it will increase the power required to advertise a service; for low-power devices, the additional flexibility this provides does not justify the additional use of power. It is also fairly typical of such networks that some network service information is obtained as part of the process of joining the network, and so this can be relied upon to provide nodes with the information they need. Networks that are not constrained networks can more complicated topologies at the Internet layer. Nodes connected to such networks can be assumed to be able to do DNSSD service registration domain discovery. Such networks are generally able to provide registration domain discovery and routing. By requiring the use of TCP, the possibility of off-network spoofing is eliminated. We will discuss several parts to this process: how to know what to publish, how to know where to publish it (under what name), how to publish it, how to secure its publication, and how to maintain the information once published.

We refer to the DNS Update message sent by services using SRP as an SRP update. Three types of updates appear in an SRP update: Service Discovery records, Service Description records, and Host Description records. Service Discovery records are one or more PTR RRs, mapping from the generic service type (or subtype) to the specific Service Instance Name. Service Description records are exactly one SRV RR, exactly one KEY RR, and one or more TXT RRs, all with the same name, the Service Instance Name ( section 4.1). In principle Service Description records can include other record types, with the same Service Instance Name, though in practice they rarely do. The Service Instance Name MUST be referenced by one or more Service Discovery PTR records, unless it is a placeholder service registration for an intentionally non-discoverable service name. The Host Description records for a service are a KEY RR, used to claim exclusive ownership of the service registration, and one or more RRs of type A or AAAA, giving the IPv4 or IPv6 address(es) of the host where the service resides. RFC 6763 describes the details of what each of these types of updates contains and is the definitive source for information about what to publish; the reason for summarizing this here is to provide the reader with enough information about what will be published that the service registration process can be understood at a high level without first learning the full details of DNS&nbhy;SD. Also, the "Service Instance Name" is an important aspect of first-come, first-serve naming, which we describe later on in this document.

Multicast DNS uses a single namespace, ".local", which is valid on the local link. This convenience is not available for DNS&nbhy;SD using the DNS protocol: services must exist in some specific unicast namespace. As described above, full-featured devices are responsible for knowing in what domain they should register their services. Devices made for Constrained-Node Networks register in the (proposed) special use domain name "default.service.arpa", and let the SRP server handle rewriting that to a different domain if necessary.

It is possible to issue a DNS Update that does several things at once; this means that it's possible to do all the work of adding a PTR resource record to the PTR RRset on the Service Name, and creating or updating the Service Instance Name and Host Description, in a single transaction. An SRP update takes advantage of this: it is implemented as a single DNS Update message that contains a service's Service Discovery records, Service Description records, and Host Description records. Updates done according to this specification are somewhat different than regular DNS Updates as defined in RFC2136. The RFC2136 update process can involve many update attempts: you might first attempt to add a name if it doesn't exist; if that fails, then in a second message you might update the name if it does exist but matches certain preconditions. Because the registration protocol uses a single transaction, some of this adaptability is lost. In order to allow updates to happen in a single transaction, SRP updates do not include update prerequisites. The requirements specified in are implicit in the processing of SRP updates, and so there is no need for the service sending the SRP update to put in any explicit prerequisites.

DNS&nbhy;SD Service Registration is based on standard RFC2136 DNS Update, with some differences: It implements first-come first-served name allocation, protected using SIG(0) . It enforces policy about what updates are allowed. It optionally performs rewriting of "default.service.arpa" to some other domain. It optionally performs automatic population of the address-to-name reverse mapping domains. An SRP server is not required to implement general DNS Update prerequsite processing. Clients are allowed to send updates to the generic domain "default.service.arpa"

Traditional DNS update is secured using the TSIG protocol, which uses a secret key shared between the client (which issues the update) and the server (which authenticates it). This model does not work for automatic service registration. The goal of securing the DNS&nbhy;SD Registration Protocol is to provide the best possible security given the constraint that service registration has to be automatic. It is possible to layer more operational security on top of what we describe here, but what we describe here is an improvement over the security of mDNS. The goal is not to provide the level of security of a network managed by a skilled operator.

First-Come First-Serve naming provides a limited degree of security: a service that registers its service using DNS&nbhy;SD Registration protocol is given ownership of a name for an extended period of time based on the key used to authenticate the DNS Update. As long as the registration service remembers the name and the key used to register that name, no other service can add or update the information associated with that. FCFS naming is used to protect both the Service Description and the Host Description.

The service generates a public/private key pair. This key pair MUST be stored in stable storage; if there is no writable stable storage on the client, the client MUST be pre-configured with a public/private key pair in read-only storage that can be used. This key pair MUST be unique to the device. When sending DNS updates, the service includes a KEY record containing the public portion of the key in each Host Description update and each Service Description update. Each KEY record MUST contain the same public key. The update is signed using SIG(0), using the private key that corresponds to the public key in the KEY record. The lifetimes of the records in the update is set using the EDNS(0) Update Lease option . The KEY record in Service Description updates MAY be omitted for brevity; if it is omitted, the SRP server MUST behave as if the same KEY record that is given for the Host Description is also given for each Service Description for which no KEY record is provided. Omitted KEY records are not used when computing the SIG(0) signature. The lifetime of the DNS&nbhy;SD PTR, SRV, A, AAAA and TXT records uses the LEASE field of the Update Lease option, and is typically set to two hours. This means that if a device is disconnected from the network, it does not appear in the user interfaces of devices looking for services of that type for too long. The lifetime of the KEY records is set using the KEY-LEASE field of the Update Lease Option, and should be set to a much longer time, typically 14 days. The result of this is that even though a device may be temporarily unplugged, disappearing from the network for a few days, it makes a claim on its name that lasts much longer. This means that even if a device is unplugged from the network for a few days, and its services are not available for that time, no other device can come along and claim its name the moment it disappears from the network. In the event that a device is unplugged from the network and permanently discarded, then its name is eventually cleaned up and made available for re-use.

To remove a service registration, the client retransmits its most recent update with an Update Lease option that has a LEASE value of zero. If the registration is to be permanently removed, KEY-LEASE should also be zero. Otherwise, it should have the same value it had previously; this holds the name in reserve for when the client is once again able to provide the service. SRP clients are normally expected to remove all service instances when removing a host. However, in some cases a client may not have retained sufficient state to know that some service instance is pointing to a host that it is removing. Nevertheless, removing the host can be assumed to mean that all service instances pointing to it are no longer valid. Therefore, SRP servers MAY remove all service instances pointing to a host when a host is removed, even if the client doesn't remove them explicitly.

The SRP server first validates that the DNS Update is a syntactically and semantically valid DNS Update according to the rules specified in RFC2136. SRP Updates consist of a set of Instructions that together add one or more services. Each instruction consists either of a single add, or a delete followed by an add. When an instruction contains a delete and an add, the delete MUST precede the add. The SRP server checks each Instruction in the SRP update to see that it is either a Service Discovery update, a Service Description update, or a Host Description update. Order matters in DNS updates. Specifically, deletes must precede adds for records that the deletes would affect; otherwise the add will have no effect. This is the only ordering constraint; aside from this constraint, updates may appear in whatever order is convenient when constructing the update. Because the SRP update is a DNS update, it MUST contain a single question that indicates the zone to be updated. Every delete and update in an SRP update MUST be within the zone that is specified for the SRP Update. An Instruction is a Service Discovery Instruction if it contains exactly one "Add to an RRSet" ( Section 2.5.1) RR, which is a PTR RR, which points to a Service Instance Name for which a Service Description Instruction is present in the SRP Update. Service Discovery Instructions do not contain any deletes, and do not contain any other adds. An Instruction is a Service Description Instruction if, for the appropriate Service Instance Name, it contains exactly one "Delete all RRsets from a name" update for the service instance name Section 2.5.3, exactly one "Add to an RRset" SRV RR, zero or one "Add to an RRset" KEY RR that contains the public key corresponding to the private key that was used to sign the message (if present, the KEY MUST match the KEY RR given in the Host Description), one or more "Add to an RRset" TXT RRs, and the target of the SRV RR Add points to a hostname for which there is a Host Description Instruction in the SRP Update. Service Descriptions Instructions do not modify any other RRs. An Instruction is a Host Description Instruction if, for the appropriate hostname, it contains exactly one "Delete all RRsets from a name" RR, one or more "Add to an RRset" RRs of type A and/or AAAA, exactly one "Add to an RRset" RR that adds a KEY RR that contains the public key corresponding to the private key that was used to sign the message, there is a Service Instance Name Instruction in the SRP update for which the SRV RR that is added points to the hostname being updated by this update. Host Description updates do not modify any other records. An SRP Update MUST include at least one Service Discovery Instruction, at least one Service Description Instruction, and exactly one Host Description Instruction. A DNS Update that does not is not an SRP update. A DNS Update that contains any other adds, any other deletes, or any prerequisites, is not an SRP update. Such messages should either be processed as regular RFC2136 updates, including access control checks and constraint checks, if supported, or else rejected with RCODE=REFUSED. Note that if the definitions of each of these update types are followed carefully, this means that many things that look very much like SRP updates nevertheless are not. For example, a DNS update that contains an RRset Add to a Service Name and an RRset Add to a Service Instance Name, where the Service Name does not reference the Service Instance Name, is not a valid SRP update message, but may be a valid RFC2136 update. Assuming that a DNS Update message has been validated with these conditions and is a valid SRP Update, the server checks that the name in the Host Description Instruction exists. If so, then the server checks to see if the KEY record on that name is the same as the KEY record in the Host Description Instruction. The server performs the same check for the KEY records in any Service Description Instrructions. For KEY records that were omitted from Service Description Instructions, the KEY from the Host Description Instruction is used. If any existing KEY record corresponding to a KEY record in the SRP Update does not match the KEY same record in the SRP Update (whether provided or taken from the Host Description Instruction), then the server MUST reject the SRP Update with the YXDOMAIN RCODE. Otherwise, the server validates the SRP Update using SIG(0) on the public key in the KEY record of the Host Description update. If the validation fails, the server MUST reject the SRP Update with the REFUSED RCODE. Otherwise, the SRP update is considered valid and authentic, and is processed according to the method described in RFC2136. KEY record updates omitted from Service Description update are processed as if they had been explicitly present: every Service Description that is updated MUST, after the update, have a KEY RR, and it must be the same KEY RR that is present in the Host Description to which the Service Description refers. The status that is returned depends on the result of processing the update, and can be either SUCCESS or SERVFAIL: all other possible outcomes should already have been accounted for when applying the constraints that qualify the update as an SRP Update. The server MAY add a Reverse Mapping that corresponds to the Host Description. This is not required because the Reverse Mapping serves no protocol function, but it may be useful for debugging, e.g. in annotating network packet traces or logs. In order for the server to add a reverse mapping update, it must be authoritative for the zone or have credentials to do the update. The client MAY also do a reverse mapping update if it has credentials to do so. The server MAY apply additional criteria when accepting updates. In some networks, it may be possible to do out-of-band registration of keys, and only accept updates from pre-registered keys. In this case, an update for a key that has not been registered should be rejected with the REFUSED RCODE. There are at least two benefits to doing this rather than simply using normal SIG(0) DNS updates. First, the same registration protocol can be used in both cases, so both use cases can be addressed by the same service implementation. Second, the registration protocol includes maintenance functionality not present with normal DNS updates. Note that the semantics of using SRP in this way are different than for typical RFC2136 implementations: the KEY used to sign the SRP update only allows the client to update records that refer to its Host Description. RFC2136 implementations do not normally provide a way to enforce a constraint of this type. The server may also have a dictionary of names or name patterns that are not permitted. If such a list is used, updates for Service Instance Names that match entries in the dictionary are rejected with YXDOMAIN.

All RRs within an RRset are required to have the same TTL (Clarifications to the DNS Specification, Section 5.2). In order to avoid inconsistencies, SRP places restrictions on TTLs sent by services and requires that SRP Servers enforce consistency. Services sending SRP updates MUST use consistent TTLs in all RRs within the SRP update. SRP update servers MUST check that the TTLs for all RRs within the SRP update are the same. If they are not, the SRP update MUST be rejected with a REFUSED RCODE. Additionally, when adding RRs to an RRset, for example when processing Service Discovery records, the server MUST use the same TTL on all RRs in the RRset. How this consistency is enforced is up to the implementation. TTLs sent in SRP updates are advisory: they indicate the client's guess as to what a good TTL would be. SRP servers may override these TTLs. SRP servers SHOULD ensure that TTLs are reasonable: neither too long nor too short. The TTL should never be longer than the lease time . Shorter TTLs will result in more frequent data refreshes; this increases latency on the client side, and increases load on any caching resolvers and on the authoritative server. Longer TTLs will increase the likelihood that data in caches will be stale. TTL minimums and maximums SHOULD be configurable by the operator of the SRP server.

Because the DNS&nbhy;SD registration protocol is automatic, and not managed by humans, some additional bookkeeping is required. When an update is constructed by the client, it MUST include include an EDNS(0) Update Lease Option . The Update Lease Option contains two lease times: the Lease Time and the Key Lease Time. These leases are promises, similar to DHCP leases, from the client that it will send a new update for the service registration before the lease time expires. The Lease time is chosen to represent the time after the update during which the registered records other than the KEY record should be assumed to be valid. The Key Lease time represents the time after the update during which the KEY record should be assumed to be valid. The reasoning behind the different lease times is discussed in the section on first-come, first-served naming . SRP servers may be configured with limits for these values. A default limit of two hours for the Lease and 14 days for the SIG(0) KEY are currently thought to be good choices. Clients that are going to continue to use names on which they hold leases should update well before the lease ends, in case the registration service is unavailable or under heavy load. The SRP server MUST include an EDNS(0) Update Lease option in the response if the lease time proposed by the service has been shortened or lengthened. The service MUST check for the EDNS(0) Update Lease option in the response and MUST use the lease times from that option in place of the options that it sent to the server when deciding when to update its registration. The times may be shorter or longer than those specified in the SRP update; the client must honor them in either case. Clients should assume that each lease ends N seconds after the update was first transmitted, where N is the lease duration. Servers should assume that each lease ends N seconds after the update that was successfully processed was received. Because the server will always receive the update after the client sent it, this avoids the possibility of misunderstandings. SRP servers MUST reject updates that do not include an EDNS(0) Update Lease option. Dual-use servers MAY accept updates that don't include leases, but SHOULD differentiate between SRP updates and other updates, and MUST reject updates that would otherwise be SRP updates updates if they do not include leases. Lease times have a completely different function than TTLs. On an authoritative DNS server, the TTL on a resource record is a constant: whenever that RR is served in a DNS response, the TTL value sent in the answer is the same. The lease time is never sent as a TTL; its sole purpose is to determine when the authoritative DNS server will delete stale records. It is not an error to send a DNS response with a TTL of 'n' when the remaining time on the lease is less than 'n'.

Another use of SRP is for devices that sleep to reduce power consumption. In this case, in addition to the DNS Update Lease option described above, the device includes an EDNS(0) OWNER Option. The EDNS(0) Update Lease option constitutes a promise by the device that it will wake up before this time elapses, to renew its registration and thereby demonstrate that it is still attached to the network. If it fails to renew the registration by this time, that indicates that it is no longer attached to the network, and its registration (except for the KEY in the Host Description) should be deleted. The EDNS(0) OWNER Option indicates that the device will be asleep, and will not be receptive to normal network traffic. When a DNS server receives a DNS Update with an EDNS(0) OWNER Option, that signifies that the SRP server should set up a proxy for any IPv4 or IPv6 address records in the DNS Update message. This proxy should send ARP or ND messages claiming ownership of the IPv4 and/or IPv6 addresses in the records in question. In addition, proxy should answer future ARP or ND requests for those IPv4 and/or IPv6 addresses, claiming ownership of them. When the DNS server receives a TCP SYN or UDP packet addressed to one of the IPv4 or IPv6 addresses for which it proxying, it should then wake up the sleeping device using the information in the EDNS(0) OWNER Option. At present version 0 of the OWNER Option specifies the “Wake-on-LAN Magic Packet” that needs to be sent; future versions could be extended to specify other wakeup mechanisms. Note that although the authoritative DNS server that implements the SRP function need not be on the same link as the sleeping host, the Sleep Proxy must be on the same link. It is not required that sleepy nodes on a Constrained-Node Network support sleep proxy. Such devices may have different mechanisms for dealing with sleep and wakeup. An SRP registration for such a device will be useful regardless of the mechanism whereby messages are delivered to the sleepy end device. For example, the message might be held in a buffer for an extended period of time by an intermediate device on a mesh network, and then delivered to the device when it wakes up. The exact details of such behaviors are out of scope for this document.