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 RFC 2119.
This document is released into the public domain.
This document was authored by Jeremy Rand, with input from Ryan Castellucci and Hugo Landau.
<!-- START doctoc generated TOC please keep comment here to allow auto update -->
<!-- DON'T EDIT THIS SECTION, INSTEAD RE-RUN doctoc TO UPDATE -->
This document extends the Domain Names specification, IFA-0001. It specifies an additional format of encoding the `tls` item type. This format is more suited for integrating `tls` item support into mainstream certificate trust stores, given the API limitations of those trust stores.
IFA-0001 specifies that a TLS item is a JSON list. This specification redefines TLS items to instead be a JSON object, of the form `{encoding: data}`. *encoding* can be `"dane"`, in which case `data` is a JSON list in the form specified by IFA-0001. *encoding* can also be `"d8"`, in which case `data` is in the form specified below. If multiple *encoding* keys exist for a single TLS item (e.g. both `dane` and `d8` are available), then implementations MUST pick one of their choice. Implementations MAY assume that all *encoding* options for a TLS item are equivalent. If a TLS item that contains a JSON list instead of a JSON object is encountered, implementations MUST treat the list `list` as the JSON object `{"dane": list}`, which ensures backward-compatibility with IFA-0001.
1. an integer representing a version number. Currently `1` is the only defined version. Dehydrated certificates with an unknown version MUST be ignored.
2. a base64 string representing the public key. Typically an EC public key, not RSA, because EC is much more compact. The public key is not compressed, because compressed EC public keys are not widely supported by libraries (particularly the Golang standard library).
3. an integer representing the "not before" part of the validity period, in units of 5 minutes. 5 minutes essentially divides the number by 300, which eliminates slightly more than 1 byte.
4. an integer representing the "not after" part of the validity period, in units of 5 minutes. 5 minutes essentially divides the number by 300, which eliminates slightly more than 1 byte.
5. an integer representing the signature algorithm. Typically ECDSA with SHA256.
6. a base64 string representing the signature. Typically ECDSA with SHA256.
1. A valid certificate can be rehydrated from these 6 items (plus a domain name). This means that TLS applications that require possessing a full certificate (not just a public key hash) in order to set up an override can be configured without out-of-band retrieval of a certificate.
2. These items are much smaller (they are circa 255 bytes of JSON) than the rehydrated certificate would be, and therefore can comfortably fit into a datastore where storage is expensive (e.g. a Namecoin value).
3. Rehydrated certificates are minimally trusted. For example, rehydrated certificates cannot be valid for other domain names or uses (via SAN, the CA flag, etc.) than the one which the user retrieving a TLSA record expects. As a result, these certificates can be added to a root CA trust store with minimal security implications.
To rehydrate into a full certificate, construct a self-signed certificate with the following fields:
* NotBefore: based on dehydrated data, after converting units.
* NotAfter: based on dehydrated data, after converting units.
* SerialNumber: first 19 bytes of SHA256(SHA256(the .bit domain in use) | SHA256(PublicKey) | SHA256(big-endian int64 of NotBefore in units of 5 minutes) | SHA256(big-endian int64 of NotAfter in units of 5 minutes))
Note that a canonical implementation (in Go) of this scheme is being added to ncdns. Because the signatures will fail if the certificate is not constructed deterministically, the canonical implementation of certificate construction takes precedence over this spec and any examples. (Please report any inconsistencies.)
An IFA-0001-based JSON value which uses the above dehydrated certificate looks like this:
The primary goal is to fit comfortably into a Namecoin value (which is limited to 520 bytes) when combined with IP addresses or other data. We believe that it achieves this. The goal is **not** to achieve optimal compression at this time. We expect to improve on the scalability of Namecoin values in the future, via proposals such as CBOR encoding, value compression, and segregated name values; improvements in dehydrated certificate encoding (e.g. compressed public keys and removing the overhead of base64 and decimal encoding) are fair game for those future efforts. We are choosing not to fully optimize at this time because we believe that the current proposal is currently sufficient, and future revisions to this spec are affected by ecosystem elements out of the scope of this proposal, including uncertainty about the above scalability improvements, and also including uncertainty about how TLS implementations may change their requirements in the future (for example, Mozilla's WebExtensions spec is still up in the air).
Changes to Processing Rules
---------------------------
There are no changes to item suppression or other processing rules.
Dehydrated `tls` items are more widely compatible with currently deployed TLS implementations, e.g. Google Chrome and Mozilla Firefox. However, because their encoding contains more data than a public key hash, dehydrated `tls` items will incur higher storage costs than the encoding specified in IFA-0001. Domain name owners are advised to consider the tradeoffs between compatibility and storage costs when choosing how to encode their `tls` items. It is our belief that, in most publicly deployed website configurations, the improved compatibility of dehydrated `tls` items makes their usage preferable, and that their increased storage costs are negligible in such configurations.
