Binary Encodings for JavaScript Object Notation: JSON-B, JSON-C, JSON-D

JavaScript Object Notation (JSON) is a simple text encoding for the JavaScript Data model that has found wide application beyond its original field of use. In particular JSON has rapidly become a preferred encoding for Web Services. JSON encoding supports just four fundamental data types (integer, floating point, string and boolean), arrays and objects which consist of a list of tag-value pairs. Although the JSON encoding is sufficient for many purposes it is not always efficient. In particular there is no efficient representation for blocks of binary data. Use of base64 encoding increases data volume by 33%. This overhead increases exponentially in applications where nested binary encodings are required making use of JSON encoding unsatisfactory in cryptographic applications where nested binary structures are frequently required. Another source of inefficiency in JSON encoding is the repeated occurrence of object tags. A JSON encoding containing an array of a hundred objects such as {"first":1,"second":2} will contain a hundred occurrences of the string "first" (seven bytes) and a hundred occurrences of the string "second" (eight bytes). Using two byte code sequences in place of strings allows a saving of 11 bytes per object without loss of information, a saving of 50%. A third objection to the use of JSON encoding is that floating point numbers can only be represented in decimal form and this necessarily involves a loss of precision when converting between binary and decimal representations. While such issues are rarely important in network applications they can be critical in scientific applications. It is not acceptable for saving and restoring a data set to change the result of a calculation.

The following were identified as core objectives for a binary JSON encoding: Easy to convert existing encoders and decoders to add binary support Efficient encoding of binary data Ability to convert from JSON to binary encoding in a streaming mode (i.e. without reading the entire binary data block before beginning encoding. Lossless encoding of JavaScript data types The ability to support JSON tag compression and extended data types are considered desirable but not essential for typical network applications. Three binary encodings are defined: Encodes JSON data in binary. Only the JavaScript data model is supported (i.e. atomic types are integers, double or string). Integers may be 8, 16, 32 or 64 bits either signed or unsigned. Floating points are IEEE 754 binary64 format [IEEE754] . Supports chunked encoding for binary and UTF-8 string types. As JSON-B but with support for representing JSON tags in numeric code form (16 bit code space). This is done for both compact encoding and to allow simplification of encoders/decoders in constrained environments. Codes may be defined inline or by reference to a known dictionary of codes referenced via a digest value. As JSON-C but with support for representing additional data types without loss of precision. In particular other IEEE 754 floating point formats, both binary and decimal and Intel's 80 bit floating point, plus 128 bit integers and bignum integers. Each encoding is a proper superset of JSON, JSON-C is a proper superset of JSON-B and JSON-D is a proper superset of JSON-C. Thus a single decoder MAY be used for all three new encodings and for JSON. Figure 1 shows these relationships graphically:

[[This figure is not viewable in this format. The figure is available at http://prismproof.org/Documents/draft-hallambaker-jsonbcd.html.]] Encoding Relationships.

This section presents the related specifications and standard, the terms that are used as terms of art within the documents and the terms used as requirements language.

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].

The terms of art used in this document are described in the Mesh Architecture Guide [draft-hallambaker-mesh-architecture] .

The JSON-B, JSON-C and JSON-D encodings are all based on the JSON grammar [RFC7159] . IEEE 754 Floating Point Standard is used for encoding floating point numbers [IEEE754] ,

No new terms of art are defined

The JSON-B, JSON-C and JSON-D encodings are all based on the JSON grammar [RFC7159] using the same syntactic structure but different lexical encodings. JSON-B0 and JSON-C0 replace the JSON lexical encodings for strings and numbers with binary encodings. JSON-B1 and JSON-C1 allow either lexical encoding to be used. Thus any valid JSON encoding is a valid JSON-B1 or JSON-C1 encoding. The grammar of JSON-B, JSON-C and JSON-D is a superset of the JSON grammar. The following productions are added to the grammar: Binary encodings for data values. As the binary value encodings are all self delimiting An object member where the value is specified as an X-value and thus does not require a value-separator. Binary data encodings defined in JSON-B. Defined length string encoding defined in JSON-B. Tag code definition defined in JSON-C. These may only appear before the beginning of an Object or Array and before any preceding white space. Tag code value defined in JSON-C. Additional binary data encodings defined in JSON-D for use in scientific data applications. The JSON grammar is modified to permit the use of x-value productions in place of ( value value-separator ) :

The productions number and string are defined as before:

The JSON-B encoding defines the b-value and b-string productions:

The lexical encodings of the productions are defined in the following tables where the column 'tag' specifies the byte code that begins the production, 'Fixed' specifies the number of data bytes that follow and 'Length' specifies the number of bytes used to define the length of a variable length field following the data bytes: Production Tag Fixed Length Data Description string-term x80 - 1 Terminal String 8 bit length string-term x81 - 2 Terminal String 16 bit length string-term x82 - 4 Terminal String 32 bit length string-term x83 - 8 Terminal String 64 bit length string-chunk x84 - 1 Terminal String 8 bit length string-chunk x85 - 2 Terminal String 16 bit length string-chunk x86 - 4 Terminal String 32 bit length string-chunk x87 - 8 Terminal String 64 bit length data-term x88 - 1 Terminal String 8 bit length data-term x89 - 2 Terminal String 16 bit length data-term x8A - 4 Terminal String 32 bit length data-term x8B - 8 Terminal String 64 bit length data-term X8C - 1 Terminal String 8 bit length data-term x8D - 2 Terminal String 16 bit length data-term x8E - 4 Terminal String 32 bit length data-term x8F - 8 Terminal String 64 bit length Table 1: Codes for String and Data items Production Tag Fixed Length Data Description p-int8 xA0 1 - Positive 8 bit Integer p-int16 Xa1 2 - Positive 16 bit Integer p-int32 Xa2 4 - Positive 32 bit Integer p-int64 Xa3 8 - Positive 64 bit Integer p-bignum16 Xa5 - 2 Positive Bignum n-int8 xA8 1 - Negative 8 bit Integer n-int16 xA9 2 - Negative 16 bit Integer n-int32 xAA 4 - Negative 32 bit Integer n-int64 xAB 8 - Negative 64 bit Integer n-bignum16 xAD - 2 Negative Bignum binary64 x92 8 - IEEE 754 Floating Point Binary 64 bit b-value xB0 - - True b-value xB1 - - False b-value xB2 - - Null Table 2: Codes for Integers, 64 Bit Floating Point, Boolean and Null items. A data type commonly used in networking that is not defined in this scheme is a datetime representation. To define such a data type, a string containing a date-time value in Internet type format is typically used.

The following examples show examples of using JSON-B encoding:

JSON-C (Compressed) permits numeric code values to be substituted for strings and binary data. Tag codes MAY be 8, 16 or 32 bits long encoded in network byte order. Tag codes MUST be defined before they are referenced. A Tag code MAY be defined before the corresponding data or string value is used or at the same time that it is used. A dictionary is a list of tag code definitions. An encoding MAY incorporate definitions from a dictionary using the dict-hash production. The dict hash production specifies a (positive) offset value to be added to the entries in the dictionary followed by the UDF fingerprint [draft-hallambaker-udf] of the dictionary to be used. Production Tag Fixed Length Data Description c-tag xC0 1 - 8 bit tag code c-tag xC1 2 - 16 bit tag code c-tag xC2 4 - 32 bit tag code c-def xC4 1 - 8 bit tag definition c-def xC5 2 - 16 bit tag definition c-def xC6 4 - 32 bit tag definition c-tag xC8 1 - 8 bit tag code and definition c-tag xC9 2 - 16 bit tag code and definition c-tag xCA 4 - 32 bit tag code and definition c-def xCC 1 - 8 bit tag dictionary definition c-def xCD 2 - 16 bit tag dictionary definition c-def xCE 4 - 32 bit tag dictionary definition dict-hash xD0 4 1 UDF fingerprint of dictionary Table 3: Codes Used for Compression All integer values are encoded in Network Byte Order (most significant byte first).

The following examples show examples of using JSON-C encoding:

JSON-B and JSON-C only support the two numeric types defined in the JavaScript data model: Integers and 64 bit floating point values. JSON-D (Data) defines binary encodings for additional data types that are commonly used in scientific applications. These comprise positive and negative 128 bit integers, six additional floating point representations defined by IEEE 754 [IEEE754] and the Intel extended precision 80 bit floating point representation [INTEL] . Should the need arise, even bigger bignums could be defined with the length specified as a 32 bit value permitting bignums of up to 2^35 bits to be represented.

The codes for these values are as follows: Production Tag Fixed Length Data Description p-int128 xA4 16 - Positive 128 bit Integer n-int128 xAC 16 - Negative 128 bit Integer binary16 x90 2 - IEEE 754 Floating Point Binary 16 bit binary32 x91 4 - IEEE 754 Floating Point Binary 32 bit binary128 x94 16 - IEEE 754 Floating Point Binary 64 bit Intel80 x95 10 - Intel extended Floating Point 80 bit decimal32 x96 4 - IEEE 754 Floating Point Decimal 32 Decimal64 x97 8 - IEEE 754 Floating Point Decimal 64 Decimal128 x98 16 - IEEE 754 Floating Point Decimal 128 Table 4: Additional Codes for Scientific Data

This work was assisted by conversations with Nico Williams and other participants on the applications area mailing list.

A correctly implemented data encoding mechanism should not introduce new security vulnerabilities. However, experience demonstrates that some data encoding approaches are more prone to introduce vulnerabilities when incorrectly implemented than others. In particular, whenever variable length data formats are used, the possibility of a buffer overrun vulnerability is introduced. While best practice suggests that a coding language with native mechanisms for bounds checking is the best protection against such errors, such approaches are not always followed. While such vulnerabilities are most commonly seen in the design of decoders, it is possible for the same vulnerabilities to be exploited in encoders. A common source of such errors is the case where nested length encodings are used. For example, a decoder relies on an outermost length encoding that specifies a length on 50 bytes to allocate memory for the entire result and then attempts to copy a string with a declared length of 1000 bytes within the sequence. The extensions to the JSON encoding described in this document are designed to avoid such errors. Length encodings are only used to define the length of x-value constructions which are always terminal and cannot have nested data entries.

[TBS list out all the code points that require an IANA registration]