## Libra Canonical Serialization (LCS) LCS defines a deterministic means for translating a message or data structure into bytes irrespective of platform, architecture, or programming language. ### Background In Libra, participants pass around messages or data structures that often times need to be signed by a prover and verified by one or more verifiers. Serialization in this context refers to the process of converting a message into a byte array. Many serialization approaches support loose standards such that two implementations can produce two different byte streams that would represent the same, identical message. While for many applications, non-deterministic serialization causes no issues, it does so for applications using serialization for cryptographic purposes. For example, given a signature and a message, a verifier may not unable to produce the same serialized byte array constructed by the prover when the prover signed the message resulting in a non-verifiable message. In other words, to ensure message verifiability when using non-deterministic serialization, participants must either retain the original serialized bytes or risk losing the ability to verify messages. This creates a burden requiring participants to maintain both a copy of the serialized bytes and the deserialized message often leading to confusion about safety and correctness. While there exist a handful of existing deterministic serialization formats, there is no obvious choice. To address this, we propose Libra Canonical Serialization that defines a deterministic means for translating a message into bytes and back again. ### Specification LCS supports the following data types: * Booleans * Signed 8-bit, 16-bit, 32-bit, 64-bit, and 128-bit integers * Unsigned 8-bit, 16-bit, 32-bit, 64-bit, and 128-bit integers * Option * Unit (an empty value) * Fixed and variable length sequences * UTF-8 Encoded Strings * Tuples * Structures (aka "structs") * Externally tagged enumerations (aka "enums") * Maps ### General structure LCS is not a self-describing format and as such, in order to deserialize a message, one must know the message type and layout ahead of time. Unless specified, all numbers are stored in little endian, two's complement format. ### Recursion and Depth of LCS Data Recursive data-structures (e.g. trees) are allowed. However, because of the possibility of stack overflow during (de)serialization, the *container depth* of any valid LCS data cannot exceed the constant `MAX_CONTAINER_DEPTH`. Formally, we define *container depth* as the number of structs and enums traversed during (de)serialization. This definition aims to minimize the number of operations while ensuring that (de)serialization of a known LCS format cannot cause arbitrarily large stack allocations. As an example, if `v1` and `v2` are values of depth `n1` and `n2`, * a struct value `Foo { v1, v2 }` has depth `1 + max(n1, n2)`; * an enum value `E::Foo { v1, v2 }` has depth `1 + max(n1, n2)`; * a pair `(v1, v2)` has depth `max(n1, n2)`; * the value `Some(v1)` has depth `n1`. All string and integer values have depths `0`. #### Booleans and Integers |Type |Original data |Hex representation |Serialized format | |--- |--- |--- |--- | |Boolean |True / False |0x01 / 0x00 |[01] / [00] | |8-bit signed integer |-1 |0xFF |[FF] | |8-bit unsigned integer |1 |0x01 |[01] | |16-bit signed integer |-4660 |0xEDCC |[CCED] | |16-bit unsigned integer |4660 |0x1234 |[3412] | |32-bit signed integer |-305419896 |0xEDCBA988 |[88A9CBED] | |32-bit unsigned integer |305419896 |0x12345678 |[78563412] | |64-bit signed integer |-1311768467750121216 |0xEDCBA98754321100 |[0011325487A9CBED] | |64-bit unsigned integer |1311768467750121216 |0x12345678ABCDEF00 |[00EFCDAB78563412] | #### ULEB128-Encoded Integers The LCS format also uses the [ULEB128 encoding](https://en.wikipedia.org/wiki/LEB128) internally to represent unsigned 32-bit integers in two cases where small values are usually expected: (1) lengths of variable-length sequences and (2) tags of enum values (see the corresponding sections below). |Type |Original data |Hex representation |Serialized format | |--- |--- |--- |--- | |ULEB128-encoded u32-integer|2^0 = 1 |0x00000001 |[01] | | |2^7 = 128 |0x00000080 |[8001] | | |2^14 = 16384 |0x00004000 |[808001] | | |2^21 = 2097152 |0x00200000 |[80808001] | | |2^28 = 268435456 |0x10000000 |[8080808001] | | |9487 |0x0000250f |[8f4a] | In general, a ULEB128 encoding consists of a little-endian sequence of base-128 (7-bit) digits. Each digit is completed into a byte by setting the highest bit to 1, except for the last (highest-significance) digit whose highest bit is set to 0. In LCS, the result of decoding ULEB128 bytes is required to fit into a 32-bit unsigned integer and be in canonical form. For instance, the following values are rejected: * `[808080808001]` (2^36) is too large. * `[8080808010]` (2^33) is too large. * `[8000]` is not a minimal encoding of 0. #### Optional Data Optional or nullable data either exists in its full representation or does not. LCS represents this as a single byte representing the presence `0x01` or absence `0x00` of data. If the data is present then the serialized form of that data follows. For example: ```rust let some_data: Option = Some(8); assert_eq!(to_bytes(&some_data)?, vec![1, 8]); let no_data: Option = None; assert_eq!(to_bytes(&no_data)?, vec![0]); ``` #### Fixed and Variable Length Sequences Sequences can be made of up of any LCS supported types (even complex structures) but all elements in the sequence must be of the same type. If the length of a sequence is fixed and well known then LCS represents this as just the concatenation of the serialized form of each individual element in the sequence. If the length of the sequence can be variable, then the serialized sequence is length prefixed with a ULEB128-encoded unsigned integer indicating the number of elements in the sequence. All variable length sequences must be `MAX_SEQUENCE_LENGTH` elements long or less. ```rust let fixed: [u16; 3] = [1, 2, 3]; assert_eq!(to_bytes(&fixed)?, vec![1, 0, 2, 0, 3, 0]); let variable: Vec = vec![1, 2]; assert_eq!(to_bytes(&variable)?, vec![2, 1, 0, 2, 0]); let large_variable_length: Vec<()> = vec![(); 9_487]; assert_eq!(to_bytes(&large_variable_length)?, vec![0x8f, 0x4a]); ``` #### Strings Only valid UTF-8 Strings are supported. LCS serializes such strings as a variable length byte sequence, i.e. length prefixed with a ULEB128-encoded unsigned integer followed by the byte representation of the string. ```rust // Note that this string has 10 characters but has a byte length of 24 let utf8_str = "çå∞≠¢õß∂ƒ∫"; let expecting = vec![ 24, 0xc3, 0xa7, 0xc3, 0xa5, 0xe2, 0x88, 0x9e, 0xe2, 0x89, 0xa0, 0xc2, 0xa2, 0xc3, 0xb5, 0xc3, 0x9f, 0xe2, 0x88, 0x82, 0xc6, 0x92, 0xe2, 0x88, 0xab, ]; assert_eq!(to_bytes(&utf8_str)?, expecting); ``` #### Tuples Tuples are typed composition of objects: `(Type0, Type1)` Tuples are considered a fixed length sequence where each element in the sequence can be a different type supported by LCS. Each element of a tuple is serialized in the order it is defined within the tuple, i.e. [tuple.0, tuple.2]. ```rust let tuple = (-1i8, "libra"); let expecting = vec![0xFF, 5, b'l', b'i', b'b', b'r', b'a']; assert_eq!(to_bytes(&tuple)?, expecting); ``` #### Structures Structures are fixed length sequences consisting of fields with potentially different types. Each field within a struct is serialized in the order specified by the canonical structure definition. Structs can exist within other structs and as such, LCS recurses into each struct and serializes them in order. There are no labels in the serialized format, the struct ordering defines the organization within the serialization stream. ```rust struct MyStruct { boolean: bool, bytes: Vec, label: String, } struct Wrapper { inner: MyStruct, name: String, } let s = MyStruct { boolean: true, bytes: vec![0xC0, 0xDE], label: "a".to_owned(), }; let s_bytes = to_bytes(&s)?; let mut expecting = vec![1, 2, 0xC0, 0xDE, 1, b'a']; assert_eq!(s_bytes, expecting); let w = Wrapper { inner: s, name: "b".to_owned(), }; let w_bytes = to_bytes(&w)?; assert!(w_bytes.starts_with(&s_bytes)); expecting.append(&mut vec![1, b'b']); assert_eq!(w_bytes, expecting); ``` #### Externally Tagged Enumerations An enumeration is typically represented as a type that can take one of potentially many different variants. In LCS, each variant is mapped to a variant index, a ULEB128-encoded 32-bit unsigned integer, followed by serialized data if the type has an associated value. An associated type can be any LCS supported type. The variant index is determined based on the ordering of the variants in the canonical enum definition, where the first variant has an index of `0`, the second an index of `1`, etc. ```rust enum E { Variant0(u16), Variant1(u8), Variant2(String), } let v0 = E::Variant0(8000); let v1 = E::Variant1(255); let v2 = E::Variant2("e".to_owned()); assert_eq!(to_bytes(&v0)?, vec![0, 0x40, 0x1F]); assert_eq!(to_bytes(&v1)?, vec![1, 0xFF]); assert_eq!(to_bytes(&v2)?, vec![2, 1, b'e']); ``` If you need to serialize a C-style enum, you should use a primitive integer type. #### Maps (Key / Value Stores) Maps are represented as a variable-length, sorted sequence of (Key, Value) tuples. Keys must be unique and the tuples sorted by increasing lexicographical order on the LCS bytes of each key. The representation is otherwise similar to that of a variable-length sequence. In particular, it is preceded by the number of tuples, encoded in ULEB128. ```rust let mut map = HashMap::new(); map.insert(b'e', b'f'); map.insert(b'a', b'b'); map.insert(b'c', b'd'); let expecting = vec![(b'a', b'b'), (b'c', b'd'), (b'e', b'f')]; assert_eq!(to_bytes(&map)?, to_bytes(&expecting)?); ``` ### Backwards compatibility Complex types dependent upon the specification in which they are used. LCS does not provide direct provisions for versioning or backwards / forwards compatibility. A change in an objects structure could prevent historical clients from understanding new clients and vice-versa. ## Contributing See the [CONTRIBUTING](CONTRIBUTING.md) file for how to help out. ## License This project is available under the terms of either the [Apache 2.0 license](LICENSE).