Permalink
ed32d61 Jul 1, 2018
1 contributor

Users who have contributed to this file

1323 lines (1042 sloc) 54.9 KB

FlatBuffers Binary Format

Overview

Memory Blocks

A FlatBuffers layout consists of the following types of memory blocks:

  • header
  • table
  • vector
  • string
  • vtable
  • (structs)

Each of these have a contigous memory layout. The header references the root table. Every table references a vtable and stores field in a single block of memory. Some of these fields can be offsets to vectors, strings and vectors. Vectors are one-dimensional blocks where each element is self-contained or stores reference to table or a string based on schema type information. A vtable decides which fields are present in table and where they are stored. Vtables are the key to support schema evolution. A table has no special rules about field ordering except fields must be properly aligned, and if they are ordered by size the will pack more densely but possibly more complicated to construct and the schema may provide guidelines on the preferred order. Two vtables might mean different things but accidentally have the same structure. Such vtables can be shared between different tables. vtable sharing is important when vectors store many similar tables. Structs a dense memory regions of scalar fields and smaller structs. They are mostly found embedded in tables but they are independent blocks when referenced from a union or union vector, or when used as a buffer root. Structs hold no references.

Space between the above blocks are zero padded and present in order to ensure proper alignment. Structs must be packed as densely as possible according the alignment rules that apply - this ensures that all implementations will agree on the layout. The blocks must not overlap in memory but two blocks may be shared if they represent the same data such as sharing a string.

FlatBuffers are constructed back to front such that lower level objects such as sub-tables and strings are created first in stored last and such that the root object is stored last and early in the buffer. See also (Stream Buffers](#stream-buffers) for a proposed variation over this theme.

All addressing in FlatBuffers are relative. The reason for this? When navigating the buffer you don't need to store the buffer start or the buffer length, you can also find a new reference relative the reference you already have. vtables require the table location to find a field, but a table field referencing a string field only needs the table field location, not the table start in order to find the referenced string. This results in efficient navigation.

Example

Uoffsets add their content to their address and are positive while a tables offset to its vtable is signed and is substracted. A vtables element is added to the start of the table referring to the vtable.

Schema (eclectic.fbs) :

    namespace Eclectic;

    enum Fruit : byte { Banana = -1, Orange = 42 }
    table FooBar {
        meal      : Fruit = Banana;
        density   : long (deprecated);
        say       : string;
        height    : short;
    }
    file_identifier "NOOB";
    root_type FooBar;

JSON :

    { "meal": "Orange", "say": "hello", "height": -8000 }

Buffer :

    header:

        +0x0000 00 01 00 00 ; find root table at offset +0x0000100.
        +0x0004 'N', 'O', 'O', 'B' ; possibly our file identifier

        ...

    table:

        +0x0100 e0 ff ff ff ; 32-bit soffset to vtable location
                            ; two's complement: 2^32 - 0xffffffe0 = -0x20
                            ; effective address: +0x0100 - (-0x20) = +0x0120
        +0x0104 00 01 00 00 ; 32-bit uoffset string field (FooBar.say)
                            ; find string +0x100 = 256 bytes _from_ here
                            ; = +0x0104 + 0x100 = +0x0204.
        +0x0108 42d         ; 8-bit (FooBar.meal)
        +0x0109 0           ; 8-bit padding
        +0x010a -8000d      ; 16-bit (FooBar.height)
        +0x010c  ...        ; (first byte after table end)

        ...

    vtable:

        +0x0120 0c 00       ; vtable length = 12 bytes
        +0x0122 0c 00       ; table length = 12 bytes
        +0x0124 08 00       ; field id 0: +0x08 (meal)
        +0x0126 00 00       ; field id 1: <missing> (density)
        +0x0128 04 00       ; field id 2: +0004 (say)
        +0x012a 0a 00       ; field id 3: +0x0a (height)

        ...

    string:

        +0x0204 05 00 00 00 ; vector element count (5 ubyte elements)
        +0x0208 'h' 'e'     ; vector data
        +0x020a 'l' 'l'     ; vector data
        +0x020c 'o'         ; vector data
        +0x020d  00         ; zero termination
                            ; special case for string vectors

        ...

Actual FlatCC generated buffer :

Eclectic.FooBar:
  0000  08 00 00 00 4e 4f 4f 42  e8 ff ff ff 08 00 00 00  ....NOOB........
  0010  2a 00 c0 e0 05 00 00 00  68 65 6c 6c 6f 00 00 00  *.......hello...
  0020  0c 00 0c 00 08 00 00 00  04 00 0a 00              ............

Note that FlatCC often places vtables last resulting in e0 ff ff ff style vtable offsets, while Googles flatc builder typically places them before the table resulting in 20 00 00 00 style vtable offsets which might help understand why the soffset is subtracted from and not added to the table start. Both forms are equally valid.

Primitives

Primitives are data used to build more complex structures such as strings, vectors, vtables and tables. Although structs are not strictly a primitive, it helps to view them as self-contained primitives.

Numerics

FlatBuffers are based on the following primitives that are 8, 16, 32 and 64 bits in size respectively (IEEE-754, two's complement, little endian):

uint8, uint16, uint32, uint64 (unsigned)
int8, int16, int32, int64     (two's complement)
float32, float64              (IEEE-754)

Boolean

flatbuffers_bool (uint8)
flatbuffers_true (flatbuffers_bool assign as 1, read as != 0)
flatbuffers_false (flatbuffers_bool = 0)

Note that a C99 bool type has no defined size or sign so it is not an exact representation of a flatbuffers boolean encoding.

When a stored value is interpreted as boolean it should not be assumed to be either 1 or 0 but rather as not equal to 0. When storing a boolean value or when converting a boolean value to integer before storing, the value should be 1 for true and 0 for false, In C this can be done using !!x.

Format Internal Types

flatbuffers_union_type_t (uint8, NONE = 0, 0 <= type <= 255)
flatbuffers_identifier_t (uint8[4])
flatbuffers_uoffset_t (uin32)
flatbuffers_soffset_t (int32),
flatbuffers_voffset_t (uint16)

These types can change in FlatBuffer derived formats, but when we say FlatBuffers without further annotations, we mean the above sizes in little endian encoding. We always refer to specific field type and not a specific field size because this allows us to derive new formats easily.

Scalars

To simpify discussion we use the term scalars to mean integers, boolean, floating point and enumerations. Enumerations always have an underlying signed or unsigned integer type used for its representation.

Scalars are primitives that has a size between 1 and 8 bytes. Scalars are stored aligned to the own size.

The format generally supports vectors and structs of any scalar type but some language bindings might not have support for all combinations such as arrays of booleans.

An special case is the encoding of a unions type code which internally is an enumeration but it is not normally permitted in places where we otherwise allow for scalars.

Another special case is enumerations of type boolean which may not be widely supported, but possible. The binary format is not concerned with this distinction because a boolean is just an integer at this level.

Structs

A struct is a fixed sized block of a fixed number of fields in a specific order defined by a schema. A field is either a scalar or another struct. A struct cannot contain fields that contain itself directly or indirectly. A struct is self-contained and has no references. A struct can be empty (note that this does not map cleanly to C structs).

A schema cannot change the layout of a struct without breaking binary compatibility, unlike tables.

When used as a table field, a struct is embedded within the table block unless it is union value. A vector of structs are placed in a separate memory block, similar to vectors of scalars. A vector of unions that has a struct as member will reference the struct as an offset, and the struct is then an independent memory block like a table.

FlatCC supports that a struct can be the root object of a FlatBuffer, but most implementations likely won't support this. Structs as root are very resource efficient.

Internals

All content is little endian and offsets are 4 bytes (uoffset_t). A new buffer location is found by adding a uoffset to the location where the offset is stored. The first location (offset 0) points to the root table. uoffset_t based references are to tables, vectors, and strings. References to vtables and to table fields within a table have a different calculations as discussed below.

A late addition to the format allow for adding a size prefix before the standard header. When this is done, the builder must know about so it can align content according to the changed starting position. Receivers must also know about the size field just as they must know about the excpected buffer type.

The next 4 bytes (sizeof(uoffset_t)) might represent a 4 byte buffer identifier, or it might be absent but there is no obvious way to know which. The file identifier is typically ASCII characters from the schemas file_identifier field padded with 0 bytes but may also contain any custom binary identifier in little endian encoding. See Type-Hashes. The 0 identifier should be avoided because the buffer might accidentally contain zero padding when an identifier is absent and because 0 can be used by API's to speficy that no identifier should be stored.

When reading a buffer, it should be checked that the length is at least 8 bytes (2 * sizeof(uoffset_t)) Otherwise it is not safe to check the file identifier.

The root table starts with a 4 byte vtable offset (soffset_t). The soffset_t has the same size as uoffset_t but is signed.

The vtable is found by subtracting the signed 4 byte offset to the location where the vtable offset is stored. Note that the FlatCC builder typically stores vtables at the end of the buffer (clustered vtables) and therefore the vtable offset is normally negative. Other builders often store the vtable before the table unless reusing an existing vtable and this makes the soffset positive. (Nested FlatBuffers will not store vtables at the end because it would break compartmentalization).

The vtable is a table of 2 byte offsets (voffset_t). The first two entreis are the vtable size in bytes and the table size in bytes. The next offset is the vtable entry for table field 0. A vtable will always have as many entries as the largest stored field in the table, but it might skip entries that are not needed or known (version independence) - therefore the vtable length must be checked. The table length is only needed by verifiers. A vtable lookup of an out of range table id is the same as a vtable entry that has a zero entry and results in a default value for the expected type. Multiple tables may shared the same vtable, even if they have different types. This is done via deduplication during buffer construction. A table field id maps to a vtable offset using the formula vtable-start + sizeof(voffset_t) * (field-id + 2), iff the result is within the vtable size.

The vtable entry stores a byte offset relative to the location where the soffset_t where stored at the start of the table. A table field is stored immediately after the soffset_t or may contain 0 padding. A table field is always aligned to at least its own size, or more if the schema demands it. Strings, sub-tables and vectors are stored as a 4-byte uoffset_t. Such sub-elements are always found later in the buffer because uoffset_t is unsigned. A struct is stored in-place. Conceptually a struct is not different from an integer in this respect, just larger.

If a sub-table or vector is absent and not just without fields, 0 length, the containing table field pointing the sub-table or vector should be absent. Note that structs cannot contain sub-tables or vectors.

A struct is always 0 padded up to its alignment. A structs alignment is given as the largest size of any non-struct member, or the alignment of a struct member (but not necessarily the size of a struct member), or the schema specified aligment.

A structs members are stored in-place so the struct fields are known via the the schema, not via information in the binary format

An enum is represent as its underlying integer type and can be a member of structs, fields and vectors just like integer and floating point scalars.

A table is always aligned to sizeof(uoffset_t), but may contain internal fields with larger alignment. That is, the table start or end are not affected by alignment requirements of field members, unlike structs.

A string is a special case of a vector with the extra requirement that a 0 byte must follow the counted content, so the following also applies to strings.

A vector starts with a uoffset_t field the gives the length in element counts, not in bytes. Table fields points to the length field of a vector, not the first element. A vector may have 0 length. Note that the FlatCC C binding will return a pointer to a vectors first element which is different from the buffer internal reference.

All elements of a vector has the same type which is either a scalar type including enums, a struct, or a uoffset reference where all the reference type is to tables all of the same type, all strings, or, for union vectors, references to members of the same union. Because a union member can be a struct, it is possible to have vectors that reference structs instead of embeddding them, but only via unions. It is not possible to have vectors of vectors other than string vectors, except indirectly via vectors containing tables.

A vectors first element is aligned to the size of uoffset_t or the alignment of its element type, or the alignment required by the schema, whichever is larger. Note that the vector itself might be aligned to 4 bytes while the first element is aligned to 8 bytes.

(Currently the schema semantics do no support aligning vectors beyond their element size, but that might change and can be forced when building buffers via dedicated api calls).

Strings are stored as vectors of type ubyte_t, i.e. 8-bit elements. In addition, a trailing zero is always stored. The trailing zero is not counted in the length field of the vector. Technically a string is supposed to be valid UTF-8, but in praxis it is permitted that strings contain 0 bytes or other invalid sequences. It is recommended to explicitly check strings for UTF-8 conformance when this is required rather than to expect this to alwways be true. However, the ubyte vector should be preferred for binary data.

A string, vector or table may be referenced by other tables and vectors. This is known as a directed acyclic graph (DAG). Because uoffsets are unsigned and because uoffsets are never stored with a zero value (except for null entries in union vectors), it is not possible for a buffer to contain cycles which makes them safe to traverse with too much fear of excessive recursion. This makes it possible to efficiently verify that a buffer does not contain references or content outside of its expected boundaries.

A vector can also hold unions, but it is not supported by all implementations. A union vector is in reality two separate vectors: a type vector and an offset vector in place of a single unions type and value fields in table. See unions.

Type Hashes

A type hash is a 32-bit value defined as the fnv1a32 hash of a tables or a structs fully qualified name. If the fnv1a32 hash returns 0 it should instead hash the empty string. 0 is used to indicate that a buffer should not store an identifier.

Every table and struct has a name and optionally also a namespace of one or more levels. The fully qualified name is optional namespace followed by the type name using '.' as a separator. For example "MyGame.Sample.Monster", or "Eclectic.FooBar".

The type hash can be used as the buffer identifier instead of the schema provided file_identifier. The type hash makes it possible to distinguish between different root types from the same schema, and even across schema as long as the namespace is unique.

Type hashes introduce no changes to the binary format but the application interface must choose to support user defined identifiers or explicitly support type hashes. Alternatively an application can peak directly into the buffer at offset 4 (when uoffset_t is 4 bytes long).

FlatCC generates the following identifier for the "MyGame.Sample.Monster" table:

#define MyGame_Sample_Monster_type_hash ((flatbuffers_thash_t)0xd5be61b)
#define MyGame_Sample_Monster_type_identifier "\x1b\xe6\x5b\x0d"

But we can also compute one online for our example buffer:

    fnv1a32("Eclectic.FooBar") = 0a604f58

Thus we can open a hex editor and locate

        +0x0000 00 01 00 00 ; find root table at offset +0x0000100.
        +0x0004 'N', 'O', 'O', 'B' ; possibly our file identifier

and replace it with

        +0x0000 00 01 00 00 ; find root table at offset +0x0000100.
        +0x0004 58 4f 60 0a ; very likely our file identifier identifier

or generate it with flatcc:

    $ bin/flatcc --stdout doc/eclectic.fbs | grep FooBar_type
    #define Eclectic_FooBar_type_hash ((flatbuffers_thash_t)0xa604f58)
    #define Eclectic_FooBar_type_identifier "\x58\x4f\x60\x0a"

The following snippet implements fnv1a32, and returns the empty string hash if the hash accidentially should return 0:

static inline flatbuffers_thash_t flatbuffers_type_hash_from_name(const char *name)
{
    uint32_t hash = 2166136261UL;
    while (*name) {
        hash ^= (uint32_t)*name;
        hash = hash * 16777619UL;
        ++name;
    }
    if (hash == 0) {
        hash = 2166136261UL;
    }
    return hash;
}

Conflicts

It is possible to have conficts between two type hashes although the risk is small. Conflicts are not important as long as an application can distinguish between all the types it may encouter in actual use. A switch statement in C will error at compile time for two cases that have the same value, so the problem is easily detectable and fixable by modifying a name or a namespace.

For global conflict resolution, a type should be identified by its fully qualified name using adequate namespaces. This obviously requires it to be stored separate from the buffer identifier due to size constraints.

Type Hash Variants

If an alternative buffer format is used, the type hash should be modified. For example, if uoffset_t is defined as a 64-bit value, the fnv1a64 hash should be used instead. For big endian variants the hash remains unchanged but is byteswapped. The application will use the same id while the acces layer will handle the translation.

For buffers using structs as roots, the hash remains unchanged because the struct is a unique type in schema. In this way a receiver that does not handle struct roots can avoid trying to read the root as a table.

For futher variations of the format, a format identifier can be inserted in front of the namespace when generating the hash. There is no formal approach to this, but as an example, lets say we want to use only 1 byte per vtable entry and identify these buffers with type hash using the prefix "ebt:" for example buffer type. We then have the type hash:

#define type_hash_prefix "ebt:"

hash = fnv1a32(type_hash_prefix "Eclectic.FooBar");
hash = hash ? hash : fnv1a32(type_hash_prefix);

If the hash returns 0 we hash the prefix.

Unions

A union is a contruction on top of the above primitives. It consists of a type and a value.

In the schema a union type is a set of table types with each table name assigned a type enumeration starting from 1. 0 is the type NONE meaning the union has not value assigned. The union type is represented as a ubyte enum type, or in the binary format as a value of type union_type_t which for standard FlatBuffers is an 8-bit unsigned code with 0 indicating the union stores not value and a non-zero value indicating the type of the stored union.

A union is stored in a table as a normal sub-table reference with the only difference being that the offset does not always point to a table of the same type. The 8-bit union type is stored as a separate table field conventially named the same as the union value field except for a _type suffix. The value (storing the table offset) MUST have a field ID that is exactly one larger than the type field. If value field is present the type field MUST also be present. If the type is NONE the value field MUST be absent and the type field MAY be absent because a union type always defaults to the value NONE.

Vectors of unions is a late addition (mid 2017) to the FlatBuffers format. FlatCC supports union vectors as of v0.5.0.

Vectors of unions have the same two fields as normal unions but they both store a vector and both vectors MUST have the same length or both be absent from the table. The type vector is a vector of 8-bit enums and the value vector is a vector of table offsets. Obviously each type vector element represents the type of the table in the corresponding value element. If an element is of type NONE the value offset must be stored as 0 which is a circular reference. This is the only offset that can have the value 0.

A later addition (mid 2017) to the format allows for structs and strings to also be member of a union. A union value is always an offset to an independent memory block. For strings this is just the offset to the string. For tables it is the offset to the table, naturally, and for structs, it is an offset to an separate aligned memory block that holds a struct and not an offset to memory inside any other table or struct. FlatCC supports mixed type unions and vectors of these as of v0.5.0.

Alignment

All alignments are powers of two between 1 and 256. Large alignments are only possible via schema specified alignments. The format naturally has a maximum alignment of the largest scalar it stores, which is a 64-bit integer or floating point value. Because C malloc typically returns buffers aligned to a least 8 bytes, it is often safe to place buffers in heap allocated memory, but if the target system does not permit unaligned access, or is slow on unaligned access, a buffer should be placed in sufficiently aligned memory. Typically it is a good idea to place buffer in cacheline aligned boundary anyway.

A buffers alignment is the same as the largest alignment of any object or struct it contains.

The buffer size is not guaranteed to be aligned to its own alignment unlike struct. Googles flatc builder does this, at least when size prefixed. The FlatCC tool currently does not, but it might later add an option to pad up to alignment. This would make it simpler to stack similar typed buffers in file - but users can retrieve the buffers alignment and do this manually. Thus, when stacking size prefixed buffers, each buffer should start aligned to its own size starting at the size field, and should also be zero padded up to its own alignment.

Default Values and Deprecated Values

A table can can omit storing any field that is not a required field. For strings, vectors and tables this result in returning a null value different from an empty value when reading the buffer. Struct fields not present in table are also returned as null.

All fields that can return null do not have a default value. Other values, which are integers, floats and enumerations, can all have default values. The default value is returned if not found in the table.

If a default value is not specified the default defaults to zero for the corresponding type. If an enumeration does not have a 0 value and no explicit default value, this is a schema error.

When building a buffer the builder will compare a stored field to the known default value. If it matches the field will simple be skipped. Some builder API's makes it possible to force a default value to be stored and to check if a field is missing when reading the buffer. This can be used to handle NULL values differently from default or missing values.

A deprecated field is a schema construct. The binary format either stores a table field, or it does not.

A deprecated field should be treated as not available, as in no way to read the value as opposed to returning a default value regardless of whether the field is present or not. If they for some reason are made accessible the verifier must also understand and verify these fields.

A deprecated field also cannot be written to a new buffer, although if it against guidelines remains possible to do so, it should be done as if the field was not deprecated.

Structs cannot have default values and cannot have deprecated fields in stadard FlatBuffers. FlatCC supports marking a struct field as deprecated. This implies the field will always be zeroed and with no trivial accessors. A struct can never change size without breaking support for schema evolution.

FlatCC JSON parsers allow structs to only set some values. Remaining values will be implicitly zeroed. The C API for writing buffers do not have this concern because a struct can just be written as a C struct so there is no control over which fields a user choose to set or zero. However, structs should be zeroed and padded where data is not otherwise set. This makes it possible to hash and integrity check structs (though this is not an integral part of the format).

Schema Evolution

A table has a known set of low-valued field identifiers. Any unused field id can be used in a future version. If a field (as is normal) is implicitly assigned an id, new fields can only be added at the end of the table. Internally this translates into new versions getting ever larger vtables. Note that vtables are only stored as large as needed for the actual content of table, so a rarely used new field will not cause vtables to grew when unused.

Enumarations may not change values in future versions. Unions may only added new table names to the end of the union list.

Structs cannot change size nor content. They cannot evolve. FlatCC permits deprecating fields which means old fields will be zeroed.

Names can be changed to the extend it make sense to the applications already written around the schema, but it may still break applications relying on some reflective information. For example, a reader may provide the string representation of a numeric enum value.

New types can be added. For example adding a new table is always safe as long as it does not conflict with any existing schemas using the same namespace.

Required fields cannot stop being required and they cannot be deprecated.

Various attributes and other changes form a gray area that will not make the binary format unsafe but may still break due to changes in code generation, serialization to JSON, or similar. For example, a generated constructor that creates a table from a list of positional arguments might break if the field order changes or grows or have fields deprecated. JSON parsers could cease to work on old formats if base64 serialization is added subsequently, and so on.

Keys and Sorting

Keys and sorting is a meta construct driven by the schema. The binary format has no special concept of keys and sorting and a vector can be sorted by one of several keys so it makes no sense to enforce a specific order.

The basic FlatBuffers format only permit at most one key and generally sorts vectors by that key during buffer construction. FlatCC does not do this both because sorting is not practical while building the buffer and because FlatCC supports sorting by one of several keys. Thus, in general it is not safe to assume that a vector is sorted, but it can be sorted if needed.

Size Limits

A buffer should never be larger than 2^(sizeof(soffset_t) * 8 - 1) - 1 or 2^31 - 1 i.e. 2GB for standard FlatBuffers. Beyond this safe it is not safe to represent vtable offsets and implementations can no longer use signed types to store uoffset_t values. This limit also ensures that all vectors can be represented safely with a signed 32-bit length type.

The application interface is likely to use a native type for representing sizes and vector indices. If this type is smaller that sizeof(soffset_t) or equivalently sizeof(uoffset_t) there is a risk of overflow. The simplest way to avoid any issues is to limit the accepted buffer size of the native size type. For example, on some embedded microprocessor systems a C compiler might have a 16-bit int and size_t type, even if it supports uint32_t as well. In this the safe assumption is to limit buffers to 2^15 - 1 which is very likely more than sufficient on such systems.

A builder API might also want to prevent vectors from being created when they cannot stay within the size type of the platform when the element size is multipled by the element count. This is deceiving because the element count might easily be within range. Such issues will be rare in praxis but they can be the source of magnificent vulnerabilites if not handled appropriately.

Verification

Verification as discussed here is not just about implementing a verifier. It is as much requirements that any builder must fulfill.

The objective of verification is to make it safe to read from an untrusted buffer, or a trusted buffer that accidentally has an unexpected type. The verifier does not guarantee that the type is indeed the expeced type exact it can read the buffer identifier if present which is still no guarantee. The verifier cannot make it safe to modify buffers in-place because the cost of doing such a check would be prohitive in the average case.

A buffer verifier is expected to verify that all objects (strings, vectors and tables) do not have an end position beyond the externally specified buffer size and that all offset are aligned relative to offset zero, and sometimes also relative to actually specified buffer (but sometimes it is desireable to verify buffers that are not stored aligned, such as in network buffers).

A verifier primarily checks that:

  • the buffer is at least 2 * sizeof(uoffset_t) such that the root root table and buffer identifier can be checked safely.
  • any offset being chased is inside the buffer that any data accessed from that resuling location is entirely inside the buffer and aligned notably the vtables first entry must be valid before the vtable can safely validate itself, but this also applies to table, string and vector fields.
  • any uoffset has size of at least sizeof(uoffse_t) to aviod self-reference and is no larger than the largest positive soffset_t value of same size - this ensures that implementations can safely add uoffset_t even if converting to signed first. It also, incidentally, ensures compatibility with StreamBuffers - see below.
  • vtable size is aligned and does not end outside buffer.
  • vtable size is at least the two header fields (2 *sizeof(voffset_t)`).
  • required table fields are present.
  • recursively verify all known fields and ignore other fields. Unknown fields are vtable entries after the largest known field ID of a table. These should be ignored in order to support forward versioning.
  • deprecated fields are valid if accessors are available to do so, or ignore if the there is no way to access the field by application code.
  • vectors end within the buffer.
  • strings end within the buffer and has a zero byte after the end which is also within the buffer.
  • table fields are aligned relative to buffer start - both structs, scalars, and offset types.
  • table field size is aligned relative to field start.
  • any table field does not end outside the tables size as given by the vtable.
  • table end (without chasing offsets) is not outside buffer.
  • all data referenced by offsets are also valid within the buffer according the type given by the schema.
  • verify that recursion depth is limited to a configurable acceptable level for the target system both to protect itself and such that general recursive buffer operations need not be concerned with stack overflow checks (a depth of 100 or so would normally do).
  • verify that if the union type is NONE the value (offset) field is absent and if it is not NONE that the value field is present. If the union type is known, the table should be verified. If the type is not known the table field should be ignored. A reader using the same schema would see the union as NONE. An unknown union is not an error in order to support forward versioning.
  • verifies the union value according to the type just like any other field or element.
  • verify that a union vector always has type vector if the offset vector is present and vice versa.

A verifier may choose to reject unknown fields and union types, but this should only be an user selectable option, otherwise schema evolution will not be possible.

A verifier needs to be very careful in how it deals with overflow and signs. Vector elements multiplied by element size can overflow. Adding an invalid offset might become valid due to overflow. In C math on unsigned types yield predictable two's complement overflow while signed overflow is undefined behavior and can and will result in unpredictable values with modern optimizing compilers. The upside is that if the verifier handles all this correctly, the application reader logic can be much simpler while staying safe.

A verifier does not enforce that:

  • structs and other table fields are aligned relative to table start because tables are only aligned to their soffset field. This means a table cannot be copied naively into a new buffer even if it has no offset fields.
  • the order of individual fields within a table. Even if a schema says something about ordering this should be considered advisory. A verifier may additionally check for ordering for specific applications. Table order does not affect safety nor general buffer expectations. Ordering might affect size and performance.
  • sorting as specified by schema attributes. A table might be sorted in different ways and an implementation might avoid sorting for performance reasons and other practical reasons. A verifier may, however, offer additional verification to ensure specific vectors or all vectors are sorted according to schema or other guidelines. Lack of sorting might affect expected behavior but will not make the buffer unsafe to access.
  • that structures do not overlap. Overlap can result in vulnerabilities if a buffer is modified, and no sane builder should create overlaps other than proper DAGs except via a separate compression/decopression stage of no interest to the verifier.
  • strings are UTF-8 compliant. Lack of compliance may affect expections or may make strings appear shorter or garbled. Worst case a naive UTF-8 reader might reach beyond the end when observing a lead byte suggest data after buffer end, but such a read should discover the terminal 0 before things get out of hand. Being relaxed permits specific use cases where UTF-8 is not suitable without giving up the option to verify buffers. Implementations can add additional verification for specific usecases for example by providing a strict-UTF8 flag to a verifier or by verifying at the application level. This also avoids unnecessary duplicate validation, for example when an API first verifies the buffer then converts strings to an internal heap representation where UTF-8 is validated anyway.
  • default values are not stored. It is common to force default values to be stored. This may be used to implement NULL values as missing values different from default values.
  • enum values are only among the enum values of its type. There are many use cases where it is convenient to add enum values for example flags or enums as units e.g. `2 * kilo + 3 * ounce. More generally ordinary integers may have value range restrictions which is also out of scope for verifier. An application may provide additional verification when requested.

More generally we can say that a verifier is a basic fast assurance that the buffer is safe to access. Any additional verification is application specific. The verifier makes it safe to apply secondary validation. Seconary validation could be automated via schema attributes and may be very useful as such, but it is a separate problem and out of scope for a core binary format verifier.

A buffer identifier is optional so the verifier should be informed whether an identifier must match a given id. It should check both ASCII text and zero padding not just a string compare. It is non-trivial to decide if the second buffer field is an identifier, or some other data, but if the field does not match the expected identifier, it certainly isn't what is expected.

Note that it is not entirely trivial to check vector lengths because the element size must be mulplied by the stored element count. For large elements this can lead to overflows.

Risks

Because a buffer can contain DAGs constructed to explode exponentiall, it can be dangerous to print JSON or otherwise copy content blindly if there is no upper limit on the export size.

In-place modification cannot be trusted because a standard buffer verifier will detect safe read, but will not detect if two objects are overlapping. For example, a table could be stored inside another table. Modifing one table might cause access to the contained table to go out of bounds, for example by directing the vtable elsewhere.

The platform native integer and size type might not be able to handle large FlatBuffers - see Size Limits.

Becaue FlatCC requires buffers to be sorted after builiding, there is risk due to buffer modifications. It is not sufficient to verify buffers after sorting because sorting is done inline. Therefore buffers must be trusted or rewritten before sorting.

Nested FlatBuffers

FlatBuffers can be nested inside other FlatBuffers. In concept this is very simple: a nested buffer is just a chunk of binary data stored in a ubyte vector, typically with some convenience methods generated to access a stored buffer. In praxis it adds a lot of complexity. Either a nested buffer must be created strictly separately and copied in as binary data, but often users mix the two builder contexts accidentally storing strings from one buffer inside another. And when done right, the containing ubyte vector might not be aligned appropriately making it invalid to access the buffer without first copying out of the containing buffer except where unaligned access is permitted. Further, a nested FlatBuffer necessarily has a length prefix because any ubyte vector has a length field at its start. Therefore, size prefixed flatbuffers should not normally be stored as nested buffers, but sometimes it is necessary in order have the buffer properly aligned after extraction.

The FlatCC builder makes it possible to build nested FlatBuffers while the containing table of the parent buffer is still open. It is very careful to ensure alignment and to ensure that vtables are not shared between the two (or more) buffers, otherwise a buffer could not safely be copied out. Users can still make mistakes by storing references from one buffer in another.

Still, this area is so complex that several bugs have been found. Thus, it is advice to use nested FlatBuffers with some care.

On the other hand, nested FlatBuffers make it possible to trivially extract parts of FlatBuffer data. Extracting a table would require chasing pointers and could potentially explode due to shared sub-graphs, if not handled carefully.

Big Endian FlatBuffers

FlatBuffers are formally always little endian and even on big-endian platforms they are reasonably efficient to access.

However it is possible to compile FlatBuffers with native big endian suppport using the FlatCC tool. The procedure is out of scope for this text, but the binary format is discussed here:

All fields have exactly the same type and size as the little endian format but all scalars including floating point values are stored byteswapped within their field size. Offset types are also byteswapped.

The buffer identifier is stored byte swapped if present. For example the 4-byte "MONS" identifier becomes "SNOM" in big endian format. It is therefore reasonably easy to avoid accidentially mixing little- and big-endian buffers. However, it is not trivial to handle identifers that are not exactly 4-bytes. "HI\0\0" could be "IH\0\0" or "\0\0IH". It is recommended to always use 4-byte identifies to avoid this problem. See the FlatCC release 0.4.0 big endian release for details.

When using type hash identifiers the identifier is stored as a big endian encoded hash value. The user application will the hash in its native form and accessor code will do the conversion as for other values.

StructBuffers

Unlike tables, structs are are usually embedded in in a fixed memory block representing a table, in a vector, or embedded inline in other structs, but can also be independent when used in a union.

The root object in FlatBuffers is conventionally expected to be a table, but it can also be struct. FlatCC supports StructBuffers. Since structs do not contain references, such buffers are truly flat. Most implementations are not likely to support structs are root but even so they are very useful:

It is possible to create very compact and very fast buffers this way. They can be used where one would otherwise consider using manual structs or memory blocks but with the advantage of a system and language independent schema.

StructBuffers may be particularly interesting for the Big Endian variant of FlatBuffers for two reasons: the first being that performance likely matters when using such buffers and thus Big Endian platforms might want them. The second reason is the network byte order is traditionally big endian, and this has proven very difficult to change, even in new evolving IETF standards. StructBuffers can be used to manage non-trival big endian encoded structs, especially structs containing other structs, even when the receiver does not understand FlatBuffers as concept since the header can just be dropped or trivially documented.

StreamBuffers

StreamBuffers are so far only a concept although some implementations may already be able to read them. The format is documented to aid possible future implementations.

StreamBuffers makes it possible to store partially completed buffers for example by writing directly to disk or by sending partial buffer data over a network. Traditional FlatBuffers require an extra copying step to make this possible, and if the writes are partial, each section written must also store the segment length to support reassembly. StreamBuffers avoid this problem.

StreamBuffers treat uoffset_t the same as soffset_t when the special limitation that uoffset_t is always negative when viewed as two's complement values.

The implication is that everthing a table or vector references must be stored earlier in the buffer, except vtables that can be stored freely. Existing reader implementations that treat uoffset_t as signed, such as Javascript, will be able to read StreamBuffers with no modification. Other readers can easily be modified by casting the uoffset value to signed before it.

Verifiers must ensure that any buffer verified always stores all uoffset_t with the same sign. This ensures the DAG structure is preserved without cycles.

The root table offset remains positive as an exception and points to the root buffer. A stream buffer root table will be the last table in the buffer.

The root table offset may be replaced with a root table offset at the end of the buffer instead of the start. An implementation may also zero the initial offset and update it later. In either case the buffer should be aligned accordingly.

Some may prefer the traditional FlatBuffer approach because the root table is stored and it is somehow easier or faster to access, but modern CPU cache systems do not care much about the order of access as long as as their is some aspect of locality of reference and the same applies to disk access while network access likely will have the end of the buffer in hot cache as this is the last sent. The average distance between objects will be the same for both FlatBuffers and StreamBuffers.

Bidirectional Buffers

Bidirectional Buffers is a generalization of StreamBuffers and FlatBuffers and so far only exists as an idea.

FlatBuffers and StreamBuffers only allow one direction because it guarantees easy cycle rejection. Cycles are unwanted because it is expensive to verify buffers with cycles and because recursive navigation might never terminate.

As it happens, but we can also easily reject cycles in bidirectional buffers if we apply certain constraints that are fully backwards compatible with existing FlatBuffers that have a size below 2GB.

The rule is that when an offset changes direction relative to a parent objects direction, the object creating the change of direction becomes a new start or end of buffer for anything reachable via that offset.

The root table R is reached via forward reference from the buffer header. Its boundaries are itself and the buffer end. If the this table has an offset pointing back, for example to new table X, then table X must see the buffer start and R as two boundaries. X must not directly or indirectly reach outside this region. Likewise, if the table R points to new table Y in the forward direction, then Y is bounded by itself and the buffer end.

A lower bound is the end of a table or a vector althoug we just say the table or vector is a boundary. An upper bound is until the start of the table or vector, or the end of the buffer.

When a buffer is verified, the boundaries are initially the buffer start and buffer end. When the references from the root table are followed, there new boundaries are either buffer start to root table, or root table to end, depending on the offsets direction. And so forth recursively.

We can relax the requirement that simple vectors and vtables cannot cross these boundaries, except for the hard buffer limits, but it does compicate things and is likely not necessary.

Nested FlatBuffers already follow similar boundary rules.

The point of Bidirectional buffers is not that it would be interesting to store things in both directions, but to provide a single coherent buffer model for both ways of storing buffers without making a special case. This will allow StreamBuffers to co-exist with FlatBuffers without any change, and it will allow procedures to build larger buffers to make their own changes on how to build their subsection of the buffer.

We still need to make allowance for keeping the buffer headers root pointer implicit by context, at least as an option. Otherwise it is not, in general, possible to start streaming buffer content before the entire buffer is written.

Possible Future Features

This is highly speculative, but documents some ideas that have been floating in order to avoid incompatible custom extensions on the same theme. Still these might never be implemented or end up being implemented differently. Be warned.

Force Align

force_align attribute supported on fields of structs, scalars and and vectors of fixed size elements.

Fixed Length Vectors

A fixed length vector is a vector of fixed length containing inline fixed size elements and no size prefix and they are stored inline in tables and structs. An element can be a struct or a scalar including enums.

The binary format of a fixed size vector of length n and type t can be precisely emulated by created a struct that holds exactly n fields of type t, n >= 0. This means that a fixed length vector does not store any length information in a header and that it is stored inline in tables similar to structs. Alignment follows the rules of structs and also supports the force_align attribute.

It is not valid to ever change the size of a fixed length vector when staying compliant with schema evolution.

A fixed length vector cannot hold offsets because fixed length vectors are essentially structs with alternative accessors.

Unions of fixed length vectors are not valid unless wrapped in a struct.

Careful: subject to likely change: A new type char is used to create fixed length utf8 or ASCII strings that are always zero terminated and zeropadded. A [char:4] takes up 5 bytes including a zero termination byte and may hold any valid zero terminated ASCII or UTF-8 string of that length. While [char:1] is a valid type, char is not.

Note: the syntax [string:4] has been proposed to mean a string of 4 characters, but it actually means a fixed size vector of 4 offsets to variable length strings which is not supported.

The following example

struct Basics {
    int a;
    int b;
}

struct MyStruct {
    x: int;
    z: [short:3];
    y: float;
    w: [Basics:2];
    name: [char:4];
}

table MyTable {
    t: [ubyte:2];
    m: [MyStruct:2];
}

can replace MyStruct with the binary equivalent (but different codegeneration):

struct MyStructEquivalent {
    x: int;
    z1: short;
    z2: short;
    z3: short;
    y: float;
    wa1: int;
    wa2: int;
    name1: ubyte;
    name2: ubyte;
    name3: ubyte;
    name4: ubyte;
    wname_zterm: ubyte;
    x2: int;
    z21: short;
    z22: short;
    z23: short;
    y2: float;
    wa21: int;
    wa22: int;
    name21: ubyte;
    name22: ubyte;
    name23: ubyte;
    name24: ubyte;
    wname2_zterm: ubyte;
}

struct tEquivalent {
    t1: ubyte; 
    t2: ubyte; 
}

table MyTableEquivalent {
    t: tEquivalent;
    m: MyStructEquivalent;
}

Mixins

A mixin is a table or a struct that is apparently expanded into the table or struct that contains it.

Mixins add new accessors to make it appear as if the fields of the mixed in type is copied into the containing type although physically this isn't the case. Mixins can include types that themselves have mixins and these mixed in fields are also expanded.

A mixin is an attribute applied to table or a struct field when that field has a struct or a table type. The binary format is unchanged and the table or struct will continue to work as if it wasn't a mixin and is therefore fully backwards compatible.

Example:

struct Position {
    spawned: bool(required);
    x: int;
    y: int;
}

table Motion {
    place: Position(mixin);
    vx: int = 0;
    vy: int = 0;
}

table Status {
    title: string;
    energy: int;
    sleep: int;
}

table NPC {
    npcid: int;
    motion: Motion(mixin);
    stat: Status(mixin);
}

table Rock {
    here: Position(mixin);
    color: uint32 = 0xa0a0a000;
}

table Main {
    npc1: NPC;
    rock1: Rock;
}

root_type Main;

Here the table NPC and Rock will appear with read accessors is if they have the fields:

table NPC {
    npcid: int;
    spawned: bool(required);
    x: int;
    y: int;
    vx: int = 0;
    vy: int = 0;
    title: string;
    energy: int;
    sleep: int;
    place: Position;
    motion: Motion(required);
    stat: Status;
}

table Rock {
    spawned: bool(required);
    x: int;
    y: int;
    here: Position(required);
    color: uint32 = 0xa0a0a000;
}

table Main {
    npc1: NPC;
    rock1: Rock;
}

root_type Main;

or in JSON:

{
    "npc1": {
        "npcid": 1,
        "spawned": true;
        "x": 2,
        "y": 3,
        "vx": -4,
        "vy": 5,
        "title": "Monster",
        "energy": 6,
        "sleep": 0
    },
    "rock1": {
        "spawned": false;
        "x": 0,
        "y": 0,
        "color": 0xa0a0a000
    }
}

Note that there is some redundancy here which makes it possible to access the mixed in fields in different ways and to perform type casts to a mixed in type.

A cast can happen through a generated function along the lines of npc.castToPosition(), or via field a accessor npc.getPlace().

A JSON serializer excludes the intermediate container fields such as place and motion in the example.

A builder may choose to only support the basic interface and required each mixed in table or struct be created separately. A more advanced builder would alternatively accept adding fields directly, but would error if a field is set twice by mixing up the approaches.

If a mixed type has a required field, the required status propagates to the parent, but non-required siblings are not required as can be seen in the example above.

Mixins also places some constraints on the involved types. It is not possible to mix in the same type twice because names would conflict and it would no longer be possible to do trivially cast a table or struct to one of its kinds. An empty table or struct could be mixed in to provide type information but such a type can also only be added once.

Mixing in types introduces the risk of name conflicts. It is not valid to mix in a type directly or indirectly in a way that would lead to conflicting field names in a containing type.

Note that in the example it is not possible to mix in the pos struct twice, otherwise we could have mixed in a Coord class twice to have position and velocity, but in that case it would be natural to have two fields of Coord type which are not mixed in.

Schema evolution is fully supported because the vtable and field id's are kept separate. It is possible to add new fields to any type that is mixed in. However, adding fields could lead to name conficts which are then reported by the schema compiler.