Version 17 (modified by Vojtech Horky, 12 years ago) ( diff )

Make the source code example more visible

Structured Binary Data

As part of Google Summer of Code 2012, Bithenge is being created to address #317. This page describes the project’s design and implementation. The code is at lp:~wtachi/helenos/bithenge and periodic updates are posted to HelenOS-devel.


Exploring and working with structured binary data is necessary in many different situations in a project like HelenOS. For instance, when implementing a file format or filesystem, it is first necessary to explore preexisting files and disks and learn the low‐level details of the format. Debugging compiled programs, working with core dumps, and exploring network protocols also require some way of interpreting binary data.

The most basic tool for exploring binary data is the hex editor. Using a hex editor is inefficient and unpleasant because it requires manual calculation of lengths and offsets while constantly referring back to the data format. General‐purpose scripting languages can be used instead, so a structure can be defined once and decoded as often as necessary. However, even with useful tools like Python’s struct module, the programmer must specify how to read the input data, calculate lengths and offsets, and provide useful output, so there’s much more work involved than simply specifying the format of the data. This extra code will probably be rewritten every time a new script is made, due to slightly differing requirements.

The Bithenge project involves creating a powerful library and tools that will make working with structured binary data faster and easier. It will consist of:

  • A core library that manages structured data and provides basic building blocks for binary data interpretation.
  • Data providers to access various sources of raw binary data.
  • Format providers, which can load and save complex format specifications. In particular, there will be a domain‐specific language for format specifications.
  • Clients, programs which use the library to work with binary data. For instance, there will be an interactive browser.

The initial goals for the project are an interactive browser for filesystem structures as well as a debugger backend that can interpret core dumps and task memory.


  • Work in HelenOS—this means the code must be in C and/or an easily ported language like Lua.
  • View on different layers. For instance, when viewing a FAT directory entry, it should be possible to switch between viewing the formatted date and time, the integers, and the original bytes.
  • Check whether data is valid; handle broken data reasonably well.
  • Parse pieces of the data lazily; don’t try to read everything at once.
  • Work in both directions (parsing and building) without requiring too much extra effort when specifying the format.
  • Support full modifications. Ideally, allow creation of a whole filesystem from scratch.


Bithenge represents all data in the form of a data structure called a “tree,” similar to the data structure represented by JSON. A tree consists of a boolean node, integer node, string node, or blob node, or an internal node with children. A boolean node holds a boolean value, an integer node holds a signed integer, and a string holds a Unicode string.

A blob node represents an arbitrary sequence of raw bytes. Blob nodes are polymorphic, allowing any source of raw binary data to be used. Bithenge includes blob node implementations for in‐memory buffers, files, and block devices. An implementation has also been written that reads another task’s virtual memory, but it hasn’t been committed because it’s unclear whether it will be useful in its current form.

An internal node has an arbitrary number of children, each with a unique key. The key can be any node other than an internal node. Arrays can be represented by internal nodes with integer keys starting at 0. The tree node can provide children in an arbitrary order; the order will be used when displaying the tree, but should have no semantic significance. Internal nodes are polymorphic and can delay creation of child nodes until necessary, so keeping the whole tree in memory can be avoided.

Internal nodes are currently responsible for freeing their own children. In the future, it should be possible for there to be multiple references to the same node, but it isn’t clear whether this should be implemented with symbolic links, an acyclic graph with reference counting, or a full graph.

Note that all interpreted data is represented in Bithenge with nodes. Therefore, the word “blob” usually refers to a blob node, and so on.

A decoded tree for a FAT filesystem might look something like this:

│         ├───0───▶0xfff0
│         ├───1───▶0xffff
│         └───2───▶0x0000
           │        ├───name───▶README.TXT
           │        └───size───▶0x1351


A transform is a function from a tree to a tree. One example is uint32le, which takes a 4‐byte blob node as the input tree and provides an integer node as the output tree. Another example would be FAT16_filesystem, a transform that takes a blob node as the input tree and provides a complex output tree with various decoded information about the filesystem. Some transforms, like uint32le, are built in to Bithenge; more complicated transforms can be loaded from a script file.

Transforms are represented in Bithenge with a polymorphic object. The primary method is apply, which applies a transform to an input tree and creates an output tree. When a transform takes a blob node as input, it is sometimes necessary to determine the prefix of a given blob that can be used as input to the transform; the method prefix_length can be used for this.

Built‐in transforms

These transforms are implemented in C and included with Bithenge. Note that precise names are preferred; scripts can define shorter aliases if necessary.

name input output description example
ascii blob node string decodes some bytes as ASCII characters hex:6869 becomes "hi"
uint8 1‐byte blob node integer node decodes a 1‐byte unsigned integer hex:11 becomes 17
uint16be 2‐byte blob node integer node decodes a 2‐byte big‐endian unsigned integer hex:0201 becomes 513
uint16le 2‐byte blob node integer node decodes a 2‐byte little‐endian unsigned integer hex:0101 becomes 257
uint32be 4‐byte blob node integer node decodes a 4‐byte big‐endian unsigned integer hex:00000201 becomes 513
uint32le 4‐byte blob node integer node decodes a 4‐byte little‐endian unsigned integer hex:01010000 becomes 257
uint64be 8‐byte blob node integer node decodes a 8‐byte big‐endian unsigned integer hex:0000000000000201 becomes 513
uint64le 8‐byte blob node integer node decodes a 8‐byte little‐endian unsigned integer hex:0101000000000000 becomes 257
zero_terminated blob node blob node takes bytes up until the first 00 hex:7f0400 becomes hex:7f04

Basic syntax

Script files are used to define complicated transforms.

Transforms (including built‐in transforms) can be referenced by name: uint32le.

Transforms can be given a new name: transform u32 = uint32le; defines a shorter alias for uint32le.

Transforms can be composed to create a new transform that applies them in order. The transform ascii <- zero_terminated first removes the 0x00 from the end of the blob, then decodes it as ascii. Note that the order of composition is consistent with function composition and nested application in mathematics, and also consistent with the general idea that data moves from right to left as it is decoded.

Transforms can have parameters that affect how they decode data:

transform u16(little_endian) =
    if (little_endian) {
    } else {

When such a transform is used, expressions must be given to calculate its parameters. Currently, the only possible expressions are parameters given to the current transform, boolean and integer literals, and previously decoded fields in the current struct:

transform item(little_endian) = struct {
    .len <- u16(little_endian);
    .text <- ascii <- known_length(.len);
    .data <- known_length(8);


Structs are used when a blob contains multiple data fields in sequence. A struct transform applies each subtransform to sequential parts of the blob and combines the results to create an internal node. The result of each subtransform is either assigned a key or has its keys and values merged into the final tree. Each subtransform must support prefix_length, so the lengths and positions of the data fields can be determined.


transform point = struct {
    .x <- uint32le;
    .y <- uint32le;

transform labeled_point = struct {
    .id <- uint32le;
    .label <- ascii <- zero_terminated;
    <- point;

If labeled_point is applied to hex:0600000041000300000008000000, the result is {"id": 6, "label": "A", "x": 3, "y": 8}.

Using Bithenge

The Bithenge source code is in uspace/app/bithenge and is built along with HelenOS. It can be built for Linux instead with make -f Makefile.linux.

The program can be run with bithenge <script file> <source>. The script file must define a transform called main. The source can start with one of the following prefixes:

Prefix Example Description
file: file:/textdemo Read the contents of a file. This is the default if no prefix is used.
block: block:bd/initrd Read the contents of a block device. (HelenOS only.)
hex: hex:01000000 Use a string of hexadecimal characters to create a blob node.

There are some example files in uspace/dist/src/bithenge.

Using the API

An overview of the API will be written later.

Nodes, expressions, and transforms use reference counting. Functions that produce such objects (through a bithenge_xxx_t ** argument) create a new reference to the object; you are responsible for ensuring the reference count is eventually decremented. If a function’s documentation says it “takes [ownership of] a reference” to an object, the function guarantees the object’s reference count will eventually be decremented, even if an error occurs. Therefore, if you create an object only to immediately pass it to such a function, you do not need to worry about its reference count.

Future language ideas

In approximate order of priority.

Other ideas

Conditional transforms
A way to apply different transforms depending on an expression. For example, something like: if (.has_extra) { struct { .extra <- uint32le; } }.
Transforms may need to be repeated a known number of times, until the end of the data, or until the transform indicates that repetition should stop. For instance, repeat(.len) {uint32le;}. The result could be a tree like {0: 1351, 1: 17}.
When there are pointers to other offsets in the blob, the script could pass the whole blob as a parameter and apply transforms to subblobs. This is essential for non‐sequential blobs like filesystems.
struct will be extended to work with bits instead of just bytes.
Complex expressions
Expressions that use operators or call transforms.
These could be implemented as transforms that don't actually change the input. There could be multiple levels, ranging from “warning” to “fatal error”.
An easier way to handle many constant values, like enum { 0: "none", 1: "file", 2: "directory", 3: "symlink" }.
Transforming internal nodes
After binary data is decoded into a tree, it should be possible to apply further transforms to interpret the data further. For instance, after the FAT and directory entries of a FAT filesystem have been decoded, a further transform could determine the data for each file.
Hidden fields
Some fields, such as length fields, are no longer interesting after the data is decoded, so they should be hidden by default.
Decoding may require searching for a fixed sequence of bytes in the data.
Automatic parameters
It could be useful to automatically pass some parameters rather than computing and passing them explicitly. For instance, a version number that affects the format of many different parts of the file could be passed automatically, without having to write it out every time. A more advanced automatic parameter could keep track of current offset being decoded within a blob. There would need to be some sort of scoping to determine which transforms have the automatic parameters.

Constraint‐based version

This and most other projects use an imperative design, where the format specification is always used in a fixed order, one step at a time. The imperative design causes problems when the user wants to modify a field, because arbitrary changes to other fields may be necessary that cannot be determined from the format specification.

It may be possible to solve this with a constraint-based design, where the format specification consists of statements that must be true about the raw and interpreted data, and the program figures out how to solve these constraints. Unfortunately, this approach seems too open-ended and unpredictable to fit within GSoC.

Interesting formats

These formats will be interesting and/or difficult to handle. I will keep them in mind when designing the library.

  • Filesystem allocation tables, which should be kept consistent with the actual usage of the disk.
  • Filesystem logs, which should be applied to the rest of the disk before interpreting it.
  • Formats where the whole file can have either endianness depending on a field in the header.
  • The Blender file format is especially dynamic. When Blender saves a file, it just copies the structures from memory and translates the pointers. Since each Blender version and architecture will have different structures, the output file includes a header describing the fields and binary layout of each structure. When the file is loaded, the header is read first and the structures will be translated as necessary.
  • If the language is powerful enough, it might be possible to have a native description of Zlib and other compression formats.
  • It could be interesting to parse ARM or x86 machine code.

Existing Tools

I researched existing tools related to my project, so they can be used for inspiration.


Construct is a Python library for creating declarative structure definitions. Each instance of the Construct class has a name, and knows how to read from a stream, write to a stream, and determine its length. Some predefined Construct subclasses use an arbitrary Python function evaluated at runtime, or behave differently depending on whether sub‐Constructs throw exceptions. Const uses a sub‐Construct and makes sure the value is correct. Also has lazy Constructs.

Unfortunately, if you change the size of a structure, you still have to change everything else manually.


BinData makes good use of Ruby syntax; it mostly has the same features as Construct.

Imperative DSLs

DSLs in this category are used in an obvious, deterministic manner, and complex edits (changing the length of a structure) are difficult or impossible. They are simple imperative languages in which fields, structures, bitstructures, and arrays can be defined. The length, decoded value, and presence of fields can be determined by expressions using any previously decoded field, and structures can use if/while/continue/break and similar statements. Structures can inherit from other structures, meaning that the parent’s fields are present at the beginning of the child. Statements can move to a different offset in the input data. There may be a real programming language that can be used along with the DSL.

Uses a simple stack‐based VM to calculate variable locations.
“Grammar‐based specification and parsing of binary file formats”
Actually uses an attribute grammar, but it isn’t terribly different from an imperative language.
Lets you create or modify files instead of just reading them. Fields can refer to blocks of data elsewhere in the file. Uses an XML format.
A popular tool for extracting files from video game archives. Its main strength is the broad number of compression formats supported. It can put modified files back in the archive in trivial situations.
Synalize It!
Not completely imperative; if you declare optional structs where part of the data is constant, the correct struct will be displayed. Has a Graphviz export of file structure. Uses an XML format.
Other free
BinPAC, Data::ParseBinary, DataScript, DataWorkshop, Wireshark Generic Dissector, Metafuzz BinStruct, and PADS.
Other proprietary
010 Editor, Andys Binary Folding Editor, Hackman Suite, Hex Editor Neo, iBored, and WinHext.

Less interesting tools

Simple formats in hex editors
These support static fields and dynamic lengths only: FlexHex, HexEdit, Hexplorer, Hex Workshop, and Okteta.
Simple formats elsewhere
CTF, ffe, Node Packet, Scapy, and stabs can only handle trivial structures. lpack, Perl’s pack, Python’s struct, and VStruct use concise string formats to describe simple structures. Hachoir uses Python for most things.
Protocol definition formats
ASN.1, MIDL, Piqi, and other IPC implementations go in the other direction: they generate a binary format from a text description of a structure. ASN.1 in particular has many features.
Wireshark and tcpdump
As the Construct wiki notes, you would expect these developers to have some sort of DSL, but they just use C for everything. Wireshark does use ASN.1, Diameter, and MIDL for protocols developed with them.

Miscellaneous ideas

Code exporter

A tool could generate C code to read and write data given a specification. A separate file could be used to specify which types should be used and which things should be read lazily or strictly.


A diff tool could show differences in the interpreted data.

Space‐filling curves

Space‐filling curves look cool, but this project is about avoiding looking at raw binary data.

Note: See TracWiki for help on using the wiki.