Zig CCSDS Primary Header

This post is a write up of the zig ccsds primary header library I wrote as a test of Zig when dealing with binary data format. It contains a basic, but essentially complete, implementation of the CCSDS Space Packet Protocol Primary Header in the Zig programming language.

The point of writing this library is to try Zig out on a binary data specification 1) without serialization and deserialization, and 2) with the format known beforehand, to see how it compares to C and Rust. Overall it performs very well, and this is my preferred encoding compared to similar libraries I've written in other languages.

Primary Header Layout

The CCSDS Primary Header is a simple binary packet header containing three sections, each of which is two bytes, which is referred to as a 'word':

  • The 'control' section (this is my name, not in the standard) which contains an identifier for the source/destination pair, a flag for whether there is another header following the primary header, a flag indicating whether the packet contains a command or telemetry, and the protocol version (always 0).
  • The 'sequence' section contains the sequence flag indicating whether the packet is part of a larger related sequence of packets, and the sequence counter whose meaning depends on the sequence flag.
  • The 'length' section is simply the packet length, not including the header itself, minus 1. The minus 1 comes from the fact that a CCSDS packet cannot have an empty data section (there must be at least 1 byte afterwards), so the standard does not account for that one required byte in the length field. Whether this was a good idea or not (I don't think so) it ensures that all length field values are valid.

This diagram gives a basic description of the format, going from most significant bit to least significant bit from left to right, across three 16-bit fields for a total of six bytes:

+--------+-------+------------+-------------------------------------+
|Version |Packet |Secondary   |              APID                   |
|        |Type   |Header Flag |                                     |
+--------+-------+------------+-------------------------------------+
|3 bits  | 1 bit |   1 bit    |             11 bit                  |
+--------+--------------------+-------------------------------------+
|                      16 bit word, bits 0..15                      |
+-------------------------------------------------------------------+
 
+---------+---------------------------------------------------------+
|Sequence |                     Sequence Count                      |
|Flags    |                                                         |
+---------+---------------------------------------------------------+
| 2 bits  |                        14 bit                           |
+---------+---------------------------------------------------------+
|                     16 bit word, bits 16..31                      |
+-------------------------------------------------------------------+

+-------------------------------------------------------------------+
|                           Packet Length                           |
|                                                                   |
+-------------------------------------------------------------------+
|                              16 bits                              |
+-------------------------------------------------------------------+
|                      16 bit word, bits 32..47                     |
+-------------------------------------------------------------------+

Background

As a test of a new language, I sometimes write this library for encoding and decoding CCSDS primary headers. This is a fairly simple packet header, containing only 6 bytes, but forces you to think about how you will handle endianness and bit field data in a language. These are important for embedded systems programmers (and other systems programmes), so I see this as a litmus test for how a language will handle these kinds of concerns.

There are many ways to handle this kind of data encoding. Many libraries, especially in managed languages, is to deserialize the packet from bytes into a native structure, and then serialize it back when it is ready to be sent on the network. A similar process would be necessary for controlling hardware registers.

However, I always write this library in such a way that the binary data is kept in the format defined in the standard. I prefer this to avoid creating duplicates of the data in a system (the "true" form within a byte buffer, and copy deserialized and manipulated).

C

To accomplish this in C we have to keep the data as byte pairs and construct a 16 bit unsigned integer out of them (this keeps the layout fixed and works on both big and little endian systems). We can then do the bit mask and shift operations to extract the correct bits, and then the opposite mask/shift/OR operations to place bits within these 16-bit words. This is error prone, but it does work, its cross platform, and its preferrable to C bitfields.

When getting and setting fields, an integer with at least the field width is used. For example, the APID (Application Identifier) is 11 bits, but is read out as a uint16_t in C as C does not have the ability to describe integers that are not powers of two.

Rust

Rust has a number of possible solutions- it does not currently have bitfields, although perhaps it may in the future, so we have to create the same effect through other means. The C solution works just as well in Rust as in C, and there are other solutions if you are willing to accept libraries which provide macros that extend the Rust language.

There are a number of these macro libraries. However, some of them seem to do serialization/deserialization, and all of them require your types to be defined within the macro context. I would much prefer a language that allows me to do this kind of thing without this kind of extension. This is a complex topic, and many people may not feel this way, but I personally avoid language extension for this specific use-case.

Zig

For this library, I found that there are many options, especially around dealing with endianness. Zig does better with bitfields then C or Rust- they have a defined layout and should only access bytes that are required (which is important for hardware access). Zig also has the ability to work with integers of bit sizes that are not powers of 2, like a u3 as an unsigned three bit integer, which is nice as it describes the fields of a bit field without worrying as much about invalid values when getting and setting fields. Knowing this, I was determined to use bitfields for this library and finally get bit level access to a binary format.

However, these bit fields have one limitation for this particular binary layout- while Zig bit fields pack bits in a big-endian order, within each byte they pack Least Significant Bit (LSB) to Most Significant Bit (MSB). This is the packing in some formats, but the binary data I usually deal with packs from MSB to LSB in the first byte, moving to the next byte when all bits are used up, and this is the case in CCSDS. This can also be seen as packing LSB to MSB, but starting at the least significant byte of the byte aligned sequence, which is harder to state.

It turns out that the Zig bitfield packing can still be used in this situation, but with a little bit of extra work. It looks like if you data packs the way Zig bitfields natively pack it, then the format can be easily described directly in bitfields. Otherwise, a system like the one used in this library can be used (described below).

Accessing Bit Packed Data

Getting and setting fields within a bit field structure is an occasional task of embedded programming, so we want a strategy we can apply to these situations. They do not come up as often as other tasks, but when they do they are usually very important- perhaps part of the core requirements of a system.

Zig Native Bit Packing

As stated above, if your data is bit packed in the native format provided by Zig, just encode the fields in a bit field and you are done. access the fields as normal, and learn Zig's rules on non-byte-aligned data.

If you want, fields can be wrapped in types, so the u3 for the version number can be renamed to Version, to help communicate intent and avoid mixing fields with the same width, but this is a design choice that is up to you.

An example of this for the sequence flags would be:

pub const CcsdsSequence = packed struct {
    seq_flag: SeqFlag = SeqFlag.Unsegmented,
    sequence: Sequence = 0,
};

Which in bits, from MSB to LSB of the most significant byte to least significant byte, where where F is a bit cooresponding to the sequence flag, and S is a bit cooresponding to the sequence count:

SSSS SFFF SSSS SSSS

CCSDS Bit Packing

If your fields bit pack in the opposite order, you need a way to create the packing:

FFFS SSSS SSSS SSSS

One attempt might be to swap the field order, like so:

pub const CcsdsSequence = packed struct {
    sequence: Sequence = 0,
    seq_flag: SeqFlag = SeqFlag.Unsegmented,
};

This creates the following bit packing:

SSSS SSSS FFFS SSSS

So close! In fact, we also have to swap bytes to get:

FFFS SSSS SSSS SSSS

which is what we want.

This means that we need to wrap field accesses in some way to ensure that the byte-aligned integer that contains the field is read, byte swapped, used, and then (for setters) byte swapped again and placed in memory. In other words, we need to ensure that swapping occurs before and after using a field to get the designed bit packing.

I played around with various designs, such as wrapping types to contain type-level information about endianness and size, but I ended up no a fairly simple design that can automate some of the details through comptime programming.

Creating Fields

The design here is that each byte-aligned sequence, in this case the 'control' word, 'sequence' word, and length word, each get their own type, which is a bitfield. I choose to rename some bit types like Apid instead of using u11, but again, that is a design decision and not essential.

Once each bit field was defined, I created getters and setters for each field. These are required to ensure that the fields are accessed on correctly laid out fields. This is the one place where language features like properties (where accessing the fields can implcitly happen through a function), or injecting generated functions in a structure (kind of Ruby style monkey patching) would make the use of these structure slightly simplier.

However, I've very, very glad that Zig does not include these features- what I want out of ZIg is not programmer convienence, not is it every cool programming feature of modern languages. I want something small and predictable, where I control the flow of the program, the layout of data in memory, and I do not get bogged down in complex language feature. There are plenty of languages with fancy features for when we want them- I want to see languages choose discipline and control for times when we need it.

Speaking of fancy language features, the getters are setters are done using the set_field_swapped, and get_field_swapped function. These use comptime information to generally access a field of structure on a byte swapped version of the structure. They are not perfect- I believe they can be written more generically by using byte arrays instead of integer types for swapping, but get do automate the getter/setter pairs.

Comptime

Overall I am finding that comptime is perhaps the most challenging and interesting feature of Zig. I very much like its ability to perform introspection (I wish so much for this in C), to create polymorphic types, and its dependent type style capabilities. Its possible that it could make things difficult down the line- I truely don't know- such as for code analysis which needs to run a Turing complete language in order to analyze a file. Perhaps there could be some restrictions on the use of comptime, or perhaps comptime itself can perform the required analysis, but I think its too early to tell.

Either way, it turned out to be useful in this scenario, even though some of the error messages when I got it wrong where quite difficult to understand, and I ran into an not-yet-implemented feature of inferring return types of comptime functions, which lead me to take an extra type parameter that would otherwise not be needed.

Defining Structures

Once each byte-aligned set of fields was broken out into a packed struct, the CcsdsPrimary structure is just each field in sequence:

pub const CcsdsPrimary = packed struct {
    control: CcsdsControl,
    sequence: CcsdsSequence,
    length: CcsdsLength,
};

There are functions in here as well, but for simplicity this is the structure.

To access, say, the APID of a variable with type CcsdsPrimary, you would access its 'control' field and use the 'get_apid' function:

const primary = /* Create Primary Header */;
std.debug.print("APID = {}", .{ primary.control.get_apid() });

This is not the most convienent, and the getters and setters could have been moved into the CcsdsPrimary structure itself to avoid that field access, but I felt like this design is the most generally applicable- create these substructures with field accessors, combine them into a full structure with all fields, and access your data through field and accessor functions.

Conclusion

Overall I think this is the best CCSDS Primary Header library I've ever written.

My C version has to constantly deal with potential errors due to NULL pointers (which I check in every function, every time), and has to do a lot of manual bit manipulation. My Rust version does not have the error cases of the C version, but does the same bit manipulation as I did not want to extend the language with macros. The Zig version is the only one that does not have error conditions and does not require bit manipulation, although at the cost of extra syntax in getters and setters and a pair of perhaps tricky generic accessor functions to help write the accessor functions.

Both the Rust and Zig version are great for unit testing- I find that I benefit a great deal from being able to write simple tests alongside my code and run them quickly. My C version does have tests, using the Unity test framework from ThrowTheSwitch, but I have to deal with building it myself, and my tests are in a separate file so I can compile them out of release builds.

I did not tackle other concerns for a library like this, like validating headers, handling secondary headers or data sections (which requires casting byte arrays to different types), or data integrity checking such as checksums or CRCs. I would add these if I were using this library for real work, but for now I'm quite pleased with how this worked in Zig. I will continue to withhold judgement on the language until I've used it in anger, but it continues to impress me, even having now worked out something simple but non-trivial with it.