core-lightning/doc/STYLE.md

Care And Feeding of Your Fellow Coders

Style is an individualistic thing, but working on software is group activity, so consistency is important. Generally our coding style is similar to the Linux coding style.

Communication

We communicate with each other via code; we polish each others code, and give nuanced feedback. Exceptions to the rules below always exist: accept them. Particularly if they're funny!

Prefer Short Names

num_foos is better than number_of_foos, and i is better than counter. But bool found; is better than bool ret;. Be as short as you can but still descriptive.

Prefer 80 Columns

We have to stop somewhere. The two tools here are extracting deeply-indented code into their own functions, and use of short-cuts using early returns or continues, eg:

	for (i = start; i != end; i++) {
		if (i->something)
			continue;

		if (!i->something_else)
			continue;

		do_something(i);
}

Tabs and indentaion

The C code uses TAB charaters with a visual indentation of 8 whitespaces. If you submit code for a review, make sure your editor knows this.

When breaking a line with more than 80 characters, align parameters and arguments like so:

static void subtract_received_htlcs(const struct channel *channel,
				    struct amount_msat *amount)

Note: For more details, the files .clang-format and .editorconfig are located in the projects root directory.

Prefer Simple Statements

Notice the statement above uses separate tests, rather than combining them. We prefer to only combine conditionals which are fundamentally related, eg:

	if (i->something != NULL && *i->something < 100)

Use of take()

Some functions have parameters marked with TAKES, indicating that they can take lifetime ownership of a parameter which is passed using take(). This can be a useful optimization which allows the function to avoid making a copy, but if you hand take(foo) to something which doesn't support take() you'll probably leak memory!

In particular, our automatically generated marshalling code doesn't support take().

If you're allocating something simply to hand it via take() you should use NULL as the parent for clarity, eg:

	msg = towire_shutdown(NULL, &peer->channel_id, peer->final_scriptpubkey);
	enqueue_peer_msg(peer, take(msg));

Use of tmpctx

There's a convenient temporary tal context which gets cleaned regularly: you should use this for throwaways rather than (as you'll see some of our older code do!) grabbing some passing object to hang your temporaries off!

Enums and Switch Statements

If you handle various enumerated values in a switch, don't use default: but instead mention every enumeration case-by-case. That way when a new enumeration case is added, most compilers will warn that you don't cover it. This is particularly valuable for code auto-generated from the specification!

Initialization of Variables

Avoid double-initialization of variables; it's better to set them when they're known, eg:

	bool is_foo;
	
	if (bar == foo)
		is_foo = true;
	else
		is_foo = false;

	...
	if (is_foo)...

This way the compiler will warn you if you have one path which doesn't set the variable. If you initialize with bool is_foo = false; then you'll simply get that value without warning when you change the code and forget to set it on one path.

Initialization of Memory

valgrind warns about decisions made on uninitialized memory. Prefer tal and tal_arr to talz and tal_arrz for this reason, and initialize only the fields you expect to be used.

Similarly, you can use memcheck(mem, len) to explicitly assert that memory should have been initialized, rather than having valgrind trigger later. We use this when placing things on queues, for example.

Use of static and const

Everything should be declared static and const by default. Note that tal_free() can free a const pointer (also, that it returns NULL, for convenience).

Typesafety Is Worth Some Pain

If code is typesafe, refactoring is as simple as changing a type and compiling to find where to refactor. We rely on this, so most places in the code will break if you hand the wrong type, eg type_to_string and structeq.

The two tools we have to help us are complicated macros in ccan/typesafe_cb allow you to create callbacks which must match the type of their argument, rather than using void *. The other is ARRAY_SIZE, a macro which won't compile if you hand it a pointer instead of an actual array.

Use of FIXME

There are two cases in which you should use a /* FIXME: */ comment: one is where an optimization is possible but it's not clear that it's yet worthwhile, and the second one is to note an ugly corner case which could be improved (and may be in a following patch).

There are always compromises in code: eventually it needs to ship. FIXME is grep-fodder for yourself and others, as well as useful warning signs if we later encounter an issue in some part of the code.

If You Don't Know The Right Thing, Do The Simplest Thing

Sometimes the right way is unclear, so it's best not to spend time on it. It's far easier to rewrite simple code than complex code, too.

Write For Today: Unused Code Is Buggy Code

Don't overdesign: complexity is a killer. If you need a fancy data structure, start with a brute force linked list. Once that's working, perhaps consider your fancy structure, but don't implement a generic thing. Use /* FIXME: ...*/ to salve your conscience.

Keep Your Patches Reviewable

Try to make a single change at a time. It's tempting to do "drive-by" fixes as you see other things, and a minimal amount is unavoidable, but you can end up shaving infinite yaks. This is a good time to drop a /* FIXME: ...*/ comment and move on.

Creating JSON APIs

Our JSON RPCs always return a top-level object. This allows us to add warnings (e.g. that we're still starting up) or other optional fields later.

Prefer to use JSON names which are already in use, or otherwise names from the BOLT specifications.

The same command should always return the same JSON format: this is why e.g. listchannels return an array even if given an argument so there's only zero or one entries.

All warning fields should have unique names which start with warning_, the value of which should be an explanation. This allows for programs to deal with them sanely, and also perform translations.

Documenting JSON APIs

We use JSON schemas to validate that JSON-RPC returns are in the correct form, and also to generate documentation. See doc/schemas/WRITING_SCHEMAS.md.

Changing JSON APIs

All JSON API changes need a Changelog line (see below).

You can always add a new output JSON field (Changelog-Added), but you cannot remove one without going through a 6-month deprecation cycle (Changelog-Deprecated)

So, only output it if allow-deprecated-apis is true, so users can test their code is futureproof. In 6 months remove it (Changelog-Removed).

Changing existing input parameters is harder, and should generally be avoided. Adding input parameters is possible, but should be done cautiously as too many parameters get unwieldy quickly.

Github Workflows

We have adopted a number of workflows to facilitate the development of Core Lightning, and to make things more pleasant for contributors.

Changelog Entries in Commit Messages

We are maintaining a changelog in the top-level directory of this project. However since every pull request has a tendency to touch the file and therefore create merge-conflicts we decided to derive the changelog file from the pull requests that were added between releases. In order for a pull request to show up in the changelog at least one of its commits will have to have a line with one of the following prefixes:

• Changelog-Added: if the pull request adds a new feature
• Changelog-Changed: if a feature has been modified and might require changes on the user side
• Changelog-Deprecated: if a feature has been marked for deprecation, but not yet removed
• Changelog-Fixed: if a bug has been fixed
• Changelog-Removed: if a (previously deprecated) feature has been removed
• Changelog-Experimental: if it only affects --enable-experimental-features builds, or experimental- config options.

In case you think the pull request is small enough not to require a changelog entry please use Changelog-None in one of the commit messages to opt out.

Under some circumstances a feature may be removed even without deprecation warning if it was not part of a released version yet, or the removal is urgent.

In order to ensure that each pull request has the required Changelog-*: line for the changelog our trusty @bitcoin-bot will check logs whenever a pull request is created or updated and search for the required line. If there is no such line it'll mark the pull request as pending to call out the need for an entry.