FIDL Tuning Proposal 043

Documentation Comment Format — Mark me up, mark me down

Field Value
Status Accepted
Authors jeremeymanson@google.com
Submitted 2019-05-06
Reviewed 2019-05-30

Summary

This FTP standardizes on a single format for writing documentation comments. This format can be consumed by tools that generate API documentation in other human readable and machine consumable formats (e.g., HTML, Markdown).

Motivation

We currently have an API documentation rubric that is very clear about what aspects of FIDL code need to be documented, including parameters, return values, and errors. We also encourage developers to provide examples of API usage. However, we have not given developers a clear way of expressing those features in FIDL API documentation.

With the advent of the fidldoc tool, it becomes more important to provide developers a way to express formatting in their comments. Developers writing comments should know how to, for example, format a list in their output. They must also know how to indicate that something is a return value or a parameter, so that it can be surfaced correctly in the output.

Design

TL;DR: We want to use markdown for comments. The devil, of course, is in the details.

This FTP modifies the API documentation rubric, as well as tooling we use to process FIDL API documentation. It does not affect the FIDL language, in that the set of legal FIDL remains the same.

Why Markdown?

The solution space for doc comments can be divided into two parts: "developing our own solution," and "using an existing solution." We feel that FIDL is not going to be a large enough ecosystem to warrant development of a separate standard for comment syntax. By using an existing solution, developers will be able to leverage external documentation and tooling (and potentially their existing knowledge). In addition, by using an existing solution, we may save development time.

If we commit to using an existing solution, we must pick one. There are several language specific solutions that could be extended (e.g., javadoc and python doc strings). There are also general-purpose solutions (e.g., LaTeX, RST, Markdown).

We believe that Markdown is the best choice. Unlike the language specific solutions, there are a number of tools that allow markdown integration into new languages. Markdown is also widely used and understood by developers: it's used, for example, by Github for its documentation. Finally, a number of languages (e.g., Rust and Kotlin) are standardizing on Markdown for their syntax, and it is starting to supplant existing solutions in other languages, as well (for example, LLVM is going to be migrating from RST to Markdown).

What do you mean, Markdown?

Markdown has a variety of implementations that have slightly different behaviors. Any number of them are reasonable choices. We choose CommonMark, because it is the closest we have to a standard. For developers whose tools need to target both CommonMark and another markdown implementation, we recommend keeping their docs compatible with both, to the extent possible.

Markdown is not extensible, and so does not help you express language elements. We add special-purpose extensions of markdown that can be processed by fidldoc (and other API doc consuming tools).

A doc comment is written in markdown and precedes the element it is documenting. It contains a description, optionally followed by documentation about parameters, errors, and "see" (which indicates that the reader should look at the referenced API for more information).

Parameters and Errors

Request parameters should be documented:

* request `paramname` Description of parameter

Response parameters should be documented:

* response `paramname` Description of parameter

We also considered param and return, or in and out, as the keywords instead of request and response.

In cases where the method does not use the same identifier as a parameter for both the request and response, the words request and response are optional.

Methods that return with no parameter value (Foo() -> ()) can use the term response without a corresponding parameter for documentation.

Error clauses should be documented:

* error Description of error values

Fully Qualified Names

Fully qualified names are of the form:

<library>/<top level declaration>.<member>

This will uniquely identify any member because there is no overloading.

Currently, ordinal hashing is based on names of the form <library>.<top level declaration>/<member> (see FTP-020), and fidlc reports errors using the form <library>.<top level declaration>/<member>. It's our intent to align these on the unambiguous format described above. We will amend FTP-029: Increasing Method Ordinals to use <library>/<top level declaration>.<member> as the name hashed, and modify fidlc to report errors consistently.

Links to other FIDL language elements that have documentation associated with them (or documented entities) can be made by adding [`link-target`]. For example, [`fidl.io/NodeInfo`] links to docs on said library. The resolution rules are as follows:

  1. First, nested elements will be checked. If you are documenting a struct Object, and it contains a member event, you may refer to it as [`event`].
  2. Next, elements at the same scope level as the documented element will be checked. For example, if you are documenting a protocol method foo(), and the same protocol contains a method bar(), you may refer to it as [`bar`].
  3. Next, elements of the enclosing scope are checked, if there is an enclosing scope. For example, if you are documenting a protocol method foo(), and there is another protocol in the same library called Info, you may refer to it (and its elements) by saying [`Info`].
  4. 3 is repeated at successively enclosed scopes until you are at the top level scope. If you are documenting a protocol method foo(), and you write [`fuchsia.io/NodeInfo`], it will refer to the union fuchsia.io/NodeInfo.

Fully qualified names are of the form <library>/<top level declaration>.<member>, see details above.

For other link shortcuts, you can specify the link target, e.g.:


That line will not appear in tooling output.

If the given FIDL target type is known to the tool at runtime, the location does not need to be specified. For example, it's likely that docs for fuchsia.sys.ComponentController and fuchsia.sys.EnvironmentController will be generated as part of the same tool invocation. The tool will know about links between them.

Developers may also use the following to indicate there is a related API:

* see [`fidl.io`]

Where appropriate.

Implementation Strategy

Implementation will include adding this to the FIDL rubric, publicizing it, and incorporating the special annotations into the fidldoc tool. We can also add lint checks for fidldoc syntax, either to the fidl-lint tool, or to a separate tool.

Documentation and Examples

A complete example is shown below. Note that this API does not currently look like this; the status has been changed to an error for illustration.

library fuchsia.io;

protocol File {
    /// Acquires a [`fuchsia.mem.Buffer`] representing this file, if
    /// there is one, with the requested access rights.
    ///
    /// ## Rights
    ///
    /// This method requires the following rights:
    ///
    /// * [`OPEN_RIGHT_WRITABLE`] if `flags` includes
    ///   [`VMO_FLAG_WRITE`].
    /// * [`OPEN_RIGHT_READABLE`] if `flags` includes
    ///   [`VMO_FLAG_READ`] or [`VMO_FLAG_EXEC`].
    ///
    /// + request `flags` a bit field composing any of
    ///   `VMO_FLAG_READ`, `VMO_FLAG_WRITE`, or `VMO_FLAG_EXEC`.
    /// - response `buffer` the requested [`fuchsia.mem.Buffer`], or
    ///     null if there was an error, or the buffer does not exist.
    /// * error a zx_status value indicating success or failure.
    /// * see [`fuchsia.mem.Buffer`]
    /// [`fuchsia.mem.Buffer`]:
    ///    http://fuchsia.dev/fidldoc/fuchsia/mem/Buffer.html
    GetBuffer(uint32 flags) ->
        (fuchsia.mem.Buffer? buffer) error zx.status;
};

Note that, by convention, you only need to link the first reference to an element in a given doc comment. The first reference to VMO_FLAG_READ above is linked, and the second one is not.

Backwards Compatibility

No significant backwards compatibility issues. Current doc uses the C++ style |param| notation to indicate parameters and return values. This can be changed relatively easily.

Performance

This will impact developer velocity by making them type more, but also understand more.

Drawbacks, Alternatives, and Unknowns

The assumption is that having a format is strictly better than not having a format. Therefore, there are few drawbacks.

Alternatives might include other API doc formats. Java uses Javadoc, which is very verbose and relies on inline HTML. Developers find it painful. Other languages use RST. However, this is becoming less popular; developers are simply more familiar with Markdown. Notably, the LLVM project is migrating from RST to Markdown.

We considered using a different variant of Markdown. We decided to use CommonMark because it is the best specified and standardized. Developers who need their code to work in other Markdown rendering systems should try to write doc comments that comply both with CommonMark and the systems they are targeting.

We considered not inventing a new syntax for linking documented entities. The alternatives that were considered were:

  • Auto-detection. Experience with auto-detection mechanisms in other contexts have shown that they rarely detect what the developer intended. In addition, auto-detection prevents tooling from surfacing the fact that a link is wrong. We therefore defer work on auto-detection until a future date.
  • Using existing syntax. This has the same problems as auto-detection, but the symptoms are somewhat less awful. If we were to use fuchsia.io/NodeInfo as a syntax, then if it were misspelled, the link would not be present, and we would simply get code font. We would like tooling that detects broken links, instead of having a fallback behavior.

Items that we should consider in the future, but are out of scope for this FTP, include:

  • Ways of embedding images or flowcharts into generated docs.
  • Ways of embedding automatically checked examples into docs.

Prior Art and References

This proposal is heavily influenced by the documentation styles for Rust and Kotlin.