CONTRIBUTING.md

Contributing

Thanks for considering helping this project. There are many ways you can help: using the library and reporting bugs, reporting usability issues, making additions and improvements to the library, documentation and finding security bugs.

Reporting bugs

Please file a github issue. Include as much information as possible. Suspected protocol bugs are easier debugged with a pcap or reproduction steps.

Feel free to file github issues to get help, or ask a question.

If you believe you've found a security bug please open a draft security advisory in GitHub, and not as a regular repository issue. See SECURITY.md for more information.

Code changes

Some ideas and guidelines for contributions:

• For large features, file an issue prior to starting work. This means everyone can see what is in progress prior to a PR.
• Feel free to submit a PR even if the work is not totally finished, for feedback or to hand-over.
• Prefer not to reference github issue or PR numbers in commits.
• Try to keep code formatting commits separate from functional commits.
• See .github/workflows/build.yml for how to run the various test suites, and how to make coverage measurements.
• I run cargo outdated prior to major releases; but PRs to update specific dependencies are welcome.

Commit history

We prefer to keep the commit history clean and easy to follow. As such, we prefer small commits that do one thing. In particular:

• Avoid mixing refactoring and functional changes in the same commit if possible
• Make mechanical changes (like renaming or moving code around) in a separate commit
• Isolate updates to Cargo.lock in their own commits

Our default workflow is to rebase clean commit history from a PR to main.

Security bugs

Please report security bugs by opening a draft security advisory in GitHub, and not as a regular repository issue.

See SECURITY.md for more information.

If you're looking for security bugs, this crate is set up for cargo fuzz but would benefit from more runtime, targets and corpora.

Testing

• Features involving additions to the public API should have (at least) API-level tests (see rustls/tests/api.rs).
• Protocol additions should have some coverage -- consider enabling corresponding tests in the bogo suite, or writing some ad hoc tests.

PRs which cause test failures or a significant coverage decrease are unlikely to be accepted.

Testing with multiple CryptoProviders

Generally any test that relies on a CryptoProvider anywhere, should be run against all CryptoProviders, such that cargo test --all-features runs the test several times. To achieve that we have two methods:

• For unit tests, see rustls/src/test_macros.rs which provides the test_for_each_provider! macro. This can be placed around normal tests and instantiates the tests once per provider.

  Note that rustfmt does not format code inside a macro invocation: when developing test code, sed test_for_each_provider! { to mod tests {, format the code, then sed it back.

• For integration tests -- where the amount of test code is more significant, and lack of rustfmt support is more painful -- we instantiate the tests by importing them multiple times, and then the tests resolve the provider module to use via super::provider. For example, see rustls/tests/runners/unbuffered.rs and rustls/tests/unbuffered.rs.

Style guide

Ordering

Top-down ordering within modules

Within a module, we prefer to order items top-down. This means that items within a module will depend on items defined below them, but not (usually) above them. The idea here is that the public API, with more internal dependencies, will be read (and changed) more often, and putting it closer to the top of the module makes it more accessible.

This can be surprising to many engineers who are used to the bottom-up ordering used in languages like Python, where items can have a run-time dependency on other items defined in the same module.

Usually const values will thus go on the bottom of the module (least complex, usually no dependencies of their own), although in larger modules it can make sense to place a const directly below the user (especially if there is a single user, or just a few co-located users).

The #[cfg(test)] mod tests {} module goes on the very bottom, if present. Other module definitions (like mod foo { .. }) can be ordered among other items as it makes sense in the context of the items imported from them. Module declarations (like mod foo;) should be ordered before other items but after imports. Imports from local modules (both declared and defined) should be kept close to the module declaration/definition.

Files that have substantial amounts of code inside inline modules should probably avoid also having much code outside of these modules.

Ordering for a given type

For a given type, we prefer to order items as follows:

1. The type definition (struct or enum)
2. The inherent impl block (that is, not a trait implementation)
3. impl blocks for traits, from most specific to least specific. The least specific would be something like a Debug or Clone impl.

Ordering associated functions within an inherent impl block

Here's a guide to how we like to order associated functions:

0. Associated functions (that is, fn foo() {} instead of fn foo(&self) {})
1. Constructors, starting with the constructor that takes the least arguments
2. Public API that takes a &mut self
3. Public API that takes a &self
4. Private API that takes a &mut self
5. Private API that takes a &self
6. const values

Note that we usually also practice top-down ordering here; where these are in conflict, make a choice that you think makes sense. For getters and setters, the order should typically mirror the order of the fields in the type definition.

Functions

Consider avoiding short single-use functions

While single-use functions can make sense if the algorithm is sufficiently complex that it warrants an explicit name and interface, using many short single-use functions can make the code harder to follow, due to having to jump around in order to gain an understanding of what's going on. When writing a single-use function, consider whether it needs the dedicated interface, or if it could be inlined into its caller instead.

Consider avoiding free-standing functions

If a function's semantics or implementation are strongly dependent on one of its arguments, and the argument is defined in a type within the current crate, prefer using a method on the type. Similarly, if a function is taking multiple arguments that originate from the same common type in all call-sites it is a strong candidate for becoming a method on the type.

Order arguments from most specific to least specific

When writing a function, we prefer to order arguments from most specific to least specific. This means that an image_id might go before the domain, which will go before the app context. More specific arguments are more differentiating between a given function and other functions, so putting them first makes it easier to infer the context/meaning of the function (compared to starting with a number of generic context-like types).

Use impl where possible

We prefer to use impl ... for arguments and return types when there's a single use of the type. Generic type argument bounds add a level of indirection that's harder to read in one pass.

Avoid type elision for fully qualified function calls

We prefer to write fully qualified function calls with types included, rather than elided. For example:

// Incorrect:
<_>::default()

// Correct:
CertificateChain::default()

Validation

Where possible, avoid writing validate or check type functions that try to check for error conditions based on the state of a populated object. Prefer "parse, don't validate" style and try to use the type system to make it impossible for invalid states to be represented.

Error handling

We use Result types pervasively throughout the code to signal error cases. Outside of unit/integration tests we prefer to avoid unwrap() and expect() calls unless there is a clear invariant which can be locally validated by the structure of the code. If there is such an invariant, we usually add a comment explaining how the invariant is upheld. In other cases (especially for error cases which can arise from network traffic, which could represent an attacker), we always prefer to handle errors and ultimately return an error to the network peer or close the connection.

Expressions

Avoid single-use bindings

We generally make full use of the expression-oriented nature of Rust. For example, when using iterators we prefer to use map and other combinators instead of for-loops when possible, and will often avoid variable bindings if a variable is only used once. Naming variables takes cognitive efforts, and so does tracking references to bindings in your mind. One metric we like to minimize is the number of mutable bindings in a given scope.

Remember that the overall goal is to make the code easy to understand. Combinators can help with this by eliding boilerplate (like replacing a None => None arm with a map() call), but they can also make it harder to understand the code. One example is that a combinator chain like .map().map_err() might be harder to understand than a match statement (since, in this case, both of the arms have a significant transformation).

Use early return and continue to reduce nesting

The typed nature of Rust can cause some code to end up at deeply indented levels, which we call "rightward drift". This makes lines shorter, making the code harder to read. To avoid this, try to return early for error cases, or continue early in a loop to skip an iteration.

Hoist common expression returns

When writing a match or if expression that has arms that each share a return type (e.g. Ok(...)), hoist the commonality outside the match. This helps separate out the important differences and reduces code duplication.

// Incorrect:
match foo {
    1..10 => Ok(do_one_thing()),
    _ => Ok(do_another()),
}

// Correct:
Ok(match foo {
    1..10 => do_one_thing(),
    _ => do_another(),
})

Avoid ref in match patterns

When writing match expressions, try to avoid using ref in patterns. Prefer taking a reference on the scrutinee of the match.

Since the addition of binding modes for improved match ergonomics the ref keyword is unidiomatic and can be unfamiliar to readers.

Naming

Use concise names

We prefer concise names, especially for local variables, but prefer to expand acronyms/abbreviations that are not very well known (e.g. prefer key_usage instead of ku, anonymous instead of anon). Extremely common short-forms like url are acceptable.

Avoid adding a suffix for a variable that describes its type (provided that its type is hard to confuse with other types -- for example, we do still use _id suffixes because we usually use numeric IDs for database entities). The precision/conciseness trade-off for variable names also depends on the scope of the binding.

Avoid get_ prefixes

Per the API guidelines, get_() prefixes are discouraged.

Enum variants

When implementing or modifying an enum type, list its variants in alphabetical order. It's acceptable to ignore this advice when matching the order imposed by an external source, e.g. a standards document.

Prefer active verbs for variant names. E.g. Allow instead of Allowed, Forbid instead of Forbidden. Avoid faux-bools like Yes and No, instead preferring variant names that are descriptive of the different states.

Don't elide generic lifetimes

We prefer not to elide lifetimes when naming types that are generic over lifetimes. Always include a lifetime placeholder (e.g. <'_>) to avoid confusion.

Imports

We use 3 blocks of imports in our Rust files:

1. std imports
2. Imports from external crates
3. Crate-internal imports

We believe that this makes it easier to see where a particular import comes from.

Within the import blocks we prefer to separate imports that don't share a parent module. For example,

// Incorrect
use alloc::{format, vec::Vec};

// Correct
use alloc::format;
use alloc::vec::Vec;

We prefer to reference types and traits by an imported symbol name instead of using qualified references. Qualification paths generally add noise and are unnecessary. The one exception to this is when the symbol name is overly generic, or easily confused between different crates. In this case we prefer to import the symbol name under an alias, or if the parent module name is short, using a one-level qualified path. E.g. for a crate with a local Error type, prefer to import std::error::Error as StdError.

Exports

We prefer to export types under a single name, avoiding re-exporting types from the top-level lib.rs. The exception to this are "paved path" exports that we expect every user will need. The canonical example of such types are client::ClientConfig and server::ServerConfig. In general this sort of type is rare and most new types should be exported only from the module in which they are defined.

Misc

Numeric literals

Prefer a numeric base that fits with the domain of the value being used. E.g. use hexadecimal for protocol message literals, and octal for UNIX privileges. Use digit grouping to make larger numeric constants easy to read, e.g. use 100_000_000 instead of 100000000.

Avoid type aliases

We prefer to avoid type aliases as they obfuscate the underlying type and don't provide additional type safety. Using the newtype idiom is one alternative when an abstraction boundary is worth the added complexity.

Design and Architecture

Some general concepts about how the library should fit together:

• Linker friendliness
• Small mandatory API
• Safe and sensible defaults
• Separation of mechanism and policy

Linker friendliness

When a program incorporates rustls, we should try to ensure that parts that are not used can be discarded by the linker.

The linker can discard code if it is unreachable by the code that is referenced by the downstream program. This generally means that:

• runtime trait-based or function pointer dynamic dispatch is good,
• compile-time trait-based generics are good (so long as the same code is not monomorphized multiple times in a typical program),
• enum-based dispatch is bad.

Here is an example that is less good:

enum Algorithm {
    Aes128,
    ChaCha20,
}

fn encrypt(alg: Algorithm, buffer: &mut [u8]) {
    match alg {
        Algorithm::Aes128 => encrypt_aes128(buffer),
        Algorithm::ChaCha20 => encrypt_chacha20(buffer),
    }
}

A program that only used encrypt(Algorithm::ChaCha20, ..) would end up with a copy of AES inside. (If such a function was inlined into its caller, the reference to encrypt_aes128 could theoretically be deleted if the compiler can be certain of the value of alg. This is not guaranteed to happen.)

Instead, prefer:

trait Algorithm {
    fn encrypt(&self, buffer: &mut [u8]);
}

struct Aes128;
struct ChaCha20;

impl Algorithm for Aes128 {
    fn encrypt(&self, buffer: &mut [u8]) {
        encrypt_aes128(buffer)
    }
}

impl Algorithm for ChaCha20 {
    fn encrypt(&self, buffer: &mut [u8]) {
        encrypt_chacha20(buffer)
    }
}

(or a function-pointer equivalent, which would be less idiomatic Rust.)

Some judgement is needed: this is only worth it if a reasonable program would not use all the possibilities, and the individual parts are large enough to have an impact on the size of a final program.

Small mandatory API

We should try to keep the API for achieving the most common goals as simple as possible. Good models for this are simpleclient and simpleserver. The purpose of this is to allow people interacting with the library for the first time to make progress.

Safe and sensible defaults

We should be confident to make decisions on behalf of our users and express that in the library's defaults. It is good to encode our experience of TLS in this way, and we should not require that users are experts in TLS to end up with a result that is working, secure, stable, and performant.

Separation of mechanism and policy

A configuration knob with a good default is preferable to a fixed behavior, even if the end result is the same for 99% of users.

With that said, we should take care to avoid overloading users with choices, and direct them (with examples, documentation, and simplified top-level APIs) away from that complexity.

There is quite a lot of nuance here: configuration should avoid allowing bad outcomes (in the cryptographic, memory safety, and "good netizen" senses). Where that is not possible we typically include danger in API naming to assist code reviewers of end-user code.

Licensing

Contributions are made under rustls's licenses.