Skip to content

Latest commit

 

History

History
270 lines (166 loc) · 14.6 KB

DESIGN.md

File metadata and controls

270 lines (166 loc) · 14.6 KB

cargo-mutants design

See README.md for an overview of how the project works.

See also CONTRIBUTING.md for more advice on style, approach, etc.

Physical structure / source tree overview

main.rs -- the cargo mutants entry point and command-line parsing.

cargo.rs -- Knows how to compose Cargo commands.

Actually running subprocesses is delegated to process.rs, so that we can later potentially run different build tools to Cargo.

build_dir.rs -- Manage temporary build directories.

console.rs -- colored output to the console including drawing progress bars. The interface to the console and indicatif crates is localized here.

copy_tree.rs -- Copy a source file tree into a build dir, with gitignore and other exclusions.

interrupt.rs -- Handle Ctrl-C signals by setting a global atomic flag, which is checked during long-running operations.

lab.rs -- A mutants "lab": manages generating and testing mutants. Contains effectively the main loop of the program: build and test every mutant.

log_file.rs -- Manage one log file per mutant scenario, within the output dir.

mutate.rs -- Different types of mutations we can apply, based on the AST from visit.rs, including generating a diff for the mutation and generating a tree with the mutation applied.

options.rs -- Global options for timeouts, etc. main.rs has the command line flags; this has an internal version of the options that have a pervasive effect through the program.

outcome.rs -- The result of running a single test or build, including distinguishing which type of command was run (check/build/test), where the log file is, what happened (success/failure/timeout/etc), and whether a mutation was applied.

output.rs -- Manages the mutants.out directory.

scenario.rs -- Each of the build/test cycles is a "scenario": either building the source tree, testing the baseline, or testing a mutant.

source.rs -- A source tree and files within it, including visiting each source file to find mutations.

span.rs -- A (line, column) addressing within a source file, a range between two positions, and edits to the content based on those addresses.

visit.rs -- Walk a source file's AST, and guess at likely-legal-but-wrong replacements. The interface to the syn parser is localized here, and also the core of cargo-mutants logic to guess at valid replacements.

Major processing stages

  1. Find the workspace enclosing the start directory, and the packages within it.
  2. Determine which packages to mutate.
  3. Generate mutants by walking each package.
  4. Copy the source tree.
  5. Run baseline tests.
  6. Test each mutant in parallel.

Finding the workspace and packages

cargo-mutants is invoked from within, or given with -d, a single directory, called the start directory. To find mutants and run tests we first need to find the enclosing workspace and the packages within it.

This is done basically by parsing the output of cargo locate-project and cargo metadata.

We often want to test only one or a subset of packages in the workspace. This can be set explicitly with --package and --workspace, or heuristically depending on the project metadata and the start directory.

For each package, cargo tells us the build targets including tests and the main library or binary. The tests are not considered for mutation, so this leaves us with some targets of interest, and for each of them cargo tells us one top source file, typically something like src/lib.rs or src/main.rs.

Discovering mutants

After discovering packages and before running any tests, we discover all the potential mutants.

Starting from the top files for each package, we parse each source file using syn and then walk its AST. In the course of that walk we can find three broad categories of patterns:

  • A mod statement (without a block), which tells us of another source file we must remember to walk.
  • A source pattern that cargo-mutants knows how to mutate, such as a function returning a value.
  • A pattern that tells cargo-mutants not to look further into this branch of the tree, such as #[test] or #[mutants::skip].

For baseline builds and tests, we test all the packages that will later be mutated. For mutant builds and tests, we pass --package to build and test only the package containing the mutant, on the assumption that each mutant should be caught by its own package's tests.

We may later mutate at a granularity smaller than a single function, for example by cutting out an if statement or a loop, but that is not yet implemented. (#73)

Copying the source tree

Mutations are tested in copies of the source tree. (An option could be added to test in-place, which would be nice for CI.)

Initially, one copy is made to run baseline tests; if they succeed then additional copies are made as necessary for each parallel job.

After copying the tree, cargo-mutants scans the top-level Cargo.toml and any .cargo/config.toml for relative dependencies. If there are any, the paths are rewritten to be absolute, so that they still work when cargo is run in the scratch directory.

Currently, the whole workspace tree is copied. In future, possibly only the package to be mutated could be copied: this would require changes to the code that fixes up dependencies.

Copies by default respect gitignore, but this can be turned off.

Each parallel build dir is copied from the original source so that it sees any gitignore files in parent directories.

(This current approach assumes that all the packages are under the workspace directory, which is common but not actually required.)

Handling timeouts

Mutations can cause a program to go into an infinite (or just very long) loop: for example we might mutate a function in if should_terminate() { break } to return false.

It's also possible that the un-mutated program has a bug that makes its test suite loop forever sometimes. Obviously this is a bug but we want cargo-mutants to be safe and easy to use on arbitrary trees that might have bugs.

We want to handle timeouts internally for a few reasons, including:

  • If one mutation hangs we still want to go on and try others. (So it's not so good if the cargo mutants process is killed by the user or a CI timeout.)

  • The fact that the mutation hung is a potentially interesting signal about the program to report. (Possibly the user will just have to mark should_terminate as skipped, but at least they can do that once and then have other builds go faster.)

  • For either CI or interactive use it's better if cargo mutants finishes in a bounded time.

(We are primarily concerned here with timeouts on tests; let's assume that cargo build will never get stuck; if it does then the whole environment probably has problems that need user investigation.)

The timeout for running tests is controlled by Options::timeout.

The timeout can be set by the user with --timeout, in which case it's simply used as is. If it's not specified, it is auto-set from the time to run the baseline tests, with a multiplier and a floor.

Detecting that a program has run too long is simple: we just watch the clock while waiting for it to finish. Terminating it, however, is more complicated:

The immediate child process spawned by cargo-mutants is cargo test .... This in turn spawns its own children running the various test binaries. It is these grandchild processes that are most likely stuck in a loop.

(It's also possible, and not unlikely, that the test binaries themselves start children: the cargo-mutants CLI tests do this. And those great-grand-children might get stuck. But the same logic applies.)

cargo mutants ....
  cargo test ...
    target/debug/someprog_api_test
    target/debug/someprog_cli_test
      target/debug/someprog ...

When we decide to stop the long-running test, we need to terminate the whole tree of processes. Unix provides a "process group" concept for doing this: we put the immediate child in a new process group, and then all its descendents will also be in that process group. We can stop the whole lot using killpg.

However, the test processes are then not in cargo-mutants's process group. So if the user hits ctrl-c on cargo mutants, that signal will not get to the test processes: cargo mutants would stop but the test process that's actually chewing up the CPU will continue.

Therefore we need to also intercept the signal to cargo-mutants and manually pass it on to the subprocess group.

Output directory handling

Various output files, including the text output from all the cargo commands are written into mutants.out within the directory specified by --output, or by default the source directory.

Cargo quirks

To build the tests, we actually run cargo test --no-test (or similarly for Nextest) so that the build uses the same profile as the tests. This is because the tests are built with the test profile, which can have different settings from the dev profile.

Handling strict lints

Some trees are configured so that any unused variable is an error. This is a reasonable choice to keep the tree very clean in CI, but if unhandled it would be a problem for cargo mutants. Many mutants -- in fact at the time of writing all generated mutants -- ignore function parameters and return a static value. Rejecting them due to the lint failure is a missed opportunity to consider a similar but more subtle potential bug.

Therefore when running rustc we configure all warnings off, with --cap-lints.

Testing

Cargo-mutants is primarily tested on its public interface, which is the command line. These tests live in tests/cli and generally have the form of:

  1. Make a copy of a testdata tree, so that it's not accidentally modified.
  2. Run a cargo-mutants command on it.
  3. Inspect the stdout, return code, or mutants.out.

cargo-mutants runs as a subprocess of the test process so that we get the most realistic view of its behavior. In some cases it is run via the cargo command to test that this level of indirection works properly.

Test performance

Since cargo-mutants itself runs the test suite of the program under test many times there is a risk that the test suite can get slow. Aside from slowing down developers and CI, this has a multiplicative effect on the time to run cargo mutants on itself.

To manage test time:

  • Although key behaviour should be tested through integration tests that run the CLI, it's OK to handle additional cases with unit tests that run much faster.

  • Whenever reasonable, CLI tests can only list mutants with --list rather than actually testing all of them: we have just a small set of tests that check that the mutants that are listed are actually run.

  • Use relatively small testdata trees that are sufficient to test the right behavior.

testdata trees

The primary means of testing is Rust source trees under testdata: you can copy an existing tree and modify it to show the new behavior that you want to test.

A selection of test trees are available for testing different scenarios. If there is an existing suitable tree, please use it. If you need to test a situation that is not covered yet, please add a new tree.

Please describe the purpose of the testdata tree inside the tree, either in Cargo.toml or a README.md file.

To make a new tree you can copy an existing tree, but make sure to change the package name in its Cargo.toml.

All the trees need to be excluded from the overall workspace in the top-level Cargo.toml.

--list tests

There is a general test that runs cargo mutants --list on each tree and compares the result to an Insta snapshot, for both the text and json output formats.

Many features can be tested adequately by only looking at the list of mutants produced, with no need to actually test the mutants. In this case the generic list tests might be enough.

Unit tests

Although we primarily want to test the public interface (which is the command line), unit tests can be added in a mod test {} within the source tree for any behavior that is inconvenient or overly slow to exercise from the command line.

Nextest tests

cargo-mutants tests require nextest to be installed.

UI Style

Always print paths with forward slashes, even on Windows. Use path_slash.

Logging

cargo-mutants writes two types of log files into the mutants.out directory:

Within the log directory, there is a file for each mutant scenario. This mostly contains the output from the Cargo commands, interspersed with some lines showing what command was run and the outcome. This is opened in append mode so that both cargo-mutants and cargo can write to it. This log is watched and the last line is shown in the progress bar to indicate what's going on.

Also, a file mutants.out/debug.log is written using tracing and written using the tracing macros. This contains more detailed information about what's happening in cargo-mutants itself.

Parallelism

cargo-mutants supports running multiple tests in parallel. Although both Cargo builds and Rust tests can parallelize internally, they can also both end up waiting for a single compilation unit, link, or straggling test. Running multiple tests in parallel makes better use of machine resources when each individual test becomes serialized.

The initial baseline build is done in a single job, with no parallelism at the cargo-mutants layer.

If the baseline test completes successfully, its build directory is copied to make a total of one per parallel job. Unlike the initial copy from the source directory, this includes the target directory, since it should now be up to date for the compiler options that cargo-mutants will use.

We then launch the appropriate number of threads, each of which has its own build directory. They each build and test new mutants until everything is done, or until there's an error that stops all processing.

RUSTFLAGS

Cargo has a somewhat complex system for controlling flags passed to rustc. cargo-mutants uses this to pass --cap-lints to avoid failures due to strict lints.

However, for other settings, we want to keep using whatever flags the user has configured in their environment, in the source tree, or in their per-user configuration. Unfortunately the Cargo RUSTFLAGS or CARGO_ENCODED_RUSTFLAGS variables entirely replace, instead of appending to, the user's flags.

Fully handling this seems to require us to reimplement some of Cargo's logic from https://doc.rust-lang.org/cargo/reference/config.html#buildrustflags to find the right target and config, in the right files, determine the user's flags, and then append our own.

For now, cargo-mutants appends to the environment variables but does not attempt to read the config files.