cargo-mutants

https://github.com/sourcefrog/cargo-mutants

cargo-mutants is a mutation testing tool for Rust. It guides you to missing test coverage by finding functions whose implementation could be replaced by something trivial and the tests would all still pass.

Coverage measurements can be helpful, but they really tell you what code is reached by a test, and not whether the test really assert anything about the behavior of the code. Mutation tests give different, perhaps richer information, about whether the tests really check the code's behavior.

CAUTION: This tool builds and runs code with machine-generated modifications. If the code under test, or the test suite, has side effects such as writing or deleting files, running it with mutations may be dangerous. Think first about what side effects the test suite could possibly have, and/or run it in a restricted or disposable environment.

NOTE: cargo-mutants is still pretty new! It can find some interesting results, but because it has a very basic idea of which functions to mutate and how, it generates significant false positives and false negatives. The proof-of-concept is successful, though, and I think the results can be iteratively improved.

Install

cargo install cargo-mutants

Using cargo-mutants

Just run cargo mutants in a Rust source directory, and it will point out functions that may be inadequately tested:

% cargo mutants --dir ~/src/unix_mode/
baseline test with no mutations ... ok
replace type_bits with Default::default() in src/lib.rs:42:32 ... caught
replace is_file with Default::default() in src/lib.rs:52:35 ... caught
replace is_dir with Default::default() in src/lib.rs:62:34 ... caught
replace is_symlink with Default::default() in src/lib.rs:72:38 ... caught
replace is_fifo with Default::default() in src/lib.rs:77:35 ... caught
replace is_char_device with Default::default() in src/lib.rs:82:42 ... caught
replace is_block_device with Default::default() in src/lib.rs:87:43 ... NOT CAUGHT!
...

In this version of the unix_mode crate, the is_block_device function was indeed untested.

To see what mutants could be generated without running them, use --list. --list also supports a --json option to make the output more machine-readable, and a --diff option to show the replacement.

Understanding the results

If tests fail in a clean copy of the tree, there might be an (intermittent) failure in the source directory, or there might be some problem that stops them passing when run from a different location, such as a relative path in Cargo.toml. Fix this first.

Otherwise, cargo mutants generates every mutant it can and prints the result of trying each one:

• caught -- A test failed with this mutant applied. This is a good sign about test coverage. You can look in mutants.out/log to see which tests failed.

• not caught -- No test failed with this mutation applied, which seems to indicate a gap in test coverage. Or, it may be that the mutant is undistinguishable from the correct code. You may wish to add a better test, or mark that the function should be skipped.

• check failed -- cargo check failed on the mutated code, probably because the mutation does not typecheck. This is inconclusive about test coverage and no action is needed, but indicates an opportunity for cargo-mutants to either generate better mutants, or at least not generate unviable mutants.

• build failed -- Similarly, but cargo build failed. This should be rare.

Skipping functions

To mark functions so they are not mutated:

1. Add a Cargo dependency on the mutants crate.

2. Mark functions with #[mutants::skip].

The crate is tiny and the attribute has no effect on the compiled code. It only flags the function for cargo-mutants.

Exit codes

• 0: Success. No mutants were found that weren't caught by tests.

• 1: Usage error: bad command-line arguments etc.

• 2: Found some mutants that were not covered by tests.

• 3: Some tests timed out: possibly the mutatations caused an infinite loop, or the timeout is too low.

• 4: The tests are already failing or hanging before any mutations are applied, so no mutations were tested.

mutants.out

A mutants.out directory is created in the source directory. It contains:

• A logs/ directory, with one log file for each mutation plus the baseline unmutated case. The log contains the diff of the mutation plus the output from cargo.

• A mutants.json file describing all the generated mutants.

• An outcomes.json file describing the results of all tests.

Hangs and timeouts

Some mutations to the tree can cause the test suite to hang. For example, in this code, cargo-mutants might try changing should_stop to always return false:

    while !should_stop() {
      // something
    }

cargo mutants automatically sets a timeout when running tests with mutations applied, and reports mutations that hit a timeout. The automatic timeout is the maximum of 5 seconds, or 3x the time to run tests with no mutations.

You can also set an explicit timout with the --timeout option. In this case the timeout is also applied to tests run with no mutation.

The timeout does not apply to cargo check or cargo build, only cargo test.

When a test times out, you can mark it with #[mutants::skip] so that future cargo mutants runs go faster.

Tips

• Trees that deny style lints such as unused parameters are likely to fail to build when mutated, without really saying much about the value of the tests. I suggest you don't statically deny warnings in your source code, but rather set RUSTFLAGS when you do want to check this -- and don't do this when running cargo mutants.

Performance

Anything you can do to make the cargo build and cargo test suite faster will have a multiplicative effect on cargo mutants run time, and of course will also make normal development more pleasant. There's lots of good advice on the web.

In particular, on Linux, using the Mold linker can improve build times significantly: because cargo-mutants does many incremental builds, link time is important.

Hard-to-test cases

Some functions don't cause a test suite failure if emptied, but also cannot be removed. For example, functions to do with managing caches or that have other performance side effects.

Ideally, these should be tested, but doing so in a way that's not flaky can be difficult. cargo-mutants can help in a few ways:

• It helps to at least highlight to the developer that the function is not covered by tests, and so should perhaps be treated with extra care, or tested manually.
• A #[mutants::skip] annotation can be added to suppress warnings and explain the decision.
• Sometimes these effects can be tested by making the side-effect observable with, for example, a counter of the number of memory allocations or cache misses/hits.

Goals

cargo-mutants it easy to run on any Rust source tree, and will tell you something interesting about areas where bugs might be lurking or the tests might be insufficient.

Being easy to use means:

• It requires no changes to the source tree or other setup: just install and run. So, if it does not find anything interesting to say about a well-tested tree, it didn't cost you much.

• There's no effect on the operation of the program other than when run under cargo mutants.

• It is reasonably fast even on large Rust trees. The overall run time is, roughly, the product of the number of viable mutations multiplied by the time to run the test suite for each mutation. Typically, one cargo mutants run will give you all the information it can find about missing test coverage in the tree, and you don't need to run it again as you iterate on tests, so it's relatively OK if it takes a while.

• cargo-mutants should avoid generating "unviable" mutants that "obviously" won't compile, because that wastes time. However, when it's uncertain whether the mutant will build, it's worth trying things that might find interesting results even if they might fail to build. Over time, we expect to make it smarter about avoiding useless mutations and generating more interesting mutations.

• It runs correctly on any Rust source trees that are built and tested by Cargo, that will build and run their tests in a copy of the tree, and that have hermetic tests.

• cargo-mutants doesn't crash or hang, even if it generates mutants that cause the software under test to crash or hang.

• The results are reproducible, assuming the test suite is deterministic.

Interesting results mean:

• It tells you about places where the code could be wrong (or might already be wrong) and the test suite wouldn't catch it.

• Most, ideally all, findings should indicate something that really should be tested more, or that may already be buggy.

• It complements coverage tools by finding code that might be executed by a test (and show up as covered) but where the test result does not actually depend on the behavior of the code.

• It complements fuzzing or property testing by covering code that might be hard to hook up to a fuzzer interface, or where that work just has not been done yet.

• It's easy to understand what the output is telling you. It may take some thought about how to effectively test the under-tested code, but at least it's easy to see the potential bug that wouldn't be caught.

• Although run time matters, it's worth spending more time to generate more mutants that might find interesting results, even if some of them might not compile.

• Realistically, cargo-mutants may find some mutants that aren't caught by tests but also aren't interesting, or aren't feasible to test. In those cases it should be easy to permanently dismiss them (e.g. by adding a #[mutants::skip] attribute or a config file.)

• As much as possible it should avoid generating trivial mutants, where the mutated code is equivalent to the original code, and so it's not interesting that the test suite doesn't catch the change.

• On trees that are already very well-tested, cargo-mutants may find nothing interesting, and then it should just say so.

• And for trees that are thoroughly tested, you can use cargo mutants in CI to check that they remain so.

Limitations, caveats, known bugs, and future enhancements

• cargo-mutants has a limited repertoire of mutations it can generate. As this improves, it will generate more interesting results, and we expect this can incrementally improve over time.

• It also currently has a limited understanding of function return types, and so sometimes generates "unviable" mutants that won't build, which wastes some time. These also seem easy to improve. In particular:

  • It should skip functions with #[cfg(...)] attributes that don't match the current platform, but it does not yet.

  • It should also probably skip unsafe functions, and maybe functions containing unsafe {} blocks.

  • (There are several others.)

• cargo-mutants sees the AST of the tree but doesn't fully "understand" the types. Possibly it could learn to get type information from the compiler (or rust-analyzer?), which would help it generate more interesting viable mutants, and fewer unviable mutants.

• Copying the tree to build it doesn't work well if the Cargo.toml points to dependencies by a relative path (other than in subdirectories). This could be handled by an option to mutate in-place (maybe into a copy made by the user) or possibly an option to copy a larger containing directory. You can work around this by editing Cargo.toml to make the paths absolute, before running cargo mutants.

• Copying a Rust tree and its target/ directory seems to cause the first build to be slower than an incremental build in the source directory, even while mtimes are preserved. (Perhaps the path is part of the calculation whether files need to be rebuilt?) Later incremental builds are faster. sccache might help with this but I have not yet tested it. However, copying target/ is still generally faster than not copying it.

• To make this faster on large trees, we could keep several scratch trees and test them in parallel, which is likely to exploit CPU resources more thoroughly than Cargo's own parallelism: in particular Cargo tends to fall down to a single task during linking, and often comes down to running a single straggler test at a time.

• It currently assumes all the source is in src/ of the directory, but Cargo doesn't require that, and some crates have their source in a different directory. This could be fixed by reading cargo metadata.

• Mutated functions could discard all parameters to avoid strict warnings about them being unused. (I haven't seen any crates yet that enforce this.)

How it works

The basic approach is:

• First, run cargo build --tests and cargo check in the source tree to "freshen" it so that the mutated copies will have a good starting point.

• Make a copy of the whole tree into a scratch directory. The same directory is reused across all the mutations to benefit from incremental builds.

  • Before applying any mutations, check that cargo test succeeds in the scratch directory: perhaps a test is already broken, or perhaps the tree doesn't build when copied because it relies on relative paths to find dependencies, etc.

• Build a list of mutations:

  • Walk all source files and parse each one looking for functions.
  • Skip functions that should not be mutated for any of several reasons: because they're tests, because they have a #[mutants::skip] attribute, etc.
  • For each function, depending on its return type, generate every mutation pattern that produces a result of that type.

• For each mutation:

  • Apply the mutation to the scratch tree by patching the affected file.
  • Run cargo test in the tree, saving output to a log file.
  • If the build fails or the tests fail, that's good: the mutation was somehow caught.
  • If the build and tests succeed, that might mean test coverage was inadequate, or it might mean we accidentally generated a no-op mutation. (Doing so is a shortcoming in this tool.)
  • Revert the mutation to return the tree to its clean state.

The nice thing about Default is that it's defined on many commonly-used types including (), so cargo-mutants does not need to really understand the function return type at this early stage. Some functions will fail to build because they return a type that does not implement Default, and that's OK.

The file is parsed using the syn crate, but mutations are applied textually, rather than to the token stream, so that unmutated code retains its prior formatting, comments, line numbers, etc. This makes it possible to show a text diff of the mutation and should make it easier to understand any error messages from the build of the mutated code.

For more details, see DESIGN.md.

Related work

cargo-mutants was inspired by reading about the Descartes mutation-testing tool for Java described in Increment magazine's testing issue.

It's an interesting insight that mutation at the level of a whole function is a practical sweet-spot to discover missing tests, while (at least at moderate-size trees) still making it feasible to exhaustively generate every mutant.

Mutagen

There's an existing Rust mutation testing tool called Mutagen.

Some differences are:

• Mutagen seems more mature.

• Mutagen requires changes to the source tree, and for functions to be mutated to be marked with an attribute. cargo-mutants can work with any unmodified tree.

• Mutagen builds the tree only once; cargo-mutants does an incremental build for each mutation.

  This is slower, although for some trees the incremental build may be relatively cheap compared to running the test suite.

  On the up side building for each mutation gives cargo-mutants the freedom to try mutations it's not sure will compile.

• Mutagen has a neat system to use coverage information to run only the tests that could possibly be affected. This could potentially be ported.

• Mutagen needs a nightly compiler; cargo-mutants should work with any reasonably-recent compiler and is tested on stable.

• (Probably there are more. Please let me know.)

Stability

cargo-mutants is in alpha and behavior, output formats, command-line syntax, json output formats, etc, may change from one release to the next.