M0031 fuzz-corpus-test¶
Rule: Every fuzz target has a corpus tracked in git and a test that replays it.
| Code | M0031 (stable, never reused) |
| Id | fuzz-corpus-test |
| Category | repo |
| Source | spirit_level |
| Fix | by hand |
Level per tier¶
| Tier | Level | Reason |
|---|---|---|
| small | off | |
| small-workspace | off | |
| medium | warn |
a fuzz target's interesting inputs are a regression suite only once they are in git and replayed on every ordinary test run |
| big | warn |
a fuzz target's interesting inputs are a regression suite only once they are in git and replayed on every ordinary test run |
| ai | off | |
| embedded | off | |
| ui | off | |
| egui | off | |
| units-strict | off | |
| units-paranoid | off |
Why¶
A fuzzer earns nothing if the inputs it found are thrown away. SQLite and TigerBeetle keep their interesting fuzz
cases as files and replay them on every ordinary test run, so a bug a fuzzer took hours to find never comes back
silently. A fuzz/fuzz_targets/*.rs with no corpus in git, or no test replaying it, has already lost that
regression suite: cargo fuzz run alone needs a fuzzing engine nobody runs by hand before every commit.
How to fix¶
- Commit a handful of minimised seed inputs under
fuzz/corpus/<target>/(the file names cargo-fuzz itself would write). - Add an ordinary
#[test]whose name mentions the target, reading every file in that corpus directory and calling the target's entry function directly, no fuzzing engine involved:
#[test]
fn replays_parse_bytes_corpus() {
for entry in std::fs::read_dir("fuzz/corpus/parse_bytes").unwrap() {
let data = std::fs::read(entry.unwrap().path()).unwrap();
parse_bytes::run(&data);
}
}
The lint only checks that a corpus directory is tracked and that some #[test] names the target; it cannot see
whether the test actually replays the corpus.
Configure¶
Allow one finding in the code with // spirit-level: allow(fuzz-corpus-test) <reason> (waivers); the code works there too, as in [lints.M0031].