lez-fuzzing/docs/fuzzing.md
Roman 8bd0a1a612
fix: add new fuzz targets
- template for adding targets
2026-04-15 15:47:01 +08:00

276 lines
9.0 KiB
Markdown

# Fuzzing Guide
This document covers how to run fuzz targets, add new targets, minimise failures,
and convert findings into regression tests.
The fuzzing infrastructure lives in a **separate repository** (`lez-fuzzing/`) which
reads the Logos Execution Zone (LEZ) codebase from `../logos-execution-zone/` (a sibling
directory that must be cloned separately).
---
## Prerequisites
```bash
# Rust nightly is required by cargo-fuzz / libFuzzer
rustup install nightly
rustup component add llvm-tools-preview --toolchain nightly
cargo install cargo-fuzz
```
---
## Repository Setup
`lez-fuzzing` is a **standalone repository** — it does **not** use git submodules.
It expects the LEZ codebase to be cloned at `../logos-execution-zone` relative to itself.
```bash
# Clone both repositories side-by-side into the same parent directory:
git clone <LEZ_REPO_URL> logos-execution-zone
git clone <LEZ_FUZZING_REPO_URL> lez-fuzzing
# The directory layout must be:
# <parent>/
# ├── logos-execution-zone/
# └── lez-fuzzing/
```
---
## How to Run
All fuzz targets must be run with `RISC0_DEV_MODE=1` to disable expensive ZK
proof generation. The `just` recipes handle this automatically.
```bash
# From lez-fuzzing/
# Run all targets for 30 s each
just fuzz
# Run a specific target for 120 s
RISC0_DEV_MODE=1 cargo fuzz run fuzz_state_transition -- -max_total_time=120
# Run the saved corpus (regression mode, no mutations)
just fuzz-regression
```
---
## Available Fuzz Targets
| Target | What it fuzzes | Entry point |
|--------|---------------|-------------|
| `fuzz_transaction_decoding` | borsh decoding of all transaction and block types | `fuzz/fuzz_targets/fuzz_transaction_decoding.rs` |
| `fuzz_stateless_verification` | `transaction_stateless_check()` signature validation | `fuzz/fuzz_targets/fuzz_stateless_verification.rs` |
| `fuzz_state_transition` | `V03State::transition_from_*()` with invariant checks | `fuzz/fuzz_targets/fuzz_state_transition.rs` |
| `fuzz_block_verification` | Block hash integrity + replayer pipeline | `fuzz/fuzz_targets/fuzz_block_verification.rs` |
---
## How to Add a New Fuzz Target
### Step 1 — Scaffold with `just new-target`
```bash
just new-target my_feature
```
This single command does four things automatically:
| What | Where |
|---|---|
| Creates the corpus directory | `fuzz/corpus/fuzz_my_feature/` |
| Writes a typed fuzz target template | `fuzz/fuzz_targets/fuzz_my_feature.rs` |
| Appends `[[bin]]` entry | `fuzz/Cargo.toml` |
| Inserts target into every CI matrix + perf loop | `.github/workflows/fuzz.yml` |
The generated template uses `ArbNSSATransaction` from `fuzz_props::arbitrary_types`
so libfuzzer drives every field of `NSSATransaction` independently — no manual
`Unstructured` wiring required.
### Step 2 — Implement the target
Edit `fuzz/fuzz_targets/fuzz_my_feature.rs`. Replace the placeholder with the
function under test and any invariant assertions. Use the typed wrappers from
[`fuzz_props::arbitrary_types`](../fuzz_props/src/arbitrary_types.rs) for
structured input, or the proptest generators from
[`fuzz_props::generators`](../fuzz_props/src/generators.rs) for richer strategies.
### Step 3 — Register the binary (automated)
`just new-target` calls [`scripts/add_fuzz_target.py`](../scripts/add_fuzz_target.py)
which appends the `[[bin]]` entry to [`fuzz/Cargo.toml`](../fuzz/Cargo.toml)
automatically. Once present, `cargo fuzz list` (and therefore `just fuzz`,
`just fuzz-regression`, `just corpus-cmin`) pick up the target automatically — no
further Justfile edits required.
> **Manual fallback:** if you create a target without `just new-target`, add the
> entry yourself:
>
> ```toml
> [[bin]]
> name = "fuzz_my_feature"
> path = "fuzz_targets/fuzz_my_feature.rs"
> test = false
> bench = false
> ```
### Step 4 — Add to CI matrix (automated)
`just new-target` also inserts `fuzz_my_feature` into every strategy matrix and the
perf-baseline shell loop in [`.github/workflows/fuzz.yml`](../.github/workflows/fuzz.yml)
automatically via `scripts/add_fuzz_target.py`.
> **Manual fallback:** if you created the target without `just new-target`, add
> `- fuzz_my_feature` to the `target:` list in the three places shown in
> `.github/workflows/fuzz.yml` (smoke-fuzz, regression, perf-baseline).
### Step 5 — Verify
```bash
RISC0_DEV_MODE=1 cargo fuzz build fuzz_my_feature
just fuzz-regression # runs the new target against its (empty) corpus
```
### Quick reference: what to touch
| File | Action | Automated? |
|---|---|---|
| `fuzz/fuzz_targets/fuzz_<name>.rs` | Create | ✅ `just new-target` |
| `fuzz/corpus/fuzz_<name>/` | Create | ✅ `just new-target` |
| `fuzz/Cargo.toml` | Add `[[bin]]` | ✅ `just new-target` |
| `Justfile` | Nothing — auto-discovers | ✅ automatic |
| `.github/workflows/fuzz.yml` | Add to 3 matrix lists | ✅ `just new-target` |
---
## Updating the LEZ Dependency
`lez-fuzzing` reads LEZ source directly from `../logos-execution-zone`. To pick up LEZ
changes, simply update that repo:
```bash
cd ../logos-execution-zone
git pull --ff-only
cd ../lez-fuzzing
# Rebuild to confirm compatibility:
cargo build -p fuzz_props
RISC0_DEV_MODE=1 cargo fuzz build
```
The `just update-lez` recipe automates the pull:
```bash
just update-lez
```
---
## Minimising & Reproducing Failures
When `cargo fuzz` finds a crash it writes an artifact to
`fuzz/artifacts/fuzz_<target>/crash-<hash>`.
### Minimise
```bash
# Produces a smaller input that still triggers the same crash
just fuzz-tmin fuzz_state_transition fuzz/artifacts/fuzz_state_transition/crash-abc123
```
### Convert to a regression test
```bash
# Print the bytes as a Rust byte-literal (paste into a #[test])
cargo fuzz fmt fuzz_state_transition fuzz/artifacts/fuzz_state_transition/crash-abc123
```
Add the minimised file to the corpus so CI always reproduces it:
```bash
cp fuzz/artifacts/fuzz_state_transition/crash-abc123-minimised \
fuzz/corpus/fuzz_state_transition/regression_001
```
Open a PR. The `regression` CI job will permanently block re-introduction of this bug.
---
## Invariant Framework
Shared invariants live in `fuzz_props/src/invariants.rs`. Each invariant implements
`ProtocolInvariant` and is automatically run by `assert_invariants()`.
To add a new invariant:
1. Add a zero-size struct implementing `ProtocolInvariant`.
2. Register it in the `invariants` slice inside `assert_invariants()`.
3. Write a `#[test]` in `fuzz_props` that triggers and detects a synthetic violation.
---
## Performance Baseline
Measured on a 4-core x86_64 Linux runner with `RISC0_DEV_MODE=1`:
| Target | Throughput |
|--------|-----------|
| `fuzz_transaction_decoding` | ~200 000 exec/sec |
| `fuzz_stateless_verification` | ~30 000 exec/sec |
| `fuzz_state_transition` | ~5 000 exec/sec |
| `fuzz_block_verification` | ~50 000 exec/sec |
Recommended local settings for longer runs:
```bash
# Use all available cores
RISC0_DEV_MODE=1 cargo fuzz run fuzz_state_transition \
-- -max_total_time=3600 -jobs=$(nproc) -workers=$(nproc)
```
---
## ZK-Proof Cost Warning
`PrivacyPreservingTransaction` uses `risc0-zkvm` (seconds per proof).
All fuzz targets **must** set `RISC0_DEV_MODE=1` in the environment and the `just`
recipes handle this automatically via:
```just
export RISC0_DEV_MODE := "1"
```
Do **not** invoke full proof generation inside any fuzz target. The `RISC0_DEV_MODE=1`
flag stubs out ZK proof generation and replaces it with a fast mock implementation.
---
## Input Generators
The `fuzz_props` crate (`fuzz_props/src/generators.rs`) provides reusable input
generators for both `libfuzzer` (via `arbitrary`) and `proptest`:
| Generator | Covers |
|-----------|--------|
| `arbitrary_transaction()` | IS-2: malformed + boundary transactions |
| `arb_borsh_transaction_bytes()` | IS-2: raw borsh bytes including invalid encodings |
| `arb_invalid_account_state_tx()` | IS-3: phantom accounts + overflow amounts |
| `arb_duplicate_tx_sequence()` | IS-4: duplicated + re-ordered transaction sequences |
| `arb_pathological_sequence()` | IS-5: zero-value, self-transfer, max-nonce inputs |
---
## Known Limitations & Future Work
| Item | Notes |
|------|-------|
| `PrivacyPreservingTransaction` coverage | Currently only exercised in decoding target; a dedicated slow target with `RISC0_DEV_MODE=1` and `proptest` should be added after the four MVP targets are stable |
| `V03State` snapshot equality | If `V03State` does not implement `PartialEq`/`Clone`, implement or derive them in `lez/nssa/src/state.rs` behind a `cfg(any(test, feature = "fuzzing"))` guard |
| AFL++ integration | A `just fuzz-afl` recipe can be added later; the same corpus is compatible |
| Differential testing (sequencer vs replayer) | Add a fifth target that feeds the same block to `SequencerCore` and `indexer_core` and asserts identical state roots |
| LEZ version tracking | There is no submodule pin — `lez-fuzzing` reads `../logos-execution-zone` as checked out. Update that repo to a release tag or a tested commit, then run `just update-lez` (which does `git pull --ff-only`) and open a PR to bump it. |