Rust Sanitizers Miri

Rust Sanitizers and Miri

Purpose

Guide agents through runtime safety validation for Rust: ASan/TSan/MSan/UBSan via RUSTFLAGS, Miri for compile-time UB detection in unsafe code, Android NDK HWASan for on-device validation, and interpreting sanitizer reports.

RIPDPI Project Context

This project has ~135 unsafe occurrences across 97 Rust crates. JNI interop crates (ripdpi-android, ripdpi-tunnel-android) use heavy FFI with the JVM. Other crates use raw pointers, libc syscalls, and platform-specific code. Miri cannot execute JNI/FFI code -- see the FFI caveat in section 4.

Triggers

• "How do I run AddressSanitizer on Rust code?"
• "How do I use Miri to check my unsafe Rust?"
• "How do I run ThreadSanitizer on a Rust program?"
• "My unsafe Rust might have UB -- how do I detect it?"
• "How do I run HWASan on Android native code?"

Workflow

1. Sanitizers in Rust (nightly required)

Rust sanitizers require nightly and a compatible platform:

# Install nightly
rustup toolchain install nightly
rustup component add rust-src --toolchain nightly

# AddressSanitizer (Linux, macOS)
RUSTFLAGS="-Z sanitizer=address" \
    cargo +nightly test -Zbuild-std \
    --target x86_64-unknown-linux-gnu

# ThreadSanitizer (Linux)
RUSTFLAGS="-Z sanitizer=thread" \
    cargo +nightly test -Zbuild-std \
    --target x86_64-unknown-linux-gnu

# MemorySanitizer (Linux, requires all-instrumented build)
RUSTFLAGS="-Z sanitizer=memory -Zsanitizer-memory-track-origins" \
    cargo +nightly test -Zbuild-std \
    --target x86_64-unknown-linux-gnu

-Zbuild-std rebuilds the standard library with the sanitizer, which is necessary for accurate results.

2. Interpreting ASan output in Rust

==12345==ERROR: AddressSanitizer: heap-buffer-overflow on address 0x602000000050
READ of size 4 at 0x602000000050 thread T0
    #0 0x401234 in myapp::module::function /src/main.rs:15
    #1 0x401567 in myapp::main /src/main.rs:42

Rust-specific patterns:

3. Android NDK ASan / HWASan

For on-device testing of JNI crates cross-compiled to Android NDK:

# HWASan (ARM64 only, Android 10+, preferred over ASan on ARM64)
# In .cargo/config.toml or via env:
RUSTFLAGS="-Z sanitizer=hwaddress" \
    cargo +nightly build -Zbuild-std \
    --target aarch64-linux-android

# ASan for Android (works on both ARM and x86 emulators)
RUSTFLAGS="-Z sanitizer=address" \
    cargo +nightly build -Zbuild-std \
    --target aarch64-linux-android

HWASan is preferred on ARM64: lower overhead than ASan, catches the same bugs, and is hardware-accelerated via top-byte-ignore (TBI). Requires Android 10+ and ARM64.

To run on-device:

1. Build the .so with the sanitizer flags above
2. Push to device and set LD_PRELOAD for the sanitizer runtime, or
3. Use Android Gradle plugin android.defaultConfig.externalNativeBuild with arguments "-DANDROID_STL=c++_shared" and enable HWASan in CMake

3a. MTE (Memory Tagging Extension) — Android 14+ on supporting SoCs

MTE is the hardware-accelerated successor to HWASan, available on arm64 Android 14+ with a supporting SoC (Pixel 8+, recent Samsung flagships). Unlike HWASan (software top-byte-ignore), MTE uses dedicated CPU instructions to tag heap allocations and check on access. The runtime cost in production is near-zero in async mode.

Activation (no Rust code changes)

In AndroidManifest.xml:

<application
    android:memtagMode="async"
    ... >

Modes:

• async — production-safe. Tag mismatches detected at some delay (typically next syscall); minimal latency cost.
• sync — debug only. Tag mismatches detected at the exact access; higher cost but precise.
• off — explicit disable.

For Rust code, MTE works transparently through bionic's allocator (jemalloc-based on Android). No RUSTFLAGS, no #![cfg_attr(sanitize, ...)], no rebuild. The allocator inserts tags; the CPU verifies; mismatches deliver SIGSEGV with si_code = SEGV_MTEAERR (async) or SEGV_MTESERR (sync).

Production vs debug trade-off

For RIPDPI: enable memtagMode="async" on release builds for arm64-v8a; the cost is well within the 5% size-baseline allowance. Keep HWASan available for CI runs on emulators (which typically don't expose MTE hardware) and for non-Pixel test devices.

Detection of MTE-caught crashes

Tombstone analysis: adb pull /data/tombstones/<latest> and grep for MTEAERR / MTESERR. The tombstone includes the tagged address and the access kind (read/write). For Rust-side debugging, pair with addr2line on the symbol-bearing build (see rust-debugging skill).

What MTE catches

Same UB class as HWASan: use-after-free on heap allocations, double-free, buffer overflow into adjacent tagged allocation, type confusion that crosses allocation boundaries. NOT caught: stack-based UAF (different mechanism — Stack-MTE is a separate, less-deployed extension), uninit reads (use MSan or Miri), data races (use TSan or loom).

For RIPDPI's hot-path code with raw pointers (ripdpi-runtime/platform/linux.rs, packet parsers using ptr::read_unaligned), MTE is the right production-grade safety net once the hardware supports it.

Rollout

1. Bump targetSdkVersion to 34+ (already done in RIPDPI).
2. Add android:memtagMode="async" to <application> in AndroidManifest.xml.
3. Run the full soak suite on a Pixel 8+ to verify no false positives surface.
4. Ship.

4. Miri -- interpreter for undefined behaviour

FFI LIMITATION: Miri cannot execute extern "C", JNI, libc syscalls, or inline assembly. This means JNI interop crates (ripdpi-android, ripdpi-tunnel-android) and any code calling libc directly cannot be tested under Miri without stubs. Use #[cfg(miri)] to provide mock implementations (see below).

# Install Miri (requires nightly)
rustup +nightly component add miri

# Run tests under Miri
cargo +nightly miri test

# Run specific test
cargo +nightly miri test test_name

# Strict provenance mode (recommended for CI)
MIRIFLAGS="-Zmiri-strict-provenance" cargo +nightly miri test

# Disable isolation (allow file I/O, randomness)
MIRIFLAGS="-Zmiri-disable-isolation" cargo +nightly miri test

Stubbing FFI for Miri

For crates with JNI or libc FFI, gate real implementations behind #[cfg(not(miri))] and provide stubs:

#[cfg(not(miri))]
extern "C" {
    fn platform_specific_call(fd: i32) -> i32;
}

#[cfg(miri)]
unsafe fn platform_specific_call(_fd: i32) -> i32 {
    0 // safe stub for Miri interpretation
}

// For JNI: skip JNI tests entirely under Miri
#[test]
fn test_pure_logic() {
    // This runs under Miri -- no JNI calls
}

#[test]
#[cfg_attr(miri, ignore)]
fn test_jni_integration() {
    // This is skipped under Miri
}

5. What Miri detects

• Dangling pointers: use after free, use after reallocation
• Invalid values: bad enum discriminants, non-0/1 bools, unaligned refs
• Uninitialized memory: reading MaybeUninit before init
• Stacked Borrows / Tree Borrows violations: aliasing rule breaches
• Data races: Miri has a dedicated concurrency model that interleaves threads at yield points, detecting unsynchronized accesses to shared state. This is independent of the aliasing model (Stacked Borrows / Tree Borrows).

6. Miri configuration via MIRIFLAGS

7. CI integration

# GitHub Actions
- name: Miri
  run: |
    rustup toolchain install nightly
    rustup +nightly component add miri
    cargo +nightly miri test
  env:
    MIRIFLAGS: "-Zmiri-disable-isolation -Zmiri-strict-provenance"

- name: ASan (nightly)
  run: |
    rustup component add rust-src --toolchain nightly
    RUSTFLAGS="-Z sanitizer=address" \
    cargo +nightly test -Zbuild-std \
    --target x86_64-unknown-linux-gnu

Aliasing assumptions transit via Box

Severity: WARNING when mixing Box<T> with raw pointer FFI

Box<T> carries Unique<T> semantics: the compiler assumes that the data inside is exclusively owned by the Box and that no other pointer aliases it (the noalias LLVM attribute). If you extract a *mut T from a Box, pass it to FFI, and the FFI code stores it alongside the live Box, you have aliasing — Box and the raw pointer both claim unique access.

Concrete failure mode:

let mut boxed = Box::new(MyStruct::new());
let raw: *mut MyStruct = &mut *boxed as *mut _;
unsafe { ffi_register(raw); }   // FFI stores `raw`
boxed.field = 42;               // Box load — LLVM may reorder past FFI store
// `raw` and `boxed` now alias — Tree Borrows flags this

Correct patterns:

• Use Box::into_raw to transfer ownership to FFI; never use the Box again.
• If FFI must borrow a pointer from Rust, go through Pin<Box<T>> so Rust knows the address is stable and does not optimize based on noalias.
• Run MIRIFLAGS="-Zmiri-tree-borrows" to catch this — Tree Borrows (PLDI 2025) is more precise than Stacked Borrows and surfaces this aliasing violation earlier.

Reference: crabbook/raii_and_memory_safety.md (footnote on Box/noalias interaction)

Related skills

• rust-debugging -- GDB/LLDB debugging of Rust panics
• rust-unsafe -- unsafe Rust patterns and review checklist
• rust-security -- supply chain safety and memory-safe development
• rust-ffi -- C interoperability and safe FFI wrappers
• memory-model -- atomics, memory ordering, lock-free structures

Source: po4yka/RIPDPI — distributed by TomeVault.