The detector already used the CLR data-directory RVA to split DLLs into NativeDll / ManagedDll; EXEs were a single undifferentiated Exe kind. Apply the same CLR check to EXEs so callers can tell a protected .NET executable from a native one without unpacking. Routing is unchanged: both EXE kinds go to the EXE pipeline. - CLI per-file lines and the run log now print NativeExe / ManagedExe (the kind comes from the same Debug formatting as the DLL variants). - The web API's detect()/unpack_file() kind strings become 'native-exe' / 'managed-exe'; the web UI gains matching labels, and the trap-retry guard (DLL-probe recovery) keys off both EXE kinds. Golden corpus unchanged (35/35 byte-identical); kind is classification only and never affects output bytes.
7.5 KiB
Design
Senbei is a fully static unpacker: it replays the unpacking algorithm on the file bytes in memory and writes the recovered PE image. No code from the protected binary is ever executed, no process is launched or attached to, and no driver or proxy DLL is involved.
Crate layout
The crate is split into a pure core and a thin CLI shell:
src/unpacker/— the core. Pure functions over byte slices: no file I/O, no environment access (beyond a few debugging overrides, see development.md), panic-free at the public boundary (all internal panics are trapped and converted toUnpackError::Corrupt). This is what the WebAssembly build embeds.src/(top level) — the CLI shell: argument parsing, recursive folder scanning, per-run log file, progress bar, Explorer-friendly exit pause, and the single-file/folder orchestration injob.rs.src/metadata.rs— il2cppglobal-metadata.datmethod-token de-obfuscation (format version 31; other versions are left untouched).
src/
├── main.rs argument parsing + dispatch
├── lib.rs module roots
├── job.rs single-file + folder orchestration, out-naming,
│ companion splice, stub overlay/TLS restore,
│ pipeline routing (incl. the wasm-safe byte API)
├── scan.rs recursive Crackproof + metadata discovery
├── metadata.rs il2cpp global-metadata.dat de-obfuscation
├── logfile.rs per-run timestamped log
├── ui.rs progress bar + status lines
├── pause.rs Explorer-friendly exit pause
└── unpacker/ pure, panic-free, no-I/O core
├── mod.rs detection + unpack_auto dispatch
├── exe.rs EXE pipeline (PE32+ and PE32)
├── dll.rs native + managed DLL pipeline
├── integrity.rs static post-unpack sanity check
├── primitives.rs decrypt_data* steps, key/shift selection
├── bytecode.rs bytecode VM
├── parallel.rs deterministic block-parallel fan-out
├── tables.rs constant tables
└── crc32.rs checksum
Detection and routing
Detection is content-based (unpacker::detect), never extension-based: the
key table is derived from the file header and checked against the format
magic, then the PE characteristics classify the input as EXE or DLL and the
CLR data directory splits each into native vs managed (NativeExe /
ManagedExe / NativeDll / ManagedDll).
unpack_auto then dispatches:
NativeExe/ManagedExe→ the EXE pipeline (handles both PE32+ and PE32). Managed EXEs take the same path: their import-string table is null (imports are the CLR bootstrap stub), the entry point comes from the protected header (the config block stores 0 for managed images), and the COR20 header, BSJB metadata stream, and CLR resources are restored verbatim from the protected file, mirroring the managed-DLL restore.NativeDll/ManagedDll→ the DLL pipeline first; on failure, the EXE pipeline as a fallback. Two DLL layouts exist in the wild: an older layout the DLL pipeline parses, and a newer one that protects DLLs with the EXE-style shell layout instead. The DLL-first order keeps old-layout outputs byte-identical (the EXE pipeline also "succeeds" on old-layout DLLs but produces different bytes); the fallback handles the new layout (including the managed-DLL .NET metadata restore).
One routing shortcut bypasses unpack_auto: inputs spliced from an external
companion (job.rs, both the CLI and the wasm byte API) go straight to the
EXE pipeline. The companion layout is definitionally the EXE-style shell,
so the DLL probe can never be right for it — and the probe's rejection of
EXE-shell DLLs relies on a caught panic, which is a fatal trap on targets
without unwinding (WebAssembly). Output bytes are identical to the
probe-then-fallback route.
External-companion inputs
Some builds split a protected module into an on-disk loader stub plus an
encrypted ._ companion. When a <name>._ sibling matches the stub's header
region, job.rs splices the two before unpacking and afterwards overlays the
export table and TLS directory from the stub — pieces the encrypted companion
does not carry. All overlay steps are best-effort no-ops when their inputs
can't be mapped, so a malformed stub can never corrupt an otherwise-good
unpack.
Pipelines
Both pipelines are heuristic with trial-and-validate: where a layout leaves ambiguity (e.g. which block is the real file decryptor, or a page-XOR shift), the pipeline tries candidates and validates the result structurally (an entry-stub oracle, checksum stamps, cluster stamps) instead of trusting the first match. A validation failure falls through to the next candidate rather than producing silently wrong output.
Several protected stages are themselves little bytecode programs. The core
includes a small VM (bytecode.rs) that generates and interprets those
programs rather than hardcoding each variant's constants.
Integrity check
Every produced image passes through integrity::check — a static, execution-
free sanity check that only flags defects impossible in a correctly unpacked
image (malformed headers, unmapped/non-executable/all-zero/all-int3 entry
point, a native DLL with no base-relocation directory, any import descriptor
whose DLL name is still ciphertext, a managed image whose COR20 header or BSJB
metadata did not survive). See usage.md.
A clean report is not a proof of correctness; a non-clean report is a reliable
"broken" signal.
Parallelism
Section decrypt/decompress blocks write disjoint output spans and read only
immutable input plus snapshotted key tables, so parallel.rs fans them out
across worker threads with byte-identical output regardless of thread
count. There is no unsafe: the buffer is carved with safe split_at_mut
chains so the borrow checker proves spans never alias. Overlapping spans (only
possible on corrupt input) degrade to the sequential whole-buffer pass,
preserving the deterministic last-writer-wins behavior of the serial
pipeline. SENBEI_THREADS=1 forces the sequential path; on targets without
threads (WebAssembly) the sequential path is used automatically.
Error model
The public API never panics: every pipeline runs under a catch_unwind
wrapper (catch_unpack) that converts a trapped panic to
UnpackError::Corrupt, with the default panic hook transiently suppressed.
Size requests are bounds-checked against a 1 GiB MAX_IMAGE_SIZE before
allocation so a crafted header cannot abort the process with a huge
allocation. In folder mode each file is isolated: one file's failure is logged
and counted, never fatal to the run.
WebAssembly caveat: the prebuilt wasm std cannot unwind, so a caught
panic becomes a fatal unreachable trap there. The DLL-routing probe relies
on this mechanism to reject EXE-shell-layout DLLs, so the web build routes
around it instead of through it: spliced companion inputs skip the probe
entirely (see "Detection and routing"), and the web app isolates every unpack
in a disposable Web Worker — a trapped DLL is retried once in a fresh worker
with the forced-EXE pipeline (job::unpack_bytes_force_exe), reproducing the
probe-then-fallback outcome without a catchable panic. A trap on any other
input is reported as a clean error rather than freezing the page.