fix: security hardening — 40 findings from full codebase review
Full codebase review by 4 independent agents (security, architecture,
code quality, correctness) identified ~80 findings. This commit fixes 40
of them across all workspace crates.
Critical fixes:
- Federation service: validate origin against mTLS cert CN/SAN (C1)
- WS bridge: add DM channel auth, size limits, rate limiting (C2)
- hpke_seal: panic on error instead of silent empty ciphertext (C3)
- hpke_setup_sender_and_export: error on parse fail, no PQ downgrade (C7)
Security fixes:
- Zeroize: seed_bytes() returns Zeroizing<[u8;32]>, private_to_bytes()
returns Zeroizing<Vec<u8>>, ClientAuth.access_token, SessionState.password,
conversation hex_key all wrapped in Zeroizing
- Keystore: 0o600 file permissions on Unix
- MeshIdentity: 0o600 file permissions on Unix
- Timing floors: resolveIdentity + WS bridge resolve_user get 5ms floor
- Mobile: TLS verification gated behind insecure-dev feature flag
- Proto: from_bytes default limit tightened from 64 MiB to 8 MiB
Correctness fixes:
- fetch_wait: register waiter before fetch to close TOCTOU window
- MeshEnvelope: exclude hop_count from signature (forwarding no longer
invalidates sender signature)
- BroadcastChannel: encrypt returns Result instead of panicking
- transcript: rename verify_transcript_chain → validate_transcript_structure
- group.rs: extract shared process_incoming() for receive_message variants
- auth_ops: remove spurious RegistrationRequest deserialization
- MeshStore.seen: bounded to 100K with FIFO eviction
Quality fixes:
- FFI error classification: typed downcast instead of string matching
- Plugin HookVTable: SAFETY documentation for unsafe Send+Sync
- clippy::unwrap_used: warn → deny workspace-wide
- Various .unwrap_or("") → proper error returns
Review report: docs/REVIEW-2026-03-04.md
152 tests passing (72 core + 35 server + 14 E2E + 1 doctest + 30 P2P)
This commit is contained in:
@@ -182,10 +182,21 @@ pub struct HookVTable {
|
||||
pub destroy: Option<unsafe extern "C" fn(user_data: *mut core::ffi::c_void)>,
|
||||
}
|
||||
|
||||
// Safety: user_data is an opaque pointer managed by the plugin. The plugin is
|
||||
// responsible for its own thread safety. The server only calls hook functions
|
||||
// one at a time per plugin (wrapped in a single Arc). Plugins that mutate
|
||||
// user_data through callbacks must use interior mutability.
|
||||
// SAFETY: `HookVTable` contains raw pointers (`user_data`, function pointers)
|
||||
// which are not inherently `Send`/`Sync`. These impls are sound because:
|
||||
//
|
||||
// 1. `user_data` is an opaque pointer managed entirely by the plugin. The plugin
|
||||
// contract (documented in the module-level doc comment) requires that plugins
|
||||
// use interior mutability (Mutex/RwLock) if `user_data` is mutated through
|
||||
// callbacks. The server wraps each loaded plugin in an `Arc<HookVTable>` and
|
||||
// may invoke hooks from any Tokio worker thread.
|
||||
//
|
||||
// 2. All function pointers are `unsafe extern "C" fn` — they are plain addresses
|
||||
// with no captured state. The code they point to must be thread-safe per the
|
||||
// plugin contract.
|
||||
//
|
||||
// 3. The server guarantees that `destroy` is called exactly once during shutdown,
|
||||
// after which no further hook calls are made on the vtable.
|
||||
#[allow(unsafe_code)]
|
||||
unsafe impl Send for HookVTable {}
|
||||
#[allow(unsafe_code)]
|
||||
|
||||
Reference in New Issue
Block a user