fix: per-endpoint rate limits, request timeout and input hardening

bal-server:
- Wire independent actix-governor buckets per endpoint class (pushtxs,
  searchtx, info/stats, default) so polling a cheap endpoint can no longer
  exhaust the push-txs budget; previously every limit was dead config
- Enforce BAL_SERVER_ACTIX_TIMEOUT_SECS via a request-timeout middleware that
  returns 408 Request Timeout, covering slow request bodies
- De-duplicate txids within a single /pushtxs request body
- Replace the byte-slice xpub detection (panic on non-ASCII) with a safe
  prefix check

bal-pusher:
- Handle malformed/empty ZMQ frames without indexing panics or UTF-8 expect
- Return an error from sign_message and skip the stats report instead of
  panicking when the private key cannot be parsed

validation / db:
- Reject IPv4-mapped IPv6 targets (::ffff:a.b.c.d) in is_valid_welist_url
  to close an SSRF bypass, with unit tests
- Use Path::starts_with for the SQLite forbidden-directory check so paths
  like /etcfoo are not misclassified

misc:
- Add explicit secret patterns to .gitignore and sync the docs
This commit is contained in:
2026-10-14 18:30:23 -04:00
parent 457c143954
commit b8d87ecabd
8 changed files with 224 additions and 55 deletions

View File

@@ -41,15 +41,20 @@ Example: `BAL_SERVER_REGTEST_ADDRESS=tpub...`, `BAL_SERVER_BITCOIN_FIXED_FEE=500
| `BAL_SERVER_ACTIX_WORKERS` | `4` | Number of actix-web worker threads |
| `BAL_SERVER_ACTIX_MAX_CONNECTIONS` | `100` | Maximum concurrent connections |
| `BAL_SERVER_TRUSTED_PROXY` | `127.0.0.1` | Trusted reverse proxy IP for rate-limiting client identification |
| `BAL_SERVER_ACTIX_PUSHTXS_PER_SEC` | `1` | Rate limit: requests per second (applied to all endpoints) |
| `BAL_SERVER_ACTIX_PUSHTXS_BURST` | `3` | Rate limit: burst size (applied to all endpoints) |
| `BAL_SERVER_ACTIX_PUSHTXS_PER_SEC` | `1` | Rate limit: push txs requests per second |
| `BAL_SERVER_ACTIX_PUSHTXS_BURST` | `3` | Rate limit: push txs burst size |
| `BAL_SERVER_ACTIX_SEARCHTX_PER_SEC` | `5` | Rate limit: searchtx requests per second |
| `BAL_SERVER_ACTIX_SEARCHTX_BURST` | `10` | Rate limit: searchtx burst size |
| `BAL_SERVER_ACTIX_INFO_PER_SEC` | `20` | Rate limit: info requests per second |
| `BAL_SERVER_ACTIX_INFO_BURST` | `30` | Rate limit: info burst size |
| `BAL_SERVER_ACTIX_DEFAULT_PER_SEC` | `50` | Rate limit: default requests per second |
| `BAL_SERVER_ACTIX_INFO_PER_SEC` | `20` | Rate limit: info/stats requests per second |
| `BAL_SERVER_ACTIX_INFO_BURST` | `30` | Rate limit: info/stats burst size |
| `BAL_SERVER_ACTIX_DEFAULT_PER_SEC` | `50` | Rate limit: default requests per second (`/`, `/.pub_key.pem`, `/version`) |
| `BAL_SERVER_ACTIX_DEFAULT_BURST` | `100` | Rate limit: default burst size |
Each endpoint class has its own independent token bucket, keyed by the real
client IP (see `BAL_SERVER_TRUSTED_PROXY`). `BAL_SERVER_ACTIX_TIMEOUT_SECS` is
enforced by a request-timeout middleware that returns `408 Request Timeout` if
the handler (including request body reading) exceeds the limit.
### `bal-pusher` (prefixed `BAL_PUSHER_`)
#### Core Settings

View File

@@ -36,7 +36,7 @@
**Description:** All `unwrap()`/`expect()` calls on critical paths have been replaced with safe error handling.
**Mitigation Applied:**
- `bal-server`: `from_utf8` returns 400, `sqlite::open` uses `open_db()` with validation, all handlers return proper HTTP status codes.
- `bal-pusher`: RPC failures log errors + sleep + retry (no panic), ZMQ `recv` uses `set_rcvtimeo(5000)`, `fs::read_to_string` uses `match` + 500, `cfg.lock()` uses `poisoned.into_inner()` recovery.
- `bal-pusher`: RPC failures log errors + sleep + retry (no panic), ZMQ `recv` uses `set_rcvtimeo(5000)`, malformed ZMQ frames are handled with `get()`/`first()` and `from_utf8_lossy` (no indexing or UTF-8 `expect`), stats signing returns an error that is logged and skipped rather than `expect`-panicking.
**Regression Tests:** `tests/panic_regression_tests.rs` (2 tests).
### 3. Secret Leakage (FIXED)
@@ -52,18 +52,20 @@
**Description:** All DoS vectors mitigated via actix-web migration.
**Mitigation Applied:**
- Body size limit: `PayloadConfig::default().limit(max_body_size)` via `BAL_SERVER_ACTIX_MAX_BODY_SIZE` (default 1 MiB).
- Rate limiting: `actix-governor` with token-bucket per client IP (`BAL_SERVER_ACTIX_PUSHTXS_PER_SEC`/`BURST`). Uses `RealIpKeyExtractor` to extract real client IP from proxy headers.
- Rate limiting: `actix-governor` with a token bucket per client IP, with an **independent bucket for each endpoint class** (push txs, searchtx, info/stats, default) so that polling a cheap endpoint cannot exhaust the push-txs budget. Uses `RealIpKeyExtractor` to extract the real client IP from proxy headers.
- Connection limits: `workers(4)` and `max_connections(100)` via `BAL_SERVER_ACTIX_WORKERS`/`MAX_CONNECTIONS`.
- Body timeout: configurable via `BAL_SERVER_ACTIX_TIMEOUT_SECS`.
- Request timeout: `middleware::from_fn` wraps every handler in `tokio::time::timeout`, returning `408 Request Timeout` when `BAL_SERVER_ACTIX_TIMEOUT_SECS` (default 5) elapses, covering slow request bodies.
- ZMQ timeout: `set_rcvtimeo(5000)` prevents infinite blocking.
- RPC retry: sleep + retry on connection failure instead of panic.
- ZMQ message handling is panic-free: multipart frames are accessed with `get()`/`first()` and topics are decoded with `from_utf8_lossy`, so malformed frames cannot abort the pusher.
- Stats-report signing failures are logged and skipped instead of panicking the pusher when the key is unreadable.
### 5. SSRF / Network Abuse via `reqwest` (FIXED)
**Severity:** MEDIUM | **Status:** Fixed
**Location:** `src/bin/bal-pusher.rs`, `src/validation.rs`
**Description:** URL validation prevents redirecting requests to internal/private IPs.
**Mitigation Applied:** `is_valid_welist_url()` in `src/validation.rs` blocks localhost, loopback, RFC1918, link-local, multicast, unspecified, and IPv6 unique-local addresses. HTTPS-only scheme enforced.
**Regression Tests:** `tests/ssrf_tests.rs` (integration) + 8 unit tests in `src/validation.rs`.
**Mitigation Applied:** `is_valid_welist_url()` in `src/validation.rs` blocks localhost, loopback, RFC1918, link-local, multicast, unspecified, IPv6 unique-local addresses, and **IPv4-mapped IPv6 addresses** (`::ffff:a.b.c.d`) that route to loopback/private/link-local/unspecified IPv4 targets. HTTPS-only scheme enforced. Note: hostnames are not DNS-resolved before validation, so a domain that resolves to a private address is not caught by this check — restrict `WELIST_SERVER_URL` to trusted values.
**Regression Tests:** `tests/ssrf_tests.rs` (integration) + 11 unit tests in `src/validation.rs`.
### 6. Insecure Database Access (FIXED)
**Severity:** MEDIUM | **Status:** Fixed
@@ -130,16 +132,17 @@
6. Use read-only filesystem for the server binary.
### Application-Level
1. **Rate Limiting:** Implemented via `actix-governor` with token-bucket per client IP. `RealIpKeyExtractor` identifies clients behind reverse proxy using `X-Real-IP` / `X-Forwarded-For` headers. Trusted proxy IP configurable via `BAL_SERVER_TRUSTED_PROXY`.
1. **Rate Limiting:** Implemented via `actix-governor` with token-bucket per client IP and a **separate bucket per endpoint class**. `RealIpKeyExtractor` identifies clients behind reverse proxy using `X-Real-IP` / `X-Forwarded-For` headers. Trusted proxy IP configurable via `BAL_SERVER_TRUSTED_PROXY`.
2. **Input Validation:** Network enum check, txid hex validation, body size limits.
3. **HTTPS:** Via Nginx reverse proxy with Let's Encrypt.
4. **WAL Mode:** Enabled with retry logic for concurrent access.
5. **ZMQ Timeout:** 5-second receive timeout prevents infinite blocking.
6. **Transaction Size Limits:** Configurable via rate limiting and body size.
7. **ZMQ Retry:** Reconnect logic with timeout-based detection.
8. **Fee Limits:** Per-network `fixed_fee` configuration.
9. **Network Limits:** Only known networks accepted (bitcoin, testnet, testnet4, signet, regtest).
10. **Locktime Reasonableness:** Locktime compared against blockchain height and median time with threshold-based distinction.
3. **Request Timeout:** `tokio::time::timeout` middleware returns `408` when a handler (including body read) exceeds `BAL_SERVER_ACTIX_TIMEOUT_SECS`.
4. **HTTPS:** Via Nginx reverse proxy with Let's Encrypt.
5. **WAL Mode:** Enabled with retry logic for concurrent access.
6. **ZMQ Timeout:** 5-second receive timeout prevents infinite blocking.
7. **Transaction Size Limits:** Configurable via rate limiting and body size.
8. **ZMQ Retry:** Reconnect logic with timeout-based detection.
9. **Fee Limits:** Per-network `fixed_fee` configuration.
10. **Network Limits:** Only known networks accepted (bitcoin, testnet, testnet4, signet, regtest).
11. **Locktime Reasonableness:** Locktime compared against blockchain height and median time with threshold-based distinction.
---
@@ -153,4 +156,4 @@
| `tests/secret_leakage_tests.rs` | 3 | .gitignore, no tracked secrets, no hardcoded tokens |
| `tests/input_validation_tests.rs` | 4 | Address caching, network validation, txid hex validation |
| `tests/db_path_validation.rs` | 5 | Path traversal, forbidden dirs, symlinks, WAL pragma, valid paths |
| `src/validation.rs` (inline) | 8 | SSRF URL validation unit tests |
| `src/validation.rs` (inline) | 11 | SSRF URL validation unit tests |