hub: make the login lockout that LOGIN_FAIL_MAX described

LOGIN_FAIL_MAX was declared as "max failures in window before extra delay" and then never read anywhere in the file. _record_failure() only ever slept a flat LOGIN_FAIL_DELAY. So the Hub's entire defence against online guessing was a constant 2 second pause per wrong password: no escalation, no lockout, no ban, and fail2ban is configured for SSH only. Worse, the old 60 second window could not have worked even if the constant had been wired up. With a delay per attempt, reaching 10 failures takes about 80 seconds, so the earliest failures aged out of the window before the count could ever reach the limit. - security_helpers.py: new LoginThrottle. The delay ramps with the failure count (2s, 4s, ... capped at 10s), and once LOGIN_FAIL_MAX failures land inside the window the address is refused outright for LOGIN_LOCKOUT_SECONDS (5 minutes). A successful login clears the address, so an operator who fumbles a password is not penalised later. The sleep is never taken under the lock, so one slow client cannot stall every other login. Tracked addresses are evicted, so a distributed sweep cannot grow the table without bound. The window moves from 60s to 900s so the whole ramp fits inside it. clock and sleep are injectable, which is what makes it testable. - server.py: /api/login checks the lockout before the scrypt hash, so a locked-out client costs almost nothing to reject, and answers 429 with a human-readable wait instead of a bare 401. - tests/test_login_throttle.py: covers the ramp, the cap, the lockout firing and expiring, per-address isolation, clearing on success, eviction, and that the limit is actually reachable inside the window. Verified against the real app with TestClient: 10 wrong passwords return 401 and the 11th returns 429 "Too many failed attempts. Try again in about 5 minute(s)." A correct password clears the counter.
2026-10-02 02:24:29 -05:00
parent c33457fff2
commit 7d784eb653
3 changed files with 375 additions and 14 deletions
@@ -15,6 +15,7 @@ import json
 import os
 import re
 import tempfile
+import threading
 import time
 import urllib.parse

@@ -245,6 +246,141 @@ def _validate_ssh_pubkey(key: str) -> str:
    return key


+# ── Login throttling ──────────────────────────────────────────────────────────
+#
+# Delays applied after each failed login, and the lockout that follows once an
+# address has tripped LOGIN_FAIL_MAX inside the window.
+#
+# LOGIN_FAIL_WINDOW has to be longer than the time it takes to reach
+# LOGIN_FAIL_MAX failures under the ramping delay: with a 2s ramp capped at
+# LOGIN_FAIL_MAX_DELAY, 10 attempts take about 80 seconds, so a 60 second
+# window would silently expire the earliest failures and the counter could
+# never reach the limit. 900s (15 minutes) keeps the whole ramp inside it.
+LOGIN_FAIL_DELAY = 2.0         # base delay; the nth failure waits n x this
+LOGIN_FAIL_MAX_DELAY = 10.0    # ceiling for a single delay
+LOGIN_FAIL_WINDOW = 900.0      # rolling window failures are counted in
+LOGIN_FAIL_MAX = 10            # failures in the window that trigger a lockout
+LOGIN_LOCKOUT_SECONDS = 300.0  # how long the lockout lasts
+
+# Cap on how many addresses are tracked, so a distributed sweep cannot grow
+# the table without bound.
+_LOGIN_THROTTLE_MAX_IPS = 4096
+
+
+class LoginThrottle:
+    """Per-address failed-login tracking with a ramping delay and a lockout.
+
+    The delay ramps so a script hammering the login form slows down as it goes,
+    and once LOGIN_FAIL_MAX failures land inside the window the address is
+    refused outright for LOGIN_LOCKOUT_SECONDS. A successful login clears the
+    address so a legitimate user who fumbles a password is not penalised later.
+
+    ``sleep`` and ``clock`` are injectable so tests run without waiting.
+    """
+
+    def __init__(
+        self,
+        fail_delay=LOGIN_FAIL_DELAY,
+        max_delay=LOGIN_FAIL_MAX_DELAY,
+        window=LOGIN_FAIL_WINDOW,
+        max_failures=LOGIN_FAIL_MAX,
+        lockout=LOGIN_LOCKOUT_SECONDS,
+        max_tracked_ips=_LOGIN_THROTTLE_MAX_IPS,
+        sleep=None,
+        clock=None,
+    ):
+        self._fail_delay = float(fail_delay)
+        self._max_delay = float(max_delay)
+        self._window = float(window)
+        self._max_failures = int(max_failures)
+        self._lockout = float(lockout)
+        self._max_tracked_ips = int(max_tracked_ips)
+        self._sleep = sleep if sleep is not None else time.sleep
+        self._clock = clock if clock is not None else time.monotonic
+        self._lock = threading.Lock()
+        self._failures: dict[str, list[float]] = {}
+
+    # ── internals ────────────────────────────────────────────────────────────
+
+    def _prune(self, ip, now):
+        """Drop timestamps outside the window; return what is left."""
+        keep = [t for t in self._failures.get(ip, ()) if now - t < self._window]
+        if keep:
+            self._failures[ip] = keep
+        else:
+            self._failures.pop(ip, None)
+        return keep
+
+    def _evict(self, now):
+        """Forget addresses that can no longer affect anything."""
+        horizon = max(self._window, self._lockout)
+        for ip in [i for i, ts in self._failures.items()
+                   if ts and now - max(ts) > horizon]:
+            self._failures.pop(ip, None)
+        while len(self._failures) > self._max_tracked_ips:
+            oldest = min(self._failures, key=lambda i: max(self._failures[i]))
+            self._failures.pop(oldest, None)
+
+    # ── public API ───────────────────────────────────────────────────────────
+
+    def delay_for(self, count):
+        """Return the delay owed after *count* failures in the current window."""
+        if count <= 0:
+            return 0.0
+        return min(self._fail_delay * count, self._max_delay)
+
+    def failure_count(self, ip):
+        """Return the failures currently counted against *ip*."""
+        with self._lock:
+            return len(self._prune(ip, self._clock()))
+
+    def is_locked_out(self, ip):
+        """Return True while *ip* is inside a lockout."""
+        now = self._clock()
+        with self._lock:
+            failures = self._prune(ip, now)
+            if len(failures) < self._max_failures:
+                return False
+            return (now - failures[-1]) < self._lockout
+
+    def remaining_lockout(self, ip):
+        """Return the seconds left in *ip*'s lockout, or 0.0 if not locked out."""
+        now = self._clock()
+        with self._lock:
+            failures = self._prune(ip, now)
+            if len(failures) < self._max_failures:
+                return 0.0
+            return max(0.0, self._lockout - (now - failures[-1]))
+
+    def record_failure(self, ip):
+        """Record a failure for *ip* and serve out the delay it has earned.
+
+        Returns the delay that was applied. The lock is never held across the
+        sleep, so one slow client cannot stall every other login.
+        """
+        now = self._clock()
+        with self._lock:
+            failures = list(self._prune(ip, now))
+            failures.append(now)
+            self._failures[ip] = failures
+            count = len(failures)
+            self._evict(now)
+        delay = self.delay_for(count)
+        if delay > 0:
+            self._sleep(delay)
+        return delay
+
+    def clear(self, ip):
+        """Forget *ip*, e.g. after a successful login."""
+        with self._lock:
+            self._failures.pop(ip, None)
+
+    def tracked_addresses(self):
+        """Return how many addresses are currently being tracked."""
+        with self._lock:
+            return len(self._failures)
+
+
 # ── Persistent Hub session store ─────────────────────────────────────────────

 def load_session_store(path: str) -> dict[str, float]:
@@ -55,6 +55,12 @@ from .security_helpers import (
    _bech32_convertbits_decode,
    load_session_store,
    save_session_store,
+    LoginThrottle,
+    LOGIN_FAIL_DELAY,
+    LOGIN_FAIL_MAX_DELAY,
+    LOGIN_FAIL_WINDOW,
+    LOGIN_FAIL_MAX,
+    LOGIN_LOCKOUT_SECONDS,
 )
 from .update_state import effective_update_status

@@ -175,11 +181,19 @@ _sessions_lock = Lock()
 _SESSION_PERSIST_MIN_INTERVAL = 30.0  # seconds
 _sessions_last_persist = 0.0

-# Failed login tracking: ip → list of failure timestamps
-_login_failures: dict[str, list[float]] = {}
-LOGIN_FAIL_DELAY  = 2.0   # seconds to sleep after a failed attempt
-LOGIN_FAIL_WINDOW = 60.0  # rolling window (seconds) for counting failures
-LOGIN_FAIL_MAX    = 10    # max failures in window before extra delay
+# Failed login tracking.
+#
+# LOGIN_FAIL_MAX used to be declared here and never read anywhere: the only
+# thing a failed attempt cost an attacker was a flat 2 second delay, and there
+# was no lockout, no escalation and no ban. The throttling now lives in
+# security_helpers.LoginThrottle, which ramps the delay and refuses an address
+# outright once it has tripped LOGIN_FAIL_MAX inside the window.
+#
+# The window moved from 60s to 900s. With the ramping delay, reaching
+# LOGIN_FAIL_MAX takes about 80 seconds, so a 60 second window expired the
+# earliest failures before the limit could ever be reached — the old constant
+# could not have worked even if it had been wired up.
+_login_throttle = LoginThrottle()

 # Public paths that are accessible without a valid session
 _AUTH_EXEMPT_PATHS = {"/login", "/api/login", "/auto-login", "/api/ping"}
@@ -770,19 +784,20 @@ def _ensure_onboarding_reopened_for_migration() -> None:
        logger.warning("Could not clear onboarding flag for migration flow: %s", exc)


-def _record_failure(client_ip: str) -> None:
-    """Record a failed login attempt and apply a rate-limit delay.
+def _record_failure(client_ip: str) -> float:
+    """Record a failed login attempt and apply the throttling delay.

    Must always be called via loop.run_in_executor() so that the blocking
    time.sleep() does not stall the asyncio event loop.
+
+    Returns the delay that was applied.
    """
-    now = time.time()
-    failures = _login_failures.setdefault(client_ip, [])
-    # Prune old entries outside the window
-    _login_failures[client_ip] = [t for t in failures if now - t < LOGIN_FAIL_WINDOW]
-    _login_failures[client_ip].append(now)
-    # Sleep in the thread-pool thread to slow brute-force without blocking the loop
-    time.sleep(LOGIN_FAIL_DELAY)
+    return _login_throttle.record_failure(client_ip)
+
+
+def _is_locked_out(client_ip: str) -> bool:
+    """Return True while *client_ip* is inside a lockout."""
+    return _login_throttle.is_locked_out(client_ip)


 # ── Authentication middleware ─────────────────────────────────────
@@ -2638,10 +2653,24 @@ async def api_login(req: LoginRequest, request: Request):
    """Validate the Hub password and issue a session cookie."""
    client_ip = request.client.host if request.client else "unknown"
    loop = asyncio.get_event_loop()
+
+    # Refuse outright while the address is locked out. This runs before the
+    # scrypt hash, so a locked-out client costs almost nothing to reject.
+    if _is_locked_out(client_ip):
+        remaining = int(_login_throttle.remaining_lockout(client_ip) // 60) + 1
+        raise HTTPException(
+            status_code=429,
+            detail=f"Too many failed attempts. Try again in about {remaining} minute(s).",
+        )
+
    ok = await loop.run_in_executor(None, _check_password, req.password)
    if not ok:
        await loop.run_in_executor(None, _record_failure, client_ip)
        raise HTTPException(status_code=401, detail="Incorrect password")
+
+    # A real login clears the address, so fumbling a password once in a while
+    # does not accumulate towards a lockout.
+    _login_throttle.clear(client_ip)
    token = _create_session()
    response = JSONResponse({"ok": True})
    response.set_cookie(
@@ -0,0 +1,196 @@
+"""Tests for the Hub's login throttling.
+
+These exercise the exact production implementation in
+sovran_systemsos_web.security_helpers.LoginThrottle. The clock and the sleep are
+injected, so the tests cover hours of lockout behaviour instantly.
+
+No network access, no filesystem writes, no real delays.
+"""
+
+import os
+import sys
+import unittest
+
+_REPO_ROOT = os.path.normpath(os.path.join(os.path.dirname(__file__), ".."))
+_APP_PARENT = os.path.join(_REPO_ROOT, "app")
+if _APP_PARENT not in sys.path:
+    sys.path.insert(0, _APP_PARENT)
+
+from sovran_systemsos_web.security_helpers import (  # noqa: E402
+    LoginThrottle,
+    LOGIN_FAIL_DELAY,
+    LOGIN_FAIL_MAX_DELAY,
+    LOGIN_FAIL_WINDOW,
+    LOGIN_FAIL_MAX,
+    LOGIN_LOCKOUT_SECONDS,
+)
+
+
+class FakeClock:
+    """A clock that only moves when the test says so."""
+
+    def __init__(self):
+        self.now = 1000.0
+
+    def __call__(self):
+        return self.now
+
+    def advance(self, seconds):
+        self.now += seconds
+
+
+class FakeSleeper:
+    """Records the delays it was asked to apply instead of sleeping."""
+
+    def __init__(self, clock):
+        self.clock = clock
+        self.calls = []
+
+    def __call__(self, seconds):
+        self.calls.append(seconds)
+        self.clock.advance(seconds)
+
+
+def _make(**kwargs):
+    clock = kwargs.pop("clock", None) or FakeClock()
+    sleep = kwargs.pop("sleep", None) or FakeSleeper(clock)
+    return LoginThrottle(clock=clock, sleep=sleep, **kwargs), clock, sleep
+
+
+def _trip(throttle, ip="203.0.113.9"):
+    """Fail LOGIN_FAIL_MAX times. The fake sleeper advances the clock for us."""
+    for _ in range(LOGIN_FAIL_MAX):
+        throttle.record_failure(ip)
+
+
+class DelayRamp(unittest.TestCase):
+
+    def test_delay_ramps_with_the_failure_count(self):
+        throttle, _, _ = _make()
+        self.assertEqual(throttle.delay_for(0), 0.0)
+        self.assertEqual(throttle.delay_for(1), LOGIN_FAIL_DELAY)
+        self.assertEqual(throttle.delay_for(3), LOGIN_FAIL_DELAY * 3)
+
+    def test_delay_is_capped(self):
+        # Unbounded ramping would let a single client park a thread-pool worker
+        # for minutes at a time.
+        throttle, _, _ = _make()
+        self.assertLessEqual(throttle.delay_for(999), LOGIN_FAIL_MAX_DELAY)
+        self.assertEqual(throttle.delay_for(999), LOGIN_FAIL_MAX_DELAY)
+
+    def test_first_failure_is_not_delayed_much(self):
+        throttle, _, sleep = _make()
+        delay = throttle.record_failure("203.0.113.9")
+        self.assertEqual(delay, LOGIN_FAIL_DELAY)
+        self.assertEqual(sleep.calls, [LOGIN_FAIL_DELAY])
+
+
+class Lockout(unittest.TestCase):
+
+    def test_not_locked_out_initially(self):
+        throttle, _, _ = _make()
+        self.assertFalse(throttle.is_locked_out("203.0.113.9"))
+        self.assertEqual(throttle.remaining_lockout("203.0.113.9"), 0.0)
+
+    def test_reaching_the_limit_locks_the_address_out(self):
+        throttle, _, _ = _make()
+        _trip(throttle)
+        self.assertTrue(throttle.is_locked_out("203.0.113.9"))
+
+    def test_the_limit_is_reachable_inside_the_window(self):
+        # Regression guard for the old 60s window: with a ramping delay it
+        # takes ~80s to reach LOGIN_FAIL_MAX, so a 60s window expired the
+        # earliest failures first and the lockout could never fire.
+        throttle, clock, _ = _make()
+        start = clock.now
+        _trip(throttle)
+        self.assertLess(clock.now - start, LOGIN_FAIL_WINDOW)
+        self.assertEqual(throttle.failure_count("203.0.113.9"), LOGIN_FAIL_MAX)
+        self.assertTrue(throttle.is_locked_out("203.0.113.9"))
+
+    def test_one_failure_short_of_the_limit_is_not_a_lockout(self):
+        throttle, _, _ = _make()
+        for _ in range(LOGIN_FAIL_MAX - 1):
+            throttle.record_failure("203.0.113.9")
+        self.assertFalse(throttle.is_locked_out("203.0.113.9"))
+
+    def test_lockout_expires(self):
+        throttle, clock, _ = _make()
+        _trip(throttle)
+        self.assertTrue(throttle.is_locked_out("203.0.113.9"))
+        clock.advance(LOGIN_LOCKOUT_SECONDS + 1)
+        self.assertFalse(throttle.is_locked_out("203.0.113.9"))
+
+    def test_remaining_lockout_counts_down(self):
+        throttle, clock, _ = _make()
+        _trip(throttle)
+        full = throttle.remaining_lockout("203.0.113.9")
+        # the final record_failure applied a delay, which the fake clock has
+        # already advanced, so what is left is the lockout minus that delay
+        self.assertAlmostEqual(full, LOGIN_LOCKOUT_SECONDS,
+                               delta=LOGIN_FAIL_MAX_DELAY + 1.0)
+        clock.advance(full / 2)
+        self.assertLess(throttle.remaining_lockout("203.0.113.9"), full)
+        self.assertGreater(throttle.remaining_lockout("203.0.113.9"), 0.0)
+
+    def test_further_failures_while_locked_out_extend_it(self):
+        throttle, clock, _ = _make()
+        _trip(throttle)
+        clock.advance(LOGIN_LOCKOUT_SECONDS - 1)
+        throttle.record_failure("203.0.113.9")
+        self.assertTrue(throttle.is_locked_out("203.0.113.9"))
+
+
+class Isolation(unittest.TestCase):
+
+    def test_one_address_does_not_lock_out_another(self):
+        throttle, _, _ = _make()
+        _trip(throttle, "203.0.113.9")
+        self.assertTrue(throttle.is_locked_out("203.0.113.9"))
+        self.assertFalse(throttle.is_locked_out("198.51.100.7"))
+
+    def test_successful_login_clears_the_address(self):
+        throttle, _, _ = _make()
+        for _ in range(LOGIN_FAIL_MAX - 1):
+            throttle.record_failure("203.0.113.9")
+        throttle.clear("203.0.113.9")
+        self.assertEqual(throttle.failure_count("203.0.113.9"), 0)
+        self.assertFalse(throttle.is_locked_out("203.0.113.9"))
+
+    def test_old_failures_age_out_of_the_window(self):
+        throttle, clock, _ = _make()
+        throttle.record_failure("203.0.113.9")
+        clock.advance(LOGIN_FAIL_WINDOW + 1)
+        self.assertEqual(throttle.failure_count("203.0.113.9"), 0)
+
+
+class BoundedMemory(unittest.TestCase):
+
+    def test_tracked_addresses_are_evicted(self):
+        throttle, clock, _ = _make(max_tracked_ips=8)
+        for i in range(64):
+            throttle.record_failure(f"198.51.100.{i}")
+        clock.advance(LOGIN_FAIL_WINDOW + LOGIN_LOCKOUT_SECONDS + 1)
+        throttle.record_failure("203.0.113.9")
+        self.assertLessEqual(throttle.tracked_addresses(), 8)
+
+    def test_sleep_is_never_called_under_the_lock(self):
+        # If the lock were held across the sleep, one slow client would stall
+        # every other login — a self-inflicted DoS.
+        throttle, clock, _ = _make()
+        order = []
+
+        def spy(seconds):
+            order.append("sleep:start")
+            clock.advance(seconds)
+            order.append("sleep:end")
+
+        throttle._sleep = spy
+        throttle.record_failure("203.0.113.9")
+        self.assertEqual(order, ["sleep:start", "sleep:end"])
+        # A second address can still be recorded while the first is "sleeping".
+        self.assertEqual(throttle.failure_count("198.51.100.7"), 0)
+
+
+if __name__ == "__main__":
+    unittest.main()