Five findings from the latest review on #758 (2 medium, 3 nit): Medium: - browser_oauth_routes.oauth_login_callback + oauth_routes.oauth_callback_nextcloud: fail closed with 400 when the oauth_session row is unknown/expired. Previously both callbacks fell through with code_verifier="" and expected_nonce=None, silently bypassing the PKCE + nonce protections introduced in earlier rounds. Symmetric unit tests pin both contracts. - token_utils.verify_id_token: use secrets.compare_digest for the nonce check instead of short-circuit !=. Mirrors the sibling PKCE verifier comparison; closes the last secret-equality timing-side-channel surface in the auth path. Nit: - Tighten the comment at all 4 mcp_authorization_code/code_verifier store + retrieve sites so a future refactor sees the field reuse immediately (renaming the column requires a schema migration). - _should_use_secure_cookies: explicit string normalisation instead of bool(settings.cookie_secure). Dynaconf normally coerces but tests / direct settings.set calls can leave the raw string in place — bool("false") is True. New parametrized unit tests cover the coercion matrix + http/https fallback. - oauth_routes.py:591 f-string log converted to lazy %s formatting (folded into the Flow 2 callback rewrite). Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
296 lines
12 KiB
Python
296 lines
12 KiB
Python
"""Token utility functions for extracting user identity from MCP access tokens.
|
|
|
|
Extracted from server/oauth_tools.py to break circular import dependencies
|
|
between server/ and auth/ layers.
|
|
"""
|
|
|
|
import logging
|
|
import secrets
|
|
import time
|
|
from typing import Any
|
|
|
|
import anyio
|
|
import jwt
|
|
from jwt import PyJWKSet
|
|
from mcp.server.auth.middleware.auth_context import get_access_token
|
|
from mcp.server.auth.provider import AccessToken
|
|
from mcp.server.fastmcp import Context
|
|
from mcp.shared.exceptions import McpError
|
|
from mcp.types import ErrorData
|
|
|
|
from ..http import nextcloud_httpx_client
|
|
|
|
logger = logging.getLogger(__name__)
|
|
|
|
|
|
# OIDC discovery + JWKS caches keyed by URL → (expires_at, data). Single
|
|
# source of truth for the codebase: oauth_routes / browser_oauth_routes both
|
|
# go through ``get_oidc_discovery`` which reads/writes _discovery_cache, so
|
|
# the first discovery fetch primes the cache for all later callers (PR #758
|
|
# round-2 nit 3). 5-minute TTL.
|
|
_discovery_cache: dict[str, tuple[float, dict[str, Any]]] = {}
|
|
_jwks_cache: dict[str, tuple[float, dict[str, Any]]] = {}
|
|
_OIDC_CACHE_TTL = 300
|
|
|
|
# Per-URL fetch locks coalesce concurrent cache misses into a single HTTP
|
|
# request, preventing thundering-herd against the IdP at cache expiry
|
|
# (PR #758 round-3 review). Mirrors the lock-dict + meta-lock idiom from
|
|
# token_broker.py.
|
|
_fetch_locks: dict[str, anyio.Lock] = {}
|
|
_fetch_locks_lock = anyio.Lock()
|
|
|
|
|
|
async def _get_fetch_lock(url: str) -> anyio.Lock:
|
|
"""Return the per-URL lock used to serialise cache-miss fetches."""
|
|
async with _fetch_locks_lock:
|
|
lock = _fetch_locks.get(url)
|
|
if lock is None:
|
|
lock = anyio.Lock()
|
|
_fetch_locks[url] = lock
|
|
return lock
|
|
|
|
|
|
class IdTokenVerificationError(Exception):
|
|
"""Raised when an OIDC ID token fails signature or claim verification."""
|
|
|
|
|
|
async def _get_cached(
|
|
cache: dict[str, tuple[float, dict[str, Any]]],
|
|
url: str,
|
|
*,
|
|
follow_redirects: bool = False,
|
|
) -> dict[str, Any]:
|
|
"""Return cached JSON response for *url* or fetch + cache on miss/expiry.
|
|
|
|
``follow_redirects`` is forwarded to ``nextcloud_httpx_client``: discovery
|
|
fetches against Nextcloud without pretty URLs need it (the configured
|
|
``/.well-known/openid-configuration`` path issues a 301), but JWKS
|
|
fetches deliberately stay strict — the URL came from the discovery
|
|
document we already trust, so a redirect there would be suspicious.
|
|
|
|
Concurrent callers seeing the same cache miss are coalesced via a
|
|
per-URL ``anyio.Lock``: only one fetch runs, the rest wait and read the
|
|
populated cache.
|
|
"""
|
|
entry = cache.get(url)
|
|
if entry is not None and time.time() < entry[0]:
|
|
return entry[1]
|
|
lock = await _get_fetch_lock(url)
|
|
try:
|
|
async with lock:
|
|
# Re-check inside the lock — a concurrent waiter may have already
|
|
# populated the cache before we acquired it.
|
|
entry = cache.get(url)
|
|
if entry is not None and time.time() < entry[0]:
|
|
return entry[1]
|
|
async with nextcloud_httpx_client(
|
|
follow_redirects=follow_redirects
|
|
) as http_client:
|
|
response = await http_client.get(url)
|
|
response.raise_for_status()
|
|
data = response.json()
|
|
cache[url] = (time.time() + _OIDC_CACHE_TTL, data)
|
|
return data
|
|
finally:
|
|
# Drop the dict entry so a misconfigured deployment hitting
|
|
# arbitrary URLs can't grow ``_fetch_locks`` without bound (PR #758
|
|
# round-4 review nit 4). Already-queued waiters share our local
|
|
# ``lock`` reference and remain coalesced; new arrivals lazily
|
|
# recreate a lock — by which time the cache is populated, so they
|
|
# short-circuit before reaching the lock anyway.
|
|
async with _fetch_locks_lock:
|
|
if _fetch_locks.get(url) is lock:
|
|
del _fetch_locks[url]
|
|
|
|
|
|
async def get_oidc_discovery(discovery_url: str) -> dict[str, Any]:
|
|
"""Return the cached OIDC discovery document for *discovery_url*.
|
|
|
|
Shares the 5-minute discovery cache used by `verify_id_token`, so a
|
|
callback that does discovery → token-exchange → ID-token verification
|
|
reuses one HTTP round-trip instead of three. The fetch follows
|
|
redirects because Nextcloud without pretty URLs returns 301 from
|
|
``/.well-known/openid-configuration`` to ``/index.php/.well-known/...``.
|
|
Single source of truth for OIDC discovery in the codebase
|
|
(PR #758 round-2 nit 3).
|
|
"""
|
|
return await _get_cached(_discovery_cache, discovery_url, follow_redirects=True)
|
|
|
|
|
|
async def verify_id_token(
|
|
id_token: str | None,
|
|
*,
|
|
discovery_url: str,
|
|
expected_audience: str,
|
|
expected_nonce: str | None = None,
|
|
) -> dict[str, Any]:
|
|
"""Verify an OIDC ID token's signature and standard claims.
|
|
|
|
Implements the verification steps required by OIDC core spec section
|
|
3.1.3.7 (ID Token Validation) for the authorization-code flow:
|
|
- Signature against JWKS (RS256)
|
|
- Issuer matches the OP that issued the token
|
|
- Audience contains the expected client_id
|
|
- Token is not expired (`exp`)
|
|
- `iat` is well-formed (PyJWT default)
|
|
- `nonce` matches when one was included in the auth request
|
|
|
|
Replaces the prior `jwt.decode(id_token, options={"verify_signature": False})`
|
|
pattern (issue #626 finding 1) on the OAuth callback paths.
|
|
|
|
Args:
|
|
id_token: Raw ID token (JWT) string.
|
|
discovery_url: OIDC `.well-known/openid-configuration` URL of the IdP.
|
|
expected_audience: The MCP-server-side OAuth client_id used for this
|
|
authorization request.
|
|
expected_nonce: When the auth request included a nonce, the same value
|
|
so it can be checked here. None disables the nonce check (callers
|
|
that didn't bind a nonce in the auth request).
|
|
|
|
Returns:
|
|
Decoded, verified ID-token claims.
|
|
|
|
Raises:
|
|
IdTokenVerificationError: On any verification failure.
|
|
"""
|
|
if not id_token:
|
|
raise IdTokenVerificationError("ID token missing from token response")
|
|
|
|
try:
|
|
discovery = await get_oidc_discovery(discovery_url)
|
|
|
|
issuer = discovery.get("issuer")
|
|
jwks_uri = discovery.get("jwks_uri")
|
|
if not issuer or not jwks_uri:
|
|
raise IdTokenVerificationError(
|
|
"OIDC discovery response missing issuer or jwks_uri"
|
|
)
|
|
|
|
jwks_data = await _get_cached(_jwks_cache, jwks_uri)
|
|
except IdTokenVerificationError:
|
|
raise
|
|
except Exception as e:
|
|
raise IdTokenVerificationError(
|
|
f"Failed to fetch OIDC discovery / JWKS: {e}"
|
|
) from e
|
|
|
|
try:
|
|
jwks = PyJWKSet.from_dict(jwks_data)
|
|
unverified_header = jwt.get_unverified_header(id_token)
|
|
kid = unverified_header.get("kid")
|
|
if not kid:
|
|
raise IdTokenVerificationError("ID token header missing 'kid'")
|
|
try:
|
|
signing_key = jwks[kid]
|
|
except KeyError:
|
|
# Cache miss may indicate IdP key rotation. Refresh JWKS once
|
|
# before giving up, per OIDC core §10.1.1: when an unrecognised
|
|
# `kid` arrives the relying party should refetch the JWKS rather
|
|
# than waiting for cache TTL to elapse.
|
|
_jwks_cache.pop(jwks_uri, None)
|
|
try:
|
|
jwks_data = await _get_cached(_jwks_cache, jwks_uri)
|
|
jwks = PyJWKSet.from_dict(jwks_data)
|
|
signing_key = jwks[kid]
|
|
except KeyError as e:
|
|
raise IdTokenVerificationError(
|
|
f"No JWKS key matches ID token kid {kid!r}"
|
|
) from e
|
|
except Exception as e:
|
|
raise IdTokenVerificationError(
|
|
f"Failed to refresh JWKS after kid miss: {e}"
|
|
) from e
|
|
|
|
# PyJWT verifies the JWT with the algorithm declared in its header,
|
|
# cross-checked against this allowlist (so an attacker can't downgrade
|
|
# to ``none`` or HMAC). The allowlist covers the OIDC algorithms
|
|
# most cloud IdPs ship by default:
|
|
# - RS256: Nextcloud user_oidc, Keycloak default, Auth0, Google.
|
|
# - PS256: Azure AD on newer keys.
|
|
# - ES256: some Keycloak realms, AWS Cognito user pools.
|
|
# Symmetric (HSxxx) and ``none`` are intentionally absent.
|
|
payload: dict[str, Any] = jwt.decode(
|
|
id_token,
|
|
signing_key.key,
|
|
algorithms=["RS256", "PS256", "ES256"],
|
|
audience=expected_audience,
|
|
issuer=issuer,
|
|
options={
|
|
"verify_signature": True,
|
|
"verify_exp": True,
|
|
"verify_iat": True,
|
|
"verify_aud": True,
|
|
"verify_iss": True,
|
|
"require": ["sub", "iss", "aud", "exp", "iat"],
|
|
},
|
|
)
|
|
except IdTokenVerificationError:
|
|
raise
|
|
except jwt.PyJWTError as e:
|
|
raise IdTokenVerificationError(f"ID token verification failed: {e}") from e
|
|
except Exception as e:
|
|
raise IdTokenVerificationError(
|
|
f"Unexpected error verifying ID token: {e}"
|
|
) from e
|
|
|
|
# Constant-time comparison mirrors the PKCE verifier check
|
|
# (oauth_routes.py:1029) — short-circuit `!=` is avoided in
|
|
# security-sensitive equality even when the secret is server-generated
|
|
# (round-6 review).
|
|
if expected_nonce is not None and not secrets.compare_digest(
|
|
payload.get("nonce", "") or "", expected_nonce
|
|
):
|
|
raise IdTokenVerificationError("ID token nonce does not match request nonce")
|
|
|
|
return payload
|
|
|
|
|
|
async def extract_user_id_from_token(_ctx: Context) -> str:
|
|
"""Extract user_id from the verified MCP access token.
|
|
|
|
Reads the `sub` claim from `AccessToken.resource`, which is populated by
|
|
`UnifiedTokenVerifier` after JWT signature verification (or token
|
|
introspection for opaque tokens). We never re-decode the raw token here:
|
|
the verifier has already validated the signature and extracted the
|
|
identity claim.
|
|
|
|
Args:
|
|
_ctx: MCP context with access token. Intentionally unused — kept on
|
|
the public signature so call sites can pass the FastMCP Context
|
|
they already hold without rewriting; identity is read from the
|
|
verifier-populated AccessToken via get_access_token().
|
|
|
|
Returns:
|
|
user_id from the verified token, or ``"default_user"`` when no
|
|
access token is present at all (BasicAuth mode — there is no
|
|
OAuth identity to extract, so the sentinel is returned and the
|
|
caller's BasicAuth branch handles it).
|
|
|
|
Raises:
|
|
McpError: An access token was present but had no ``sub`` claim
|
|
(``access_token.resource`` empty). Failing closed prevents a
|
|
malformed IdP token from silently bucketing every request
|
|
under the ``"default_user"`` key in SQLite, which would risk
|
|
cross-tenant data exposure (PR #758 follow-up review).
|
|
"""
|
|
access_token: AccessToken | None = get_access_token()
|
|
|
|
if not access_token:
|
|
logger.warning("No access token found via get_access_token()")
|
|
return "default_user"
|
|
|
|
user_id = access_token.resource
|
|
if not user_id:
|
|
logger.error(
|
|
"Access token has no resource (sub) claim — verifier should have rejected it"
|
|
)
|
|
raise McpError(
|
|
ErrorData(
|
|
# JSON-RPC 2.0 reserves -32000..-32099 for application errors.
|
|
code=-32001,
|
|
message="Cannot determine user identity from access token",
|
|
)
|
|
)
|
|
|
|
return user_id
|