fix(auth): address PR #758 round-3 final review

Seven findings from the latest review on #758, plus a regression test
catching the substance of the cache-stampede fix:

- verify_id_token: widen id_token annotation to str | None to match
  callers passing nc_token_response.get("id_token")
- extract_user_id_from_token: use JSON-RPC reserved error code -32001
  instead of -1
- _get_cached: per-URL anyio.Lock dict + meta-lock coalesces concurrent
  cache misses into a single IdP fetch (mirrors token_broker.py idiom)
- delete_browser_session: collapse SELECT+DELETE into atomic
  DELETE ... RETURNING user_id (SQLite >= 3.35)
- new test_origin_normalise.py: parametrized port/scheme/host equivalence
  cases for the CSRF Origin guard
- browser_oauth_routes: correct misleading "PR #758 finding 5" cross-
  references (finding 5 was Fernet-key hardening, not CSRF)
- ASProxySession.nonce: make required, drop spurious "legacy session"
  default; reword the in-flight `or None` comment to reflect that
  ASProxySession is purely in-memory
- new test_get_cached_coalesces_concurrent_misses: pins the
  cache-stampede protection — fires 10 concurrent _get_cached calls and
  asserts exactly one HTTP fetch

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
This commit is contained in:
Chris Coutinho
2026-05-02 23:14:07 +02:00
co-authored by Claude Opus 4.7
parent 9d0e7dcebe
commit 3a4fa8adc8
7 changed files with 152 additions and 30 deletions
+42 -13
View File
@@ -8,6 +8,7 @@ import logging
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
@@ -30,6 +31,23 @@ _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."""
@@ -48,19 +66,29 @@ async def _get_cached(
``/.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.
"""
now = time.time()
entry = cache.get(url)
if entry is not None:
expires_at, data = entry
if now < expires_at:
return data
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] = (now + _OIDC_CACHE_TTL, data)
return data
if entry is not None and time.time() < entry[0]:
return entry[1]
lock = await _get_fetch_lock(url)
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
async def get_oidc_discovery(discovery_url: str) -> dict[str, Any]:
@@ -78,7 +106,7 @@ async def get_oidc_discovery(discovery_url: str) -> dict[str, Any]:
async def verify_id_token(
id_token: str,
id_token: str | None,
*,
discovery_url: str,
expected_audience: str,
@@ -240,7 +268,8 @@ async def extract_user_id_from_token(_ctx: Context) -> str:
)
raise McpError(
ErrorData(
code=-1,
# JSON-RPC 2.0 reserves -32000..-32099 for application errors.
code=-32001,
message="Cannot determine user identity from access token",
)
)