madhushreeag commented on code in PR #44365: URL: https://github.com/apache/superset/pull/44365#discussion_r4198022677
########## superset/security/login_token.py: ########## @@ -0,0 +1,290 @@ +# Licensed to the Apache Software Foundation (ASF) under one +# or more contributor license agreements. See the NOTICE file +# distributed with this work for additional information +# regarding copyright ownership. The ASF licenses this file +# to you under the Apache License, Version 2.0 (the +# "License"); you may not use this file except in compliance +# with the License. You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, +# software distributed under the License is distributed on an +# "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY +# KIND, either express or implied. See the License for the +# specific language governing permissions and limitations +# under the License. +"""One-time login tokens for establishing a session inside an iframe. + +An SSO redirect flow cannot run inside an iframe: identity providers commonly +refuse to be framed, and the redirect chain depends on cookies that browsers +treat as third-party in an embedded context. A parent application that already +holds a trustworthy proof of the user's identity uses these tokens to turn that +proof into an ordinary Superset session in two steps: + +1. its backend mints a token, presenting a credential that an operator-supplied + resolver validates (server-to-server, so the credential never reaches the + browser), and +2. the browser navigates the iframe to the consume endpoint, which exchanges the + token for a session cookie. + +The token is an opaque handle to a short-lived server-side record; no identity +data travels in the URL. +""" + +from __future__ import annotations + +import logging +import re +from datetime import datetime, timedelta +from typing import Any, Callable, cast, TypedDict +from uuid import UUID, uuid4 + +from flask import current_app, Request + +from superset.daos.key_value import KeyValueDAO +from superset.key_value.types import JsonKeyValueCodec, KeyValueResource + +logger = logging.getLogger(__name__) + +LOGIN_TOKEN_RESOURCE = KeyValueResource.LOGIN_TOKEN +LOGIN_TOKEN_CODEC = JsonKeyValueCodec() + +# Kept short deliberately: the token is handed to a browser as an iframe URL, so +# it lands in access logs and the parent page's DOM. A lifetime measured in +# seconds bounds the window in which an observer could replay it. +DEFAULT_LOGIN_TOKEN_TTL_SECONDS = 60 + + +class LoginTokenUserInfo(TypedDict, total=False): + """The identity a resolver returns for a one-time login token. + + This mirrors the ``userinfo`` contract of Flask-AppBuilder's + ``auth_user_oauth``, which the consume step delegates to. ``username`` + identifies the user, falling back to ``email`` when absent. ``role_keys`` are + resolved through ``AUTH_ROLES_MAPPING``, so a caller can only ever request + roles the operator has already mapped -- authorization stays with the + operator rather than moving to the parent application. + """ + + username: str + email: str + first_name: str + last_name: str + role_keys: list[str] + + +LoginTokenIdentityResolver = Callable[..., LoginTokenUserInfo | None] + +# Characters a WHATWG URL parser strips before resolving, plus their +# percent-encoded forms, which some browsers also remove when following a +# Location header. Normalizing them first stops `/\tx` or `/%09x` smuggling a +# different target past the checks below. +_URL_STRIPPED_CONTROL_CHARS = re.compile(r"[\t\n\r]|%09|%0[ADad]") + + +def is_safe_next_path(url: str) -> bool: + """Whether ``url`` is a site-relative path this endpoint may redirect to. + + Deliberately stricter than :func:`superset.utils.link_redirect. + is_safe_redirect_url`, which resolves "internal" against + ``WEBDRIVER_BASEURL`` / ``WEBDRIVER_BASEURL_USER_FRIENDLY``. Those are + report-worker settings defaulting to ``http://0.0.0.0:8080/`` and need bear + no relation to the public origin, so a deployment that has never configured + reports would see a legitimate same-origin absolute URL rejected and be + redirected to ``/`` instead of the requested dashboard. + + Requiring a relative path avoids the question entirely: the parent + application always knows the path it wants, the browser resolves it against + Superset's own origin, and there is no host to compare. + """ + if not url or not url.strip(): + return False + + normalized = _URL_STRIPPED_CONTROL_CHARS.sub("", url.strip()) + # Browsers treat backslashes as forward slashes in special schemes, so + # `/\evil.com` would resolve as the protocol-relative `//evil.com`. + normalized = normalized.replace("\\", "/") + + # A single leading slash, and nothing that could be read as a host or a + # scheme. `//host`, `https://host` and `mailto:x` are all rejected. + return normalized.startswith("/") and not normalized.startswith("//") + + +def is_enabled() -> bool: + """Whether the flow is usable: feature flag on and a resolver configured.""" + # Deferred: ``superset/__init__`` imports the app factory, so a module-level + # import here would be circular via superset.daos / superset.security. + from superset import ( # pylint: disable=import-outside-toplevel + is_feature_enabled, + ) + + if not is_feature_enabled("LOGIN_TOKEN"): + return False + + return current_app.config.get("LOGIN_TOKEN_IDENTITY_RESOLVER") is not None + + +def get_ttl_seconds() -> int: + """The configured token lifetime, falling back to the default.""" + configured = current_app.config.get( + "LOGIN_TOKEN_TTL_SECONDS", DEFAULT_LOGIN_TOKEN_TTL_SECONDS + ) + try: + ttl = int(configured) + except (TypeError, ValueError): + logger.warning( + "LOGIN_TOKEN_TTL_SECONDS is not an integer (%r); using %s", + configured, + DEFAULT_LOGIN_TOKEN_TTL_SECONDS, + ) + return DEFAULT_LOGIN_TOKEN_TTL_SECONDS + + return ttl if ttl > 0 else DEFAULT_LOGIN_TOKEN_TTL_SECONDS + + +# Each rejection below is a distinct failure with its own diagnostic — an +# unconfigured resolver is normal and silent, a misconfigured or misbehaving one +# is logged. Merging the branches to satisfy the return-count limit would lose +# that distinction, which is the only signal an operator gets. +def resolve_identity( # pylint: disable=too-many-return-statements + request: Request, **kwargs: Any +) -> LoginTokenUserInfo | None: + """Run the operator-supplied resolver against an inbound mint request. + + The resolver is the whole authentication boundary for minting: it decides + what counts as proof of identity (an OIDC id token, an existing session, a + credential checked against an internal service). A resolver that raises, or + returns something other than a mapping, is treated as a rejection rather + than a server error, so a failing validation can never be mistaken for a + successful one. + """ + resolver = current_app.config.get("LOGIN_TOKEN_IDENTITY_RESOLVER") + if resolver is None: + return None + + if not callable(resolver): + logger.error("LOGIN_TOKEN_IDENTITY_RESOLVER is not callable") + return None + + try: + userinfo = resolver(request, **kwargs) + except Exception: # pylint: disable=broad-except + logger.exception("LOGIN_TOKEN_IDENTITY_RESOLVER rejected the request") + return None + + if not userinfo: + return None + + if not isinstance(userinfo, dict): + # A truthy non-mapping would otherwise raise on the ``.get`` below and + # surface as a 500, contradicting the rejection contract above. + logger.error( + "LOGIN_TOKEN_IDENTITY_RESOLVER returned %s, expected a mapping", + type(userinfo).__name__, + ) + return None + + # Normalize before validating, because ``auth_user_oauth`` selects on key + # *presence* rather than truthiness: ``if "username" in userinfo`` wins even + # when the value is empty, and the empty username is then rejected outright. + # A resolver returning {"username": "", "email": "[email protected]"} would + # otherwise mint a perfectly good token that always fails redemption with a + # 401, instead of falling back to the email. Stripping and dropping empty + # string values keeps this check and FAB's in agreement. + userinfo = { + key: value.strip() if isinstance(value, str) else value + for key, value in userinfo.items() + if not (isinstance(value, str) and not value.strip()) + } + + if not (userinfo.get("username") or userinfo.get("email")): + # auth_user_oauth derives the username from one of these two; without + # either there is no identity to provision. + logger.error("LOGIN_TOKEN_IDENTITY_RESOLVER returned no username or email") + return None + + # The resolver is operator-supplied and typed Any, so the isinstance guard + # above narrows it only as far as a plain dict. Its keys are the resolver's + # contract, checked here to the extent that matters (a usable identifier); + # anything extra is ignored by auth_user_oauth. + return cast(LoginTokenUserInfo, userinfo) + + +def mint(userinfo: LoginTokenUserInfo) -> tuple[str, datetime]: + """Store ``userinfo`` behind a fresh opaque token. + + Returns the token and its expiry. The token is a ``uuid4``, so it carries no + identity data and is not guessable. + """ + token = uuid4() + expires_on = datetime.now() + timedelta(seconds=get_ttl_seconds()) + + # Opportunistic GC rather than collision avoidance: a fresh uuid4 cannot + # collide, but with a TTL measured in seconds these rows turn over quickly, + # and the scheduled prune job runs far less often than tokens are minted. + # Minting is not a hot path, so one indexed DELETE here is cheap insurance + # against the table filling with dead entries between prunes. + KeyValueDAO.delete_expired_entries(LOGIN_TOKEN_RESOURCE) + KeyValueDAO.create_entry( + resource=LOGIN_TOKEN_RESOURCE, + value=dict(userinfo), + codec=LOGIN_TOKEN_CODEC, + key=token, + expires_on=expires_on, + ) + return str(token), expires_on + + +def consume(token: str) -> LoginTokenUserInfo | None: + """Exchange a token for its ``userinfo``, burning it durably. + + The entry is row-locked before it is read, so two concurrent requests cannot + both observe it: the second blocks until the first commits and then finds it + gone. + + The delete is then **committed here**, rather than left to the caller's unit + of work. Flask-AppBuilder's ``add_user`` / ``update_user`` catch their own + failures, call ``rollback()`` on this same session and return ``False`` + without raising -- and ``update_user_auth_stat`` ignores that return value + entirely. A provisioning error, or a failing ``user_updating`` signal + handler, would therefore silently undo a pending delete and leave a token + that has already been handed to a browser redeemable again. Committing the + burn first makes it independent of anything provisioning does, including a + login that ultimately succeeds. + + Returns ``None`` for an unknown, malformed, or expired token -- callers must + not distinguish between those cases in their response. + """ + from superset import db # pylint: disable=import-outside-toplevel + + try: + key = UUID(token) + except (AttributeError, TypeError, ValueError): + return None + + entry = KeyValueDAO.get_entry(LOGIN_TOKEN_RESOURCE, key, for_update=True) + if entry is None: + return None + + # Delete regardless of expiry: an expired token is spent either way, and + # leaving it behind would only wait for the pruning job. + expired = entry.is_expired() + try: + userinfo = LOGIN_TOKEN_CODEC.decode(entry.value) + except Exception: # pylint: disable=broad-except + logger.exception("Unable to decode stored login token payload") + userinfo = None + + KeyValueDAO.delete_entry(LOGIN_TOKEN_RESOURCE, key) Review Comment: Right! Also, `delete_entry` returns the result of a SELECT (if entry := get_entry(...)), not the DELETE's rowcount- so both racers get True. Checking its return value wouldn't fix this. The gate has to be a conditional DELETE. As per your suggestion, the burn is now the gate. A single conditional `DELETE` whose affected-row count must be exactly 1, and the identity is returned only to the request that owns that row - otherwise `None`, which reaches the browser as the same 401 as every other failure. That holds on every backend rather than only where `FOR UPDATE` compiles to something. `with_for_update` stays as the backstop you describe, so on Postgres and MySQL the contenders still serialize up front instead of the loser decoding and discarding. -- This is an automated message from the Apache Git Service. To respond to the message, please log on to GitHub and use the URL above to go to the specific comment. To unsubscribe, e-mail: [email protected] For queries about this service, please contact Infrastructure at: [email protected] --------------------------------------------------------------------- To unsubscribe, e-mail: [email protected] For additional commands, e-mail: [email protected]
