Clean up caching/locking of OIDC metadata load (#9362)

Ensure that we lock correctly to prevent multiple concurrent metadata load
requests, and generally clean up the way we construct the metadata cache.

1 files changed, 129 insertions, 0 deletions

diff --git a/synapse/util/caches/cached_call.py b/synapse/util/caches/cached_call.py
new file mode 100644
index 0000000000..3ee0f2317a
--- /dev/null
+++ b/synapse/util/caches/cached_call.py
@@ -0,0 +1,129 @@
+# -*- coding: utf-8 -*-
+# Copyright 2021 The Matrix.org Foundation C.I.C.
+#
+# Licensed 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.
+
+from typing import Awaitable, Callable, Generic, Optional, TypeVar, Union
+
+from twisted.internet.defer import Deferred
+from twisted.python.failure import Failure
+
+from synapse.logging.context import make_deferred_yieldable, run_in_background
+
+TV = TypeVar("TV")
+
+
+class CachedCall(Generic[TV]):
+    """A wrapper for asynchronous calls whose results should be shared
+
+    This is useful for wrapping asynchronous functions, where there might be multiple
+    callers, but we only want to call the underlying function once (and have the result
+    returned to all callers).
+
+    Similar results can be achieved via a lock of some form, but that typically requires
+    more boilerplate (and ends up being less efficient).
+
+    Correctly handles Synapse logcontexts (logs and resource usage for the underlying
+    function are logged against the logcontext which is active when get() is first
+    called).
+
+    Example usage:
+
+        _cached_val = CachedCall(_load_prop)
+
+        async def handle_request() -> X:
+            # We can call this multiple times, but it will result in a single call to
+            # _load_prop().
+            return await _cached_val.get()
+
+        async def _load_prop() -> X:
+            await difficult_operation()
+
+
+    The implementation is deliberately single-shot (ie, once the call is initiated,
+    there is no way to ask for it to be run). This keeps the implementation and
+    semantics simple. If you want to make a new call, simply replace the whole
+    CachedCall object.
+    """
+
+    __slots__ = ["_callable", "_deferred", "_result"]
+
+    def __init__(self, f: Callable[[], Awaitable[TV]]):
+        """
+        Args:
+            f: The underlying function. Only one call to this function will be alive
+                at once (per instance of CachedCall)
+        """
+        self._callable = f  # type: Optional[Callable[[], Awaitable[TV]]]
+        self._deferred = None  # type: Optional[Deferred]
+        self._result = None  # type: Union[None, Failure, TV]
+
+    async def get(self) -> TV:
+        """Kick off the call if necessary, and return the result"""
+
+        # Fire off the callable now if this is our first time
+        if not self._deferred:
+            self._deferred = run_in_background(self._callable)
+
+            # we will never need the callable again, so make sure it can be GCed
+            self._callable = None
+
+            # once the deferred completes, store the result. We cannot simply leave the
+            # result in the deferred, since if it's a Failure, GCing the deferred
+            # would then log a critical error about unhandled Failures.
+            def got_result(r):
+                self._result = r
+
+            self._deferred.addBoth(got_result)
+
+        # TODO: consider cancellation semantics. Currently, if the call to get()
+        #    is cancelled, the underlying call will continue (and any future calls
+        #    will get the result/exception), which I think is *probably* ok, modulo
+        #    the fact the underlying call may be logged to a cancelled logcontext,
+        #    and any eventual exception may not be reported.
+
+        # we can now await the deferred, and once it completes, return the result.
+        await make_deferred_yieldable(self._deferred)
+
+        # I *think* this is the easiest way to correctly raise a Failure without having
+        # to gut-wrench into the implementation of Deferred.
+        d = Deferred()
+        d.callback(self._result)
+        return await d
+
+
+class RetryOnExceptionCachedCall(Generic[TV]):
+    """A wrapper around CachedCall which will retry the call if an exception is thrown
+
+    This is used in much the same way as CachedCall, but adds some extra functionality
+    so that if the underlying function throws an exception, then the next call to get()
+    will initiate another call to the underlying function. (Any calls to get() which
+    are already pending will raise the exception.)
+    """
+
+    slots = ["_cachedcall"]
+
+    def __init__(self, f: Callable[[], Awaitable[TV]]):
+        async def _wrapper() -> TV:
+            try:
+                return await f()
+            except Exception:
+                # the call raised an exception: replace the underlying CachedCall to
+                # trigger another call next time get() is called
+                self._cachedcall = CachedCall(_wrapper)
+                raise
+
+        self._cachedcall = CachedCall(_wrapper)
+
+    async def get(self) -> TV:
+        return await self._cachedcall.get()