Fix stack overflow in `_PerHostRatelimiter` due to synchronous requests (#14812)

When there are many synchronous requests waiting on a
`_PerHostRatelimiter`, each request will be started recursively just
after the previous request has completed. Under the right conditions,
this leads to stack exhaustion.

A common way for requests to become synchronous is when the remote
client disconnects early, because the homeserver is overloaded and slow
to respond.

Avoid stack exhaustion under these conditions by deferring subsequent
requests until the next reactor tick.

Fixes #14480.

Signed-off-by: Sean Quah <seanq@matrix.org>

1 files changed, 25 insertions, 9 deletions

diff --git a/synapse/util/ratelimitutils.py b/synapse/util/ratelimitutils.py
index 2aceb1a47f..bd72947bfe 100644
--- a/synapse/util/ratelimitutils.py
+++ b/synapse/util/ratelimitutils.py
@@ -34,6 +34,7 @@ from prometheus_client.core import Counter
 from typing_extensions import ContextManager
 
 from twisted.internet import defer
+from twisted.internet.interfaces import IReactorTime
 
 from synapse.api.errors import LimitExceededError
 from synapse.config.ratelimiting import FederationRatelimitSettings
@@ -146,12 +147,14 @@ class FederationRateLimiter:
 
     def __init__(
         self,
+        reactor: IReactorTime,
         clock: Clock,
         config: FederationRatelimitSettings,
         metrics_name: Optional[str] = None,
     ):
         """
         Args:
+            reactor
             clock
             config
             metrics_name: The name of the rate limiter so we can differentiate it
@@ -163,7 +166,7 @@ class FederationRateLimiter:
 
         def new_limiter() -> "_PerHostRatelimiter":
             return _PerHostRatelimiter(
-                clock=clock, config=config, metrics_name=metrics_name
+                reactor=reactor, clock=clock, config=config, metrics_name=metrics_name
             )
 
         self.ratelimiters: DefaultDict[
@@ -194,12 +197,14 @@ class FederationRateLimiter:
 class _PerHostRatelimiter:
     def __init__(
         self,
+        reactor: IReactorTime,
         clock: Clock,
         config: FederationRatelimitSettings,
         metrics_name: Optional[str] = None,
     ):
         """
         Args:
+            reactor
             clock
             config
             metrics_name: The name of the rate limiter so we can differentiate it
@@ -207,6 +212,7 @@ class _PerHostRatelimiter:
                 for this rate limiter.
                 from the rest in the metrics
         """
+        self.reactor = reactor
         self.clock = clock
         self.metrics_name = metrics_name
 
@@ -364,12 +370,22 @@ class _PerHostRatelimiter:
 
     def _on_exit(self, request_id: object) -> None:
         logger.debug("Ratelimit(%s) [%s]: Processed req", self.host, id(request_id))
-        self.current_processing.discard(request_id)
-        try:
-            # start processing the next item on the queue.
-            _, deferred = self.ready_request_queue.popitem(last=False)
 
-            with PreserveLoggingContext():
-                deferred.callback(None)
-        except KeyError:
-            pass
+        # When requests complete synchronously, we will recursively start the next
+        # request in the queue. To avoid stack exhaustion, we defer starting the next
+        # request until the next reactor tick.
+
+        def start_next_request() -> None:
+            # We only remove the completed request from the list when we're about to
+            # start the next one, otherwise we can allow extra requests through.
+            self.current_processing.discard(request_id)
+            try:
+                # start processing the next item on the queue.
+                _, deferred = self.ready_request_queue.popitem(last=False)
+
+                with PreserveLoggingContext():
+                    deferred.callback(None)
+            except KeyError:
+                pass
+
+        self.reactor.callLater(0.0, start_next_request)