Merge pull request #1765 from matrix-org/markjh/timeout_get_missing_events

cherrypick #1744: limit total timeout for get_missing_events to 10s

3 files changed, 27 insertions, 3 deletions

diff --git a/synapse/federation/federation_client.py b/synapse/federation/federation_client.py
index 6851f2376d..b4bcec77ed 100644
--- a/synapse/federation/federation_client.py
+++ b/synapse/federation/federation_client.py
@@ -707,7 +707,7 @@ class FederationClient(FederationBase):
 
     @defer.inlineCallbacks
     def get_missing_events(self, destination, room_id, earliest_events_ids,
-                           latest_events, limit, min_depth):
+                           latest_events, limit, min_depth, timeout):
         """Tries to fetch events we are missing. This is called when we receive
         an event without having received all of its ancestors.
 
@@ -721,6 +721,7 @@ class FederationClient(FederationBase):
                 have all previous events for.
             limit (int): Maximum number of events to return.
             min_depth (int): Minimum depth of events tor return.
+            timeout (int): Max time to wait in ms
         """
         try:
             content = yield self.transport_layer.get_missing_events(
@@ -730,6 +731,7 @@ class FederationClient(FederationBase):
                 latest_events=[e.event_id for e in latest_events],
                 limit=limit,
                 min_depth=min_depth,
+                timeout=timeout,
             )
 
             events = [
diff --git a/synapse/federation/federation_server.py b/synapse/federation/federation_server.py
index f4c60e67e3..800f04189f 100644
--- a/synapse/federation/federation_server.py
+++ b/synapse/federation/federation_server.py
@@ -425,6 +425,7 @@ class FederationServer(FederationBase):
                 " limit: %d, min_depth: %d",
                 earliest_events, latest_events, limit, min_depth
             )
+
             missing_events = yield self.handler.on_get_missing_events(
                 origin, room_id, earliest_events, latest_events, limit, min_depth
             )
@@ -567,6 +568,25 @@ class FederationServer(FederationBase):
                                 len(prevs - seen), pdu.room_id, list(prevs - seen)[:5]
                             )
 
+                            # XXX: we set timeout to 10s to help workaround
+                            # https://github.com/matrix-org/synapse/issues/1733.
+                            # The reason is to avoid holding the linearizer lock
+                            # whilst processing inbound /send transactions, causing
+                            # FDs to stack up and block other inbound transactions
+                            # which empirically can currently take up to 30 minutes.
+                            #
+                            # N.B. this explicitly disables retry attempts.
+                            #
+                            # N.B. this also increases our chances of falling back to
+                            # fetching fresh state for the room if the missing event
+                            # can't be found, which slightly reduces our security.
+                            # it may also increase our DAG extremity count for the room,
+                            # causing additional state resolution?  See #1760.
+                            # However, fetching state doesn't hold the linearizer lock
+                            # apparently.
+                            #
+                            # see https://github.com/matrix-org/synapse/pull/1744
+
                             missing_events = yield self.get_missing_events(
                                 origin,
                                 pdu.room_id,
@@ -574,6 +594,7 @@ class FederationServer(FederationBase):
                                 latest_events=[pdu],
                                 limit=10,
                                 min_depth=min_depth,
+                                timeout=10000,
                             )
 
                             # We want to sort these by depth so we process them and
diff --git a/synapse/federation/transport/client.py b/synapse/federation/transport/client.py
index 491cdc29e1..915af34409 100644
--- a/synapse/federation/transport/client.py
+++ b/synapse/federation/transport/client.py
@@ -386,7 +386,7 @@ class TransportLayerClient(object):
     @defer.inlineCallbacks
     @log_function
     def get_missing_events(self, destination, room_id, earliest_events,
-                           latest_events, limit, min_depth):
+                           latest_events, limit, min_depth, timeout):
         path = PREFIX + "/get_missing_events/%s" % (room_id,)
 
         content = yield self.client.post_json(
@@ -397,7 +397,8 @@ class TransportLayerClient(object):
                 "min_depth": int(min_depth),
                 "earliest_events": earliest_events,
                 "latest_events": latest_events,
-            }
+            },
+            timeout=timeout,
         )
 
         defer.returnValue(content)