Merge remote-tracking branch 'origin/develop' into rav/warn_on_logcontext_fail

1 files changed, 40 insertions, 17 deletions

diff --git a/synapse/federation/federation_server.py b/synapse/federation/federation_server.py
index bea7fd0b71..247ddc89d5 100644
--- a/synapse/federation/federation_server.py
+++ b/synapse/federation/federation_server.py
@@ -1,5 +1,6 @@
 # -*- coding: utf-8 -*-
 # Copyright 2015, 2016 OpenMarket Ltd
+# Copyright 2018 New Vector Ltd
 #
 # Licensed under the Apache License, Version 2.0 (the "License");
 # you may not use this file except in compliance with the License.
@@ -30,9 +31,10 @@ import synapse.metrics
 from synapse.types import get_domain_from_id
 from synapse.util import async
 from synapse.util.caches.response_cache import ResponseCache
-from synapse.util.logcontext import make_deferred_yieldable, preserve_fn
 from synapse.util.logutils import log_function
 
+from six import iteritems
+
 # when processing incoming transactions, we try to handle multiple rooms in
 # parallel, up to this limit.
 TRANSACTION_CONCURRENCY_LIMIT = 10
@@ -65,7 +67,7 @@ class FederationServer(FederationBase):
 
         # We cache responses to state queries, as they take a while and often
         # come in waves.
-        self._state_resp_cache = ResponseCache(hs, timeout_ms=30000)
+        self._state_resp_cache = ResponseCache(hs, "state_resp", timeout_ms=30000)
 
     @defer.inlineCallbacks
     @log_function
@@ -212,16 +214,17 @@ class FederationServer(FederationBase):
         if not in_room:
             raise AuthError(403, "Host not in room.")
 
-        result = self._state_resp_cache.get((room_id, event_id))
-        if not result:
-            with (yield self._server_linearizer.queue((origin, room_id))):
-                d = self._state_resp_cache.set(
-                    (room_id, event_id),
-                    preserve_fn(self._on_context_state_request_compute)(room_id, event_id)
-                )
-                resp = yield make_deferred_yieldable(d)
-        else:
-            resp = yield make_deferred_yieldable(result)
+        # we grab the linearizer to protect ourselves from servers which hammer
+        # us. In theory we might already have the response to this query
+        # in the cache so we could return it without waiting for the linearizer
+        # - but that's non-trivial to get right, and anyway somewhat defeats
+        # the point of the linearizer.
+        with (yield self._server_linearizer.queue((origin, room_id))):
+            resp = yield self._state_resp_cache.wrap(
+                (room_id, event_id),
+                self._on_context_state_request_compute,
+                room_id, event_id,
+            )
 
         defer.returnValue((200, resp))
 
@@ -425,9 +428,9 @@ class FederationServer(FederationBase):
             "Claimed one-time-keys: %s",
             ",".join((
                 "%s for %s:%s" % (key_id, user_id, device_id)
-                for user_id, user_keys in json_result.iteritems()
-                for device_id, device_keys in user_keys.iteritems()
-                for key_id, _ in device_keys.iteritems()
+                for user_id, user_keys in iteritems(json_result)
+                for device_id, device_keys in iteritems(user_keys)
+                for key_id, _ in iteritems(device_keys)
             )),
         )
 
@@ -494,13 +497,33 @@ class FederationServer(FederationBase):
     def _handle_received_pdu(self, origin, pdu):
         """ Process a PDU received in a federation /send/ transaction.
 
+        If the event is invalid, then this method throws a FederationError.
+        (The error will then be logged and sent back to the sender (which
+        probably won't do anything with it), and other events in the
+        transaction will be processed as normal).
+
+        It is likely that we'll then receive other events which refer to
+        this rejected_event in their prev_events, etc.  When that happens,
+        we'll attempt to fetch the rejected event again, which will presumably
+        fail, so those second-generation events will also get rejected.
+
+        Eventually, we get to the point where there are more than 10 events
+        between any new events and the original rejected event. Since we
+        only try to backfill 10 events deep on received pdu, we then accept the
+        new event, possibly introducing a discontinuity in the DAG, with new
+        forward extremities, so normal service is approximately returned,
+        until we try to backfill across the discontinuity.
+
         Args:
             origin (str): server which sent the pdu
             pdu (FrozenEvent): received pdu
 
         Returns (Deferred): completes with None
-        Raises: FederationError if the signatures / hash do not match
-    """
+
+        Raises: FederationError if the signatures / hash do not match, or
+            if the event was unacceptable for any other reason (eg, too large,
+            too many prev_events, couldn't find the prev_events)
+        """
         # check that it's actually being sent from a valid destination to
         # workaround bug #1753 in 0.18.5 and 0.18.6
         if origin != get_domain_from_id(pdu.event_id):