Add some comments

synapse/federation/federation_server.py

@@ -494,13 +494,33 @@ def _transaction_from_pdus(self, pdu_list):
     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 like 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):

synapse/handlers/federation.py

@@ -542,9 +542,16 @@ def redact_disallowed(event, state):
     def backfill(self, dest, room_id, limit, extremities):
         """ Trigger a backfill request to `dest` for the given `room_id`
 
-        This will attempt to get more events from the remote. This may return
-        be successfull and still return no events if the other side has no new
-        events to offer.
+        This will attempt to get more events from the remote. If the other side
+        has no new events to offer, this will return an empty list.
+
+        As the events are received, we check their signatures, and also do some
+        sanity-checking on them. If any of the backfilled events are invalid,
+        this method throws a SynapseError.
+
+        TODO: make this more useful to distinguish failures of the remote
+        server from invalid events (there is probably no point in trying to
+        re-fetch invalid events from every other HS in the room.)
         """
         if dest == self.server_name:
             raise SynapseError(400, "Can't backfill from self.")
@@ -556,6 +563,8 @@ def backfill(self, dest, room_id, limit, extremities):
             extremities=extremities,
         )
 
+        # do some sanity-checking of the received events, before we go and
+        # do state resolution across 1000 events.
         for ev in events:
             self._sanity_check_event(ev)