Move FederationServer._handle_new_pdu to FederationHandler

Unfortunately this significantly increases the size of the already-rather-big
FederationHandler, but the code fits more naturally here, and it paves the way
for the tighter integration that I need between handling incoming PDUs and
doing the join dance.

Other than renaming the existing `FederationHandler.on_receive_pdu` to
`_process_received_pdu` to make way for it, this just consists of the move, and
replacing `self.handler` with `self` and `self` with `self.replication_layer`.

1 files changed, 197 insertions, 5 deletions

diff --git a/synapse/handlers/federation.py b/synapse/handlers/federation.py
index ed0fa51e7f..d0c2b4d6ed 100644
--- a/synapse/handlers/federation.py
+++ b/synapse/handlers/federation.py
@@ -31,7 +31,7 @@ from synapse.util.logcontext import (
 )
 from synapse.util.metrics import measure_func
 from synapse.util.logutils import log_function
-from synapse.util.async import run_on_reactor
+from synapse.util.async import run_on_reactor, Linearizer
 from synapse.util.frozenutils import unfreeze
 from synapse.crypto.event_signing import (
     compute_event_signature, add_hashes_and_signatures,
@@ -79,12 +79,204 @@ class FederationHandler(BaseHandler):
 
         # When joining a room we need to queue any events for that room up
         self.room_queues = {}
+        self._room_pdu_linearizer = Linearizer("fed_room_pdu")
+
+    @defer.inlineCallbacks
+    @log_function
+    def on_receive_pdu(self, origin, pdu, get_missing=True):
+        """ Process a PDU received via a federation /send/ transaction, or
+        via backfill of missing prev_events
+
+        Args:
+            origin (str): server which initiated the /send/ transaction. Will
+                be used to fetch missing events or state.
+            pdu (FrozenEvent): received PDU
+            get_missing (bool): True if we should fetch missing prev_events
+
+        Returns (Deferred): completes with None
+        """
+
+        # We reprocess pdus when we have seen them only as outliers
+        existing = yield self.get_persisted_pdu(
+            origin, pdu.event_id, do_auth=False
+        )
+
+        # FIXME: Currently we fetch an event again when we already have it
+        # if it has been marked as an outlier.
+
+        already_seen = (
+            existing and (
+                not existing.internal_metadata.is_outlier()
+                or pdu.internal_metadata.is_outlier()
+            )
+        )
+        if already_seen:
+            logger.debug("Already seen pdu %s", pdu.event_id)
+            return
+
+        state = None
+
+        auth_chain = []
+
+        have_seen = yield self.store.have_events(
+            [ev for ev, _ in pdu.prev_events]
+        )
+
+        fetch_state = False
+
+        # Get missing pdus if necessary.
+        if not pdu.internal_metadata.is_outlier():
+            # We only backfill backwards to the min depth.
+            min_depth = yield self.get_min_depth_for_context(
+                pdu.room_id
+            )
+
+            logger.debug(
+                "_handle_new_pdu min_depth for %s: %d",
+                pdu.room_id, min_depth
+            )
+
+            prevs = {e_id for e_id, _ in pdu.prev_events}
+            seen = set(have_seen.keys())
+
+            if min_depth and pdu.depth < min_depth:
+                # This is so that we don't notify the user about this
+                # message, to work around the fact that some events will
+                # reference really really old events we really don't want to
+                # send to the clients.
+                pdu.internal_metadata.outlier = True
+            elif min_depth and pdu.depth > min_depth:
+                if get_missing and prevs - seen:
+                    # If we're missing stuff, ensure we only fetch stuff one
+                    # at a time.
+                    logger.info(
+                        "Acquiring lock for room %r to fetch %d missing events: %r...",
+                        pdu.room_id, len(prevs - seen), list(prevs - seen)[:5],
+                    )
+                    with (yield self._room_pdu_linearizer.queue(pdu.room_id)):
+                        logger.info(
+                            "Acquired lock for room %r to fetch %d missing events",
+                            pdu.room_id, len(prevs - seen),
+                        )
+
+                        yield self._get_missing_events_for_pdu(
+                            origin, pdu, prevs, min_depth
+                        )
+
+            prevs = {e_id for e_id, _ in pdu.prev_events}
+            seen = set(have_seen.keys())
+            if prevs - seen:
+                logger.info(
+                    "Still missing %d events for room %r: %r...",
+                    len(prevs - seen), pdu.room_id, list(prevs - seen)[:5]
+                )
+                fetch_state = True
+
+        if fetch_state:
+            # We need to get the state at this event, since we haven't
+            # processed all the prev events.
+            logger.debug(
+                "_handle_new_pdu getting state for %s",
+                pdu.room_id
+            )
+            try:
+                state, auth_chain = yield self.replication_layer.get_state_for_room(
+                    origin, pdu.room_id, pdu.event_id,
+                )
+            except:
+                logger.exception("Failed to get state for event: %s", pdu.event_id)
+
+        yield self._process_received_pdu(
+            origin,
+            pdu,
+            state=state,
+            auth_chain=auth_chain,
+        )
+
+    @defer.inlineCallbacks
+    def _get_missing_events_for_pdu(self, origin, pdu, prevs, min_depth):
+        """
+        Args:
+            origin (str): Origin of the pdu. Will be called to get the missing events
+            pdu: received pdu
+            prevs (str[]): List of event ids which we are missing
+            min_depth (int): Minimum depth of events to return.
+
+        Returns:
+            Deferred<dict(str, str?)>: updated have_seen dictionary
+        """
+        # We recalculate seen, since it may have changed.
+        have_seen = yield self.store.have_events(prevs)
+        seen = set(have_seen.keys())
+
+        if not prevs - seen:
+            # nothing left to do
+            defer.returnValue(have_seen)
+
+        latest = yield self.store.get_latest_event_ids_in_room(
+            pdu.room_id
+        )
+
+        # We add the prev events that we have seen to the latest
+        # list to ensure the remote server doesn't give them to us
+        latest = set(latest)
+        latest |= seen
+
+        logger.info(
+            "Missing %d events for room %r: %r...",
+            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.replication_layer.get_missing_events(
+            origin,
+            pdu.room_id,
+            earliest_events_ids=list(latest),
+            latest_events=[pdu],
+            limit=10,
+            min_depth=min_depth,
+            timeout=10000,
+        )
+
+        # We want to sort these by depth so we process them and
+        # tell clients about them in order.
+        missing_events.sort(key=lambda x: x.depth)
+
+        for e in missing_events:
+            yield self.on_receive_pdu(
+                origin,
+                e,
+                get_missing=False
+            )
+
+        have_seen = yield self.store.have_events(
+            [ev for ev, _ in pdu.prev_events]
+        )
+        defer.returnValue(have_seen)
 
     @log_function
     @defer.inlineCallbacks
-    def on_receive_pdu(self, origin, pdu, state=None, auth_chain=None):
-        """ Called by the ReplicationLayer when we have a new pdu. We need to
-        do auth checks and put it through the StateHandler.
+    def _process_received_pdu(self, origin, pdu, state=None, auth_chain=None):
+        """ Called when we have a new pdu. We need to do auth checks and put it
+        through the StateHandler.
 
         auth_chain and state are None if we already have the necessary state
         and prev_events in the db
@@ -738,7 +930,7 @@ class FederationHandler(BaseHandler):
                     continue
 
                 try:
-                    self.on_receive_pdu(origin, p)
+                    self._process_received_pdu(origin, p)
                 except:
                     logger.exception("Couldn't handle pdu")