1 files changed, 29 insertions, 17 deletions
diff --git a/synapse/util/ratelimitutils.py b/synapse/util/ratelimitutils.py
index c5a45cef7c..7deb38f2a7 100644
--- a/synapse/util/ratelimitutils.py
+++ b/synapse/util/ratelimitutils.py
@@ -13,20 +13,19 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 
+import collections
+import contextlib
+import logging
+
 from twisted.internet import defer
 
 from synapse.api.errors import LimitExceededError
-
 from synapse.util.logcontext import (
-    run_in_background, make_deferred_yieldable,
     PreserveLoggingContext,
+    make_deferred_yieldable,
+    run_in_background,
 )
 
-import collections
-import contextlib
-import logging
-
-
 logger = logging.getLogger(__name__)
 
 
@@ -93,13 +92,22 @@ class _PerHostRatelimiter(object):
 
         self.window_size = window_size
         self.sleep_limit = sleep_limit
-        self.sleep_msec = sleep_msec
+        self.sleep_sec = sleep_msec / 1000.0
         self.reject_limit = reject_limit
         self.concurrent_requests = concurrent_requests
 
+        # request_id objects for requests which have been slept
         self.sleeping_requests = set()
+
+        # map from request_id object to Deferred for requests which are ready
+        # for processing but have been queued
         self.ready_request_queue = collections.OrderedDict()
+
+        # request id objects for requests which are in progress
         self.current_processing = set()
+
+        # times at which we have recently (within the last window_size ms)
+        # received requests.
         self.request_times = []
 
     @contextlib.contextmanager
@@ -118,11 +126,15 @@ class _PerHostRatelimiter(object):
 
     def _on_enter(self, request_id):
         time_now = self.clock.time_msec()
+
+        # remove any entries from request_times which aren't within the window
         self.request_times[:] = [
             r for r in self.request_times
             if time_now - r < self.window_size
         ]
 
+        # reject the request if we already have too many queued up (either
+        # sleeping or in the ready queue).
         queue_size = len(self.ready_request_queue) + len(self.sleeping_requests)
         if queue_size > self.reject_limit:
             raise LimitExceededError(
@@ -135,9 +147,13 @@ class _PerHostRatelimiter(object):
 
         def queue_request():
             if len(self.current_processing) > self.concurrent_requests:
-                logger.debug("Ratelimit [%s]: Queue req", id(request_id))
                 queue_defer = defer.Deferred()
                 self.ready_request_queue[request_id] = queue_defer
+                logger.info(
+                    "Ratelimiter: queueing request (queue now %i items)",
+                    len(self.ready_request_queue),
+                )
+
                 return queue_defer
             else:
                 return defer.succeed(None)
@@ -149,10 +165,9 @@ class _PerHostRatelimiter(object):
 
         if len(self.request_times) > self.sleep_limit:
             logger.debug(
-                "Ratelimit [%s]: sleeping req",
-                id(request_id),
+                "Ratelimiter: sleeping request for %f sec", self.sleep_sec,
             )
-            ret_defer = run_in_background(self.clock.sleep, self.sleep_msec / 1000.0)
+            ret_defer = run_in_background(self.clock.sleep, self.sleep_sec)
 
             self.sleeping_requests.add(request_id)
 
@@ -201,11 +216,8 @@ class _PerHostRatelimiter(object):
         )
         self.current_processing.discard(request_id)
         try:
-            request_id, deferred = self.ready_request_queue.popitem()
-
-            # XXX: why do we do the following? the on_start callback above will
-            # do it for us.
-            self.current_processing.add(request_id)
+            # start processing the next item on the queue.
+            _, deferred = self.ready_request_queue.popitem(last=False)
 
             with PreserveLoggingContext():
                 deferred.callback(None)