23 files changed, 334 insertions, 204 deletions

diff --git a/synapse/util/__init__.py b/synapse/util/__init__.py
index fc11e26623..680ea928c7 100644
--- a/synapse/util/__init__.py
+++ b/synapse/util/__init__.py
@@ -13,14 +13,14 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 
-from synapse.util.logcontext import PreserveLoggingContext
+import logging
+from itertools import islice
 
-from twisted.internet import defer, reactor, task
+import attr
 
-import time
-import logging
+from twisted.internet import defer, task
 
-from itertools import islice
+from synapse.util.logcontext import PreserveLoggingContext
 
 logger = logging.getLogger(__name__)
 
@@ -31,16 +31,27 @@ def unwrapFirstError(failure):
     return failure.value.subFailure
 
 
+@attr.s
 class Clock(object):
-    """A small utility that obtains current time-of-day so that time may be
-    mocked during unit-tests.
+    """
+    A Clock wraps a Twisted reactor and provides utilities on top of it.
 
-    TODO(paul): Also move the sleep() functionality into it
+    Args:
+        reactor: The Twisted reactor to use.
     """
+    _reactor = attr.ib()
+
+    @defer.inlineCallbacks
+    def sleep(self, seconds):
+        d = defer.Deferred()
+        with PreserveLoggingContext():
+            self._reactor.callLater(seconds, d.callback, seconds)
+            res = yield d
+        defer.returnValue(res)
 
     def time(self):
         """Returns the current system time in seconds since epoch."""
-        return time.time()
+        return self._reactor.seconds()
 
     def time_msec(self):
         """Returns the current system time in miliseconds since epoch."""
@@ -56,6 +67,7 @@ class Clock(object):
             msec(float): How long to wait between calls in milliseconds.
         """
         call = task.LoopingCall(f)
+        call.clock = self._reactor
         call.start(msec / 1000.0, now=False)
         return call
 
@@ -73,7 +85,7 @@ class Clock(object):
                 callback(*args, **kwargs)
 
         with PreserveLoggingContext():
-            return reactor.callLater(delay, wrapped_callback, *args, **kwargs)
+            return self._reactor.callLater(delay, wrapped_callback, *args, **kwargs)
 
     def cancel_call_later(self, timer, ignore_errs=False):
         try:
diff --git a/synapse/util/async.py b/synapse/util/async.py
index 9dd4e6b5bc..5d0fb39130 100644
--- a/synapse/util/async.py
+++ b/synapse/util/async.py
@@ -13,41 +13,26 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 
+import logging
+from contextlib import contextmanager
 
-from twisted.internet import defer, reactor
+from six.moves import range
+
+from twisted.internet import defer
 from twisted.internet.defer import CancelledError
 from twisted.python import failure
 
+from synapse.util import Clock, logcontext, unwrapFirstError
+
 from .logcontext import (
-    PreserveLoggingContext, make_deferred_yieldable, run_in_background
+    PreserveLoggingContext,
+    make_deferred_yieldable,
+    run_in_background,
 )
-from synapse.util import logcontext, unwrapFirstError
-
-from contextlib import contextmanager
-
-import logging
-
-from six.moves import range
 
 logger = logging.getLogger(__name__)
 
 
-@defer.inlineCallbacks
-def sleep(seconds):
-    d = defer.Deferred()
-    with PreserveLoggingContext():
-        reactor.callLater(seconds, d.callback, seconds)
-        res = yield d
-    defer.returnValue(res)
-
-
-def run_on_reactor():
-    """ This will cause the rest of the function to be invoked upon the next
-    iteration of the main loop
-    """
-    return sleep(0)
-
-
 class ObservableDeferred(object):
     """Wraps a deferred object so that we can add observer deferreds. These
     observer deferreds do not affect the callback chain of the original
@@ -180,13 +165,18 @@ class Linearizer(object):
             # do some work.
 
     """
-    def __init__(self, name=None):
+    def __init__(self, name=None, clock=None):
         if name is None:
             self.name = id(self)
         else:
             self.name = name
         self.key_to_defer = {}
 
+        if not clock:
+            from twisted.internet import reactor
+            clock = Clock(reactor)
+        self._clock = clock
+
     @defer.inlineCallbacks
     def queue(self, key):
         # If there is already a deferred in the queue, we pull it out so that
@@ -227,7 +217,7 @@ class Linearizer(object):
             # the context manager, but it needs to happen while we hold the
             # lock, and the context manager's exit code must be synchronous,
             # so actually this is the only sensible place.
-            yield run_on_reactor()
+            yield self._clock.sleep(0)
 
         else:
             logger.info("Acquired uncontended linearizer lock %r for key %r",
@@ -404,7 +394,7 @@ class DeferredTimeoutError(Exception):
     """
 
 
-def add_timeout_to_deferred(deferred, timeout, on_timeout_cancel=None):
+def add_timeout_to_deferred(deferred, timeout, reactor, on_timeout_cancel=None):
     """
     Add a timeout to a deferred by scheduling it to be cancelled after
     timeout seconds.
@@ -419,6 +409,7 @@ def add_timeout_to_deferred(deferred, timeout, on_timeout_cancel=None):
     Args:
         deferred (defer.Deferred): deferred to be timed out
         timeout (Number): seconds to time out after
+        reactor (twisted.internet.reactor): the Twisted reactor to use
 
         on_timeout_cancel (callable): A callable which is called immediately
             after the deferred times out, and not if this deferred is
diff --git a/synapse/util/caches/__init__.py b/synapse/util/caches/__init__.py
index 183faf75a1..7b065b195e 100644
--- a/synapse/util/caches/__init__.py
+++ b/synapse/util/caches/__init__.py
@@ -13,15 +13,25 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 
-from prometheus_client.core import Gauge, REGISTRY, GaugeMetricFamily
-
 import os
 
-from six.moves import intern
 import six
+from six.moves import intern
+
+from prometheus_client.core import REGISTRY, Gauge, GaugeMetricFamily
 
 CACHE_SIZE_FACTOR = float(os.environ.get("SYNAPSE_CACHE_FACTOR", 0.5))
 
+
+def get_cache_factor_for(cache_name):
+    env_var = "SYNAPSE_CACHE_FACTOR_" + cache_name.upper()
+    factor = os.environ.get(env_var)
+    if factor:
+        return float(factor)
+
+    return CACHE_SIZE_FACTOR
+
+
 caches_by_name = {}
 collectors_by_name = {}
 
diff --git a/synapse/util/caches/descriptors.py b/synapse/util/caches/descriptors.py
index 8a9dcb2fc2..f8a07df6b8 100644
--- a/synapse/util/caches/descriptors.py
+++ b/synapse/util/caches/descriptors.py
@@ -13,25 +13,26 @@
 # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
 # See the License for the specific language governing permissions and
 # limitations under the License.
+import functools
+import inspect
 import logging
+import threading
+from collections import namedtuple
 
+import six
+from six import itervalues, string_types
+
+from twisted.internet import defer
+
+from synapse.util import logcontext, unwrapFirstError
 from synapse.util.async import ObservableDeferred
-from synapse.util import unwrapFirstError, logcontext
-from synapse.util.caches import CACHE_SIZE_FACTOR
+from synapse.util.caches import get_cache_factor_for
 from synapse.util.caches.lrucache import LruCache
 from synapse.util.caches.treecache import TreeCache, iterate_tree_cache_entry
 from synapse.util.stringutils import to_ascii
 
 from . import register_cache
 
-from twisted.internet import defer
-from collections import namedtuple
-
-import functools
-import inspect
-import threading
-
-
 logger = logging.getLogger(__name__)
 
 
@@ -205,7 +206,7 @@ class Cache(object):
     def invalidate_all(self):
         self.check_thread()
         self.cache.clear()
-        for entry in self._pending_deferred_cache.itervalues():
+        for entry in itervalues(self._pending_deferred_cache):
             entry.invalidate()
         self._pending_deferred_cache.clear()
 
@@ -310,7 +311,7 @@ class CacheDescriptor(_CacheDescriptorBase):
             orig, num_args=num_args, inlineCallbacks=inlineCallbacks,
             cache_context=cache_context)
 
-        max_entries = int(max_entries * CACHE_SIZE_FACTOR)
+        max_entries = int(max_entries * get_cache_factor_for(orig.__name__))
 
         self.max_entries = max_entries
         self.tree = tree
@@ -392,9 +393,10 @@ class CacheDescriptor(_CacheDescriptorBase):
 
                 ret.addErrback(onErr)
 
-                # If our cache_key is a string, try to convert to ascii to save
-                # a bit of space in large caches
-                if isinstance(cache_key, basestring):
+                # If our cache_key is a string on py2, try to convert to ascii
+                # to save a bit of space in large caches. Py3 does this
+                # internally automatically.
+                if six.PY2 and isinstance(cache_key, string_types):
                     cache_key = to_ascii(cache_key)
 
                 result_d = ObservableDeferred(ret, consumeErrors=True)
@@ -565,7 +567,7 @@ class CacheListDescriptor(_CacheDescriptorBase):
                     return results
 
                 return logcontext.make_deferred_yieldable(defer.gatherResults(
-                    cached_defers.values(),
+                    list(cached_defers.values()),
                     consumeErrors=True,
                 ).addCallback(update_results_dict).addErrback(
                     unwrapFirstError
diff --git a/synapse/util/caches/dictionary_cache.py b/synapse/util/caches/dictionary_cache.py
index bdc21e348f..6c0b5a4094 100644
--- a/synapse/util/caches/dictionary_cache.py
+++ b/synapse/util/caches/dictionary_cache.py
@@ -13,12 +13,13 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 
-from synapse.util.caches.lrucache import LruCache
-from collections import namedtuple
-from . import register_cache
-import threading
 import logging
+import threading
+from collections import namedtuple
 
+from synapse.util.caches.lrucache import LruCache
+
+from . import register_cache
 
 logger = logging.getLogger(__name__)
 
@@ -107,29 +108,28 @@ class DictionaryCache(object):
         self.sequence += 1
         self.cache.clear()
 
-    def update(self, sequence, key, value, full=False, known_absent=None):
+    def update(self, sequence, key, value, fetched_keys=None):
         """Updates the entry in the cache
 
         Args:
             sequence
-            key
-            value (dict): The value to update the cache with.
-            full (bool): Whether the given value is the full dict, or just a
-                partial subset there of. If not full then any existing entries
-                for the key will be updated.
-            known_absent (set): Set of keys that we know don't exist in the full
-                dict.
+            key (K)
+            value (dict[X,Y]): The value to update the cache with.
+            fetched_keys (None|set[X]): All of the dictionary keys which were
+                fetched from the database.
+
+                If None, this is the complete value for key K. Otherwise, it
+                is used to infer a list of keys which we know don't exist in
+                the full dict.
         """
         self.check_thread()
         if self.sequence == sequence:
             # Only update the cache if the caches sequence number matches the
             # number that the cache had before the SELECT was started (SYN-369)
-            if known_absent is None:
-                known_absent = set()
-            if full:
-                self._insert(key, value, known_absent)
+            if fetched_keys is None:
+                self._insert(key, value, set())
             else:
-                self._update_or_insert(key, value, known_absent)
+                self._update_or_insert(key, value, fetched_keys)
 
     def _update_or_insert(self, key, value, known_absent):
         # We pop and reinsert as we need to tell the cache the size may have
diff --git a/synapse/util/caches/expiringcache.py b/synapse/util/caches/expiringcache.py
index ff04c91955..4abca91f6d 100644
--- a/synapse/util/caches/expiringcache.py
+++ b/synapse/util/caches/expiringcache.py
@@ -13,11 +13,10 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 
-from synapse.util.caches import register_cache
-
-from collections import OrderedDict
 import logging
+from collections import OrderedDict
 
+from synapse.util.caches import register_cache
 
 logger = logging.getLogger(__name__)
 
diff --git a/synapse/util/caches/lrucache.py b/synapse/util/caches/lrucache.py
index 1c5a982094..b684f24e7b 100644
--- a/synapse/util/caches/lrucache.py
+++ b/synapse/util/caches/lrucache.py
@@ -14,8 +14,8 @@
 # limitations under the License.
 
 
-from functools import wraps
 import threading
+from functools import wraps
 
 from synapse.util.caches.treecache import TreeCache
 
diff --git a/synapse/util/caches/stream_change_cache.py b/synapse/util/caches/stream_change_cache.py
index a7fe0397fa..8637867c6d 100644
--- a/synapse/util/caches/stream_change_cache.py
+++ b/synapse/util/caches/stream_change_cache.py
@@ -13,12 +13,11 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 
-from synapse.util.caches import register_cache, CACHE_SIZE_FACTOR
-
-
-from blist import sorteddict
 import logging
 
+from sortedcontainers import SortedDict
+
+from synapse.util import caches
 
 logger = logging.getLogger(__name__)
 
@@ -32,16 +31,18 @@ class StreamChangeCache(object):
     entities that may have changed since that position. If position key is too
     old then the cache will simply return all given entities.
     """
-    def __init__(self, name, current_stream_pos, max_size=10000, prefilled_cache={}):
-        self._max_size = int(max_size * CACHE_SIZE_FACTOR)
+
+    def __init__(self, name, current_stream_pos, max_size=10000, prefilled_cache=None):
+        self._max_size = int(max_size * caches.CACHE_SIZE_FACTOR)
         self._entity_to_key = {}
-        self._cache = sorteddict()
+        self._cache = SortedDict()
         self._earliest_known_stream_pos = current_stream_pos
         self.name = name
-        self.metrics = register_cache("cache", self.name, self._cache)
+        self.metrics = caches.register_cache("cache", self.name, self._cache)
 
-        for entity, stream_pos in prefilled_cache.items():
-            self.entity_has_changed(entity, stream_pos)
+        if prefilled_cache:
+            for entity, stream_pos in prefilled_cache.items():
+                self.entity_has_changed(entity, stream_pos)
 
     def has_entity_changed(self, entity, stream_pos):
         """Returns True if the entity may have been updated since stream_pos
@@ -65,22 +66,26 @@ class StreamChangeCache(object):
         return False
 
     def get_entities_changed(self, entities, stream_pos):
-        """Returns subset of entities that have had new things since the
-        given position. If the position is too old it will just return the given list.
+        """
+        Returns subset of entities that have had new things since the given
+        position.  Entities unknown to the cache will be returned.  If the
+        position is too old it will just return the given list.
         """
         assert type(stream_pos) is int
 
         if stream_pos >= self._earliest_known_stream_pos:
-            keys = self._cache.keys()
-            i = keys.bisect_right(stream_pos)
+            not_known_entities = set(entities) - set(self._entity_to_key)
 
-            result = set(
-                self._cache[k] for k in keys[i:]
-            ).intersection(entities)
+            result = (
+                {self._cache[k] for k in self._cache.islice(
+                    start=self._cache.bisect_right(stream_pos))}
+                .intersection(entities)
+                .union(not_known_entities)
+            )
 
             self.metrics.inc_hits()
         else:
-            result = entities
+            result = set(entities)
             self.metrics.inc_misses()
 
         return result
@@ -90,12 +95,13 @@ class StreamChangeCache(object):
         """
         assert type(stream_pos) is int
 
+        if not self._cache:
+            # If we have no cache, nothing can have changed.
+            return False
+
         if stream_pos >= self._earliest_known_stream_pos:
             self.metrics.inc_hits()
-            keys = self._cache.keys()
-            i = keys.bisect_right(stream_pos)
-
-            return i < len(keys)
+            return self._cache.bisect_right(stream_pos) < len(self._cache)
         else:
             self.metrics.inc_misses()
             return True
@@ -107,10 +113,8 @@ class StreamChangeCache(object):
         assert type(stream_pos) is int
 
         if stream_pos >= self._earliest_known_stream_pos:
-            keys = self._cache.keys()
-            i = keys.bisect_right(stream_pos)
-
-            return [self._cache[k] for k in keys[i:]]
+            return [self._cache[k] for k in self._cache.islice(
+                start=self._cache.bisect_right(stream_pos))]
         else:
             return None
 
@@ -129,8 +133,10 @@ class StreamChangeCache(object):
             self._entity_to_key[entity] = stream_pos
 
             while len(self._cache) > self._max_size:
-                k, r = self._cache.popitem()
-                self._earliest_known_stream_pos = max(k, self._earliest_known_stream_pos)
+                k, r = self._cache.popitem(0)
+                self._earliest_known_stream_pos = max(
+                    k, self._earliest_known_stream_pos,
+                )
                 self._entity_to_key.pop(r, None)
 
     def get_max_pos_of_last_change(self, entity):
diff --git a/synapse/util/caches/treecache.py b/synapse/util/caches/treecache.py
index fcc341a6b7..dd4c9e6067 100644
--- a/synapse/util/caches/treecache.py
+++ b/synapse/util/caches/treecache.py
@@ -1,3 +1,5 @@
+from six import itervalues
+
 SENTINEL = object()
 
 
@@ -49,7 +51,7 @@ class TreeCache(object):
         if popped is SENTINEL:
             return default
 
-        node_and_keys = zip(nodes, key)
+        node_and_keys = list(zip(nodes, key))
         node_and_keys.reverse()
         node_and_keys.append((self.root, None))
 
@@ -76,7 +78,7 @@ def iterate_tree_cache_entry(d):
     can contain dicts.
     """
     if isinstance(d, dict):
-        for value_d in d.itervalues():
+        for value_d in itervalues(d):
             for value in iterate_tree_cache_entry(value_d):
                 yield value
     else:
diff --git a/synapse/util/file_consumer.py b/synapse/util/file_consumer.py
index 3380970e4e..629ed44149 100644
--- a/synapse/util/file_consumer.py
+++ b/synapse/util/file_consumer.py
@@ -13,11 +13,11 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 
-from twisted.internet import threads, reactor
+from six.moves import queue
 
-from synapse.util.logcontext import make_deferred_yieldable, run_in_background
+from twisted.internet import threads
 
-from six.moves import queue
+from synapse.util.logcontext import make_deferred_yieldable, run_in_background
 
 
 class BackgroundFileConsumer(object):
@@ -27,6 +27,7 @@ class BackgroundFileConsumer(object):
     Args:
         file_obj (file): The file like object to write to. Closed when
             finished.
+        reactor (twisted.internet.reactor): the Twisted reactor to use
     """
 
     # For PushProducers pause if we have this many unwritten slices
@@ -34,9 +35,11 @@ class BackgroundFileConsumer(object):
     # And resume once the size of the queue is less than this
     _RESUME_ON_QUEUE_SIZE = 2
 
-    def __init__(self, file_obj):
+    def __init__(self, file_obj, reactor):
         self._file_obj = file_obj
 
+        self._reactor = reactor
+
         # Producer we're registered with
         self._producer = None
 
@@ -71,7 +74,10 @@ class BackgroundFileConsumer(object):
         self._producer = producer
         self.streaming = streaming
         self._finished_deferred = run_in_background(
-            threads.deferToThread, self._writer
+            threads.deferToThreadPool,
+            self._reactor,
+            self._reactor.getThreadPool(),
+            self._writer,
         )
         if not streaming:
             self._producer.resumeProducing()
@@ -109,7 +115,7 @@ class BackgroundFileConsumer(object):
                 # producer.
                 if self._producer and self._paused_producer:
                     if self._bytes_queue.qsize() <= self._RESUME_ON_QUEUE_SIZE:
-                        reactor.callFromThread(self._resume_paused_producer)
+                        self._reactor.callFromThread(self._resume_paused_producer)
 
                 bytes = self._bytes_queue.get()
 
@@ -121,7 +127,7 @@ class BackgroundFileConsumer(object):
                 # If its a pull producer then we need to explicitly ask for
                 # more stuff.
                 if not self.streaming and self._producer:
-                    reactor.callFromThread(self._producer.resumeProducing)
+                    self._reactor.callFromThread(self._producer.resumeProducing)
         except Exception as e:
             self._write_exception = e
             raise
diff --git a/synapse/util/frozenutils.py b/synapse/util/frozenutils.py
index f497b51f4a..581c6052ac 100644
--- a/synapse/util/frozenutils.py
+++ b/synapse/util/frozenutils.py
@@ -13,19 +13,20 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 
+from six import string_types
+
+from canonicaljson import json
 from frozendict import frozendict
-import simplejson as json
 
 
 def freeze(o):
-    t = type(o)
-    if t is dict:
+    if isinstance(o, dict):
         return frozendict({k: freeze(v) for k, v in o.items()})
 
-    if t is frozendict:
+    if isinstance(o, frozendict):
         return o
 
-    if t is str or t is unicode:
+    if isinstance(o, string_types):
         return o
 
     try:
@@ -37,11 +38,10 @@ def freeze(o):
 
 
 def unfreeze(o):
-    t = type(o)
-    if t is dict or t is frozendict:
+    if isinstance(o, (dict, frozendict)):
         return dict({k: unfreeze(v) for k, v in o.items()})
 
-    if t is str or t is unicode:
+    if isinstance(o, string_types):
         return o
 
     try:
diff --git a/synapse/util/httpresourcetree.py b/synapse/util/httpresourcetree.py
index e9f0f292ee..2d7ddc1cbe 100644
--- a/synapse/util/httpresourcetree.py
+++ b/synapse/util/httpresourcetree.py
@@ -12,10 +12,10 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 
-from twisted.web.resource import NoResource
-
 import logging
 
+from twisted.web.resource import NoResource
+
 logger = logging.getLogger(__name__)
 
 
diff --git a/synapse/util/logcontext.py b/synapse/util/logcontext.py
index a58c723403..f6c7175f74 100644
--- a/synapse/util/logcontext.py
+++ b/synapse/util/logcontext.py
@@ -22,10 +22,10 @@ them.
 See doc/log_contexts.rst for details on how this works.
 """
 
-from twisted.internet import defer
-
-import threading
 import logging
+import threading
+
+from twisted.internet import defer
 
 logger = logging.getLogger(__name__)
 
@@ -49,17 +49,107 @@ except Exception:
         return None
 
 
+class ContextResourceUsage(object):
+    """Object for tracking the resources used by a log context
+
+    Attributes:
+        ru_utime (float): user CPU time (in seconds)
+        ru_stime (float): system CPU time (in seconds)
+        db_txn_count (int): number of database transactions done
+        db_sched_duration_sec (float): amount of time spent waiting for a
+            database connection
+        db_txn_duration_sec (float): amount of time spent doing database
+            transactions (excluding scheduling time)
+        evt_db_fetch_count (int): number of events requested from the database
+    """
+
+    __slots__ = [
+        "ru_stime", "ru_utime",
+        "db_txn_count", "db_txn_duration_sec", "db_sched_duration_sec",
+        "evt_db_fetch_count",
+    ]
+
+    def __init__(self, copy_from=None):
+        """Create a new ContextResourceUsage
+
+        Args:
+            copy_from (ContextResourceUsage|None): if not None, an object to
+                copy stats from
+        """
+        if copy_from is None:
+            self.reset()
+        else:
+            self.ru_utime = copy_from.ru_utime
+            self.ru_stime = copy_from.ru_stime
+            self.db_txn_count = copy_from.db_txn_count
+
+            self.db_txn_duration_sec = copy_from.db_txn_duration_sec
+            self.db_sched_duration_sec = copy_from.db_sched_duration_sec
+            self.evt_db_fetch_count = copy_from.evt_db_fetch_count
+
+    def copy(self):
+        return ContextResourceUsage(copy_from=self)
+
+    def reset(self):
+        self.ru_stime = 0.
+        self.ru_utime = 0.
+        self.db_txn_count = 0
+
+        self.db_txn_duration_sec = 0
+        self.db_sched_duration_sec = 0
+        self.evt_db_fetch_count = 0
+
+    def __iadd__(self, other):
+        """Add another ContextResourceUsage's stats to this one's.
+
+        Args:
+            other (ContextResourceUsage): the other resource usage object
+        """
+        self.ru_utime += other.ru_utime
+        self.ru_stime += other.ru_stime
+        self.db_txn_count += other.db_txn_count
+        self.db_txn_duration_sec += other.db_txn_duration_sec
+        self.db_sched_duration_sec += other.db_sched_duration_sec
+        self.evt_db_fetch_count += other.evt_db_fetch_count
+        return self
+
+    def __isub__(self, other):
+        self.ru_utime -= other.ru_utime
+        self.ru_stime -= other.ru_stime
+        self.db_txn_count -= other.db_txn_count
+        self.db_txn_duration_sec -= other.db_txn_duration_sec
+        self.db_sched_duration_sec -= other.db_sched_duration_sec
+        self.evt_db_fetch_count -= other.evt_db_fetch_count
+        return self
+
+    def __add__(self, other):
+        res = ContextResourceUsage(copy_from=self)
+        res += other
+        return res
+
+    def __sub__(self, other):
+        res = ContextResourceUsage(copy_from=self)
+        res -= other
+        return res
+
+
 class LoggingContext(object):
     """Additional context for log formatting. Contexts are scoped within a
     "with" block.
 
+    If a parent is given when creating a new context, then:
+        - logging fields are copied from the parent to the new context on entry
+        - when the new context exits, the cpu usage stats are copied from the
+          child to the parent
+
     Args:
         name (str): Name for the context for debugging.
+        parent_context (LoggingContext|None): The parent of the new context
     """
 
     __slots__ = [
-        "previous_context", "name", "ru_stime", "ru_utime",
-        "db_txn_count", "db_txn_duration_sec", "db_sched_duration_sec",
+        "previous_context", "name", "parent_context",
+        "_resource_usage",
         "usage_start",
         "main_thread", "alive",
         "request", "tag",
@@ -90,24 +180,21 @@ class LoggingContext(object):
         def add_database_scheduled(self, sched_sec):
             pass
 
+        def record_event_fetch(self, event_count):
+            pass
+
         def __nonzero__(self):
             return False
         __bool__ = __nonzero__  # python3
 
     sentinel = Sentinel()
 
-    def __init__(self, name=None):
+    def __init__(self, name=None, parent_context=None):
         self.previous_context = LoggingContext.current_context()
         self.name = name
-        self.ru_stime = 0.
-        self.ru_utime = 0.
-        self.db_txn_count = 0
-
-        # sec spent waiting for db txns, excluding scheduling time
-        self.db_txn_duration_sec = 0
 
-        # sec spent waiting for db txns to be scheduled
-        self.db_sched_duration_sec = 0
+        # track the resources used by this context so far
+        self._resource_usage = ContextResourceUsage()
 
         # If alive has the thread resource usage when the logcontext last
         # became active.
@@ -118,6 +205,8 @@ class LoggingContext(object):
         self.tag = ""
         self.alive = True
 
+        self.parent_context = parent_context
+
     def __str__(self):
         return "%s@%x" % (self.name, id(self))
 
@@ -155,6 +244,10 @@ class LoggingContext(object):
                 self.previous_context, old_context
             )
         self.alive = True
+
+        if self.parent_context is not None:
+            self.parent_context.copy_to(self)
+
         return self
 
     def __exit__(self, type, value, traceback):
@@ -176,6 +269,13 @@ class LoggingContext(object):
         self.previous_context = None
         self.alive = False
 
+        # if we have a parent, pass our CPU usage stats on
+        if self.parent_context is not None:
+            self.parent_context._resource_usage += self._resource_usage
+
+            # reset them in case we get entered again
+            self._resource_usage.reset()
+
     def copy_to(self, record):
         """Copy logging fields from this context to a log record or
         another LoggingContext
@@ -200,39 +300,43 @@ class LoggingContext(object):
             logger.warning("Stopped logcontext %s on different thread", self)
             return
 
-        # When we stop, let's record the resource used since we started
-        if self.usage_start:
-            usage_end = get_thread_resource_usage()
+        # When we stop, let's record the cpu used since we started
+        if not self.usage_start:
+            logger.warning(
+                "Called stop on logcontext %s without calling start", self,
+            )
+            return
+
+        usage_end = get_thread_resource_usage()
 
-            self.ru_utime += usage_end.ru_utime - self.usage_start.ru_utime
-            self.ru_stime += usage_end.ru_stime - self.usage_start.ru_stime
+        self._resource_usage.ru_utime += usage_end.ru_utime - self.usage_start.ru_utime
+        self._resource_usage.ru_stime += usage_end.ru_stime - self.usage_start.ru_stime
 
-            self.usage_start = None
-        else:
-            logger.warning("Called stop on logcontext %s without calling start", self)
+        self.usage_start = None
 
     def get_resource_usage(self):
-        """Get CPU time used by this logcontext so far.
+        """Get resources used by this logcontext so far.
 
         Returns:
-            tuple[float, float]: The user and system CPU usage in seconds
+            ContextResourceUsage: a *copy* of the object tracking resource
+                usage so far
         """
-        ru_utime = self.ru_utime
-        ru_stime = self.ru_stime
+        # we always return a copy, for consistency
+        res = self._resource_usage.copy()
 
         # If we are on the correct thread and we're currently running then we
         # can include resource usage so far.
         is_main_thread = threading.current_thread() is self.main_thread
         if self.alive and self.usage_start and is_main_thread:
             current = get_thread_resource_usage()
-            ru_utime += current.ru_utime - self.usage_start.ru_utime
-            ru_stime += current.ru_stime - self.usage_start.ru_stime
+            res.ru_utime += current.ru_utime - self.usage_start.ru_utime
+            res.ru_stime += current.ru_stime - self.usage_start.ru_stime
 
-        return ru_utime, ru_stime
+        return res
 
     def add_database_transaction(self, duration_sec):
-        self.db_txn_count += 1
-        self.db_txn_duration_sec += duration_sec
+        self._resource_usage.db_txn_count += 1
+        self._resource_usage.db_txn_duration_sec += duration_sec
 
     def add_database_scheduled(self, sched_sec):
         """Record a use of the database pool
@@ -241,7 +345,15 @@ class LoggingContext(object):
             sched_sec (float): number of seconds it took us to get a
                 connection
         """
-        self.db_sched_duration_sec += sched_sec
+        self._resource_usage.db_sched_duration_sec += sched_sec
+
+    def record_event_fetch(self, event_count):
+        """Record a number of events being fetched from the db
+
+        Args:
+            event_count (int): number of events being fetched
+        """
+        self._resource_usage.evt_db_fetch_count += event_count
 
 
 class LoggingContextFilter(logging.Filter):
diff --git a/synapse/util/logformatter.py b/synapse/util/logformatter.py
index 3e42868ea9..a46bc47ce3 100644
--- a/synapse/util/logformatter.py
+++ b/synapse/util/logformatter.py
@@ -14,10 +14,11 @@
 # limitations under the License.
 
 
-from six import StringIO
 import logging
 import traceback
 
+from six import StringIO
+
 
 class LogFormatter(logging.Formatter):
     """Log formatter which gives more detail for exceptions
diff --git a/synapse/util/logutils.py b/synapse/util/logutils.py
index 03249c5dc8..62a00189cc 100644
--- a/synapse/util/logutils.py
+++ b/synapse/util/logutils.py
@@ -14,13 +14,11 @@
 # limitations under the License.
 
 
-from inspect import getcallargs
-from functools import wraps
-
-import logging
 import inspect
+import logging
 import time
-
+from functools import wraps
+from inspect import getcallargs
 
 _TIME_FUNC_ID = 0
 
diff --git a/synapse/util/manhole.py b/synapse/util/manhole.py
index 97e0f00b67..14be3c7396 100644
--- a/synapse/util/manhole.py
+++ b/synapse/util/manhole.py
@@ -12,11 +12,11 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 
-from twisted.conch.manhole import ColoredManhole
-from twisted.conch.insults import insults
 from twisted.conch import manhole_ssh
-from twisted.cred import checkers, portal
+from twisted.conch.insults import insults
+from twisted.conch.manhole import ColoredManhole
 from twisted.conch.ssh.keys import Key
+from twisted.cred import checkers, portal
 
 PUBLIC_KEY = (
     "ssh-rsa AAAAB3NzaC1yc2EAAAABIwAAAGEArzJx8OYOnJmzf4tfBEvLi8DVPrJ3/c9k2I/Az"
diff --git a/synapse/util/metrics.py b/synapse/util/metrics.py
index 1ba7d65c7c..6ba7107896 100644
--- a/synapse/util/metrics.py
+++ b/synapse/util/metrics.py
@@ -13,14 +13,14 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 
-from twisted.internet import defer
+import logging
+from functools import wraps
 
 from prometheus_client import Counter
-from synapse.util.logcontext import LoggingContext
 
-from functools import wraps
-import logging
+from twisted.internet import defer
 
+from synapse.util.logcontext import LoggingContext
 
 logger = logging.getLogger(__name__)
 
@@ -60,10 +60,9 @@ def measure_func(name):
 
 class Measure(object):
     __slots__ = [
-        "clock", "name", "start_context", "start", "new_context", "ru_utime",
-        "ru_stime",
-        "db_txn_count", "db_txn_duration_sec", "db_sched_duration_sec",
+        "clock", "name", "start_context", "start",
         "created_context",
+        "start_usage",
     ]
 
     def __init__(self, clock, name):
@@ -81,10 +80,7 @@ class Measure(object):
             self.start_context.__enter__()
             self.created_context = True
 
-        self.ru_utime, self.ru_stime = self.start_context.get_resource_usage()
-        self.db_txn_count = self.start_context.db_txn_count
-        self.db_txn_duration_sec = self.start_context.db_txn_duration_sec
-        self.db_sched_duration_sec = self.start_context.db_sched_duration_sec
+        self.start_usage = self.start_context.get_resource_usage()
 
     def __exit__(self, exc_type, exc_val, exc_tb):
         if isinstance(exc_type, Exception) or not self.start_context:
@@ -108,15 +104,12 @@ class Measure(object):
             logger.warn("Expected context. (%r)", self.name)
             return
 
-        ru_utime, ru_stime = context.get_resource_usage()
-
-        block_ru_utime.labels(self.name).inc(ru_utime - self.ru_utime)
-        block_ru_stime.labels(self.name).inc(ru_stime - self.ru_stime)
-        block_db_txn_count.labels(self.name).inc(context.db_txn_count - self.db_txn_count)
-        block_db_txn_duration.labels(self.name).inc(
-            context.db_txn_duration_sec - self.db_txn_duration_sec)
-        block_db_sched_duration.labels(self.name).inc(
-            context.db_sched_duration_sec - self.db_sched_duration_sec)
+        usage = context.get_resource_usage() - self.start_usage
+        block_ru_utime.labels(self.name).inc(usage.ru_utime)
+        block_ru_stime.labels(self.name).inc(usage.ru_stime)
+        block_db_txn_count.labels(self.name).inc(usage.db_txn_count)
+        block_db_txn_duration.labels(self.name).inc(usage.db_txn_duration_sec)
+        block_db_sched_duration.labels(self.name).inc(usage.db_sched_duration_sec)
 
         if self.created_context:
             self.start_context.__exit__(exc_type, exc_val, exc_tb)
diff --git a/synapse/util/msisdn.py b/synapse/util/msisdn.py
index 607161e7f0..a6c30e5265 100644
--- a/synapse/util/msisdn.py
+++ b/synapse/util/msisdn.py
@@ -14,6 +14,7 @@
 # limitations under the License.
 
 import phonenumbers
+
 from synapse.api.errors import SynapseError
 
 
diff --git a/synapse/util/ratelimitutils.py b/synapse/util/ratelimitutils.py
index 0ab63c3d7d..5ac33b2132 100644
--- a/synapse/util/ratelimitutils.py
+++ b/synapse/util/ratelimitutils.py
@@ -13,21 +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.async import sleep
 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__)
 
 
@@ -153,7 +151,7 @@ class _PerHostRatelimiter(object):
                 "Ratelimit [%s]: sleeping req",
                 id(request_id),
             )
-            ret_defer = run_in_background(sleep, self.sleep_msec / 1000.0)
+            ret_defer = run_in_background(self.clock.sleep, self.sleep_msec / 1000.0)
 
             self.sleeping_requests.add(request_id)
 
diff --git a/synapse/util/retryutils.py b/synapse/util/retryutils.py
index 4e93f69d3a..8a3a06fd74 100644
--- a/synapse/util/retryutils.py
+++ b/synapse/util/retryutils.py
@@ -12,14 +12,13 @@
 # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
 # See the License for the specific language governing permissions and
 # limitations under the License.
-import synapse.util.logcontext
-from twisted.internet import defer
-
-from synapse.api.errors import CodeMessageException
-
 import logging
 import random
 
+from twisted.internet import defer
+
+import synapse.util.logcontext
+from synapse.api.errors import CodeMessageException
 
 logger = logging.getLogger(__name__)
 
diff --git a/synapse/util/rlimit.py b/synapse/util/rlimit.py
index f4a9abf83f..6c0f2bb0cf 100644
--- a/synapse/util/rlimit.py
+++ b/synapse/util/rlimit.py
@@ -13,9 +13,8 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 
-import resource
 import logging
-
+import resource
 
 logger = logging.getLogger("synapse.app.homeserver")
 
diff --git a/synapse/util/stringutils.py b/synapse/util/stringutils.py
index b98b9dc6e4..43d9db67ec 100644
--- a/synapse/util/stringutils.py
+++ b/synapse/util/stringutils.py
@@ -15,6 +15,7 @@
 
 import random
 import string
+
 from six.moves import range
 
 _string_with_symbols = (
diff --git a/synapse/util/versionstring.py b/synapse/util/versionstring.py
index 52086df465..1fbcd41115 100644
--- a/synapse/util/versionstring.py
+++ b/synapse/util/versionstring.py
@@ -14,9 +14,9 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 
-import subprocess
-import os
 import logging
+import os
+import subprocess
 
 logger = logging.getLogger(__name__)