summary refs log tree commit diff
path: root/synapse/util/caches/descriptors.py
diff options
context:
space:
mode:
Diffstat (limited to 'synapse/util/caches/descriptors.py')
-rw-r--r--synapse/util/caches/descriptors.py102
1 files changed, 60 insertions, 42 deletions
diff --git a/synapse/util/caches/descriptors.py b/synapse/util/caches/descriptors.py
index af65bfe7b8..f8a07df6b8 100644
--- a/synapse/util/caches/descriptors.py
+++ b/synapse/util/caches/descriptors.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.
@@ -12,25 +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__)
 
 
@@ -39,12 +41,11 @@ _CacheSentinel = object()
 
 class CacheEntry(object):
     __slots__ = [
-        "deferred", "sequence", "callbacks", "invalidated"
+        "deferred", "callbacks", "invalidated"
     ]
 
-    def __init__(self, deferred, sequence, callbacks):
+    def __init__(self, deferred, callbacks):
         self.deferred = deferred
-        self.sequence = sequence
         self.callbacks = set(callbacks)
         self.invalidated = False
 
@@ -62,7 +63,6 @@ class Cache(object):
         "max_entries",
         "name",
         "keylen",
-        "sequence",
         "thread",
         "metrics",
         "_pending_deferred_cache",
@@ -75,13 +75,16 @@ class Cache(object):
         self.cache = LruCache(
             max_size=max_entries, keylen=keylen, cache_type=cache_type,
             size_callback=(lambda d: len(d)) if iterable else None,
+            evicted_callback=self._on_evicted,
         )
 
         self.name = name
         self.keylen = keylen
-        self.sequence = 0
         self.thread = None
-        self.metrics = register_cache(name, self.cache)
+        self.metrics = register_cache("cache", name, self.cache)
+
+    def _on_evicted(self, evicted_count):
+        self.metrics.inc_evictions(evicted_count)
 
     def check_thread(self):
         expected_thread = self.thread
@@ -109,11 +112,10 @@ class Cache(object):
         callbacks = [callback] if callback else []
         val = self._pending_deferred_cache.get(key, _CacheSentinel)
         if val is not _CacheSentinel:
-            if val.sequence == self.sequence:
-                val.callbacks.update(callbacks)
-                if update_metrics:
-                    self.metrics.inc_hits()
-                return val.deferred
+            val.callbacks.update(callbacks)
+            if update_metrics:
+                self.metrics.inc_hits()
+            return val.deferred
 
         val = self.cache.get(key, _CacheSentinel, callbacks=callbacks)
         if val is not _CacheSentinel:
@@ -133,12 +135,9 @@ class Cache(object):
         self.check_thread()
         entry = CacheEntry(
             deferred=value,
-            sequence=self.sequence,
             callbacks=callbacks,
         )
 
-        entry.callbacks.update(callbacks)
-
         existing_entry = self._pending_deferred_cache.pop(key, None)
         if existing_entry:
             existing_entry.invalidate()
@@ -146,13 +145,25 @@ class Cache(object):
         self._pending_deferred_cache[key] = entry
 
         def shuffle(result):
-            if self.sequence == entry.sequence:
-                existing_entry = self._pending_deferred_cache.pop(key, None)
-                if existing_entry is entry:
-                    self.cache.set(key, result, entry.callbacks)
-                else:
-                    entry.invalidate()
+            existing_entry = self._pending_deferred_cache.pop(key, None)
+            if existing_entry is entry:
+                self.cache.set(key, result, entry.callbacks)
             else:
+                # oops, the _pending_deferred_cache has been updated since
+                # we started our query, so we are out of date.
+                #
+                # Better put back whatever we took out. (We do it this way
+                # round, rather than peeking into the _pending_deferred_cache
+                # and then removing on a match, to make the common case faster)
+                if existing_entry is not None:
+                    self._pending_deferred_cache[key] = existing_entry
+
+                # we're not going to put this entry into the cache, so need
+                # to make sure that the invalidation callbacks are called.
+                # That was probably done when _pending_deferred_cache was
+                # updated, but it's possible that `set` was called without
+                # `invalidate` being previously called, in which case it may
+                # not have been. Either way, let's double-check now.
                 entry.invalidate()
             return result
 
@@ -164,25 +175,29 @@ class Cache(object):
 
     def invalidate(self, key):
         self.check_thread()
+        self.cache.pop(key, None)
 
-        # Increment the sequence number so that any SELECT statements that
-        # raced with the INSERT don't update the cache (SYN-369)
-        self.sequence += 1
+        # if we have a pending lookup for this key, remove it from the
+        # _pending_deferred_cache, which will (a) stop it being returned
+        # for future queries and (b) stop it being persisted as a proper entry
+        # in self.cache.
         entry = self._pending_deferred_cache.pop(key, None)
+
+        # run the invalidation callbacks now, rather than waiting for the
+        # deferred to resolve.
         if entry:
             entry.invalidate()
 
-        self.cache.pop(key, None)
-
     def invalidate_many(self, key):
         self.check_thread()
         if not isinstance(key, tuple):
             raise TypeError(
                 "The cache key must be a tuple not %r" % (type(key),)
             )
-        self.sequence += 1
         self.cache.del_multi(key)
 
+        # if we have a pending lookup for this key, remove it from the
+        # _pending_deferred_cache, as above
         entry_dict = self._pending_deferred_cache.pop(key, None)
         if entry_dict is not None:
             for entry in iterate_tree_cache_entry(entry_dict):
@@ -190,8 +205,10 @@ class Cache(object):
 
     def invalidate_all(self):
         self.check_thread()
-        self.sequence += 1
         self.cache.clear()
+        for entry in itervalues(self._pending_deferred_cache):
+            entry.invalidate()
+        self._pending_deferred_cache.clear()
 
 
 class _CacheDescriptorBase(object):
@@ -294,7 +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
@@ -376,9 +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)
@@ -549,7 +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