Fix overzealous cache invalidation

Fixes an issue where a cache invalidation would invalidate *all* pending
entries, rather than just the entry that we intended to invalidate.

1 files changed, 38 insertions, 26 deletions

diff --git a/synapse/util/caches/descriptors.py b/synapse/util/caches/descriptors.py
index bf3a66eae4..68285a7594 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.
@@ -39,12 +40,11 @@ _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 +62,6 @@ class Cache(object):
         "max_entries",
         "name",
         "keylen",
-        "sequence",
         "thread",
         "metrics",
         "_pending_deferred_cache",
@@ -80,7 +79,6 @@ class Cache(object):
 
         self.name = name
         self.keylen = keylen
-        self.sequence = 0
         self.thread = None
         self.metrics = register_cache(name, self.cache)
 
@@ -113,11 +111,10 @@ 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:
@@ -137,12 +134,9 @@ 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()
@@ -150,13 +144,25 @@ 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
 
@@ -168,25 +174,29 @@ 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):
@@ -194,8 +204,10 @@ class Cache(object):
 
     def invalidate_all(self):
         self.check_thread()
-        self.sequence += 1
         self.cache.clear()
+        for entry in self._pending_deferred_cache.itervalues():
+            entry.invalidate()
+        self._pending_deferred_cache.clear()
 
 
 class _CacheDescriptorBase(object):