Bug Fixed: Failures in tpc_finish of client-storages weren't handled

  correctly, leaving the client storage in an inconsistent state.

src/ZEO/ClientStorage.py

@@ -1075,20 +1075,28 @@ class ClientStorage(object):
             # tpc_cond condition variable prevents more than one
             # thread from calling tpc_finish() at a time.
             tid = self._server.tpc_finish(id(txn))
-            self._lock.acquire()  # for atomic processing of invalidations
+
             try:
-                self._update_cache(tid)
-                if f is not None:
-                    f(tid)
-            finally:
-                self._lock.release()
+                self._lock.acquire()  # for atomic processing of invalidations
+                try:
+                    self._update_cache(tid)
+                    if f is not None:
+                        f(tid)
+                finally:
+                    self._lock.release()
+
+                r = self._check_serials()
+                assert r is None or len(r) == 0, "unhandled serialnos: %s" % r
+            except:
+                # The server successfully committed.  If we get a failure
+                # here, our own state will be in question, so reconnect.
+                self._connection.close()
+                raise
 
-            r = self._check_serials()
-            assert r is None or len(r) == 0, "unhandled serialnos: %s" % r
+            self.end_transaction()
         finally:
             self._load_lock.release()
             self._iterator_gc()
-            self.end_transaction()
 
     def _update_cache(self, tid):
         """Internal helper to handle objects modified by a transaction.

src/ZEO/tests/ConnectionTests.py

@@ -1075,6 +1075,8 @@ class TimeoutTests(CommonSetupTearDown):
         self.assert_(storage.is_connected())
         # We expect finish to fail.
         self.assertRaises(ClientDisconnected, storage.tpc_finish, t)
+        storage.tpc_abort(t)
+
         # Now we think we've committed the second transaction, but we really
         # haven't.  A third one should produce a POSKeyError on the server,
         # which manifests as a ConflictError on the client.

src/ZEO/tests/testZEO.py

@@ -1006,6 +1006,104 @@ transaction, we'll get a result:
 
     """
 
+def tpc_finish_error():
+    r"""Server errors in tpc_finish weren't handled properly.
+
+    >>> import ZEO.ClientStorage
+
+    >>> class Connection:
+    ...     def __init__(self, client):
+    ...         self.client = client
+    ...     def get_addr(self):
+    ...         return 'server'
+    ...     def is_async(self):
+    ...         return True
+    ...     def register_object(self, ob):
+    ...         pass
+    ...     def close(self):
+    ...         print 'connection closed'
+
+    >>> class ConnectionManager:
+    ...     def __init__(self, addr, client, tmin, tmax):
+    ...         self.client = client
+    ...     def connect(self, sync=1):
+    ...         self.client.notifyConnected(Connection(self.client))
+
+    >>> class StorageServer:
+    ...     should_fail = True
+    ...     def __init__(self, conn):
+    ...         self.conn = conn
+    ...         self.t = None
+    ...     def get_info(self):
+    ...         return {}
+    ...     def endZeoVerify(self):
+    ...         self.conn.client.endVerify()
+    ...     def lastTransaction(self):
+    ...         return '\0'*8
+    ...     def tpc_begin(self, t, *args):
+    ...         if self.t is not None:
+    ...             raise TypeError('already trans')
+    ...         self.t = t
+    ...         print 'begin', args
+    ...     def vote(self, t):
+    ...         if self.t != t:
+    ...             raise TypeError('bad trans')
+    ...         print 'vote'
+    ...     def tpc_finish(self, *args):
+    ...         if self.should_fail:
+    ...             raise TypeError()
+    ...         print 'finish'
+    ...     def tpc_abort(self, t):
+    ...         if self.t != t:
+    ...             raise TypeError('bad trans')
+    ...         self.t = None
+    ...         print 'abort'
+    ...     def iterator_gc(*args):
+    ...         pass
+
+    >>> class ClientStorage(ZEO.ClientStorage.ClientStorage):
+    ...     ConnectionManagerClass = ConnectionManager
+    ...     StorageServerStubClass = StorageServer
+
+    >>> class Transaction:
+    ...     user = 'test'
+    ...     description = ''
+    ...     _extension = {}
+
+    >>> cs = ClientStorage(('', ''))
+    >>> t1 = Transaction()
+    >>> cs.tpc_begin(t1)
+    begin ('test', '', {}, None, ' ')
+
+    >>> cs.tpc_vote(t1)
+    vote
+
+    >>> cs.tpc_finish(t1)
+    Traceback (most recent call last):
+    ...
+    TypeError
+
+    >>> cs.tpc_abort(t1)
+    abort
+
+    >>> t2 = Transaction()
+    >>> cs.tpc_begin(t2)
+    begin ('test', '', {}, None, ' ')
+    >>> cs.tpc_vote(t2)
+    vote
+
+    If client storage has an internal error after the storage finish
+    succeeeds, it will close the connection, which will force a
+    restart and reverification.
+
+    >>> StorageServer.should_fail = False
+    >>> cs._update_cache = lambda : None
+    >>> try: cs.tpc_finish(t2)
+    ... except: pass
+    ... else: print "Should have failed"
+    finish
+    connection closed
+    """
 
 test_classes = [FileStorageTests, FileStorageRecoveryTests,
                 MappingStorageTests, DemoStorageTests,