[Twisted-Python] cReactor fixes

Brian Warner warner at lothar.com
Sun Jan 5 17:34:10 EST 2003

Hi all..

For some reason not yet entirely clear to me, I started to put some time
into bringing cReactor up to date. My hope was to make it functional enough
to pass the full test suite. At the moment, it fails 120 of the 413 tests in
the suite (that doesn't include test_threads, which I had to disable to
prevent the test suite from hanging forever).

In tracking down the reasons for these failures, I've come across a number
of places where the reactor interface (as defined by IReactorCore and
friends in twisted.internet.interfaces) was ambiguous about something, and
the default reactor has implemented different functionality than the
cReactor. Several tests have come to rely upon default.py's implementation,
and fail when run under cReactor even though it appears to implement the
same interface.

To resolve these, we need to nail down the ambiguities in the Interface, so
that all reactors can implement the same thing. The two issues I've
encountered so far:

 IReactorCore.addSystemEventTrigger()   [1 test failure]

  The interface mentions three (startup, shutdown, persist) as examples of
  "system events", and says they will be fired internally by the Reactor. It
  also says:

        An implementor of this interface must only implement those events
        described here.

  The cReactor implements this literally, and refuses to accept calls to
  addSystemEventTrigger or fireSystemEvent with event names outside of
  ['startup', 'shutdown', 'persist']. The default reactor (and the ones that
  inherit system event code from it, which is all of them except cReactor)
  use a dict for the events, and therefore accept any string as a system
  event type. test_internet.InterfaceTestCase.testTriggerSystemEvent creates
  several new event types ("test", "defer", "remove") and runs tests with
  them. This test thus fails under cReactor.

  RESOLUTION: I think we decided that cReactor should accept arbitrary event
  names, and I'm working on code to implement that. That restrictive
  sentence should be removed from the docs. I'd also like to see a better
  explanation of what event types will be fired by whom: if the API allows
  an arbitrary string, but the Reactor will only fire "startup" and
  "shutdown" for you, then you have to know that it is your own
  responsibility to fire the new triggers that you've added.

 IReactorCore.stop() [38 test failures]

  The interface description for .stop and .crash make the claim that once
  the Reactor is shut down with .stop(), it may not be restarted with
  .run(). Reactors which have been stopped with .crash() *may* be restarted.

  cReactor implements this literally. Once reactor.stop() has been called,
  the reactor is moved into a state from which .run() and .iterate() raise a
  RuntimeError "the reactor is shut down!" exception. The default reactor
  does not enforce this: .run and .stop may be called as many times as you

  All tests in a given invocation of bin/trial or admin/runtests are
  executed with the same Reactor instance. Indeed everything done by a
  single python process is performed with the same Reactor instance. Normal
  apps don't care: they start the reactor in main() or under twistd, and it
  runs until the app shuts down. The tests, however, start and stop the
  reactor all the time, generally once per test case.

  Under cReactor, the first test to do reactor.stop() will terminate the
  reactor permanently, causing failures of all later tests that attempt
  reactor.run() or reactor.iterate(). This also tends to kill tests which
  re-bind to the same port multiple times, as the reactor.iterate() needed
  to release the port fails, leaving the port bound, causing a "Address
  already in use" error on the second test.

  POTENTIAL RESOLUTION: We need to decide whether the default reactor should
  obey the restriction defined by IReactorCore, or whether the interface
  (and cReactor) should be changed to match the behavior of the default

  Another possibility which might be easier to implement would be to enforce
  the no-restart restriction, but also create IReactorCore.reset(), and
  declare that it will return the reactor to the state it had at boot time,
  allowing .run() to be called on it once again. This method would also
  remove all DelayedCall events, remove any TCP/UDP sockets or listening
  ports, delete any systemEventTriggers that were added, and generally get
  rid of any other state which might have accumulated since startup. We
  could have the test suite use this to reset the reactor just after each

The other test failures are the result of functionality that is missing from
cReactor. In general, cReactor has not been maintained as new features were
added to the other reactors. Here is a summary of all the failures. (note
that some tests fail for multiple reasons).

 IReactorTCP.connectTCP [8 failures]
  client support (outbound sockets) is missing

 IListeningPort.getHost [22 failures]
  simple missing function, mostly used to determine the port number of a
  system-selected listening port

 IReactorProcess [9 failures]
  launching child processes is unimplemented 

 rebuild attempts: missing __dict__ [5 failures]
  twisted.python.rebuild attempts to rebuild a C module with no .__dict__

 IReactorMulticast [5 failures]
  multicast support is not implemented

 IReactorUDP [23 failures]
  UDP support is unimplemented

 IReactorCore.addSystemEventTrigger()   [1 test failure]

 IReactorCore.stop() [38 test failures]

Some of these aren't hard to fix: IListeningPort.getHost should only take a
few lines, connectTCP is pretty straightforward, and IReactorUDP should be a
fairly easy cut-and-paste job from the TCP code. But others are tricky:
IReactorProcess is fairly high-level, and IReactorMulticast might be relying
upon some portability code in the python socket module that wouldn't be
available at the C layer.

I think it is now pretty clear what work needs to be done. Everyone is
welcome to take a crack at it :). I will probably finish the getDelayedCall
and addSystemEventTrigger changes I'm working on, commit them, and then let
others pursue the pieces that are useful to them. (the default reactor is
currently good enough for my needs; I only put this much time into cReactor
because I wanted to understand the test failures well enough to know that my
changes weren't the cause).


More information about the Twisted-Python mailing list