python-peps/pep-0442.txt

PEP: 442
Title: Safe object finalization
Version: $Revision$
Last-Modified: $Date$
Author: Antoine Pitrou <solipsis@pitrou.net>
Status: Draft
Type: Standards Track
Content-Type: text/x-rst
Created: 2013-05-18
Python-Version: 3.4
Post-History: 2013-05-18
Resolution: TBD


Abstract
========

This PEP proposes to deal with the current limitations of object
finalization.  The goal is to be able to define and run finalizers
for any object, regardless of their position in the object graph.

This PEP doesn't call for any change in Python code.  Objects
with existing finalizers will benefit automatically.


Definitions
===========

Reference
    A directional link from an object to another.  The target of the
    reference is kept alive by the reference, as long as the source is
    itself alive and the reference isn't cleared.

Weak reference
    A directional link from an object to another, which doesn't keep
    alive its target.  This PEP focusses on non-weak references.

Reference cycle
    A cyclic subgraph of directional links between objects, which keeps
    those objects from being collected in a pure reference-counting
    scheme.

Cyclic isolate (CI)
    A standalone subgraph of objects in which no object is referenced
    from the outside, containing one or several reference cycles, *and*
    whose objects are still in a usable, non-broken state: they can
    access each other from their respective finalizers.

Cyclic garbage collector (GC)
    A device able to detect cyclic isolates and turn them into cyclic
    trash.  Objects in cyclic trash are eventually disposed of by
    the natural effect of the references being cleared and their
    reference counts dropping to zero.

Cyclic trash (CT)
    A former cyclic isolate whose objects have started being cleared
    by the GC.  Objects in cyclic trash are potential zombies; if they
    are accessed by Python code, the symptoms can vary from weird
    AttributeErrors to crashes.

Zombie / broken object
    An object part of cyclic trash.  The term stresses that the object
    is not safe: its outgoing references may have been cleared, or one
    of the objects it references may be zombie.  Therefore,
    it should not be accessed by arbitrary code (such as finalizers).

Finalizer
    A function or method called when an object is intended to be
    disposed of.  The finalizer can access the object and release any
    resource held by the object (for example mutexes or file
    descriptors).  An example is a ``__del__`` method.

Resurrection
    The process by which a finalizer creates a new reference to an
    object in a CI.  This can happen as a quirky but supported
    side-effect of ``__del__`` methods.


Impact
======

While this PEP discusses CPython-specific implementation details, the
change in finalization semantics is expected to affect the Python
ecosystem as a whole.  In particular, this PEP obsoletes the current
guideline that "objects with a ``__del__`` method should not be part of a
reference cycle".


Benefits
========

The primary benefits of this PEP regard objects with finalizers, such
as objects with a ``__del__`` method and generators with a ``finally``
block.  Those objects can now be reclaimed when they are part of a
reference cycle.

The PEP also paves the way for further benefits:

* The module shutdown procedure may not need to set global variables to
  None anymore.  This could solve a well-known class of irritating issues.

The PEP doesn't change the semantics of:

* Weak references caught in reference cycles.

* C extension types with a custom ``tp_dealloc`` function.


Description
===========

Reference-counted disposal
--------------------------

In normal reference-counted disposal, an object's finalizer is called
just before the object is deallocated.  If the finalizer resurrects
the object, deallocation is aborted.

*However*, if the object was already finalized, then the finalizer isn't
called.  This prevents us from finalizing zombies (see below).

Disposal of cyclic isolates
---------------------------

Cyclic isolates are first detected by the garbage collector, and then
disposed of.  The detection phase doesn't change and won't be described
here.  Disposal of a CI traditionally works in the following order:

1. Weakrefs to CI objects are cleared, and their callbacks called. At
   this point, the objects are still safe to use.

2. The CI becomes a CT as the GC systematically breaks all
   known references inside it (using the ``tp_clear`` function).

3. Nothing.  All CT objects should have been disposed of in step 2
   (as a side-effect of clearing references); this collection is
   finished.

This PEP proposes to turn CI disposal into the following sequence (new
steps are in bold):

1. Weakrefs to CI objects are cleared, and their callbacks called. At
   this point, the objects are still safe to use.

2. **The finalizers of all CI objects are called.**

3. **The CI is traversed again to determine if it is still isolated.
   If it is determined that at least one object in CI is now reachable
   from outside the CI, this collection is aborted and the whole CI
   is resurrected.  Otherwise, proceed.**

4. The CI becomes a CT as the GC systematically breaks all
   known references inside it (using the ``tp_clear`` function).

5. Nothing.  All CT objects should have been disposed of in step 4
   (as a side-effect of clearing references); this collection is
   finished.

.. note::
    The GC doesn't recalculate the CI after step 2 above, hence the need
    for step 3 to check that the whole subgraph is still isolated.


C-level changes
===============

Type objects get a new ``tp_finalize`` slot to which ``__del__`` methods
are mapped (and reciprocally).  Generators are modified to use this slot,
rather than ``tp_del``.  A ``tp_finalize`` function is a normal C
function which will be called with a valid and alive ``PyObject`` as its
only argument.  It doesn't need to manipulate the object's reference count,
as this will be done by the caller.  However, it must ensure that the
original exception state is restored before returning to the caller.

For compatibility, ``tp_del`` is kept in the type structure.  Handling
of objects with a non-NULL ``tp_del`` is unchanged: when part of a CI,
they are not finalized and end up in ``gc.garbage``.  However, a non-NULL
``tp_del`` is not encountered anymore in the CPython source tree (except
for testing purposes).

Two new C API functions are provided to ease calling of ``tp_finalize``,
especially from custom deallocators.

On the internal side, a bit is reserved in the GC header for GC-managed
objects to signal that they were finalized.  This helps avoid finalizing
an object twice (and, especially, finalizing a CT object after it was
broken by the GC).

.. note::
    Objects which are not GC-enabled can also have a ``tp_finalize`` slot.
    They don't need the additional bit since their ``tp_finalize`` function
    can only be called from the deallocator: it therefore cannot be called
    twice, except when resurrected.


Discussion
==========

Predictability
--------------

Following this scheme, an object's finalizer is always called exactly
once.  The only exception is if an object is resurrected: the finalizer
will be called again when the object becomes unreachable again.

For CI objects, the order in which finalizers are called (step 2 above)
is undefined.

Safety
------

It is important to explain why the proposed change is safe.  There
are two aspects to be discussed:

* Can a finalizer access zombie objects (including the object being
  finalized)?

* What happens if a finalizer mutates the object graph so as to impact
  the CI?

Let's discuss the first issue.  We will divide possible cases in two
categories:

* If the object being finalized is part of the CI: by construction, no
  objects in CI are zombies yet, since CI finalizers are called before
  any reference breaking is done.  Therefore, the finalizer cannot
  access zombie objects, which don't exist.

* If the object being finalized is not part of the CI/CT: by definition,
  objects in the CI/CT don't have any references pointing to them from
  outside the CI/CT.  Therefore, the finalizer cannot reach any zombie
  object (that is, even if the object being finalized was itself
  referenced from a zombie object).

Now for the second issue.  There are three potential cases:

* The finalizer clears an existing reference to a CI object.  The CI
  object may be disposed of before the GC tries to break it, which
  is fine (the GC simply has to be aware of this possibility).

* The finalizer creates a new reference to a CI object.  This can only
  happen from a CI object's finalizer (see above why).  Therefore, the
  new reference will be detected by the GC after all CI finalizers are
  called (step 3 above), and collection will be aborted without any
  objects being broken.

* The finalizer clears or creates a reference to a non-CI object.  By
  construction, this is not a problem.


Implementation
==============

An implementation is available in branch ``finalize`` of the repository
at http://hg.python.org/features/finalize/.


Validation
==========

Besides running the normal Python test suite, the implementation adds
test cases for various finalization possibilities including reference cycles,
object resurrection and legacy ``tp_del`` slots.

The implementation has also been checked to not produce any regressions on
the following test suites:

* `Tulip <http://code.google.com/p/tulip/>`_, which makes an extensive
  use of generators

* `Tornado <http://www.tornadoweb.org>`_

* `SQLAlchemy <http://www.sqlalchemy.org/>`_

* `Django <https://www.djangoproject.com/>`_

* `zope.interface <http://pypi.python.org/pypi/zope.interface>`_


References
==========

Notes about reference cycle collection and weak reference callbacks:
http://hg.python.org/cpython/file/4e687d53b645/Modules/gc_weakref.txt

Generator memory leak: http://bugs.python.org/issue17468

Allow objects to decide if they can be collected by GC:
http://bugs.python.org/issue9141

Module shutdown procedure based on GC
http://bugs.python.org/issue812369

Copyright
=========

This document has been placed in the public domain.

..
   Local Variables:
   mode: indented-text
   indent-tabs-mode: nil
   sentence-end-double-space: t
   fill-column: 70
   coding: utf-8
   End: