Modules/gc_weakref.txt

Intro
=====

The basic rule for dealing with weakref callbacks (and __del__ methods too,
for that matter) during cyclic gc:

    Once gc has computed the set of unreachable objects, no Python-level
    code can be allowed to access an unreachable object.

If that can happen, then the Python code can resurrect unreachable objects
too, and gc can't detect that without starting over.  Since gc eventually
runs tp_clear on all unreachable objects, if an unreachable object is
resurrected then tp_clear will eventually be called on it (or may already
have been called before resurrection).  At best (and this has been an
historically common bug), tp_clear empties an instance's __dict__, and
"impossible" AttributeErrors result.  At worst, tp_clear leaves behind an
insane object at the C level, and segfaults result (historically, most
often by setting a new-style class's mro pointer to NULL, after which
attribute lookups performed by the class can segfault).

OTOH, it's OK to run Python-level code that can't access unreachable
objects, and sometimes that's necessary.  The chief example is the callback
attached to a reachable weakref W to an unreachable object O.  Since O is
going away, and W is still alive, the callback must be invoked.  Because W
is still alive, everything reachable from its callback is also reachable,
so it's also safe to invoke the callback (although that's trickier than it
sounds, since other reachable weakrefs to other unreachable objects may
still exist, and be accessible to the callback -- there are lots of painful
details like this covered in the rest of this file).

Python 2.4/2.3.5
================

The "Before 2.3.3" section below turned out to be wrong in some ways, but
I'm leaving it as-is because it's more right than wrong, and serves as a
wonderful example of how painful analysis can miss not only the forest for
the trees, but also miss the trees for the aphids sucking the trees
dry <wink>.

The primary thing it missed is that when a weakref to a piece of cyclic
trash (CT) exists, then any call to any Python code whatsoever can end up
materializing a strong reference to that weakref's CT referent, and so
possibly resurrect an insane object (one for which cyclic gc has called-- or
will call before it's done --tp_clear()).  It's not even necessarily that a
weakref callback or __del__ method does something nasty on purpose:  as
soon as we execute Python code, threads other than the gc thread can run
too, and they can do ordinary things with weakrefs that end up resurrecting
CT while gc is running.

    http://www.python.org/sf/1055820

shows how innocent it can be, and also how nasty.  Variants of the three
focussed test cases attached to that bug report are now part of Python's
standard Lib/test/test_gc.py.

Jim Fulton gave the best nutshell summary of the new (in 2.4 and 2.3.5)
approach:

    Clearing cyclic trash can call Python code.  If there are weakrefs to
    any of the cyclic trash, then those weakrefs can be used to resurrect
    the objects.  Therefore, *before* clearing cyclic trash, we need to
    remove any weakrefs.  If any of the weakrefs being removed have
    callbacks, then we need to save the callbacks and call them *after* all
    of the weakrefs have been cleared.

Alas, doing just that much doesn't work, because it overlooks what turned
out to be the much subtler problems that were fixed earlier, and described
below.  We do clear all weakrefs to CT now before breaking cycles, but not
all callbacks encountered can be run later.  That's explained in horrid
detail below.

Older text follows, with a some later comments in [] brackets:

Before 2.3.3
============

Before 2.3.3, Python's cyclic gc didn't pay any attention to weakrefs.
Segfaults in Zope3 resulted.

weakrefs in Python are designed to, at worst, let *other* objects learn
that a given object has died, via a callback function.  The weakly
referenced object itself is not passed to the callback, and the presumption
is that the weakly referenced object is unreachable trash at the time the
callback is invoked.

That's usually true, but not always.  Suppose a weakly referenced object
becomes part of a clump of cyclic trash.  When enough cycles are broken by
cyclic gc that the object is reclaimed, the callback is invoked.  If it's
possible for the callback to get at objects in the cycle(s), then it may be
possible for those objects to access (via strong references in the cycle)
the weakly referenced object being torn down, or other objects in the cycle
that have already suffered a tp_clear() call.  There's no guarantee that an
object is in a sane state after tp_clear().  Bad things (including
segfaults) can happen right then, during the callback's execution, or can
happen at any later time if the callback manages to resurrect an insane
object.

[That missed that, in addition, a weakref to CT can exist outside CT, and
 any callback into Python can use such a non-CT weakref to resurrect its CT
 referent.  The same bad kinds of things can happen then.]

Note that if it's possible for the callback to get at objects in the trash
cycles, it must also be the case that the callback itself is part of the
trash cycles.  Else the callback would have acted as an external root to
the current collection, and nothing reachable from it would be in cyclic
trash either.

[Except that a non-CT callback can also use a non-CT weakref to get at
 CT objects.]

More, if the callback itself is in cyclic trash, then the weakref to which
the callback is attached must also be trash, and for the same kind of
reason:  if the weakref acted as an external root, then the callback could
not have been cyclic trash.

So a problem here requires that a weakref, that weakref's callback, and the
weakly referenced object, all be in cyclic trash at the same time.  This
isn't easy to stumble into by accident while Python is running, and, indeed,
it took quite a while to dream up failing test cases.  Zope3 saw segfaults
during shutdown, during the second call of gc in Py_Finalize, after most
modules had been torn down.  That creates many trash cycles (esp. those
involving new-style classes), making the problem much more likely.  Once you
know what's required to provoke the problem, though, it's easy to create
tests that segfault before shutdown.

In 2.3.3, before breaking cycles, we first clear all the weakrefs with
callbacks in cyclic trash.  Since the weakrefs *are* trash, and there's no
defined-- or even predictable --order in which tp_clear() gets called on
cyclic trash, it's defensible to first clear weakrefs with callbacks.  It's
a feature of Python's weakrefs too that when a weakref goes away, the
callback (if any) associated with it is thrown away too, unexecuted.

[In 2.4/2.3.5, we first clear all weakrefs to CT objects, whether or not
 those weakrefs are themselves CT, and whether or not they have callbacks.
 The callbacks (if any) on non-CT weakrefs (if any) are invoked later,
 after all weakrefs-to-CT have been cleared.  The callbacks (if any) on CT
 weakrefs (if any) are never invoked, for the excruciating reasons
 explained here.]

Just that much is almost enough to prevent problems, by throwing away
*almost* all the weakref callbacks that could get triggered by gc.  The
problem remaining is that clearing a weakref with a callback decrefs the
callback object, and the callback object may *itself* be weakly referenced,
via another weakref with another callback.  So the process of clearing
weakrefs can trigger callbacks attached to other weakrefs, and those
latter weakrefs may or may not be part of cyclic trash.

So, to prevent any Python code from running while gc is invoking tp_clear()
on all the objects in cyclic trash,

[That was always wrong:  we can't stop Python code from running when gc
 is breaking cycles.  If an object with a __del__ method is not itself in
 a cycle, but is reachable only from CT, then breaking cycles will, as a
 matter of course, drop the refcount on that object to 0, and its __del__
 will run right then.  What we can and must stop is running any Python
 code that could access CT.]
                                     it's not quite enough just to invoke
tp_clear() on weakrefs with callbacks first.  Instead the weakref module
grew a new private function (_PyWeakref_ClearRef) that does only part of
tp_clear():  it removes the weakref from the weakly-referenced object's list
of weakrefs, but does not decref the callback object.  So calling
_PyWeakref_ClearRef(wr) ensures that wr's callback object will never
trigger, and (unlike weakref's tp_clear()) also prevents any callback
associated *with* wr's callback object from triggering.

[Although we may trigger such callbacks later, as explained below.]

Then we can call tp_clear on all the cyclic objects and never trigger
Python code.

[As above, not so:  it means never trigger Python code that can access CT.]

After we do that, the callback objects still need to be decref'ed.  Callbacks
(if any) *on* the callback objects that were also part of cyclic trash won't
get invoked, because we cleared all trash weakrefs with callbacks at the
start.  Callbacks on the callback objects that were not part of cyclic trash
acted as external roots to everything reachable from them, so nothing
reachable from them was part of cyclic trash, so gc didn't do any damage to
objects reachable from them, and it's safe to call them at the end of gc.

[That's so.  In addition, now we also invoke (if any) the callbacks on
 non-CT weakrefs to CT objects, during the same pass that decrefs the
 callback objects.]

An alternative would have been to treat objects with callbacks like objects
with __del__ methods, refusing to collect them, appending them to gc.garbage
instead.  That would have been much easier.  Jim Fulton gave a strong
argument against that (on Python-Dev):

    There's a big difference between __del__ and weakref callbacks.
    The __del__ method is "internal" to a design.  When you design a
    class with a del method, you know you have to avoid including the
    class in cycles.

    Now, suppose you have a design that makes has no __del__ methods but
    that does use cyclic data structures.  You reason about the design,
    run tests, and convince yourself you don't have a leak.

    Now, suppose some external code creates a weakref to one of your
    objects.  All of a sudden, you start leaking.  You can look at your
    code all you want and you won't find a reason for the leak.

IOW, a class designer can out-think __del__ problems, but has no control
over who creates weakrefs to his classes or class instances.  The class
user has little chance either of predicting when the weakrefs he creates
may end up in cycles.

Callbacks on weakref callbacks are executed in an arbitrary order, and
that's not good (a primary reason not to collect cycles with objects with
__del__ methods is to avoid running finalizers in an arbitrary order).
However, a weakref callback on a weakref callback has got to be rare.
It's possible to do such a thing, so gc has to be robust against it, but
I doubt anyone has done it outside the test case I wrote for it.

[The callbacks (if any) on non-CT weakrefs to CT objects are also executed
 in an arbitrary order now.  But they were before too, depending on the
 vagaries of when tp_clear() happened to break enough cycles to trigger
 them.  People simply shouldn't try to use __del__ or weakref callbacks to
 do fancy stuff.]