python-peps/pep-0556.rst

PEP: 556
Title: Threaded garbage collection
Author: Antoine Pitrou <solipsis@pitrou.net>
Status: Deferred
Type: Standards Track
Content-Type: text/x-rst
Created: 2017-09-08
Python-Version: 3.7
Post-History: 2017-09-08


Deferral Notice
===============

This PEP is currently not being actively worked on.  It may be revived
in the future.  The main missing steps are:

* polish the implementation, adapting the test suite where necessary;

* ensure setting threaded garbage collection does not disrupt existing
  code in unexpected ways (expected impact includes lengthening the
  lifetime of objects in reference cycles).


Abstract
========

This PEP proposes a new optional mode of operation for CPython's cyclic
garbage collector (GC) where implicit (i.e. opportunistic) collections
happen in a dedicated thread rather than synchronously.


Terminology
===========

An "implicit" GC run (or "implicit" collection) is one that is triggered
opportunistically based on a certain heuristic computed over allocation
statistics, whenever a new allocation is requested.  Details of the
heuristic are not relevant to this PEP, as it does not propose to change it.

An "explicit" GC run (or "explicit" collection) is one that is requested
programmatically by an API call such as ``gc.collect``.

"Threaded" refers to the fact that GC runs happen in a dedicated thread
separate from sequential execution of application code.  It does not mean
"concurrent" (the Global Interpreter Lock, or GIL, still serializes
execution among Python threads *including* the dedicated GC thread)
nor "parallel" (the GC is not able to distribute its work onto several
threads at once to lower wall-clock latencies of GC runs).


Rationale
=========

The mode of operation for the GC has always been to perform implicit
collections synchronously.  That is, whenever the aforementioned heuristic
is activated, execution of application code in the current thread is
suspended and the GC is launched in order to reclaim dead reference
cycles.

There is a catch, though.  Over the course of reclaiming dead reference
cycles (and any ancillary objects hanging at those cycles), the GC can
execute arbitrary finalization code in the form of ``__del__`` methods
and ``weakref`` callbacks.  Over the years, Python has been used for more
and more sophisticated purposes, and it is increasingly common for
finalization code to perform complex tasks, for example in distributed
systems where loss of an object may require notifying other (logical
or physical) nodes.

Interrupting application code at arbitrary points to execute finalization
code that may rely on a consistent internal state and/or on acquiring
synchronization primitives gives rise to reentrancy issues that even the
most seasoned experts have trouble fixing properly [#queue-reentrancy-bug]_.

This PEP bases itself on the observation that, despite the apparent
similarities, same-thread reentrancy is a fundamentally harder
problem than multi-thread synchronization.  Instead of letting each
developer or library author struggle with extremely hard reentrancy
issues, one by one, this PEP proposes to allow the GC to run in a
separate thread where well-known multi-thread synchronization practices
are sufficient.


Proposal
========

Under this PEP, the GC has two modes of operation:

* "serial", which is the default and legacy mode, where an implicit GC
  run is performed immediately in the thread that detects such an implicit
  run is desired (based on the aforementioned allocation heuristic).

* "threaded", which can be explicitly enabled at runtime on a per-process
  basis, where implicit GC runs are *scheduled* whenever the allocation
  heuristic is triggered, but run in a dedicated background thread.

Hard reentrancy problems which plague sophisticated uses of finalization
callbacks in the "serial" mode become relatively easy multi-thread
synchronization problems in the "threaded" mode of operation.

The GC also traditionally allows for explicit GC runs, using the Python
API ``gc.collect`` and the C API ``PyGC_Collect``.  The visible semantics
of these two APIs are left unchanged: they perform a GC run immediately
when called, and only return when the GC run is finished.


New public APIs
---------------

Two new Python APIs are added to the ``gc`` module:

* ``gc.set_mode(mode)`` sets the current mode of operation (either "serial"
  or "threaded").  If setting to "serial" and the current mode is
  "threaded", then the function also waits for the GC thread to end.

* ``gc.get_mode()`` returns the current mode of operation.

It is allowed to switch back and forth between modes of operation.


Intended use
------------

Given the per-process nature of the switch and its repercussions on
semantics of all finalization callbacks, it is recommended that it is
set at the beginning of an application's code (and/or in initializers
for child processes e.g. when using ``multiprocessing``).  Library functions
should probably not mess with this setting, just as they shouldn't call
``gc.enable`` or ``gc.disable``, but there's nothing to prevent them from
doing so.


Non-goals
---------

This PEP does not address reentrancy issues with other kinds of
asynchronous code execution (for example signal handlers registered
with the ``signal`` module).  The author believes that the overwhelming
majority of painful reentrancy issues occur with finalizers.  Most of the
time, signal handlers are able to set a single flag and/or wake up a
file descriptor for the main program to notice.  As for those signal
handlers which raise an exception, they *have* to execute in-thread.

This PEP also does not change the execution of finalization callbacks
when they are called as part of regular reference counting, i.e. when
releasing a visible reference drops an object's reference count to zero.
Since such execution happens at deterministic points in code, it is usually
not a problem.


Internal details
================

TODO: Update this section to conform to the current implementation.

``gc`` module
-------------

An internal flag ``gc_is_threaded`` is added, telling whether GC is serial
or threaded.

An internal structure ``gc_mutex`` is added to avoid two GC runs at once:

.. code-block::

   static struct {
       PyThread_type_lock lock;  /* taken when collecting */
       PyThreadState *owner;  /* whichever thread is currently collecting
                                 (NULL if no collection is taking place) */
   } gc_mutex;

An internal structure ``gc_thread`` is added to handle synchronization with
the GC thread:

.. code-block::

   static struct {
      PyThread_type_lock wakeup; /* acts as an event
                                    to wake up the GC thread */
      int collection_requested; /* non-zero if collection requested */
      PyThread_type_lock done; /* acts as an event signaling
                                  the GC thread has exited */
   } gc_thread;


``threading`` module
--------------------

Two private functions are added to the ``threading`` module:

* ``threading._ensure_dummy_thread(name)`` creates and registers a ``Thread``
  instance for the current thread with the given *name*, and returns it.

* ``threading._remove_dummy_thread(thread)`` removes the given *thread*
  (as returned by ``_ensure_dummy_thread``) from the threading module's
  internal state.

The purpose of these two functions is to improve debugging and introspection
by letting ``threading.current_thread()`` return a more meaningfully-named
object when called inside a finalization callback in the GC thread.


Pseudo-code
-----------

Here is a proposed pseudo-code for the main primitives, public and internal,
required for implementing this PEP.  All of them will be implemented in C
and live inside the ``gc`` module, unless otherwise noted:

.. code-block::

   def collect_with_callback(generation):
       """
       Collect up to the given *generation*.
       """
       # Same code as currently (see collect_with_callback() in gcmodule.c)


   def collect_generations():
       """
       Collect as many generations as desired by the heuristic.
       """
       # Same code as currently (see collect_generations() in gcmodule.c)


   def lock_and_collect(generation=-1):
       """
       Perform a collection with thread safety.
       """
       me = PyThreadState_GET()
       if gc_mutex.owner == me:
           # reentrant GC collection request, bail out
           return
       Py_BEGIN_ALLOW_THREADS
       gc_mutex.lock.acquire()
       Py_END_ALLOW_THREADS
       gc_mutex.owner = me
       try:
           if generation >= 0:
               return collect_with_callback(generation)
           else:
               return collect_generations()
       finally:
           gc_mutex.owner = NULL
           gc_mutex.lock.release()


   def schedule_gc_request():
       """
       Ask the GC thread to run an implicit collection.
       """
       assert gc_is_threaded == True
       # Note this is extremely fast if a collection is already requested
       if gc_thread.collection_requested == False:
           gc_thread.collection_requested = True
           gc_thread.wakeup.release()


   def is_implicit_gc_desired():
       """
       Whether an implicit GC run is currently desired based on allocation
       stats.  Return a generation number, or -1 if none desired.
       """
       # Same heuristic as currently (see _PyObject_GC_Alloc in gcmodule.c)


   def PyGC_Malloc():
       """
       Allocate a GC-enabled object.
       """
       # Update allocation statistics (same code as currently, omitted for brevity)
       if is_implicit_gc_desired():
           if gc_is_threaded:
               schedule_gc_request()
           else:
               lock_and_collect()
       # Go ahead with allocation (same code as currently, omitted for brevity)


   def gc_thread(interp_state):
       """
       Dedicated loop for threaded GC.
       """
       # Init Python thread state (omitted, see t_bootstrap in _threadmodule.c)
       # Optional: init thread in Python threading module, for better introspection
       me = threading._ensure_dummy_thread(name="GC thread")

       while gc_is_threaded == True:
           Py_BEGIN_ALLOW_THREADS
           gc_thread.wakeup.acquire()
           Py_END_ALLOW_THREADS
           if gc_thread.collection_requested != 0:
               gc_thread.collection_requested = 0
               lock_and_collect(generation=-1)

       threading._remove_dummy_thread(me)
       # Signal we're exiting
       gc_thread.done.release()
       # Free Python thread state (omitted)


   def gc.set_mode(mode):
       """
       Set current GC mode.  This is a process-global setting.
       """
       if mode == "threaded":
           if not gc_is_threaded == False:
               # Launch thread
               gc_thread.done.acquire(block=False)  # should not fail
               gc_is_threaded = True
               PyThread_start_new_thread(gc_thread)
       elif mode == "serial":
           if gc_is_threaded == True:
               # Wake up thread, asking it to end
               gc_is_threaded = False
               gc_thread.wakeup.release()
               # Wait for thread exit
               Py_BEGIN_ALLOW_THREADS
               gc_thread.done.acquire()
               Py_END_ALLOW_THREADS
               gc_thread.done.release()
       else:
           raise ValueError("unsupported mode %r" % (mode,))


   def gc.get_mode(mode):
       """
       Get current GC mode.
       """
       return "threaded" if gc_is_threaded else "serial"


   def gc.collect(generation=2):
       """
       Schedule collection of the given generation and wait for it to
       finish.
       """
       return lock_and_collect(generation)


Discussion
==========

Default mode
------------

One may wonder whether the default mode should simply be changed to "threaded".
For multi-threaded applications, it would probably not be a problem:
those applications must already be prepared for finalization handlers to
be run in arbitrary threads.  In single-thread applications, however, it
is currently guaranteed that finalizers will always be called in the main
thread.  Breaking this property may induce subtle behaviour changes or bugs,
for example if finalizers rely on some thread-local values.

Another problem is when a program uses ``fork()`` for concurrency.
Calling ``fork()`` from a single-threaded program is safe,
but it's fragile (to say the least) if the program is multi-threaded.

Explicit collections
--------------------

One may ask whether explicit collections should also be delegated to the
background thread.  The answer is it doesn't really matter: since
``gc.collect`` and ``PyGC_Collect`` actually *wait* for the collection to
end (breaking this property would break compatibility), delegating the
actual work to a background thread wouldn't ease synchronization with the
thread requesting an explicit collection.

In the end, this PEP choses the behaviour that seems simpler to implement
based on the pseudo-code above.

Impact on memory use
--------------------

The "threaded" mode incurs a slight delay in implicit collections compared
to the default "serial" mode.  This obviously may change the memory profile
of certain applications.  By how much remains to be measured in real-world
use, but we expect the impact to remain minor and bearable.  First because
implicit collections are based on a *heuristic* whose effect does not result
in deterministic visible behaviour anyway.  Second because the GC deals
with reference cycles while many objects are reclaimed immediately when their
last visible reference disappears.

Impact on CPU consumption
-------------------------

The pseudo-code above adds two lock operations for each implicit collection
request in "threaded" mode: one in the thread making the request (a
``release`` call) and one in the GC thread (an ``acquire`` call).
It also adds two other lock operations, regardless of the current mode,
around each actual collection.

We expect the cost of those lock operations to be very small, on modern
systems, compared to the actual cost of crawling through the chains of
pointers during the collection itself ("pointer chasing" being one of
the hardest workloads on modern CPUs, as it lends itself poorly to
speculation and superscalar execution).

Actual measurements on worst-case mini-benchmarks may help provide
reassuring upper bounds.

Impact on GC pauses
-------------------

While this PEP does not concern itself with GC pauses, there is a
practical chance that releasing the GIL at some point during an implicit
collection (for example by virtue of executing a pure Python finalizer)
will allow application code to run in-between, lowering the *visible* GC
pause time for some applications.

If this PEP is accepted, future work may try to better realize this potential
by speculatively releasing the GIL during collections, though it is unclear
how doable that is.


Open issues
===========

* ``gc.set_mode`` should probably be protected against multiple concurrent
  invocations.  Also, it should raise when called from *inside* a GC run
  (i.e. from a finalizer).

* What happens at shutdown?  Does the GC thread run until ``_PyGC_Fini()``
  is called?


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

A draft implementation is available in the ``threaded_gc`` branch
[#implementation-branch]_ of the author's Github fork [#cpython-pitrou-fork]_.


References
==========

.. [#queue-reentrancy-bug] https://bugs.python.org/issue14976

.. [#implementation-branch] https://github.com/pitrou/cpython/tree/threaded_gc

.. [#cpython-pitrou-fork] https://github.com/pitrou/cpython/


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: