repoview: fix memory leak of filtered repo classes
The leak occurs in long-running server processes with
extensions, and is measured at 110kB per request.
Before this change, the contents of the `_filteredrepotypes`
cache are not properly garbage collected, despite it begin
a `WeakKeyDictionary`.
Extensions have a tendency to generate a new repository class
for each `localrepo` instantiation. Server processes based
on `hgwebdir_mod` will instantiate a new `localrepo` for each
HTTP request that involves a repository.
As a result, with a testing process that repeatedly opens a
repository with several extensions activated
(`topic` notably among them), we see a steady increase in
resident memory of 110kB per repository instantiation before this
change. This is also true, if we call `gc.collect()` at each
instantiation, like `hgwebdir_mod` does, or not.
The cause of the leak is that the *values* aren't weak references.
This change uses `weakref.ref` for the values, and this makes
in our measurements the resident size increase drop to 5kB per
repository instantiation, with no explicit call of `gc.collect()`
at all.
There is currently no reason to believe that this remaining leak
of 5kB is related to or even due to Mercurial core.
We've also seen evidence that `ui.ui` instances weren't properly
garbage collected before the change (with the change, they are).
This could explain why the figures are relatively high.
In theory, the collection of weak references could lead to
much more misses in the cache, so we measured the impact on
the original case that was motivation for introducing that cache
in
7e89bd0cfb86 (see also
issue5043): `hg convert` of the
mozilla-central repository. The bad news here is that there is a
major memory leak there, both with and without the present changeset.
There were no more cache misses, and we could see no
more memory leak with this change: the resident size after importing
roughly 100000 changesets was at 12.4GB before, and 12.5GB after.
The small increase is mentioned for completeness only, and we
believe that it should be ignored, at least as long as the main
leak isn't fixed. At less than 1% of the main leak, even finding out
whether it is merely noise would be wasteful.
Original context where this was spotted and first mitigated:
https://foss.heptapod.net/heptapod/heptapod/-/issues/466
The leak reduction was also obtained in Heptapod inner HTTP server,
which amounts to the same as `hgwebdir_mod` for these questions.
The measurements done with Python 3.9, similar figures seen with 3.8.
More work on our side would be needed to give measurements with 2.7,
because of testing server process does not support it.
# demandimportpy3 - global demand-loading of modules for Mercurial
#
# Copyright 2017 Facebook Inc.
#
# This software may be used and distributed according to the terms of the
# GNU General Public License version 2 or any later version.
"""Lazy loading for Python 3.6 and above.
This uses the new importlib finder/loader functionality available in Python 3.5
and up. The code reuses most of the mechanics implemented inside importlib.util,
but with a few additions:
* Allow excluding certain modules from lazy imports.
* Expose an interface that's substantially the same as demandimport for
Python 2.
This also has some limitations compared to the Python 2 implementation:
* Much of the logic is per-package, not per-module, so any packages loaded
before demandimport is enabled will not be lazily imported in the future. In
practice, we only expect builtins to be loaded before demandimport is
enabled.
"""
# This line is unnecessary, but it satisfies test-check-py3-compat.t.
from __future__ import absolute_import
import contextlib
import importlib.util
import sys
from . import tracing
_deactivated = False
# Python 3.5's LazyLoader doesn't work for some reason.
# https://bugs.python.org/issue26186 is a known issue with extension
# importing. But it appears to not have a meaningful effect with
# Mercurial.
_supported = sys.version_info[0:2] >= (3, 6)
class _lazyloaderex(importlib.util.LazyLoader):
"""This is a LazyLoader except it also follows the _deactivated global and
the ignore list.
"""
def exec_module(self, module):
"""Make the module load lazily."""
with tracing.log('demandimport %s', module):
if _deactivated or module.__name__ in ignores:
self.loader.exec_module(module)
else:
super().exec_module(module)
class LazyFinder(object):
"""A wrapper around a ``MetaPathFinder`` that makes loaders lazy.
``sys.meta_path`` finders have their ``find_spec()`` called to locate a
module. This returns a ``ModuleSpec`` if found or ``None``. The
``ModuleSpec`` has a ``loader`` attribute, which is called to actually
load a module.
Our class wraps an existing finder and overloads its ``find_spec()`` to
replace the ``loader`` with our lazy loader proxy.
We have to use __getattribute__ to proxy the instance because some meta
path finders don't support monkeypatching.
"""
__slots__ = ("_finder",)
def __init__(self, finder):
object.__setattr__(self, "_finder", finder)
def __repr__(self):
return "<LazyFinder for %r>" % object.__getattribute__(self, "_finder")
# __bool__ is canonical Python 3. But check-code insists on __nonzero__ being
# defined via `def`.
def __nonzero__(self):
return bool(object.__getattribute__(self, "_finder"))
__bool__ = __nonzero__
def __getattribute__(self, name):
if name in ("_finder", "find_spec"):
return object.__getattribute__(self, name)
return getattr(object.__getattribute__(self, "_finder"), name)
def __delattr__(self, name):
return delattr(object.__getattribute__(self, "_finder"))
def __setattr__(self, name, value):
return setattr(object.__getattribute__(self, "_finder"), name, value)
def find_spec(self, fullname, path, target=None):
finder = object.__getattribute__(self, "_finder")
try:
find_spec = finder.find_spec
except AttributeError:
loader = finder.find_module(fullname, path)
if loader is None:
spec = None
else:
spec = importlib.util.spec_from_loader(fullname, loader)
else:
spec = find_spec(fullname, path, target)
# Lazy loader requires exec_module().
if (
spec is not None
and spec.loader is not None
and getattr(spec.loader, "exec_module", None)
):
spec.loader = _lazyloaderex(spec.loader)
return spec
ignores = set()
def init(ignoreset):
global ignores
ignores = ignoreset
def isenabled():
return not _deactivated and any(
isinstance(finder, LazyFinder) for finder in sys.meta_path
)
def disable():
new_finders = []
for finder in sys.meta_path:
new_finders.append(
finder._finder if isinstance(finder, LazyFinder) else finder
)
sys.meta_path[:] = new_finders
def enable():
if not _supported:
return
new_finders = []
for finder in sys.meta_path:
new_finders.append(
LazyFinder(finder) if not isinstance(finder, LazyFinder) else finder
)
sys.meta_path[:] = new_finders
@contextlib.contextmanager
def deactivated():
# This implementation is a bit different from Python 2's. Python 3
# maintains a per-package finder cache in sys.path_importer_cache (see
# PEP 302). This means that we can't just call disable + enable.
# If we do that, in situations like:
#
# demandimport.enable()
# ...
# from foo.bar import mod1
# with demandimport.deactivated():
# from foo.bar import mod2
#
# mod2 will be imported lazily. (The converse also holds -- whatever finder
# first gets cached will be used.)
#
# Instead, have a global flag the LazyLoader can use.
global _deactivated
demandenabled = isenabled()
if demandenabled:
_deactivated = True
try:
yield
finally:
if demandenabled:
_deactivated = False