homepage

When Louie Lu posted a link to
https://blog.louie.lu/2017/06/12/diagnosing-and-fixing-reference-leaks-in-cpython/
on core-mentorship list, I tested idlelib.
python -m test -ugui test_idle # SUCCESS, no extraneous output
python -m test -R: test_idle # SUCCESS, no extraneous output
python -m test -R: -ugui test_idle # error output, FAILURE
[So people who leaktest without a screen see nothing in idlelib.]
Error output is about 20 copies of the following:
can't invoke "event" command: application has been destroyed
 while executing
"event generate $w <<ThemeChanged>>"
 (procedure "ttk::ThemeChanged" line 6)
 invoked from within
"ttk::ThemeChanged"
At the end:
test_idle leaked [471, 471, 471, 471] references, sum=1884
test_idle leaked [209, 211, 211, 211] memory blocks, sum=842
[similar for python 3.6]
In a response email, I noted that test_idle gathers tests from idlelib.idle_test.test_* and that something extra is needed to pin leaks to specific test modules.
I don't know whether the absence of 'invoke event' error messages when not running -R means that there are also no refleaks, or not.
---
import os
import subprocess
os.chdir('f:/dev/3x/Lib/idlelib/idle_test')
testfiles = [name for name in os.listdir() if name.startswith('test_')]
for name in testfiles:
 os.rename(name, 'x'+name)
for name in testfiles:
 os.rename('x'+name, name)
 try:
 res = subprocess.run(
 ['f:/dev/3x/python.bat', '-m', 'test', '-R:', '-ugui', 'test_idle'],
 stdout=subprocess.PIPE, stderr=subprocess.PIPE)
 if res.returncode:
 print(name)
 print(res.stderr)
 except Exception as err:
 print(name, err)
 os.rename(name, 'x'+name)
for name in testfiles:
 os.rename('x'+name, name)
---
reports
test_macosx.py
b'beginning 9 repetitions\r\n123456789\r\n\r\ntest_idle leaked [31, 31, 31, 31] references, sum=124\r\ntest_idle leaked [19, 21, 21, 21] memory blocks, sum=82\r\n'
test_query.py
b'beginning 9 repetitions\r\n123456789\r\n\r\ntest_idle leaked [429, 429, 429, 429] references, sum=1716\r\ntest_idle leaked [190, 192, 192, 192] memory blocks, sum=766\r\n'
There are no 'invoke event' messages.
For further testing within each file, by commenting out code, as suggested in the link above, I replaced 'testfiles' in the middle loop with ['testmacosx.py'] or ['test_query.py']. For test_macosx, the culprit is class SetupTest. For test_query, the culprit is class QueryGuiTest. Adding cls.root.update_idletasks did not solve the problem by itself (as it has in other cases). I plan to continue another time.

test_query were fixed in PR 2147, which is leak by not removing mock.Mock() in dialog.

New changeset b070fd275b68df5c5ba9f6f43197b8d7066f0b18 by terryjreedy (mlouielu) in branch 'master':
bpo-30642: IDLE: Fix test_query refleak (#2147)
https://github.com/python/cpython/commit/b070fd275b68df5c5ba9f6f43197b8d7066f0b18

New changeset 2bfb45d447c445b3c3afc19d16b4cd4773975993 by terryjreedy in branch '3.6':
bpo-30642: IDLE: Fix test_query refleak (#2147) (#2161)
https://github.com/python/cpython/commit/2bfb45d447c445b3c3afc19d16b4cd4773975993

The macosx leak came from repeated calls to
 root.createcommand("::tk::mac::OpenDocument", doOpenFile)
in macosx.addOpenEventSupport(root, flist), which is called from macosx.setupApp, which is called in test_macosx.SetupTest.test_setupapp.
Attached is an improved version of findleak.py, which made it easy to isolate one test file by renaming, repeatedly determine whether the leak remained after modifying either macosx.py or test_macosx.py, and revert all the other names when done.

New changeset b0efd493b6af24a6ae744e7e02f4b69c70e88f3d by terryjreedy in branch '3.6':
[3.6]bpo-30642: Fix ref leak in idle_test.test_macosx (#2163) (#2165)
https://github.com/python/cpython/commit/b0efd493b6af24a6ae744e7e02f4b69c70e88f3d

f:\dev36円>python -m test -R: -ugui test_idle
gives about 40 invoke event messages, but the test passes. So the messages and leaks are not connected.

Unlinked old findleak.py after uploading much improved version to #31130.