gh-102251: add missing cleanups for test_import

[sunmy2019]

Leak reasons are described here.

#103879 (comment)

For 1, the fix is straightforward.

For 2, I added 2 clean-up functions and trigger them manually.

For 3, I haven't figured out a clean solution, but I would prefer to skip it under ref leak modes only. So I introduce a new method because AFAIK there isn't a way to skip only ref leak tests.

• Issue: Fix Refleak in test_import #102251

[Eclips4]

As I understand, this doesn't solve problem, just "shuts up" the noise from known refleaks on test_import, right?

[sunmy2019]

As I understand, this doesn't solve problem, just "shuts up" the noise from known refleaks on test_import, right?

That is not true. I do think I did resolve all ref leaks on Linux.

But I did not test it on Windows.

[Eclips4]

As I understand, this doesn't solve problem, just "shuts up" the noise from known refleaks on test_import, right?

That is not true. I do think I did resolve all ref leaks on Linux.

But I did not test it on Windows.

Oh, seems on Windows problem doesn't solved. I rebuilt intepreter with your changes and got it:

 ./python-mtest-R3:3test_importRunningDebug|x64interpreter... 0:00:00Runtestssequentially0:00:00 [1/1] test_importbeginning6repetitions123456 ...... test_importleaked [41, 43, 40] references, sum=124test_importleaked [41, 41, 40] memoryblocks, sum=122test_importfailed (referenceleak) ==Testsresult: FAILURE==1testfailed: test_importTotalduration: 28.7secTestsresult: FAILURE

[sunmy2019]

This trick cannot be used with ./python -m test -j n, thus causing failures on CI.

[sunmy2019]

All ref leaks on Linux come from SinglephaseInitTests, but running

 .\python_d.exe -m test test_import -R 3:3 -v -i *SinglephaseInitTests* 

on Windows does give ref leaks

OK (skipped=5) . test_import leaked [41, 43, 40] references, sum=124 test_import leaked [41, 41, 40] memory blocks, sum=122 

[sunmy2019]

@Eclips4 Can you bisect which test leaks ref?

[Eclips4]

@Eclips4 Can you bisect which test leaks ref?

Sure!
If we delete ImportTests/test_concurrency refleak tests are passed. Seems that problem in this test

[Eclips4]

Also, I tried to found commit which introduce this reference leak ( in test_concurrency)
and again bisected to #94504..

[Eclips4]

Oh, seems your solution also works on Windows. ( I forgot to add -j 1 parameter 🤦‍♂️ , sorry)
But problem with test_concurrency still there, and that's interesting, because it doesn't reproducible on Linux.

Probably, it should be a another PR (at the current moment I still haven't found a solution).
My some notes about it:
If we delete these lines from test_concurrency:

sys.modules.pop('package', None) sys.modules.pop('package.submodule', None)

Refleak tests will pass.
Also, this problem confirms that Windows runner in buildbots fleet really needed.

[erlend-aasland]

The libregrtest modification seems way too hacky for my taste. What happens if you apply my suggestion for _testsinglephase.c, and revert all other changes introduced in this PR?

[sunmy2019]

revert all other changes introduced in this PR?

There are 3 leak points when running in Linux. Applying your suggestions only fixes one of them.

[erlend-aasland]

revert all other changes introduced in this PR?

There are 3 leak points when running in Linux. Applying your suggestions only fixes one of them.

I suggest you turn this PR¹ into a fix for this one issue, so we can land that fix. The fix is obvious; I can land that for you today. The rest of the changes are controversial; I think it would be wise to separate them out.

Also: please don't tag people with in commit messages. Those commits ends up as noise in the GitHub notifications page.

Footnotes

1. or create a new PR for this ↩

[sunmy2019]

1. or create a new PR for this ↩

I will create a new one.

---

#105082

[erlend-aasland]

Status for buildbot run for commit 20307d0:

• buildbot/AMD64 Fedora Stable Refleaks PR: test_peg_generator leaked
• buildbot/AMD64 RHEL7 Refleaks PR: test_peg_generator leaked
• buildbot/AMD64 RHEL8 Refleaks PR: test_peg_generator leaked
• buildbot/PPC64LE RHEL8 Refleaks PR: test_peg_generator leaked
• buildbot/aarch64 RHEL8 Refleaks PR: test_peg_generator leaked
• buildbot/s390x Fedora Refleaks PR: test_peg_generator leaked
• buildbot/s390x RHEL7 Refleaks PR: test_peg_generator leaked
• buildbot/s390x RHEL8 Refleaks PR: test_peg_generator leaked

[ericsnowcurrently]

LGTM

[ericsnowcurrently]

GitHub took me straight to "Files changed", so I didn't see the other comments before I gave an approving review. The change looks fine to me, but please be sure to address any other comments before this is merged.

[sunmy2019]

The second part is split to #105085

I will make this PR focus on the 3rd
part of the problem.

[sunmy2019]

Now this PR only has changes related to the 3rd part.

---

test_basic_multiple_interpreters_deleted_no_reset deliberately leaks things. And we have a "no-leak" version test right after this: test_basic_multiple_interpreters_reset_each.

I think the correct thing to do here is to skip this specific test when hunting ref leaks.

---

However, the previous approach does not work under -j. And the test main has no way to pass the information to the actual test. So I came up with the name trick.

I am open to any other methods.

[bedevere-bot]

🤖 New build scheduled with the buildbot fleet by @sunmy2019 for commit 5934e50 🤖

If you want to schedule another build, you need to add the 🔨 test-with-refleak-buildbots label again.

[bedevere-bot]

🤖 New build scheduled with the buildbot fleet by @sunmy2019 for commit 60b4921 🤖

If you want to schedule another build, you need to add the 🔨 test-with-refleak-buildbots label again.

[erlend-aasland]

Closing in favour of gh-106013. Thanks for the effort you've put into this @sunmy2019; highly appreciated!

Also, thanks for keeping up with me; sorry for pushing you in different directions, but my opinions changed.