Issue 30028: make test.support.temp_cwd() fork-safe

The context manager test.support.temp_cwd() creates a temporary directory and removes it on exit. The test runner test.regrtest uses this context manager.

I observed an annoying behaviour of test.support.temp_cwd() on Linux/UNIX: if the code, that runs in the temp_cwd() context forks and if the forked child terminates (without calling exec), then the temporary directory will be removed twice: by the child and by the parent. This can cause errors in the parent, if the parent tries to access the no longer existing directory.

I discovered this problem, when a test in test_multiprocessing_fork failed and the test directory for the complete test.regrtest-run got removed. Of course all other tests failed too.

I propose to modify test.support.temp_cwd() to remove the created directory only, if the process id (os.getpid()) is unchanged. I'll create a pull request.

Is os.getpid always available? I have found hasattr(os, 'getpid') in the logging module.

The line is from 2006. Some parts of the logging module are much older, maybe even from the time Python had DOS support.

I had the same concerns about os.getpid(), but test.support uses it unconditionally in various places. See https://github.com/python/cpython/blob/master/Lib/test/support/__init__.py#L803 for an example.

Then LGTM.

I proposed PR 1074 to remove the hasattr(os, 'getpid') from logging.LogRecord constructor.

#28156 added HAVE_GETPID check.
Maybe, we should have dummy getpid() for CloudABI?

Naoki:
> #28156 added HAVE_GETPID check.
> Maybe, we should have dummy getpid() for CloudABI?

Oh. If os.getpid() is not always available, the documentation should be updated.

> Oh. If os.getpid() is not always available, the documentation should be updated.

Yes, but CloudABI is not officially supported platform.
They have many patches to run Python on CloudABI. https://github.com/NuxiNL/cloudabi-ports/tree/master/packages/python

How can we document about it?  I hadn't used it.

os.getlogin document says `Availability: Unix, Windows.` [1]
Is it proper way to indicate "there may be some minor platforms which doesn't provide this."?
[1] https://docs.python.org/3.6/library/os.html#os.getlogin

> os.getlogin document says `Availability: Unix, Windows.`

Originally this meant "not in Mac OS, DOS, OS/2, VMS". But now the support of all these platforms is dropped and the note is outdated.

New changeset 33dddac00ba8d9b72cf21b8698504077eb3c23ad by Gregory P. Smith (Anselm Kruis) in branch 'master':
bpo-30028: make test.support.temp_cwd() fork-safe (GH-1066)
https://github.com/python/cpython/commit/33dddac00ba8d9b72cf21b8698504077eb3c23ad

New changeset 694c5e0e1f7f23595fab314f26b89e241477ff18 by Miss Islington (bot) in branch '3.7':
bpo-30028: make test.support.temp_cwd() fork-safe (GH-1066)
https://github.com/python/cpython/commit/694c5e0e1f7f23595fab314f26b89e241477ff18

New changeset 61bd4d2e6319a3c5c3b9ce5f807b44a45cc1d4a1 by Gregory P. Smith (Anselm Kruis) in branch '2.7':
[2.7] bpo-30028: make test.support.temp_cwd() fork-safe (GH-1066) (GH-5825)
https://github.com/python/cpython/commit/61bd4d2e6319a3c5c3b9ce5f807b44a45cc1d4a1

New changeset 9c819a6a7d34594779fea3a25fd69ec5745e185e by Gregory P. Smith (Anselm Kruis) in branch '3.6':
[3.6] bpo-30028: make test.support.temp_cwd() fork-safe (GH-1066) (GH-5826)
https://github.com/python/cpython/commit/9c819a6a7d34594779fea3a25fd69ec5745e185e

messages: + msg291413
stage: patch review