bpo-38692: Skip test_posix.test_pidfd_open() on EPERM

[vstinner]

Skip test_posix.test_pidfd_open() test if os.pidfd_open() fails with
a PermissionError. It can happen in a Linux sandbox using a whitelist
of syscalls which doesn't allow pidfd_open syscall yet.

https://bugs.python.org/issue38692

[vstinner]

Concrete example such whitelist sandbox: systemd-nspawn doesn't allow pidfd_open syscall yet. This issue affects mock tool used to build packages on Fedora.
https://bugs.python.org/issue38692#msg357054

The issue was already discussed at:
https://bugs.python.org/issue38692#msg356289
@benjaminp wrote " It seems like systemd-nspawn is just breaking everything: https://sourceware.org/ml/libc-alpha/2019-11/msg00277.html "

Honestly, I'm not interested to discuss if a sandbox is "broken" or not. Such sandboxes exist, it's just a fact. IMHO it's not a big deal to add 2 lines to test_posix just to not be annoyed by the sandbox policy.

[hroncok]

This indeed workarounds the issue we have in Fedora. I will also try get pidfd_open whitelisted in systemd-nspawn, but that can take a while. In the meantime, this is IMHO the best we can get.

[vstinner]

@benjaminp @pablogsal @nanjekyejoannah : does it look like reasonable to you to skip the test if Python is running in a sandbox?

[pablogsal]

@benjaminp @pablogsal @nanjekyejoannah : does it look like reasonable to you to skip the test if Python is running in a sandbox?

This sounds very reasonable to me :)

[yan12125]

I got another test failure during our weekly python 3.9 update on Arch Linux:

======================================================================
FAIL: test_pidfd_send_signal (test.test_signal.PidfdSignalTest)
----------------------------------------------------------------------
Traceback (most recent call last):
  File "/build/python-git/src/cpython/Lib/test/test_signal.py", line 1287, in test_pidfd_send_signal
    self.assertEqual(cm.exception.errno, errno.EBADF)
AssertionError: 1 != 9

----------------------------------------------------------------------

Could that test be skipped in case of EPERM from the system call pidfd_send_signal, too?

[vstinner]

I got another test failure during our weekly python 3.9 update on Arch Linux: (...)

This PR fix it, no? You get a EPERM error (errno 1).

Could that test be skipped in case of EPERM from the system call pidfd_send_signal, too?

My PR skip pidfd_open() test if the syscall fails with EPERM. Python doesn't expose pidfd_send_signal() yet, so I don't understand your question.

[vstinner]

I got another test failure during our weekly python 3.9 update on Arch Linux: (...)

Would yo mind to describe your issue in https://bugs.python.org/issue38692 ? What is your kernel version? Are you running the Python test suite in any sandbox? If you are not using a sandbox, I'm surprised that you got a EPERM error.

[vstinner]

I got another test failure during our weekly python 3.9 update on Arch Linux: (...)

Oh. Is it a similar issue than https://bugs.python.org/issue38692#msg356235 where you ran the Python test suite using systemd-nspawn sandbox?

[vstinner]

Thanks for your review @pablogsal.