Re: F37 kernel 6.0.16/6.0.18 breaking Python tests: Allows to bind a socket twice

[Justin Forbes]

On Wed, Jan 11, 2023 at 6:53 AM Miro Hrončok <mhroncok(a)redhat.com&gt; wrote:
>  Hello.
>
>  On Monday I observed that the Fedora 37 and 36 Python CI tests on the Testing
>  farm fail.
>
>
> https://lists.fedoraproject.org/archives/list/[e-mail]/...
>
>  The same tests work on Rawhide and worked on Fedora 37/36 previously --
>  unfortunately this could have been a while ago, maybe even couple weeks, due to
>  the holidays.
>
>  See https://src.fedoraproject.org/rpms/python3.11/pull-request/99
>
>  The test failures look like this:
>
>  ======================================================================
>  FAIL: test_bind (test.test_asyncore.TestAPI_UseIPv4Poll.test_bind)
>  ----------------------------------------------------------------------
>  Traceback (most recent call last):
>     File "/usr/lib64/python3.11/test/test_asyncore.py", line 752, in test_bind
>       self.assertRaises(OSError, s2.bind, (self.addr[0], port))
>  AssertionError: OSError not raised by bind
>
>  ======================================================================
>  FAIL: test_bind (test.test_asyncore.TestAPI_UseIPv6Select.test_bind)
>  ----------------------------------------------------------------------
>  Traceback (most recent call last):
>     File "/usr/lib64/python3.11/test/test_asyncore.py", line 752, in test_bind
>       self.assertRaises(OSError, s2.bind, (self.addr[0], port))
>  AssertionError: OSError not raised by bind
>
>
>  All the failed tests are related to either multiprocessing or asyncio.
>  All of them are either:
>
>     AssertionError: OSError not raised by bind
>
>  Or:
>
>     AssertionError: OSError not raised by Listener
>
>
>  The same tests pass during the Koji build when the package builds.
>
>  The kernel in F37 Koji is 6.0.10.
>  The kernel on the F37 Testing Farm is 6.0.16.
>
>
>  I've checked and all the failing tests basically do:
>
>    - bind a socket
>    - bind it again (and assert failure, which fails)
>
>  Petr Viktorin found out that SO_REUSEPORT socket option exists which when
>  enabled might cause the assertions to fail.
>
>  But that should not be possible to set from outside of the process.
>
>  https://lwn.net/Articles/542629/
>
>  I've installed kernel 6.0.16-300.fc37.x86_64 and I can reproduce this with e.g.:
>
>  $ sudo dnf install python3-test
>  $ python3 -m test -v test_asyncore
>
>  Same with 6.0.18-300.fc37.x86_64.
>
>  I cannot reproduce this with Rawhide's
>  6.2.0-0.rc2.20230105git41c03ba9beea.20.fc38.x86_64.
>
>
>  Is this a bug that needs to be fixed or do the tests need to be changed not to
>  assert this?
>
>  The fact that 6.2.0 makes this go away makes me think this is a kernel bug but
>  I am not sure. 
This does seem a bug. The big question, is does 6.1 make it go away?
kernel-6.1.4-200.fc37 is available in koji. The 6.0 series is end of
support for Fedora, and I expect 6.1.5 to be available as an update
this week.

Thanks,
Justin


>
> --
> Miro Hrončok
> --
> Phone: [phone]
> IRC: mhroncok
>