test_pair() and test_speech128() tests of test_zlib fail on the s390x architecture if zlib uses the s390x hardware accelerator.

RHEL8 downstream issues (with most details):
https://bugzilla.redhat.com/show_bug.cgi?id=1974658

Fedora downstream issues:
https://bugzilla.redhat.com/show_bug.cgi?id=2038848

The s390x has a hardware accelerator for zlib. Depending if the hardware accelerator is used or not, the output (compress data) is different.

Also, test_zlib compress data in two different ways and then expect the same output. test_zlib pass with the software implementation which creates a single (final) compressed block. test_zlib fails with the hardware implementation which creates multiple compressed blocks (the last one is a final block).

Another reason the output differs is the FHT/DHT heuristic. The zlib deflate algorithm can analyze the data distribution and decide whether it wants to use a fixed-Huffman table (FHT) or a dynamic-Huffman table (DHT) for the next block, but the accelerator can't. Furthermore, looking at data in software would kill the accelerator performance. Therefore the following heuristic is used on s390x: the first 4k are compressed with FHT and the rest of the data with DHT. So, compress() creates a single FHT block. compressobj() creates a FHT block, a DHT block and a trailing block.

It is *not a bug* in zlib: the decompression gives back the original content as expected in all cases. The issue is that Python test_zlib makes too many assumptions on how "streamed" data should be compressed. The test expected that compressed data using different ways to call zlib would return the exact same compressed data. If an accelarator is used, it's not always the case.

I checked os.uname() value on two buildbots, by looking at their test.pythoninfo:

os.uname: posix.uname_result(sysname='Linux', nodename='ztcpip3.pok.ibm.com', release='3.10.0-1160.53.1.el7.s390x', version='#1 SMP Thu Dec 16 04:33:52 EST 2021', machine='s390x')

os.uname: posix.uname_result(sysname='Linux', nodename='dje', release='5.10.0-11-s390x', version='#1 SMP Debian 5.10.92-1 (2022-01-18)', machine='s390x')

So the machine is 's390x'.

New changeset 9475dc0 by Victor Stinner in branch 'main':
bpo-46623: Skip two test_zlib tests on s390x (GH-31096)
9475dc0

Even if Python 3.9 and 3.10 are also affected, I prefer to not backport the change since it's not ideal.

The Python 3.7 package on Fedora also skips these two tests using --ignore option of regrtest:

%ifarch s390x
--ignore test_speech128 \\
--ignore test_pair \\
%endif

• https://src.fedoraproject.org/rpms/python3.7/pull-request/38#request_diff
• https://src.fedoraproject.org/rpms/python3.7/blob/rawhide/f/python3.7.spec#_1168

The current fix skips all of the tests, rather than just the parts that fail.

The current fix skips all of the tests, rather than just the parts that fail.

Right. In short, I copied the Fedora fix to Python as 9475dc0 because I know that it works on s390x.

But I don't have access to a s390x machine with hardware accelerator, so I cannot develop a change which only skips the tests if hardware accelerator is present, or only skip the part which is affected by the hardware accelerator.

test_zlib: test_pair() and test_speech128() fail with s390x hardware accelerator

The initial scope of the issue is fixed, so I just close the issue.

If you manage to write and test a better fix, please open a separated issue/PR.