Remove 31/32-bit s390 Linux support (s390-linux-gnu triplet) #87345
Comments
|
configure.ac contains triplet definitions for s390 and s390x mainframes. While s390x (Linux on IBM Z) is still widely supported, s390 is not. s390 was released in 1990 and discontinued in 1998, https://en.wikipedia.org/wiki/IBM_System/390 Ariadne Conill from Alpine pointed out that: pyca/cryptography#5771 (comment) I propose to remove check for __s390__ from configure.ac and configure. We might also want to remove __alpha__, __hppa__, and __m68k__ at a later point. DEC hasn't been around for a long time. |
tiran
added
3.10
|
I'm a Debian Developer and maintainer for multiple Debian Ports architectures. Please don't remove support for Alpha, HPPA, m68k, ia64, PowerPC, SuperH, SPARC as we're still maintaining these in Debian. Here are the latest build results for Python3.9 in Debian with build logs for all these architectures:
There are no issues reported and we heavily rely on Python. |
|
And, FWIW, I generally don't quite understand what the problem with old triplet definitions in configure.ac are. I assume they don't hurt anyone, do they? You never know what usecases your users have and as long as a code snippet doesn't interfere with other code, I don't see little point in removing it. |
|
It's about setting expectations. For some users, the presence of a platform triplet implies support for a platform. |
|
That's an argument I have personally never heard before and I have been dealing with a lot of architecture support in many packages. FWIW, lots of upstream projects have targets which are officially supported and others which are there but not supported and that was never really a problem. As long as the architectures are being maintained in the Linux kernel, GCC, binutils and glibc, they can be considered usable and maintained. The m68k architecture for example has a very active and large community due to the popularity of retro computing. This has lead to efforts which will soon bring Rust support to M68k and the Amiga. I therefore don't really understand what you gain when you make it harder for downstreams to use Python. |
I opened this ticket after a user told me that they grepped the source code of Python, found the string "s390", and concluded that s390 is still supported by us. This ticket is really just about s390. I said that we *might* want to remove other platform at a *later* point. For now I don't mind to keep them. We could also put unsupported platforms behind a --enable-unsupported-platforms guard. The option would allow you to compile on retro platforms and serve as indicator that the platform is not supported by us. |
I don't see a big benefit of removing a few lines of configure.ac. It's not like it's a big maintenance burden.
I maintain https://pythondev.readthedocs.io/platforms.html : unofficial list of platforms and C compilateurs supported by Python, and the different kinds of support levels (supported, best effort, not supported). |
|
Does AIX support the s390 architecture? |
Because one user was surprised by a few lines in configure.ac, the conclusion is to remove support for that architecture? That's a very odd justification. If it really just affects configure.ac, I don't really see a point in removing them. Even in OpenJDK we keep such unofficial architectures in configure.ac (with my OpenJDK upstream maintainer hat on). |
The platform triplet is s390-linux-gnu, not AIX.
You are misinterpreting my words. My initial posting explains that the hardware platform is no longer supported by any of our supported operating systems. I want to make it obvious that the platform has been dropped half a decade ago. |
vstinner
changed the title
vstinner
changed the title
|
I checked buildbots, we only have s390x workers (SLES, Debian, Fedora, RHEL7, RHEL8): There is no s390 worker running, so I would not say that we still support this platform. Since 2010, there were 5 commits (all made by me ;-) to fix tests) mentioning "s390" but all of them are about the 64-bit s390x arch, not the 32-bit arch. |
vstinner
changed the title
vstinner
changed the title
|
I created PR 24534 "Drop 32-bit Linux s390 platform support".
I would prefer to have one ticket per platform support removal. |
|
s390 is a 31-bit platform, not a 32-bit platform. I also don't see what this change achieves other than making the use of Python 3.10 on s390 harder. It's not like the removal of support for non-threaded builds which actually saved quite some code: |
That's a political statement, not a technical one. The change has zero functional impact on the other targets. It just makes the use of Python in a potential s390 VM harder. s390 packages are still being built for SUSE Linux Enterprise 12 which is still actively supported. I assume the same applies to RHEL LTS releases but I can't verify that as I have no insight into RedHat's internal build system. |
|
The guidelines for platform support are explained in PEP-11 (https://www.python.org/dev/peps/pep-0011/#supporting-platforms). We don't support platforms unless we have maintainers and CI (builtbots) in place for the platform. |
|
+1 from me. Whatever few users s390 still has, they can keep using Python 3.9 which is 5 years newer than the latest kernel they can use. Moving forward, s390 will be unambiguously unsupported as we cannot test against this platform. Unless we get a buildbot provided for this purpose, as well as somebody willing to fix broken builds on that buildbot long-term, it is what it is. |
There is no s390-specifc code in Python in the first place, is there? |
|
This has nothing to do with AIX. This conversation should include Charalampos Stratakis, but I don't see him as an option for Nosy. It probably is easy to add a s390 31-bit build to one of the buildbots. |
ARM64 only uses 48 bits for the address, but it uses 64-bit CPU words. Usually, we refer to an architecture by its CPU word, 32 or 64 bits. s390 uses 32-bit CPU words, no? The Wikipedia article says: "ESA/390 is arguably a 32-bit architecture; as with System/360, System/370, 370-XA, and ESA/370, the general-purpose registers are 32 bits long, and the arithmetic instructions support 32-bit arithmetic. Only byte-addressable real memory (Central Storage) and Virtual Storage addressing is limited to 31 bits."
Red Hat maintains Fedora and RHEL, none is supporting s390, but both support s390x:
Note: The correct name is Red Hat ;-) |
|
Łukasz Langa:
Nowadays, it became trivial to maintain downstream patches. It is easy to fork the Python Git repository and put your patches on top of it. Here it's only a matter of doing the maintenance upstream or not.
Wait. I would prefer to not have a buildbot worker unless there is a volunteer to fix all s390 specific issues. Otherwise, setting a buildbot worker will not solve any problem. I'm talking about build failures, test failures and issues on the worker directly. |
|
For RHEL7 which is the older OS that buildbots are still running, only the System Z architecture is supported. From the release notes [0]: Note that Red Hat Enterprise Linux 7 supports IBM zEnterprise 196 hardware or later; IBM System z10 mainframe systems are no longer supported and will not boot Red Hat Enterprise Linux 7. Also 32 (31 bits) in this case are only supported through virtualization with older OS's. So IMO it's fine to remove the support. |
|
And to dig a bit further with a semi-official answer. RHEL4 had standalone support for s390, while since RHEL5+ we've had only multilib support (64 bits kernel and possibility of s390 userspace packages). RHEL7 that is the oldest currently supported RHEL OS, does have multilib support, meaning that 32 bit (s390) userspace packages are available for s390x booting on 64 bit kernel. Although a later base python version for RHEL7 will not be shipped as we already have python2.7.5 and python3.8.6 supported there, which already builds for s390 for the aforementioned multilib support. On Software Collections where we actually sometimes ship later Python versions, we compile only for 64 bits so the removal of the s390 pieces wouldn't pose an issue here. Hence the only problem I can figure out from my analysis would be for users on s390x who would download the necessary 32bit libraries and dependencies from the repos and use the -m32 CFLAGS and LDFLAGS to get a 32 bits build. |
What is the use case or benefit of building Python for 32-bit rather than 64-bit? |
That's not really the question. The question is whether an upstream project should prevent downstreams from using unsupported target configurations and I think the answer to that question is no. Python is free software (as opposed to just open source software) and one of the key features of free software is that you don't tell your users how they use your software. Open source licenses that limit use cases of software are considered non-free by most if not all Linux distributions for that very reason. There are valid reasons for preventing your software from being built on certain targets - such as the maintenance burden for architecture-specific code - but none of them apply here. A few lines of autoconf plus some preprocessor macros don't pose any burden and therefore the choice should be in favor of allowing downstreams to build unsupported configurations. As for providing a CI: Setting up a CI machine for individual upstream projects is not a problem for big corporations like IBM or Intel, but it is certainly a hassle for individual open source developers and hobbyists. And while we (Debian Ports) have provided some CI machines for individual upstream projects such as GCC and LLVM, it should be sufficient for most upstream projects to rely on Debian's buildd infrastructure as we simply don't have the resources that big corporations have. As for your original question: We still maintain multiple 32-bit ports in Debian, both as official and unofficial releases, and the same is done in other Linux distributions such as Gentoo, openSUSE, Void and others. Lots of hobbyists are pouring a lot of lifeblood and efforts into these ports such as m68k - which has still a surprisingly large user base thanks to retro-computing fans - which is why I am kindly asking you to not put up any obstacles into our ways. As I said before, the Python interpreter is one of these excellent works of engineering that just work. Other interpreters/compilers such as OpenJDK, Ruby, Go or Rust require much more attention to keep them portable while the Python interpreter has never caused any issues which is something I am very grateful for, in particular given the fact how much other code directly depends on the Python interpreter to work (just think of the many package managers and other system tools written in Python). So I think I can speak for Debian, Gentoo and many other downstream projects that it is important for many that it stays that way. Of course, that shouldn't Python development keep from moving forward and if the dependence on architecture-dependent code should increase at some point, we can still discuss this issue again and we will be more than happy to help with the porting efforts. Thank You! |
We are not (actively) prevent unsupported target. We are merely removing unused code that we cannot test. Downstream is free to re-apply or maintain additional patches to enable platforms that we don't support. The rules for platform support are explained in PEP-11.
Free software doesn't mean free labor. Upstream is free to choose how we spend our time on the project or which patches we accept. For platform support we have (IMHO) reasonable rules for code: Python must compile, work, and pass a sufficient set of its test suite. While "Python" software is free, the trademark, logo, and rights to the name "Python" are owned by the PSF. The trademark rules are also reasonable. On very few occasions the PSF has asked developers to choose a different name to avoid confusion. For example we asked the developer of a Python 2.8 fork to rename. The fork is now known a Tauthon. Platform compatibility patches are fine. I see three ways to resolve the dispute:
I'm going to hold off and delay PR 24534 for at least two weeks to give you time to work on (1) or (3). |
|
I already am running a Debian s390x buildbot for Python. Someone can adjust the rules for the buildbot to include a 31-bit builder. The Debian buildbot has relatively few builder variants relative to the other s390x targets. |
|
David, this bug is about s390, not s390x. The s390x platform is supported and tested. |
|
I understand the issue is s390, not s390x. I am offering that there already is an s390x worker, so would it be sufficient to build and test Python in 31 bit mode on that worker as opposed to installing a complete s390 Debian system? |
To get a platform supported by Python, we also need a volunteer to fix issues specific to the 31 bit s390 platform: see PEP-11. The policy for platform support evolved last years. |
|
David, could you please provide the output of "gcc -dM -E - < /dev/null" on s390x in 31 bit mode? I'm curious and would like to see the platform constants. |
You do not need to support every platform. Just allow your users to use them. If something breaks downstream on a certain platform, distribution maintainers are usually happy to send patches to fix these issues. I think I can speak for myself here and many of my colleagues at SUSE, my fellow Debian Developers and the guys over at Gentoo who have also lots of talented folk who keep sending fixes for many upstream projects. My colleague Andreas, for example, has contributed multiple fixes in cpython on m68k:
|
This is kinda missing the point though. For example I've dealt a lot with the CPython codebase (and I'm also one of the Red Hat maintainers for RHEL and Fedora) and although I'm not a core developer, the codebase can be quite intimidating. I agree that some lines of code do not seem like much but they add up. If noone actually has a use case here, removing them would be the best option overall just from a cognitive perspective. Now I agree there are hobbyist's and so forth but you present an example, which is irrelevant to this case (m68k is not s390, right?). Do you actually know someone who is actively interested in the usecase of building a 32 bit python on an s390x system? Or do you know someone who owns an s390 system? Maybe someone who sends related fixes to another project? You claim to speak for other maintainers, yet I'd like to actually hear their position on that. Feel free to add them to the nosy list. If the latest kernel is not even booting, I don't think many people actually care about this arch to at least keep it running. On the other hand I do think though here that David would be the best person to speak about those use cases, as he is the most experienced here with the s390(x) architecture. Now from my personal point of view, I don't mind actually keeping the support there, but it will never come up downstream for me, hence it won't affect me as Fedora doesn't build anymore for s390 (and already posted before the case for RHEL). So again, do you know of this specific usecase, of someone either owning an s390 machine and compiling python, or actually utilizing an s390x machine and using the multilib packages to build a 32 bit python? If so, would you (or they) step up to fix related issues when they come up and assuming a buildbot would be set? Or would you know someone who would? While your arguments make sense up to a point, claiming that others will step up or saying what others have done for different architectures, doesn't really help your arguments to hold. |
|
gcc -m31 -dM -E - < /dev/null |
|
gcc -dM -E - < /dev/null |
|
I'm not sure why people insists to say that it's 31 bit CPU, gcc-s390.txt shows that it uses a classic 32-bit ABI (32-bit long/size_t): #define __SIZE_MAX__ 0xffffffffUL
#define __LONG_WIDTH__ 3231-bit is a limit of the address space (memory), not the ALU. |
|
I am not aware of significant use of 31 bit mode. But I don't see the benefit of annoying and discouraging users who want to experiment with Python and with Linux on Z in 31 bit mode. Yes, maintenance theoretically is a burden, but there have been no recent issues that were specific to Linux on Z in 31 bit mode. In fact, most of the original Linux on Z support issues that I opened were endianness issues, which aren't ameliorated by removing 31 bit support. As others have expressed, deprecating 31 bit mode only removes the configuration option with no other code simplification. It seems that it would be better to leave the configuration alone until there actually was an unresolved issue that motivated removal. But I am not aware of any client requirement to continue the support. Leaving the 31 bit configuration unchanged is more about maintaining good will among the volunteers who are interested in the platform. |
|
Victor: https://en.wikipedia.org/wiki/31-bit_computing :-) |
Well, yes, I have just got it confirmed from our PM, SUSE has living and breathing (and what’s even more interesting, paying) clients using SLE-12 on s390 (i.e., s390x kernel with 32bit libraries). Of course, it doesn’t mean they are eager to run Python 3.10 in such configuration (how do you call such package? python3a so that “python3a” > “python39”?), we will have to evaluate with them what are their plans in the post-36 world with the modern Python there. |
Fully agree.
As I have mentioned before, the Python interpreter in general has been very very unproblematic on any platform. The only real issue that I am aware of is that the testsuite can get stuck on machines with a large number of CPUs (we're seeing that on our SPARC T5 in Debian).
And there is still MIPS-BE, PPC-BE, M68k, HPPA among other which are all big-endian.
Exactly my point. If it removed a considerable amount of code, I would actually see a point. But just removing a few lines of autoconf or preprocessor code makes no differences from a maintainer's point of view.
Jepp, fully agree.
Sure. But free software shouldn't be solely about commercial customers. If someone wants to play with Python on the s390 emulator Hercules, for example, upstream projects shouldn't be keeping them from doing that.
Absolutely. And about not limiting choices when there is no technical argument for it. |
|
I created a thread on python-dev: "Move support of legacy platforms/architectures outside Python" |
Motorola 68k isn't a 16-bit architecture, it's a 32-bit architecture. Also, I am starting to get the feeling that you are trying to escalate a conflict here. None of the code that you are complaining about is hurting anyone. Yet, you think it is important to keep this discussion and ruining my day. The fact that you take sides for OpenBSD but consider m68k legacy and unmaintained (which isn't correct at all - look at the kernel), shows that you are biased in this discussion. |
|
Oh, and LLVM is currently gaining support M68k which you consider "legacy": It might be a good idea to do some research first before making such statements. |
|
For the last time: This ticket is solely about s390 platform. Please stop derailing this ticket with comments about unrelated platforms like m68k. |
|
Christian, The Python Community Code of Conduct also states: Being respectful of differing viewpoints and experiences. Various developers are passionate about this topic and the entire series of comments has been respectful, IMHO. I don't believe that it is helpful when you cherry-pick a rule from the code of conduct in response to someone providing additional evidence. That is intimidation and shutting down the conversation with an implicit threat. You specifically mentioned m68k when you opened the issue. This conversation feels very much that you and Victor have pre-determined an outcome and are not open to other points of view. I and others have proposed compromises, but the response appears, at least to me, as shifting requirements. |
|
My offer still stands: If you can fulfill the requirements of PEP-11 for s390, then I'm fine with keeping the code for s390 around. Victor has a different opinion, so you have to contact the Steering Council and get their approval, too. Our ticket system doesn't permit amending comments. Otherwise I would have clarified my comment. Also I used the words "might" and "at a later point" in my initial ticket. |
|
https://bugs.python.org/issue43179 and https://mail.python.org/archives/list/python-dev@python.org/thread/F5BXISYP7RAINXUMYJSEYG7GCFRFAENF/ discussions didn't reach kind of consensus. I'm tired of these discussions, so I just closed my PR 24534. If someone wants to drop support of the 31-bit Linux s390 platform, please go ahead, feel free to copy my patch. |
|
I think there is one productive result of this discussion which is this patch by Jessica Clark which gets rid of architecture-specific alignment code: Unfortunately, it has not seen any positive reviews yet. Getting this merged would remove some maintenance burden from the CPython maintainers. |
|
The report/justification for the removal is simply incorrect. Linux still supports s390-*-linux-gnu in user mode. The GNU toolchain is maintained, and the glibc testsuite is in good shape. Some distributions still support s390 (31-bit) applications. Others (Fedora 24 or later, for example) no longer do so. I expect that 31-bit s390 Linux Python is really unusual these days, so removing it could still be fine, but lack of support from the GNU toolchain or the Linux kernel is not the correct justification for its removal. If you have questions about the maintenance status of various parts of the GNU toolchain, please feel free to reach out to any of the GNU lists. https://gcc.gnu.org/mailman/listinfo/gcc |
|
FWIW, Gentoo is still building 31bit s390 stages. Since our package manager is python-based we'd probably notice bad breakage... |
|
@akhuettel is there a timeline for Gentoo dropping support for this architecture? |
Note: these values reflect the state of the issue at the time it was migrated and might not reflect the current state.
Show more details
GitHub fields:
bugs.python.org fields:
The text was updated successfully, but these errors were encountered: