Skip to content

PyThread assumes pthread_key_t is an integer, which is against POSIX #69844

Closed

Description

@EdSchouten
BPO 25658
Nosy @ncoghlan, @scoder, @vstinner, @bitdancer, @embray, @ma8ma, @EdSchouten
PRs
  • bpo-25658: Implement PEP 539 for Thread Specific Storage (TSS) API #1362
  • bpo-31828: make Py_tss_NEEDS_INIT usable in non-static initialisers #4060
  • Files
  • issue25658-1.patch
  • key-constant-time.diff
  • pythread-key-intptr_t.patch
  • configure-pthread_key_t.patch
  • pythread-tss.patch
  • pythread-tss-2.patch
  • pythread-tss-3.patch: minor changes for version 2
  • 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:

    assignee = None
    closed_at = <Date 2017-10-06.10:45:57.315>
    created_at = <Date 2015-11-18.15:50:16.862>
    labels = ['interpreter-core', 'build', '3.7']
    title = 'PyThread assumes pthread_key_t is an integer, which is against POSIX'
    updated_at = <Date 2017-10-20.16:09:32.004>
    user = 'https://github.com/EdSchouten'

    bugs.python.org fields:

    activity = <Date 2017-10-20.16:09:32.004>
    actor = 'scoder'
    assignee = 'none'
    closed = True
    closed_date = <Date 2017-10-06.10:45:57.315>
    closer = 'ncoghlan'
    components = ['Interpreter Core']
    creation = <Date 2015-11-18.15:50:16.862>
    creator = 'EdSchouten'
    dependencies = []
    files = ['44269', '44303', '45427', '45548', '45763', '46227', '46379']
    hgrepos = []
    issue_num = 25658
    keywords = ['patch']
    message_count = 42.0
    messages = ['254846', '273441', '273443', '273451', '273453', '273460', '273464', '273480', '273916', '273919', '274030', '274035', '274040', '280511', '280512', '280525', '280526', '280527', '281227', '281990', '282024', '282033', '282404', '282410', '285039', '285328', '285478', '286012', '290035', '292625', '292841', '298342', '298347', '298674', '298676', '299542', '301768', '303814', '304650', '304652', '304657', '304663']
    nosy_count = 7.0
    nosy_names = ['ncoghlan', 'scoder', 'vstinner', 'r.david.murray', 'erik.bray', 'masamoto', 'EdSchouten']
    pr_nums = ['1362', '4060']
    priority = 'normal'
    resolution = 'fixed'
    stage = 'resolved'
    status = 'closed'
    superseder = None
    type = 'compile error'
    url = 'https://bugs.python.org/issue25658'
    versions = ['Python 3.7']

    Activity

    1. EdSchouten commented on Nov 18, 2015

      EdSchoutenmannequin
      MannequinAuthor

      While trying to port Python over to a new platform (CloudABI), I noticed a couple of compiler errors in PyThread_create_key(), PyThread_delete_key(), PyThread_delete_key_value() and PyThread_set_key_value() caused by fact that pthread_key_t is converted to an integer (and vice versa)

      POSIX doesn't seem to require that pthread_key_t is an integer or any other arithmetic type:

      http://pubs.opengroup.org/onlinepubs/9699919799/basedefs/sys_types.h.html

      Old revisions of the standard did require pthread_t to be an arithmetic type, but this requirement was dropped later on.

      In my opinion we should strongly consider changing the API, so that we can treat the key created by pthread_key_create() or returned by TlsAlloc() as an opaque type.

    2. embray commented on Aug 23, 2016

      @embray
      Contributor

      I agree--this has the same problem on Cygwin, where pthread_key_t is not just a typedef'd integer (in fact it's a pointer to an instance of a class).

      Anyways as Ed wrote above POSIX says this is supposed to be an opaque type and there's no reason to assume it's an integer. Linux could change its definition tomorrow and we'd have no one to blame but ourselves if it breaks Python.

      If it's too hard to change the Python API, at the very least the change in bpo-22206 should be reverted or reworked somehow, because there's no reason to assume that pthread_key_t can even be compared safely to an integer, much less that it would be less than INT_MAX.

    3. vstinner commented on Aug 23, 2016

      @vstinner
      Member

      What do you suggest? Python C code bases uses "long" everywhere.

    4. embray commented on Aug 23, 2016

      @embray
      Contributor

      I'm not really sure what "long" has to do with it...

      The problem is that the PyThread API uses ints to represent TLS keys, and has for about as long as it's existed (maybe what you meant by "long"). But when support for native TLS was added (bpo-9786 for pthread, sometime earlier for Windows) , the faulty assumption as made in several places that this API (i.e. the type of key is "int") should always map perfectly onto native APIs, and it doesn't.

      There are several places for example where an int key is passed to pthread_getspecific and pthread_setspecific (http://pubs.opengroup.org/onlinepubs/009695399/functions/pthread_getspecific.html). On Linux the compiler happens to allow this because pthread_key_t is defined as "unsigned int" So yeah there's an implicit cast, but since it counts up from zero it usually works. Likewise TlsAlloc on Windows expects the key to be a DWORD but the compiler will accept an int.

      This is really an unsafe assumption though, especially when PyThread_create_key casts the key to an int, and later reuses the (possibly not safely cast) key with PyThread_delete_key/get_key_value/set_key_value). This was brought up at the time, where MvL wrote:

      In principle, this is really difficult to get right. AFAICT,
      pthread_key_t doesn't even have to be an integral type, and even
      if it is, it may well become -1. However, we can probably worry
      about this when a system comes along where this implementation
      breaks.

      One possible workaround without changing the existing API would be this: Each native support wrapper should also provide a *safe* mapping between its native key types and ints, to support the PyThread API.

      For example, the pthread interface could maintain a linked list or an even an array of pthread_key_t pointers, and use the int "key" as the index into that list. If I understand correctly this should be basically harmless since the same key (and hence key -> native-key mapping) can be shared across threads.

    5. embray commented on Aug 23, 2016

      @embray
      Contributor

      (Of course, maintaining such a list might take some care, but only when creating and deleting keys--it wouldn't add any overhead to using them to get/set values.)

    6. bitdancer commented on Aug 23, 2016

      @bitdancer
      Member

      I'd say that sounds reasonable, but most likely it will only be someone working with one of the impacted platforms who will have the motivation to come up with a patch. Especially since neither of the impacted platforms is current formally supported (meaning, we don't yet have anyone on the core team working with those platforms).

      We have, it seems, arrived at the time that MvL foresaw.... :)

    7. embray commented on Aug 23, 2016

      @embray
      Contributor

      The good news about this (in the pthread case) is that it does not need to be seen as some workaround for unusual platforms, but rather making the existing code more POSIX-compliant (and hence correct).

      The Win32 side I'm a little less worried about because the TLS key is documented to be a DWORD, and so the existing casting to int should still work fine most of the time. However, the docs for TlsAlloc() (https://msdn.microsoft.com/en-us/library/windows/desktop/ms686801(v=vs.85).aspx) do state:

      The value of the TLS index should be treated as an opaque value; do not assume that it is an index into a zero-based array.

      which again makes bpo-22206 suspect :(

    8. embray commented on Aug 23, 2016

      @embray
      Contributor

      FWIW I've created an initial patch for this. Seems to work fine, but it's a bit of a mess and I have a few small implementation concerns. I'll try to get it cleaned up sometime in the next few days and I'll post it.

    9. embray commented on Aug 30, 2016

      @embray
      Contributor

      Here's a first stab at a patch for this. A linked list is maintained which maps pthread_key_t instances to an int "key_id" that is used for the PyThread API. Each function needs to map a given key_id to the actual pthread_key_t. This adds a little bit of overhead of course, but in typical usage I don't think there are many entries in this list to loop over.

      This also reverts the change from bpo-22206 which is no longer relevant. No synchronization is provided for PyThread_create_key and PyThread_delete_key, but in typical usage that would be the user's responsibility anyways. Otherwise every PyThread_*_key function would have to have a mutex around it.

    10. vstinner commented on Aug 30, 2016

      @vstinner
      Member

      -1 for your change bpo-25658-1.patch.

      It adds a O(n) slowdown to PyThread_set_key_value() whereas the performance of this function matters. Moreover, the code currently works fine on Linux, I fail to see why we should make Python slower to support a new platform.

      The PEP-11 explains the criteria to support a new platform in Python. Before we start to officially support CloudABI, you should start to work on a fork. When you will get enough users and have a better view of all requires changes, you can come back with a full plan to support officially this platform.

    11. embray commented on Aug 31, 2016

      @embray
      Contributor

      Hi,

      First of all, I should be clear this is not just about CloudABI. In fact I don't even know what CloudABI is. I'm personally working toward getting a working/stable buildbot for Cygwin, so that I can move Python toward supporting it again. Cygwin has not been officially unsupported, but support for it is broken due to lack of a buildbot and lack of a maintainer, both of which I'm willing to offer. I can't get a stable buildbot though until some issues with the build are resolved, and this is a major blocker. If you prefer I can also work on a more formal plan to clarify Python's support for Cygwin which is currently in limbo.

      Not saying a solution to this needs to be merged ASAP as I can work on a branch, but in the meantime it's nice to have an initial fix for the issue so that it can be worked around--I'm sure other platforms that have this issue, such as the CloudABI folks, will appreciate that as well.

      Second of all, the point must be made that I'm not suggesting this be changed just in order to support some specific platform. The fact remains that the existing code is misusing the pthread API and is fragile to begin with. "It works on Linux" is not an excuse--not only does it suggest a bias, but the fact that it even works on Linux is an accident of implementation. The implementation of the pthread_key_t type could change tomorrow, and it would be Python's fault if that breaks things (not that I think that's at all likely to happen, though I'd be less surprised if OSX suddenly changed :)

      As for the performance I agree that PyThread_(get|set)key_value should be fast. This may be O(n) but how big, typically, is n? In a normal run of the Python interpreter n=1, and the autoTLSkey is always the first in the list. It is only larger if some third-party code is using the PyThread API for TLS (and in most typical uses this will only be a few keys at the most, though I couldn't even find any examples in the wild where third-party code is using these functions).

      n could also grow on forks, or manual Py_Finialize/Py_Initialize. This patch didn't implement PyThread_ReInitTLS but it could, in which case it would reset next_key_id to 0 for each child process. Likewise, PyThread_ReInitTLS could also be called from Py_Initialize, which should be harmless.

      So TL;DR I'm not really convinced by the performance argument. But even if that were an issue other implementations are possible; this was just among the simplest and least wasteful.

    12. EdSchouten commented on Aug 31, 2016

      EdSchoutenmannequin
      MannequinAuthor

      I think that PEP-11 also doesn't rule out changes in this area. For example, consider the paragraph starting with the sentence "This policy does not disqualify supporting other platforms indirectly. [...]"

      Attached is a patch that accomplishes the same with a constant running time for operations on keys by using an array. Allocation of new keys runs in expected constant time by using a randomised allocation pattern.

      If we're not big fans of using randomised access, changing this to a linear scan would already be an improvement: keys are only allocated infrequently, but hopefully accessed quite often.

    13. embray commented on Aug 31, 2016

      @embray
      Contributor

      Ah, I wasn't thinking clearly toward the bottom of my last message. I see now that after a fork, _PyGILState_Reinit calls PyThread_delete_key followed by a new PyThread_create_key--in the current version of my patch that would result in putting the autoTLSkey at the end of the linked list, which is no good. That could be worked around, but...

      Ed's version looks good to me. I had the same idea as an alternative, though was a little concerned with the possibility that the array could grow too large. But as I wrote in my last message that would be an extreme case. And regardless his version will at least maintain constant time, so +1.

    14. ma8ma commented on Nov 10, 2016

      ma8mamannequin
      Mannequin

      Hi, I came from bpo-28656.
      I have read past discussions, And I've understood the pthread_key_t has possible of either integer or pointer. So I think there is a simple solution that replaces key type to intptr_t.
      Thus I wrote a patch, but it maybe need configure check for intptr_t if CPytyhon still be supporting for before C99 compilers.
      I confirmed patch on Cygwin x86. Would you be able to confirm for other platforms?

    15. 17 remaining items

    16. ma8ma commented on Apr 30, 2017

      ma8mamannequin
      Mannequin

      Victor,
      The tracemalloc module is not available on platforms that provides own TLS implementation that uses PyMem_* functions. In the own implementation, it occurs deadlock by recursion call in thread key search. PyOnceVar API handles extendable array by using PyMem_Reallloc, therefore, I think TSS key initialization using the API will raise deadlock likewise.

    17. ncoghlan commented on May 3, 2017

      @ncoghlan
      Contributor

      Noting a design consideration that I only picked up in the latest PR review: when exposed as part of a non-opaque struct, the type used for TSS keys becomes part of Python's ABI, which means the API design in the PEP is going to have to handle making the struct fully opaque when Py_LIMITED_API is defined.

      My proposal for handling that:

      • update all of the API functions to accept Py_tss_t by pointer rather than by value
      • update the Py_tss_t definition to be an opaque struct when Py_LIMITED_API is defined
      • add PyThread_tss_alloc and PyThread_tss_free functions for dynamic allocation of key storage (since static allocation won't be possible in the Py_LIMITED_API case)

      My main rationale for going to that level of effort is that the current thread local storage API is available when Py_LIMITED_API is defined, so its replacement should be available as well. However, exposing NATIVE_TSS_KEY_T as part of the stable ABI would prevent us from switching from int to a platform specific implementation for platforms that start out using the generic emulation.

      If issue bpo-29881 ever became a public API, it would then be useful for managing the lifecycle of dynamically allocated TSS keys in extension modules that restrict themselves to the stable ABI.

    18. ma8ma commented on Jul 14, 2017

      ma8mamannequin
      Mannequin

      Hi, I attempted Nick's proposal and removed unused codes from TLS implementation (bpo-30279, bpo-30832). This change looks good to me (PR 1362).
      As the result, I think ready to be more slim API because the own implementation for TLS was removed. Therefore, additional proposal to two API functions:

      • Omit PyThread_ReInitTSS function
      • Omit PyThread_tss_delete_value function

      These functions have handled memory deallocation to key-value storage on the own implementation, but above said, currently the own implementation is gone. It means:

      1. PyThread_ReInitTSS makes to be no-op for all supported platforms (Windows and pthreads)
      2. likewise PyThread_tss_delete_value just does to empty the storage (i.e. "PyThread_tss_delete_value(key)" makes to be equal to "PyThread_tss_set(key, NULL)").

      We (or at least I) won't get back to the own implementation, therefore, I suggest to omit these functions for new TSS API.

      See proposed updates ma8ma#1

    19. ncoghlan commented on Jul 14, 2017

      @ncoghlan
      Contributor

      Nice :)

      With the legacy code cleanups merged, I'd say the next step would be to update the PEP with the simplified API and the explanation for why the removed functions are no longer needed (i.e. we're making native TSS support a hard dependency for 3.7+).

      I'm also thinking we may want to update PEP-11, akin to the entry for the legacy ASCII-only C locale: https://www.python.org/dev/peps/pep-0011/#legacy-c-locale

      I'll handle the PEP-11 update after PEP-539 is approved, the PEP itself would just need a new Platform Support Changes section like the one in PEP-538: https://www.python.org/dev/peps/pep-0538/#platform-support-changes

    20. ma8ma commented on Jul 19, 2017

      ma8mamannequin
      Mannequin

      Nick and Erik, thank you for the comments! I merged proposal into the PR.

      Well, I'm interested in the hide implementation detail for TSS API (lately, I've read the python-ideas thread "PEP: Hide implementation details in the C API" which Victor opened [1]).

      Present, the draft of new API has given two methods for thread key initialization for the non-limited API (i.e. Py_tss_NEEDS_INIT for statically, PyThread_tss_alloc for dynamic). The static initialization needs implementation detail for thread key, but Py_tss_t is designed as an opaque data type based on the stable ABI and we don't feel like to open the implementation detail to the API client. On the other hand, we'd provide newly thread key (de-)allocation functions for the limited API, the API client is able to get an initialized value without knowing thread key detail. And also the functions can be used on the non-limited API.

      Therefore, I think it makes more sense that all API clients use PyThread_tss_(alloc|free) regardless of the limited API. The reason is which are (1) Py_tss_t behaves as an opaque data type as expected for any API client (cannot read and write directly the fields in any case), (2) the API gets more consistency (just one method for key initialization on the API client).

      TL;DR: I'd suggest to make key type strict, what do you think about below changes?

      [1] https://mail.python.org/pipermail/python-ideas/2017-July/046399.html

    21. ma8ma commented on Jul 19, 2017

      ma8mamannequin
      Mannequin

      oh, I found a mistake.

    22. ma8ma commented on Jul 31, 2017

      ma8mamannequin
      Mannequin

      Since previous comment, I've studied the switch for show/hide implementation detail. As the result, I have understood the Py_BUILD_CORE macro hasn't been generally used for hiding implementation detail (and Py_LIMITED_API does the part).
      Therefore, I withdraw previous proposal for implementation detail, and I'm going to fix PEP a few. After that, I'd like to post to python-dev to know the developer's view of the PEP draft.

    23. ma8ma commented on Sep 9, 2017

      ma8mamannequin
      Mannequin

      FYI, PEP-539 was accepted (see python-dev threads [1] [2]).
      The only thing missing is a reference implementation, next I complete it.

      [1] https://mail.python.org/pipermail/python-dev/2017-August/149091.html
      [2] https://mail.python.org/pipermail/python-dev/2017-September/149354.html

    24. ncoghlan commented on Oct 6, 2017

      @ncoghlan
      Contributor

      This has been merged now: 731e189

      Thank you for the PEP & implementation! :)

    25. scoder commented on Oct 20, 2017

      @scoder
      Contributor

      Would it be possible to define Py_tss_NEEDS_INIT as a constant variable instead of a static initialiser? That would enable its use also for non-static initialisations.

    26. ncoghlan commented on Oct 20, 2017

      @ncoghlan
      Contributor

      Stefan: I'd expect so, but that's best covered by a new RFE and an associated PR (I'm not sure what you mean from the text description, but I assume it will be obvious as C code)

    27. scoder commented on Oct 20, 2017

      @scoder
      Contributor

      It seems that there's a simpler way that uses a cast on the literal. I added a pull request. Looks simple enough to not merit its own ticket, I think.

    28. scoder commented on Oct 20, 2017

      @scoder
      Contributor

      Seems like this isn't trivial, so I created a new ticket for this. See bpo-31828.

    29. transferred this issue fromon Apr 10, 2022
    Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

    Metadata

    Metadata

    Assignees

    No one assigned

      Labels

      3.7 (EOL)end of lifebuildThe build process and cross-buildinterpreter-core(Objects, Python, Grammar, and Parser dirs)

      Projects

      No projects

        Milestone

        No milestone

        Relationships

        None yet

        Development

        No branches or pull requests

        Issue actions