assert(PyGILState_Check()) cannot be used in functions called by Py_NewInterpreter(): the assertion fails.

I propose to add a flag to disable PyGILState_Check() while Py_NewInterpreter() is run to be able to add assertions in these functions.

The assertion helps to detect complex bugs like race condition in multithreaded applications.

gil_check.patch: add more checks on the GIL

* PyGILState_Check() now returns 1 (success) before the creation of the GIL and
  after the destruction of the GIL
* Add a flag to disable PyGILState_Check(). Disable PyGILState_Check() when
  Py_NewInterpreter() is called
* Add assert(PyGILState_Check()) to:

  - _Py_dup()
  - _Py_fstat()
  - _Py_read()
  - _Py_write()
  - PyObject_Malloc()
  - PyObject_Calloc()
  - PyObject_Realloc()
  - PyObject_Free()

I created this issue while working on the issue #26249 "Change PyMem_Malloc to use pymalloc allocator". I would like to check that application call PyObject_Malloc() and PyMem_Malloc() with the GIL held.

Hum, in fact I prefer to implement GIL check in debug hooks of Python memory allocators. So it becomes to get these checks on Python compiled in debug mode.

Patch 2: Don't touch Python memory allocators.

Instead of I wrote a second patch, pymem_gil_check.patch (based on gil_check-2.patch).

--

Oh, I forgot to mention that the gil_check patch changes PyThreadState_DeleteCurrent(): I reordered instructions to be able call PyObject_Free() with the GIL check.

New changeset c24630367767 by Victor Stinner in branch 'default':
Fix Py_FatalError() if called without the GIL
https://hg.python.org/cpython/rev/c24630367767

New changeset 8a5a0e223559 by Victor Stinner in branch 'default':
Issue #26558: Remove useless check in tracemalloc
https://hg.python.org/cpython/rev/8a5a0e223559

See also issues:

* #10915: Make the PyGILState API compatible with multiple interpreters 
* #15751: Support subinterpreters in the GIL state API

This issue contains is a compromise: it's a workaround until the root issue is solve. Since #10915 is open since 2011, I'm not sure that it's possible to fix the issue, nor that it's worth :-) To me, subinterpreters still looks like an experimental feature.

New changeset e590c632c9fa by Victor Stinner in branch 'default':
Add more checks on the GIL
https://hg.python.org/cpython/rev/e590c632c9fa

New changeset cde3d986da00 by Victor Stinner in branch 'default':
Check the GIL in PyObject_Malloc()
https://hg.python.org/cpython/rev/cde3d986da00

Ok, it looks like buildbots are happy. I close the issue.

New changeset b394fc71f92a by Victor Stinner in branch '3.5':
Fix Py_FatalError() if called without the GIL
https://hg.python.org/cpython/rev/b394fc71f92a

New changeset c298c6d8b324 by Victor Stinner in branch '3.5':
faulthandler: Test Py_FatalError() with GIL released
https://hg.python.org/cpython/rev/c298c6d8b324

New changeset 3a228ab17c2a9cffd1a2f15f30d6209768de20a6 by Victor Stinner in branch 'master':
bpo-26558: Fix Py_FatalError() with GIL released (GH-10267)
https://github.com/python/cpython/commit/3a228ab17c2a9cffd1a2f15f30d6209768de20a6

New changeset 192c54713b4c67a8d14b2d98056771feba40ca37 by Miss Islington (bot) in branch '3.7':
bpo-26558: Fix Py_FatalError() with GIL released (GH-10267)
https://github.com/python/cpython/commit/192c54713b4c67a8d14b2d98056771feba40ca37

New changeset ff56597151206a03ce421712516323430b7749c8 by Victor Stinner in branch '3.6':
bpo-26558: Fix Py_FatalError() with GIL released (GH-10267) (GH-10270)
https://github.com/python/cpython/commit/ff56597151206a03ce421712516323430b7749c8