homepage

_PyGILState_Init() initialize autoInterpreterState variable. This variable have to be set before the first call to PyGILState_Ensure().
The problem is that _PyGILState_Init() is called late: at the end of Py_InitializeEx(). It's called after initsite(), whereas initsite() may need to call _PyGILState_Init().
Example: add the following lines at the end of site.py:
---
import readline
import rlcompleter
readline.parse_and_bind("tab: complete")
raw_input("press TAB")
---
Run an interpreter and press TAB:
---
$ ./python
press TABpython: Python/pystate.c:595: PyGILState_Ensure: Assertion `autoInterpreterState' failed.
Abandon
---
Other example of functiions using _PyGILState_Init(): _PyObject_Dump() (useful for debug), sqlite module, os.popen*(), ctypes Python callbacks, etc.
I don't know the right place for _PyGILState_Init() in Py_InitializeEx(). _PyGILState_Init() calls the following functions:
 - PyThread_get_thread_ident()
 - PyThread_allocate_lock()
 - PyThread_acquire_lock()
 - PyThread_release_lock()
 - Py_FatalError() (on error)
None of these function use Python functions, so _PyGILState_Init() can be called very early. The earliest position is just after the call to PyThreadState_New(), before PyThreadState_Swap(). It tested it and it works correctly.
If _PyGILState_Init() is called before PyThreadState_Swap(), PyThreadState_Swap() can be simplified (it doesn't need to check if GIL is initialized or not). Attached patch implement that.
--
I hit this bug when I was debuging Python: I added a call to _PyObject_Dump() which stopped Python (assertion error) because the GIL was not initialized :-/
I already hitted this bug some weeks ago when I was working on the thread state preallocation when creating a new thread: #7544.

haypo> I already hitted this bug some weeks ago when I was working 
haypo> on the thread state preallocation when creating a new thread:
haypo> #7544.
According to msg97715: it was indirectly the same issue, call _PyObject_Dump() while the GIL is not initialized.

Bootstrap issues are always quite delicate, and I'd suggest being careful in this case.

See also #3137.

I would keep the call to PyThreadState_Swap() next to PyThreadState_New(): create the thread state and install it. _PyGILState_Init() may come after.

haypo, it seems you removed the initial message...

> haypo, it seems you removed the initial message...
I know... My mouse clicked on [remove] button, it wasn't me! I don't know how to restore it. The message: msg100432.

I found the following issue in Roundup tracker to restore a deleted message:
http://psf.upfronthosting.co.za/roundup/meta/issue267
I tried to write the URL, but it doesn't work:
http://bugs.python.org/issue8063?@action=edit&@add@messages=100432
I tried also using Poster Firefox extension (to post a POST request, not a GET request), but it doesn't work. My URL should be wrong :-/

My SIGINT.patch for #3137 moves the call to _PyGILState_Init() just before initsite(), so it doesn't change too much Py_InitializeEx() and it's enough for me :-)

While working on #9425, I usually hit two annoying issues:
 - _PyObject_Dump() crashs (assertion error) if I call it (with gdb) in Py_InitializeEx()
 - because of python-gdb.py, gdb does segfault (I don't know yet where it does come from)
So I'm back on the GIL topic: I still would like to initialize the GIL earlier in Py_InitializeEx(). As Amaury wrote, I think that the right place is just after "(void) PyThreadState_Swap(tstate);". This is exactly what does my new patch (for py3k).
I think that only python 3.2 should be patched.

Commited as r84163 to 3.2. Don't backport because it is not really a bug and I prefer to avoid touching stable branches with such minor details.