Logged In: YES user_id=1038590 The patch doesn't apply cleanly to current CVS (no actual conflict that I can see - just the patch machinery getting confused by a cautionary comment added to the top of pystate.c) Inspecting the patch manually, it looks like it does the right thing, given that Python instructs that Py_Initialise and PyEval_InitThreads *both* be called prior while the application is still single-threaded. However, _PyThread_Started seems like an odd bit of code - prior to this patch, it is set (initialised to 0 in pythonrun.c, set to 1 by PyEval_InitThreads), but never referenced anywhere (despite the comment saying that it is a "Flag for Py_Exit"). It's also not part of any API (even an internal one). It is defined in pythonrun.c, then declared as extern in ceval.c. The patch adds another extern declaration in pystate.c It would seem better (and only slightly more work) to add a real API (such as adding PyThread_IsInitialized() to pythread.h), rather than entrenching further use of _PyThread_Started.
Logged In: YES user_id=1038590 Patch #1044089 has been created with an alternate implementation of the same concept. Rather than using _PyThread_Started, it: - uses interpreter_lock within ceval.c, which matches the existing ceval.c code that uses the existence of the GIL to determine if PyEval_InitThreads has been called - adds a new API function PyEval_ThreadsInitialized which allows external code to check if the GIL exists or not - uses this new API in the PyGILState_* functions The same functions get modified as in the patch attached here - the only difference is in how the code asks the question "Has PyEval_InitThreads been called?"
Logged In: YES user_id=31435 Sorry, I'm opposed to this patch. Slowing every lock operation in the core by an additional test+branch makes everyone pay for what you want, and muddies semantics that haven't changed since 1991 . If you need to do something here, it would be better to go back to your first python-dev msg's approach. Note that the core never calls lock operations unless ceval.c's interperter_lock is non-NULL. That's how *it* avoids the expense of lock fiddling so long as only one thread is in use. It's not a goal to make that easy to spell (we're coding in C here, not Python). It may be a goal to expose enough machinery so that motivated extension module authors can endure the same kind of pain.
