Error scope constructor causes access violation in ~error_already_set()

[AndreasSchoenle]

When using error_already_set() in code that is called from a C++ program, the error_scope constructor crashes due to non-existent thread state. The C++ Code acquires the thread state through PyGILState_Ensure(), when throwing the error_already_set() PyGILState_Release() is called during thread exit.

It seems to me that when the destructor is called, the crash occurs because the current thread does not hold the GIL.

About the destructor:

pybind11/include/pybind11/pybind11.h

Line 1991 in e2b884c

error_scope scope;

Could it be that the two calls:

error_scope scope;
gil_scoped_acquire gil;

have to be reversed? Also, why is error_scope needed - as I understand we just need the GIL to decrease the ref count on the held error objects?

[eacousineau]

Hm... Seems like it might be that. Any chance you have a simple minimal reproduction, where you can toggle the ordering / remove error_scope to see if that fixes it?

[wjakob]

I've just run into this myself, and the fix that you suggested looks good to me.
The error_scope is needed for module initialization, see #1362.