In PyThreadState_Clear() and PyInterpreterState_Clear() the code expects a "current" thread state (i.e. set by PyThreadState_Swap()) to be set. Presumably this is for the following:

• to ensure the GIL is held
• to use for logging audit events (actually only needs the current interpreter?)
• to use in object finalizers for objects held by the target thread/interpreter state

For PyThreadState_Clear(), this is a problem when the current thread state is also the target thread state, since the function is progressively rendering the thread state unusable. For PyInterpreterState_Clear() the "current" thread state must belong to the target interpreter, but will be rendered unusable before it's potentially needed.

(FWIW, it isn't clear to me yet why we don't see frequent crashes from this already (at least when we're clearing the main interpreter during finalization). I do suspect that this contributes to some of the occasional CI failures we see relative to finalization.)

Ideally auditing and the various object finalizers could work without having a current thread state, especially in the case where the current thread state could only possibly be the one getting cleared. Could we split up PyThreadState so at least the problematic part is guaranteed not be used in PyThreadState_Clear() or PyInterpreterState_Clear()? Are there other complexities we have to deal with?

(Note that we'd still require that the GIL be held by the current OS thread when PyThreadState_Clear() or PyInterpreterState_Clear() are called.)