> Wouldn't it be better to expose and re-use the HEAD_LOCK and HEAD_UNLOCK
> macros from pystate.c?

I don't like holding locks before calling "alien" code, it's a recipe
for deadlocks: for example, if another thread-local object was
deallocated as part of the deallocation chain, we would call back into
local_clear(), and deadlock.

> That said, I doubt this is the issue here. We are removing a string key pointing
> to a localdummy object. Both are small atomic types not handled by the GC, so
> I don't see how deallocating these objects could release the GIL.

Yes, it shouldn't happen, the thread local dict is deallocated right
before (I initially thought the thread local dict was deallocated
here).

Without a proper backtrace, i'ts going to be hard to debug...