`threading.local()` implementation is not thread-safe in the free-threaded build

[colesbury]

Bug report

The implementation of Python thread local variables (threading.local() or _thread._local) has some thread-safety issues. The issues are particularly relevant to the free-threaded build, but some can affect the default build too.

local_clear loop over threads isn't safe

The local_clear() function is called when a thread local variable is destroyed. It loops over all threads in the interpreter and removes itself from their dictionaries.

https://github.com/python/cpython/blob/fd0f814ade43fa479bfbe76dc226424db14a9354/Modules/_threadmodule.c#L1568-L1581

This isn't thread-safe because after HEAD_UNLOCK(runtime), the stored tstate might be deleted concurrently. This can happen even in the default build with the GIL enabled because PyThreadState_Delete() doesn't require the GIL to be held. However, it's less likely to occur in practice because threading module created threads hold onto the GIL until they're deleted.

local_clear access to tstate->dict isn't thread-safe

In the free-threaded build, local_clear() may be run concurrently with some other thread's PyThreadState_Clear(). The access to another thread's tstate->dict isn't safe because it may be getting destroyed concurrently.

[corona10]

@mpage https://github.com/python/cpython/issues/118490 if the linked issue is related to this issue, please take a look at both of them :)