`PyThreadState` and `PyInterpreterState` are not freed

[zzzeek]

The following test program gains about 10M per second under top, on Python 3.11 only (confirmed for all development stages: 3.11.0a4, 3.11.0b1, 3.11.0rc1, 3.11.0) . 3.10 shows no memory growth.

from threading import Thread

while True:
    t1 = Thread()
    t1.start()

For direct results, here's the same program using the resource module that I just saw used over at #98467:

from threading import Thread
import resource

print("Before", resource.getrusage(resource.RUSAGE_SELF).ru_maxrss / 1024, "MB")

i = 0

while True:
    i += 1
    t1 = Thread()
    t1.start()
    if i % 10000 == 0:
        print(
            f"memory after {i} iterations:",
            resource.getrusage(resource.RUSAGE_SELF).ru_maxrss / 1024,
            "MB",
        )

The above program under Py 3.11.0 prints:

$ .venv-3.11/bin/python test5.py 
Before 8.14453125 MB
memory after 10000 iterations: 11.6875 MB
memory after 20000 iterations: 15.30078125 MB
memory after 30000 iterations: 18.65234375 MB
memory after 40000 iterations: 22.26171875 MB
memory after 50000 iterations: 25.87109375 MB
memory after 60000 iterations: 29.22265625 MB
memory after 70000 iterations: 32.83203125 MB
memory after 80000 iterations: 36.44140625 MB
memory after 90000 iterations: 39.78125 MB
memory after 100000 iterations: 43.390625 MB
...

under Python 3.10 it prints:

$ .venv-3.10/bin/python test5.py 
Before 9.25 MB
memory after 10000 iterations: 9.46484375 MB
memory after 20000 iterations: 9.47265625 MB
memory after 30000 iterations: 9.47265625 MB
memory after 40000 iterations: 9.47265625 MB
memory after 50000 iterations: 9.47265625 MB
memory after 60000 iterations: 9.47265625 MB
memory after 70000 iterations: 9.47265625 MB
memory after 80000 iterations: 9.47265625 MB

the issue looks extremely similar to another one we just fixed in greenlet, over at python-greenlet/greenlet#328, although this one is much more surprising. Issue #98467 might also be related.

Environment:

$ uname -a
Linux photon3 5.17.14-200.fc35.x86_64 #1 SMP PREEMPT Thu Jun 9 14:02:42 UTC 2022 x86_64 x86_64 x86_64 GNU/Linux

$ python3
Python 3.11.0 (main, Nov  5 2022, 23:27:22) [GCC 11.3.1 20220421 (Red Hat 11.3.1-3)] on linux

• PR: GH-99205: Mark new interpreters and threads as non-static #99268

• PR: [3.11] GH-99205: Mark new interpreters and threads as non-static (GH-99268) #99301

• PR: GH-99205: remove _static field from PyThreadState and PyInterpreterState #99385

[kumaraditya303]

Seems related to the 3.11 new frame handling.

cc @markshannon @brandtbucher

[brandtbucher]

I can reproduce this.

At first glance, I'm not certain that its related to the frame stuff: debugging seems to confirm that the allocations and deallocations of stack chunks are balanced, and the issue is present even when changing which allocator we use, etc.

Plus, the leaked memory per thread (~370 bytes) seems too small to be our 16KB chunks. It's almost exactly sizeof(PyThreadState) (360 bytes), though! So my hunch is that we're leaking PyThreadState structs.

[brandtbucher]

Found it... we don't free the memory if the thread state's _static member is nonzero:

cpython/Python/pystate.c

Lines 754 to 760 in 2eee9d9

	static void
	free_threadstate(PyThreadState *tstate)
	{
	if (!tstate->_static) {
	PyMem_RawFree(tstate);
	}
	}

It appears we're not setting it correctly. We initialize each thread by memcpy'ing the initial thread, which is statically allocated and has the member set to 1:

cpython/Python/pystate.c

Lines 847 to 850 in 2eee9d9

	// Set to _PyThreadState_INIT.
	memcpy(tstate,
	&initial._main_interpreter._initial_thread,
	sizeof(*tstate));

cpython/Include/internal/pycore_runtime_init.h

Lines 65 to 70 in 2eee9d9

	._initial_thread = _PyThreadState_INIT, \
	}
	#define _PyThreadState_INIT \
	{ \
	._static = 1, \

So I think the fix is just to make sure that new threads have tstate->_static initialized to 0.

@ericsnowcurrently: is my understanding correct? I can write a patch, if so.

[brandtbucher]

Also, it looks like PyInterpreterState has a similar leak.

[ericsnowcurrently]

I want to agree, but something seems off. I'll have to take a look tomorrow.

[brandtbucher]

Confirmed that setting tstate->_static = false; after the memcpy fixes the issue.

[gpshead]

It might be useful to define APIs to copy a PyInterpreterState and PyThreadStats where all of this can be encapsulated? granted there seems to be only one place in the code that is doing this right now so...