← Back
Editing: test_leaks.py
# -*- coding: utf-8 -*- """ Testing scenarios that may have leaked. """ import sys import gc import time import weakref import threading import greenlet from . import TestCase from . import PY314 from . import RUNNING_ON_FREETHREAD_BUILD from . import WIN from .leakcheck import fails_leakcheck from .leakcheck import ignores_leakcheck from .leakcheck import RUNNING_ON_MANYLINUX # pylint:disable=protected-access assert greenlet.GREENLET_USE_GC # Option to disable this was removed in 1.0 class HasFinalizerTracksInstances(object): EXTANT_INSTANCES = set() def __init__(self, msg): self.msg = sys.intern(msg) self.EXTANT_INSTANCES.add(id(self)) def __del__(self): self.EXTANT_INSTANCES.remove(id(self)) def __repr__(self): return "<HasFinalizerTracksInstances at 0x%x %r>" % ( id(self), self.msg ) @classmethod def reset(cls): cls.EXTANT_INSTANCES.clear() def fails_leakcheck_except_on_free_thraded(func): if RUNNING_ON_FREETHREAD_BUILD: # These all seem to pass on free threading because # of the changes to the garbage collector return func return fails_leakcheck(func) class TestLeaks(TestCase): def test_arg_refs(self): args = ('a', 'b', 'c') refcount_before = sys.getrefcount(args) # pylint:disable=unnecessary-lambda g = greenlet.greenlet( lambda *args: greenlet.getcurrent().parent.switch(*args)) for _ in range(100): g.switch(*args) self.assertEqual(sys.getrefcount(args), refcount_before) def test_kwarg_refs(self): kwargs = {} self.assertEqual(sys.getrefcount(kwargs), 2 if not PY314 else 1) # pylint:disable=unnecessary-lambda g = greenlet.greenlet( lambda **gkwargs: greenlet.getcurrent().parent.switch(**gkwargs)) for _ in range(100): g.switch(**kwargs) # Python 3.14 elides reference counting operations # in some cases. See https://github.com/python/cpython/pull/130708 self.assertEqual(sys.getrefcount(kwargs), 2 if not PY314 else 1) @staticmethod def __recycle_threads(): # By introducing a thread that does sleep we allow other threads, # that have triggered their __block condition, but did not have a # chance to deallocate their thread state yet, to finally do so. # The way it works is by requiring a GIL switch (different thread), # which does a GIL release (sleep), which might do a GIL switch # to finished threads and allow them to clean up. def worker(): time.sleep(0.001) t = threading.Thread(target=worker) t.start() time.sleep(0.001) t.join(10) def test_threaded_leak(self): gg = [] def worker(): # only main greenlet present gg.append(weakref.ref(greenlet.getcurrent())) for _ in range(2): t = threading.Thread(target=worker) t.start() t.join(10) del t greenlet.getcurrent() # update ts_current self.__recycle_threads() greenlet.getcurrent() # update ts_current gc.collect() greenlet.getcurrent() # update ts_current for g in gg: self.assertIsNone(g()) def test_threaded_adv_leak(self): gg = [] def worker(): # main and additional *finished* greenlets ll = greenlet.getcurrent().ll = [] def additional(): ll.append(greenlet.getcurrent()) for _ in range(2): greenlet.greenlet(additional).switch() gg.append(weakref.ref(greenlet.getcurrent())) for _ in range(2): t = threading.Thread(target=worker) t.start() t.join(10) del t greenlet.getcurrent() # update ts_current self.__recycle_threads() greenlet.getcurrent() # update ts_current gc.collect() greenlet.getcurrent() # update ts_current for g in gg: self.assertIsNone(g()) def assertClocksUsed(self): used = greenlet._greenlet.get_clocks_used_doing_optional_cleanup() self.assertGreaterEqual(used, 0) # we don't lose the value greenlet._greenlet.enable_optional_cleanup(True) used2 = greenlet._greenlet.get_clocks_used_doing_optional_cleanup() self.assertEqual(used, used2) self.assertGreater(greenlet._greenlet.CLOCKS_PER_SEC, 1) def _check_issue251(self, manually_collect_background=True, explicit_reference_to_switch=False): # See https://github.com/python-greenlet/greenlet/issues/251 # Killing a greenlet (probably not the main one) # in one thread from another thread would # result in leaking a list (the ts_delkey list). # We no longer use lists to hold that stuff, though. # For the test to be valid, even empty lists have to be tracked by the # GC assert gc.is_tracked([]) HasFinalizerTracksInstances.reset() greenlet.getcurrent() greenlets_before = self.count_objects(greenlet.greenlet, exact_kind=False) background_glet_running = threading.Event() background_glet_killed = threading.Event() background_greenlets = [] # XXX: Switching this to a greenlet subclass that overrides # run results in all callers failing the leaktest; that # greenlet instance is leaked. There's a bound method for # run() living on the stack of the greenlet in g_initialstub, # and since we don't manually switch back to the background # greenlet to let it "fall off the end" and exit the # g_initialstub function, it never gets cleaned up. Making the # garbage collector aware of this bound method (making it an # attribute of the greenlet structure and traversing into it) # doesn't help, for some reason. def background_greenlet(): # Throw control back to the main greenlet. jd = HasFinalizerTracksInstances("DELETING STACK OBJECT") greenlet._greenlet.set_thread_local( 'test_leaks_key', HasFinalizerTracksInstances("DELETING THREAD STATE")) # Explicitly keeping 'switch' in a local variable # breaks this test in all versions if explicit_reference_to_switch: s = greenlet.getcurrent().parent.switch s([jd]) else: greenlet.getcurrent().parent.switch([jd]) bg_main_wrefs = [] def background_thread(): glet = greenlet.greenlet(background_greenlet) bg_main_wrefs.append(weakref.ref(glet.parent)) background_greenlets.append(glet) glet.switch() # Be sure it's active. # Control is ours again. del glet # Delete one reference from the thread it runs in. background_glet_running.set() background_glet_killed.wait(10) # To trigger the background collection of the dead # greenlet, thus clearing out the contents of the list, we # need to run some APIs. See issue 252. if manually_collect_background: greenlet.getcurrent() t = threading.Thread(target=background_thread) t.start() background_glet_running.wait(10) greenlet.getcurrent() lists_before = self.count_objects(list, exact_kind=True) assert len(background_greenlets) == 1 self.assertFalse(background_greenlets[0].dead) # Delete the last reference to the background greenlet # from a different thread. This puts it in the background thread's # ts_delkey list. del background_greenlets[:] background_glet_killed.set() # Now wait for the background thread to die. t.join(10) del t # As part of the fix for 252, we need to cycle the ceval.c # interpreter loop to be sure it has had a chance to process # the pending call. self.wait_for_pending_cleanups() lists_after = self.count_objects(list, exact_kind=True) greenlets_after = self.count_objects(greenlet.greenlet, exact_kind=False) # On 2.7, we observe that lists_after is smaller than # lists_before. No idea what lists got cleaned up. All the # Python 3 versions match exactly. self.assertLessEqual(lists_after, lists_before) # On versions after 3.6, we've successfully cleaned up the # greenlet references thanks to the internal "vectorcall" # protocol; prior to that, there is a reference path through # the ``greenlet.switch`` method still on the stack that we # can't reach to clean up. The C code goes through terrific # lengths to clean that up. if not explicit_reference_to_switch \ and greenlet._greenlet.get_clocks_used_doing_optional_cleanup() is not None: # If cleanup was disabled, though, we may not find it. self.assertEqual(greenlets_after, greenlets_before) if manually_collect_background: # TODO: Figure out how to make this work! # The one on the stack is still leaking somehow # in the non-manually-collect state. self.assertEqual(HasFinalizerTracksInstances.EXTANT_INSTANCES, set()) else: # The explicit reference prevents us from collecting it # and it isn't always found by the GC either for some # reason. The entire frame is leaked somehow, on some # platforms (e.g., MacPorts builds of Python (all # versions!)), but not on other platforms (the linux and # windows builds on GitHub actions and Appveyor). So we'd # like to write a test that proves that the main greenlet # sticks around, and we can on my machine (macOS 11.6, # MacPorts builds of everything) but we can't write that # same test on other platforms. However, hopefully iteration # done by leakcheck will find it. pass if greenlet._greenlet.get_clocks_used_doing_optional_cleanup() is not None: self.assertClocksUsed() def test_issue251_killing_cross_thread_leaks_list(self): self._check_issue251() def test_issue251_with_cleanup_disabled(self): greenlet._greenlet.enable_optional_cleanup(False) try: self._check_issue251() finally: greenlet._greenlet.enable_optional_cleanup(True) @fails_leakcheck_except_on_free_thraded def test_issue251_issue252_need_to_collect_in_background(self): # Between greenlet 1.1.2 and the next version, this was still # failing because the leak of the list still exists when we # don't call a greenlet API before exiting the thread. The # proximate cause is that neither of the two greenlets from # the background thread are actually being destroyed, even # though the GC is in fact visiting both objects. It's not # clear where that leak is? For some reason the thread-local # dict holding it isn't being cleaned up. # # The leak, I think, is in the CPYthon internal function that # calls into green_switch(). The argument tuple is still on # the C stack somewhere and can't be reached? That doesn't # make sense, because the tuple should be collectable when # this object goes away. # # Note that this test sometimes spuriously passes on Linux, # for some reason, but I've never seen it pass on macOS. self._check_issue251(manually_collect_background=False) @fails_leakcheck_except_on_free_thraded def test_issue251_issue252_need_to_collect_in_background_cleanup_disabled(self): self.expect_greenlet_leak = True greenlet._greenlet.enable_optional_cleanup(False) try: self._check_issue251(manually_collect_background=False) finally: greenlet._greenlet.enable_optional_cleanup(True) @fails_leakcheck_except_on_free_thraded def test_issue251_issue252_explicit_reference_not_collectable(self): self._check_issue251( manually_collect_background=False, explicit_reference_to_switch=True) UNTRACK_ATTEMPTS = 100 def _only_test_some_versions(self): # We're only looking for this problem specifically on 3.11, # and this set of tests is relatively fragile, depending on # OS and memory management details. So we want to run it on 3.11+ # (obviously) but not every older 3.x version in order to reduce # false negatives. At the moment, those false results seem to have # resolved, so we are actually running this on 3.8+ assert sys.version_info[0] >= 3 if RUNNING_ON_MANYLINUX: self.skipTest("Slow and not worth repeating here") @ignores_leakcheck # Because we're just trying to track raw memory, not objects, and running # the leakcheck makes an already slow test slower. def test_untracked_memory_doesnt_increase(self): # See https://github.com/gevent/gevent/issues/1924 # and https://github.com/python-greenlet/greenlet/issues/328 self._only_test_some_versions() def f(): return 1 ITER = 10000 def run_it(): for _ in range(ITER): greenlet.greenlet(f).switch() # Establish baseline for _ in range(3): run_it() # uss: (Linux, macOS, Windows): aka "Unique Set Size", this is # the memory which is unique to a process and which would be # freed if the process was terminated right now. uss_before = self.get_process_uss() for count in range(self.UNTRACK_ATTEMPTS): uss_before = max(uss_before, self.get_process_uss()) run_it() uss_after = self.get_process_uss() if uss_after <= uss_before and count > 1: break self.assertLessEqual(uss_after, uss_before) def _check_untracked_memory_thread(self, deallocate_in_thread=True): self._only_test_some_versions() # Like the above test, but what if there are a bunch of # unfinished greenlets in a thread that dies? # Does it matter if we deallocate in the thread or not? # First, make sure we can get useful measurements. This will # be skipped if not. self.get_process_uss() EXIT_COUNT = [0] def f(): try: greenlet.getcurrent().parent.switch() except greenlet.GreenletExit: EXIT_COUNT[0] += 1 raise return 1 ITER = 10000 def run_it(): glets = [] for _ in range(ITER): # Greenlet starts, switches back to us. # We keep a strong reference to the greenlet though so it doesn't # get a GreenletExit exception. g = greenlet.greenlet(f) glets.append(g) g.switch() return glets test = self class ThreadFunc: uss_before = uss_after = 0 glets = () ITER = 2 def __call__(self): self.uss_before = test.get_process_uss() for _ in range(self.ITER): self.glets += tuple(run_it()) for g in self.glets: test.assertIn('suspended active', str(g)) # Drop them. if deallocate_in_thread: self.glets = () self.uss_after = test.get_process_uss() # Establish baseline uss_before = uss_after = None for count in range(self.UNTRACK_ATTEMPTS): EXIT_COUNT[0] = 0 thread_func = ThreadFunc() t = threading.Thread(target=thread_func) t.start() t.join(30) self.assertFalse(t.is_alive()) if uss_before is None: uss_before = thread_func.uss_before uss_before = max(uss_before, thread_func.uss_before) if deallocate_in_thread: self.assertEqual(thread_func.glets, ()) self.assertEqual(EXIT_COUNT[0], ITER * thread_func.ITER) del thread_func # Deallocate the greenlets; but this won't raise into them del t if not deallocate_in_thread: self.assertEqual(EXIT_COUNT[0], 0) if deallocate_in_thread: self.wait_for_pending_cleanups() uss_after = self.get_process_uss() # See if we achieve a non-growth state at some point. Break when we do. if uss_after <= uss_before and count > 1: break self.wait_for_pending_cleanups() uss_after = self.get_process_uss() # On Windows, USS can fluctuate by tens of KB between measurements # due to working set trimming, page table updates, etc. Allow a # small tolerance so OS-level noise doesn't cause false failures. # Real leaks produce MBs of growth (each iteration creates 20k # greenlets), so 512 KB is well below the detection threshold for # genuine issues. tolerance = 512 * 1024 if WIN else 0 self.assertLessEqual(uss_after, uss_before + tolerance, "after attempts %d" % (count,)) @ignores_leakcheck # Because we're just trying to track raw memory, not objects, and running # the leakcheck makes an already slow test slower. def test_untracked_memory_doesnt_increase_unfinished_thread_dealloc_in_thread(self): self._check_untracked_memory_thread(deallocate_in_thread=True) @ignores_leakcheck # Because the main greenlets from the background threads do not exit in a timely fashion, # we fail the object-based leakchecks. def test_untracked_memory_doesnt_increase_unfinished_thread_dealloc_in_main(self): # Between Feb 10 and Feb 20 2026, this test started failing on # Github Actions, windows 3.14t. With no relevant code changes on # our part. Both versions were 3.14.3 (same build). The only change # is the Github actions "Runner Image". The working one was version # 20260202.17.1, while the updated failing version was # 20260217.31.1. Both report the same version of the operating system # (Microsoft Windows Server 2025 10.0.26100). # # Reevaluate on future runner image releases. if WIN and RUNNING_ON_FREETHREAD_BUILD and PY314: self.skipTest("Windows 3.14t appears to leak. No other platform does.") self._check_untracked_memory_thread(deallocate_in_thread=False) if __name__ == '__main__': __import__('unittest').main()
Save File
Cancel