Repository navigation
shutdown (exit) can hang or segfault with daemon threads running #46164
Description
Activity
This probably applies to 3.0 as well but i have not tested it there.
Here are some sample failures:
=========== A ==============
Exception in thread Thread-0000000001 (most likely raised during
interpreter shutdown):Exception in thread Thread-0000000003 (most likely
raised during interpreter shutdown):Program received signal SIGSEGV, Segmentation fault.
[Switching to Thread 0xe4cf1bb0 (LWP 22281)]
0x080c0824 in PyEval_EvalFrameEx (f=0x820d59c, throwflag=0)
at Python/ceval.c:2633
2633 if (tstate->frame->f_exc_type != NULL)
======================================= B ==============
Exception in thread Thread-0000000001 (most likely raised during interpreter shutdown): Traceback (most recent call last): File "/home/gps/oss/python/trunk/Lib/threading.py", line 486, in __bootstrap_inner File "thread_exit.py", line 24, in run <type 'exceptions.AttributeError'>: 'NoneType' object has no attribute 'add'
Program received signal SIGINT, Interrupt.
[Switching to Thread 0xf7fe08e0 (LWP 21703)]
0xffffe410 in __kernel_vsyscall ()
(gdb) bt
#0 0xffffe410 in __kernel_vsyscall ()
#1 0x4dac1af0 in sem_wait@GLIBC_2.0 () from
/lib/tls/i686/cmov/libpthread.so.0
#2 0xf7fe08c0 in ?? ()
#3 0x080ec9d5 in PyThread_acquire_lock (lock=0x0, waitflag=1)
at Python/thread_pthread.h:349
#4 0x080be6d5 in PyEval_RestoreThread (tstate=0x8168050) at
Python/ceval.c:314
#5 0x0806caf5 in file_dealloc (f=0xf7f29d58) at Objects/fileobject.c:396
#6 0x0810be36 in frame_dealloc (f=0x8201f04) at Objects/frameobject.c:416
#7 0x080e28d2 in PyThreadState_Clear (tstate=0x8184068)
at Python/pystate.c:217
#8 0x080e2994 in PyInterpreterState_Clear (interp=0x8168008)
at Python/pystate.c:105
#9 0x080e473b in Py_Finalize () at Python/pythonrun.c:444
#10 0x080e428e in handle_system_exit () at Python/pythonrun.c:1641
#11 0x080e447a in PyErr_PrintEx (set_sys_last_vars=1)
at Python/pythonrun.c:1087
#12 0x080e510d in PyRun_SimpleFileExFlags (fp=0x817bd98,
filename=0xffffd73c "thread_exit.py", closeit=1, flags=0xffffd518)
at Python/pythonrun.c:996
#13 0x0805753b in Py_Main (argc=1, argv=0xffffd5b4) at Modules/main.c:574
#14 0x4d95cea2 in __libc_start_main () from /lib/tls/i686/cmov/libc.so.6
#15 0x080566c1 in _start () at ../sysdeps/i386/elf/start.S:119
(gdb) bt
#0 0x080c0824 in PyEval_EvalFrameEx (f=0x820d59c, throwflag=0)
at Python/ceval.c:2633
#1 0x080c451c in PyEval_EvalFrameEx (f=0x820d40c, throwflag=0)
at Python/ceval.c:3690
#2 0x080c451c in PyEval_EvalFrameEx (f=0x820d13c, throwflag=0)
at Python/ceval.c:3690
#3 0x080c5fd4 in PyEval_EvalCodeEx (co=0xf7f22d58, globals=0xf7f962d4,
locals=0x0, args=0xf7f46e18, argcount=1, kws=0x0, kwcount=0, defs=0x0,
defcount=0, closure=0x0) at Python/ceval.c:2876
#4 0x0810d4d6 in function_call (func=0xf7f41454, arg=0xf7f46e0c, kw=0x0)
at Objects/funcobject.c:524
#5 0x0805b0b4 in PyObject_Call (func=0xf7f41454, arg=0x0, kw=0x0)
at Objects/abstract.c:1872
#6 0x08061008 in instancemethod_call (func=0xf7f41454, arg=0xf7f46e0c,
kw=0x0)
at Objects/classobject.c:2509
#7 0x0805b0b4 in PyObject_Call (func=0xf7f3fdec, arg=0x0, kw=0x0)
at Objects/abstract.c:1872
#8 0x080be8f9 in PyEval_CallObjectWithKeywords (func=0x0, arg=0xf7fa002c,
kw=0x0) at Python/ceval.c:3473
#9 0x080efb4d in t_bootstrap (boot_raw=0x820c910)
at ./Modules/threadmodule.c:422
#10 0x4dabd341 in start_thread () from /lib/tls/i686/cmov/libpthread.so.0
#11 0x4da114ee in clone () from /lib/tls/i686/cmov/libc.so.6
============================And as with all problems of this sort... sometimes the program exits
normally without any problems.I ran python trunk:60012 under gdb above. But these problems occur on
older versions such as 2.4.So can we definitely rule out that this could be caused by the recent
changes to threading.py (r57216)?Yes i believe it is unrelated to any recent change.
I can reproduce both behaviors on my OS X 10.4 dual core mac using
Python 2.5.1.Python 2.3 on the mac appears to get stuck in a loop when run stand
alone but gets a memory access fault when run under gdb.python 2.4.4 seems to hang most of the time.
(all behaviors are possible i expect, i just ran it by hand under
several versions a few times)The systems i ran it on when reporting the bug was SMP. As with many
threading bugs it might be easier to reproduce on SMP systems but i
haven't checked.I uploaded a script for a similar issue:
http://bugs.python.org/issue1596321 (thread_crash.py)
which sometimes segfaults on Windows XP (trunk version, debug build),
with the same call stacks as printed by Gregory, on both threads.I explain it this way: On interpreter shutdown, the main thread clears the other's thread TreadState. There you find the instruction: Py_CLEAR(tstate->frame); But this can call arbitrary python code on deallocation of locals, and release the GIL (file.close() in our case). The other thread can then continue to run. If it is given enough processor time before the whole process shutdowns, it will reach this line in ceval.c (line 2633) if (tstate->frame->f_exc_type != NULL) since tstate->frame has been cleared by the main thread ==> boom.
There can be several ways to solve this problem:
- While clearing a thread state, prevent the "owner" thread to gain
focus. Only the main thread can "Py_END_ALLOW_THREADS". The daemon
thread is blocked forever, and will die there. - Be smarter when clearing tstate->frame: first call frame_clear(),
which will free local variables, but let the frame in a valid state.
Only after that, call Py_CLEAR(tstate->frame), when we are sure that no
more thread switch can occur. Of course there are many other fields in a
frame; we have to inspect them carefully. The first solution seems a
more robust approach.
- While clearing a thread state, prevent the "owner" thread to gain
Rhamphoryncus commented
on Jan 17, 2008 RhamphoryncusmannequinMannequinI think non-main threads should kill themselves off if they grab the
interpreter lock and the interpreter is tearing down. They're about to
get killed off anyway, when the process exits.PyGILState_Ensure would still be broken. It touches various things that
get torn down (autoInterpreterState, autoTLSkey, and the HEAD_LOCK
macros) before it grabs the GIL. Reordering should be possible of course.agreed, during shutdown the other threads should be stopped. anything
to do this complicates acquiring and releasing the GIL by adding another
check to see if we're shutting down.brainstorm: I haven't looked at the existing BEGIN_ALLOW_THREADS and
END_ALLOW_THREADS implementations but would it be possible to modify
them on the fly from the thread doing the shutdown (main or not) while
it holds the GIL such that all future calls to BEGIN_ALLOW_THREADS do
not actually release the GIL and END_ALLOW_THREADS always blocks. That
should bring other threads to a halt pretty quickly and prevent
destructors from releasing the GIL (file.close, etc).Rhamphoryncus commented
on Jan 17, 2008 RhamphoryncusmannequinMannequinI'm not sure I understand you, Gregory. Are arguing in favour of adding
extra logic to the GIL code, or against it?I'm attaching a patch that has non-main thread exit, and it seems to fix
the test case. It doesn't fix the PyGILState_Ensure problems though.Also note that PyThread_exit_thread() was completely broken, becoming a
no-op once threading was initialized, and even before then it would exit
the process rather than just the thread. I've fixed it for pthreads,
but not for any of the other platforms.PyGILState_Ensure would still be broken.
It touches various things that get torn down (autoInterpreterState,
autoTLSkey, and the HEAD_LOCK macros) before it grabs the GIL.
Reordering should be possible of course.Adam, did you notice the change on revision 59231 ? the
PyGILState_Ensure stuff should now remain valid during the
PyInterpreterState_Clear() call.Adam, your patch cover one case of the thread releasing the GIL
(Py_ALLOW_THREADS), but there are more calls to PyThread_acquire_lock
(the checkinterval for example).I have a competing patch: it makes the main thread never release the GIL
after some point. Other threads stay blocked and don't clean themselves:
the OS will take care of them.Both approaches correct the initial problem, though.
A remaining question is where exactly in Py_Finalize() we should ban
other threads. I think that this point should be as late as possible, to
allow some object/resources to be cleaned properly.Rhamphoryncus commented
on Jan 17, 2008 RhamphoryncusmannequinMannequinAdam, did you notice the change on revision 59231 ? the
PyGILState_Ensure stuff should now remain valid during the
PyInterpreterState_Clear() call.That doesn't matter. PyGILState_Ensure needs to remain valid *forever*.
Only once the process is completely gone can we be sure it won't be called.Note that PyGILState_Ensure has two behaviours: it can be called when
your thread is already running python code and has a tstate, or it can
be called when you have none. revision 59231 fixed the former, but only
when Py_Finalize hasn't finished. It doesn't fix it for any daemon
threads, or for dummy threads (created outside of python's control)....
You're right, I did forget the 3 other places that acquire the
interpreter_lock. The more I think about it the more I think our two
approaches are equivalent, but conditionalizing the release means we
don't need to touch the broken PyThread_exit_thread functions.I think the banning should be as early as possible, right after
call_sys_exitfunc() has finished. You can't do anything sane once the
interpreter is half-gone (despite the precedent of trying anyway.)That doesn't matter. PyGILState_Ensure needs to remain valid
*forever*. Only once the process is completely gone can we be sure
it won't be called.We could apply the same idea: when exiting, PyGILState_Ensure() blocks
forever, except for the main thread of course.Note that all this state must be restartable: after Py_Finalize(), it
should be possible to call Py_Initialize() again. This seems to raise
the score of the "exit_thread" approach.
I don't know if multiple interpreters are well supported, though.
Is there somewhere a list of use cases, or a test script that can
exercise this?Rhamphoryncus commented
on Jan 17, 2008 RhamphoryncusmannequinMannequinPyGILState_Ensure WOULD block forever if it acquired the GIL before
doing anything else.The only way to make Py_Initialize callable after Py_Finalize is to make
various bits of the finalization into no-ops. For instance, it's
currently impossible to unload C extension modules, so they must be left
around permanently. I'm not convinced it works currently, or that
there's use cases for it.Note that unloading python.so between Py_Finalize and Py_Initialize
would definitely break things, as you'd lose all the global variables
maintaining things. That eliminates the only "it's more elegant" argument.PyGILState_Ensure WOULD block forever if it acquired the GIL
before doing anything else.
Is it possible at all? PyThread_acquire_lock is not reentrant.8 remaining items
With Python 3.3, thread_exit.py still crashes more or less randomly:
$ ./python thread_exit.py [49205 refs] python: Python/_warnings.c:501: setup_context: Assertion `((((((PyObject*)(*filename))->ob_type))->tp_flags & ((1L<<28))) != 0)' failed. Abandon
A patch that seems to work under Linux and Windows (for 3.2/3.3).
Victor pointed out that Py_Finalize() is not necessarily called in the main Python thread. This new patch records the thread state of the finalizing thread, and also includes a test case.
python-dev commented
on May 4, 2011 python-devmannequinMannequinNew changeset 2a19d09b08f8 by Antoine Pitrou in branch '3.2':
Issue bpo-1856: Avoid crashes and lockups when daemon threads run while the
http://hg.python.org/cpython/rev/2a19d09b08f8New changeset c892b0321d23 by Antoine Pitrou in branch 'default':
Issue bpo-1856: Avoid crashes and lockups when daemon threads run while the
http://hg.python.org/cpython/rev/c892b0321d23Should be fixed in 3.2 and 3.3 now. I don't really want to bother with 2.7 and 3.1 (the GIL implementation is different), but someone can backport the patch if they want to :)
guettli commented
on Aug 7, 2013 guettlimannequinMannequinThere are some examples to work around this for Python2: http://stackoverflow.com/questions/18098475/detect-interpreter-shut-down-in-daemon-thread
python-dev commented
on Jun 17, 2014 python-devmannequinMannequinNew changeset 7741d0dd66ca by Benjamin Peterson in branch '2.7':
avoid crashes and lockups from daemon threads during interpreter shutdown (bpo-1856)
http://hg.python.org/cpython/rev/7741d0dd66cahttp://tracker.ceph.com/issues/8797 reports that the backport to 2.7 causes a regression in ceph.
python-dev commented
on Nov 21, 2014 python-devmannequinMannequinNew changeset 4ceca79d1c63 by Antoine Pitrou in branch '2.7':
Issue bpo-21963: backout issue bpo-1856 patch (avoid crashes and lockups when
https://hg.python.org/cpython/rev/4ceca79d1c63bpo-1193099 was marked as a duplicate of this issue.
Note: these values reflect the state of the issue at the time it was migrated and might not reflect the current state.
Show more details
GitHub fields:
bugs.python.org fields: