Mailing List Archive: The GIL and PyEval

The GIL and PyEval_RestoreThread

Sep 26, 2023, 6:20 AM

Post #1 of 3 (121 views)

Hi all,

I've been working on embedding Python and have an interesting case around
locking with PyEval_RestoreThread which wasn't quite doing what I expect,
hoping someone can explain what I should expect here.

I have a little example (I'm running this in parallel from two different
threads; I have some more C code for that but I don't think it's super
interesting):

void run_python(PyThreadState* thread) {
LOG("Restoring thread %p...", thread);
PyEval_RestoreThread(thread);
LOG("Restored thread %p", thread);
PyRun_SimpleString("import time; print('sleeping'); time.sleep(3.0)");
LOG("Saving thread...");
PyThreadState* saved_thread = PyEval_SaveThread();
LOG("Saved thread %p", saved_thread);
}

This produces output like
11:46:48.110058893: Restoring thread 0xabc480...
11:46:48.110121656: Restored thread 0xabc480
11:46:48.110166060: Restoring thread 0xabc480...
sleeping
11:46:48.110464194: Restored thread 0xabc480
sleeping
11:46:51.111307541: Saving thread...
11:46:51.111361075: Saved thread 0xabc480
11:46:51.113116633: Saving thread...
11:46:51.113177605: Saved thread 0xabc480

The thing that surprises me is that both threads seem to be able to pass
PyEval_RestoreThread before either reaches the corresponding
PyEval_SaveThread call, which I wasn't expecting to happen; I assumed that
since RestoreThread acquires the GIL, that thread state would remain locked
until it's released.

I understand that the system occasionally switches threads, which I guess
might well happen with that time.sleep() call, but I wasn't expecting the
same thread to become usable somewhere else. Maybe I am just confusing
things by approaching the same Python thread from multiple OS threads
concurrently and should be managing my own locking around that?

Thanks in advance,

Peter

--
Thought Machine Group Limited, a company registered in England & Wales.
Registered number: 11114277.
Registered Office: 5 New Street Square,
London EC4A 3TW
<https://maps.google.com/?q=5+New+Street+Square,+London+EC4A+3TW&entry=gmail&source=g>.

The content of this email is confidential and intended for the recipient
specified in message only. It is strictly forbidden to share any part of
this message with any third party, without a written consent of the sender.
If you received this message by mistake, please reply to this message and
follow with its deletion, so that we can ensure such a mistake does not
occur in the future.
--
https://mail.python.org/mailman/listinfo/python-list

Re: The GIL and PyEval_RestoreThread [ In reply to ]

python-list at python

Sep 26, 2023, 7:48 PM

Post #2 of 3 (121 views)

Permalink

On 2023-09-26 14:20, Peter Ebden via Python-list wrote:
> Hi all,
>
> I've been working on embedding Python and have an interesting case around
> locking with PyEval_RestoreThread which wasn't quite doing what I expect,
> hoping someone can explain what I should expect here.
>
> I have a little example (I'm running this in parallel from two different
> threads; I have some more C code for that but I don't think it's super
> interesting):
>
> void run_python(PyThreadState* thread) {
> LOG("Restoring thread %p...", thread);
> PyEval_RestoreThread(thread);
> LOG("Restored thread %p", thread);
> PyRun_SimpleString("import time; print('sleeping'); time.sleep(3.0)");
> LOG("Saving thread...");
> PyThreadState* saved_thread = PyEval_SaveThread();
> LOG("Saved thread %p", saved_thread);
> }
>
> This produces output like
> 11:46:48.110058893: Restoring thread 0xabc480...
> 11:46:48.110121656: Restored thread 0xabc480
> 11:46:48.110166060: Restoring thread 0xabc480...
> sleeping
> 11:46:48.110464194: Restored thread 0xabc480
> sleeping
> 11:46:51.111307541: Saving thread...
> 11:46:51.111361075: Saved thread 0xabc480
> 11:46:51.113116633: Saving thread...
> 11:46:51.113177605: Saved thread 0xabc480
>
> The thing that surprises me is that both threads seem to be able to pass
> PyEval_RestoreThread before either reaches the corresponding
> PyEval_SaveThread call, which I wasn't expecting to happen; I assumed that
> since RestoreThread acquires the GIL, that thread state would remain locked
> until it's released.
>
> I understand that the system occasionally switches threads, which I guess
> might well happen with that time.sleep() call, but I wasn't expecting the
> same thread to become usable somewhere else. Maybe I am just confusing
> things by approaching the same Python thread from multiple OS threads
> concurrently and should be managing my own locking around that?
>
Storing the result of PyEval_SaveThread in a local variable looks wrong
to me.

In the source for the regex module, I release the GIL with
PyEval_SaveThread and save its result. Then, when I want to claim the
GIL, I pass that saved value to PyEval_RestoreThread.

You seem to be releasing the GIL and discarding the result, so which
thread are you resuming when you call PyEval_RestoreThread?

It looks like you're resuming the same thread twice. As it's already
resumed the second time, no wonder it's not blocking!

--
https://mail.python.org/mailman/listinfo/python-list

Re: The GIL and PyEval_RestoreThread [ In reply to ]

python-list at python

Sep 27, 2023, 2:14 AM

Post #3 of 3 (118 views)

Permalink

The thread variable I'm passing in is the one I originally got from calling
Py_NewInterpreter. I'd assumed that I didn't need to particularly track the
one I get back from SaveThread since it should always be the one I restored
previously (which does seem to be the case).

> It looks like you're resuming the same thread twice. As it's already
resumed the second time, no wonder it's not blocking!

That isn't how I read the docs though? It says "If the lock has been
created, the current thread must not have acquired it, otherwise deadlock
ensues." That suggests to me that it should try to acquire the GIL again
and wait until it can (although possibly also that it's not an expected use
and Python thread states are expected to be more 1:1 with C threads).

On Wed, Sep 27, 2023 at 3:53?AM MRAB via Python-list <python-list@python.org>
wrote:

> On 2023-09-26 14:20, Peter Ebden via Python-list wrote:
> > Hi all,
> >
> > I've been working on embedding Python and have an interesting case around
> > locking with PyEval_RestoreThread which wasn't quite doing what I expect,
> > hoping someone can explain what I should expect here.
> >
> > I have a little example (I'm running this in parallel from two different
> > threads; I have some more C code for that but I don't think it's super
> > interesting):
> >
> > void run_python(PyThreadState* thread) {
> > LOG("Restoring thread %p...", thread);
> > PyEval_RestoreThread(thread);
> > LOG("Restored thread %p", thread);
> > PyRun_SimpleString("import time; print('sleeping'); time.sleep(3.0)");
> > LOG("Saving thread...");
> > PyThreadState* saved_thread = PyEval_SaveThread();
> > LOG("Saved thread %p", saved_thread);
> > }
> >
> > This produces output like
> > 11:46:48.110058893: Restoring thread 0xabc480...
> > 11:46:48.110121656: Restored thread 0xabc480
> > 11:46:48.110166060: Restoring thread 0xabc480...
> > sleeping
> > 11:46:48.110464194: Restored thread 0xabc480
> > sleeping
> > 11:46:51.111307541: Saving thread...
> > 11:46:51.111361075: Saved thread 0xabc480
> > 11:46:51.113116633: Saving thread...
> > 11:46:51.113177605: Saved thread 0xabc480
> >
> > The thing that surprises me is that both threads seem to be able to pass
> > PyEval_RestoreThread before either reaches the corresponding
> > PyEval_SaveThread call, which I wasn't expecting to happen; I assumed
> that
> > since RestoreThread acquires the GIL, that thread state would remain
> locked
> > until it's released.
> >
> > I understand that the system occasionally switches threads, which I guess
> > might well happen with that time.sleep() call, but I wasn't expecting the
> > same thread to become usable somewhere else. Maybe I am just confusing
> > things by approaching the same Python thread from multiple OS threads
> > concurrently and should be managing my own locking around that?
> >
> Storing the result of PyEval_SaveThread in a local variable looks wrong
> to me.
>
> In the source for the regex module, I release the GIL with
> PyEval_SaveThread and save its result. Then, when I want to claim the
> GIL, I pass that saved value to PyEval_RestoreThread.
>
> You seem to be releasing the GIL and discarding the result, so which
> thread are you resuming when you call PyEval_RestoreThread?
>
> It looks like you're resuming the same thread twice. As it's already
> resumed the second time, no wonder it's not blocking!
>
> --
> https://mail.python.org/mailman/listinfo/python-list
>

--
Thought Machine Group Limited, a company registered in England & Wales.
Registered number: 11114277.
Registered Office: 5 New Street Square,
London EC4A 3TW
<https://maps.google.com/?q=5+New+Street+Square,+London+EC4A+3TW&entry=gmail&source=g>.

The content of this email is confidential and intended for the recipient
specified in message only. It is strictly forbidden to share any part of
this message with any third party, without a written consent of the sender.
If you received this message by mistake, please reply to this message and
follow with its deletion, so that we can ensure such a mistake does not
occur in the future.
--
https://mail.python.org/mailman/listinfo/python-list