Condition.wait() without a timeout will never raise a KeyboardInterrupt:

cond = threading.Condition()
cond.acquire()
cond.wait()

*** Pressing Ctrl-C now does nothing ***

If you pass a timeout to Condition.wait(), however, it does behave as expected:

cond.wait()

^CTraceback (most recent call last):
  File "/usr/lib/python3.1/threading.py", line 242, in wait
    _sleep(delay)
KeyboardInterrupt

This may affect other problems reported with multiprocessing pools. Most notably:
http://bugs.python.org/issue8296
http://stackoverflow.com/questions/1408356

In 3.2, even using a timeout doesn't make the call interruptible.
The solution would be to fix the internal locking APIs so that they handle incoming signals properly, and are able to return an error status.

3.2 is interesting in that it introduces a new internal API: PyThread_acquire_lock_timed(). We can therefore change that API again before release without risking any compatibility breakage. Reid, would you want to work on this?

I'd like to fix it, but I don't know if I'll be able to in time. It was something that bugged me while running the threading tests while working on Unladen.

I'm imagining (for POSIX platforms) adding some kind of check for signals when the system call returns EINTR. If the signal handler raises an exception, like an interrupt should raise a KeyboardInterrupt, we can just give a different return code and propagate the exception.

It also seems like this behavior can be extended gradually to different platforms, since I don't have the resources to change and test every threading implementation.

I'm imagining (for POSIX platforms) adding some kind of check for
signals when the system call returns EINTR. If the signal handler
raises an exception, like an interrupt should raise a
KeyboardInterrupt, we can just give a different return code and
propagate the exception.

Yes, this is what I'm proposing too.

It also seems like this behavior can be extended gradually to
different platforms, since I don't have the resources to change and
test every threading implementation.

There is only one active POSIX threading implementation in 3.2, in
Python/thread_pthread.h.
(and the only non-POSIX one is for Windows)

Here's a patch that makes Python-level lock acquisitions interruptible for py3k. There are many users of the C-level lock API, most of whom are not set up to deal with lock acquisition failure. I decided to make a new API function and leave the others alone.

If possible, I think this should go out with 3.2.

In that case, I was wondering if I should merge PyThread_acquire_lock_timed with my new PyThread_acquire_lock_timed_intr, since PyThread_acquire_lock_timed wasn't available in 3.1. Although it did go out in 2.7, we don't promise C API compatibility with the 2.x series, so I don't think it matters.

I've tested this patch on Mac OS X and Linux. The whole test suite passes on both, along with the test that I added to test_threadsignals.py.

I added a noop compatibility wrapper to thread_nt.h, but I haven't tested it or built it. When I get around to testing/fixing the subprocess patch on Windows, I'll make sure this works and the test is skipped.

Oh, nice!

I was wondering if I should merge PyThread_acquire_lock_timed with my
new PyThread_acquire_lock_timed_intr, since PyThread_acquire_lock_timed
wasn't available in 3.1.

Yes, I think you should.

I haven't tried the patch, but it seems you got the logic right. There's a problem sometimes that you're using 2 spaces for indent rather than 4. Also, you forgot to update the RLock implementation in _threadmodule.c (and perhaps add another test for it).

(there are other modules which use the PyThread_acquire_lock API, but most of the time the locks are held for a short time (and shouldn't deadlock), which makes converting them less of a priority)

Here is a new version of a patch that updates recursive locks to have the same behavior. The pure Python RLock implementaiton should be interruptible by virtue of the base lock acquire primitive being interruptible.

I've also updated the relevant documentation I could find. I've surely missed some, though.

I also got rid of the _intr version of lock acquisition and simply added a new parameter to the _timed variant. The only two callers of it are the ones I updated in the _thread module.

The RLock version gets different semantics in your patch: it doesn't retry acquiring when a signal is caught. Is there a reason for that?

By the way, your patch works fine under Windows.

Oops, copy/paste oversight. =/ I wrote a test to verify that it handles signals, and then retries the lock acquire.

Also, thanks for the quick reviews!

The latest patch looks good to me.

Actually, there is a problem in Lock.acquire and RLock.acquire. If a signal occurs and signal handling returns successfully, acquiring the lock will be retried without decrementing the timeout first. Therefore, we may end up waiting longer than the user wanted.

I'm not sure how to tackle that: either we accept that an incoming signal will make the wait longer, or we fix it by properly decrementing the timeout (which will complicate things a bit, especially for cross-platform time querying - but see bpo-9079 which might help us).

Accepting that the timeout is not perfect in the face of signals as a caveat is fine (document the possibility) and can be improved later.

Alternatively, do you think it would be better to ignore interrupts when a timeout is passed? If a timeout is passed, the lock acquire will eventually fail in a deadlock situation, and the signal will be handled in the eval loop.

However, if the timeout is sufficiently long, this is still a problem.

I'd prefer to do that or use gettimeofday from _time than leave this as is.

Waiting until the portability hacks for gettimeofday make it into core Python.

The portability API is now available (see Include/pytime.h).

Added a patch that adds support for recomputing the timeout, plus a test for it.

Can this still make it into 3.2, or is it too disruptive at this point in the release process?

Added a patch that adds support for recomputing the timeout, plus a test for it.

Can this still make it into 3.2, or is it too disruptive at this point
in the release process?

No problem at this point, we're not yet in beta phase.
I haven't looked at the patch itself, but thank you.

Committed in r87292. Thank you for doing this!

The test fails on FreeBSD 6.2 (x86 FreeBSD py3k buildbot) since r87292 => bpo-10720. I reopen the issue.

Well, since bpo-10720 exists, while do you reopen this one?

Is it possible to backport this patch to 2.7?

Is it possible to backport this patch to 2.7?

No.

Python 2.7 supports many implementations of threads:

$ ls Python/thread_*h -1
Python/thread_atheos.h
Python/thread_beos.h
Python/thread_cthread.h
Python/thread_foobar.h
Python/thread_lwp.h
Python/thread_nt.h
Python/thread_os2.h
Python/thread_pth.h
Python/thread_pthread.h
Python/thread_sgi.h
Python/thread_solaris.h
Python/thread_wince.h

Supporting signals on all implementations would be a huge work.

Handling correctly signals is also a complex task. For example, the PEP-475 changed how Python 3.5 handles signals (retry interrupted syscall), the change is larger than just the threading module. This PEP has an impact on the whole CPython code base (C and Python).

In Python, unit tests are important. The change b750c45a772c added many unit tests, I expect that it will be tricky to make them reliable on all platforms supported by Python 2.7. I dislike the idea of backporting the feature just for a few platforms.

I mean that the change b750c45a772c alone is not enough, this change depends on many earlier changes used to cleanup/prepare the code to support this change. It is also followed by many changes to enhance how Python handles signals. It took multiple years to fix all subtle issues, race conditions, etc.

There is also a significant risk of breaking applications relying on the current behaviour of Python 2.7 (locks are not interrupted by signals).

As a data point confirming that: we tried backporting the much nicer non-polling condition implementation (https://github.com/python/cpython/commit/15801a1d52c25fa2a19d649ea2671080f138fca1.patch) to our Python 2.7 interpreter at work but ran into actual code that failed as a result of the interruption behavior changing.

While it is a nice goal, it is a non-trivial change. Not something we could ever do within a stable release (2.7).