The signal module checks the SIGINT handler on startup. It only registers a new custom handler if the default OS handler is still installed, so that when embedding python in an application the SIGINT handler of that application is not overwritten. 

But on shutdown in finisignal, it *always* sets SIGINT to SIG_DFL. The reason is that it saves the old handler in old_siginthandler, but *only* if the signal handler is overwritten in init, which only happens when it was SIG_DFL in the first place! If there was already a handler in place in init (-> no overwriting), old_siginthandler will default to the initialization value, which is also SIG_DFL.

This means that when an application embeds Python and needs a custom SIGINT handler, it will stop to work as soon as it shuts down Python since it will always be reset to SIG_DFL.

New changeset e905c84494526363086f66a979e317e155bf9536 by Antoine Pitrou (pkerling) in branch 'master':
bpo-30654: Do not reset SIGINT handler to SIG_DFL in finisignal (GH-7146)
https://github.com/python/cpython/commit/e905c84494526363086f66a979e317e155bf9536

I'd rather not backport this to 2.7 as it's quite late in the maintenance cycle and I'd like to avoid any regressions there.

> I'd rather not backport this to 2.7 as it's quite late in the maintenance cycle and I'd like to avoid any regressions there.

Since it's a subtle behavior change and the PR doesn't add a flag to opt-in for the old behaviour, I'm not sure about backporting the change (to 3.6 or 3.7).

For 3.7, we are very close to the final release. It doesn't give much time to users to test the new behavior :-( I would suggest to keep the old behavior in Python 3.7 as well.

To be honest, I don't understand well the change. So I'm not confortable with it.

I understand that Py_Initialize() + Py_Finalize() restored the SIGINT handler, but now Python *always* sets SIGINT to SIG_DFL in Py_Finalize(). So if an application has its own signal handler, Python replaces it...

But embedded Python also gives the choice of not setting Python signal handlers in Py_Initialize().

> So if an application has its own signal handler, Python replaces it...

You have it backwards.  Please read the bug report.

> For 3.7, we are very close to the final release. It doesn't give much time to users to test the new behavior 

That's a fair point.  What's the procedure here?  If I backport the fix to the 3.7 branch, will it go straight into the 3.7.0 release or will it be deferred to 3.7.1?  @Ned

> You have it backwards.  Please read the bug report.

I'm confused by the NEWS entry:

+Fixed reset of the SIGINT handler to SIG_DFL on interpreter shutdown even
+when there was a custom handler set previously. Patch by Philipp Kerling.

I read it as Python now always reset SIGINT to SIG_DFL. The commit title is more explicit: "Do not reset SIGINT (...)".

I propose to rephrase the NEWS entry as:

"""
Fix signal handlers when Python is embedded. On Python shutdown, do not reset the SIGINT handler to SIG_DFL, when a custom handler has been set before Python initialization. Patch by Philipp Kerling.
"""

--

Ok, now I understood the change. In this case, I'm ok to backport it to 3.6 and 3.7.

Python 2.7 has this behavior since 10 years, it seems like people learnt to live with it. I also dislike touching Python 2.7, to prevent any risk of regression if someone really rely on the current behavior.

> What's the procedure here?  If I backport the fix to the 3.7 branch, will it go straight into the 3.7.0 release or will it be deferred to 3.7.1??

"All fixes that have been merged into the 3.7 branch as of cutoff tomorrow
will be in 3.7.0b5 and fixes merged afterwards will be in 3.7.0rc1 up to
its cutoff point. After 3.7.0rc1 cutoff, 3.7 merges will appear in 3.7.1.
Please continue to exercise diligence when deciding whether a change is
appropriate for 3.7; as a rule of thumb, treat the 3.7 branch as if it were
already released and in maintenance mode."

Ok, I think it go into 3.7.0 after all.  3.7.1 wouldn't get more testing before it's released anyway.

New changeset 623b439abebc913bc416d92f38fe371e84b0276b by Antoine Pitrou (Miss Islington (bot)) in branch '3.7':
bpo-30654: Do not reset SIGINT handler to SIG_DFL in finisignal (GH-7146) (GH-7306)
https://github.com/python/cpython/commit/623b439abebc913bc416d92f38fe371e84b0276b

New changeset 1d5198fd41ad9185e9e6b3aa595769c3693d57be by Antoine Pitrou (Miss Islington (bot)) in branch '3.6':
bpo-30654: Do not reset SIGINT handler to SIG_DFL in finisignal (GH-7146) (GH-7307)
https://github.com/python/cpython/commit/1d5198fd41ad9185e9e6b3aa595769c3693d57be

Thanks pkerling for your bug report and your fix!

Thanks! I'm a bit disappointed that it won't make it into 2.7, but I can understand the decision.

To give some context: I came across this while working on Kodi and noticing that it does not shutdown cleanly via Ctrl+C or SIGTERM. After investigating, I came to the conclusion that it is due to this bug. Kodi shuts down the Python interpreter every time no add-on is doing active work, which is almost guaranteed to happen shortly after application startup. Then this bug caused a reset of the SIGTERM handler to the default handler, making the Kodi handler that does a clean shutdown useless.
Now there are plans to switch to Python 3 in Kodi, but it won't happen until the major release after the next, so we're stuck with 2.7 for some time. I guess we'll have to work around this in Kodi for the time being.

I'm nosy'ing Benjamin, the 2.7 release manager, in case he wants to comment on desirability of this bugfix in Python 2.7.

It seems like this only affects embeddings and for embeddings, it's strictly an improvement?

Short of any bug in the patch, yes, it's stricly an improvement.

IMHO it's a bugfix.

Reopening for the 2.7 backport discussion.

New changeset ded666ff0ce44785a495ff89b676ca68e4cc08e8 by Antoine Pitrou in branch '2.7':
[2.7] bpo-30654: Do not reset SIGINT handler to SIG_DFL in finisignal (GH-7146) (GH-7347)
https://github.com/python/cpython/commit/ded666ff0ce44785a495ff89b676ca68e4cc08e8

So this is in 2.7 finally.  Closing again.

I marked bpo-24415 as duplicate of this issue.