Reinstall signal handler when Python signal handler called

[user202729]

This would fix sagemath/sage#39601 . Not sure if it has any downside.

[user202729]

These two tests are failing: test_interrupt_bomb and test_sig_block_outside_sig_on.

The test can be fixed by adding appropriate PyErr_CheckSignals , but that feels like fixing the symptoms.

On the other hand a test could be added. It could be as simple as

import signal
from signal import SIGINT
signal.signal(SIGINT, signal.getsignal(SIGINT))

[tornaria]

The point of having a os-level interrupt handler is to rollback the process (using longjmp) to a state where the python signal handler can be called.

IOW, I don't think this will fix anything until the python handler is called; the problem is when the python handler is NOT called (this is the reason why cysignals needs to use a os-level interrupt handler).

I think there's something special about time.sleep() that allows it to be interrupted with a python signal handler (i.e. cysignals is not needed here). Your benchmark should be to be able to interrupt code like this:

$ cat forever.pyx 
from cysignals.signals cimport sig_on, sig_off

def run():
    sig_on()
    while True:
        pass
    sig_off()

---

I believe the only way to fix this is for matplotlib to restore the os-level interrupt handler. See prompt-toolkit/python-prompt-toolkit#1576 for a similar issue and prompt-toolkit/python-prompt-toolkit#1822 for the fix.

A different workaround might be to make sure the signal handler is correctly installed at each prompt (I'm not sure ipython has a hook for that... I think prompt_toolkit has it, but we don't use prompt_toolkit directly afair). This would not be good for scripts but at least an improvement for interactive use.

[user202729]

1. Fixing matplotlib would be a good idea.

2. It's true that fixes such as this one won't fix code of the following form

   plt.pause(2)
   [wait for 1 second]
   [Ctrl-C]
   longRunningCythonFunction()
   [Ctrl-C]
   

   but at least it will fix the following case (which is much more common)

   plt.pause(2)
   [wait for 1 second]
   [Ctrl-C]
   longRunningPurePythonFunction()
   [Ctrl-C]
   

3. In an ideal world the library that changes/restores signal handler should be responsible for it. In practice programs are preferably bug-proof i.e. if other programs are buggy we should try our best to gracefully handle error.

4. Looks like this pull request actually causes some trouble (when automatic pdb call on error is enabled, it ends up invoking pdb twice on keyboard interrupt…? or maybe that's an irrelevant issue)

   though any idea why the behavior change happens? Could the python signal handler be ran twice? (how exactly does PyErr_SetInterrupt and PyErr_CheckSignals work?)