Signal handling doesn't properly handle OpenMP code

[ohanar]

I was playing around with the using the new prange functionality in cython and happened across this when I tried to interrupt the execution:

sage: sage: time prime_pi(10**12)
37607912018
Time: CPU 1.25 s, Wall: 0.80 s
time prime_pi(10**12)
^C/home/sage/5.0.beta8/local/lib/libcsage.so(print_backtrace+0x31)[0x7f9a6f5c0996]
/home/sage/5.0.beta8/local/lib/libcsage.so(sigdie+0x14)[0x7f9a6f5c09c8]
/home/sage/5.0.beta8/local/lib/libcsage.so(sage_signal_handler+0x17d)[0x7f9a6f5c0587]
/lib64/libpthread.so.0(+0xfb80)[0x7f9a74c7bb80]
/home/sage/5.0.beta8/local/lib/libpython2.7.so.1.0(PyErr_Restore+0x4c)[0x7f9a74f8a99c]
/home/sage/5.0.beta8/local/lib/libpython2.7.so.1.0(PyErr_SetString+0x27)[0x7f9a74f8aad7]
/home/sage/5.0.beta8/local/lib/libcsage.so(sage_signal_handler+0xed)[0x7f9a6f5c04f7]
/lib64/libpthread.so.0(+0xfb80)[0x7f9a74c7bb80]
/home/sage/5.0.beta8/local/lib/libpython2.7.so.1.0(PyErr_Restore+0x4c)[0x7f9a74f8a99c]
/home/sage/5.0.beta8/local/lib/libcsage.so(sage_interrupt_handler+0x48)[0x7f9a6f5c03b0]
/lib64/libpthread.so.0(+0xfb80)[0x7f9a74c7bb80]
/home/sage/5.0.beta8/local/lib/python2.7/site-packages/sage/functions/prime_pi.so(+0x4099)[0x7f9a4e185099]
/home/sage/5.0.beta8/local/lib/python2.7/site-packages/sage/functions/prime_pi.so(+0x3fe8)[0x7f9a4e184fe8]
/home/sage/5.0.beta8/local/lib/python2.7/site-packages/sage/functions/prime_pi.so(+0x5095)[0x7f9a4e186095]
/home/sage/5.0.beta8/local/lib/python2.7/site-packages/sage/functions/prime_pi.so(+0x58cd)[0x7f9a4e1868cd]
/home/sage/5.0.beta8/local/lib/python2.7/site-packages/sage/functions/prime_pi.so(+0x7509)[0x7f9a4e188509]
/home/sage/5.0.beta8/local/lib/python2.7/site-packages/sage/functions/prime_pi.so(+0x8c38)[0x7f9a4e189c38]
/home/sage/5.0.beta8/local/lib/libpython2.7.so.1.0(PyObject_Call+0x53)[0x7f9a74ed6b53]
/home/sage/5.0.beta8/local/lib/libpython2.7.so.1.0(+0x4dc2b)[0x7f9a74ed6c2b]
/home/sage/5.0.beta8/local/lib/libpython2.7.so.1.0(PyObject_CallMethod+0xc1)[0x7f9a74ed6f41]
/home/sage/5.0.beta8/local/lib/libpynac-0.2.so.3(_ZNK5GiNaC8function4evalEi+0x4ce)[0x7f9a52c15e3e]
/home/sage/5.0.beta8/local/lib/libpynac-0.2.so.3(_ZN5GiNaC2ex20construct_from_basicERKNS_5basicE+0x6e)[0x7f9a52c03b4e]
/home/sage/5.0.beta8/local/lib/python2.7/site-packages/sage/symbolic/function.so(_Z16g_function_eval1jRKN5GiNaC2exEb+0xb7)[0x7f9a521da557]
/home/sage/5.0.beta8/local/lib/python2.7/site-packages/sage/symbolic/function.so(+0x21def)[0x7f9a521e1def]
/home/sage/5.0.beta8/local/lib/libpython2.7.so.1.0(PyObject_Call+0x53)[0x7f9a74ed6b53]
/home/sage/5.0.beta8/local/lib/libpython2.7.so.1.0(PyEval_CallObjectWithKeywords+0x47)[0x7f9a74f740e7]
/home/sage/5.0.beta8/local/lib/python2.7/site-packages/sage/functions/prime_pi.so(+0x939f)[0x7f9a4e18a39f]
/home/sage/5.0.beta8/local/lib/libpython2.7.so.1.0(PyObject_Call+0x53)[0x7f9a74ed6b53]
/home/sage/5.0.beta8/local/lib/libpython2.7.so.1.0(PyEval_EvalFrameEx+0x3fcd)[0x7f9a74f7869d]
/home/sage/5.0.beta8/local/lib/libpython2.7.so.1.0(PyEval_EvalCodeEx+0x855)[0x7f9a74f7b645]
/home/sage/5.0.beta8/local/lib/libpython2.7.so.1.0(PyEval_EvalCode+0x32)[0x7f9a74f7b782]
/home/sage/5.0.beta8/local/lib/libpython2.7.so.1.0(PyEval_EvalFrameEx+0x57bf)[0x7f9a74f79e8f]
/home/sage/5.0.beta8/local/lib/libpython2.7.so.1.0(PyEval_EvalCodeEx+0x855)[0x7f9a74f7b645]
/home/sage/5.0.beta8/local/lib/libpython2.7.so.1.0(PyEval_EvalFrameEx+0x51f1)[0x7f9a74f798c1]
/home/sage/5.0.beta8/local/lib/libpython2.7.so.1.0(PyEval_EvalCodeEx+0x855)[0x7f9a74f7b645]
/home/sage/5.0.beta8/local/lib/libpython2.7.so.1.0(PyEval_EvalFrameEx+0x51f1)[0x7f9a74f798c1]
/home/sage/5.0.beta8/local/lib/libpython2.7.so.1.0(PyEval_EvalFrameEx+0x5dc1)[0x7f9a74f7a491]
/home/sage/5.0.beta8/local/lib/libpython2.7.so.1.0(PyEval_EvalCodeEx+0x855)[0x7f9a74f7b645]
/home/sage/5.0.beta8/local/lib/libpython2.7.so.1.0(PyEval_EvalFrameEx+0x51f1)[0x7f9a74f798c1]
/home/sage/5.0.beta8/local/lib/libpython2.7.so.1.0(PyEval_EvalCodeEx+0x855)[0x7f9a74f7b645]
/home/sage/5.0.beta8/local/lib/libpython2.7.so.1.0(PyEval_EvalFrameEx+0x51f1)[0x7f9a74f798c1]
/home/sage/5.0.beta8/local/lib/libpython2.7.so.1.0(PyEval_EvalCodeEx+0x855)[0x7f9a74f7b645]
/home/sage/5.0.beta8/local/lib/libpython2.7.so.1.0(PyEval_EvalFrameEx+0x51f1)[0x7f9a74f798c1]
/home/sage/5.0.beta8/local/lib/libpython2.7.so.1.0(PyEval_EvalCodeEx+0x855)[0x7f9a74f7b645]
/home/sage/5.0.beta8/local/lib/libpython2.7.so.1.0(PyEval_EvalCode+0x32)[0x7f9a74f7b782]
/home/sage/5.0.beta8/local/lib/libpython2.7.so.1.0(PyRun_FileExFlags+0xb0)[0x7f9a74f9db20]
/home/sage/5.0.beta8/local/lib/libpython2.7.so.1.0(PyRun_SimpleFileExFlags+0xdf)[0x7f9a74f9e5bf]
/home/sage/5.0.beta8/local/lib/libpython2.7.so.1.0(Py_Main+0xb85)[0x7f9a74fb1275]
/lib64/libc.so.6(__libc_start_main+0xed)[0x7f9a74281fad]
python[0x4006f1]

------------------------------------------------------------------------
An error occured during signal handling.
This probably occurred because a *compiled* component of Sage has a bug
in it and is not properly wrapped with sig_on(), sig_off(). You might
want to run Sage under gdb with 'sage -gdb' to debug this.
Sage will now terminate.
------------------------------------------------------------------------
/home/sage/5.0.beta8/spkg/bin/sage: line 308:  8428 Segmentation fault      (core dumped) sage-ipython "$@" -i

I've attached a patch for the sage library that demonstrates the bug.

[ohanar]

prime_pi that causes the interrupt error

[jdemeyer]

Proposal: close as wontfix.

The problem is that Cython uses OpenMP. Even I could fix the signal handling code, it's not at all clear how to make OpenMP deal with this. The following makes me think that a fix is impossible:

• ​http://stackoverflow.com/questions/3914264/openmp-is-there-a-way-for-a-thread-to-terminate-all-other-parallel-threads
• ​http://stackoverflow.com/questions/8482651/how-do-i-conditionally-terminate-a-parallel-region-in-openmp

OpenMP simply isn't designed for this. With pthreads, it might be possible as it has clear specifications about how threads interact with signals.

[jdemeyer]

very basic code demonstrating the issue

[jdemeyer]

Using sig_check() inside the loop does work:

from cython.parallel import prange

include 'ext/interrupt.pxi'

def dumb_function():
    cdef int i,x

    while True:
        for i in prange(1<<30,nogil=True):
            with gil:
                sig_check()
            x += i

    return x

From Cython's point of view, sig_check() simply raises an exception when an interrupt occurred, which is explicitly allowed by ​http://docs.cython.org/src/userguide/parallelism.html#breaking-out-of-loops The worst-case overhead is one loop iteration per thread, which is optimal given the limitations of OpenMP.

We should probably declare sig_check() as nogil though, see #15352.