Repository navigation
Nested ctypes calls using njit(cache=True) results in segmentation fault #3621
Description
Using numba 0.41.0 and python 2.7.12 on linux, I see that using 3 levels of function calls to reach a ctypes function call results in a segmentation fault when using njit(cache=True). The first run on the python file runs fine as expected; however, when run the second time with the cached files present, a segmentation fault occurs. Interestingly, when reducing the code to use only 2 function calls to reach the ctypes function, everything works perfectly fine.
compiled using clang++ 8.0.0:
clang++ -O3 -std=c++17 -fPIC -shared crash.cpp -o /tmp/libcrash.so
crash.hpp
extern "C"
{
double e();
}
crash.cpp
#include "crash.hpp"
double e()
{
return 1.0;
}
crash.py
from numba import njit
import ctypes
libcrash = ctypes.CDLL("/tmp/libcrash.so")
c_func = libcrash.e
c_func.argtypes = []
c_func.restype = ctypes.c_double
@njit(cache=True)
def c():
return c_func()
@njit(cache=True)
def b():
return c()
@njit()
def a(cache=True):
return b()
print a()
Activity
[-]Nested ctypes calls in njit[/-][+]Nested ctypes calls using njit(cache=True) results in segmentation fault[/+]lespaul commentedon Dec 20, 2018
Here is a backtrace from gdb which probably isn't helpful but just in case
stuartarchibald commentedon Dec 20, 2018
Thanks for the report. I can reproduce. Here's a reproducer binding to
libm.run with:
this is the top of the backtrace, seems like the signal comes from the ctype bound DSO function ptr.
sklam commentedon Dec 20, 2018
The external function address reference should have disabled caching. That happened in the first layer but lost in the higher layers.
sklam commentedon Dec 20, 2018
So, there's a fix to make it properly disable caching. But, I think users also want a way to reference pre-compiled external function and have the numba function still be cache-able.
Add test for numba#3621