Segmentation fault when calling C code using GMP initialization together with OpenMP

[ederc]

When calling C code from julia I figured out a segmentation fault which appears when using in the C code GMP together with OpenMP and letting the code run in parallel. If I let the code run from the C library without calling it from julia, this error does not appear. Note that the error occurs only if I set num_threads greater than 1 in the OpenMP pragma.

Here is a minimal example:

#include <gmp.h>
#include <omp.h>
void min_example()
{
    int i;
#pragma omp parallel for num_threads(4) private(i)
    for (i = 0; i < 4; ++i) {
        mpz_t x;
        mpz_init(x);
        mpz_clear(x);
    }
}

Running the C code directly no error appears, there is no memory issue, I even tried it with valgrind.
Doing now a ccall like

function test()
    lib = Libdl.dlopen(clib)
    sym = Libdl.dlsym(lib, :min_example)
    ccall(sym, Nothing, ())
end

the call of test throws a segmentation fault:

#0  0x00007ffff7b36871 in jl_gc_counted_malloc () from /usr/bin/../lib/libjulia.so.1
#1  0x00007fffe64fb008 in __gmpz_init () from /usr/bin/../lib/julia/libgmp.so

The error always appears when calling mpz_init(), also in my bigger real example where I have encountered the problem at first.
Again, if num_threads(1) is given, no such error appears, but once the value of num_threads is greater than 1 and the computation is done in parallel I run into this error. Of course, x in the above example is private to each thread.

Here is the versioninfo:

Julia Version 1.2.0
Commit c6da87ff4b (2019-08-20 00:03 UTC)
Platform Info:
  OS: Linux (x86_64-pc-linux-gnu)
  CPU: Intel(R) Core(TM) i7-5557U CPU @ 3.10GHz
  WORD_SIZE: 64
  LIBM: libopenlibm
  LLVM: libLLVM-6.0.1 (ORCJIT, broadwell)

[stevengj]

At first I thought that maybe Julia's bundled MPFR library isn't built with the --enable-thread-safe option, but this doesn't seem to be the case:

julia> ccall((:mpfr_buildopt_tls_p,:libmpfr), Cint, ())
1

julia> ccall((:mpfr_buildopt_tls_p,:libmpfr), Cint, ())
1

However, I should note that we also replace the GMP memory allocator with our own:

        ccall((:__gmp_set_memory_functions, :libgmp), Cvoid,
              (Ptr{Cvoid},Ptr{Cvoid},Ptr{Cvoid}),
              cglobal(:jl_gc_counted_malloc),
              cglobal(:jl_gc_counted_realloc_with_old_size),
              cglobal(:jl_gc_counted_free_with_size))

and your segfault is occurring in jl_gc_counted_malloc.

@vtjnash, is jl_gc_counted_malloc thread-safe?

[thofma]

@ederc I think I ran into something similar when enabling threading for flint in Nemo. Can you try the hack at https://github.com/Nemocas/Nemo.jl/blob/bb40f2055d2fdd50e58ac032937ce75f1a1ea5fa/src/Nemo.jl#L226? (As a workaround)

[ederc]

@thofma Thanks Tommy, your hack works for me.

Is this really the way one should do this or should we keep this issue open?

[thofma]

I don't think this is the way one should do it. I hope that someone knowledgeable of the GC internals can help with this.