Regression: MPI_Info_get_nthkey() broken in master

[dalcinl]

This is a trivial reproducer:

In [1]: from mpi4py import MPI

In [2]: info = MPI.Info.Create()

In [3]: info.Set('key', 'value')

In [4]: info.Get_nthkey(0)

Output:

python3: pinfo_get_nthkey.c:97: PMPI_Info_get_nthkey: Assertion `OPAL_OBJ_MAGIC_ID == ((opal_object_t *) (key))->obj_magic_id' failed.

See mpi4py testsuite failure.

[devreal]

I just tested this with a recent master:

#include <mpi.h>
#include <stdio.h>
#include <string.h>

int main()
{
  MPI_Init(NULL, NULL);
  MPI_Info info;
  MPI_Info_create(&info);
  MPI_Info_set(info, "key", "value");
  char buf[MPI_MAX_INFO_KEY];
  memset(buf, MPI_MAX_INFO_KEY, '\0');
  MPI_Info_get_nthkey(info, 0, buf);

  printf("info key 0: %s\n", buf);

  MPI_Finalize();

  return 0;
}

Correctly prints:

$ mpirun -n 1 ./test_info_get_nthkey
info key 0: key

@dalcini can you please test and see if you can reproduce it with this C test?

[dalcinl]

Looks like your master was not recent enough.

$ mpicc tmp.c 
$ ./a.out 
a.out: pinfo_get_nthkey.c:97: PMPI_Info_get_nthkey: Assertion `OPAL_OBJ_MAGIC_ID == ((opal_object_t *) (key))->obj_magic_id' failed.
[kw60439:1760662] *** Process received signal ***
[kw60439:1760662] Signal: Aborted (6)
[kw60439:1760662] Signal code:  (-6)
[kw60439:1760662] [ 0] /lib64/libpthread.so.0(+0x141e0)[0x7fbda3c6a1e0]
[kw60439:1760662] [ 1] /lib64/libc.so.6(gsignal+0x145)[0x7fbda3ac89d5]
[kw60439:1760662] [ 2] /lib64/libc.so.6(abort+0x116)[0x7fbda3ab18a4]
[kw60439:1760662] [ 3] /lib64/libc.so.6(+0x26789)[0x7fbda3ab1789]
[kw60439:1760662] [ 4] /lib64/libc.so.6(+0x36026)[0x7fbda3ac1026]
[kw60439:1760662] [ 5] /home/devel/mpi/openmpi/dev/lib/libmpi.so.0(MPI_Info_get_nthkey+0x2ad)[0x7fbda3d75f6e]
[kw60439:1760662] [ 6] ./a.out[0x4011c4]
[kw60439:1760662] [ 7] /lib64/libc.so.6(__libc_start_main+0xf2)[0x7fbda3ab31e2]
[kw60439:1760662] [ 8] ./a.out[0x4010be]
[kw60439:1760662] *** End of error message ***
Aborted (core dumped)

[dalcinl]

I built from a git clone on Fedora 33 (GCC 10.21), and configured with:

./configure --disable-dependency-tracking --enable-debug --enable-mem-debug --disable-man-pages --disable-mpi-fortran

The mpi4py testsuite failure I linked is running on GitHub Actions Ubuntu 20.04 runners.

[devreal]

Thanks for testing, I found the bug: #8665

[dalcinl]

@jsquyres This kind of regressions are the consequence of insufficient testing. I'm setting this thing up in GitHub Actions [link] to make the point that integrating mpi4py testsuite in Open MPI testing infrastructure should be relatively easy. Porting this stuff to Azure Pipelines should trivial (actually, the mpi4py part of it is mostly done [link]).

[jsquyres]

I just merged the PR which fixes this issue, but... ZOMG. Thank you @dalcinl for the GitHub Actions file! I'm slightly tweaking it and testing it on my own setup, but I can foresee putting this on the main Open MPI repo in the very near future.

[dalcinl]

I'm slightly tweaking it and testing it on my own setup

Please share your tweaks once you are done! I'm particularly interested on adding a matrix with different configure options, or perhaps better, sets of runtime MCA options.