segv in type_dup_fn_* Fortran tests

[jsquyres]

On master head, I see that the IBM tests for duplicating datatypes are consistently segv'ing:

$ cd ompi-tests/ibm/datatype
$ ./type_dup_fn_mpifh
...segv...

This happens in all 3 of the Fortran tests (type_dup_fn_[mpih|usempi|usempif08]). It does not seem to happen in the C version of this same test (type_dup_fn), which is... weird.

Here's a stack trace from a resulting corefile:

#0  0x0000003fed232495 in raise () from /lib64/libc.so.6
#1  0x0000003fed233c75 in abort () from /lib64/libc.so.6
#2  0x0000003fed2703a7 in __libc_message () from /lib64/libc.so.6
#3  0x0000003fed275dee in malloc_printerr () from /lib64/libc.so.6
#4  0x0000003fed278c3d in _int_free () from /lib64/libc.so.6
#5  0x00002aaaac3e029a in opal_datatype_destruct (datatype=0x731d80)
    at opal_datatype_create.c:83
#6  0x00002aaaab1f0de1 in opal_obj_run_destructors (object=0x731d80)
    at ../../opal/class/opal_object.h:462
#7  0x00002aaaab1f1269 in ompi_datatype_destroy (type=0x7fffffffcf10)
    at ompi_datatype_create.c:90
#8  0x00002aaaab2c3d67 in PMPI_Type_free (type=0x7fffffffcf10) at ptype_free.c:60
#9  0x00002aaaaaf1fdff in ompi_type_free_f (type=0x7fffffffcf6c, 
    ierr=0x7fffffffcf68) at type_free_f.c:76
#10 0x00002aaaaaf1fdcf in mpi_type_free_ (type=0x7fffffffcf6c, 
    ierr=0x7fffffffcf68) at type_free_f.c:56
#11 0x0000000000400dde in mpi_type_dup_fn_mpifh () at type_dup_fn_mpifh.f90:22
#12 0x0000000000400e58 in main (argc=1, argv=0x7fffffffd5e5)
    at type_dup_fn_mpifh.f90:27
#13 0x0000003fed21ed1d in __libc_start_main () from /lib64/libc.so.6
#14 0x0000000000400c29 in _start ()

I notice that a free() is failing because it appears to be freeing datatype->ptypes, which appears to be a non-malloc'ed pointer somehow:

83              free(datatype->ptypes);
(gdb) p datatype->ptypes
$1 = (size_t *) 0x2aaaab516000 <__compound_literal.2>

@bosilca I tried to dig into this but couldn't figure out where ptypes came from...

[jsquyres]

@ggouaillardet @bosilca git bisect shows that this problem originated from commit 7c938f0, just a few days ago ("opal/datatype: plug a memory leak in opal_datatype_t destructor").

Can you please investigate?

[ggouaillardet]

@jsquyres will do !

[ggouaillardet]

@jsquyres thanks for the report, this is fixed in #6348

FWIW, the C test uses MPI_INT but the Fortran test uses MPI_INTEGER.
ompi_mpi_int.dt.super.ptypes is NULL, but ompi_mpi_integer.dt.super.ptypes is not, hence the crash in Fortran but not in C