Retain info references on public info dup

[devreal]

We need to keep the references when duplicating info through MPI_Info_dup so that subsequent calls to MPI_Info_dup see the same info entries.

Test case below was provided by Neil Fortner. Is there a way to add this test to the test harness?

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

#define ERROR(msg) \
do { \
    printf(msg "\n"); \
    exit(1); \
} while(0)

int main(void) {
    MPI_Info info1 = MPI_INFO_NULL, info2 = MPI_INFO_NULL, info3 = MPI_INFO_NULL;
    int nkeys1, nkeys2, nkeys3;

    if (MPI_Info_create(&info1) != MPI_SUCCESS)
        ERROR("MPI_Info_create failed");

    if (MPI_Info_set(info1, "custom_key", "custom_value") != MPI_SUCCESS)
        ERROR("MPI_Info_set failed");

    if (MPI_Info_get_nkeys(info1, &nkeys1) != MPI_SUCCESS)
        ERROR("MPI_Info_get_nkeys(info1, &nkeys1) failed");

    if (MPI_Info_dup(info1, &info2) != MPI_SUCCESS)
        ERROR("MPI_Info_dup failed");

    if (MPI_Info_free(&info1) != MPI_SUCCESS)
        ERROR("MPI_Info_free(&info1) failed");

    if (MPI_Info_get_nkeys(info2, &nkeys2) != MPI_SUCCESS)
        ERROR("MPI_Info_get_nkeys(info2, &nkeys2) failed");

    if (nkeys1 != nkeys2)
        ERROR("Number of keys on duplicated MPI info object does not match that on original");

    if (MPI_Info_dup(info2, &info3) != MPI_SUCCESS)
        ERROR("MPI_Info_dup failed");

    if (MPI_Info_free(&info2) != MPI_SUCCESS)
        ERROR("MPI_Info_free(&info2) failed");

    if (MPI_Info_get_nkeys(info3, &nkeys3) != MPI_SUCCESS)
        ERROR("MPI_Info_get_nkeys(info3, &nkeys3) failed");

    if (nkeys1 != nkeys3)
        ERROR("Number of keys on double duplicated MPI info object does not match that on original");

    if (MPI_Info_free(&info3) != MPI_SUCCESS)
        ERROR("MPI_Info_free(&info3) failed");

    printf("test passed\n");

    return 0;
}

[edgargabriel]

@devreal I plan to add running the HDF5 testsuite to the ompi CI through a github action, that should cover this testcase as well.

[bosilca]

Why the name change for the public_only argument ? I find the old naming more clear.

You don't need to check for public, at that point in the code it doesn't matter anymore. Moreover, while trying to understand the code I don't even see that's the need for ie_referenced, as all set entries have it set. Are you seeing a case where this field is not set ?

[edgargabriel]

I can confirm btw. that this PR fixes the issue #12742 , i.e. the HDF5 testphdf5 test is passing again

[bosilca]

I looked at the code and I see now why you need to copy ie_referenced. It is only set for set/get and not by dup, so if you try to dup an info twice it will not be duplicated the second time.

[devreal]

@bosilca I changed the parameter back to public_only.