coll-base-allgather: fix MPI_IN_PLACE processing

[mkurnosov]

The call of MPI_Allgather with sendbuf and sendtype parameters equal to MPI_IN_PLACE and NULL correspondingly, produces the segmentation fault.

The problem is that sendtype is used even when sendbuf value is MPI_IN_PLACE. But according to the standard, sendtype and sendcount parameters should be ignored in this case.

Signed-off-by: Mikhail Kurnosov mkurnosov@gmail.com

[mkurnosov]

Reproducer

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

int main(int argc, char **argv)
{
    int rank, comm_size;
    MPI_Init(&argc, &argv);
    MPI_Comm_size(MPI_COMM_WORLD, &comm_size);
    MPI_Comm_rank(MPI_COMM_WORLD, &rank);
    int count = 5;

    int *rbuf = malloc(count * comm_size * sizeof(*rbuf));
    if (!rbuf) {
        fprintf(stderr, "Error: malloc failed\n");
        MPI_Abort(MPI_COMM_WORLD, 1);
    }
    for (int i = 0; i < count; i++)
        rbuf[rank * count + i] = rank + i + 1;

    MPI_Allgather(MPI_IN_PLACE, 0, NULL, rbuf, count, MPI_INT, MPI_COMM_WORLD);

    for (int i = 0; i < comm_size; i++) {
        for (int j = 0; j < count; j++) {
            if (rbuf[i * count + j] != i + j + 1) {
                fprintf(stderr, "*** Validation error: [%d]=%d, need %d\n", 
                        i * count + j, rbuf[i * count + j], i + j + 1);
            }
        }
    }
        
    free(rbuf);
    MPI_Finalize();
    return 0;
}

[jsquyres]

bot:mellanox:retest

[jladd-mlnx]

bot:mellanox:retest

[hppritcha]

@mkurnosov please open a PR against 4.0.x with a cherry-pick of 540c2d1 once this PR is merged into master.

[mkurnosov]

@hppritcha done, I reopend the PR #5474

[aravindksg]

In conjunction with the fix in this PR, would the tuned allgather algorithms also need to be fixed?

I ran into a segfault with the above reproducer code on master and the fault was in ompi_coll_tuned_allgather_intra_dec_fixed() where we have a call to ompi_datatype_type_size(sdtype, &dsize). With sdtype being NULL, it produces a segfault. It seems we need to check for MPI_IN_PLACE here too before calling ompi_datatype_type_size() ?

Also, MPI_IN_PLACE check would be needed in ompi_coll_tuned_allgatherv_intra_dec_fixed() as the spec states sendbuf and sendtype needs to be ignored for Allgatherv.

[aravindksg]

@hppritcha , @bosilca : Any thoughts on the above comment related to the segfault in the tuned algorithm as well? If this is a valid issue, I can create a PR for review.

[bosilca]

A fix for tuned would be more than welcome.