MPI_ERROR field in status struct when calling MPI_Test on a null request

[simonbyrne]

Problem

When calling MPI_Test on a null request, what should the value of the MPI_ERROR field of the status struct be set to? For example consider the following:

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

int main(int argc, char** argv) {
    MPI_Init(NULL, NULL);

    MPI_Request request;
    int flag;
    MPI_Status status;

    request = MPI_REQUEST_NULL;

    // poison the MPI_ERROR value
    status.MPI_ERROR = 1;
    
    MPI_Test(&request, &flag, &status);

    printf("status.MPI_ERROR == MPI_SUCCESS: %d\n", status.MPI_ERROR == MPI_SUCCESS);

    MPI_Finalize();
}

The standard has 2 conflicting statements:

Section 3.2.5 (Return Status) states that

In general, message-passing calls do not modify the value of the error code field of
status variables. This field may be updated only by the functions in Section 3.7.5 which
return multiple statuses. The field is updated if and only if such function returns with an
error code of MPI_ERR_IN_STATUS.

Rationale. The error field in status is not needed for calls that return only one status,
such as MPI_WAIT, since that would only duplicate the information returned by the
function itself. The current design avoids the additional overhead of setting it, in such
cases. The field is needed for calls that return multiple statuses, since each request
may have had a different failure. (End of rationale.)

suggests that MPI_ERROR should be unmodified by MPI_Test in the above code.

On the other hand, section 3.7.3 (Communication Completion) states that

An empty status is a status which is set to
return tag = MPI_ANY_TAG, source = MPI_ANY_SOURCE, error = MPI_SUCCESS, and is
also internally configured so that calls to MPI_GET_COUNT, MPI_GET_ELEMENTS, and
MPI_GET_ELEMENTS_X return count = 0 and MPI_TEST_CANCELLED returns false.

and

One is allowed to call MPI_TEST with a null or inactive request argument. In such a
case the procedure returns with flag = true and empty status.

suggests that MPI_ERROR should be set to MPI_SUCCESS.

As you might guess, MPICH does the former, and Open MPI does the latter.

Proposal

Either:
a) being an "empty status" should not depend on the value of MPI_ERROR,
b) MPI_Test (and MPI_Wait) should modify MPI_ERROR when the request is null or inactive.
c) both (a) and allow (b)
d) require the field to be initialized as MPI_SUCCESS

Changes to the Text

Under option a), section 3.7.3 becomes:

An empty status is a status which is set to
return tag = MPI_ANY_TAG, source = MPI_ANY_SOURCE, error = MPI_SUCCESS, and is
also internally configured so that calls to MPI_GET_COUNT, MPI_GET_ELEMENTS, and
MPI_GET_ELEMENTS_X return count = 0 and MPI_TEST_CANCELLED returns false.

under option b), section 3.2.5 would need to be modified to allow functions to modify the value of the error field. I'm not sure the exact wording, but something like the following

In general, message-passing calls may or may not modify the value of the error code field of
status variables. This field may be updated by the functions in Section 3.7.5 which
return multiple statuses. The field is updated if such function returns with an
error code of MPI_ERR_IN_STATUS.

though this doesn't sound quite right.

I'm not sure how option d) would be worded.

Impact on Implementations

a) would require changes to Open MPI
b) would require changes to MPICH
c) would require no changes.
d) would require no changes.

Impact on Users

a) would change how users check if a status is empty
b) would require no changes
c) same as (a)
d) would require user to initialize status objects before use or reuse.

References and Pull Requests

[wgropp]

It would be odd for MPI_ERROR to be set ONLY when a NULL request is used. While there is no perfect choice here, I'd define an Empty Status to have MPI_ERROR set to MPI_SUCCESS only if that field would have meaning with a non-null request - i.e., not for MPI_Wait and MPI_Test, which already make an exception for MPI_ERROR for when a status is returned.

Note that the only fields that a user can use in MPI_Status to determine that the status is empty is tag and source, since the wild cards aren't valid values on success. The other values, for count and error, are valid, and there values only allow a user to distinguish between a valid empty status and an invalid status - one that should never be returned by a valid MPI program.

I do think that we could choose to allow an implementation to set MPI_ERROR to MPI_SUCCESS in this case, but not require it. While this would mean that a user that bother to confirm that MPI_ERROR was MPI_SUCCESS in an empty status would have to remove that test, I suspect, outside of test cases, this is very close to the null set.

[simonbyrne]

One additional point to consider is if the error field is not set by MPI_Test, then accessing it could be undefined behavior in C (which I think is the intent of the warning in section 3.2.5), e.g.

MPI_Status status;
MPI_Test(&request, &flag, &status);
// do something with status.MPI_ERROR would be undefined behavior

In this case, the user would need to explicitly zero-initialize the struct

MPI_Status status = {};
MPI_Test(&request, &flag, &status);
// do something with status.MPI_ERROR would now be safe

which, since MPI_SUCCESS = 0 (section 9.4), would give the desired result.