4.1.x ERROR field of MPI_Status uninitialized after MPI_Recv call

[blattms]

Background information

What version of Open MPI are you using?

The issue is there in versions 4.1.6 (Debian sid) and 4.1.4 (Debian bookworm), but not on 4.1.0 (Debian oldstable/bullseye).

Describe how Open MPI was installed (e.g., from a source/distribution tarball, from a git clone, from an operating system distribution package, etc.)

openmpi is installed from packages either on Debian sid or Debian 12.2

Please describe the system on which you are running

• Operating system/version: Debian GNU/Linux 12.2 and sid
• Computer hardware: amd64
• Network type: only one node without network

---

Details of the problem

If I compile the following test file with debugging

#include <stdio.h>
#include <stdlib.h>
#include <mpi.h>
 
/**
 * @brief Display information contained in the MPI_Status.
 **/
int main(int argc, char* argv[])
{
    MPI_Init(&argc, &argv);
 
    int my_rank;
    MPI_Comm_rank(MPI_COMM_WORLD, &my_rank);
 
    if(my_rank == 0)
    {
        // The "master" MPI process issues the MPI_Bsend.
        int buffer_sent = 12345;
        int tag = 67890;
        printf("MPI process %d sends value %d with tag %d.\n", my_rank, buffer_sent, tag);
        MPI_Ssend(&buffer_sent, 1, MPI_INT, 1, tag, MPI_COMM_WORLD);
    }
    else
    {
        // The "slave" MPI process receives the message.
        int buffer_received;
        MPI_Status status;
        MPI_Recv(&buffer_received, 1, MPI_INT, MPI_ANY_SOURCE, MPI_ANY_TAG, MPI_COMM_WORLD, &status);
	if(status.MPI_ERROR)
          printf("MPI process %d received value %d from rank %d, with tag %d and error code %d.\n", 
               my_rank,
               buffer_received,
               status.MPI_SOURCE,
               status.MPI_TAG,
               status.MPI_ERROR);
    }
 
    MPI_Finalize();
 
    return EXIT_SUCCESS;
}

using

shell$ mpicc -o tester -g -O0 mpi_test.c

and run under valgrind using:

shell$ mpirun -np 2 valgrind --suppressions=/usr/share/openmpi/openmpi-valgrind.supp ./tester

Then valgrind shows that the ERROR field is not initialized during the if(status.MPI_ERROR):

MPI process 0 sends value 12345 with tag 67890.
==38889== Thread 1:
==38889== Conditional jump or move depends on uninitialised value(s)
==38889==    at 0x10925D: main (mpi_test.c:29)
==38889== 

Shouldn't MPI_Recv initialize the MPI_Status field completely? We have asserts like assert(!status.MPI_ERROR); that are failing because of this. The workaround is to always default initialize the status field

MPI_Status status = {};

Is this intended or are we doing something wrong?

[devreal]

This is intended behavior. Section 3.2.5 states:

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 that
return multiple statuses.

I agree that this is not ideal since it leads to spurious warnings like this. Feel free to open an issue against the standard: https://github.com/mpi-forum/mpi-issues

[blattms]

Thanks a lot. I had a hard time figuring out the issue. I only read the man page. From that this was not clear.

[blattms]

Closing this as it is answered.

[jsquyres]

Thanks a lot. I had a hard time figuring out the issue. I only read the man page. From that this was not clear.

Fair enough. Got any suggestions for text to add to the man page about this? (we'll likely need to include that same text on a bunch of receive-flavored man pages)

[blattms]

Thanks a lot. I had a hard time figuring out the issue. I only read the man page. From that this was not clear.

Fair enough. Got any suggestions for text to add to the man page about this? (we'll likely need to include that same text on a bunch of receive-flavored man pages)

I guess a comment would fit after this paragraph about status in the Notes section

If your application does not need to examine the status field, you can save resources by using the
predefined constant  MPI_STATUS_IGNORE as a special value for the status argument.

maybe something along these lines:

Not the the status field is mainly meant to be used to provide the number of elements received via function 
MPI_Get_count, the rank of the sending process and the tag used for sending. Other fields of status might
be uninitialized after the function returns.

[jsquyres]

@blattms Well this certainly caused me to go down a rabbit hole. 🐇 (hey, who knew that Github had a Rabbit emoji?)

I took your suggestion and tweaked it, and noticed that we actually had related text about this issue in some other man pages. So I ended up stealing that text and using it. Check out #12058 (which rendered to a temporary RTD location here: https://ompi--12058.org.readthedocs.build/en/12058/)

[blattms]

Thanks a lot for. That text is really perfect. 👍

[bosilca]

Not that long ago we had a very similar discussion that ended up in us making the change. Check #9128 for more info.