bug: uninitialized MPI_Status object

[mpichbot]

Originally by jhammond on 2016-03-28 13:14:51 -0500

---

Message 1```

I notice an uninitialized MPI_Status object can make MPI_Get_count return a wrong result,
when the data amount in the MPI operation is zero (although I only checked MPI-IO).
Attached is a test program that uses MPI collective read where only root process has
non-zero length data to read. The expected result from MPI_Get_count is 0 for all
non-root process. To mimic an uninitialized MPI_Status object, I call memset to make
the object non-zero.

Here is the code fragment.

if (rank =# 0) len10;
else len = 0;

MPI_File_read_all(fh, buf, len, MPI_BYTE, &status);

MPI_Get_count(&status, MPI_BYTE, &get_size);

For process rank > 0, get_size may not be 0.

Message 2```

My test program can be compiled with command "mpicc get_count.c -o get_count"
and run "mpiexec -n 4 get_count". For correct (expected) result, nothing
will be printed on stdout. Otherwise, error messages will be printed.

My point is MPI_Get_count does not report correct result because
MPI collective I/O call fails to initialize the MPI_Status object.

I found that OpenMPI and earlier version of MPICH (2-1.2.1) can run
this test code correctly.

The test failed when I ran MPICH 3.1.4 and the current from GIT repo.

[mpichbot]

Originally by jhammond on 2016-03-28 13:15:02 -0500

---

Attachment added: get_count.c (1.4 KiB)

[mpichbot]

Originally by robl on 2016-03-28 16:31:19 -0500

---

Wei-keng continues his investigation:

I found the cause for the MPI_Get_count warning message.

Valgrind considers the argument status passed to MPI_Get_count
has not been initialized. However, in PnetCDF, status is returned
from a call to either MPI_File_write_all or MPI_File_read_all,
which should initialize the status object "entirely". Entirely means
all members defined in the C struct MPI_Status. I guess MPICH
fails to do that and hence valgrid complains it.

typedef struct MPI_Status {
    int count_lo;
    int count_hi_and_cancelled;
    int MPI_SOURCE;
    int MPI_TAG;
    int MPI_ERROR;
} MPI_Status;

It appears count_lo and count_hi_and_cancelled are not set in MPI-IO calls
or MPI_Iprobe. I tested this on MPI_Iprobe and it has the same problem.

To verify, I added the following 2 lines before calling MPI-IO and MPI_Iprobe
and valgrind stops complaining.

    mpistatus.count_lo = 0;
    mpistatus.count_hi_and_cancelled = 0;

[mpichbot]

Originally by balaji on 2016-03-29 12:45:36 -0500

---

This needs a similar patch as what Yanfei did for valgrind warnings.

http://git.mpich.org/mpich.git/commitdiff/b7db27e2b52595ce089f0cd81a1ed89e4e31056e

[wkliao]

I am wondering the status of this issue.

[raffenet]

This was fixed in 4b7d553.