On 8/7/2012 5:45 AM, Jeff Squyres wrote:
So the issue is when (for example) Fortran MPI_Recv says "hey, C ints are the same
as Fortran INEGERs, so I don't need a temporary MPI_Status buffer; I'll just use the
INTEGER array that I was given, and pass it to the back-end C MPI_Recv() routine."
Then C MPI_Recv() tries to write to the size_t variable, and it might be poorly aligned.
Kaboom.
Right. Kind of. Read on...
So, specifically, what I propose is getting rid of the short-cuts that try to
use Fortran statuses in-place if Fortran INTEGERs are as big as C ints. I can
make the changes. Sanity checks on all that are welcome.
Hmm. I'm not excited about this -- the whole point is that if we don't need to
do an extra copy, let's not do it.
Is there a better way to fix this?
Off the top of my head -- for example, could we change some of those
compile-time checks to run-time checks, and add in an alignment check? E.g.:
----
#if OMPI_SIZEOF_FORTRAN_INTEGER == SIZEOF_INT
c_status = (MPI_Status *) status;
#else
c_status =&c_status2;
#endif
-----
to
----
/* The constant checks will be resolved at compile time; assume
alignment_is_good() is an inline macro checking for "good" alignment
on platforms where alignment(int) != alignment(size_t) */
if (OMPI_SIZEOF_FORTRAN_INTEGER == SIZEOF_INT&&
alignment_is_good(status)) {
c_status = (MPI_Status *) status;
} else {
c_status =&c_status2;
}
-----
Would something like that work?
I'm thinking that the benefit here is that we only penalize platforms (with an extra
"if" statement) where alignment(int) != alignment(size_t).
I did quite a bit more poking around. It appears this issue is already
"well known." That is, due to this issue, we're not allowed to assume
that a status that the user passed to us (ompi/mpi/c layer) has proper
alignment since it might have come from Fortran. So, we should use
OMPI_STATUS_SET and OMPI_STATUS_SET_COUNT instead of doing direct status
assignments. Not only does ompi/request/request.h define these macros,
but it also gives a nice description of the issue and points us to trac
2526.
It seems to me there are a number of places where we do direct status
assignments. I found a couple of places in ompi/mca/pml/ob1/*.c and
quite a few more in ompi/mpi/c/*.c. If I'm sufficiently inspired
tomorrow, I might look around to see if I can identify other places to
look. I can also confirm this leads to failures -- not only the mprobe
stuff I reported but even vanilla MPI_Recv if you tweak the conditions
just right. I'll try to get to this stuff tomorrow.
