[mpich-discuss] segfault when calling MPI_Init

[Rajeev Thakur]

The C standard specifies that argv[argc] is a null pointer. MPIDI_PG_Init in
src/mpid/ch3/src/mpidi_pg.c assumed that. Someone sent in a patch to relax
that assumption to make it more tolerant of user error. We have included
that patch in the 1.0.7 rc. So the code shouldn't seg fault with 1.0.7. 

Rajeev


> -----Original Message-----
> From: owner-mpich-discuss at mcs.anl.gov 
> [mailto:owner-mpich-discuss at mcs.anl.gov] On Behalf Of Martin Siegert
> Sent: Thursday, April 03, 2008 10:53 PM
> To: mpich-discuss at mcs.anl.gov
> Subject: [mpich-discuss] segfault when calling MPI_Init
> 
> Hi,
> 
> I am trying to compile GAMESS (Ameslab) using the MPI interface
> and mpich2-1.0.6p1. I get it to compile just fine, but I get
> a segfault (signal 11) when it calls MPI_Init.
> 
> The main program is written in Fortran, but in order to process
> command line arguments that then can be passed to MPI_Init a
> C routine is used.
> Here are the relevant pieces from the code:
> 
> ddi_fortran.c:
> 
>    void F77_Init() {
>       int i,j,lenmax=256,argc=iargc_();
>       char **argv = NULL;
>       char arg[256];
> 
>       STD_DEBUG((stdout," DDI: Entering F77 DDI_Init.\n"))
> 
>    /* -------------------------- *\
>       Get command line arguments
>    \* -------------------------- */
>       if(argc) {
>          argc++;
>          argv = (char **)Malloc(argc*sizeof(char*));
>          for(i=0; i<argc; i++) {
>            for(j=0; j<256; j++) arg[j]=' ';
> 
>          # if defined CRAY
>              getarg_(&i,arg,&lenmax);
>          # else
>              getarg_(&i,arg,lenmax);
>          # endif
> 
>            for(j=0; j<256 && arg[j] != ' '; j++);
>            arg[j] = 0;
>            argv[i] = (char *) strdup(arg);
>          }
>       }
> 
>       MAX_DEBUG((stdout," DDI: Calling DDI_Init.\n"))
> 
>    /* -------------- *\
>       Initialize DDI
>    \* -------------- */
>       DDI_Init(argc,argv);
> 
>    }
> 
> ddi_init.c:
> 
>    void DDI_Init(int argc,char **argv) {
> ...
>     # if defined DDI_MPI
>       Init_mpi(argc,argv);
>     # endif
> ...
> 
>    static void Init_mpi(int targc,char *targv[]) {
> ...
>       int argc = targc;
>       char **argv = targv;
> ...
>    /* -------------- *\
>       Initialize MPI
>    \* -------------- */
>       if(MPI_Init(&argc,&argv) != MPI_SUCCESS) {
>          fprintf(stdout," DDI: MPI_Init failed.\n");
>          fflush(stdout); exit(911);
>       }
> 
> Thus the fortran program calls F77_Init which in turn uses getarg_i
> to get the command line arguments and then basically calls
> MPI_Init(&argc,&argv).
> The segmentation fault happens when MPI_Init is called.
> 
> I vaguely remember seeing a discussion (on this list?) about hidden
> arguments in MPI_Init (when called from Fortran?) but I failed to
> find anything in the mail archive of the list. Anyway, I played
> around somewhat with the code (this is pure guess work on my part)
> and changed it to:
> 
>       if(argc) {
>          argc++;
>          argv = (char **)Malloc((argc+1)*sizeof(char*));
>          for(i=0; i<argc; i++) {
>            for(j=0; j<256; j++) arg[j]=' ';
> 
>          # if defined CRAY
>              getarg_(&i,arg,&lenmax);
>          # else
>              getarg_(&i,arg,lenmax);
>          # endif
> 
>            for(j=0; j<256 && arg[j] != ' '; j++);
>            arg[j] = 0;
>            argv[i] = (char *) strdup(arg);
>          }
>       }
>       arg[0] = 0;
>       argv[argc] = (char *) strdup(arg);
> 
> i.e., added an additional array element to argv which I set to '\0'.
> After that the segmentation fault disappears.
> 
> Can somebody explain to me whether this is appropriate or am I totally
> off base here? Any hints about why I am getting the segmentation fault
> in the first place?
> 
> Cheers,
> Martin
> 
> -- 
> Martin Siegert
> Head, Research Computing
> WestGrid Site Lead
> Client and Research Services               phone: 778 782-4691
> Simon Fraser University                    fax:   778 782-4242
> Burnaby, British Columbia                  email: siegert at sfu.ca
> Canada  V5A 1S6
> 
>