failure to dup communicator in pnetcdf

[Rob Latham]

I noticed last week that parallel-netcdf does not duplicate the
communicator passed to the library as part of ncmpi_open/nfmpi_open and
ncmpi_create/nfmpi_create.  

It's been 8 years and no one has complained, and a lot of the heavy
lifting is done in the MPI-IO library (which does duplicate the
communicator).  Still, I'm about to commit this patch unless someone
can remind me why we do not need it...


Index: src/lib/mpincio.c
===================================================================
--- src/lib/mpincio.c	(revision 923)
+++ src/lib/mpincio.c	(working copy)
@@ -136,7 +136,7 @@ ncmpiio_create(MPI_Comm     comm,
 
     nciop->mpiomode  = MPI_MODE_RDWR;
     nciop->mpioflags = 0;
-    nciop->comm      = comm;
+    MPI_Comm_dup(comm, &(nciop->comm));
 
     ncmpiio_extract_hints(nciop, info);
 
@@ -147,7 +147,7 @@ ncmpiio_create(MPI_Comm     comm,
         /* to avoid calling MPI_File_set_size() later, let process 0 check
            if the file exists. If not, no need to call MPI_File_set_size */
         int rank;
-        MPI_Comm_rank(comm, &rank);
+        MPI_Comm_rank(nciop->comm, &rank);
         if (rank == 0) { /* check if file exists */
             if (access(path, F_OK) == 0) { /* but is this only available in Linux? */
                 /* file does exist, so delete it */
@@ -164,7 +164,8 @@ ncmpiio_create(MPI_Comm     comm,
 #endif
     }
 
-    mpireturn = MPI_File_open(comm, (char *)path, mpiomode, info, &nciop->collective_fh);
+    mpireturn = MPI_File_open(nciop->comm, (char *)path, mpiomode, 
+            info, &nciop->collective_fh);
     if (mpireturn != MPI_SUCCESS) {
         int rank;
         MPI_Comm_rank(comm, &rank);
@@ -215,11 +216,12 @@ ncmpiio_open(MPI_Comm     comm,
  
     nciop->mpiomode  = mpiomode;
     nciop->mpioflags = 0;
-    nciop->comm      = comm;
+    MPI_Comm_dup(comm, &(nciop->comm));
  
     ncmpiio_extract_hints(nciop, info);
 
-    mpireturn = MPI_File_open(comm, (char *)path, mpiomode, info, &nciop->collective_fh);
+    mpireturn = MPI_File_open(nciop->comm, (char *)path, mpiomode, 
+            info, &nciop->collective_fh);
     if (mpireturn != MPI_SUCCESS) {
         int rank;
         MPI_Comm_rank(comm, &rank);
@@ -321,6 +323,9 @@ ncmpiio_close(ncio *nciop, int doUnlink) {
     MPI_Info_free(&(nciop->mpiinfo));
 #endif
 
+  if (nciop->comm != MPI_COMM_NULL) {
+      MPI_Comm_free(&(nciop->comm));
+  }
 
   ncmpiio_free(nciop);
 

-- 
Rob Latham
Mathematics and Computer Science Division
Argonne National Lab, IL USA