[svn-r7754] Purpose:

    Code cleanup

Description:
    Straighten out more goofiness in the MPI code dealing with collective I/O
transfers - mostly make certain that a view is set if-and-only-if collective
I/O is occurring on raw data (and vice versa for views and independent I/O)

Platforms tested:
    FreeBSD 4.9 (sleipnir) w/parallel & FPHDF5
    too minor to repquire h5committest

5 files changed, 89 insertions, 57 deletions

diff --git a/src/H5Dio.c b/src/H5Dio.c
index 0fcc7ea..12d4dec 100644
--- a/src/H5Dio.c
+++ b/src/H5Dio.c
@@ -515,6 +515,7 @@ H5D_read(H5D_t *dataset, const H5T_t *mem_type, const H5S_t *mem_space,
     H5S_conv_t	*sconv=NULL;	        /*space conversion funcs*/
 #ifdef H5_HAVE_PARALLEL
     H5FD_mpio_xfer_t xfer_mode;         /*xfer_mode for this request */
+    hbool_t     use_par_opt_io=FALSE;   /* Whether the 'optimized' I/O routines with be parallel */
     hbool_t     xfer_mode_changed=FALSE;    /* Whether the transfer mode was changed */
 #endif /*H5_HAVE_PARALLEL*/
     H5P_genplist_t *dx_plist=NULL;      /* Data transfer property list */
@@ -637,11 +638,13 @@ H5D_read(H5D_t *dataset, const H5T_t *mem_type, const H5S_t *mem_space,
     } /* end switch */
 
     /* Get dataspace functions */
-    if (NULL==(sconv=H5S_find(mem_space, file_space, sconv_flags)))
+    if (NULL==(sconv=H5S_find(mem_space, file_space, sconv_flags, &use_par_opt_io)))
         HGOTO_ERROR (H5E_DATASET, H5E_UNSUPPORTED, FAIL, "unable to convert from file to memory data space")
 
 #ifdef H5_HAVE_PARALLEL
-    H5D_io_assist_mpio(dx_plist, xfer_mode, &xfer_mode_changed);
+    /* Don't reset the transfer mode if we can't or won't use it */
+    if(!use_par_opt_io || !H5T_path_noop(tpath))
+        H5D_io_assist_mpio(dx_plist, xfer_mode, &xfer_mode_changed);
 #endif /*H5_HAVE_PARALLEL*/
 
     /* Determine correct I/O routine to invoke */
@@ -737,6 +740,7 @@ H5D_write(H5D_t *dataset, const H5T_t *mem_type, const H5S_t *mem_space,
     H5S_conv_t	*sconv=NULL;		/*space conversion funcs*/
 #ifdef H5_HAVE_PARALLEL
     H5FD_mpio_xfer_t xfer_mode;         /*xfer_mode for this request */
+    hbool_t     use_par_opt_io=FALSE;   /* Whether the 'optimized' I/O routines with be parallel */
     hbool_t     xfer_mode_changed=FALSE;    /* Whether the transfer mode was changed */
 #endif /*H5_HAVE_PARALLEL*/
     H5P_genplist_t *dx_plist=NULL;      /* Data transfer property list */
@@ -857,11 +861,13 @@ H5D_write(H5D_t *dataset, const H5T_t *mem_type, const H5S_t *mem_space,
     } /* end switch */
 
     /* Get dataspace functions */
-    if (NULL==(sconv=H5S_find(mem_space, file_space, sconv_flags)))
+    if (NULL==(sconv=H5S_find(mem_space, file_space, sconv_flags, &use_par_opt_io)))
 	HGOTO_ERROR (H5E_DATASET, H5E_UNSUPPORTED, FAIL, "unable to convert from memory to file data space")
         
 #ifdef H5_HAVE_PARALLEL
-    H5D_io_assist_mpio(dx_plist, xfer_mode, &xfer_mode_changed);
+    /* Don't reset the transfer mode if we can't or won't use it */
+    if(!use_par_opt_io || !H5T_path_noop(tpath))
+        H5D_io_assist_mpio(dx_plist, xfer_mode, &xfer_mode_changed);
 #endif /*H5_HAVE_PARALLEL*/
 
     /* Determine correct I/O routine to invoke */
diff --git a/src/H5FDfphdf5.c b/src/H5FDfphdf5.c
index 59dbecf..577f51d 100644
--- a/src/H5FDfphdf5.c
+++ b/src/H5FDfphdf5.c
@@ -1161,7 +1161,6 @@ H5FD_fphdf5_read(H5FD_t *_file, H5FD_mem_t mem_type, hid_t dxpl_id,
 {
     H5FD_fphdf5_t      *file = (H5FD_fphdf5_t*)_file;
     MPI_Offset          mpi_off;
-    MPI_Offset          mpi_disp;
     MPI_Status          status;
     int                 mrc;
     MPI_Datatype        buf_type;
@@ -1216,6 +1215,9 @@ H5FD_fphdf5_read(H5FD_t *_file, H5FD_mem_t mem_type, hid_t dxpl_id,
             HGOTO_ERROR(H5E_PLIST, H5E_CANTGET, FAIL, "can't get MPI-I/O type property")
 
     if (use_view_this_time) {
+        /* Sanity check that views will only be used by collective I/O */
+        assert(xfer_mode==H5FD_MPIO_COLLECTIVE);
+
         /* Prepare for a full-blown xfer using btype, ftype, and disp */
         if (H5P_get(plist, H5FD_FPHDF5_XFER_MEM_MPI_TYPE_NAME, &buf_type) < 0)
             HGOTO_ERROR(H5E_PLIST, H5E_CANTGET, FAIL, "can't get MPI-I/O type property")
@@ -1223,26 +1225,27 @@ H5FD_fphdf5_read(H5FD_t *_file, H5FD_mem_t mem_type, hid_t dxpl_id,
         if (H5P_get(plist, H5FD_FPHDF5_XFER_FILE_MPI_TYPE_NAME, &file_type) < 0)
             HGOTO_ERROR(H5E_PLIST, H5E_CANTGET, FAIL, "can't get MPI-I/O type property")
 
+        /* Set the file view when we are using MPI derived types */
+        if ((mrc = MPI_File_set_view(file->f, mpi_off, MPI_BYTE,
+                                     file_type, H5FD_mpio_native,
+                                     file->info)) != MPI_SUCCESS)
+            HMPI_GOTO_ERROR(FAIL, "MPI_File_set_view failed", mrc)
+
         /*
          * When using types, use the address as the displacement for
          * MPI_File_set_view and reset the address for the read to zero
          */
-        mpi_disp = mpi_off;
         mpi_off = 0;
-
-        /* Set the file view when we are using MPI derived types */
-        if ((mrc = MPI_File_set_view(file->f, (MPI_Offset)mpi_disp, MPI_BYTE,
-                                     file_type, H5FD_mpio_native,
-                                     file->info)) != MPI_SUCCESS)
-            HMPI_GOTO_ERROR(FAIL, "MPI_File_set_view failed", mrc)
     } else {
+        /* Sanity check that independent I/O must be occuring */
+        assert(xfer_mode==H5FD_MPIO_INDEPENDENT);
+
         /*
          * Prepare for a simple xfer of a contiguous block of bytes. The
          * btype, ftype, and disp fields are not used.
          */
         buf_type = MPI_BYTE;
         file_type = MPI_BYTE;
-        mpi_disp = 0;   /* mpi_off is already set */
     }
 
     /* If metadata, check the metadata cache first */
@@ -1271,9 +1274,7 @@ HDfprintf(stderr, "%s:%d: Metadata cache read failed!\n", FUNC, __LINE__);
     }
 
     /* Read the data. */
-    assert(xfer_mode == H5FD_MPIO_INDEPENDENT || xfer_mode == H5FD_MPIO_COLLECTIVE);
-
-    if (xfer_mode == H5FD_MPIO_INDEPENDENT) {
+    if (!use_view_this_time) {
         if ((mrc = MPI_File_read_at(file->f, mpi_off, buf, size_i,
                                     buf_type, &status)) != MPI_SUCCESS)
             HMPI_GOTO_ERROR(FAIL, "MPI_File_read_at failed", mrc)
@@ -1481,7 +1482,8 @@ H5FD_fphdf5_write_real(H5FD_t *_file, hid_t dxpl_id, haddr_t addr, int size,
 
 
     if (use_view_this_time) {
-        MPI_Offset          mpi_disp;
+        /* Sanity check that views will only be used by collective I/O */
+        assert(xfer_mode==H5FD_MPIO_COLLECTIVE);
 
         /* Prepare for a full-blown xfer using btype, ftype, and disp */
         if (H5P_get(plist, H5FD_FPHDF5_XFER_MEM_MPI_TYPE_NAME, &buf_type) < 0)
@@ -1490,19 +1492,21 @@ H5FD_fphdf5_write_real(H5FD_t *_file, hid_t dxpl_id, haddr_t addr, int size,
         if (H5P_get(plist, H5FD_FPHDF5_XFER_FILE_MPI_TYPE_NAME, &file_type) < 0)
             HGOTO_ERROR(H5E_PLIST, H5E_CANTGET, FAIL, "can't get MPI-I/O type property")
 
+        /* Set the file view when we are using MPI derived types */
+        if ((mrc = MPI_File_set_view(file->f, mpi_off, MPI_BYTE,
+                                     file_type, H5FD_mpio_native,
+                                     file->info)) != MPI_SUCCESS)
+            HMPI_GOTO_ERROR(FAIL, "MPI_File_set_view failed", mrc)
+
         /*
          * When using types, use the address as the displacement for
          * MPI_File_set_view and reset the address for the read to zero
          */
-        mpi_disp = mpi_off;
         mpi_off = 0;
-
-        /* Set the file view when we are using MPI derived types */
-        if ((mrc = MPI_File_set_view(file->f, mpi_disp, MPI_BYTE,
-                                     file_type, H5FD_mpio_native,
-                                     file->info)) != MPI_SUCCESS)
-            HMPI_GOTO_ERROR(FAIL, "MPI_File_set_view failed", mrc)
     } else {
+        /* Sanity check that independent I/O must be occuring */
+        assert(xfer_mode==H5FD_MPIO_INDEPENDENT);
+
         /*
          * Prepare for a simple xfer of a contiguous block of bytes. The
          * btype and ftype.
@@ -1517,9 +1521,7 @@ H5FD_fphdf5_write_real(H5FD_t *_file, hid_t dxpl_id, haddr_t addr, int size,
         xfer_mode = H5P_peek_unsigned(plist, H5D_XFER_IO_XFER_MODE_NAME);
  
     /* Write the data. */
-    assert(xfer_mode == H5FD_MPIO_INDEPENDENT || xfer_mode == H5FD_MPIO_COLLECTIVE);
-
-    if (xfer_mode == H5FD_MPIO_INDEPENDENT) {
+    if (!use_view_this_time) {
         /*OKAY: CAST DISCARDS CONST QUALIFIER*/
         if ((mrc = MPI_File_write_at(file->f, mpi_off, (void*)buf,
                                      size, buf_type, &status)) != MPI_SUCCESS)
diff --git a/src/H5FDmpio.c b/src/H5FDmpio.c
index 5bd714b..4fefbde 100644
--- a/src/H5FDmpio.c
+++ b/src/H5FDmpio.c
@@ -1470,7 +1470,7 @@ H5FD_mpio_read(H5FD_t *_file, H5FD_mem_t UNUSED type, hid_t dxpl_id, haddr_t add
 	       void *buf/*out*/)
 {
     H5FD_mpio_t			*file = (H5FD_mpio_t*)_file;
-    MPI_Offset			mpi_off, mpi_disp;
+    MPI_Offset			mpi_off;
     MPI_Status  		mpi_stat;
     int				mpi_code;	/* mpi return code */
     MPI_Datatype		buf_type, file_type;
@@ -1528,40 +1528,41 @@ H5FD_mpio_read(H5FD_t *_file, H5FD_mem_t UNUSED type, hid_t dxpl_id, haddr_t add
             HGOTO_ERROR(H5E_PLIST, H5E_CANTGET, FAIL, "can't get MPI-I/O type property")
 
     if (use_view_this_time) {
+        /* Sanity check that views will only be used by collective I/O */
+        assert(xfer_mode==H5FD_MPIO_COLLECTIVE);
+
         /* prepare for a full-blown xfer using btype, ftype, and disp */
         if(H5P_get(plist,H5FD_MPIO_XFER_MEM_MPI_TYPE_NAME,&buf_type)<0)
             HGOTO_ERROR(H5E_PLIST, H5E_CANTGET, FAIL, "can't get MPI-I/O type property")
         if(H5P_get(plist,H5FD_MPIO_XFER_FILE_MPI_TYPE_NAME,&file_type)<0)
             HGOTO_ERROR(H5E_PLIST, H5E_CANTGET, FAIL, "can't get MPI-I/O type property")
 
+        /*
+         * Set the file view when we are using MPI derived types
+         */
+        /*OKAY: CAST DISCARDS CONST QUALIFIER*/
+        if (MPI_SUCCESS != (mpi_code=MPI_File_set_view(file->f, mpi_off, MPI_BYTE, file_type, (char*)"native",  file->info)))
+            HMPI_GOTO_ERROR(FAIL, "MPI_File_set_view failed", mpi_code)
+
         /* When using types, use the address as the displacement for
          * MPI_File_set_view and reset the address for the read to zero
          */
-        mpi_disp=mpi_off;
         mpi_off=0;
     } /* end if */
     else {
+        /* Sanity check that independent I/O must be occuring */
+        assert(xfer_mode==H5FD_MPIO_INDEPENDENT);
+
         /*
          * Prepare for a simple xfer of a contiguous block of bytes. The
          * btype, ftype, and disp fields are not used.
          */
         buf_type = MPI_BYTE;
         file_type = MPI_BYTE;
-        mpi_disp = 0;		/* mpi_off is alread set */
     } /* end else */
 
-    /*
-     * Set the file view when we are using MPI derived types
-     */
-    if (use_view_this_time) {
-        /*OKAY: CAST DISCARDS CONST QUALIFIER*/
-        if (MPI_SUCCESS != (mpi_code=MPI_File_set_view(file->f, mpi_disp, MPI_BYTE, file_type, (char*)"native",  file->info)))
-            HMPI_GOTO_ERROR(FAIL, "MPI_File_set_view failed", mpi_code)
-    } /* end if */
-    
     /* Read the data. */
-    assert(H5FD_MPIO_INDEPENDENT==xfer_mode || H5FD_MPIO_COLLECTIVE==xfer_mode);
-    if (H5FD_MPIO_INDEPENDENT==xfer_mode) {
+    if (!use_view_this_time) {
         if (MPI_SUCCESS!= (mpi_code=MPI_File_read_at(file->f, mpi_off, buf, size_i, buf_type, &mpi_stat)))
             HMPI_GOTO_ERROR(FAIL, "MPI_File_read_at failed", mpi_code)
     } else {
@@ -1776,7 +1777,7 @@ H5FD_mpio_write(H5FD_t *_file, H5FD_mem_t type, hid_t dxpl_id, haddr_t addr,
 		size_t size, const void *buf)
 {
     H5FD_mpio_t			*file = (H5FD_mpio_t*)_file;
-    MPI_Offset 		 	mpi_off, mpi_disp;
+    MPI_Offset 		 	mpi_off;
     MPI_Status			mpi_stat;
     MPI_Datatype		buf_type, file_type;
     int			        mpi_code;	/* MPI return code */
@@ -1835,37 +1836,39 @@ H5FD_mpio_write(H5FD_t *_file, H5FD_mem_t type, hid_t dxpl_id, haddr_t addr,
             HGOTO_ERROR(H5E_PLIST, H5E_CANTGET, FAIL, "can't get MPI-I/O type property")
 
     if (use_view_this_time) {
+        /* Sanity check that views will only be used by collective I/O */
+        assert(xfer_mode==H5FD_MPIO_COLLECTIVE);
+
         /* prepare for a full-blown xfer using btype, ftype, and disp */
         if(H5P_get(plist,H5FD_MPIO_XFER_MEM_MPI_TYPE_NAME,&buf_type)<0)
             HGOTO_ERROR(H5E_PLIST, H5E_CANTGET, FAIL, "can't get MPI-I/O type property")
         if(H5P_get(plist,H5FD_MPIO_XFER_FILE_MPI_TYPE_NAME,&file_type)<0)
             HGOTO_ERROR(H5E_PLIST, H5E_CANTGET, FAIL, "can't get MPI-I/O type property")
 
+        /*
+         * Set the file view when we are using MPI derived types
+         */
+        /*OKAY: CAST DISCARDS CONST QUALIFIER*/
+        if (MPI_SUCCESS != (mpi_code=MPI_File_set_view(file->f, mpi_off, MPI_BYTE, file_type, (char*)"native", file->info)))
+            HMPI_GOTO_ERROR(FAIL, "MPI_File_set_view failed", mpi_code)
+
         /* When using types, use the address as the displacement for
          * MPI_File_set_view and reset the address for the read to zero
          */
-        mpi_disp=mpi_off;
         mpi_off=0;
     } /* end if */
     else {
+        /* Sanity check that independent I/O must occur */
+        assert(xfer_mode==H5FD_MPIO_INDEPENDENT);
+
         /*
          * Prepare for a simple xfer of a contiguous block of bytes.
          * The btype, ftype, and disp fields are not used.
          */
         buf_type = MPI_BYTE;
         file_type = MPI_BYTE;
-        mpi_disp = 0;		/* mpi_off is already set */
     } /* end else */
 
-    /*
-     * Set the file view when we are using MPI derived types
-     */
-    if (use_view_this_time) {
-        /*OKAY: CAST DISCARDS CONST QUALIFIER*/
-        if (MPI_SUCCESS != (mpi_code=MPI_File_set_view(file->f, mpi_disp, MPI_BYTE, file_type, (char*)"native", file->info)))
-            HMPI_GOTO_ERROR(FAIL, "MPI_File_set_view failed", mpi_code)
-    } /* end if */
-    
     /* Metadata specific actions */
     if(type!=H5FD_MEM_DRAW) {
         /* Check if we need to syncronize all processes before attempting metadata write
@@ -1873,6 +1876,10 @@ H5FD_mpio_write(H5FD_t *_file, H5FD_mem_t type, hid_t dxpl_id, haddr_t addr,
          * and writes the metadata to the file before all the processes have
          * read the data, "transmitting" data from the "future" to the reading
          * process. -QAK )
+         *
+         * The only time we don't want to block before a metdata write is when
+         * we are flushing out a bunch of metadata.  Then, we block before the
+         * first write and don't block for further writes in the sequence.
          */
         if(H5P_exist_plist(plist,H5AC_BLOCK_BEFORE_META_WRITE_NAME)>0)
             if(H5P_get(plist,H5AC_BLOCK_BEFORE_META_WRITE_NAME,&block_before_meta_write)<0)
@@ -1898,8 +1905,7 @@ H5FD_mpio_write(H5FD_t *_file, H5FD_mem_t type, hid_t dxpl_id, haddr_t addr,
     } /* end if */
 
     /* Write the data. */
-    assert(H5FD_MPIO_INDEPENDENT==xfer_mode || H5FD_MPIO_COLLECTIVE==xfer_mode);
-    if (H5FD_MPIO_INDEPENDENT==xfer_mode) {
+    if (!use_view_this_time) {
         /*OKAY: CAST DISCARDS CONST QUALIFIER*/
         if (MPI_SUCCESS != (mpi_code=MPI_File_write_at(file->f, mpi_off, (void*)buf, size_i, buf_type, &mpi_stat)))
             HMPI_GOTO_ERROR(FAIL, "MPI_File_write_at failed", mpi_code)
diff --git a/src/H5S.c b/src/H5S.c
index f69f8f5..e1d0327 100644
--- a/src/H5S.c
+++ b/src/H5S.c
@@ -1478,7 +1478,11 @@ H5S_find (const H5S_t *mem_space, const H5S_t *file_space, unsigned
 #ifndef H5_HAVE_PARALLEL
 UNUSED
 #endif /* H5_HAVE_PARALLEL */
-flags)
+flags, hbool_t
+#ifndef H5_HAVE_PARALLEL
+UNUSED
+#endif /* H5_HAVE_PARALLEL */
+*use_par_opt_io)
 {
     H5S_conv_t	*path=NULL;  /* Space conversion path */
 #ifdef H5_HAVE_PARALLEL
@@ -1521,10 +1525,17 @@ flags)
 
             /* Check if we can use the optimized parallel I/O routines */
             if(opt==TRUE) {
+                /* Set the pointers to the MPI-specific routines */
                 H5S_conv_g[i]->read = H5S_mpio_spaces_read;
                 H5S_conv_g[i]->write = H5S_mpio_spaces_write;
+
+                /* Indicate that the I/O will be parallel */
+                *use_par_opt_io=TRUE;
             } /* end if */
             else {
+                /* Indicate that the I/O will _NOT_ be parallel */
+                *use_par_opt_io=FALSE;
+
 #endif /* H5_HAVE_PARALLEL */
                 H5S_conv_g[i]->read = H5S_select_read;
                 H5S_conv_g[i]->write = H5S_select_write;
@@ -1556,10 +1567,17 @@ flags)
 
     /* Check if we can use the optimized parallel I/O routines */
     if(opt==TRUE) {
+        /* Set the pointers to the MPI-specific routines */
         path->read = H5S_mpio_spaces_read;
         path->write = H5S_mpio_spaces_write;
+
+        /* Indicate that the I/O will be parallel */
+        *use_par_opt_io=TRUE;
     } /* end if */
     else {
+        /* Indicate that the I/O will _NOT_ be parallel */
+        *use_par_opt_io=FALSE;
+
 #endif /* H5_HAVE_PARALLEL */
         path->read = H5S_select_read;
         path->write = H5S_select_write;
diff --git a/src/H5Sprivate.h b/src/H5Sprivate.h
index 389b3b1..46d302f 100644
--- a/src/H5Sprivate.h
+++ b/src/H5Sprivate.h
@@ -164,7 +164,7 @@ typedef struct H5S_conv_t {
 H5_DLL H5S_t *H5S_copy(const H5S_t *src);
 H5_DLL herr_t H5S_close(H5S_t *ds);
 H5_DLL H5S_conv_t *H5S_find(const H5S_t *mem_space, const H5S_t *file_space,
-                unsigned flags);
+                unsigned flags, hbool_t *use_par_opt_io);
 H5_DLL H5S_class_t H5S_get_simple_extent_type(const H5S_t *ds);
 H5_DLL hssize_t H5S_get_simple_extent_npoints(const H5S_t *ds);
 H5_DLL hsize_t H5S_get_npoints_max(const H5S_t *ds);