[svn-r438] Changes for PHDF5 to support extendable datasets.

H5D.c:
    If a dataset, using chunk storage, is created/opened and if its data
    file is accessed via MPIO, allocate file space for the chunks immediately.
    Also do the allocation when a dataset is extended.  Must do it at these
    points because H5Dcreate, H5Dopen and H5Dextend are collective calls
    in MPIO access mode and are the only chance to get all MPI processes
    to synchonize their meta-data.
    Also disallow opening/creating datasets with compression option if
    MPIO is used.  Not easy to support concurrent compression writes in
    an efficient way.  But read-only access should be fine.  Will implement
    that later.

H5F.c:
    Disable file cache if MPIO is used by forcing cache size and number
    of elements to zero.

H5Fistore.c:
    Added two new routines, H5F_istore_get_addr and H5F_istore_allocate.
    H5F_istore_allocate supports allocating file space for the chunks
    of a dataset.  H5F_istore_get_addr returns the file space address
    of a chunk that has been allocated.
    Changed H5F_istore_read and H5F_istore_write to bypass the chunk
    cache read/write mechanism.  Instead, they do read/write directly
    with the file for only the elements requested.
    (Some debug statements, enclosed by #ifdef AKC, are still in the
    file.  I am still working on some code.  They would be cleaned
    out soon.)

1 files changed, 288 insertions, 1 deletions

diff --git a/src/H5Fistore.c b/src/H5Fistore.c
index 4cb808b..a97e542 100644
--- a/src/H5Fistore.c
+++ b/src/H5Fistore.c
@@ -398,7 +398,9 @@ H5F_istore_new_node(H5F_t *f, H5B_ins_t op,
     intn		i;
 
     FUNC_ENTER(H5F_istore_new_node, FAIL);
-
+#ifdef AKC
+printf("%s: Called\n", FUNC);
+#endif
     /* check args */
     assert(f);
     assert(lt_key);
@@ -409,6 +411,9 @@ H5F_istore_new_node(H5F_t *f, H5B_ins_t op,
 
     /* Allocate new storage */
     assert (udata->key.nbytes > 0);
+#ifdef AKC
+printf("calling H5MF_alloc for new chunk\n");
+#endif
     if (H5MF_alloc(f, H5MF_RAW, udata->key.nbytes, addr /*out */ ) < 0) {
 	HRETURN_ERROR(H5E_IO, H5E_CANTINIT, FAIL,
 		      "couldn't allocate new file storage");
@@ -539,6 +544,9 @@ H5F_istore_insert(H5F_t *f, const haddr_t *addr, void *_lt_key,
     H5B_ins_t		ret_value = H5B_INS_ERROR;
 
     FUNC_ENTER(H5F_istore_insert, H5B_INS_ERROR);
+#ifdef AKC
+printf("%s: Called\n", FUNC);
+#endif
 
     /* check args */
     assert(f);
@@ -568,6 +576,9 @@ H5F_istore_insert(H5F_t *f, const haddr_t *addr, void *_lt_key,
 	 * then we should reallocate storage.
 	 */
 	if (lt_key->nbytes != udata->key.nbytes) {
+#ifdef AKC
+printf("calling H5MF_realloc for new chunk\n");
+#endif
 	    if (H5MF_realloc (f, H5MF_RAW, lt_key->nbytes, addr,
 			      udata->key.nbytes, new_node/*out*/)<0) {
 		HRETURN_ERROR (H5E_STORAGE, H5E_WRITEERROR, H5B_INS_ERROR,
@@ -601,6 +612,9 @@ H5F_istore_insert(H5F_t *f, const haddr_t *addr, void *_lt_key,
 	/*
 	 * Allocate storage for the new chunk
 	 */
+#ifdef AKC
+printf("calling H5MF_alloc for new chunk\n");
+#endif
 	if (H5MF_alloc(f, H5MF_RAW, udata->key.nbytes, new_node/*out*/)<0) {
 	    HRETURN_ERROR(H5E_IO, H5E_CANTINIT, H5B_INS_ERROR,
 			  "file allocation failed");
@@ -1329,6 +1343,49 @@ H5F_istore_read(H5F_t *f, const H5O_layout_t *layout,
 	    sub_offset_m[i] = chunk_offset[i] + offset_wrt_chunk[i] +
 			      offset_m[i] - offset_f[i];
 	}
+#ifdef HAVE_PARALLEL
+	/*
+	 * If MPIO is used, must bypass the chunk-cache scheme
+	 * because other MPI processes could be writing to other
+	 * elements in the same chunk.
+	 * Do a direct write-through of only the elements requested.
+	 */
+	 if (f->shared->access_parms->driver==H5F_LOW_MPIO){
+	    H5F_istore_ud1_t	udata;
+	    H5O_layout_t	l;	/* temporary layout */
+	    if (H5F_istore_get_addr(f, layout, chunk_offset, &udata)==FAIL){
+		HRETURN_ERROR (H5E_IO, H5E_WRITEERROR, FAIL,
+				"unable to locate raw data chunk");
+	    };
+	    /*
+	     * use default transfer mode as we do not support collective
+	     * transfer mode since each data write could decompose into
+	     * multiple chunk writes and we are not doing the calculation
+	     * yet.
+	     */
+	    l.type = H5D_CONTIGUOUS;
+	    l.ndims = layout->ndims;
+	    for (i=l.ndims; i-- > 0;)
+		l.dim[i] = layout->dim[i];
+	    l.addr = udata.addr;
+	    if (H5F_arr_read(f, &l,
+			    comp, NULL /* no efl */,
+			    sub_size, size_m,
+			    sub_offset_m, offset_wrt_chunk,
+			    H5D_XFER_DFLT, buf)==FAIL){
+		HRETURN_ERROR (H5E_IO, H5E_READERROR, FAIL,
+			     "unable to read raw data from file");
+	    };
+	}else{
+#endif
+
+#ifdef AKC
+printf("Locking chunk( ");
+for (i=0; i<layout->ndims; i++){
+printf("%ld ", chunk_offset[i]);
+}
+printf(")\n");
+#endif
 
 	/*
 	 * Lock the chunk, transfer data to the application, then unlock the
@@ -1346,6 +1403,9 @@ H5F_istore_read(H5F_t *f, const H5O_layout_t *layout,
 	    HRETURN_ERROR (H5E_IO, H5E_READERROR, FAIL,
 			   "unable to unlock raw data chunk");
 	}
+#ifdef HAVE_PARALLEL
+	}
+#endif
 
 	/* Increment indices */
 	for (i=layout->ndims-1, carry=1; i>=0 && carry; --i) {
@@ -1458,6 +1518,49 @@ H5F_istore_write(H5F_t *f, const H5O_layout_t *layout,
 			      offset_m[i] - offset_f[i];
 	}
 
+#ifdef HAVE_PARALLEL
+	/*
+	 * If MPIO is used, must bypass the chunk-cache scheme
+	 * because other MPI processes could be writing to other
+	 * elements in the same chunk.
+	 * Do a direct write-through of only the elements requested.
+	 */
+	 if (f->shared->access_parms->driver==H5F_LOW_MPIO){
+	    H5F_istore_ud1_t	udata;
+	    H5O_layout_t	l;	/* temporary layout */
+	    if (H5F_istore_get_addr(f, layout, chunk_offset, &udata)==FAIL){
+		HRETURN_ERROR (H5E_IO, H5E_WRITEERROR, FAIL,
+				"unable to locate raw data chunk");
+	    };
+	    /*
+	     * use default transfer mode as we do not support collective
+	     * transfer mode since each data write could decompose into
+	     * multiple chunk writes and we are not doing the calculation
+	     * yet.
+	     */
+	    l.type = H5D_CONTIGUOUS;
+	    l.ndims = layout->ndims;
+	    for (i=l.ndims; i-- > 0;)
+		l.dim[i] = layout->dim[i];
+	    l.addr = udata.addr;
+	    if (H5F_arr_write(f, &l,
+			    comp, NULL /* no efl */,
+			    sub_size, size_m,
+			    sub_offset_m, offset_wrt_chunk,
+			    H5D_XFER_DFLT, buf)==FAIL){
+		HRETURN_ERROR (H5E_IO, H5E_WRITEERROR, FAIL,
+			     "unable to write raw data to file");
+	    };
+	}else{
+#endif
+
+#ifdef AKC
+printf("Locking chunk( ");
+for (i=0; i<layout->ndims; i++){
+printf("%ld ", chunk_offset[i]);
+}
+printf(")\n");
+#endif
 	/*
 	 * Lock the chunk, copy from application to chunk, then unlock the
 	 * chunk.
@@ -1475,6 +1578,9 @@ H5F_istore_write(H5F_t *f, const H5O_layout_t *layout,
 	    HRETURN_ERROR (H5E_IO, H5E_WRITEERROR, FAIL,
 			   "uanble to unlock raw data chunk");
 	}
+#ifdef HAVE_PARALLEL
+	}
+#endif
 	
 	/* Increment indices */
 	for (i=layout->ndims-1, carry=1; i>=0 && carry; --i) {
@@ -1625,3 +1731,184 @@ H5F_istore_debug(H5F_t *f, const haddr_t *addr, FILE * stream, intn indent,
 
     FUNC_LEAVE (SUCCEED);
 }
+
+
+#ifdef HAVE_PARALLEL
+/*-------------------------------------------------------------------------
+ * Function:	H5F_istore_get_addr
+ *
+ * Purpose:	Get the file address of a chunk if file space has been
+ *		assigned.  Save the retrieved information in the udata
+ *		supplied.
+ *
+ * Return:	Success:	SUCCEED
+ *
+ *		Failure:	FAIL
+ *
+ * Programmer:	Albert Cheng
+ *              June 27, 1998
+ *
+ * Modifications:
+ *
+ *-------------------------------------------------------------------------
+ */
+herr_t
+H5F_istore_get_addr (H5F_t *f, const H5O_layout_t *layout,
+		 const hssize_t offset[], H5F_istore_ud1_t *udata/*out*/)
+{
+    intn		i;
+    herr_t		status;			/*func return status	*/
+    herr_t		ret_value=FAIL;		/*return value		*/
+    
+    FUNC_ENTER (H5F_istore_get_addr, NULL);
+
+    assert(f);
+    assert(layout && (layout->ndims > 0));
+    assert(offset);
+    assert(udata);
+
+
+    for (i=0; i<layout->ndims; i++) {
+	udata->key.offset[i] = offset[i];
+    }
+    udata->mesg = *layout;
+    H5F_addr_undef (&(udata->addr));
+    status = H5B_find (f, H5B_ISTORE, &(layout->addr), udata);
+    H5E_clear ();
+    if (status>=0 && H5F_addr_defined (&(udata->addr)))
+	HRETURN(SUCCEED);
+
+    FUNC_LEAVE (FAIL);
+}
+
+
+/*-------------------------------------------------------------------------
+ * Function:	H5F_istore_allocate
+ *
+ * Purpose:	Allocate file space for all chunks that are not allocated yet.
+ *		Return SUCCEED if all needed allocation succeed, otherwise
+ *		FAIL.
+ *
+ * Return:	Success:	SUCCEED
+ *
+ *		Failure:	FAIL
+ *
+ * Note:	Current implementation relies on cache_size being 0,
+ *		thus no chunk is cashed and written to disk immediately
+ *		when a chunk is unlocked (via H5F_istore_unlock)
+ *		This should be changed to do a direct flush independent
+ *		of the cache value.
+ *
+ * Programmer:	Albert Cheng
+ *		June 26, 1998
+ *
+ * Modifications:
+ *
+ *-------------------------------------------------------------------------
+ */
+herr_t
+H5F_istore_allocate (H5D_t *dataset, H5F_t *f, const H5O_layout_t *layout,
+		const H5S_t *space, const H5O_compress_t *comp)
+{
+    hsize_t		space_dim[H5O_LAYOUT_NDIMS];
+    intn		space_ndims;
+
+    intn		i, carry;
+    hssize_t		chunk_offset[H5O_LAYOUT_NDIMS];
+    uint8		*chunk=NULL;
+    intn		idx_hint=0;
+    size_t		chunk_size;
+    H5F_istore_ud1_t	udata;
+    
+    FUNC_ENTER(H5F_istore_allocate, FAIL);
+#ifdef AKC
+printf("Enter %s:\n", FUNC);
+#endif
+
+    /* Check args */
+    assert(dataset);
+    assert(layout);
+    assert(f);
+    assert(space);
+
+    if (layout->type != H5D_CHUNKED)
+	HRETURN(SUCCEED);		/*nothing to do or should we FAIL? */
+
+    assert(layout->ndims>0 && layout->ndims<=H5O_LAYOUT_NDIMS);
+    assert(H5F_addr_defined(&(layout->addr)));
+
+    /* get current dims of dataset */
+    if ((space_ndims=H5S_get_dims(space, space_dim, NULL)) <= 0 ||
+	space_ndims+1 != layout->ndims){
+	HRETURN_ERROR(H5E_DATASET, H5E_CANTINIT, NULL,
+		    "unable to allocate chunk storage");
+    }
+    /* copy the element size over */
+    space_dim[space_ndims] = layout->dim[space_ndims];
+
+    /*
+     * Setup indice to go through all chunks. (Future improvement
+     * should allocate only chunks that have no file space assigned yet.
+     */
+    for (i=0, chunk_size=1; i<layout->ndims; i++) {
+	chunk_offset[i]=0;
+	chunk_size *= layout->dim[i];
+    }
+
+    /* Loop over all chunks */
+    while (1) {
+	
+#ifdef AKC
+printf("Checking allocation for chunk( ");
+for (i=0; i<layout->ndims; i++){
+printf("%ld ", chunk_offset[i]);
+}
+printf(")\n");
+#endif
+#ifdef NO
+	if (H5F_istore_get_addr(f, layout, chunk_offset, &udata)==FAIL){
+#endif
+	    /* No file space assigned yet.  Allocate it. */
+	    /* The following needs improvement like calling the */
+	    /* allocation directly rather than indirectly using the */
+	    /* allocation effect in the unlock process. */
+
+#ifdef AKC
+printf("need allocation\n");
+#endif
+	    /*
+	     * Lock the chunk, copy from application to chunk, then unlock the
+	     * chunk.
+	     */
+	    if (NULL==(chunk=H5F_istore_lock (f, layout, comp, chunk_offset,
+					      FALSE, &idx_hint))) {
+		HRETURN_ERROR (H5E_IO, H5E_WRITEERROR, FAIL,
+			       "unable to read raw data chunk");
+	    }
+	    if (H5F_istore_unlock (f, layout, comp, TRUE, chunk_offset, &idx_hint,
+				   chunk, chunk_size)<0) {
+		HRETURN_ERROR (H5E_IO, H5E_WRITEERROR, FAIL,
+			       "uanble to unlock raw data chunk");
+	    }
+#ifdef NO
+	}
+	else{
+#ifdef AKC
+printf("NO need for allocation\n");
+printf("udata.addr.offset=%d\n", udata.addr.offset);
+#endif
+	}
+#endif
+	
+	/* Increment indices */
+	for (i=layout->ndims-1, carry=1; i>=0 && carry; --i) {
+	    chunk_offset[i] += layout->dim[i];
+	    if (chunk_offset[i] >= space_dim[i]) chunk_offset[i] = 0;
+	    else carry = 0;
+	}
+	if (carry) break;
+    }
+
+    FUNC_LEAVE(SUCCEED);
+}
+#endif