[svn-r13997] Moved H5D_istore_lock and H5D_istore_unlock from H5D_istore_writevv and H5D_istore_readvv to

H5D_chunk_write and H5D_chunk_read to avoid frequent lock and unlock and to improve some
performance.

Tested with h5committest on THG machines.

8 files changed, 301 insertions, 179 deletions

diff --git a/src/H5Dcompact.c b/src/H5Dcompact.c
index d18eef9..2e4d3f7 100644
--- a/src/H5Dcompact.c
+++ b/src/H5Dcompact.c
@@ -139,7 +139,7 @@ ssize_t
 H5D_compact_readvv(const H5D_io_info_t *io_info,
     size_t dset_max_nseq, size_t *dset_curr_seq, size_t dset_size_arr[], hsize_t dset_offset_arr[],
     size_t mem_max_nseq, size_t *mem_curr_seq, size_t mem_size_arr[], hsize_t mem_offset_arr[],
-    void *buf)
+    haddr_t UNUSED addr, void UNUSED *pointer/*in*/, void *buf)
 {
     ssize_t ret_value;          /* Return value */
 
@@ -181,7 +181,7 @@ ssize_t
 H5D_compact_writevv(const H5D_io_info_t *io_info,
     size_t dset_max_nseq, size_t *dset_curr_seq, size_t dset_size_arr[], hsize_t dset_offset_arr[],
     size_t mem_max_nseq, size_t *mem_curr_seq, size_t mem_size_arr[], hsize_t mem_offset_arr[],
-    const void *buf)
+    haddr_t UNUSED addr, void UNUSED *pointer/*in*/, const void *buf)
 {
     ssize_t ret_value;          /* Return value */
 
diff --git a/src/H5Dcontig.c b/src/H5Dcontig.c
index c83a88b..e1a02cb 100644
--- a/src/H5Dcontig.c
+++ b/src/H5Dcontig.c
@@ -359,8 +359,8 @@ H5D_contig_write(H5D_t *dset, const H5D_dxpl_cache_t *dxpl_cache,
     assert (buf);
 
     H5D_BUILD_IO_INFO(&io_info,dset,dxpl_cache,dxpl_id,store);
-    if(H5D_contig_writevv(&io_info, (size_t)1, &dset_curr_seq, &dset_len,
-            &dset_off, (size_t)1, &mem_curr_seq, &mem_len, &mem_off, buf) < 0)
+    if(H5D_contig_writevv(&io_info, (size_t)1, &dset_curr_seq, &dset_len, &dset_off, 
+            (size_t)1, &mem_curr_seq, &mem_len, &mem_off, (haddr_t)0, NULL, buf) < 0)
         HGOTO_ERROR(H5E_IO, H5E_WRITEERROR, FAIL, "vector write failed")
 
 done:
@@ -390,7 +390,7 @@ ssize_t
 H5D_contig_readvv(const H5D_io_info_t *io_info,
     size_t dset_max_nseq, size_t *dset_curr_seq, size_t dset_len_arr[], hsize_t dset_offset_arr[],
     size_t mem_max_nseq, size_t *mem_curr_seq, size_t mem_len_arr[], hsize_t mem_offset_arr[],
-    void *_buf)
+    haddr_t UNUSED address, void UNUSED *pointer, void *_buf)
 {
     H5F_t *file = io_info->dset->oloc.file;        /* File for dataset */
     H5D_rdcdc_t *dset_contig=&(io_info->dset->shared->cache.contig); /* Cached information about contiguous data */
@@ -655,7 +655,7 @@ ssize_t
 H5D_contig_writevv(const H5D_io_info_t *io_info,
     size_t dset_max_nseq, size_t *dset_curr_seq, size_t dset_len_arr[], hsize_t dset_offset_arr[],
     size_t mem_max_nseq, size_t *mem_curr_seq, size_t mem_len_arr[], hsize_t mem_offset_arr[],
-    const void *_buf)
+    haddr_t UNUSED address, void UNUSED *pointer, const void *_buf)
 {
     H5F_t *file = io_info->dset->oloc.file;        /* File for dataset */
     H5D_rdcdc_t *dset_contig=&(io_info->dset->shared->cache.contig); /* Cached information about contiguous data */
diff --git a/src/H5Defl.c b/src/H5Defl.c
index b7ba045..c8d1098 100644
--- a/src/H5Defl.c
+++ b/src/H5Defl.c
@@ -256,7 +256,7 @@ ssize_t
 H5D_efl_readvv(const H5D_io_info_t *io_info,
     size_t dset_max_nseq, size_t *dset_curr_seq, size_t dset_len_arr[], hsize_t dset_offset_arr[],
     size_t mem_max_nseq, size_t *mem_curr_seq, size_t mem_len_arr[], hsize_t mem_offset_arr[],
-    void *_buf)
+    haddr_t UNUSED address, void UNUSED *pointer/*in*/, void *_buf)
 {
     const H5O_efl_t *efl=&(io_info->store->efl); /* Pointer to efl info */
     unsigned char *buf;         /* Pointer to buffer to write */
@@ -340,7 +340,7 @@ ssize_t
 H5D_efl_writevv(const H5D_io_info_t *io_info,
     size_t dset_max_nseq, size_t *dset_curr_seq, size_t dset_len_arr[], hsize_t dset_offset_arr[],
     size_t mem_max_nseq, size_t *mem_curr_seq, size_t mem_len_arr[], hsize_t mem_offset_arr[],
-    const void *_buf)
+    haddr_t UNUSED address, void UNUSED *pointer/*in*/, const void *_buf)
 {
     const H5O_efl_t *efl=&(io_info->store->efl); /* Pointer to efl info */
     const unsigned char *buf;   /* Pointer to buffer to write */
diff --git a/src/H5Dio.c b/src/H5Dio.c
index 1bf5b04..7472921 100644
--- a/src/H5Dio.c
+++ b/src/H5Dio.c
@@ -821,10 +821,10 @@ H5D_contig_read(H5D_io_info_t *io_info, hsize_t nelmts,
 	}
 	else
 #endif
-         {
+        {
             if((io_info->ops.read)(io_info, (size_t)nelmts,
                     H5T_get_size(dataset->shared->type), file_space, mem_space,
-                    (haddr_t)0, buf/*out*/) < 0)
+                    (haddr_t)0, NULL, buf/*out*/) < 0)
                 HGOTO_ERROR(H5E_DATASET, H5E_CANTGET, FAIL, "contiguous read failed ");
 	}
 
@@ -926,7 +926,7 @@ H5D_contig_read(H5D_io_info_t *io_info, hsize_t nelmts,
                 || (dataset->shared->layout.type == H5D_CHUNKED && H5F_addr_defined(dataset->shared->layout.u.chunk.addr)))
             || dataset->shared->dcpl_cache.efl.nused > 0 ||
              dataset->shared->layout.type == H5D_COMPACT);
-        n = H5D_select_fgath(io_info, file_space, &file_iter, smine_nelmts, tconv_buf/*out*/);
+        n = H5D_select_fgath(io_info, file_space, &file_iter, smine_nelmts, (haddr_t)0, NULL, tconv_buf/*out*/);
 
 #ifdef H5S_DEBUG
 	H5_timer_end(&(io_info->stats->stats[1].gath_timer), &timer);
@@ -1096,7 +1096,7 @@ H5D_contig_write(H5D_io_info_t *io_info, hsize_t nelmts,
         {
             if((io_info->ops.write)(io_info, (size_t)nelmts,
                     H5T_get_size(dataset->shared->type), file_space, mem_space,
-                    (haddr_t)0, buf/*out*/) < 0)
+                    (haddr_t)0, NULL, buf/*out*/) < 0)
                 HGOTO_ERROR(H5E_DATASET, H5E_CANTGET, FAIL, "contiguous write failed ")
 	}
 
@@ -1221,9 +1221,8 @@ H5D_contig_write(H5D_io_info_t *io_info, hsize_t nelmts,
 #ifdef H5S_DEBUG
                 H5_timer_begin(&timer);
 #endif
-                n = H5D_select_fgath(io_info,
-                    file_space, &bkg_iter, smine_nelmts,
-                    bkg_buf/*out*/);
+                n = H5D_select_fgath(io_info, file_space, &bkg_iter, smine_nelmts,
+                                    (haddr_t)0, NULL, bkg_buf/*out*/);
 
 #ifdef H5S_DEBUG
                 H5_timer_end(&(io_info->stats->stats[0].bkg_timer), &timer);
@@ -1253,7 +1252,7 @@ H5D_contig_write(H5D_io_info_t *io_info, hsize_t nelmts,
         H5_timer_begin(&timer);
 #endif
 	status = H5D_select_fscat(io_info, file_space, &file_iter, smine_nelmts,
-                    tconv_buf);
+                    (haddr_t)0, NULL, tconv_buf);
 #ifdef H5S_DEBUG
         H5_timer_end(&(io_info->stats->stats[0].scat_timer), &timer);
         io_info->stats->stats[0].scat_nbytes += smine_nelmts * dst_type_size;
@@ -1297,6 +1296,12 @@ done:
  * Programmer:	Raymond Lu
  *		Thursday, April 10, 2003
  *
+ * Modification: 
+ *              Raymond Lu
+ *              20 July 2007
+ *              Moved H5D_istore_lock and H5D_istore_unlock to this level
+ *              from H5D_istore_readvv to avoid frequent lock and unlock
+ *              and to improve performance.
  *-------------------------------------------------------------------------
  */
 static herr_t
@@ -1320,6 +1325,7 @@ H5D_chunk_read(H5D_io_info_t *io_info, hsize_t nelmts,
     size_t	request_nelmts;		/*requested strip mine	*/
     hsize_t     smine_start;            /*strip mine start loc  */
     size_t      n, smine_nelmts;        /*elements per strip    */
+    size_t      accessed_bytes;         /*total accessed size in a chunk */
     H5S_sel_iter_t mem_iter;            /*memory selection iteration info*/
     hbool_t	mem_iter_init = FALSE;	/*memory selection iteration info has been initialized */
     H5S_sel_iter_t bkg_iter;            /*background iteration info*/
@@ -1330,6 +1336,10 @@ H5D_chunk_read(H5D_io_info_t *io_info, hsize_t nelmts,
     uint8_t	*tconv_buf = NULL;	/*datatype conv buffer	*/
     uint8_t	*bkg_buf = NULL;	/*background buffer	*/
     H5D_storage_t store;                /*union of EFL and chunk pointer in file space */
+    void        *chunk = NULL;
+    haddr_t	chunk_addr;             /* Chunk address on disk */
+    H5D_istore_ud1_t udata;		/*B-tree pass-through	*/
+    unsigned    idx_hint=0;             /* Cache index hint      */
     herr_t	ret_value = SUCCEED;	/*return value		*/
 
     FUNC_ENTER_NOAPI_NOINIT(H5D_chunk_read)
@@ -1383,11 +1393,27 @@ H5D_chunk_read(H5D_io_info_t *io_info, hsize_t nelmts,
                 store.chunk.offset = chunk_info->coords;
                 store.chunk.index = chunk_info->index;
 
+                /* Load the chunk into cache and lock it. */
+		chunk_addr = H5D_istore_get_addr(io_info, &udata);
+
+                if (H5D_istore_if_load(dataset, chunk_addr)) {
+		    if(NULL == (chunk = H5D_istore_lock(io_info, &udata, FALSE, &idx_hint)))
+			HGOTO_ERROR(H5E_IO, H5E_READERROR, FAIL, "unable to read raw data chunk")
+		} else
+		    chunk = NULL;
+
                 /* Perform the actual read operation */
                 if((io_info->ops.read)(io_info, chunk_info->chunk_points,
                         H5T_get_size(dataset->shared->type), chunk_info->fspace,
-                        chunk_info->mspace, (haddr_t)0, buf) < 0)
+                        chunk_info->mspace, chunk_addr, chunk, buf) < 0)
                     HGOTO_ERROR(H5E_DATASET, H5E_READERROR, FAIL, " chunked read failed")
+  
+                /* Release the cache lock on the chunk. */
+                if (H5D_istore_if_load(dataset, chunk_addr)) {
+                    accessed_bytes = chunk_info->chunk_points * H5T_get_size(dataset->shared->type);
+		    if(H5D_istore_unlock(io_info, FALSE, idx_hint, chunk, accessed_bytes) < 0)
+			HGOTO_ERROR(H5E_IO, H5E_READERROR, FAIL, "unable to unlock raw data chunk")
+		}
 
                 /* Advance to next chunk in list */
                 chunk_node = H5SL_next(chunk_node);
@@ -1492,6 +1518,15 @@ H5D_chunk_read(H5D_io_info_t *io_info, hsize_t nelmts,
         store.chunk.offset = chunk_info->coords;
         store.chunk.index = chunk_info->index;
 
+        /* Load the chunk into cache and lock it. */
+        chunk_addr = H5D_istore_get_addr(io_info, &udata);
+
+        if (H5D_istore_if_load(dataset, chunk_addr)) {
+            if(NULL == (chunk = H5D_istore_lock(io_info, &udata, FALSE, &idx_hint)))
+                HGOTO_ERROR(H5E_IO, H5E_READERROR, FAIL, "unable to read raw data chunk")
+        } else
+            chunk = NULL;
+
         for (smine_start=0; smine_start<chunk_info->chunk_points; smine_start+=smine_nelmts) {
             /* Go figure out how many elements to read from the file */
             assert(H5S_SELECT_ITER_NELMTS(&file_iter)==(chunk_info->chunk_points-smine_start));
@@ -1509,7 +1544,14 @@ H5D_chunk_read(H5D_io_info_t *io_info, hsize_t nelmts,
             HDassert(((dataset->shared->layout.type == H5D_CONTIGUOUS && H5F_addr_defined(dataset->shared->layout.u.contig.addr))
                     || (dataset->shared->layout.type == H5D_CHUNKED && H5F_addr_defined(dataset->shared->layout.u.chunk.addr)))
                 || dataset->shared->dcpl_cache.efl.nused > 0 || dataset->shared->layout.type == H5D_COMPACT);
-            n = H5D_select_fgath(io_info, chunk_info->fspace, &file_iter, smine_nelmts, tconv_buf/*out*/);
+
+            if(chunk) {
+                n = H5D_select_mgath(chunk, chunk_info->fspace, &file_iter, 
+                                 smine_nelmts, dxpl_cache, tconv_buf/*out*/);
+            } else {
+                n = H5D_select_fgath(io_info, chunk_info->fspace, &file_iter, smine_nelmts, 
+                                 (haddr_t)0, NULL, tconv_buf/*out*/);
+            }
 
 #ifdef H5S_DEBUG
             H5_timer_end(&(io_info->stats->stats[1].gath_timer), &timer);
@@ -1577,6 +1619,13 @@ H5D_chunk_read(H5D_io_info_t *io_info, hsize_t nelmts,
                 HGOTO_ERROR(H5E_DATASET, H5E_READERROR, FAIL, "scatter failed")
         } /* end for */
 
+        /* Release the cache lock on the chunk. */
+        if (H5D_istore_if_load(dataset, chunk_addr)) {
+            accessed_bytes = chunk_info->chunk_points * H5T_get_size(dataset->shared->type);
+            if(H5D_istore_unlock(io_info, FALSE, idx_hint, chunk, accessed_bytes) < 0)
+                HGOTO_ERROR(H5E_IO, H5E_READERROR, FAIL, "unable to unlock raw data chunk")
+        }
+
         /* Release selection iterators */
         if(file_iter_init) {
             if(H5S_SELECT_ITER_RELEASE(&file_iter) < 0)
@@ -1636,6 +1685,12 @@ done:
  * Programmer:	Raymond Lu
  *		Thursday, April 10, 2003
  *
+ * Modification: 
+ *              Raymond Lu
+ *              20 July 2007
+ *              Moved H5D_istore_lock and H5D_istore_unlock to this level
+ *              from H5D_istore_writevv to avoid frequent lock and unlock
+ *              and to improve performance.
  *-------------------------------------------------------------------------
  */
 static herr_t
@@ -1657,6 +1712,7 @@ H5D_chunk_write(H5D_io_info_t *io_info, hsize_t nelmts,
     size_t	max_type_size;	        /* Size of largest source/destination type */
     size_t	target_size;		/*desired buffer size	*/
     size_t	request_nelmts;		/*requested strip mine	*/
+    size_t      accessed_bytes;         /*total accessed size in a chunk */
     hsize_t     smine_start;            /*strip mine start loc  */
     size_t      n, smine_nelmts;        /*elements per strip    */
     H5S_sel_iter_t mem_iter;            /*memory selection iteration info*/
@@ -1669,7 +1725,11 @@ H5D_chunk_write(H5D_io_info_t *io_info, hsize_t nelmts,
     uint8_t	*tconv_buf = NULL;	/*datatype conv buffer	*/
     uint8_t	*bkg_buf = NULL;	/*background buffer	*/
     H5D_storage_t store;                /*union of EFL and chunk pointer in file space */
-
+    void     *chunk = NULL;
+    haddr_t	chunk_addr;             /* Chunk address on disk */
+    H5D_istore_ud1_t udata;		/*B-tree pass-through	*/
+    unsigned    idx_hint=0;             /* Cache index hint      */
+    hbool_t     relax=TRUE;             /* Whether whole chunk is selected */
     herr_t	ret_value = SUCCEED;	/*return value		*/
 
     FUNC_ENTER_NOAPI_NOINIT(H5D_chunk_write)
@@ -1681,6 +1741,16 @@ H5D_chunk_write(H5D_io_info_t *io_info, hsize_t nelmts,
     /* Set dataset storage for I/O info */
     io_info->store=&store;
 
+#ifdef H5_HAVE_PARALLEL
+    /* Additional sanity checks when operating in parallel */
+    if(IS_H5FD_MPI(dataset->oloc.file)) {
+        if (chunk_addr==HADDR_UNDEF)
+            HGOTO_ERROR(H5E_IO, H5E_WRITEERROR, FAIL, "unable to locate raw data chunk")
+        if (dataset->shared->dcpl_cache.pline.nused>0)
+            HGOTO_ERROR(H5E_IO, H5E_WRITEERROR, FAIL, "cannot write to chunked storage with filters in parallel")
+    } /* end if */
+#endif /* H5_HAVE_PARALLEL */
+
     /*
      * If there is no type conversion then write directly from the
      * application's buffer.  This saves at least one mem-to-mem copy.
@@ -1718,12 +1788,33 @@ H5D_chunk_write(H5D_io_info_t *io_info, hsize_t nelmts,
                 store.chunk.offset = chunk_info->coords;
                 store.chunk.index = chunk_info->index;
 
+                /* Load the chunk into cache.  But if the whole chunk is written,
+                 * simply allocate space instead of load the chunk. */
+                chunk_addr = H5D_istore_get_addr(io_info, &udata);
+
+                if (H5D_istore_if_load(dataset, chunk_addr)) {
+                    accessed_bytes = chunk_info->chunk_points * H5T_get_size(dataset->shared->type);
+                    if(accessed_bytes != dataset->shared->layout.u.chunk.size)
+                        relax=FALSE;
+
+		    if(NULL == (chunk = H5D_istore_lock(io_info, &udata, relax, &idx_hint)))
+			HGOTO_ERROR(H5E_IO, H5E_READERROR, FAIL, "unable to read raw data chunk")
+		} else
+		    chunk = NULL;
+
                 /* Perform the actual read operation */
                 if((io_info->ops.write)(io_info, chunk_info->chunk_points,
                         H5T_get_size(dataset->shared->type), chunk_info->fspace,
-                        chunk_info->mspace, (haddr_t)0, buf) < 0)
+                        chunk_info->mspace, chunk_addr, chunk, buf) < 0)
                     HGOTO_ERROR(H5E_DATASET, H5E_READERROR, FAIL, " chunked write failed")
 
+                /* Release the cache lock on the chunk. */
+                if (H5D_istore_if_load(dataset, chunk_addr)) {
+		    if(H5D_istore_unlock(io_info, TRUE, idx_hint, chunk, accessed_bytes) < 0)
+			HGOTO_ERROR(H5E_IO, H5E_READERROR, FAIL, "unable to unlock raw data chunk")
+		}
+                relax=TRUE;
+
                 /* Advance to next chunk in list */
                 chunk_node = H5SL_next(chunk_node);
             } /* end while */
@@ -1830,6 +1921,25 @@ H5D_chunk_write(H5D_io_info_t *io_info, hsize_t nelmts,
         store.chunk.offset = chunk_info->coords;
         store.chunk.index = chunk_info->index;
 
+        /* Load the chunk into cache.  But if the whole chunk is written,
+         * simply allocate space instead of load the chunk. */
+        chunk_addr = H5D_istore_get_addr(io_info, &udata);
+
+        if (H5D_istore_if_load(dataset, chunk_addr)) {
+            accessed_bytes = chunk_info->chunk_points * H5T_get_size(dataset->shared->type);
+            if(accessed_bytes != dataset->shared->layout.u.chunk.size)
+                relax=FALSE;
+            if(relax) {
+                accessed_bytes = H5S_GET_SELECT_NPOINTS(chunk_info->mspace)*H5T_get_size(mem_type);
+                if(accessed_bytes != dataset->shared->layout.u.chunk.size)
+                    relax = FALSE;
+            }
+
+	    if(NULL == (chunk = H5D_istore_lock(io_info, &udata, relax, &idx_hint)))
+		HGOTO_ERROR(H5E_IO, H5E_READERROR, FAIL, "unable to read raw data chunk")
+	} else
+	    chunk = NULL;
+
         for (smine_start=0; smine_start<chunk_info->chunk_points; smine_start+=smine_nelmts) {
             /* Go figure out how many elements to read from the file */
             assert(H5S_SELECT_ITER_NELMTS(&file_iter)==(chunk_info->chunk_points-smine_start));
@@ -1868,8 +1978,13 @@ H5D_chunk_write(H5D_io_info_t *io_info, hsize_t nelmts,
 #ifdef H5S_DEBUG
                     H5_timer_begin(&timer);
 #endif
-                    n = H5D_select_fgath(io_info, chunk_info->fspace, &bkg_iter, smine_nelmts,
-                            bkg_buf/*out*/);
+                    if(chunk) {
+                        n = H5D_select_mgath(chunk, chunk_info->fspace, &bkg_iter, 
+                                 smine_nelmts, dxpl_cache, bkg_buf/*out*/);
+                    } else {
+                        n = H5D_select_fgath(io_info, chunk_info->fspace, &bkg_iter, smine_nelmts, 
+                                 (haddr_t)0, NULL, bkg_buf/*out*/);
+                    }
 
 #ifdef H5S_DEBUG
                     H5_timer_end(&(io_info->stats->stats[0].bkg_timer), &timer);
@@ -1900,7 +2015,7 @@ H5D_chunk_write(H5D_io_info_t *io_info, hsize_t nelmts,
 #endif
             status = H5D_select_fscat(io_info,
                 chunk_info->fspace, &file_iter, smine_nelmts,
-                tconv_buf);
+                chunk_addr, chunk, tconv_buf);
 
 #ifdef H5S_DEBUG
             H5_timer_end(&(io_info->stats->stats[0].scat_timer), &timer);
@@ -1911,6 +2026,14 @@ H5D_chunk_write(H5D_io_info_t *io_info, hsize_t nelmts,
                 HGOTO_ERROR(H5E_IO, H5E_WRITEERROR, FAIL, "scatter failed")
         } /* end for */
 
+        /* Release the cache lock on the chunk. */
+        if (H5D_istore_if_load(dataset, chunk_addr)) {
+            accessed_bytes = chunk_info->chunk_points * H5T_get_size(dataset->shared->type);
+	    if(H5D_istore_unlock(io_info, TRUE, idx_hint, chunk, accessed_bytes) < 0)
+		HGOTO_ERROR(H5E_IO, H5E_READERROR, FAIL, "unable to unlock raw data chunk")
+	}
+        relax=TRUE;
+
         /* Release selection iterators */
         if(file_iter_init) {
             if(H5S_SELECT_ITER_RELEASE(&file_iter) < 0)
diff --git a/src/H5Distore.c b/src/H5Distore.c
index 5d744e5..687e6d6 100644
--- a/src/H5Distore.c
+++ b/src/H5Distore.c
@@ -129,37 +129,6 @@ typedef struct H5D_rdcc_ent_t {
 typedef H5D_rdcc_ent_t *H5D_rdcc_ent_ptr_t; /* For free lists */
 
 /*
- * B-tree key.	A key contains the minimum logical N-dimensional address and
- * the logical size of the chunk to which this key refers.  The
- * fastest-varying dimension is assumed to reference individual bytes of the
- * array, so a 100-element 1-d array of 4-byte integers would really be a 2-d
- * array with the slow varying dimension of size 100 and the fast varying
- * dimension of size 4 (the storage dimensionality has very little to do with
- * the real dimensionality).
- *
- * Only the first few values of the OFFSET and SIZE fields are actually
- * stored on disk, depending on the dimensionality.
- *
- * The chunk's file address is part of the B-tree and not part of the key.
- */
-typedef struct H5D_istore_key_t {
-    size_t	nbytes;				/*size of stored data	*/
-    hsize_t	offset[H5O_LAYOUT_NDIMS];	/*logical offset to start*/
-    unsigned	filter_mask;			/*excluded filters	*/
-} H5D_istore_key_t;
-
-/*
- * Common data exchange structure for indexed storage nodes.  This structure is
- * passed through the B-link tree layer to the methods for the objects
- * to which the B-link tree points.
- */
-typedef struct H5D_istore_bt_ud_common_t {
-    /* downward */
-    H5D_istore_key_t	key;	                /*key values		*/
-    const H5O_layout_t	*mesg;		        /*layout message	*/
-} H5D_istore_bt_ud_common_t;
-
-/*
  * Data exchange structure for indexed storage nodes.  This structure is
  * passed through the B-link tree layer to the methods for the objects
  * to which the B-link tree points for operations which require no
@@ -169,12 +138,6 @@ typedef struct H5D_istore_bt_ud_common_t {
  */
 typedef H5D_istore_bt_ud_common_t H5D_istore_ud0_t;
 
-/* B-tree callback info for various operations */
-typedef struct H5D_istore_ud1_t {
-    H5D_istore_bt_ud_common_t common;           /* Common info for B-tree user data (must be first) */
-    haddr_t		addr;			/*file address of chunk */
-} H5D_istore_ud1_t;
-
 /* B-tree callback info for iteration to total allocated space */
 typedef struct H5D_istore_it_ud1_t {
     H5D_istore_bt_ud_common_t common;           /* Common info for B-tree user data (must be first) */
@@ -1715,7 +1678,7 @@ done:
  *
  *-------------------------------------------------------------------------
  */
-static void *
+void *
 H5D_istore_lock(const H5D_io_info_t *io_info, H5D_istore_ud1_t *udata,
     hbool_t relax, unsigned *idx_hint/*in,out*/)
 {
@@ -1781,6 +1744,10 @@ H5D_istore_lock(const H5D_io_info_t *io_info, H5D_istore_ud1_t *udata,
 #endif
         if(NULL == (chunk = H5D_istore_chunk_alloc(chunk_size, pline)))
             HGOTO_ERROR(H5E_RESOURCE, H5E_NOSPACE, NULL, "memory allocation failed for raw data chunk")
+	
+        /* In the case that some dataset functions look through this data,
+         * clear it to all 0s. */
+        HDmemset(chunk, 0, chunk_size);
 
     } /* end if */
     else {
@@ -1897,7 +1864,7 @@ H5D_istore_lock(const H5D_io_info_t *io_info, H5D_istore_ud1_t *udata,
             ent->offset[u] = io_info->store->chunk.offset[u];
         ent->rd_count = chunk_size;
         ent->wr_count = chunk_size;
-        ent->chunk = chunk;
+        ent->chunk = (uint8_t*)chunk;
 
         /* Add it to the cache */
         assert(NULL==rdcc->slot[idx]);
@@ -1998,9 +1965,9 @@ done:
  *
  *-------------------------------------------------------------------------
  */
-static herr_t
+herr_t
 H5D_istore_unlock(const H5D_io_info_t *io_info,
-    hbool_t dirty, unsigned idx_hint, uint8_t *chunk, size_t naccessed)
+    hbool_t dirty, unsigned idx_hint, void *chunk, size_t naccessed)
 {
     const H5O_layout_t *layout=&(io_info->dset->shared->layout); /* Dataset layout */
     const H5D_rdcc_t	*rdcc = &(io_info->dset->shared->cache.chunk);
@@ -2029,7 +1996,7 @@ H5D_istore_unlock(const H5D_io_info_t *io_info,
             assert(layout->u.chunk.size>0);
             H5_ASSIGN_OVERFLOW(x.chunk_size,layout->u.chunk.size,hsize_t,size_t);
             x.alloc_size = x.chunk_size;
-            x.chunk = chunk;
+            x.chunk = (uint8_t*)chunk;
 
             if (H5D_istore_flush_entry(io_info, &x, TRUE)<0)
                 HGOTO_ERROR(H5E_IO, H5E_WRITEERROR, FAIL, "cannot flush indexed storage buffer")
@@ -2063,6 +2030,48 @@ done:
 
 
 /*-------------------------------------------------------------------------
+ * Function:	H5D_istore_if_load
+ *
+ * Purpose:	A small internal function to if it's necessary to load the
+ *              chunk into cache.  If the chunk is too large to load into 
+ *              the cache and it has no filters in the pipeline (i.e. not 
+ *              compressed) and if the address for the chunk has been 
+ *              defined, then don't load the chunk into the cache, just 
+ *              read the data from it directly. If MPI based VFD 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 read-through of only the elements 
+ *              requested.
+ *
+ * Return:	TRUE or FALSE
+ *
+ * Programmer:	Raymond Lu
+ *		17 July 2007
+ *
+ *-------------------------------------------------------------------------
+ */
+hbool_t 
+H5D_istore_if_load(H5D_t *dataset, haddr_t caddr)
+{
+    hbool_t ret_value;
+ 
+    FUNC_ENTER_NOAPI_NOINIT(H5D_istore_if_load)
+
+    assert(dataset);
+
+    if (dataset->shared->dcpl_cache.pline.nused==0 && ((dataset->shared->layout.u.chunk.size > 
+            dataset->shared->cache.chunk.nbytes && caddr!=HADDR_UNDEF)
+            || (IS_H5FD_MPI(dataset->oloc.file) && (H5F_ACC_RDWR & 
+            H5F_get_intent(dataset->oloc.file))))) {
+        ret_value = FALSE;
+    } else
+        ret_value = TRUE;
+
+    FUNC_LEAVE_NOAPI(ret_value)
+}
+
+
+/*-------------------------------------------------------------------------
  * Function:	H5D_istore_readvv
  *
  * Purpose:	Reads a multi-dimensional buffer from (part of) an indexed raw
@@ -2073,19 +2082,25 @@ done:
  * Programmer:	Quincey Koziol
  *		Wednesday, May  7, 2003
  *
+ * Modification: 
+ *              Raymond Lu
+ *              20 July 2007
+ *              Moved H5D_istore_lock and H5D_istore_unlock to H5D_chunk_read
+ *              from this function to avoid frequent lock and unlock.
+ *
  *-------------------------------------------------------------------------
  */
 ssize_t
 H5D_istore_readvv(const H5D_io_info_t *io_info,
     size_t chunk_max_nseq, size_t *chunk_curr_seq, size_t chunk_len_arr[], hsize_t chunk_offset_arr[],
     size_t mem_max_nseq, size_t *mem_curr_seq, size_t mem_len_arr[], hsize_t mem_offset_arr[],
-    void *buf)
+    haddr_t chunk_addr, void *chunk, void *buf)
 {
     H5D_t *dset=io_info->dset;          /* Local pointer to the dataset info */
     H5D_istore_ud1_t udata;		/*B-tree pass-through	*/
-    haddr_t	        chunk_addr;     /* Chunk address on disk */
     size_t		u;              /* Local index variables */
     ssize_t             ret_value;      /* Return value */
+        ssize_t         naccessed;      /* Number of bytes accessed in chunk */
 
     FUNC_ENTER_NOAPI(H5D_istore_readvv, FAIL)
 
@@ -2106,9 +2121,7 @@ H5D_istore_readvv(const H5D_io_info_t *io_info,
 HDfprintf(stderr,"%s: io_info->store->chunk.offset={",FUNC);
 for(u=0; u<dset->shared->layout.u.chunk.ndims; u++)
     HDfprintf(stderr,"%Hd%s",io_info->store->chunk.offset[u],(u<(dset->shared->layout.u.chunk.ndims-1) ? ", " : "}\n"));
-#endif /* QAK */
-    chunk_addr = H5D_istore_get_addr(io_info, &udata);
-#ifdef QAK
+
 HDfprintf(stderr,"%s: chunk_addr=%a, chunk_size=%Zu\n",FUNC,chunk_addr,dset->shared->layout.u.chunk.size);
 HDfprintf(stderr,"%s: chunk_len_arr[%Zu]=%Zu\n",FUNC,*chunk_curr_seq,chunk_len_arr[*chunk_curr_seq]);
 HDfprintf(stderr,"%s: chunk_offset_arr[%Zu]=%Hu\n",FUNC,*chunk_curr_seq,chunk_offset_arr[*chunk_curr_seq]);
@@ -2128,8 +2141,7 @@ HDfprintf(stderr,"%s: buf=%p\n",FUNC,buf);
      * writing to other elements in the same chunk.  Do a direct
      * read-through of only the elements requested.
      */
-    if (dset->shared->dcpl_cache.pline.nused==0 && ((dset->shared->layout.u.chunk.size>dset->shared->cache.chunk.nbytes && chunk_addr!=HADDR_UNDEF)
-            || (IS_H5FD_MPI(dset->oloc.file) && (H5F_ACC_RDWR & H5F_get_intent(dset->oloc.file))))) {
+    if (!H5D_istore_if_load(dset, chunk_addr)) {
         H5D_io_info_t chk_io_info;      /* Temporary I/O info object */
         H5D_storage_t chk_store;        /* Chunk storage information */
 
@@ -2141,12 +2153,11 @@ HDfprintf(stderr,"%s: buf=%p\n",FUNC,buf);
         H5D_BUILD_IO_INFO(&chk_io_info,dset,io_info->dxpl_cache,io_info->dxpl_id,&chk_store);
 
         /* Do I/O directly on chunk without reading it into the cache */
-        if ((ret_value=H5D_contig_readvv(&chk_io_info, chunk_max_nseq, chunk_curr_seq, chunk_len_arr, chunk_offset_arr, mem_max_nseq, mem_curr_seq, mem_len_arr, mem_offset_arr, buf))<0)
+        if ((ret_value=H5D_contig_readvv(&chk_io_info, chunk_max_nseq, chunk_curr_seq, chunk_len_arr, 
+                chunk_offset_arr, mem_max_nseq, mem_curr_seq, mem_len_arr, mem_offset_arr, (haddr_t)0, NULL, buf))<0)
             HGOTO_ERROR(H5E_IO, H5E_READERROR, FAIL, "unable to read raw data to file")
     } /* end if */
     else {
-        uint8_t         *chunk;         /* Pointer to cached chunk in memory */
-        unsigned        idx_hint=0;     /* Cache index hint      */
         ssize_t         naccessed;      /* Number of bytes accessed in chunk */
 
         /* If the chunk address is not defined, check if the fill value is
@@ -2229,20 +2240,11 @@ HDfprintf(stderr,"%s: buf=%p\n",FUNC,buf);
             } /* end if */
         } /* end if */
 
-        /*
-         * Lock the chunk, copy from application to chunk, then unlock the
-         * chunk.
-         */
-        if(NULL == (chunk = H5D_istore_lock(io_info, &udata, FALSE, &idx_hint)))
-            HGOTO_ERROR(H5E_IO, H5E_READERROR, FAIL, "unable to read raw data chunk")
-
         /* Use the vectorized memory copy routine to do actual work */
         if((naccessed = H5V_memcpyvv(buf, mem_max_nseq, mem_curr_seq, mem_len_arr, mem_offset_arr, chunk, chunk_max_nseq, chunk_curr_seq, chunk_len_arr, chunk_offset_arr)) < 0)
             HGOTO_ERROR(H5E_IO, H5E_READERROR, FAIL, "vectorized memcpy failed")
 
         H5_CHECK_OVERFLOW(naccessed, ssize_t, size_t);
-        if(H5D_istore_unlock(io_info, FALSE, idx_hint, chunk, (size_t)naccessed) < 0)
-            HGOTO_ERROR(H5E_IO, H5E_READERROR, FAIL, "unable to unlock raw data chunk")
 
         /* Set return value */
         ret_value = naccessed;
@@ -2264,17 +2266,22 @@ done:
  * Programmer:	Quincey Koziol
  *		Friday, May  2, 2003
  *
+ * Modification: 
+ *              Raymond Lu
+ *              20 July 2007
+ *              Moved H5D_istore_lock and H5D_istore_unlock to H5D_chunk_write
+ *              from this function to avoid frequent lock and unlock.
+ *
  *-------------------------------------------------------------------------
  */
 ssize_t
 H5D_istore_writevv(const H5D_io_info_t *io_info,
     size_t chunk_max_nseq, size_t *chunk_curr_seq, size_t chunk_len_arr[], hsize_t chunk_offset_arr[],
     size_t mem_max_nseq, size_t *mem_curr_seq, size_t mem_len_arr[], hsize_t mem_offset_arr[],
-    const void *buf)
+    haddr_t chunk_addr, void *chunk, const void *buf)
 {
     H5D_t *dset = io_info->dset;          /* Local pointer to the dataset info */
     H5D_istore_ud1_t udata;		/*B-tree pass-through	*/
-    haddr_t	        chunk_addr;     /* Chunk address on disk */
     size_t		u;              /* Local index variables */
     ssize_t             ret_value;      /* Return value */
 
@@ -2292,14 +2299,11 @@ H5D_istore_writevv(const H5D_io_info_t *io_info,
     HDassert(mem_offset_arr);
     HDassert(buf);
 
-    /* Get the address of this chunk on disk */
 #ifdef QAK
 HDfprintf(stderr,"%s: io_info->store->chunk.offset={",FUNC);
 for(u=0; u<dset->shared->layout.u.chunk.ndims; u++)
     HDfprintf(stderr,"%Hd%s",io_info->store->chunk.offset[u],(u<(dset->shared->layout.u.chunk.ndims-1) ? ", " : "}\n"));
-#endif /* QAK */
-    chunk_addr = H5D_istore_get_addr(io_info, &udata);
-#ifdef QAK
+
 HDfprintf(stderr,"%s: chunk_addr=%a, chunk_size=%Zu\n",FUNC,chunk_addr,dset->shared->layout.u.chunk.size);
 HDfprintf(stderr,"%s: chunk_len_arr[%Zu]=%Zu\n",FUNC,*chunk_curr_seq,chunk_len_arr[*chunk_curr_seq]);
 HDfprintf(stderr,"%s: chunk_offset_arr[%Zu]=%Hu\n",FUNC,*chunk_curr_seq,chunk_offset_arr[*chunk_curr_seq]);
@@ -2318,18 +2322,7 @@ HDfprintf(stderr,"%s: mem_offset_arr[%Zu]=%Hu\n",FUNC,*mem_curr_seq,mem_offset_a
      * writing to other elements in the same chunk.  Do a direct
      * write-through of only the elements requested.
      */
-#ifdef H5_HAVE_PARALLEL
-    /* Additional sanity checks when operating in parallel */
-    if(IS_H5FD_MPI(dset->oloc.file)) {
-        if (chunk_addr==HADDR_UNDEF)
-            HGOTO_ERROR(H5E_IO, H5E_WRITEERROR, FAIL, "unable to locate raw data chunk")
-        if (dset->shared->dcpl_cache.pline.nused>0)
-            HGOTO_ERROR(H5E_IO, H5E_WRITEERROR, FAIL, "cannot write to chunked storage with filters in parallel")
-    } /* end if */
-#endif /* H5_HAVE_PARALLEL */
-
-    if (dset->shared->dcpl_cache.pline.nused==0 && ((dset->shared->layout.u.chunk.size>dset->shared->cache.chunk.nbytes && chunk_addr!=HADDR_UNDEF)
-            || (IS_H5FD_MPI(dset->oloc.file) && (H5F_ACC_RDWR & H5F_get_intent(dset->oloc.file))))) {
+    if (!H5D_istore_if_load(dset, chunk_addr)) {
         H5D_io_info_t chk_io_info;      /* Temporary I/O info object */
         H5D_storage_t chk_store;        /* Chunk storage information */
 
@@ -2341,56 +2334,17 @@ HDfprintf(stderr,"%s: mem_offset_arr[%Zu]=%Hu\n",FUNC,*mem_curr_seq,mem_offset_a
         H5D_BUILD_IO_INFO(&chk_io_info,dset,io_info->dxpl_cache,io_info->dxpl_id,&chk_store);
 
         /* Do I/O directly on chunk without reading it into the cache */
-        if ((ret_value=H5D_contig_writevv(&chk_io_info, chunk_max_nseq, chunk_curr_seq, chunk_len_arr, chunk_offset_arr, mem_max_nseq, mem_curr_seq, mem_len_arr, mem_offset_arr, buf))<0)
+        if ((ret_value=H5D_contig_writevv(&chk_io_info, chunk_max_nseq, chunk_curr_seq, chunk_len_arr, chunk_offset_arr, mem_max_nseq, mem_curr_seq, mem_len_arr, mem_offset_arr, (haddr_t)0, NULL, buf))<0)
             HGOTO_ERROR(H5E_IO, H5E_WRITEERROR, FAIL, "unable to write raw data to file")
     } /* end if */
     else {
-        uint8_t         *chunk;         /* Pointer to cached chunk in memory */
-        unsigned        idx_hint=0;     /* Cache index hint      */
         ssize_t         naccessed;      /* Number of bytes accessed in chunk */
-        size_t          total_bytes;    /* Total # of bytes accessed on disk & memory */
-        hbool_t         relax;          /* Whether whole chunk is selected */
-
-        /*
-         * Lock the chunk, copy from application to chunk, then unlock the
-         * chunk.
-         */
-#ifdef OLD_WAY
-/* Note that this is technically OK, since eventually all the data in the chunk
- * will be overwritten.  However, it seems risky and a better approach would
- * be to lock the chunk in the dataset I/O routine (setting the relax flag
- * appropriately) and then unlock it after all the I/O the chunk was finished. -QAK
- */
-        if(chunk_max_nseq==1 && chunk_len_arr[0] == dset->shared->layout.u.chunk.size)
-            relax = TRUE;
-        else
-            relax = FALSE;
-#else /* OLD_WAY */
-        relax=TRUE;
-        total_bytes=0;
-        for(u=*chunk_curr_seq; u<chunk_max_nseq; u++)
-            total_bytes+=chunk_len_arr[u];
-        if(total_bytes!=dset->shared->layout.u.chunk.size)
-            relax=FALSE;
-        if(relax) {
-            total_bytes=0;
-            for(u=*mem_curr_seq; u<mem_max_nseq; u++)
-                total_bytes+=mem_len_arr[u];
-            if(total_bytes!=dset->shared->layout.u.chunk.size)
-                relax=FALSE;
-        } /* end if */
-#endif /* OLD_WAY */
-
-        if (NULL==(chunk=H5D_istore_lock(io_info, &udata, relax, &idx_hint)))
-            HGOTO_ERROR(H5E_IO, H5E_WRITEERROR, FAIL, "unable to read raw data chunk")
 
         /* Use the vectorized memory copy routine to do actual work */
         if((naccessed=H5V_memcpyvv(chunk,chunk_max_nseq,chunk_curr_seq,chunk_len_arr,chunk_offset_arr,buf,mem_max_nseq,mem_curr_seq,mem_len_arr,mem_offset_arr))<0)
             HGOTO_ERROR(H5E_IO, H5E_WRITEERROR, FAIL, "vectorized memcpy failed")
 
         H5_CHECK_OVERFLOW(naccessed,ssize_t,size_t);
-        if (H5D_istore_unlock(io_info, TRUE, idx_hint, chunk, (size_t)naccessed)<0)
-            HGOTO_ERROR(H5E_IO, H5E_WRITEERROR, FAIL, "uanble to unlock raw data chunk")
 
         /* Set return value */
         ret_value=naccessed;
diff --git a/src/H5Dmpio.c b/src/H5Dmpio.c
index f44bbaa..90f8974 100644
--- a/src/H5Dmpio.c
+++ b/src/H5Dmpio.c
@@ -338,6 +338,7 @@ H5D_mpio_select_read(H5D_io_info_t *io_info,
 		     const H5S_t UNUSED *file_space, 
 		     const H5S_t UNUSED *mem_space,
 		     haddr_t addr,		     
+                     void UNUSED *pointer,
 		     void *buf/*out*/)
 {
     herr_t ret_value = SUCCEED;
@@ -369,6 +370,7 @@ H5D_mpio_select_write(H5D_io_info_t *io_info,
 		      const H5S_t UNUSED *file_space, 
 		      const H5S_t UNUSED *mem_space,
 		      haddr_t addr,
+                      void UNUSED *pointer,
 		      const void *buf)
 {
     herr_t ret_value = SUCCEED;
@@ -1168,15 +1170,13 @@ H5D_multi_chunk_collective_io(H5D_io_info_t *io_info,fm_map *fm,const void *buf,
             if(do_write) {
                 if((io_info->ops.write)(io_info,
                          chunk_info->chunk_points,H5T_get_size(io_info->dset->shared->type),
-                         chunk_info->fspace,chunk_info->mspace,0,
-                         buf) < 0)
+                         chunk_info->fspace,chunk_info->mspace,0, NULL, buf) < 0)
                     HGOTO_ERROR(H5E_DATASET, H5E_WRITEERROR, FAIL, "optimized write failed")
             }
             else {
                  if((io_info->ops.read)(io_info,
                           chunk_info->chunk_points,H5T_get_size(io_info->dset->shared->type),
-                          chunk_info->fspace,chunk_info->mspace,0,
-                          buf) < 0)
+                          chunk_info->fspace,chunk_info->mspace,0, NULL, buf) < 0)
                     HGOTO_ERROR(H5E_DATASET, H5E_READERROR, FAIL, "optimized read failed")
               }
 #else
@@ -1322,15 +1322,13 @@ H5D_multi_chunk_collective_io_no_opt(H5D_io_info_t *io_info,fm_map *fm,const voi
               if(do_write) {
                 if((io_info->ops.write)(io_info,
                          chunk_info->chunk_points,H5T_get_size(io_info->dset->shared->type),
-                         chunk_info->fspace,chunk_info->mspace, (hsize_t)0,
-                         buf) < 0)
+                         chunk_info->fspace,chunk_info->mspace, (hsize_t)0, NULL, buf) < 0)
                     HGOTO_ERROR(H5E_DATASET, H5E_WRITEERROR, FAIL, "optimized write failed")
               }
               else {
                 if((io_info->ops.read)(io_info,
                           chunk_info->chunk_points,H5T_get_size(io_info->dset->shared->type),
-                          chunk_info->fspace,chunk_info->mspace, (hsize_t)0,
-                          buf) < 0)
+                          chunk_info->fspace,chunk_info->mspace, (hsize_t)0, NULL, buf) < 0)
                   HGOTO_ERROR(H5E_DATASET, H5E_READERROR, FAIL, "optimized read failed")
               }
             }
@@ -1470,13 +1468,13 @@ H5D_final_collective_io(H5D_io_info_t *io_info,MPI_Datatype*mpi_file_type,MPI_Da
     if(do_write) {
         if((io_info->ops.write)(io_info,
 		coll_info->mpi_buf_count,0,NULL,NULL,coll_info->chunk_addr,
-                buf) < 0)
+                NULL, buf) < 0)
 	    HGOTO_ERROR(H5E_DATASET, H5E_WRITEERROR, FAIL, "optimized write failed")
     }
     else {
         if((io_info->ops.read)(io_info,
 	        coll_info->mpi_buf_count,0,NULL,NULL,coll_info->chunk_addr,
-                buf) < 0)
+                NULL, buf) < 0)
             HGOTO_ERROR(H5E_DATASET, H5E_READERROR, FAIL, "optimized read failed")
     }
 
diff --git a/src/H5Dpkg.h b/src/H5Dpkg.h
index f28a5db..1f92bf8 100644
--- a/src/H5Dpkg.h
+++ b/src/H5Dpkg.h
@@ -71,24 +71,24 @@ struct H5D_io_info_t;
 typedef herr_t (*H5D_io_read_func_t)(struct H5D_io_info_t *io_info,
     size_t nelmts, size_t elmt_size,
     const H5S_t *file_space, const H5S_t *mem_space, haddr_t addr,
-    void *buf/*out*/);
+    void *chunk, void *buf/*out*/);
 
 
 /* Write directly from app buffer to file */
 typedef herr_t (*H5D_io_write_func_t)(struct H5D_io_info_t *io_info,
     size_t nelmts, size_t elmt_size,
     const H5S_t *file_space, const H5S_t *mem_space, haddr_t addr,
-    const void *buf);
+    void *chunk, const void *buf);
 
 /* Function pointers for I/O on particular types of dataset layouts */
 typedef ssize_t (*H5D_io_readvv_func_t)(const struct H5D_io_info_t *io_info,
     size_t dset_max_nseq, size_t *dset_curr_seq, size_t dset_len_arr[], hsize_t dset_offset_arr[],
     size_t mem_max_nseq, size_t *mem_curr_seq, size_t mem_len_arr[], hsize_t mem_offset_arr[],
-    void *buf);
+    haddr_t chunk_addr, void *chunk, void *buf);
 typedef ssize_t (*H5D_io_writevv_func_t)(const struct H5D_io_info_t *io_info,
     size_t dset_max_nseq, size_t *dset_curr_seq, size_t dset_len_arr[], hsize_t dset_offset_arr[],
     size_t mem_max_nseq, size_t *mem_curr_seq, size_t mem_len_arr[], hsize_t mem_offset_arr[],
-    const void *buf);
+    haddr_t chunk_addr, void *chunk, const void *buf);
 
 /* Typedef for raw data I/O framework info */
 typedef struct H5D_io_ops_t {
@@ -250,6 +250,46 @@ typedef struct {
     hbool_t     has_vlen_fill_type;     /* Whether the datatype for the fill value has a variable-length component */
 } H5D_fill_buf_info_t;
 
+/*************************/
+/* For chunk lock        */
+/*************************/
+/*
+ * B-tree key.	A key contains the minimum logical N-dimensional address and
+ * the logical size of the chunk to which this key refers.  The
+ * fastest-varying dimension is assumed to reference individual bytes of the
+ * array, so a 100-element 1-d array of 4-byte integers would really be a 2-d
+ * array with the slow varying dimension of size 100 and the fast varying
+ * dimension of size 4 (the storage dimensionality has very little to do with
+ * the real dimensionality).
+ *
+ * Only the first few values of the OFFSET and SIZE fields are actually
+ * stored on disk, depending on the dimensionality.
+ *
+ * The chunk's file address is part of the B-tree and not part of the key.
+ */
+typedef struct H5D_istore_key_t {
+    size_t	nbytes;				/*size of stored data	*/
+    hsize_t	offset[H5O_LAYOUT_NDIMS];	/*logical offset to start*/
+    unsigned	filter_mask;			/*excluded filters	*/
+} H5D_istore_key_t;
+
+
+/*
+ * Common data exchange structure for indexed storage nodes.  This structure is
+ * passed through the B-link tree layer to the methods for the objects
+ * to which the B-link tree points.
+ */
+typedef struct H5D_istore_bt_ud_common_t {
+    /* downward */
+    H5D_istore_key_t	key;	                /*key values		*/
+    const H5O_layout_t	*mesg;		        /*layout message	*/
+} H5D_istore_bt_ud_common_t;
+
+/* B-tree callback info for various operations */
+typedef struct H5D_istore_ud1_t {
+    H5D_istore_bt_ud_common_t common;           /* Common info for B-tree user data (must be first) */
+    haddr_t		addr;			/*file address of chunk */
+} H5D_istore_ud1_t; 
 
 /*****************************/
 /* Package Private Variables */
@@ -272,10 +312,10 @@ H5_DLL herr_t H5D_alloc_storage(H5F_t *f, hid_t dxpl_id, H5D_t *dset, H5D_time_a
 /* Functions that perform serial I/O operations */
 H5_DLL herr_t H5D_select_fscat (H5D_io_info_t *io_info,
     const H5S_t *file_space, H5S_sel_iter_t *file_iter, size_t nelmts,
-    const void *_buf);
+    haddr_t chunk_addr, void *chunk, const void *_buf);
 H5_DLL size_t H5D_select_fgath (H5D_io_info_t *io_info,
     const H5S_t *file_space, H5S_sel_iter_t *file_iter, size_t nelmts,
-    void *buf);
+    haddr_t chunk_addr, void *chunk, void *buf);
 H5_DLL herr_t H5D_select_mscat (const void *_tscat_buf,
     const H5S_t *space, H5S_sel_iter_t *iter, size_t nelmts,
     const H5D_dxpl_cache_t *dxpl_cache, void *_buf/*out*/);
@@ -285,11 +325,11 @@ H5_DLL size_t H5D_select_mgath (const void *_buf,
 H5_DLL herr_t H5D_select_read(H5D_io_info_t *io_info,
     size_t nelmts, size_t elmt_size,
     const H5S_t *file_space, const H5S_t *mem_space,
-    haddr_t addr, void *buf/*out*/);
+    haddr_t addr, void *chunk/*in*/, void *buf/*out*/);
 H5_DLL herr_t H5D_select_write(H5D_io_info_t *io_info,
     size_t nelmts, size_t elmt_size,
     const H5S_t *file_space, const H5S_t *mem_space,
-    haddr_t addr, const void *buf/*out*/);
+    haddr_t addr, void *chunk/*in*/, const void *buf/*out*/);
 
 /* Functions that operate on contiguous storage */
 H5_DLL herr_t H5D_contig_create(H5F_t *f, hid_t dxpl_id, H5O_layout_t *layout);
@@ -298,11 +338,11 @@ H5_DLL haddr_t H5D_contig_get_addr(const H5D_t *dset);
 H5_DLL ssize_t H5D_contig_readvv(const H5D_io_info_t *io_info,
     size_t dset_max_nseq, size_t *dset_curr_seq, size_t dset_len_arr[], hsize_t dset_offset_arr[],
     size_t mem_max_nseq, size_t *mem_curr_seq, size_t mem_len_arr[], hsize_t mem_offset_arr[],
-    void *buf);
+    haddr_t UNUSED address, void UNUSED *pointer, void *buf);
 H5_DLL ssize_t H5D_contig_writevv(const H5D_io_info_t *io_info,
     size_t dset_max_nseq, size_t *dset_curr_seq, size_t dset_len_arr[], hsize_t dset_offset_arr[],
     size_t mem_max_nseq, size_t *mem_curr_seq, size_t mem_len_arr[], hsize_t mem_offset_arr[],
-    const void *buf);
+    haddr_t UNUSED address, void UNUSED *pointer, const void *buf);
 H5_DLL herr_t H5D_contig_copy(H5F_t *f_src, const H5O_layout_t *layout_src, H5F_t *f_dst, 
     H5O_layout_t *layout_dst, H5T_t *src_dtype, H5O_copy_t *cpy_info, hid_t dxpl_id);
 
@@ -311,11 +351,11 @@ H5_DLL herr_t H5D_compact_fill(H5D_t *dset, hid_t dxpl_id);
 H5_DLL ssize_t H5D_compact_readvv(const H5D_io_info_t *io_info,
     size_t dset_max_nseq, size_t *dset_curr_seq, size_t dset_size_arr[], hsize_t dset_offset_arr[],
     size_t mem_max_nseq, size_t *mem_curr_seq, size_t mem_size_arr[], hsize_t mem_offset_arr[],
-    void *buf);
+    haddr_t UNUSED addr, void UNUSED *pointer/*in*/, void *buf);
 H5_DLL ssize_t H5D_compact_writevv(const H5D_io_info_t *io_info,
     size_t dset_max_nseq, size_t *dset_curr_seq, size_t dset_size_arr[], hsize_t dset_offset_arr[],
     size_t mem_max_nseq, size_t *mem_curr_seq, size_t mem_size_arr[], hsize_t mem_offset_arr[],
-    const void *buf);
+    haddr_t UNUSED addr, void UNUSED *pointer/*in*/, const void *buf);
 H5_DLL herr_t H5D_compact_copy(H5F_t *f_src, H5O_layout_t *layout_src,
     H5F_t *f_dst, H5O_layout_t *layout_dst, H5T_t *src_dtype, H5O_copy_t *cpy_info, hid_t dxpl_id);
 
@@ -344,26 +384,32 @@ H5_DLL herr_t H5D_istore_stats (H5D_t *dset, hbool_t headers);
 H5_DLL ssize_t H5D_istore_readvv(const H5D_io_info_t *io_info,
     size_t chunk_max_nseq, size_t *chunk_curr_seq, size_t chunk_len_arr[],
     hsize_t chunk_offset_arr[], size_t mem_max_nseq, size_t *mem_curr_seq,
-    size_t mem_len_arr[], hsize_t mem_offset_arr[], void *buf);
+    size_t mem_len_arr[], hsize_t mem_offset_arr[], haddr_t chunk_addr, void *chunk, void *buf);
 H5_DLL ssize_t H5D_istore_writevv(const H5D_io_info_t *io_info,
     size_t chunk_max_nseq, size_t *chunk_curr_seq, size_t chunk_len_arr[],
     hsize_t chunk_offset_arr[], size_t mem_max_nseq, size_t *mem_curr_seq,
-    size_t mem_len_arr[], hsize_t mem_offset_arr[], const void *buf);
+    size_t mem_len_arr[], hsize_t mem_offset_arr[], haddr_t chunk_addr, void *chunk, 
+    const void *buf);
 H5_DLL haddr_t H5D_istore_get_addr(const H5D_io_info_t *io_info,
     struct H5D_istore_ud1_t *_udata);
 H5_DLL herr_t H5D_istore_copy(H5F_t *f_src, H5O_layout_t *layout_src,
     H5F_t *f_dst, H5O_layout_t *layout_dst, H5T_t *src_dtype,
     H5O_copy_t *cpy_info, H5O_pline_t *pline, hid_t dxpl_id);
+H5_DLL void * H5D_istore_lock(const H5D_io_info_t *io_info, H5D_istore_ud1_t *udata,
+    hbool_t relax, unsigned *idx_hint/*in,out*/);
+H5_DLL herr_t H5D_istore_unlock(const H5D_io_info_t *io_info,
+    hbool_t dirty, unsigned idx_hint, void *chunk, size_t naccessed);
+H5_DLL hbool_t H5D_istore_if_load(H5D_t *dataset, haddr_t caddr);
 
 /* Functions that operate on external file list (efl) storage */
 H5_DLL ssize_t H5D_efl_readvv(const H5D_io_info_t *io_info,
     size_t dset_max_nseq, size_t *dset_curr_seq, size_t dset_len_arr[], hsize_t dset_offset_arr[],
     size_t mem_max_nseq, size_t *mem_curr_seq, size_t mem_len_arr[], hsize_t mem_offset_arr[],
-    void *buf);
+    haddr_t UNUSED addr, void UNUSED *pointer/*in*/, void *buf);
 H5_DLL ssize_t H5D_efl_writevv(const H5D_io_info_t *io_info,
     size_t dset_max_nseq, size_t *dset_curr_seq, size_t dset_len_arr[], hsize_t dset_offset_arr[],
     size_t mem_max_nseq, size_t *mem_curr_seq, size_t mem_len_arr[], hsize_t mem_offset_arr[],
-    const void *buf);
+    haddr_t UNUSED addr, void UNUSED *pointer/*in*/, const void *buf);
 
 /* Functions that perform fill value operations on datasets */
 H5_DLL herr_t H5D_fill(const void *fill, const H5T_t *fill_type, void *buf,
@@ -390,13 +436,13 @@ H5_DLL herr_t H5D_fill_term(H5D_fill_buf_info_t *fb_info);
 H5_DLL herr_t H5D_mpio_select_read(H5D_io_info_t *io_info,
     size_t nelmts, size_t elmt_size,
     const struct H5S_t *file_space, const struct H5S_t *mem_space,
-    haddr_t addr,void *buf/*out*/);
+    haddr_t addr, void UNUSED *pointer/*in*/, void *buf/*out*/);
 
 /* MPI-IO function to read , it will select either regular or irregular read */
 H5_DLL herr_t H5D_mpio_select_write(H5D_io_info_t *io_info,
     size_t nelmts, size_t elmt_size,
     const struct H5S_t *file_space, const struct H5S_t *mem_space,
-    haddr_t addr,const void *buf);
+    haddr_t addr, void UNUSED *pointer/*in*/, const void *buf);
 
 /* MPI-IO function to handle contiguous collective IO */
 H5_DLL herr_t
diff --git a/src/H5Dselect.c b/src/H5Dselect.c
index d451bbd..105b4e5 100644
--- a/src/H5Dselect.c
+++ b/src/H5Dselect.c
@@ -80,7 +80,7 @@ H5FL_SEQ_DEFINE_STATIC(hsize_t);
 herr_t
 H5D_select_fscat (H5D_io_info_t *io_info,
     const H5S_t *space, H5S_sel_iter_t *iter, size_t nelmts,
-    const void *_buf)
+    haddr_t chunk_addr, void *chunk/*in*/, const void *_buf)
 {
     const uint8_t *buf=_buf;       /* Alias for pointer arithmetic */
     hsize_t _off[H5D_IO_VECTOR_SIZE];             /* Array to store sequence offsets */
@@ -129,7 +129,7 @@ H5D_select_fscat (H5D_io_info_t *io_info,
         mem_off=0;
 
         /* Write sequence list out */
-        if((*io_info->ops.writevv)(io_info, nseq, &dset_curr_seq, len, off, (size_t)1, &mem_curr_seq, &mem_len, &mem_off, buf) < 0)
+        if((*io_info->ops.writevv)(io_info, nseq, &dset_curr_seq, len, off, (size_t)1, &mem_curr_seq, &mem_len, &mem_off, chunk_addr, chunk, buf) < 0)
             HGOTO_ERROR(H5E_DATASPACE, H5E_WRITEERROR, FAIL, "write error");
 
         /* Update buffer */
@@ -174,9 +174,9 @@ done:
 size_t
 H5D_select_fgath (H5D_io_info_t *io_info,
     const H5S_t *space, H5S_sel_iter_t *iter, size_t nelmts,
-    void *_buf/*out*/)
+    haddr_t chunk_addr, void *chunk/*in*/, void *_buf/*out*/)
 {
-    uint8_t *buf=_buf;          /* Alias for pointer arithmetic */
+    uint8_t *buf=(uint8_t*)_buf;          /* Alias for pointer arithmetic */
     hsize_t _off[H5D_IO_VECTOR_SIZE];          /* Array to store sequence offsets */
     hsize_t *off=NULL;          /* Pointer to sequence offsets */
     hsize_t mem_off;            /* Offset in memory */
@@ -224,7 +224,8 @@ H5D_select_fgath (H5D_io_info_t *io_info,
         mem_off=0;
 
         /* Read sequence list in */
-        if((*io_info->ops.readvv)(io_info, nseq, &dset_curr_seq, len, off, (size_t)1, &mem_curr_seq, &mem_len, &mem_off, buf) < 0)
+        if((*io_info->ops.readvv)(io_info, nseq, &dset_curr_seq, len, off, (size_t)1, 
+                                  &mem_curr_seq, &mem_len, &mem_off, chunk_addr, chunk, buf) < 0)
             HGOTO_ERROR(H5E_DATASPACE, H5E_READERROR, 0, "read error");
 
         /* Update buffer */
@@ -433,7 +434,7 @@ herr_t
 H5D_select_read(H5D_io_info_t *io_info,
     size_t nelmts, size_t elmt_size,
     const H5S_t *file_space, const H5S_t *mem_space,
-    haddr_t UNUSED addr,
+    haddr_t addr, void *chunk/*in*/,
     void *buf/*out*/)
 {
     H5S_sel_iter_t mem_iter;    /* Memory selection iteration info */
@@ -525,7 +526,7 @@ H5D_select_read(H5D_io_info_t *io_info,
         if ((tmp_file_len=(*io_info->ops.readvv)(io_info,
                 file_nseq, &curr_file_seq, file_len, file_off,
                 mem_nseq, &curr_mem_seq, mem_len, mem_off,
-                buf))<0)
+                addr, chunk, buf))<0)
             HGOTO_ERROR(H5E_DATASPACE, H5E_READERROR, FAIL, "read error");
 
         /* Decrement number of elements left to process */
@@ -577,7 +578,7 @@ herr_t
 H5D_select_write(H5D_io_info_t *io_info,
     size_t nelmts, size_t elmt_size,
     const H5S_t *file_space, const H5S_t *mem_space,
-    haddr_t UNUSED addr,
+    haddr_t addr, void *chunk/*in*/,
     const void *buf/*out*/)
 {
     H5S_sel_iter_t mem_iter;    /* Memory selection iteration info */
@@ -668,7 +669,7 @@ H5D_select_write(H5D_io_info_t *io_info,
         if ((tmp_file_len=(*io_info->ops.writevv)(io_info,
                 file_nseq, &curr_file_seq, file_len, file_off,
                 mem_nseq, &curr_mem_seq, mem_len, mem_off,
-                buf))<0)
+                addr, chunk, buf))<0)
             HGOTO_ERROR(H5E_DATASPACE, H5E_WRITEERROR, FAIL, "write error");
 
         /* Decrement number of elements left to process */