[svn-r9358] Purpose:

    Bug fix

Description:
    Relax restrictions on parallel I/O to allow compressed, chunked datasets
to be read in parallel (collective access will be degraded to independent
access, but will retrieve the information still).

Platforms tested:
    FreeBSD 4.10 (sleipnir) w/parallel
    Solaris 2.7 (arabica)
    IRIX64 6.5 (modi4)
    h5committest

5 files changed, 26 insertions, 27 deletions

diff --git a/src/H5D.c b/src/H5D.c
index 41d3b53..456020b 100644
--- a/src/H5D.c
+++ b/src/H5D.c
@@ -2642,10 +2642,6 @@ H5D_open_oid(const H5G_entry_t *ent, hid_t dxpl_id)
     if(H5P_set(plist, H5D_CRT_DATA_PIPELINE_NAME, &pline) < 0)
         HGOTO_ERROR(H5E_DATASET, H5E_CANTSET, NULL, "can't set pipeline")
 
-    /* If MPI VFD is used, no filter support yet. */
-    if(IS_H5FD_MPI(dataset->ent.file) && pline.nused > 0)
-        HGOTO_ERROR (H5E_DATASET, H5E_UNSUPPORTED, NULL, "Parallel IO does not support filters yet")
-    
     /*
      * Get the raw data layout info.  It's actually stored in two locations:
      * the storage message of the dataset (dataset->storage) and certain
diff --git a/src/H5Dio.c b/src/H5Dio.c
index fbbf4dc..5ad5b53 100644
--- a/src/H5Dio.c
+++ b/src/H5Dio.c
@@ -3294,7 +3294,7 @@ H5D_ioinfo_init(H5D_t *dset, const H5D_dxpl_cache_t *dxpl_cache, hid_t dxpl_id,
     /*
      * Check if we can set direct MPI-IO read/write functions
      */
-    opt=H5D_mpio_opt_possible(dset->ent.file,mem_space,file_space,flags,&dset->shared->layout);
+    opt=H5D_mpio_opt_possible(dset,mem_space,file_space,flags);
     if(opt==FAIL)
         HGOTO_ERROR(H5E_DATASPACE, H5E_BADRANGE, NULL, "invalid check for direct IO dataspace ");
 
diff --git a/src/H5Distore.c b/src/H5Distore.c
index e84488f..9fc906d 100644
--- a/src/H5Distore.c
+++ b/src/H5Distore.c
@@ -1918,17 +1918,11 @@ 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->layout.u.chunk.size>dset->shared->cache.chunk.nbytes && dset->shared->dcpl_cache.pline.nused==0 && chunk_addr!=HADDR_UNDEF)
-            || (IS_H5FD_MPI(dset->ent.file) && (H5F_ACC_RDWR & H5F_get_intent(dset->ent.file)))) {
+    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->ent.file) && (H5F_ACC_RDWR & H5F_get_intent(dset->ent.file))))) {
         H5D_io_info_t chk_io_info;      /* Temporary I/O info object */
         H5D_storage_t chk_store;        /* Chunk storage information */
 
-#ifdef H5_HAVE_PARALLEL
-        /* Additional sanity check when operating in parallel */
-        if (chunk_addr==HADDR_UNDEF || dset->shared->dcpl_cache.pline.nused>0)
-            HGOTO_ERROR (H5E_IO, H5E_WRITEERROR, FAIL, "unable to locate raw data chunk");
-#endif /* H5_HAVE_PARALLEL */
-
         /* Set up the storage information for the chunk */
         chk_store.contig.dset_addr=chunk_addr;
         chk_store.contig.dset_size=(hsize_t)dset->shared->layout.u.chunk.size;
@@ -2121,17 +2115,21 @@ 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.
      */
-    if ((dset->shared->layout.u.chunk.size>dset->shared->cache.chunk.nbytes && dset->shared->dcpl_cache.pline.nused==0 && chunk_addr!=HADDR_UNDEF)
-            || (IS_H5FD_MPI(dset->ent.file) && (H5F_ACC_RDWR & H5F_get_intent(dset->ent.file)))) {
-        H5D_io_info_t chk_io_info;      /* Temporary I/O info object */
-        H5D_storage_t chk_store;        /* Chunk storage information */
-
 #ifdef H5_HAVE_PARALLEL
-        /* Additional sanity check when operating in parallel */
-        if (chunk_addr==HADDR_UNDEF || dset->shared->dcpl_cache.pline.nused>0)
+    /* Additional sanity checks when operating in parallel */
+    if(IS_H5FD_MPI(dset->ent.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->ent.file) && (H5F_ACC_RDWR & H5F_get_intent(dset->ent.file))))) {
+        H5D_io_info_t chk_io_info;      /* Temporary I/O info object */
+        H5D_storage_t chk_store;        /* Chunk storage information */
+
         /* Set up the storage information for the chunk */
         chk_store.contig.dset_addr=chunk_addr;
         chk_store.contig.dset_size=(hsize_t)dset->shared->layout.u.chunk.size;
diff --git a/src/H5Dmpio.c b/src/H5Dmpio.c
index 032dea9..08c6a48 100644
--- a/src/H5Dmpio.c
+++ b/src/H5Dmpio.c
@@ -63,7 +63,7 @@ H5D_mpio_spaces_xfer(H5D_io_info_t *io_info, size_t elmt_size,
  *-------------------------------------------------------------------------
  */
 htri_t
-H5D_mpio_opt_possible( const H5F_t *file, const H5S_t *mem_space, const H5S_t *file_space, const unsigned flags,const H5O_layout_t *layout)
+H5D_mpio_opt_possible( const H5D_t *dset, const H5S_t *mem_space, const H5S_t *file_space, const unsigned flags)
 {
     htri_t c1,c2;               /* Flags whether a selection is optimizable */
     htri_t ret_value=TRUE;
@@ -71,6 +71,7 @@ H5D_mpio_opt_possible( const H5F_t *file, const H5S_t *mem_space, const H5S_t *f
     FUNC_ENTER_NOAPI(H5D_mpio_opt_possible, FAIL);
 
     /* Check args */
+    assert(dset);
     assert(mem_space);
     assert(file_space);
 
@@ -116,10 +117,14 @@ H5D_mpio_opt_possible( const H5F_t *file, const H5S_t *mem_space, const H5S_t *f
         int mpi_code;               /* MPI return code */
         unsigned u;                 /* Local index variable */
 
+        /* Disallow collective I/O if there are any I/O filters on chunks */
+        if(dset->shared->dcpl_cache.pline.nused>0)
+            HGOTO_DONE(FALSE)
+
         /* Getting MPI communicator and rank */
-        if((comm = H5F_mpi_get_comm(file))==MPI_COMM_NULL)
+        if((comm = H5F_mpi_get_comm(dset->ent.file))==MPI_COMM_NULL)
             HGOTO_ERROR(H5E_DATASPACE, H5E_CANTGET, FAIL, "can't retrieve MPI communicator")
-        if((mpi_rank = H5F_mpi_get_rank(file))<0)
+        if((mpi_rank = H5F_mpi_get_rank(dset->ent.file))<0)
             HGOTO_ERROR(H5E_DATASPACE, H5E_CANTGET, FAIL, "can't retrieve MPI rank")
 
       /* Currently collective chunking storage 
@@ -150,8 +155,8 @@ H5D_mpio_opt_possible( const H5F_t *file, const H5S_t *mem_space, const H5S_t *f
         if(H5S_SELECT_BOUNDS(file_space,startf,endf)==FAIL)
             HGOTO_ERROR(H5E_DATASPACE, H5E_BADRANGE,FAIL, "invalid check for single selection blocks");
 
-        for(u=0; u < layout->u.chunk.ndims; u++) 
-            chunk_dim[u] = layout->u.chunk.dim[u];
+        for(u=0; u < dset->shared->layout.u.chunk.ndims; u++) 
+            chunk_dim[u] = dset->shared->layout.u.chunk.dim[u];
 
         /* Case 1: check whether all hyperslab in this process is inside one chunk.
            Note: we don't handle when starting point is less than zero since that may cover
diff --git a/src/H5Dpkg.h b/src/H5Dpkg.h
index a5bbeaf..499ceb0 100644
--- a/src/H5Dpkg.h
+++ b/src/H5Dpkg.h
@@ -290,8 +290,8 @@ H5_DLL herr_t H5D_mpio_spaces_write(H5D_io_info_t *io_info,
 
 /* MPI-IO function to check if a direct I/O transfer is possible between
  * memory and the file */
-H5_DLL htri_t H5D_mpio_opt_possible(const H5F_t *file, const H5S_t *mem_space,
-    const H5S_t *file_space, const unsigned flags, const H5O_layout_t *layout);
+H5_DLL htri_t H5D_mpio_opt_possible(const H5D_t *file, const H5S_t *mem_space,
+    const H5S_t *file_space, const unsigned flags);
 #endif /* H5_HAVE_PARALLEL */
 
 /* Testing functions */