/*-------------------------------------------------------------------------
* Function: H5D_mpio_spaces_xfer
*
* Purpose: Use MPI-IO to transfer data efficiently
* directly between app buffer and file.
*
* Return: non-negative on success, negative on failure.
*
* Programmer: rky 980813
*
* Notes:
* For collective data transfer only since this would eventually call
* H5FD_mpio_setup to do setup to eveually call MPI_File_set_view in
* H5FD_mpio_read or H5FD_mpio_write. MPI_File_set_view is a collective
* call. Letting independent data transfer use this route would result in
* hanging.
*
* The preconditions for calling this routine are located in the
* H5S_mpio_opt_possible() routine, which determines whether this routine
* can be called for a given dataset transfer.
*
* Modifications:
* rky 980918
* Added must_convert parameter to let caller know we can't optimize
* the xfer.
*
* Albert Cheng, 001123
* Include the MPI_type freeing as part of cleanup code.
*
* QAK - 2002/04/02
* Removed the must_convert parameter and move preconditions to
* H5S_mpio_opt_possible() routine
*
* QAK - 2002/06/17
* Removed 'disp' parameter from H5FD_mpio_setup routine and use the
* address of the dataset in MPI_File_set_view() calls, as necessary.
*
* QAK - 2002/06/18
* Removed 'dc_plist' parameter, since it was not used. Also, switch to
* getting the 'extra_offset' setting for each selection.
*
*-------------------------------------------------------------------------
*/
static herr_t
H5D_mpio_spaces_xfer(H5D_io_info_t *io_info, size_t elmt_size,
const H5S_t *file_space, const H5S_t *mem_space,
void *_buf /*out*/, hbool_t do_write )
{
haddr_t addr; /* Address of dataset (or selection) within file */
size_t mpi_buf_count, mpi_file_count; /* Number of "objects" to transfer */
hsize_t mpi_buf_offset, mpi_file_offset; /* Offset within dataset where selection (ie. MPI type) begins */
MPI_Datatype mpi_buf_type, mpi_file_type; /* MPI types for buffer (memory) and file */
hbool_t mbt_is_derived=0, /* Whether the buffer (memory) type is derived and needs to be free'd */
mft_is_derived=0; /* Whether the file type is derived and needs to be free'd */
hbool_t plist_is_setup=0; /* Whether the dxpl has been customized */
uint8_t *buf=(uint8_t *)_buf; /* Alias for pointer arithmetic */
int mpi_code; /* MPI return code */
herr_t ret_value = SUCCEED; /* Return value */
FUNC_ENTER_NOAPI_NOINIT(H5D_mpio_spaces_xfer);
/* Check args */
assert (io_info);
assert (io_info->dset);
assert (file_space);
assert (mem_space);
assert (buf);
assert (IS_H5FD_MPIO(io_info->dset->ent.file));
/* Make certain we have the correct type of property list */
assert(TRUE==H5P_isa_class(io_info->dxpl_id,H5P_DATASET_XFER));
/* create the MPI buffer type */
if (H5S_mpio_space_type( mem_space, elmt_size,
/* out: */
&mpi_buf_type,
&mpi_buf_count,
&mpi_buf_offset,
&mbt_is_derived )<0)
HGOTO_ERROR(H5E_DATASPACE, H5E_BADTYPE, FAIL,"couldn't create MPI buf type");
/* create the MPI file type */
if ( H5S_mpio_space_type( file_space, elmt_size,
/* out: */
&mpi_file_type,
&mpi_file_count,
&mpi_file_offset,
&mft_is_derived )<0)
HGOTO_ERROR(H5E_DATASPACE, H5E_BADTYPE, FAIL,"couldn't create MPI file type");
/* Get the base address of the contiguous dataset or the chunk */
if(io_info->dset->shared->layout.type == H5D_CONTIGUOUS)
addr = H5D_contig_get_addr(io_info->dset) + mpi_file_offset;
else {
haddr_t chunk_addr; /* for collective chunk IO */
assert(io_info->dset->shared->layout.type == H5D_CHUNKED);
chunk_addr=H5D_istore_get_addr(io_info,NULL);
addr = H5F_BASE_ADDR(io_info->dset->ent.file) + chunk_addr + mpi_file_offset;
}
/*
* Pass buf type, file type to the file driver. Request an MPI type
* transfer (instead of an elementary byteblock transfer).
*/
if(H5FD_mpi_setup_collective(io_info->dxpl_id, mpi_buf_type, mpi_file_type)<0)
HGOTO_ERROR(H5E_PLIST, H5E_CANTSET, FAIL, "can't set MPI-I/O properties");
plist_is_setup=1;
/* Adjust the buffer pointer to the beginning of the selection */
buf+=mpi_buf_offset;
/* transfer the data */
if (do_write) {
if (H5F_block_write(io_info->dset->ent.file, H5FD_MEM_DRAW, addr, mpi_buf_count, io_info->dxpl_id, buf) <0)
HGOTO_ERROR(H5E_IO, H5E_WRITEERROR, FAIL,"MPI write failed");
} else {
if (H5F_block_read (io_info->dset->ent.file, H5FD_MEM_DRAW, addr, mpi_buf_count, io_info->dxpl_id, buf) <0)
HGOTO_ERROR(H5E_IO, H5E_READERROR, FAIL,"MPI read failed");
}
done:
/* Reset the dxpl settings */
if(plist_is_setup) {
if(H5FD_mpi_teardown_collective(io_info->dxpl_id)<0)
HDONE_ERROR(H5E_DATASPACE, H5E_CANTFREE, FAIL, "unable to reset dxpl values");
} /* end if */
/* free the MPI buf and file types */
if (mbt_is_derived) {
if (MPI_SUCCESS != (mpi_code= MPI_Type_free( &mpi_buf_type )))
HMPI_DONE_ERROR(FAIL, "MPI_Type_free failed", mpi_code);
}
if (mft_is_derived) {
if (MPI_SUCCESS != (mpi_code= MPI_Type_free( &mpi_file_type )))
HMPI_DONE_ERROR(FAIL, "MPI_Type_free failed", mpi_code);
}
FUNC_LEAVE_NOAPI(ret_value);
} /* end H5D_mpio_spaces_xfer() */
/*------------------------------------------------------------------------- * Function: H5B2_cache_hdr_flush * * Purpose: Flushes a dirty B-tree header to disk. * * Return: Non-negative on success/Negative on failure * * Programmer: Quincey Koziol * [email protected] * Feb 1 2005 * Changes: JRM -- 8/21/06 * Added the flags_ptr parameter. This parameter exists to * allow the flush routine to report to the cache if the * entry is resized or renamed as a result of the flush. * *flags_ptr is set to H5C_CALLBACK__NO_FLAGS_SET on entry. * *------------------------------------------------------------------------- */ static herr_t H5B2_cache_hdr_flush(H5F_t *f, hid_t dxpl_id, hbool_t destroy, haddr_t addr, H5B2_t *bt2, unsigned UNUSED * flags_ptr) { H5WB_t *wb = NULL; /* Wrapped buffer for header data */ uint8_t hdr_buf[H5B2_HDR_BUF_SIZE]; /* Buffer for header */ herr_t ret_value = SUCCEED; /* Return value */ FUNC_ENTER_NOAPI(H5B2_cache_hdr_flush, FAIL) /* check arguments */ HDassert(f); HDassert(H5F_addr_defined(addr)); HDassert(bt2); if (bt2->cache_info.is_dirty) { H5B2_shared_t *shared; /* Shared B-tree information */ uint8_t *hdr; /* Pointer to header buffer */ uint8_t *p; /* Pointer into raw data buffer */ size_t size; /* Header size on disk */ uint32_t metadata_chksum; /* Computed metadata checksum value */ /* Get the pointer to the shared B-tree info */ shared = (H5B2_shared_t *)H5RC_GET_OBJ(bt2->shared); HDassert(shared); /* Wrap the local buffer for serialized header info */ if(NULL == (wb = H5WB_wrap(hdr_buf, sizeof(hdr_buf)))) HGOTO_ERROR(H5E_BTREE, H5E_CANTINIT, FAIL, "can't wrap buffer") /* Compute the size of the serialized B-tree header on disk */ size = H5B2_HEADER_SIZE(f); /* Get a pointer to a buffer that's large enough for header */ if(NULL == (hdr = (uint8_t *)H5WB_actual(wb, size))) HGOTO_ERROR(H5E_BTREE, H5E_NOSPACE, FAIL, "can't get actual buffer") /* Get temporary pointer to serialized header */ p = hdr; /* Magic number */ HDmemcpy(p, H5B2_HDR_MAGIC, (size_t)H5_SIZEOF_MAGIC); p += H5_SIZEOF_MAGIC; /* Version # */ *p++ = H5B2_HDR_VERSION; /* B-tree type */ *p++ = shared->type->id; /* Node size (in bytes) */ UINT32ENCODE(p, shared->node_size); /* Raw key size (in bytes) */ UINT16ENCODE(p, shared->rrec_size); /* Depth of tree */ UINT16ENCODE(p, shared->depth); /* Split & merge %s */ H5_CHECK_OVERFLOW(shared->split_percent, /* From: */ unsigned, /* To: */ uint8_t); *p++ = (uint8_t)shared->split_percent; H5_CHECK_OVERFLOW(shared->merge_percent, /* From: */ unsigned, /* To: */ uint8_t); *p++ = (uint8_t)shared->merge_percent; /* Root node pointer */ H5F_addr_encode(f, &p, bt2->root.addr); UINT16ENCODE(p, bt2->root.node_nrec); H5F_ENCODE_LENGTH(f, p, bt2->root.all_nrec); /* Compute metadata checksum */ metadata_chksum = H5_checksum_metadata(hdr, (size - H5B2_SIZEOF_CHKSUM), 0); /* Metadata checksum */ UINT32ENCODE(p, metadata_chksum); /* Write the B-tree header. */ HDassert((size_t)(p - hdr) == size); if(H5F_block_write(f, H5FD_MEM_BTREE, addr, size, dxpl_id, hdr) < 0) HGOTO_ERROR(H5E_BTREE, H5E_CANTFLUSH, FAIL, "unable to save B-tree header to disk") bt2->cache_info.is_dirty = FALSE; } /* end if */
