int mca_io_ompio_file_set_view (ompi_file_t *fp,
OMPI_MPI_OFFSET_TYPE disp,
ompi_datatype_t *etype,
ompi_datatype_t *filetype,
char *datarep,
ompi_info_t *info)
{
mca_io_ompio_data_t *data;
mca_io_ompio_file_t *fh;
data = (mca_io_ompio_data_t *) fp->f_io_selected_data;
fh = &data->ompio_fh;
if (NULL != fh->f_decoded_iov) {
free (fh->f_decoded_iov);
fh->f_decoded_iov = NULL;
}
if (NULL != fh->f_datarep) {
free (fh->f_datarep);
fh->f_datarep = NULL;
}
/* Reset the flags first */
fh->f_flags = 0;
fh->f_flags |= OMPIO_FILE_VIEW_IS_SET;
fh->f_datarep = strdup (datarep);
mca_io_ompio_set_view_internal (fh,
disp,
etype,
filetype,
datarep,
info);
if (OMPI_SUCCESS != mca_fcoll_base_file_select (&data->ompio_fh,
NULL)) {
opal_output(1, "mca_fcoll_base_file_select() failed\n");
return OMPI_ERROR;
}
return OMPI_SUCCESS;
}
int
ompio_io_ompio_file_open (ompi_communicator_t *comm,
const char *filename,
int amode,
ompi_info_t *info,
mca_io_ompio_file_t *ompio_fh, bool use_sharedfp)
{
int ret = OMPI_SUCCESS;
int remote_arch;
ompio_fh->f_iov_type = MPI_DATATYPE_NULL;
ompio_fh->f_comm = MPI_COMM_NULL;
if ( ((amode&MPI_MODE_RDONLY)?1:0) + ((amode&MPI_MODE_RDWR)?1:0) +
((amode&MPI_MODE_WRONLY)?1:0) != 1 ) {
return MPI_ERR_AMODE;
}
if ((amode & MPI_MODE_RDONLY) &&
((amode & MPI_MODE_CREATE) || (amode & MPI_MODE_EXCL))) {
return MPI_ERR_AMODE;
}
if ((amode & MPI_MODE_RDWR) && (amode & MPI_MODE_SEQUENTIAL)) {
return MPI_ERR_AMODE;
}
ompio_fh->f_rank = ompi_comm_rank (comm);
ompio_fh->f_size = ompi_comm_size (comm);
remote_arch = opal_local_arch;
ompio_fh->f_convertor = opal_convertor_create (remote_arch, 0);
if ( true == use_sharedfp ) {
ret = ompi_comm_dup (comm, &ompio_fh->f_comm);
if ( OMPI_SUCCESS != ret ) {
goto fn_fail;
}
}
else {
/* No need to duplicate the communicator if the file_open is called
from the sharedfp component, since the comm used as an input
is already a dup of the user level comm. */
ompio_fh->f_flags |= OMPIO_SHAREDFP_IS_SET;
ompio_fh->f_comm = comm;
}
ompio_fh->f_fstype = NONE;
ompio_fh->f_amode = amode;
ompio_fh->f_info = info;
ompio_fh->f_atomicity = 0;
ompi_io_ompio_set_file_defaults (ompio_fh);
ompio_fh->f_filename = filename;
ompio_fh->f_split_coll_req = NULL;
ompio_fh->f_split_coll_in_use = false;
/*Initialize the print_queues queues here!*/
coll_write_time = (mca_io_ompio_print_queue *) malloc (sizeof(mca_io_ompio_print_queue));
coll_read_time = (mca_io_ompio_print_queue *) malloc (sizeof(mca_io_ompio_print_queue));
ompi_io_ompio_initialize_print_queue(coll_write_time);
ompi_io_ompio_initialize_print_queue(coll_read_time);
/* set some function pointers required for fcoll, fbtls and sharedfp modules*/
ompio_fh->f_decode_datatype=ompi_io_ompio_decode_datatype;
ompio_fh->f_generate_current_file_view=ompi_io_ompio_generate_current_file_view;
ompio_fh->f_sort=ompi_io_ompio_sort;
ompio_fh->f_sort_iovec=ompi_io_ompio_sort_iovec;
ompio_fh->f_allgather_array=ompi_io_ompio_allgather_array;
ompio_fh->f_allgatherv_array=ompi_io_ompio_allgatherv_array;
ompio_fh->f_gather_array=ompi_io_ompio_gather_array;
ompio_fh->f_gatherv_array=ompi_io_ompio_gatherv_array;
ompio_fh->f_get_num_aggregators=mca_io_ompio_get_num_aggregators;
ompio_fh->f_get_bytes_per_agg=mca_io_ompio_get_bytes_per_agg;
ompio_fh->f_set_aggregator_props=ompi_io_ompio_set_aggregator_props;
ompio_fh->f_full_print_queue=ompi_io_ompio_full_print_queue;
ompio_fh->f_register_print_entry=ompi_io_ompio_register_print_entry;
/* This fix is needed for data seiving to work with
two-phase collective I/O */
if ((amode & MPI_MODE_WRONLY)){
amode -= MPI_MODE_WRONLY;
amode += MPI_MODE_RDWR;
}
/*--------------------------------------------------*/
if (OMPI_SUCCESS != (ret = mca_fs_base_file_select (ompio_fh,
NULL))) {
opal_output(1, "mca_fs_base_file_select() failed\n");
goto fn_fail;
}
if (OMPI_SUCCESS != (ret = mca_fbtl_base_file_select (ompio_fh,
NULL))) {
opal_output(1, "mca_fbtl_base_file_select() failed\n");
goto fn_fail;
}
if (OMPI_SUCCESS != (ret = mca_fcoll_base_file_select (ompio_fh,
NULL))) {
opal_output(1, "mca_fcoll_base_file_select() failed\n");
goto fn_fail;
}
ompio_fh->f_sharedfp_component = NULL; /*component*/
ompio_fh->f_sharedfp = NULL; /*module*/
ompio_fh->f_sharedfp_data = NULL; /*data*/
if ( true == use_sharedfp ) {
if (OMPI_SUCCESS != (ret = mca_sharedfp_base_file_select (ompio_fh, NULL))) {
opal_output ( ompi_io_base_framework.framework_output,
"mca_sharedfp_base_file_select() failed\n");
ompio_fh->f_sharedfp = NULL; /*module*/
/* Its ok to not have a shared file pointer module as long as the shared file
** pointer operations are not used. However, the first call to any file_read/write_shared
** function will return an error code.
*/
}
/* open the file once more for the shared file pointer if required.
** Per default, the shared file pointer specific actions are however
** only performed on first access of the shared file pointer, except
** for the addproc sharedfp component.
**
** Lazy open does not work for the addproc sharedfp
** component since it starts by spawning a process using MPI_Comm_spawn.
** For this, the first operation has to be collective which we can
** not guarantuee outside of the MPI_File_open operation.
*/
if ( NULL != ompio_fh->f_sharedfp &&
true == use_sharedfp &&
(!mca_io_ompio_sharedfp_lazy_open ||
!strcmp (ompio_fh->f_sharedfp_component->mca_component_name,
"addproc") )) {
ret = ompio_fh->f_sharedfp->sharedfp_file_open(comm,
filename,
amode,
info,
ompio_fh);
if ( OMPI_SUCCESS != ret ) {
goto fn_fail;
}
}
}
/*Determine topology information if set*/
if (ompio_fh->f_comm->c_flags & OMPI_COMM_CART){
ret = mca_io_ompio_cart_based_grouping(ompio_fh);
if(OMPI_SUCCESS != ret ){
ret = MPI_ERR_FILE;
}
}
ret = ompio_fh->f_fs->fs_file_open (comm,
filename,
amode,
info,
ompio_fh);
if ( OMPI_SUCCESS != ret ) {
ret = MPI_ERR_FILE;
goto fn_fail;
}
/* If file has been opened in the append mode, move the internal
file pointer of OMPIO to the very end of the file. */
if ( ompio_fh->f_amode & MPI_MODE_APPEND ) {
OMPI_MPI_OFFSET_TYPE current_size;
ompio_fh->f_fs->fs_file_get_size( ompio_fh,
¤t_size);
ompi_io_ompio_set_explicit_offset (ompio_fh, current_size);
}
return OMPI_SUCCESS;
fn_fail:
/* no need to free resources here, since the destructor
* is calling mca_io_ompio_file_close, which actually gets
*rid of all allocated memory items */
return ret;
}
int mca_io_ompio_set_view_internal(mca_io_ompio_file_t *fh,
OMPI_MPI_OFFSET_TYPE disp,
ompi_datatype_t *etype,
ompi_datatype_t *filetype,
char *datarep,
ompi_info_t *info)
{
size_t max_data = 0;
int i;
int num_groups = 0;
contg *contg_groups;
size_t ftype_size;
OPAL_PTRDIFF_TYPE ftype_extent, lb, ub;
ompi_datatype_t *newfiletype;
if ( NULL != fh->f_etype ) {
ompi_datatype_destroy (&fh->f_etype);
}
if ( NULL != fh->f_filetype ) {
ompi_datatype_destroy (&fh->f_filetype);
}
if ( NULL != fh->f_orig_filetype ) {
ompi_datatype_destroy (&fh->f_orig_filetype);
}
if (NULL != fh->f_decoded_iov) {
free (fh->f_decoded_iov);
fh->f_decoded_iov = NULL;
}
if (NULL != fh->f_datarep) {
free (fh->f_datarep);
fh->f_datarep = NULL;
}
/* Reset the flags first */
fh->f_flags = 0;
fh->f_flags |= OMPIO_FILE_VIEW_IS_SET;
fh->f_datarep = strdup (datarep);
ompi_datatype_duplicate (filetype, &fh->f_orig_filetype );
opal_datatype_get_extent(&filetype->super, &lb, &ftype_extent);
opal_datatype_type_size (&filetype->super, &ftype_size);
if ( etype == filetype &&
ompi_datatype_is_predefined (filetype ) &&
ftype_extent == (OPAL_PTRDIFF_TYPE)ftype_size ){
ompi_datatype_create_contiguous(MCA_IO_DEFAULT_FILE_VIEW_SIZE,
&ompi_mpi_byte.dt,
&newfiletype);
ompi_datatype_commit (&newfiletype);
}
else {
newfiletype = filetype;
}
fh->f_iov_count = 0;
fh->f_disp = disp;
fh->f_offset = disp;
fh->f_total_bytes = 0;
ompi_io_ompio_decode_datatype (fh,
newfiletype,
1,
NULL,
&max_data,
&fh->f_decoded_iov,
&fh->f_iov_count);
opal_datatype_get_extent(&newfiletype->super, &lb, &fh->f_view_extent);
opal_datatype_type_ub (&newfiletype->super, &ub);
opal_datatype_type_size (&etype->super, &fh->f_etype_size);
opal_datatype_type_size (&newfiletype->super, &fh->f_view_size);
ompi_datatype_duplicate (etype, &fh->f_etype);
ompi_datatype_duplicate (newfiletype, &fh->f_filetype);
fh->f_cc_size = get_contiguous_chunk_size (fh);
if (opal_datatype_is_contiguous_memory_layout(&etype->super,1)) {
if (opal_datatype_is_contiguous_memory_layout(&filetype->super,1) &&
fh->f_view_extent == (OPAL_PTRDIFF_TYPE)fh->f_view_size ) {
fh->f_flags |= OMPIO_CONTIGUOUS_FVIEW;
}
}
contg_groups = (contg*) calloc ( 1, fh->f_size * sizeof(contg));
if (NULL == contg_groups) {
opal_output (1, "OUT OF MEMORY\n");
return OMPI_ERR_OUT_OF_RESOURCE;
}
for( i = 0; i < fh->f_size; i++){
contg_groups[i].procs_in_contg_group = (int*)calloc (1,fh->f_size * sizeof(int));
if(NULL == contg_groups[i].procs_in_contg_group){
int j;
opal_output (1, "OUT OF MEMORY\n");
for(j=0; j<i; j++) {
free(contg_groups[j].procs_in_contg_group);
}
free(contg_groups);
return OMPI_ERR_OUT_OF_RESOURCE;
}
}
if( OMPI_SUCCESS != mca_io_ompio_fview_based_grouping(fh,
&num_groups,
contg_groups)){
opal_output(1, "mca_io_ompio_fview_based_grouping() failed\n");
free(contg_groups);
return OMPI_ERROR;
}
if( !( (fh->f_comm->c_flags & OMPI_COMM_CART) &&
(num_groups == 1 || num_groups == fh->f_size)) ) {
mca_io_ompio_finalize_initial_grouping(fh,
num_groups,
contg_groups);
}
for( i = 0; i < fh->f_size; i++){
free(contg_groups[i].procs_in_contg_group);
}
free(contg_groups);
if ( etype == filetype &&
ompi_datatype_is_predefined (filetype ) &&
ftype_extent == (OPAL_PTRDIFF_TYPE)ftype_size ){
ompi_datatype_destroy ( &newfiletype );
}
if (OMPI_SUCCESS != mca_fcoll_base_file_select (fh, NULL)) {
opal_output(1, "mca_fcoll_base_file_select() failed\n");
return OMPI_ERROR;
}
return OMPI_SUCCESS;
}
int
ompio_io_ompio_file_open (ompi_communicator_t *comm,
char *filename,
int amode,
ompi_info_t *info,
mca_io_ompio_file_t *ompio_fh, bool use_sharedfp)
{
int ret = OMPI_SUCCESS;
int remote_arch;
if ( ((amode&MPI_MODE_RDONLY)?1:0) + ((amode&MPI_MODE_RDWR)?1:0) +
((amode&MPI_MODE_WRONLY)?1:0) != 1 ) {
return MPI_ERR_AMODE;
}
if ((amode & MPI_MODE_RDONLY) &&
((amode & MPI_MODE_CREATE) || (amode & MPI_MODE_EXCL))) {
return MPI_ERR_AMODE;
}
if ((amode & MPI_MODE_RDWR) && (amode & MPI_MODE_SEQUENTIAL)) {
return MPI_ERR_AMODE;
}
ompio_fh->f_iov_type = MPI_DATATYPE_NULL;
ompio_fh->f_rank = ompi_comm_rank (comm);
ompio_fh->f_size = ompi_comm_size (comm);
remote_arch = opal_local_arch;
ompio_fh->f_convertor = opal_convertor_create (remote_arch, 0);
ret = ompi_comm_dup (comm, &ompio_fh->f_comm);
if ( ret != OMPI_SUCCESS ) {
goto fn_fail;
}
ompio_fh->f_fstype = NONE;
ompio_fh->f_amode = amode;
ompio_fh->f_info = info;
ompio_fh->f_atomicity = 0;
ompi_io_ompio_set_file_defaults (ompio_fh);
ompio_fh->f_filename = filename;
/*Initialize the print_queues queues here!*/
coll_write_time = (print_queue *) malloc (sizeof(print_queue));
coll_read_time = (print_queue *) malloc (sizeof(print_queue));
ompi_io_ompio_initialize_print_queue(coll_write_time);
ompi_io_ompio_initialize_print_queue(coll_read_time);
/*
if (MPI_INFO_NULL != info) {
ret = ompi_info_dup (info, &ompio_fh->f_info);
if (OMPI_SUCCESS != ret) {
goto fn_fail;
}
}
*/
/* This fix is needed for data seiving to work with
two-phase collective I/O */
if ((amode & MPI_MODE_WRONLY)) {
amode -= MPI_MODE_WRONLY;
amode += MPI_MODE_RDWR;
}
/*--------------------------------------------------*/
if (OMPI_SUCCESS != (ret = mca_fs_base_file_select (ompio_fh,
NULL))) {
opal_output(1, "mca_fs_base_file_select() failed\n");
goto fn_fail;
}
if (OMPI_SUCCESS != (ret = mca_fbtl_base_file_select (ompio_fh,
NULL))) {
opal_output(1, "mca_fbtl_base_file_select() failed\n");
goto fn_fail;
}
if (OMPI_SUCCESS != (ret = mca_fcoll_base_file_select (ompio_fh,
NULL))) {
opal_output(1, "mca_fcoll_base_file_select() failed\n");
goto fn_fail;
}
ompio_fh->f_sharedfp_component = NULL; /*component*/
ompio_fh->f_sharedfp = NULL; /*module*/
ompio_fh->f_sharedfp_data = NULL; /*data*/
if (OMPI_SUCCESS != (ret = mca_sharedfp_base_file_select (ompio_fh, NULL))) {
opal_output(1, "mca_sharedfp_base_file_select() failed\n");
goto fn_fail;
}
/*Determine topology information if set*/
if (ompio_fh->f_comm->c_flags & OMPI_COMM_CART) {
ret = mca_io_ompio_cart_based_grouping(ompio_fh);
if(OMPI_SUCCESS != ret ) {
ret = MPI_ERR_FILE;
}
}
ret = ompio_fh->f_fs->fs_file_open (comm,
filename,
amode,
info,
ompio_fh);
if ( OMPI_SUCCESS != ret ) {
ret = MPI_ERR_FILE;
goto fn_fail;
}
/* open the file once more for the shared file pointer if required.
** Per default, the shared file pointer specific actions are however
** only performed on first access of the shared file pointer, except
** for the addproc sharedfp component.
**
** Lazy open does not work for the addproc sharedfp
** component since it starts by spawning a process using MPI_Comm_spawn.
** For this, the first operation has to be collective which we can
** not guarantuee outside of the MPI_File_open operation.
*/
if ( true == use_sharedfp &&
(!mca_io_ompio_sharedfp_lazy_open ||
!strcmp (ompio_fh->f_sharedfp_component->mca_component_name,
"addproc") )) {
ret = ompio_fh->f_sharedfp->sharedfp_file_open(comm,
filename,
amode,
info,
ompio_fh);
if ( OMPI_SUCCESS != ret ) {
goto fn_fail;
}
}
/* If file has been opened in the append mode, move the internal
file pointer of OMPIO to the very end of the file. */
if ( ompio_fh->f_amode & MPI_MODE_APPEND ) {
OMPI_MPI_OFFSET_TYPE current_size;
ompio_fh->f_fs->fs_file_get_size( ompio_fh,
¤t_size);
ompi_io_ompio_set_explicit_offset (ompio_fh, current_size);
}
return OMPI_SUCCESS;
fn_fail:
/* no need to free resources here, since the destructor
* is calling mca_io_ompio_file_close, which actually gets
*rid of all allocated memory items */
return ret;
}
int mca_io_ompio_file_set_view (ompi_file_t *fp,
OMPI_MPI_OFFSET_TYPE disp,
ompi_datatype_t *etype,
ompi_datatype_t *filetype,
char *datarep,
ompi_info_t *info)
{
mca_io_ompio_data_t *data;
mca_io_ompio_file_t *fh;
size_t ftype_size;
OPAL_PTRDIFF_TYPE ftype_extent, lb;
data = (mca_io_ompio_data_t *) fp->f_io_selected_data;
fh = &data->ompio_fh;
ompi_datatype_destroy (&fh->f_etype);
ompi_datatype_destroy (&fh->f_filetype);
ompi_datatype_destroy (&fh->f_orig_filetype);
if (NULL != fh->f_decoded_iov) {
free (fh->f_decoded_iov);
fh->f_decoded_iov = NULL;
}
if (NULL != fh->f_datarep) {
free (fh->f_datarep);
fh->f_datarep = NULL;
}
/* Reset the flags first */
fh->f_flags = 0;
fh->f_flags |= OMPIO_FILE_VIEW_IS_SET;
fh->f_datarep = strdup (datarep);
ompi_datatype_duplicate (filetype, &fh->f_orig_filetype );
opal_datatype_get_extent(&filetype->super, &lb, &ftype_extent);
opal_datatype_type_size (&filetype->super, &ftype_size);
if ( etype == filetype &&
ompi_datatype_is_predefined (filetype ) &&
ftype_extent == (OPAL_PTRDIFF_TYPE)ftype_size ) {
ompi_datatype_t *newfiletype;
ompi_datatype_create_contiguous(MCA_IO_DEFAULT_FILE_VIEW_SIZE,
&ompi_mpi_byte.dt,
&newfiletype);
ompi_datatype_commit (&newfiletype);
mca_io_ompio_set_view_internal (fh,
disp,
etype,
newfiletype,
datarep,
info);
ompi_datatype_destroy ( &newfiletype );
}
else {
mca_io_ompio_set_view_internal (fh,
disp,
etype,
filetype,
datarep,
info);
}
if (OMPI_SUCCESS != mca_fcoll_base_file_select (&data->ompio_fh,
NULL)) {
opal_output(1, "mca_fcoll_base_file_select() failed\n");
return OMPI_ERROR;
}
return OMPI_SUCCESS;
}