int mca_common_ompio_file_iread_at (ompio_file_t *fh,
OMPI_MPI_OFFSET_TYPE offset,
void *buf,
int count,
struct ompi_datatype_t *datatype,
ompi_request_t **request)
{
int ret = OMPI_SUCCESS;
OMPI_MPI_OFFSET_TYPE prev_offset;
mca_common_ompio_file_get_position (fh, &prev_offset );
mca_common_ompio_set_explicit_offset (fh, offset);
ret = mca_common_ompio_file_iread (fh,
buf,
count,
datatype,
request);
/* An explicit offset file operation is not suppsed to modify
** the internal file pointer. So reset the pointer
** to the previous value
** It is OK to reset the position already here, althgouth
** the operation might still be pending/ongoing, since
** the entire array of <offset, length, memaddress> have
** already been constructed in the file_iread operation
*/
mca_common_ompio_set_explicit_offset (fh, prev_offset);
return ret;
}
int mca_common_ompio_file_iread_at_all (ompio_file_t *fp,
OMPI_MPI_OFFSET_TYPE offset,
void *buf,
int count,
struct ompi_datatype_t *datatype,
ompi_request_t **request)
{
int ret = OMPI_SUCCESS;
OMPI_MPI_OFFSET_TYPE prev_offset;
mca_common_ompio_file_get_position (fp, &prev_offset );
mca_common_ompio_set_explicit_offset (fp, offset);
if ( NULL != fp->f_fcoll->fcoll_file_iread_all ) {
ret = fp->f_fcoll->fcoll_file_iread_all (fp,
buf,
count,
datatype,
request);
}
else {
/* this fcoll component does not support non-blocking
collective I/O operations. WE fake it with
individual non-blocking I/O operations. */
ret = mca_common_ompio_file_iread ( fp, buf, count, datatype, request );
}
mca_common_ompio_set_explicit_offset (fp, prev_offset);
return ret;
}
int mca_common_ompio_file_read_at (ompio_file_t *fh,
OMPI_MPI_OFFSET_TYPE offset,
void *buf,
int count,
struct ompi_datatype_t *datatype,
ompi_status_public_t * status)
{
int ret = OMPI_SUCCESS;
OMPI_MPI_OFFSET_TYPE prev_offset;
mca_common_ompio_file_get_position (fh, &prev_offset );
mca_common_ompio_set_explicit_offset (fh, offset);
ret = mca_common_ompio_file_read (fh,
buf,
count,
datatype,
status);
// An explicit offset file operation is not suppsed to modify
// the internal file pointer. So reset the pointer
// to the previous value
mca_common_ompio_set_explicit_offset (fh, prev_offset);
return ret;
}
int mca_io_ompio_file_get_position (ompi_file_t *fd,
OMPI_MPI_OFFSET_TYPE *offset)
{
int ret=OMPI_SUCCESS;
mca_io_ompio_data_t *data=NULL;
mca_io_ompio_file_t *fh=NULL;
data = (mca_io_ompio_data_t *) fd->f_io_selected_data;
fh = &data->ompio_fh;
ret = mca_common_ompio_file_get_position (fh, offset);
return ret;
}
int mca_io_ompio_file_seek (ompi_file_t *fh,
OMPI_MPI_OFFSET_TYPE off,
int whence)
{
int ret = OMPI_SUCCESS;
mca_common_ompio_data_t *data;
OMPI_MPI_OFFSET_TYPE offset, temp_offset;
data = (mca_common_ompio_data_t *) fh->f_io_selected_data;
OPAL_THREAD_LOCK(&fh->f_lock);
offset = off * data->ompio_fh.f_etype_size;
switch(whence) {
case MPI_SEEK_SET:
if (offset < 0) {
OPAL_THREAD_UNLOCK(&fh->f_lock);
return OMPI_ERROR;
}
break;
case MPI_SEEK_CUR:
ret = mca_common_ompio_file_get_position (&data->ompio_fh,
&temp_offset);
offset += temp_offset;
if (offset < 0) {
OPAL_THREAD_UNLOCK(&fh->f_lock);
return OMPI_ERROR;
}
break;
case MPI_SEEK_END:
ret = data->ompio_fh.f_fs->fs_file_get_size (&data->ompio_fh,
&temp_offset);
offset += temp_offset;
if (offset < 0 || OMPI_SUCCESS != ret) {
OPAL_THREAD_UNLOCK(&fh->f_lock);
return OMPI_ERROR;
}
break;
default:
OPAL_THREAD_UNLOCK(&fh->f_lock);
return OMPI_ERROR;
}
ret = mca_common_ompio_set_explicit_offset (&data->ompio_fh,
offset/data->ompio_fh.f_etype_size);
OPAL_THREAD_UNLOCK(&fh->f_lock);
return ret;
}
/* Infrastructure for collective operations */
int mca_common_ompio_file_read_at_all (ompio_file_t *fh,
OMPI_MPI_OFFSET_TYPE offset,
void *buf,
int count,
struct ompi_datatype_t *datatype,
ompi_status_public_t * status)
{
int ret = OMPI_SUCCESS;
OMPI_MPI_OFFSET_TYPE prev_offset;
mca_common_ompio_file_get_position (fh, &prev_offset );
mca_common_ompio_set_explicit_offset (fh, offset);
ret = fh->f_fcoll->fcoll_file_read_all (fh,
buf,
count,
datatype,
status);
mca_common_ompio_set_explicit_offset (fh, prev_offset);
return ret;
}
int mca_io_ompio_file_preallocate (ompi_file_t *fh,
OMPI_MPI_OFFSET_TYPE diskspace)
{
int ret = OMPI_SUCCESS, cycles, i;
OMPI_MPI_OFFSET_TYPE tmp, current_size, size, written, len;
mca_io_ompio_data_t *data;
char *buf = NULL;
ompi_status_public_t *status = NULL;
data = (mca_io_ompio_data_t *) fh->f_io_selected_data;
OPAL_THREAD_LOCK(&fh->f_lock);
tmp = diskspace;
ret = data->ompio_fh.f_comm->c_coll->coll_bcast (&tmp,
1,
OMPI_OFFSET_DATATYPE,
OMPIO_ROOT,
data->ompio_fh.f_comm,
data->ompio_fh.f_comm->c_coll->coll_bcast_module);
if ( OMPI_SUCCESS != ret ) {
OPAL_THREAD_UNLOCK(&fh->f_lock);
return OMPI_ERROR;
}
if (tmp != diskspace) {
OPAL_THREAD_UNLOCK(&fh->f_lock);
return OMPI_ERROR;
}
ret = data->ompio_fh.f_fs->fs_file_get_size (&data->ompio_fh,
¤t_size);
if ( OMPI_SUCCESS != ret ) {
OPAL_THREAD_UNLOCK(&fh->f_lock);
return OMPI_ERROR;
}
if ( current_size > diskspace ) {
OPAL_THREAD_UNLOCK(&fh->f_lock);
return OMPI_SUCCESS;
}
/* ROMIO explanation
On file systems with no preallocation function, we have to
explicitly write to allocate space. Since there could be holes in the file,
we need to read up to the current file size, write it back,
and then write beyond that depending on how much
preallocation is needed.
*/
if (OMPIO_ROOT == data->ompio_fh.f_rank) {
OMPI_MPI_OFFSET_TYPE prev_offset;
mca_common_ompio_file_get_position (&data->ompio_fh, &prev_offset );
size = diskspace;
if (size > current_size) {
size = current_size;
}
cycles = (size + OMPIO_PREALLOC_MAX_BUF_SIZE - 1)/
OMPIO_PREALLOC_MAX_BUF_SIZE;
buf = (char *) malloc (OMPIO_PREALLOC_MAX_BUF_SIZE);
if (NULL == buf) {
opal_output(1, "OUT OF MEMORY\n");
ret = OMPI_ERR_OUT_OF_RESOURCE;
goto exit;
}
written = 0;
for (i=0; i<cycles; i++) {
len = OMPIO_PREALLOC_MAX_BUF_SIZE;
if (len > size-written) {
len = size - written;
}
ret = mca_common_ompio_file_read (&data->ompio_fh, buf, len, MPI_BYTE, status);
if (ret != OMPI_SUCCESS) {
goto exit;
}
ret = mca_common_ompio_file_write (&data->ompio_fh, buf, len, MPI_BYTE, status);
if (ret != OMPI_SUCCESS) {
goto exit;
}
written += len;
}
if (diskspace > current_size) {
memset(buf, 0, OMPIO_PREALLOC_MAX_BUF_SIZE);
size = diskspace - current_size;
cycles = (size + OMPIO_PREALLOC_MAX_BUF_SIZE - 1) /
OMPIO_PREALLOC_MAX_BUF_SIZE;
for (i=0; i<cycles; i++) {
len = OMPIO_PREALLOC_MAX_BUF_SIZE;
if (len > diskspace-written) {
len = diskspace - written;
}
ret = mca_common_ompio_file_write (&data->ompio_fh, buf, len, MPI_BYTE, status);
if (ret != OMPI_SUCCESS) {
goto exit;
}
written += len;
}
}
// This operation should not affect file pointer position.
mca_common_ompio_set_explicit_offset ( &data->ompio_fh, prev_offset);
}
exit:
free ( buf );
fh->f_comm->c_coll->coll_bcast ( &ret, 1, MPI_INT, OMPIO_ROOT, fh->f_comm,
fh->f_comm->c_coll->coll_bcast_module);
if ( diskspace > current_size ) {
data->ompio_fh.f_fs->fs_file_set_size (&data->ompio_fh, diskspace);
}
OPAL_THREAD_UNLOCK(&fh->f_lock);
return ret;
}
