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;
}
/* 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_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;
}
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;
}
int mca_common_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;
mca_common_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!*/
mca_common_ompio_initialize_print_queue(&ompio_fh->f_coll_write_time);
mca_common_ompio_initialize_print_queue(&ompio_fh->f_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_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=mca_io_ompio_set_aggregator_props;
/* 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);
mca_common_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;
}
