/* ADIOI_GEN_IwriteContig
*
* This code handles only the case where ROMIO_HAVE_WORKING_AIO is
* defined. We post an asynchronous I/O operations using the appropriate aio
* routines. Otherwise, the ADIOI_Fns_struct will point to the FAKE
* version.
*/
void ADIOI_GEN_IwriteContig(ADIO_File fd, const void *buf, int count,
MPI_Datatype datatype, int file_ptr_type,
ADIO_Offset offset, ADIO_Request *request,
int *error_code)
{
MPI_Count len, typesize;
int aio_errno = 0;
static char myname[] = "ADIOI_GEN_IWRITECONTIG";
MPI_Type_size_x(datatype, &typesize);
len = count * typesize;
ADIOI_Assert(len == (int)((ADIO_Offset)count * (ADIO_Offset)typesize)); /* the count is an int parm */
if (file_ptr_type == ADIO_INDIVIDUAL) offset = fd->fp_ind;
/* Cast away the const'ness of 'buf' as ADIOI_GEN_aio is used for
* both read and write calls */
aio_errno = ADIOI_GEN_aio(fd, (char *) buf, len, offset, 1, request);
if (file_ptr_type == ADIO_INDIVIDUAL) fd->fp_ind += len;
fd->fp_sys_posn = -1;
/* --BEGIN ERROR HANDLING-- */
if (aio_errno != 0) {
MPIO_ERR_CREATE_CODE_ERRNO(myname, aio_errno, error_code);
return;
}
/* --END ERROR HANDLING-- */
*error_code = MPI_SUCCESS;
}
void ADIOI_TESTFS_ReadContig(ADIO_File fd, void *buf, int count,
MPI_Datatype datatype, int file_ptr_type,
ADIO_Offset offset, ADIO_Status * status, int
*error_code)
{
int myrank, nprocs;
MPI_Count datatype_size;
*error_code = MPI_SUCCESS;
MPI_Comm_size(fd->comm, &nprocs);
MPI_Comm_rank(fd->comm, &myrank);
MPI_Type_size_x(datatype, &datatype_size);
FPRINTF(stdout, "[%d/%d] ADIOI_TESTFS_ReadContig called on %s\n", myrank, nprocs, fd->filename);
if (file_ptr_type != ADIO_EXPLICIT_OFFSET) {
offset = fd->fp_ind;
fd->fp_ind += datatype_size * count;
fd->fp_sys_posn = fd->fp_ind;
} else {
fd->fp_sys_posn = offset + datatype_size * count;
}
FPRINTF(stdout, "[%d/%d] reading (buf = %p, loc = %lld, sz = %lld)\n",
myrank, nprocs, buf, (long long) offset, (long long) datatype_size * count);
#ifdef HAVE_STATUS_SET_BYTES
MPIR_Status_set_bytes(status, datatype, datatype_size * count);
#endif
}
/* ADIOI_TESTFS_IreadContig()
*
* Implemented by immediately calling ReadContig()
*/
void ADIOI_TESTFS_IreadContig(ADIO_File fd, void *buf, int count,
MPI_Datatype datatype, int file_ptr_type,
ADIO_Offset offset, ADIO_Request *request, int
*error_code)
{
ADIO_Status status;
int myrank, nprocs;
MPI_Count typesize, len;
*error_code = MPI_SUCCESS;
MPI_Comm_size(fd->comm, &nprocs);
MPI_Comm_rank(fd->comm, &myrank);
MPI_Type_size_x(datatype, &typesize);
FPRINTF(stdout, "[%d/%d] ADIOI_TESTFS_IreadContig called on %s\n",
myrank, nprocs, fd->filename);
FPRINTF(stdout, "[%d/%d] calling ADIOI_TESTFS_ReadContig\n",
myrank, nprocs);
len = count * typesize;
ADIOI_TESTFS_ReadContig(fd, buf, len, MPI_BYTE, file_ptr_type,
offset, &status, error_code);
MPIO_Completed_request_create(&fd, len, error_code, request);
}
/* ADIOI_GEN_IreadContig
*
* This code handles two distinct cases. If ROMIO_HAVE_WORKING_AIO is not
* defined, then I/O is performed in a blocking manner. Otherwise we post
* an asynchronous I/O operation using the appropriate aio routines.
*
* In the aio case we rely on ADIOI_GEN_aio(), which is implemented in
* common/ad_iwrite.c.
*/
void ADIOI_GEN_IreadContig(ADIO_File fd, void *buf, int count,
MPI_Datatype datatype, int file_ptr_type,
ADIO_Offset offset, MPI_Request *request,
int *error_code)
{
MPI_Count len, typesize;
int aio_errno = 0;
static char myname[] = "ADIOI_GEN_IREADCONTIG";
MPI_Type_size_x(datatype, &typesize);
ADIOI_Assert((count * typesize) == ((ADIO_Offset)(unsigned)count * (ADIO_Offset)typesize));
len = count * typesize;
if (file_ptr_type == ADIO_INDIVIDUAL) offset = fd->fp_ind;
aio_errno = ADIOI_GEN_aio(fd, buf, len, offset, 0, request);
if (file_ptr_type == ADIO_INDIVIDUAL) fd->fp_ind += len;
fd->fp_sys_posn = -1;
/* --BEGIN ERROR HANDLING-- */
if (aio_errno != 0) {
MPIO_ERR_CREATE_CODE_ERRNO(myname, aio_errno, error_code);
return;
}
/* --END ERROR HANDLING-- */
*error_code = MPI_SUCCESS;
}
int testtype(MPI_Datatype type, MPI_Offset expected) {
MPI_Count size, lb, extent;
int nerrors=0;
MPI_Type_size_x(type, &size);
if (size < 0) {
printf("ERROR: type size apparently overflowed integer\n");
nerrors++;
}
if (size != expected) {
printf("reported type size %lld does not match expected %lld\n",
size, expected);
nerrors++;
}
MPI_Type_get_true_extent_x(type, &lb, &extent);
if (lb != 0) {
printf("ERROR: type should have lb of 0, reported %lld\n", lb);
nerrors ++;
}
if (extent != size) {
printf("ERROR: extent should match size, not %lld\n", extent);
nerrors ++;
}
return nerrors;
}
void ADIOI_TESTFS_IwriteStrided(ADIO_File fd, const void *buf, int count,
MPI_Datatype datatype, int file_ptr_type,
ADIO_Offset offset, ADIO_Request *request, int
*error_code)
{
ADIO_Status status;
int myrank, nprocs;
MPI_Count typesize;
*error_code = MPI_SUCCESS;
MPI_Comm_size(fd->comm, &nprocs);
MPI_Comm_rank(fd->comm, &myrank);
MPI_Type_size_x(datatype, &typesize);
FPRINTF(stdout, "[%d/%d] ADIOI_TESTFS_IwriteStrided called on %s\n",
myrank, nprocs, fd->filename);
FPRINTF(stdout, "[%d/%d] calling ADIOI_TESTFS_WriteStrided\n",
myrank, nprocs);
ADIOI_TESTFS_WriteStrided(fd, buf, count, datatype, file_ptr_type,
offset, &status, error_code);
MPIO_Completed_request_create(&fd, count*typesize, error_code, request);
}
void ADIOI_NFS_WriteContig(ADIO_File fd, const void *buf, int count,
MPI_Datatype datatype, int file_ptr_type,
ADIO_Offset offset, ADIO_Status *status, int *error_code)
{
ssize_t err=-1;
MPI_Count datatype_size, len;
ADIO_Offset bytes_xfered=0;
size_t wr_count;
static char myname[] = "ADIOI_NFS_WRITECONTIG";
char *p;
MPI_Type_size_x(datatype, &datatype_size);
len = datatype_size * (ADIO_Offset)count;
if (file_ptr_type == ADIO_INDIVIDUAL) {
offset = fd->fp_ind;
}
p = (char *)buf;
while (bytes_xfered < len) {
#ifdef ADIOI_MPE_LOGGING
MPE_Log_event( ADIOI_MPE_write_a, 0, NULL );
#endif
wr_count = len - bytes_xfered;
/* work around FreeBSD and OS X defects*/
if (wr_count > INT_MAX)
wr_count = INT_MAX;
ADIOI_WRITE_LOCK(fd, offset+bytes_xfered, SEEK_SET, wr_count);
err = pwrite(fd->fd_sys, p, wr_count, offset+bytes_xfered);
/* --BEGIN ERROR HANDLING-- */
if (err == -1) {
*error_code = MPIO_Err_create_code(MPI_SUCCESS,
MPIR_ERR_RECOVERABLE,
myname, __LINE__, MPI_ERR_IO, "**io",
"**io %s", strerror(errno));
fd->fp_sys_posn = -1;
return;
}
/* --END ERROR HANDLING-- */
#ifdef ADIOI_MPE_LOGGING
MPE_Log_event( ADIOI_MPE_write_b, 0, NULL );
#endif
ADIOI_UNLOCK(fd, offset+bytes_xfered, SEEK_SET, wr_count);
bytes_xfered += err;
p += err;
}
fd->fp_sys_posn = offset + bytes_xfered;
if (file_ptr_type == ADIO_INDIVIDUAL) {
fd->fp_ind += bytes_xfered;
}
#ifdef HAVE_STATUS_SET_BYTES
MPIR_Status_set_bytes(status, datatype, bytes_xfered);
#endif
*error_code = MPI_SUCCESS;
}
void ADIOI_HFS_ReadContig(ADIO_File fd, void *buf, int count,
MPI_Datatype datatype, int file_ptr_type,
ADIO_Offset offset, ADIO_Status *status, int *error_code)
{
MPI_Count err=-1, datatype_size, len;
#ifndef PRINT_ERR_MSG
static char myname[] = "ADIOI_HFS_READCONTIG";
#endif
MPI_Type_size_x(datatype, &datatype_size);
len = datatype_size * count;
#ifdef SPPUX
fd->fp_sys_posn = -1; /* set it to null, since we are using pread */
if (file_ptr_type == ADIO_EXPLICIT_OFFSET)
err = pread64(fd->fd_sys, buf, len, offset);
else { /* read from curr. location of ind. file pointer */
err = pread64(fd->fd_sys, buf, len, fd->fp_ind);
fd->fp_ind += err;
}
#endif
#ifdef HPUX
if (file_ptr_type == ADIO_EXPLICIT_OFFSET) {
if (fd->fp_sys_posn != offset)
lseek64(fd->fd_sys, offset, SEEK_SET);
err = read(fd->fd_sys, buf, len);
fd->fp_sys_posn = offset + err;
/* individual file pointer not updated */
}
else { /* read from curr. location of ind. file pointer */
if (fd->fp_sys_posn != fd->fp_ind)
lseek64(fd->fd_sys, fd->fp_ind, SEEK_SET);
err = read(fd->fd_sys, buf, len);
fd->fp_ind += err;
fd->fp_sys_posn = fd->fp_ind;
}
#endif
#ifdef HAVE_STATUS_SET_BYTES
if (err != -1) MPIR_Status_set_bytes(status, datatype, err);
#endif
if (err == -1 ) {
#ifdef MPICH
*error_code = MPIR_Err_create_code(MPI_SUCCESS, MPIR_ERR_RECOVERABLE, myname, __LINE__, MPI_ERR_IO, "**io",
"**io %s", strerror(errno));
#elif defined(PRINT_ERR_MSG)
*error_code = (err == -1) ? MPI_ERR_UNKNOWN : MPI_SUCCESS;
#else /* MPICH-1 */
*error_code = MPIR_Err_setmsg(MPI_ERR_IO, MPIR_ADIO_ERROR,
myname, "I/O Error", "%s", strerror(errno));
ADIOI_Error(fd, *error_code, myname);
#endif
}
else *error_code = MPI_SUCCESS;
}
void ompi_type_size_x_f(MPI_Fint *type, MPI_Count *size, MPI_Fint *ierr)
{
int c_ierr;
MPI_Datatype c_type = MPI_Type_f2c(*type);
OMPI_SINGLE_NAME_DECL(size);
c_ierr = MPI_Type_size_x(c_type, size);
if (NULL != ierr) *ierr = OMPI_INT_2_FINT(c_ierr);
}
static int verify_type(char *filename, MPI_Datatype type,
int64_t expected_extent, int do_coll)
{
int rank, canary;
MPI_Count tsize;
int compare=-1;
int errs=0, toterrs=0;
MPI_Status status;
MPI_File fh;
MPI_Comm_rank(MPI_COMM_WORLD, &rank);
CHECK( MPI_File_open(MPI_COMM_WORLD, filename,
MPI_MODE_CREATE|MPI_MODE_RDWR, MPI_INFO_NULL, &fh));
CHECK( MPI_File_set_view(fh, rank*sizeof(int),
MPI_BYTE, type, "native", MPI_INFO_NULL));
MPI_Type_size_x(type, &tsize);
canary=rank+1000000;
/* skip over first instance of type */
if (do_coll) {
CHECK( MPI_File_write_at_all(fh, tsize, &canary, 1, MPI_INT, &status));
} else {
CHECK( MPI_File_write_at(fh, tsize, &canary, 1, MPI_INT, &status));
}
CHECK( MPI_File_set_view(fh, 0, MPI_INT, MPI_INT, "native",
MPI_INFO_NULL));
if (do_coll) {
CHECK( MPI_File_read_at_all(fh, expected_extent/sizeof(int)+rank,
&compare, 1, MPI_INT, &status));
} else {
CHECK( MPI_File_read_at(fh, expected_extent/sizeof(int)+rank,
&compare, 1, MPI_INT, &status));
}
if (compare != canary)
errs=1;
MPI_Allreduce(&errs, &toterrs, 1, MPI_INT, MPI_SUM, MPI_COMM_WORLD);
MPI_File_close(&fh);
if (toterrs) {
printf("%d: got %d expected %d\n", rank, compare, canary);
/* keep file if there's an error */
} else {
if (rank == 0) MPI_File_delete(filename, MPI_INFO_NULL);
}
return (toterrs);
}
void ADIOI_PANFS_ReadContig(ADIO_File fd, void *buf, int count,
MPI_Datatype datatype, int file_ptr_type,
ADIO_Offset offset, ADIO_Status *status,
int *error_code)
{
MPI_Count err = -1, datatype_size, len;
static char myname[] = "ADIOI_PANFS_READCONTIG";
MPI_Type_size_x(datatype, &datatype_size);
len = datatype_size * count;
if (file_ptr_type == ADIO_INDIVIDUAL) {
offset = fd->fp_ind;
}
if (fd->fp_sys_posn != offset) {
err = lseek(fd->fd_sys, offset, SEEK_SET);
/* --BEGIN ERROR HANDLING-- */
if (err == -1) {
*error_code = MPIO_Err_create_code(MPI_SUCCESS,
MPIR_ERR_RECOVERABLE,
myname, __LINE__,
MPI_ERR_IO, "**io",
"**io %s", strerror(errno));
fd->fp_sys_posn = -1;
return;
}
/* --END ERROR HANDLING-- */
}
AD_PANFS_RETRY(read(fd->fd_sys, buf, len),err)
/* --BEGIN ERROR HANDLING-- */
if (err == -1) {
*error_code = MPIO_Err_create_code(MPI_SUCCESS,
MPIR_ERR_RECOVERABLE,
myname, __LINE__,
MPI_ERR_IO, "**io",
"**io %s", strerror(errno));
fd->fp_sys_posn = -1;
return;
}
/* --END ERROR HANDLING-- */
fd->fp_sys_posn = offset + err;
if (file_ptr_type == ADIO_INDIVIDUAL) {
fd->fp_ind += err;
}
#ifdef HAVE_STATUS_SET_BYTES
if (err != -1) MPIR_Status_set_bytes(status, datatype, err);
#endif
*error_code = MPI_SUCCESS;
}
static void typestats(MPI_Datatype type)
{
MPI_Aint lb, extent;
MPI_Count size;
MPI_Type_get_extent(type, &lb, &extent);
MPI_Type_size_x(type, &size);
printf("dtype %d: lb = %ld extent = %ld size = %ld...",
type, (long)lb, (long)extent, size);
}
void ADIOI_PIOFS_WriteContig(ADIO_File fd, void *buf, int count,
MPI_Datatype datatype, int file_ptr_type,
ADIO_Offset offset, ADIO_Status *status, int *error_code)
{
MPI_Count err=-1, datatype_size, len;
#ifndef PRINT_ERR_MSG
static char myname[] = "ADIOI_PIOFS_WRITECONTIG";
#endif
MPI_Type_size_x(datatype, &datatype_size);
len = datatype_size * count;
if (file_ptr_type == ADIO_EXPLICIT_OFFSET) {
if (fd->fp_sys_posn != offset) {
llseek(fd->fd_sys, offset, SEEK_SET);
}
err = write(fd->fd_sys, buf, len);
fd->fp_sys_posn = offset + err;
/* individual file pointer not updated */
}
else { /* write from curr. location of ind. file pointer */
if (fd->fp_sys_posn != fd->fp_ind) {
llseek(fd->fd_sys, fd->fp_ind, SEEK_SET);
}
err = write(fd->fd_sys, buf, len);
fd->fp_ind += err;
fd->fp_sys_posn = fd->fp_ind;
}
#ifdef HAVE_STATUS_SET_BYTES
if (err != -1) MPIR_Status_set_bytes(status, datatype, err);
#endif
if (err == -1) {
#ifdef MPICH
*error_code = MPIR_Err_create_code(MPI_SUCCESS, MPIR_ERR_RECOVERABLE, myname, __LINE__, MPI_ERR_IO, "**io",
"**io %s", strerror(errno));
#elif defined(PRINT_ERR_MSG)
*error_code = MPI_ERR_UNKNOWN;
#else
*error_code = MPIR_Err_setmsg(MPI_ERR_IO, MPIR_ADIO_ERROR,
myname, "I/O Error", "%s", strerror(errno));
ADIOI_Error(fd, *error_code, myname);
#endif
}
else *error_code = MPI_SUCCESS;
}
void ADIOI_Get_position(ADIO_File fd, ADIO_Offset *offset)
{
ADIOI_Flatlist_node *flat_file;
int i, flag;
MPI_Count filetype_size, etype_size;
int filetype_is_contig;
MPI_Aint filetype_extent;
ADIO_Offset disp, byte_offset, sum=0, size_in_file, n_filetypes, frd_size;
ADIOI_Datatype_iscontig(fd->filetype, &filetype_is_contig);
etype_size = fd->etype_size;
if (filetype_is_contig) *offset = (fd->fp_ind - fd->disp)/etype_size;
else {
flat_file = ADIOI_Flatten_and_find(fd->filetype);
MPI_Type_size_x(fd->filetype, &filetype_size);
MPI_Type_extent(fd->filetype, &filetype_extent);
disp = fd->disp;
byte_offset = fd->fp_ind;
n_filetypes = -1;
flag = 0;
while (!flag) {
sum = 0;
n_filetypes++;
for (i=0; i<flat_file->count; i++) {
sum += flat_file->blocklens[i];
if (disp + flat_file->indices[i] +
n_filetypes* ADIOI_AINT_CAST_TO_OFFSET filetype_extent + flat_file->blocklens[i]
>= byte_offset) {
frd_size = disp + flat_file->indices[i] +
n_filetypes * ADIOI_AINT_CAST_TO_OFFSET filetype_extent
+ flat_file->blocklens[i] - byte_offset;
sum -= frd_size;
flag = 1;
break;
}
}
}
size_in_file = n_filetypes * (ADIO_Offset)filetype_size + sum;
*offset = size_in_file/etype_size;
}
}
/* Generic implementation of IreadStrided calls the blocking ReadStrided
* immediately.
*/
void ADIOI_FAKE_IreadStrided(ADIO_File fd, void *buf, int count,
MPI_Datatype datatype, int file_ptr_type,
ADIO_Offset offset, ADIO_Request *request,
int *error_code)
{
ADIO_Status status;
MPI_Count typesize;
MPI_Offset nbytes=0;
/* Call the blocking function. It will create an error code
* if necessary.
*/
ADIO_ReadStrided(fd, buf, count, datatype, file_ptr_type,
offset, &status, error_code);
if (*error_code == MPI_SUCCESS) {
MPI_Type_size_x(datatype, &typesize);
nbytes = (MPI_Offset)count*(MPI_Offset)typesize;
}
MPIO_Completed_request_create(&fd, nbytes, error_code, request);
}
void ADIOI_Get_byte_offset(ADIO_File fd, ADIO_Offset offset, ADIO_Offset * disp)
{
ADIOI_Flatlist_node *flat_file;
int i;
ADIO_Offset n_filetypes, etype_in_filetype, sum, abs_off_in_filetype = 0, size_in_filetype;
MPI_Count n_etypes_in_filetype, filetype_size, etype_size;
int filetype_is_contig;
MPI_Aint filetype_extent;
ADIOI_Datatype_iscontig(fd->filetype, &filetype_is_contig);
etype_size = fd->etype_size;
if (filetype_is_contig)
*disp = fd->disp + etype_size * offset;
else {
flat_file = ADIOI_Flatten_and_find(fd->filetype);
MPI_Type_size_x(fd->filetype, &filetype_size);
n_etypes_in_filetype = filetype_size / etype_size;
n_filetypes = offset / n_etypes_in_filetype;
etype_in_filetype = offset % n_etypes_in_filetype;
size_in_filetype = etype_in_filetype * etype_size;
sum = 0;
for (i = 0; i < flat_file->count; i++) {
sum += flat_file->blocklens[i];
if (sum > size_in_filetype) {
abs_off_in_filetype = flat_file->indices[i] +
size_in_filetype - (sum - flat_file->blocklens[i]);
break;
}
}
/* abs. offset in bytes in the file */
MPI_Type_extent(fd->filetype, &filetype_extent);
*disp =
fd->disp + n_filetypes * ADIOI_AINT_CAST_TO_OFFSET filetype_extent +
abs_off_in_filetype;
}
}
void ADIOI_PFS_ReadContig(ADIO_File fd, void *buf, int count,
MPI_Datatype datatype, int file_ptr_type,
ADIO_Offset offset, ADIO_Status *status, int *error_code)
{
MPI_Count err=-1, datatype_size, len;
static char myname[] = "ADIOI_PFS_READCONTIG";
MPI_Type_size_x(datatype, &datatype_size);
len = datatype_size * count;
if (file_ptr_type == ADIO_EXPLICIT_OFFSET) {
if (fd->fp_sys_posn != offset) {
lseek(fd->fd_sys, offset, SEEK_SET);
}
err = _cread(fd->fd_sys, buf, len);
fd->fp_sys_posn = offset + err;
/* individual file pointer not updated */
}
else { /* read from curr. location of ind. file pointer */
if (fd->fp_sys_posn != fd->fp_ind) {
lseek(fd->fd_sys, fd->fp_ind, SEEK_SET);
}
err = _cread(fd->fd_sys, buf, len);
fd->fp_ind += err;
fd->fp_sys_posn = fd->fp_ind;
}
#ifdef HAVE_STATUS_SET_BYTES
if (err != -1) MPIR_Status_set_bytes(status, datatype, err);
#endif
if (err == -1) {
*error_code = MPIO_Err_create_code(MPI_SUCCESS, MPIR_ERR_RECOVERABLE,
myname, __LINE__, MPI_ERR_IO,
"**io",
"**io %s", strerror(errno));
}
else *error_code = MPI_SUCCESS;
}
/* Since ADIOI_Flatten_datatype won't add a contig datatype to the
* ADIOI_Flatlist, we can force it to do so with this function. */
ADIOI_Flatlist_node * ADIOI_Add_contig_flattened(MPI_Datatype contig_type)
{
MPI_Count contig_type_sz = -1;
ADIOI_Flatlist_node *flat_node_p = ADIOI_Flatlist;
/* Add contig type to the end of the list if it doesn't already
* exist. */
while (flat_node_p->next)
{
if (flat_node_p->type == contig_type)
return flat_node_p;
flat_node_p = flat_node_p->next;
}
if (flat_node_p->type == contig_type)
return flat_node_p;
MPI_Type_size_x(contig_type, &contig_type_sz);
if ((flat_node_p->next = (ADIOI_Flatlist_node *) ADIOI_Malloc
(sizeof(ADIOI_Flatlist_node))) == NULL)
{
fprintf(stderr, "ADIOI_Add_contig_flattened: malloc next failed\n");
}
flat_node_p = flat_node_p->next;
flat_node_p->type = contig_type;
if ((flat_node_p->blocklens = (ADIO_Offset *) ADIOI_Malloc(sizeof(ADIO_Offset))) == NULL)
{
fprintf(stderr, "ADIOI_Flatlist_node: malloc blocklens failed\n");
}
if ((flat_node_p->indices = (ADIO_Offset *)
ADIOI_Malloc(sizeof(ADIO_Offset))) == NULL)
{
fprintf(stderr, "ADIOI_Flatlist_node: malloc indices failed\n");
}
flat_node_p->blocklens[0] = contig_type_sz;
flat_node_p->indices[0] = 0;
flat_node_p->count = 1;
flat_node_p->next = NULL;
return flat_node_p;
}
static void ADIOI_GEN_IreadStridedColl_fini(ADIOI_NBC_Request *nbc_req,
int *error_code)
{
ADIOI_GEN_IreadStridedColl_vars *vars = nbc_req->data.rd.rsc_vars;
MPI_Count size;
/* This is a temporary way of filling in status. The right way is to
keep track of how much data was actually read and placed in buf
during collective I/O. */
MPI_Type_size_x(vars->datatype, &size);
nbc_req->nbytes = size * vars->count;
/* free the struct for parameters and variables */
if (nbc_req->data.rd.rsc_vars) {
ADIOI_Free(nbc_req->data.rd.rsc_vars);
nbc_req->data.rd.rsc_vars = NULL;
}
/* make the request complete */
*error_code = MPI_Grequest_complete(nbc_req->req);
nbc_req->data.rd.state = ADIOI_IRC_STATE_COMPLETE;
}
/* Generic implementation of IreadContig calls the blocking ReadContig
* immediately.
*/
void ADIOI_FAKE_IreadContig(ADIO_File fd, void *buf, int count,
MPI_Datatype datatype, int file_ptr_type,
ADIO_Offset offset, ADIO_Request *request,
int *error_code)
{
ADIO_Status status;
MPI_Count typesize;
MPI_Offset len;
MPI_Type_size_x(datatype, &typesize);
len = (MPI_Offset)count * (MPI_Offset)typesize;
/* Call the blocking function. It will create an error code
* if necessary.
*/
ADIOI_Assert(len == (int) len); /* the count is an int parm */
ADIO_ReadContig(fd, buf, (int)len, MPI_BYTE, file_ptr_type, offset,
&status, error_code);
if (*error_code != MPI_SUCCESS) {
len=0;
}
MPIO_Completed_request_create(&fd, len, error_code, request);
}
int main(int argc, char *argv[])
{
int rank, size;
MPI_Datatype type;
int errs = 0, mpi_errno, errclass;
MTest_Init(&argc, &argv);
MPI_Comm_size(MPI_COMM_WORLD, &size);
MPI_Comm_rank(MPI_COMM_WORLD, &rank);
MPI_Errhandler_set(MPI_COMM_WORLD, MPI_ERRORS_RETURN);
/* Checking type_size_x for NULL variable */
type = MPI_INT;
mpi_errno = MPI_Type_size_x(type, NULL);
MPI_Error_class(mpi_errno, &errclass);
if (errclass != MPI_ERR_ARG)
++errs;
MPI_Type_free(&type);
MTest_Finalize(errs);
return 0;
}
void ADIOI_PVFS2_AIO_contig(ADIO_File fd, void *buf, int count,
MPI_Datatype datatype, int file_ptr_type,
ADIO_Offset offset, MPI_Request *request,
int flag, int *error_code)
{
int ret;
MPI_Count datatype_size, len;
ADIOI_PVFS2_fs *pvfs_fs;
ADIOI_AIO_Request *aio_req;
static char myname[] = "ADIOI_PVFS2_AIO_contig";
pvfs_fs = (ADIOI_PVFS2_fs*)fd->fs_ptr;
aio_req = (ADIOI_AIO_Request*)ADIOI_Calloc(sizeof(ADIOI_AIO_Request), 1);
MPI_Type_size_x(datatype, &datatype_size);
len = datatype_size * count;
ret = PVFS_Request_contiguous(len, PVFS_BYTE, &(aio_req->mem_req));
/* --BEGIN ERROR HANDLING-- */
if (ret != 0) {
*error_code = MPIO_Err_create_code(MPI_SUCCESS,
MPIR_ERR_RECOVERABLE,
myname, __LINE__,
ADIOI_PVFS2_error_convert(ret),
"Error in pvfs_request_contig (memory)", 0);
return;
}
/* --END ERROR HANDLING-- */
ret = PVFS_Request_contiguous(len, PVFS_BYTE, &(aio_req->file_req));
/* --BEGIN ERROR HANDLING-- */
if (ret != 0) {
*error_code = MPIO_Err_create_code(MPI_SUCCESS,
MPIR_ERR_RECOVERABLE,
myname, __LINE__,
ADIOI_PVFS2_error_convert(ret),
"Error in pvfs_request_contig (file)", 0);
return;
}
/* --END ERROR HANDLING-- */
if (file_ptr_type == ADIO_INDIVIDUAL) {
/* copy individual file pointer into offset variable, continue */
offset = fd->fp_ind;
}
if (flag == READ) {
#ifdef ADIOI_MPE_LOGGING
MPE_Log_event( ADIOI_MPE_iread_a, 0, NULL );
#endif
ret = PVFS_isys_read(pvfs_fs->object_ref, aio_req->file_req, offset,
buf, aio_req->mem_req, &(pvfs_fs->credentials),
&(aio_req->resp_io), &(aio_req->op_id), NULL);
#ifdef ADIOI_MPE_LOGGING
MPE_Log_event( ADIOI_MPE_iread_b, 0, NULL );
#endif
} else if (flag == WRITE) {
#ifdef ADIOI_MPE_LOGGING
MPE_Log_event( ADIOI_MPE_iwrite_a, 0, NULL );
#endif
ret = PVFS_isys_write(pvfs_fs->object_ref, aio_req->file_req, offset,
buf, aio_req->mem_req, &(pvfs_fs->credentials),
&(aio_req->resp_io), &(aio_req->op_id), NULL);
#ifdef ADIOI_MPE_LOGGING
MPE_Log_event( ADIOI_MPE_iwrite_b, 0, NULL );
#endif
}
/* --BEGIN ERROR HANDLING-- */
if (ret < 0 ) {
*error_code = MPIO_Err_create_code(MPI_SUCCESS,
MPIR_ERR_RECOVERABLE,
myname, __LINE__,
ADIOI_PVFS2_error_convert(ret),
"Error in PVFS_isys_io", 0);
goto fn_exit;
}
/* --END ERROR HANDLING-- */
#ifdef HAVE_MPI_GREQUEST_EXTENSIONS
/* posted. defered completion */
if (ret == 0) {
if (ADIOI_PVFS2_greq_class == 0) {
MPIX_Grequest_class_create(ADIOI_GEN_aio_query_fn,
ADIOI_PVFS2_aio_free_fn, MPIU_Greq_cancel_fn,
ADIOI_PVFS2_aio_poll_fn, ADIOI_PVFS2_aio_wait_fn,
&ADIOI_PVFS2_greq_class);
}
MPIX_Grequest_class_allocate(ADIOI_PVFS2_greq_class, aio_req, request);
memcpy(&(aio_req->req), request, sizeof(*request));
}
#else
/* if generalized request extensions not available, we will have to process
* this operation right here */
int error;
ret = PVFS_sys_wait(aio_req->op_id, "ADIOI_PVFS2_AIO_Contig", &error);
if (ret == 0) {
MPIO_Completed_request_create(&fd, len, error_code, request);
}
#endif
/* immediate completion */
if (ret == 1) {
MPIO_Completed_request_create(&fd, len, error_code, request);
}
if (file_ptr_type == ADIO_INDIVIDUAL) {
fd->fp_ind += len;
}
fd->fp_sys_posn = offset + len;
*error_code = MPI_SUCCESS;
fn_exit:
return;
}
void ADIOI_GEN_ReadStrided(ADIO_File fd, void *buf, int count,
MPI_Datatype datatype, int file_ptr_type,
ADIO_Offset offset, ADIO_Status *status, int
*error_code)
{
/* offset is in units of etype relative to the filetype. */
ADIOI_Flatlist_node *flat_buf, *flat_file;
ADIO_Offset i_offset, new_brd_size, brd_size, size;
int i, j, k, st_index=0;
MPI_Count num, bufsize;
int n_etypes_in_filetype;
ADIO_Offset n_filetypes, etype_in_filetype, st_n_filetypes, size_in_filetype;
ADIO_Offset abs_off_in_filetype=0, new_frd_size, frd_size=0, st_frd_size;
MPI_Count filetype_size, etype_size, buftype_size, partial_read;
MPI_Aint filetype_extent, buftype_extent;
int buf_count, buftype_is_contig, filetype_is_contig;
ADIO_Offset userbuf_off, req_len, sum;
ADIO_Offset off, req_off, disp, end_offset=0, readbuf_off, start_off;
char *readbuf, *tmp_buf, *value;
int info_flag;
unsigned max_bufsize, readbuf_len;
ADIO_Status status1;
if (fd->hints->ds_read == ADIOI_HINT_DISABLE) {
/* if user has disabled data sieving on reads, use naive
* approach instead.
*/
ADIOI_GEN_ReadStrided_naive(fd,
buf,
count,
datatype,
file_ptr_type,
offset,
status,
error_code);
return;
}
*error_code = MPI_SUCCESS; /* changed below if error */
ADIOI_Datatype_iscontig(datatype, &buftype_is_contig);
ADIOI_Datatype_iscontig(fd->filetype, &filetype_is_contig);
MPI_Type_size_x(fd->filetype, &filetype_size);
if ( ! filetype_size ) {
#ifdef HAVE_STATUS_SET_BYTES
MPIR_Status_set_bytes(status, datatype, 0);
#endif
*error_code = MPI_SUCCESS;
return;
}
MPI_Type_extent(fd->filetype, &filetype_extent);
MPI_Type_size_x(datatype, &buftype_size);
MPI_Type_extent(datatype, &buftype_extent);
etype_size = fd->etype_size;
ADIOI_Assert((buftype_size * count) == ((ADIO_Offset)(MPI_Count)buftype_size * (ADIO_Offset)count));
bufsize = buftype_size * count;
/* get max_bufsize from the info object. */
value = (char *) ADIOI_Malloc((MPI_MAX_INFO_VAL+1)*sizeof(char));
ADIOI_Info_get(fd->info, "ind_rd_buffer_size", MPI_MAX_INFO_VAL, value,
&info_flag);
max_bufsize = atoi(value);
ADIOI_Free(value);
if (!buftype_is_contig && filetype_is_contig) {
/* noncontiguous in memory, contiguous in file. */
flat_buf = ADIOI_Flatten_and_find(datatype);
off = (file_ptr_type == ADIO_INDIVIDUAL) ? fd->fp_ind :
fd->disp + (ADIO_Offset)etype_size * offset;
start_off = off;
end_offset = off + bufsize - 1;
readbuf_off = off;
readbuf = (char *) ADIOI_Malloc(max_bufsize);
readbuf_len = (unsigned) (MPL_MIN(max_bufsize, end_offset-readbuf_off+1));
/* if atomicity is true, lock (exclusive) the region to be accessed */
if ((fd->atomicity) && ADIO_Feature(fd, ADIO_LOCKS))
ADIOI_WRITE_LOCK(fd, start_off, SEEK_SET, end_offset-start_off+1);
ADIO_ReadContig(fd, readbuf, readbuf_len, MPI_BYTE,
ADIO_EXPLICIT_OFFSET, readbuf_off, &status1, error_code);
if (*error_code != MPI_SUCCESS) return;
for (j=0; j<count; j++)
{
for (i=0; i<flat_buf->count; i++) {
userbuf_off = (ADIO_Offset)j*(ADIO_Offset)buftype_extent + flat_buf->indices[i];
req_off = off;
req_len = flat_buf->blocklens[i];
ADIOI_BUFFERED_READ
off += flat_buf->blocklens[i];
}
}
if ((fd->atomicity) && ADIO_Feature(fd, ADIO_LOCKS))
ADIOI_UNLOCK(fd, start_off, SEEK_SET, end_offset-start_off+1);
if (file_ptr_type == ADIO_INDIVIDUAL) fd->fp_ind = off;
ADIOI_Free(readbuf);
}
else { /* noncontiguous in file */
flat_file = ADIOI_Flatten_and_find(fd->filetype);
disp = fd->disp;
if (file_ptr_type == ADIO_INDIVIDUAL) {
/* Wei-keng reworked type processing to be a bit more efficient */
offset = fd->fp_ind - disp;
n_filetypes = (offset - flat_file->indices[0]) / filetype_extent;
offset -= (ADIO_Offset)n_filetypes * filetype_extent;
/* now offset is local to this extent */
/* find the block where offset is located, skip blocklens[i]==0 */
for (i=0; i<flat_file->count; i++) {
ADIO_Offset dist;
if (flat_file->blocklens[i] == 0) continue;
dist = flat_file->indices[i] + flat_file->blocklens[i] - offset;
/* frd_size is from offset to the end of block i */
if (dist == 0) {
i++;
offset = flat_file->indices[i];
frd_size = flat_file->blocklens[i];
break;
}
if (dist > 0) {
frd_size = dist;
break;
}
}
st_index = i; /* starting index in flat_file->indices[] */
offset += disp + (ADIO_Offset)n_filetypes*filetype_extent;
}
else {
n_etypes_in_filetype = filetype_size/etype_size;
n_filetypes = offset / n_etypes_in_filetype;
etype_in_filetype = offset % n_etypes_in_filetype;
size_in_filetype = etype_in_filetype * etype_size;
sum = 0;
for (i=0; i<flat_file->count; i++) {
sum += flat_file->blocklens[i];
if (sum > size_in_filetype) {
st_index = i;
frd_size = sum - size_in_filetype;
abs_off_in_filetype = flat_file->indices[i] +
size_in_filetype - (sum - flat_file->blocklens[i]);
break;
}
}
/* abs. offset in bytes in the file */
offset = disp + (ADIO_Offset) n_filetypes*filetype_extent +
abs_off_in_filetype;
}
start_off = offset;
/* Wei-keng Liao: read request is within a single flat_file contig
* block e.g. with subarray types that actually describe the whole
* array */
if (buftype_is_contig && bufsize <= frd_size) {
/* a count of bytes can overflow. operate on original type instead */
ADIO_ReadContig(fd, buf, count, datatype, ADIO_EXPLICIT_OFFSET,
offset, status, error_code);
if (file_ptr_type == ADIO_INDIVIDUAL) {
/* update MPI-IO file pointer to point to the first byte that
* can be accessed in the fileview. */
fd->fp_ind = offset + bufsize;
if (bufsize == frd_size) {
do {
st_index++;
if (st_index == flat_file->count) {
st_index = 0;
n_filetypes++;
}
} while (flat_file->blocklens[st_index] == 0);
fd->fp_ind = disp + flat_file->indices[st_index]
+ n_filetypes*filetype_extent;
}
}
fd->fp_sys_posn = -1; /* set it to null. */
#ifdef HAVE_STATUS_SET_BYTES
MPIR_Status_set_bytes(status, datatype, bufsize);
#endif
return;
}
/* Calculate end_offset, the last byte-offset that will be accessed.
e.g., if start_offset=0 and 100 bytes to be read, end_offset=99*/
st_frd_size = frd_size;
st_n_filetypes = n_filetypes;
i_offset = 0;
j = st_index;
off = offset;
frd_size = MPL_MIN(st_frd_size, bufsize);
while (i_offset < bufsize) {
i_offset += frd_size;
end_offset = off + frd_size - 1;
j = (j+1) % flat_file->count;
n_filetypes += (j == 0) ? 1 : 0;
while (flat_file->blocklens[j]==0) {
j = (j+1) % flat_file->count;
n_filetypes += (j == 0) ? 1 : 0;
}
off = disp + flat_file->indices[j] + n_filetypes*(ADIO_Offset)filetype_extent;
frd_size = MPL_MIN(flat_file->blocklens[j], bufsize-i_offset);
}
/* if atomicity is true, lock (exclusive) the region to be accessed */
if ((fd->atomicity) && ADIO_Feature(fd, ADIO_LOCKS))
ADIOI_WRITE_LOCK(fd, start_off, SEEK_SET, end_offset-start_off+1);
readbuf_off = 0;
readbuf_len = 0;
readbuf = (char *) ADIOI_Malloc(max_bufsize);
if (buftype_is_contig && !filetype_is_contig) {
/* contiguous in memory, noncontiguous in file. should be the most
common case. */
i_offset = 0;
j = st_index;
off = offset;
n_filetypes = st_n_filetypes;
frd_size = MPL_MIN(st_frd_size, bufsize);
while (i_offset < bufsize) {
if (frd_size) {
/* TYPE_UB and TYPE_LB can result in
frd_size = 0. save system call in such cases */
/* lseek(fd->fd_sys, off, SEEK_SET);
err = read(fd->fd_sys, ((char *) buf) + i, frd_size);*/
req_off = off;
req_len = frd_size;
userbuf_off = i_offset;
ADIOI_BUFFERED_READ
}
i_offset += frd_size;
if (off + frd_size < disp + flat_file->indices[j] +
flat_file->blocklens[j] + n_filetypes*(ADIO_Offset)filetype_extent)
off += frd_size;
/* did not reach end of contiguous block in filetype.
no more I/O needed. off is incremented by frd_size. */
else {
j = (j+1) % flat_file->count;
n_filetypes += (j == 0) ? 1 : 0;
while (flat_file->blocklens[j]==0) {
j = (j+1) % flat_file->count;
n_filetypes += (j == 0) ? 1 : 0;
}
off = disp + flat_file->indices[j] +
n_filetypes*(ADIO_Offset)filetype_extent;
frd_size = MPL_MIN(flat_file->blocklens[j], bufsize-i_offset);
}
}
}
else {
/*----< main() >------------------------------------------------------------*/
int main(int argc, char **argv)
{
int i, j, err, rank, np, num_io;
char *buf, *filename;
int rank_dim[2], array_of_sizes[2];
int array_of_subsizes[2];
int count, *blocklengths, global_array_size;
MPI_Count ftype_size;
MPI_Aint *displacements;
MPI_File fh;
MPI_Datatype ftype;
MPI_Request *request;
MPI_Status *statuses;
MPI_Status status;
MPI_Offset offset = 0;
int nr_errors = 0;
#ifdef VERBOSE
int k;
#endif
MPI_Init(&argc, &argv);
MPI_Comm_rank(MPI_COMM_WORLD, &rank);
MPI_Comm_size(MPI_COMM_WORLD, &np);
if (np != 4) {
if (!rank)
printf("Please run with 4 processes. Exiting ...\n\n");
MPI_Finalize();
return 1;
}
filename = (argc > 1) ? argv[1] : "testfile";
num_io = 2;
request = (MPI_Request *) malloc(num_io * sizeof(MPI_Request));
statuses = (MPI_Status *) malloc(num_io * sizeof(MPI_Status));
/*-----------------------------------------------------------------------*/
/* process rank in each dimension */
rank_dim[0] = rank / 2;
rank_dim[1] = rank % 2;
/* global 2D array size */
array_of_sizes[0] = YLEN * 2;
array_of_sizes[1] = XLEN * 2;
global_array_size = array_of_sizes[0] * array_of_sizes[1];
array_of_subsizes[0] = YLEN / 2;
array_of_subsizes[1] = XLEN * SUB_XLEN / 5;
offset = rank_dim[0] * YLEN * array_of_sizes[1] + rank_dim[1] * XLEN;
/* define data type for file view */
count = array_of_subsizes[0] * 2; /* 2 is the no. blocks along X */
blocklengths = (int *) malloc(count * sizeof(int));
displacements = (MPI_Aint *) malloc(count * sizeof(MPI_Aint));
for (i = 0; i < count; i++)
blocklengths[i] = array_of_subsizes[1] / 2;
for (i = 0; i < array_of_subsizes[0]; i++)
for (j = 0; j < 2; j++)
displacements[i * 2 + j] = offset + i * 2 * array_of_sizes[1]
+ j * XLEN / 2;
MPI_Type_create_hindexed(count, blocklengths, displacements, MPI_CHAR, &ftype);
MPI_Type_commit(&ftype);
MPI_Type_size_x(ftype, &ftype_size);
/* subarray's layout in the global array
P0's 's layout P1's layout
[ 0 1 2 3 4 5 6 7 8 9 0 1 2 3 4 5 6 7 8 9] | [ 0 1 2 3 4 5 6 7 8 9 0 1 2 3 4 5 6 7 8 9]
[ 0] 0 1 2 3 4 5 | D E F G H I
[ 1] |
[ 2] 6 7 8 9 : ; | J K L M N O
[ 3] |
[ 4] |
[ 5] |
[ 6] |
[ 7] |
[ 8] |
[ 9] |
P2's 's layout P3's layout
[ 0 1 2 3 4 5 6 7 8 9 0 1 2 3 4 5 6 7 8 9] | [ 0 1 2 3 4 5 6 7 8 9 0 1 2 3 4 5 6 7 8 9]
[ 0] |
[ 1] |
[ 2] |
[ 3] |
[ 4] |
[ 5] X Y Z [ \ ] | l m n o p q
[ 6] |
[ 7] ^ _ ` a b c | r s t u v w
[ 8] |
[ 9] |
*/
/* initialize the write buffer */
buf = (char *) malloc(array_of_subsizes[0] * array_of_subsizes[1]);
for (i = 0; i < array_of_subsizes[0] * array_of_subsizes[1]; i++)
buf[i] = '0' + rank * 20 + i % 79;
/* zero file contents --------------------------------------------------- */
if (rank == 0) {
char *wr_buf = (char *) calloc(num_io * global_array_size, 1);
MPI_File_open(MPI_COMM_SELF, filename,
MPI_MODE_CREATE | MPI_MODE_WRONLY, MPI_INFO_NULL, &fh);
MPI_File_write(fh, wr_buf, num_io * global_array_size, MPI_CHAR, &status);
MPI_File_close(&fh);
free(wr_buf);
}
/* open the file -------------------------------------------------------- */
err = MPI_File_open(MPI_COMM_WORLD, filename,
MPI_MODE_CREATE | MPI_MODE_WRONLY, MPI_INFO_NULL, &fh);
if (err != MPI_SUCCESS) {
printf("Error: MPI_File_open() filename %s\n", filename);
MPI_Abort(MPI_COMM_WORLD, -1);
exit(1);
}
/* MPI nonblocking collective write */
for (i = 0; i < num_io; i++) {
offset = i * global_array_size;
/* set the file view */
MPI_File_set_view(fh, offset, MPI_BYTE, ftype, "native", MPI_INFO_NULL);
MPI_File_iwrite_all(fh, buf, ftype_size, MPI_CHAR, &request[i]);
}
MPI_Waitall(num_io, request, statuses);
MPI_File_close(&fh);
/* read and print file contents ----------------------------------------- */
if (rank == 0) {
char *ptr;
char *rd_buf = (char *) calloc(num_io * global_array_size, 1);
MPI_File_open(MPI_COMM_SELF, filename, MPI_MODE_RDONLY, MPI_INFO_NULL, &fh);
MPI_File_read(fh, rd_buf, num_io * global_array_size, MPI_CHAR, &status);
MPI_File_close(&fh);
#ifdef VERBOSE
printf("-------------------------------------------------------\n");
printf(" [");
for (i = 0; i < 2; i++) {
for (j = 0; j < XLEN; j++)
printf(" %d", j);
printf(" ");
}
printf("]\n\n");
ptr = rd_buf;
for (k = 0; k < num_io; k++) {
for (i = 0; i < 2 * YLEN; i++) {
printf("[%2d]", k * 2 * YLEN + i);
for (j = 0; j < 2 * XLEN; j++) {
if (j > 0 && j % XLEN == 0)
printf(" ");
if (*ptr != 0)
printf(" %c", *ptr);
else
printf(" ");
ptr++;
}
printf("\n");
}
printf("\n");
}
#endif
ptr = rd_buf;
for (i = 0; i < 2 * YLEN * num_io; i++) {
for (j = 0; j < 2 * XLEN; j++) {
if (*ptr != compare_buf[i][j]) {
fprintf(stderr, "expected %d got %d at [%d][%d]\n",
*ptr, compare_buf[i][j], i, j);
nr_errors++;
}
ptr++;
}
}
free(rd_buf);
if (nr_errors == 0)
fprintf(stdout, " No Errors\n");
else
fprintf(stderr, "Found %d errors\n", nr_errors);
}
free(blocklengths);
free(displacements);
free(buf);
free(request);
free(statuses);
MPI_Type_free(&ftype);
MPI_Finalize();
return 0;
}
void ADIOI_NFS_WriteStrided(ADIO_File fd, const void *buf, int count,
MPI_Datatype datatype, int file_ptr_type,
ADIO_Offset offset, ADIO_Status * status, int
*error_code)
{
/* offset is in units of etype relative to the filetype. */
ADIOI_Flatlist_node *flat_buf, *flat_file;
int i, j, k, err = -1, bwr_size, st_index = 0;
ADIO_Offset i_offset, sum, size_in_filetype;
ADIO_Offset num, size, n_etypes_in_filetype;
MPI_Count bufsize;
ADIO_Offset n_filetypes, etype_in_filetype;
ADIO_Offset abs_off_in_filetype = 0;
int req_len;
MPI_Count filetype_size, etype_size, buftype_size;
MPI_Aint filetype_extent, buftype_extent;
int buf_count, buftype_is_contig, filetype_is_contig;
ADIO_Offset userbuf_off;
ADIO_Offset off, req_off, disp, end_offset = 0, writebuf_off, start_off;
char *writebuf = NULL, *value;
int st_n_filetypes, writebuf_len, write_sz;
ADIO_Offset fwr_size = 0, new_fwr_size, st_fwr_size;
int new_bwr_size, err_flag = 0, info_flag, max_bufsize;
static char myname[] = "ADIOI_NFS_WRITESTRIDED";
ADIOI_Datatype_iscontig(datatype, &buftype_is_contig);
ADIOI_Datatype_iscontig(fd->filetype, &filetype_is_contig);
MPI_Type_size_x(fd->filetype, &filetype_size);
if (!filetype_size) {
#ifdef HAVE_STATUS_SET_BYTES
MPIR_Status_set_bytes(status, datatype, 0);
#endif
*error_code = MPI_SUCCESS;
return;
}
MPI_Type_extent(fd->filetype, &filetype_extent);
MPI_Type_size_x(datatype, &buftype_size);
MPI_Type_extent(datatype, &buftype_extent);
etype_size = fd->etype_size;
bufsize = buftype_size * count;
/* get max_bufsize from the info object. */
value = (char *) ADIOI_Malloc((MPI_MAX_INFO_VAL + 1) * sizeof(char));
ADIOI_Info_get(fd->info, "ind_wr_buffer_size", MPI_MAX_INFO_VAL, value, &info_flag);
max_bufsize = atoi(value);
ADIOI_Free(value);
if (!buftype_is_contig && filetype_is_contig) {
/* noncontiguous in memory, contiguous in file. */
flat_buf = ADIOI_Flatten_and_find(datatype);
off = (file_ptr_type == ADIO_INDIVIDUAL) ? fd->fp_ind : fd->disp + etype_size * offset;
start_off = off;
end_offset = off + bufsize - 1;
writebuf_off = off;
writebuf = (char *) ADIOI_Malloc(max_bufsize);
writebuf_len = (int) (MPL_MIN(max_bufsize, end_offset - writebuf_off + 1));
/* if atomicity is true, lock the region to be accessed */
if (fd->atomicity)
ADIOI_WRITE_LOCK(fd, start_off, SEEK_SET, end_offset - start_off + 1);
for (j = 0; j < count; j++)
for (i = 0; i < flat_buf->count; i++) {
userbuf_off = j * buftype_extent + flat_buf->indices[i];
req_off = off;
req_len = flat_buf->blocklens[i];
ADIOI_BUFFERED_WRITE_WITHOUT_READ off += flat_buf->blocklens[i];
}
/* write the buffer out finally */
#ifdef ADIOI_MPE_LOGGING
MPE_Log_event(ADIOI_MPE_lseek_a, 0, NULL);
#endif
lseek(fd->fd_sys, writebuf_off, SEEK_SET);
#ifdef ADIOI_MPE_LOGGING
MPE_Log_event(ADIOI_MPE_lseek_b, 0, NULL);
#endif
if (!(fd->atomicity))
ADIOI_WRITE_LOCK(fd, writebuf_off, SEEK_SET, writebuf_len);
#ifdef ADIOI_MPE_LOGGING
MPE_Log_event(ADIOI_MPE_write_a, 0, NULL);
#endif
err = write(fd->fd_sys, writebuf, writebuf_len);
#ifdef ADIOI_MPE_LOGGING
MPE_Log_event(ADIOI_MPE_write_b, 0, NULL);
#endif
if (!(fd->atomicity))
ADIOI_UNLOCK(fd, writebuf_off, SEEK_SET, writebuf_len);
if (err == -1)
err_flag = 1;
if (fd->atomicity)
ADIOI_UNLOCK(fd, start_off, SEEK_SET, end_offset - start_off + 1);
if (file_ptr_type == ADIO_INDIVIDUAL)
fd->fp_ind = off;
if (err_flag) {
*error_code = MPIO_Err_create_code(MPI_SUCCESS,
MPIR_ERR_RECOVERABLE, myname,
__LINE__, MPI_ERR_IO, "**io",
"**io %s", strerror(errno));
} else
*error_code = MPI_SUCCESS;
}
else { /* noncontiguous in file */
flat_file = ADIOI_Flatten_and_find(fd->filetype);
disp = fd->disp;
if (file_ptr_type == ADIO_INDIVIDUAL) {
/* Wei-keng reworked type processing to be a bit more efficient */
offset = fd->fp_ind - disp;
n_filetypes = (offset - flat_file->indices[0]) / filetype_extent;
offset -= (ADIO_Offset) n_filetypes *filetype_extent;
/* now offset is local to this extent */
/* find the block where offset is located, skip blocklens[i]==0 */
for (i = 0; i < flat_file->count; i++) {
ADIO_Offset dist;
if (flat_file->blocklens[i] == 0)
continue;
dist = flat_file->indices[i] + flat_file->blocklens[i] - offset;
/* fwr_size is from offset to the end of block i */
if (dist == 0) {
i++;
offset = flat_file->indices[i];
fwr_size = flat_file->blocklens[i];
break;
}
if (dist > 0) {
fwr_size = dist;
break;
}
}
st_index = i; /* starting index in flat_file->indices[] */
offset += disp + (ADIO_Offset) n_filetypes *filetype_extent;
} else {
n_etypes_in_filetype = filetype_size / etype_size;
n_filetypes = offset / n_etypes_in_filetype;
etype_in_filetype = offset % n_etypes_in_filetype;
size_in_filetype = etype_in_filetype * etype_size;
sum = 0;
for (i = 0; i < flat_file->count; i++) {
sum += flat_file->blocklens[i];
if (sum > size_in_filetype) {
st_index = i;
fwr_size = sum - size_in_filetype;
abs_off_in_filetype = flat_file->indices[i] +
size_in_filetype - (sum - flat_file->blocklens[i]);
break;
}
}
/* abs. offset in bytes in the file */
offset = disp + (ADIO_Offset) n_filetypes *filetype_extent + abs_off_in_filetype;
}
start_off = offset;
/* Wei-keng Liao:write request is within single flat_file contig block */
/* this could happen, for example, with subarray types that are
* actually fairly contiguous */
if (buftype_is_contig && bufsize <= fwr_size) {
/* though MPI api has an integer 'count' parameter, derived
* datatypes might describe more bytes than can fit into an integer.
* if we've made it this far, we can pass a count of original
* datatypes, instead of a count of bytes (which might overflow)
* Other WriteContig calls in this path are operating on data
* sieving buffer */
ADIO_WriteContig(fd, buf, count, datatype, ADIO_EXPLICIT_OFFSET,
offset, status, error_code);
if (file_ptr_type == ADIO_INDIVIDUAL) {
/* update MPI-IO file pointer to point to the first byte
* that can be accessed in the fileview. */
fd->fp_ind = offset + bufsize;
if (bufsize == fwr_size) {
do {
st_index++;
if (st_index == flat_file->count) {
st_index = 0;
n_filetypes++;
}
} while (flat_file->blocklens[st_index] == 0);
fd->fp_ind = disp + flat_file->indices[st_index]
+ (ADIO_Offset) n_filetypes *filetype_extent;
}
}
fd->fp_sys_posn = -1; /* set it to null. */
#ifdef HAVE_STATUS_SET_BYTES
MPIR_Status_set_bytes(status, datatype, bufsize);
#endif
goto fn_exit;
}
/* Calculate end_offset, the last byte-offset that will be accessed.
* e.g., if start_offset=0 and 100 bytes to be write, end_offset=99 */
st_fwr_size = fwr_size;
st_n_filetypes = n_filetypes;
i_offset = 0;
j = st_index;
off = offset;
fwr_size = MPL_MIN(st_fwr_size, bufsize);
while (i_offset < bufsize) {
i_offset += fwr_size;
end_offset = off + fwr_size - 1;
j = (j + 1) % flat_file->count;
n_filetypes += (j == 0) ? 1 : 0;
while (flat_file->blocklens[j] == 0) {
j = (j + 1) % flat_file->count;
n_filetypes += (j == 0) ? 1 : 0;
}
off = disp + flat_file->indices[j] + n_filetypes * (ADIO_Offset) filetype_extent;
fwr_size = MPL_MIN(flat_file->blocklens[j], bufsize - i_offset);
}
/* if atomicity is true, lock the region to be accessed */
if (fd->atomicity)
ADIOI_WRITE_LOCK(fd, start_off, SEEK_SET, end_offset - start_off + 1);
/* initial read for the read-modify-write */
writebuf_off = offset;
writebuf = (char *) ADIOI_Malloc(max_bufsize);
memset(writebuf, -1, max_bufsize);
writebuf_len = (int) (MPL_MIN(max_bufsize, end_offset - writebuf_off + 1));
if (!(fd->atomicity))
ADIOI_WRITE_LOCK(fd, writebuf_off, SEEK_SET, writebuf_len);
#ifdef ADIOI_MPE_LOGGING
MPE_Log_event(ADIOI_MPE_lseek_a, 0, NULL);
#endif
lseek(fd->fd_sys, writebuf_off, SEEK_SET);
#ifdef ADIOI_MPE_LOGGING
MPE_Log_event(ADIOI_MPE_lseek_b, 0, NULL);
#endif
#ifdef ADIOI_MPE_LOGGING
MPE_Log_event(ADIOI_MPE_read_a, 0, NULL);
#endif
err = read(fd->fd_sys, writebuf, writebuf_len);
#ifdef ADIOI_MPE_LOGGING
MPE_Log_event(ADIOI_MPE_read_b, 0, NULL);
#endif
if (err == -1) {
*error_code = MPIO_Err_create_code(MPI_SUCCESS,
MPIR_ERR_RECOVERABLE,
myname, __LINE__,
MPI_ERR_IO,
"ADIOI_NFS_WriteStrided: ROMIO tries to optimize this access by doing a read-modify-write, but is unable to read the file. Please give the file read permission and open it with MPI_MODE_RDWR.",
0);
goto fn_exit;
}
if (buftype_is_contig && !filetype_is_contig) {
/* contiguous in memory, noncontiguous in file. should be the most
common case. */
i_offset = 0;
j = st_index;
off = offset;
n_filetypes = st_n_filetypes;
fwr_size = MPL_MIN(st_fwr_size, bufsize);
while (i_offset < bufsize) {
if (fwr_size) {
/* TYPE_UB and TYPE_LB can result in
* fwr_size = 0. save system call in such cases */
/* lseek(fd->fd_sys, off, SEEK_SET);
* err = write(fd->fd_sys, ((char *) buf) + i, fwr_size); */
req_off = off;
req_len = fwr_size;
userbuf_off = i_offset;
ADIOI_BUFFERED_WRITE}
i_offset += fwr_size;
if (off + fwr_size < disp + flat_file->indices[j] +
flat_file->blocklens[j] + n_filetypes * (ADIO_Offset) filetype_extent)
off += fwr_size;
/* did not reach end of contiguous block in filetype.
* no more I/O needed. off is incremented by fwr_size. */
else {
j = (j + 1) % flat_file->count;
n_filetypes += (j == 0) ? 1 : 0;
while (flat_file->blocklens[j] == 0) {
j = (j + 1) % flat_file->count;
n_filetypes += (j == 0) ? 1 : 0;
}
off = disp + flat_file->indices[j] +
n_filetypes * (ADIO_Offset) filetype_extent;
fwr_size = MPL_MIN(flat_file->blocklens[j], bufsize - i_offset);
}
}
} else {
int MPIOI_File_write_all(MPI_File fh,
MPI_Offset offset,
int file_ptr_type,
const void *buf,
int count,
MPI_Datatype datatype,
char *myname,
MPI_Status *status)
{
int error_code;
MPI_Count datatype_size;
ADIO_File adio_fh;
void *e32buf=NULL;
const void *xbuf=NULL;
MPIU_THREAD_CS_ENTER(ALLFUNC,);
adio_fh = MPIO_File_resolve(fh);
/* --BEGIN ERROR HANDLING-- */
MPIO_CHECK_FILE_HANDLE(adio_fh, myname, error_code);
MPIO_CHECK_COUNT(adio_fh, count, myname, error_code);
MPIO_CHECK_DATATYPE(adio_fh, datatype, myname, error_code);
if (file_ptr_type == ADIO_EXPLICIT_OFFSET && offset < 0)
{
error_code = MPIO_Err_create_code(MPI_SUCCESS, MPIR_ERR_RECOVERABLE,
myname, __LINE__, MPI_ERR_ARG,
"**iobadoffset", 0);
error_code = MPIO_Err_return_file(adio_fh, error_code);
goto fn_exit;
}
/* --END ERROR HANDLING-- */
MPI_Type_size_x(datatype, &datatype_size);
/* --BEGIN ERROR HANDLING-- */
MPIO_CHECK_INTEGRAL_ETYPE(adio_fh, count, datatype_size, myname, error_code);
MPIO_CHECK_WRITABLE(adio_fh, myname, error_code);
MPIO_CHECK_NOT_SEQUENTIAL_MODE(adio_fh, myname, error_code);
MPIO_CHECK_COUNT_SIZE(adio_fh, count, datatype_size, myname, error_code);
/* --END ERROR HANDLING-- */
xbuf = buf;
if (adio_fh->is_external32) {
error_code = MPIU_external32_buffer_setup(buf, count, datatype, &e32buf);
if (error_code != MPI_SUCCESS)
goto fn_exit;
xbuf = e32buf;
}
ADIO_WriteStridedColl(adio_fh, xbuf, count, datatype, file_ptr_type,
offset, status, &error_code);
/* --BEGIN ERROR HANDLING-- */
if (error_code != MPI_SUCCESS)
error_code = MPIO_Err_return_file(adio_fh, error_code);
/* --END ERROR HANDLING-- */
fn_exit:
if (e32buf != NULL) ADIOI_Free(e32buf);
MPIU_THREAD_CS_EXIT(ALLFUNC,);
return error_code;
}
void ADIOI_LUSTRE_WriteStridedColl(ADIO_File fd, const void *buf, int count,
MPI_Datatype datatype,
int file_ptr_type, ADIO_Offset offset,
ADIO_Status *status, int *error_code)
{
/* Uses a generalized version of the extended two-phase method described
* in "An Extended Two-Phase Method for Accessing Sections of
* Out-of-Core Arrays", Rajeev Thakur and Alok Choudhary,
* Scientific Programming, (5)4:301--317, Winter 1996.
* http://www.mcs.anl.gov/home/thakur/ext2ph.ps
*/
ADIOI_Access *my_req;
/* array of nprocs access structures, one for each other process has
this process's request */
ADIOI_Access *others_req;
/* array of nprocs access structures, one for each other process
whose request is written by this process. */
int i, filetype_is_contig, nprocs, myrank, do_collect = 0;
int contig_access_count = 0, buftype_is_contig, interleave_count = 0;
int *count_my_req_per_proc, count_my_req_procs, count_others_req_procs;
ADIO_Offset orig_fp, start_offset, end_offset, off;
ADIO_Offset *offset_list = NULL, *st_offsets = NULL, *end_offsets = NULL;
ADIO_Offset *len_list = NULL;
int **buf_idx = NULL, *striping_info = NULL;
int old_error, tmp_error;
MPI_Comm_size(fd->comm, &nprocs);
MPI_Comm_rank(fd->comm, &myrank);
orig_fp = fd->fp_ind;
/* IO patten identification if cb_write isn't disabled */
if (fd->hints->cb_write != ADIOI_HINT_DISABLE) {
/* For this process's request, calculate the list of offsets and
lengths in the file and determine the start and end offsets. */
/* Note: end_offset points to the last byte-offset that will be accessed.
* e.g., if start_offset=0 and 100 bytes to be read, end_offset=99
*/
ADIOI_Calc_my_off_len(fd, count, datatype, file_ptr_type, offset,
&offset_list, &len_list, &start_offset,
&end_offset, &contig_access_count);
/* each process communicates its start and end offsets to other
* processes. The result is an array each of start and end offsets
* stored in order of process rank.
*/
st_offsets = (ADIO_Offset *) ADIOI_Malloc(nprocs * sizeof(ADIO_Offset));
end_offsets = (ADIO_Offset *) ADIOI_Malloc(nprocs * sizeof(ADIO_Offset));
MPI_Allgather(&start_offset, 1, ADIO_OFFSET, st_offsets, 1,
ADIO_OFFSET, fd->comm);
MPI_Allgather(&end_offset, 1, ADIO_OFFSET, end_offsets, 1,
ADIO_OFFSET, fd->comm);
/* are the accesses of different processes interleaved? */
for (i = 1; i < nprocs; i++)
if ((st_offsets[i] < end_offsets[i-1]) &&
(st_offsets[i] <= end_offsets[i]))
interleave_count++;
/* This is a rudimentary check for interleaving, but should suffice
for the moment. */
/* Two typical access patterns can benefit from collective write.
* 1) the processes are interleaved, and
* 2) the req size is small.
*/
if (interleave_count > 0) {
do_collect = 1;
} else {
do_collect = ADIOI_LUSTRE_Docollect(fd, contig_access_count,
len_list, nprocs);
}
}
ADIOI_Datatype_iscontig(datatype, &buftype_is_contig);
/* Decide if collective I/O should be done */
if ((!do_collect && fd->hints->cb_write == ADIOI_HINT_AUTO) ||
fd->hints->cb_write == ADIOI_HINT_DISABLE) {
/* use independent accesses */
if (fd->hints->cb_write != ADIOI_HINT_DISABLE) {
ADIOI_Free(offset_list);
ADIOI_Free(len_list);
ADIOI_Free(st_offsets);
ADIOI_Free(end_offsets);
}
fd->fp_ind = orig_fp;
ADIOI_Datatype_iscontig(fd->filetype, &filetype_is_contig);
if (buftype_is_contig && filetype_is_contig) {
if (file_ptr_type == ADIO_EXPLICIT_OFFSET) {
off = fd->disp + (ADIO_Offset)(fd->etype_size) * offset;
ADIO_WriteContig(fd, buf, count, datatype,
ADIO_EXPLICIT_OFFSET,
off, status, error_code);
} else
ADIO_WriteContig(fd, buf, count, datatype, ADIO_INDIVIDUAL,
0, status, error_code);
} else {
ADIO_WriteStrided(fd, buf, count, datatype, file_ptr_type,
offset, status, error_code);
}
return;
}
/* Get Lustre hints information */
ADIOI_LUSTRE_Get_striping_info(fd, &striping_info, 1);
/* calculate what portions of the access requests of this process are
* located in which process
*/
ADIOI_LUSTRE_Calc_my_req(fd, offset_list, len_list, contig_access_count,
striping_info, nprocs, &count_my_req_procs,
&count_my_req_per_proc, &my_req,
&buf_idx);
/* based on everyone's my_req, calculate what requests of other processes
* will be accessed by this process.
* count_others_req_procs = number of processes whose requests (including
* this process itself) will be accessed by this process
* count_others_req_per_proc[i] indicates how many separate contiguous
* requests of proc. i will be accessed by this process.
*/
ADIOI_Calc_others_req(fd, count_my_req_procs, count_my_req_per_proc,
my_req, nprocs, myrank, &count_others_req_procs,
&others_req);
ADIOI_Free(count_my_req_per_proc);
/* exchange data and write in sizes of no more than stripe_size. */
ADIOI_LUSTRE_Exch_and_write(fd, buf, datatype, nprocs, myrank,
others_req, my_req, offset_list, len_list,
contig_access_count, striping_info,
buf_idx, error_code);
/* If this collective write is followed by an independent write,
* it's possible to have those subsequent writes on other processes
* race ahead and sneak in before the read-modify-write completes.
* We carry out a collective communication at the end here so no one
* can start independent i/o before collective I/O completes.
*
* need to do some gymnastics with the error codes so that if something
* went wrong, all processes report error, but if a process has a more
* specific error code, we can still have that process report the
* additional information */
old_error = *error_code;
if (*error_code != MPI_SUCCESS)
*error_code = MPI_ERR_IO;
/* optimization: if only one process performing i/o, we can perform
* a less-expensive Bcast */
#ifdef ADIOI_MPE_LOGGING
MPE_Log_event(ADIOI_MPE_postwrite_a, 0, NULL);
#endif
if (fd->hints->cb_nodes == 1)
MPI_Bcast(error_code, 1, MPI_INT,
fd->hints->ranklist[0], fd->comm);
else {
tmp_error = *error_code;
MPI_Allreduce(&tmp_error, error_code, 1, MPI_INT,
MPI_MAX, fd->comm);
}
#ifdef ADIOI_MPE_LOGGING
MPE_Log_event(ADIOI_MPE_postwrite_b, 0, NULL);
#endif
if ((old_error != MPI_SUCCESS) && (old_error != MPI_ERR_IO))
*error_code = old_error;
if (!buftype_is_contig)
ADIOI_Delete_flattened(datatype);
/* free all memory allocated for collective I/O */
/* free others_req */
for (i = 0; i < nprocs; i++) {
if (others_req[i].count) {
ADIOI_Free(others_req[i].offsets);
ADIOI_Free(others_req[i].lens);
ADIOI_Free(others_req[i].mem_ptrs);
}
}
ADIOI_Free(others_req);
/* free my_req here */
for (i = 0; i < nprocs; i++) {
if (my_req[i].count) {
ADIOI_Free(my_req[i].offsets);
ADIOI_Free(my_req[i].lens);
}
}
ADIOI_Free(my_req);
for (i = 0; i < nprocs; i++) {
ADIOI_Free(buf_idx[i]);
}
ADIOI_Free(buf_idx);
ADIOI_Free(offset_list);
ADIOI_Free(len_list);
ADIOI_Free(st_offsets);
ADIOI_Free(end_offsets);
ADIOI_Free(striping_info);
#ifdef HAVE_STATUS_SET_BYTES
if (status) {
MPI_Count bufsize, size;
/* Don't set status if it isn't needed */
MPI_Type_size_x(datatype, &size);
bufsize = size * count;
MPIR_Status_set_bytes(status, datatype, bufsize);
}
/* This is a temporary way of filling in status. The right way is to
* keep track of how much data was actually written during collective I/O.
*/
#endif
fd->fp_sys_posn = -1; /* set it to null. */
}
void ADIOI_GEN_WriteStridedColl(ADIO_File fd, const void *buf, int count,
MPI_Datatype datatype, int file_ptr_type,
ADIO_Offset offset, ADIO_Status * status, int
*error_code)
{
/* Uses a generalized version of the extended two-phase method described
in "An Extended Two-Phase Method for Accessing Sections of
Out-of-Core Arrays", Rajeev Thakur and Alok Choudhary,
Scientific Programming, (5)4:301--317, Winter 1996.
http://www.mcs.anl.gov/home/thakur/ext2ph.ps */
ADIOI_Access *my_req;
/* array of nprocs access structures, one for each other process in
* whose file domain this process's request lies */
ADIOI_Access *others_req;
/* array of nprocs access structures, one for each other process
* whose request lies in this process's file domain. */
int i, filetype_is_contig, nprocs, nprocs_for_coll, myrank;
int contig_access_count = 0, interleave_count = 0, buftype_is_contig;
int *count_my_req_per_proc, count_my_req_procs, count_others_req_procs;
ADIO_Offset orig_fp, start_offset, end_offset, fd_size, min_st_offset, off;
ADIO_Offset *offset_list = NULL, *st_offsets = NULL, *fd_start = NULL,
*fd_end = NULL, *end_offsets = NULL;
MPI_Aint *buf_idx = NULL;
ADIO_Offset *len_list = NULL;
int old_error, tmp_error;
if (fd->hints->cb_pfr != ADIOI_HINT_DISABLE) {
/* Cast away const'ness as the below function is used for read
* and write */
ADIOI_IOStridedColl(fd, (char *) buf, count, ADIOI_WRITE, datatype,
file_ptr_type, offset, status, error_code);
return;
}
MPI_Comm_size(fd->comm, &nprocs);
MPI_Comm_rank(fd->comm, &myrank);
/* the number of processes that actually perform I/O, nprocs_for_coll,
* is stored in the hints off the ADIO_File structure
*/
nprocs_for_coll = fd->hints->cb_nodes;
orig_fp = fd->fp_ind;
/* only check for interleaving if cb_write isn't disabled */
if (fd->hints->cb_write != ADIOI_HINT_DISABLE) {
/* For this process's request, calculate the list of offsets and
* lengths in the file and determine the start and end offsets. */
/* Note: end_offset points to the last byte-offset that will be accessed.
* e.g., if start_offset=0 and 100 bytes to be read, end_offset=99 */
ADIOI_Calc_my_off_len(fd, count, datatype, file_ptr_type, offset,
&offset_list, &len_list, &start_offset,
&end_offset, &contig_access_count);
/* each process communicates its start and end offsets to other
* processes. The result is an array each of start and end offsets stored
* in order of process rank. */
st_offsets = (ADIO_Offset *) ADIOI_Malloc(nprocs * 2 * sizeof(ADIO_Offset));
end_offsets = st_offsets + nprocs;
MPI_Allgather(&start_offset, 1, ADIO_OFFSET, st_offsets, 1, ADIO_OFFSET, fd->comm);
MPI_Allgather(&end_offset, 1, ADIO_OFFSET, end_offsets, 1, ADIO_OFFSET, fd->comm);
/* are the accesses of different processes interleaved? */
for (i = 1; i < nprocs; i++)
if ((st_offsets[i] < end_offsets[i - 1]) && (st_offsets[i] <= end_offsets[i]))
interleave_count++;
/* This is a rudimentary check for interleaving, but should suffice
* for the moment. */
}
ADIOI_Datatype_iscontig(datatype, &buftype_is_contig);
if (fd->hints->cb_write == ADIOI_HINT_DISABLE ||
(!interleave_count && (fd->hints->cb_write == ADIOI_HINT_AUTO))) {
/* use independent accesses */
if (fd->hints->cb_write != ADIOI_HINT_DISABLE) {
ADIOI_Free(offset_list);
ADIOI_Free(st_offsets);
}
fd->fp_ind = orig_fp;
ADIOI_Datatype_iscontig(fd->filetype, &filetype_is_contig);
if (buftype_is_contig && filetype_is_contig) {
if (file_ptr_type == ADIO_EXPLICIT_OFFSET) {
off = fd->disp + (ADIO_Offset) (fd->etype_size) * offset;
ADIO_WriteContig(fd, buf, count, datatype,
ADIO_EXPLICIT_OFFSET, off, status, error_code);
} else
ADIO_WriteContig(fd, buf, count, datatype, ADIO_INDIVIDUAL, 0, status, error_code);
} else
ADIO_WriteStrided(fd, buf, count, datatype, file_ptr_type, offset, status, error_code);
return;
}
/* Divide the I/O workload among "nprocs_for_coll" processes. This is
done by (logically) dividing the file into file domains (FDs); each
process may directly access only its own file domain. */
ADIOI_Calc_file_domains(st_offsets, end_offsets, nprocs,
nprocs_for_coll, &min_st_offset,
&fd_start, &fd_end,
fd->hints->min_fdomain_size, &fd_size, fd->hints->striping_unit);
/* calculate what portions of the access requests of this process are
located in what file domains */
ADIOI_Calc_my_req(fd, offset_list, len_list, contig_access_count,
min_st_offset, fd_start, fd_end, fd_size,
nprocs, &count_my_req_procs, &count_my_req_per_proc, &my_req, &buf_idx);
/* based on everyone's my_req, calculate what requests of other
processes lie in this process's file domain.
count_others_req_procs = number of processes whose requests lie in
this process's file domain (including this process itself)
count_others_req_per_proc[i] indicates how many separate contiguous
requests of proc. i lie in this process's file domain. */
ADIOI_Calc_others_req(fd, count_my_req_procs,
count_my_req_per_proc, my_req,
nprocs, myrank, &count_others_req_procs, &others_req);
ADIOI_Free(count_my_req_per_proc);
ADIOI_Free(my_req[0].offsets);
ADIOI_Free(my_req);
/* exchange data and write in sizes of no more than coll_bufsize. */
/* Cast away const'ness for the below function */
ADIOI_Exch_and_write(fd, (char *) buf, datatype, nprocs, myrank,
others_req, offset_list,
len_list, contig_access_count, min_st_offset,
fd_size, fd_start, fd_end, buf_idx, error_code);
/* If this collective write is followed by an independent write,
* it's possible to have those subsequent writes on other processes
* race ahead and sneak in before the read-modify-write completes.
* We carry out a collective communication at the end here so no one
* can start independent i/o before collective I/O completes.
*
* need to do some gymnastics with the error codes so that if something
* went wrong, all processes report error, but if a process has a more
* specific error code, we can still have that process report the
* additional information */
old_error = *error_code;
if (*error_code != MPI_SUCCESS)
*error_code = MPI_ERR_IO;
/* optimization: if only one process performing i/o, we can perform
* a less-expensive Bcast */
#ifdef ADIOI_MPE_LOGGING
MPE_Log_event(ADIOI_MPE_postwrite_a, 0, NULL);
#endif
if (fd->hints->cb_nodes == 1)
MPI_Bcast(error_code, 1, MPI_INT, fd->hints->ranklist[0], fd->comm);
else {
tmp_error = *error_code;
MPI_Allreduce(&tmp_error, error_code, 1, MPI_INT, MPI_MAX, fd->comm);
}
#ifdef ADIOI_MPE_LOGGING
MPE_Log_event(ADIOI_MPE_postwrite_b, 0, NULL);
#endif
#ifdef AGGREGATION_PROFILE
MPE_Log_event(5012, 0, NULL);
#endif
if ((old_error != MPI_SUCCESS) && (old_error != MPI_ERR_IO))
*error_code = old_error;
/* free all memory allocated for collective I/O */
ADIOI_Free(others_req[0].offsets);
ADIOI_Free(others_req[0].mem_ptrs);
ADIOI_Free(others_req);
ADIOI_Free(buf_idx);
ADIOI_Free(offset_list);
ADIOI_Free(st_offsets);
ADIOI_Free(fd_start);
#ifdef HAVE_STATUS_SET_BYTES
if (status) {
MPI_Count bufsize, size;
/* Don't set status if it isn't needed */
MPI_Type_size_x(datatype, &size);
bufsize = size * count;
MPIR_Status_set_bytes(status, datatype, bufsize);
}
/* This is a temporary way of filling in status. The right way is to
keep track of how much data was actually written during collective I/O. */
#endif
fd->fp_sys_posn = -1; /* set it to null. */
#ifdef AGGREGATION_PROFILE
MPE_Log_event(5013, 0, NULL);
#endif
}
/*@
MPI_File_write_ordered - Collective write using shared file pointer
Input Parameters:
. fh - file handle (handle)
. buf - initial address of buffer (choice)
. count - number of elements in buffer (nonnegative integer)
. datatype - datatype of each buffer element (handle)
Output Parameters:
. status - status object (Status)
.N fortran
@*/
int MPI_File_write_ordered(MPI_File fh, ROMIO_CONST void *buf, int count,
MPI_Datatype datatype, MPI_Status *status)
{
int error_code, nprocs, myrank;
ADIO_Offset incr;
MPI_Count datatype_size;
int source, dest;
static char myname[] = "MPI_FILE_WRITE_ORDERED";
ADIO_Offset shared_fp;
ADIO_File adio_fh;
void *e32buf=NULL;
const void *xbuf;
adio_fh = MPIO_File_resolve(fh);
/* --BEGIN ERROR HANDLING-- */
MPIO_CHECK_FILE_HANDLE(adio_fh, myname, error_code);
MPIO_CHECK_COUNT(adio_fh, count, myname, error_code);
MPIO_CHECK_DATATYPE(adio_fh, datatype, myname, error_code);
/* --END ERROR HANDLING-- */
MPI_Type_size_x(datatype, &datatype_size);
/* --BEGIN ERROR HANDLING-- */
MPIO_CHECK_INTEGRAL_ETYPE(adio_fh, count, datatype_size, myname, error_code);
MPIO_CHECK_FS_SUPPORTS_SHARED(adio_fh, myname, error_code);
MPIO_CHECK_COUNT_SIZE(adio_fh, count, datatype_size, myname, error_code);
/* --END ERROR HANDLING-- */
ADIOI_TEST_DEFERRED(adio_fh, myname, &error_code);
MPI_Comm_size(adio_fh->comm, &nprocs);
MPI_Comm_rank(adio_fh->comm, &myrank);
incr = (count*datatype_size)/adio_fh->etype_size;
/* Use a message as a 'token' to order the operations */
source = myrank - 1;
dest = myrank + 1;
if (source < 0) source = MPI_PROC_NULL;
if (dest >= nprocs) dest = MPI_PROC_NULL;
MPI_Recv(NULL, 0, MPI_BYTE, source, 0, adio_fh->comm, MPI_STATUS_IGNORE);
ADIO_Get_shared_fp(adio_fh, incr, &shared_fp, &error_code);
/* --BEGIN ERROR HANDLING-- */
if (error_code != MPI_SUCCESS) {
error_code = MPIO_Err_create_code(MPI_SUCCESS, MPIR_ERR_FATAL,
myname, __LINE__, MPI_ERR_INTERN,
"**iosharedfailed", 0);
error_code = MPIO_Err_return_file(adio_fh, error_code);
goto fn_exit;
}
/* --END ERROR HANDLING-- */
MPI_Send(NULL, 0, MPI_BYTE, dest, 0, adio_fh->comm);
xbuf = buf;
if (adio_fh->is_external32) {
error_code = MPIU_external32_buffer_setup(buf, count, datatype, &e32buf);
if (error_code != MPI_SUCCESS)
goto fn_exit;
xbuf = e32buf;
}
ADIO_WriteStridedColl(adio_fh, xbuf, count, datatype, ADIO_EXPLICIT_OFFSET,
shared_fp, status, &error_code);
/* --BEGIN ERROR HANDLING-- */
if (error_code != MPI_SUCCESS)
error_code = MPIO_Err_return_file(adio_fh, error_code);
/* --END ERROR HANDLING-- */
fn_exit:
if (e32buf != NULL) ADIOI_Free(e32buf);
/* FIXME: Check for error code from WriteStridedColl? */
return error_code;
}
int MPIOI_File_iread_all(MPI_File fh,
MPI_Offset offset,
int file_ptr_type,
void *buf,
int count,
MPI_Datatype datatype,
char *myname,
MPI_Request *request)
{
int error_code;
MPI_Count datatype_size;
ADIO_File adio_fh;
void *xbuf=NULL, *e32_buf=NULL;
ROMIO_THREAD_CS_ENTER();
adio_fh = MPIO_File_resolve(fh);
/* --BEGIN ERROR HANDLING-- */
MPIO_CHECK_FILE_HANDLE(adio_fh, myname, error_code);
MPIO_CHECK_COUNT(adio_fh, count, myname, error_code);
MPIO_CHECK_DATATYPE(adio_fh, datatype, myname, error_code);
if (file_ptr_type == ADIO_EXPLICIT_OFFSET && offset < 0) {
error_code = MPIO_Err_create_code(MPI_SUCCESS, MPIR_ERR_RECOVERABLE,
myname, __LINE__, MPI_ERR_ARG,
"**iobadoffset", 0);
error_code = MPIO_Err_return_file(adio_fh, error_code);
goto fn_exit;
}
/* --END ERROR HANDLING-- */
MPI_Type_size_x(datatype, &datatype_size);
/* --BEGIN ERROR HANDLING-- */
MPIO_CHECK_INTEGRAL_ETYPE(adio_fh, count, datatype_size, myname, error_code);
MPIO_CHECK_READABLE(adio_fh, myname, error_code);
MPIO_CHECK_NOT_SEQUENTIAL_MODE(adio_fh, myname, error_code);
MPIO_CHECK_COUNT_SIZE(adio_fh, count, datatype_size, myname, error_code);
/* --END ERROR HANDLING-- */
xbuf = buf;
if (adio_fh->is_external32) {
MPI_Aint e32_size = 0;
error_code = MPIU_datatype_full_size(datatype, &e32_size);
if (error_code != MPI_SUCCESS)
goto fn_exit;
e32_buf = ADIOI_Malloc(e32_size*count);
xbuf = e32_buf;
}
ADIO_IreadStridedColl(adio_fh, xbuf, count, datatype, file_ptr_type,
offset, request, &error_code);
/* --BEGIN ERROR HANDLING-- */
if (error_code != MPI_SUCCESS)
error_code = MPIO_Err_return_file(adio_fh, error_code);
/* --END ERROR HANDLING-- */
if (e32_buf != NULL) {
error_code = MPIU_read_external32_conversion_fn(buf, datatype,
count, e32_buf);
ADIOI_Free(e32_buf);
}
fn_exit:
ROMIO_THREAD_CS_EXIT();
return error_code;
}
