static int builtin_posix_module_element_write_strided_nb (hio_element_t element, hio_request_t *request,
off_t offset, const void *ptr, size_t count,
size_t size, size_t stride) {
builtin_posix_module_dataset_t *posix_dataset = (builtin_posix_module_dataset_t *) hioi_element_dataset (element);
builtin_posix_module_t *posix_module = (builtin_posix_module_t *) posix_dataset->base.ds_module;
hio_context_t context = hioi_object_context (&element->e_object);
ssize_t bytes_written;
hio_request_t new_request;
pthread_mutex_lock (&posix_dataset->lock);
bytes_written = builtin_posix_module_element_write_strided_internal (posix_module, element, offset, ptr, count, size,
stride);
pthread_mutex_unlock (&posix_dataset->lock);
if (request) {
new_request = hioi_request_alloc (context);
if (NULL == new_request) {
return HIO_ERR_OUT_OF_RESOURCE;
}
*request = new_request;
new_request->req_transferred = bytes_written;
new_request->req_complete = true;
new_request->req_status = posix_dataset->base.ds_status;
}
return posix_dataset->base.ds_status;
}
static int builtin_posix_module_element_flush (hio_element_t element, hio_flush_mode_t mode) {
builtin_posix_module_dataset_t *posix_dataset =
(builtin_posix_module_dataset_t *) hioi_element_dataset (element);
if (!(posix_dataset->base.ds_flags & HIO_FLAG_WRITE)) {
return HIO_ERR_PERM;
}
if (HIO_FILE_MODE_OPTIMIZED == posix_dataset->base.ds_fmode) {
for (int i = 0 ; i < HIO_POSIX_MAX_OPEN_FILES ; ++i) {
if (posix_dataset->files[i].f_file.f_hndl) {
fflush (posix_dataset->files[i].f_file.f_hndl);
if (HIO_FLUSH_MODE_COMPLETE == mode) {
fsync (fileno (posix_dataset->files[i].f_file.f_hndl));
}
}
}
} else {
if (NULL != element->e_file.f_hndl) {
fflush (element->e_file.f_hndl);
if (HIO_FLUSH_MODE_COMPLETE == mode) {
fsync (fileno (element->e_file.f_hndl));
}
}
}
return HIO_SUCCESS;
}
static int builtin_posix_element_translate_strided (builtin_posix_module_t *posix_module, hio_element_t element,
uint64_t offset, size_t *size, hio_file_t **file_out) {
builtin_posix_module_dataset_t *posix_dataset = (builtin_posix_module_dataset_t *) hioi_element_dataset (element);
size_t block_id, block_base, block_bound, block_offset, file_id, file_block;
hio_context_t context = hioi_object_context (&element->e_object);
hio_file_t *file;
int32_t file_index;
char *path;
int rc;
block_id = offset / posix_dataset->ds_bs;
file_id = block_id % posix_dataset->ds_fcount;
file_block = block_id / posix_dataset->ds_fcount;
block_base = block_id * posix_dataset->ds_bs;
block_bound = block_base + posix_dataset->ds_bs;
block_offset = file_block * posix_dataset->ds_bs + offset - block_base;
hioi_log (context, HIO_VERBOSE_DEBUG_LOW, "builtin_posix_element_translate_strided: element: %s, offset: %"
PRIu64 ", file_id: %lu, file_block: %lu, block_offset: %lu, block_size: %" PRIu64,
hioi_object_identifier(element), offset, file_id, file_id, block_offset, posix_dataset->ds_bs);
if (offset + *size > block_bound) {
*size = block_bound - offset;
}
rc = asprintf (&path, "%s/data/%s_block.%08lu", posix_dataset->base_path, hioi_object_identifier(element),
(unsigned long) file_id);
if (0 > rc) {
return HIO_ERR_OUT_OF_RESOURCE;
}
/* use crc as a hash to pick a file index to use */
file_index = file_id % HIO_POSIX_MAX_OPEN_FILES;
file = posix_dataset->files + file_index;
if (file_id != file->f_bid || file->f_element != element) {
if (file->f_bid >= 0) {
POSIX_TRACE_CALL(posix_dataset, hioi_file_close (file), "file_close", file->f_bid, 0);
}
file->f_bid = -1;
file->f_element = element;
POSIX_TRACE_CALL(posix_dataset, rc = builtin_posix_open_file (posix_module, posix_dataset, path, file),
"file_open", file_id, 0);
if (HIO_SUCCESS != rc) {
return rc;
}
file->f_bid = file_id;
}
POSIX_TRACE_CALL(posix_dataset, hioi_file_seek (file, block_offset, SEEK_SET), "file_seek", file->f_bid, block_offset);
*file_out = file;
return HIO_SUCCESS;
}
static int builtin_posix_module_element_complete (hio_element_t element) {
hio_dataset_t dataset = hioi_element_dataset (element);
/* reads in this module are always blocking. need to update this code if
* that ever changes */
if (!(dataset->ds_flags & HIO_FLAG_READ)) {
return HIO_ERR_PERM;
}
return HIO_SUCCESS;
}
static int builtin_posix_element_translate (builtin_posix_module_t *posix_module, hio_element_t element, off_t offset,
size_t *size, hio_file_t **file_out, bool reading) {
builtin_posix_module_dataset_t *posix_dataset = (builtin_posix_module_dataset_t *) hioi_element_dataset (element);
if (HIO_FILE_MODE_BASIC == posix_dataset->base.ds_fmode) {
*file_out = &element->e_file;
if (offset != element->e_file.f_offset) {
fseek (element->e_file.f_hndl, offset, SEEK_SET);
element->e_file.f_offset = offset;
}
return HIO_SUCCESS;
}
if (reading && 0 == element->e_scount) {
return builtin_posix_element_translate_opt_old (posix_module, element, offset, size, file_out);
}
return builtin_posix_element_translate_opt (posix_module, element, offset, size, file_out, reading);
}
static int builtin_posix_element_translate (builtin_posix_module_t *posix_module, hio_element_t element,
uint64_t offset, size_t *size, hio_file_t **file_out, bool reading) {
builtin_posix_module_dataset_t *posix_dataset = (builtin_posix_module_dataset_t *) hioi_element_dataset (element);
int rc = HIO_SUCCESS;
switch (posix_dataset->ds_fmode) {
case HIO_FILE_MODE_BASIC:
*file_out = &element->e_file;
hioi_file_seek (&element->e_file, offset, SEEK_SET);
break;
case HIO_FILE_MODE_STRIDED:
rc = builtin_posix_element_translate_strided (posix_module, element, offset, size, file_out);
break;
case HIO_FILE_MODE_OPTIMIZED:
rc = builtin_posix_element_translate_opt (posix_module, element, offset, size, file_out, reading);
break;
}
return rc;
}
static int builtin_posix_module_element_read_strided_nb (hio_element_t element, hio_request_t *request, off_t offset,
void *ptr, size_t count, size_t size, size_t stride) {
builtin_posix_module_dataset_t *posix_dataset = (builtin_posix_module_dataset_t *) hioi_element_dataset (element);
builtin_posix_module_t *posix_module = (builtin_posix_module_t *) posix_dataset->base.ds_module;
hio_context_t context = hioi_object_context (&element->e_object);
ssize_t bytes_read;
hio_request_t new_request;
int rc = HIO_SUCCESS;
if (!(posix_dataset->base.ds_flags & HIO_FLAG_READ)) {
return HIO_ERR_PERM;
}
if (stride == 0) {
size *= count;
count = 1;
}
hioi_object_lock (&posix_dataset->base.ds_object);
bytes_read = builtin_posix_module_element_read_strided_internal (posix_module, element, offset, ptr, count, size, stride);
hioi_object_unlock (&posix_dataset->base.ds_object);
if (0 > bytes_read) {
rc = (int) bytes_read;
}
/* see if a request was requested */
if (request) {
new_request = hioi_request_alloc (context);
if (NULL == new_request) {
return HIO_ERR_OUT_OF_RESOURCE;
}
*request = new_request;
new_request->req_transferred = bytes_read;
new_request->req_complete = true;
}
return rc;
}
static int builtin_posix_module_element_flush (hio_element_t element, hio_flush_mode_t mode) {
builtin_posix_module_dataset_t *posix_dataset =
(builtin_posix_module_dataset_t *) hioi_element_dataset (element);
if (!(posix_dataset->base.ds_flags & HIO_FLAG_WRITE)) {
return HIO_ERR_PERM;
}
if (HIO_FLUSH_MODE_COMPLETE != mode) {
/* nothing to do at this time */
return HIO_SUCCESS;
}
if (HIO_FILE_MODE_BASIC != posix_dataset->ds_fmode) {
for (int i = 0 ; i < HIO_POSIX_MAX_OPEN_FILES ; ++i) {
hioi_file_flush (posix_dataset->files + i);
}
} else {
hioi_file_flush (&element->e_file);
}
return HIO_SUCCESS;
}
static ssize_t builtin_posix_module_element_read_strided_internal (builtin_posix_module_t *posix_module, hio_element_t element,
off_t offset, void *ptr, size_t count, size_t size,
size_t stride) {
builtin_posix_module_dataset_t *posix_dataset = (builtin_posix_module_dataset_t *) hioi_element_dataset (element);
size_t bytes_read = 0, ret;
hio_file_t *file;
uint64_t start, stop;
int rc;
if (!(count * size)) {
return 0;
}
errno = 0;
start = hioi_gettime ();
for (size_t i = 0 ; i < count ; ++i) {
size_t req = size, actual;
do {
actual = req;
rc = builtin_posix_element_translate (posix_module, element, offset, &actual,
&file, true);
if (HIO_SUCCESS != rc) {
break;
}
ret = fread (ptr, 1, actual, file->f_hndl);
if (ret > 0) {
bytes_read += ret;
file->f_offset += ret;
}
if (ret < actual) {
/* short read */
break;
}
req -= actual;
offset += actual;
ptr = (void *) ((intptr_t) ptr + actual);
} while (req);
if (req || HIO_SUCCESS != rc) {
break;
}
ptr = (void *) ((intptr_t) ptr + stride);
}
if (0 == bytes_read || HIO_SUCCESS != rc) {
if (0 == bytes_read) {
rc = hioi_err_errno (errno);
}
return rc;
}
stop = hioi_gettime ();
posix_dataset->base.ds_stat.s_rtime += stop - start;
posix_dataset->base.ds_stat.s_bread += bytes_read;
return bytes_read;
}
static ssize_t builtin_posix_module_element_write_strided_internal (builtin_posix_module_t *posix_module, hio_element_t element,
off_t offset, const void *ptr, size_t count, size_t size,
size_t stride) {
hio_dataset_t dataset = hioi_element_dataset (element);
size_t bytes_written = 0, ret;
hio_file_t *file;
uint64_t stop, start;
int rc;
assert (dataset->ds_flags & HIO_FLAG_WRITE);
if (!(count * size)) {
return 0;
}
if (0 == stride) {
size *= count;
count = 1;
}
start = hioi_gettime ();
errno = 0;
for (size_t i = 0 ; i < count ; ++i) {
size_t req = size, actual;
do {
actual = req;
rc = builtin_posix_element_translate (posix_module, element, offset, &actual,
&file, false);
assert (file);
if (HIO_SUCCESS != rc) {
break;
}
ret = fwrite (ptr, 1, actual, file->f_hndl);
if (ret > 0) {
bytes_written += ret;
file->f_offset += ret;
}
if (ret < actual) {
/* short write */
break;
}
req -= actual;
offset += actual;
ptr = (void *) ((intptr_t) ptr + actual);
} while (req);
if (HIO_SUCCESS != rc || req) {
break;
}
ptr = (void *) ((intptr_t) ptr + stride);
}
if (0 == bytes_written || HIO_SUCCESS != rc) {
if (0 == bytes_written) {
rc = hioi_err_errno (errno);
}
dataset->ds_status = rc;
return rc;
}
if (offset + bytes_written > element->e_size) {
element->e_size = offset + bytes_written;
}
stop = hioi_gettime ();
dataset->ds_stat.s_wtime += stop - start;
if (0 < bytes_written) {
dataset->ds_stat.s_bwritten += bytes_written;
}
hioi_log (hioi_object_context (&element->e_object), HIO_VERBOSE_DEBUG_LOW, "posix: finished write. bytes written: "
"%lu, time: %llu usec", bytes_written, stop - start);
return bytes_written;
}
static int builtin_posix_element_translate_opt (builtin_posix_module_t *posix_module, hio_element_t element, off_t offset,
size_t *size, hio_file_t **file_out, bool reading) {
builtin_posix_module_dataset_t *posix_dataset = (builtin_posix_module_dataset_t *) hioi_element_dataset (element);
hio_context_t context = hioi_object_context (&element->e_object);
builtin_posix_file_t *file;
uint64_t file_offset;
int file_index;
char *path;
int rc;
hioi_log (context, HIO_VERBOSE_DEBUG_MED, "translating element %s offset %ld size %lu",
hioi_object_identifier (&element->e_object), offset, *size);
rc = hioi_element_translate_offset (element, offset, &file_index, &file_offset, size);
if (HIO_SUCCESS != rc) {
if (reading) {
hioi_log (context, HIO_VERBOSE_DEBUG_MED, "offset not found");
/* not found */
return rc;
}
if (hioi_context_using_mpi (context)) {
rc = asprintf (&path, "%s/data.%x", posix_dataset->base_path, posix_dataset->base.ds_shared_control->s_master);
if (0 > rc) {
return HIO_ERR_OUT_OF_RESOURCE;
}
} else {
rc = asprintf (&path, "%s/data", posix_dataset->base_path);
if (0 > rc) {
return HIO_ERR_OUT_OF_RESOURCE;
}
}
file_offset = builtin_posix_reserve (posix_dataset, size);
file_index = hioi_dataset_add_file (&posix_dataset->base, strrchr (path, '/') + 1);
hioi_element_add_segment (element, file_index, file_offset, offset, *size);
} else {
hioi_log (context, HIO_VERBOSE_DEBUG_MED, "offset found in file @ index %d, offset %lu, size %lu", file_index,
file_offset, *size);
rc = asprintf (&path, "%s/%s", posix_dataset->base_path, posix_dataset->base.ds_flist[file_index].f_name);
if (0 > rc) {
return HIO_ERR_OUT_OF_RESOURCE;
}
}
/* use crc as a hash to pick a file index to use */
int internal_index = file_index % HIO_POSIX_MAX_OPEN_FILES;
file = posix_dataset->files + internal_index;
if (internal_index != file->f_bid) {
if (NULL != file->f_file.f_hndl) {
fclose (file->f_file.f_hndl);
file->f_file.f_hndl = NULL;
file->f_bid = -1;
}
rc = builtin_posix_open_file (posix_module, posix_dataset, path, &file->f_file);
if (HIO_SUCCESS != rc) {
free (path);
return rc;
}
file->f_bid = file_index;
}
free (path);
if (file_offset != file->f_file.f_offset) {
fseek (file->f_file.f_hndl, file_offset, SEEK_SET);
file->f_file.f_offset = file_offset;
}
*file_out = &file->f_file;
return HIO_SUCCESS;
}
static int builtin_posix_element_translate_opt_old (builtin_posix_module_t *posix_module, hio_element_t element, off_t offset,
size_t *size, hio_file_t **file_out) {
builtin_posix_module_dataset_t *posix_dataset = (builtin_posix_module_dataset_t *) hioi_element_dataset (element);
hio_context_t context = hioi_object_context (&element->e_object);
size_t block_id, block_base, block_bound, block_offset;
builtin_posix_file_t *file;
int32_t file_index;
char *path;
int rc, foo;
block_id = offset / posix_dataset->base.ds_bs;
block_base = block_id * posix_dataset->base.ds_bs;
block_bound = block_base + posix_dataset->base.ds_bs;
block_offset = offset - block_base;
hioi_log (context, HIO_VERBOSE_DEBUG_LOW, "builtin_posix_element_translate: element: %s, offset: %lu, block_id: %lu, "
"block_offset: %lu, block_size: %lu", hioi_object_identifier(element), (unsigned long) offset,
block_id, block_offset, posix_dataset->base.ds_bs);
if (offset + *size > block_bound) {
*size = block_bound - offset;
}
rc = asprintf (&path, "%s_block.%lu", hioi_object_identifier(element), (unsigned long) block_id);
if (0 > rc) {
return HIO_ERR_OUT_OF_RESOURCE;
}
if (HIO_FLAG_WRITE & posix_dataset->base.ds_flags) {
foo = hioi_dataset_add_file (&posix_dataset->base, path);
}
char *tmp = path;
rc = asprintf (&path, "%s/%s", posix_dataset->base_path, tmp);
free (tmp);
if (0 > rc) {
return HIO_ERR_OUT_OF_RESOURCE;
}
/* use crc as a hash to pick a file index to use */
file_index = hioi_crc32 ((uint8_t *) path, strlen (path)) % HIO_POSIX_MAX_OPEN_FILES;
file = posix_dataset->files + file_index;
if (block_id != file->f_bid || file->f_element != element) {
if (file->f_file.f_hndl != NULL) {
fclose (file->f_file.f_hndl);
file->f_file.f_hndl = NULL;
file->f_bid = -1;
}
file->f_element = element;
rc = builtin_posix_open_file (posix_module, posix_dataset, path, &file->f_file);
if (HIO_SUCCESS != rc) {
return rc;
}
file->f_bid = block_id;
}
if (block_offset != file->f_file.f_offset) {
fseek (file->f_file.f_hndl, block_offset, SEEK_SET);
file->f_file.f_offset = block_offset;
}
if (HIO_FLAG_WRITE & posix_dataset->base.ds_flags) {
hioi_element_add_segment (element, foo, block_offset, offset, *size);
}
*file_out = &file->f_file;
return HIO_SUCCESS;
}
static ssize_t builtin_posix_module_element_read_strided_internal (builtin_posix_module_t *posix_module, hio_element_t element,
uint64_t offset, void *ptr, size_t count, size_t size,
size_t stride) {
builtin_posix_module_dataset_t *posix_dataset = (builtin_posix_module_dataset_t *) hioi_element_dataset (element);
size_t bytes_read = 0, ret;
hio_file_t *file;
uint64_t start, stop;
int rc;
if (0 == count || 0 == size) {
return 0;
}
errno = 0;
start = hioi_gettime ();
for (size_t i = 0 ; i < count ; ++i) {
size_t req = size, actual;
do {
actual = req;
/* find out where the data lives */
POSIX_TRACE_CALL(posix_dataset, rc = builtin_posix_element_translate (posix_module, element, offset, &actual,
&file, true),
"element_translate", offset, req);
if (HIO_SUCCESS != rc) {
break;
}
POSIX_TRACE_CALL(posix_dataset, ret = hioi_file_read (file, ptr, actual), "file_read", offset, actual);
if (ret > 0) {
bytes_read += ret;
}
if (ret < actual) {
/* short read */
break;
}
req -= actual;
offset += actual;
ptr = (void *) ((intptr_t) ptr + actual);
} while (req);
if (req || HIO_SUCCESS != rc) {
break;
}
ptr = (void *) ((intptr_t) ptr + stride);
}
if (0 == bytes_read || HIO_SUCCESS != rc) {
if (0 == bytes_read && HIO_SUCCESS == rc) {
rc = hioi_err_errno (errno);
}
return rc;
}
stop = hioi_gettime ();
posix_dataset->base.ds_stat.s_rtime += stop - start;
posix_dataset->base.ds_stat.s_bread += bytes_read;
return bytes_read;
}
static int builtin_posix_element_translate_opt (builtin_posix_module_t *posix_module, hio_element_t element,
uint64_t offset, size_t *size, hio_file_t **file_out,
bool reading) {
builtin_posix_module_dataset_t *posix_dataset = (builtin_posix_module_dataset_t *) hioi_element_dataset (element);
hio_context_t context = hioi_object_context (&element->e_object);
hio_file_t *file;
uint64_t file_offset;
int file_index = 0;
char *path;
int rc;
hioi_log (context, HIO_VERBOSE_DEBUG_MED, "translating element %s offset %" PRIu64 " size %lu",
hioi_object_identifier (&element->e_object), offset, *size);
POSIX_TRACE_CALL(posix_dataset, rc = hioi_element_translate_offset (element, offset, &file_index, &file_offset, size),
"translate_offset", offset, *size);
#if HIO_MPI_HAVE(3)
if (HIO_SUCCESS != rc && reading) {
POSIX_TRACE_CALL(posix_dataset, rc = hioi_dataset_map_translate_offset (element, offset, &file_index, &file_offset, size),
"map_translate_offset", offset, *size);
}
#endif
if (HIO_SUCCESS != rc) {
if (reading) {
hioi_log (context, HIO_VERBOSE_DEBUG_MED, "offset %" PRIu64 " not found", offset);
/* not found */
return rc;
}
file_offset = builtin_posix_reserve (posix_dataset, size);
if (hioi_context_using_mpi (context)) {
file_index = posix_dataset->base.ds_shared_control->s_master;
} else {
file_index = 0;
}
rc = asprintf (&path, "%s/data/data.%x", posix_dataset->base_path, file_index);
if (0 > rc) {
return HIO_ERR_OUT_OF_RESOURCE;
}
hioi_element_add_segment (element, file_index, file_offset, offset, *size);
} else {
hioi_log (context, HIO_VERBOSE_DEBUG_MED, "offset found in file @ rank %d, offset %" PRIu64
", size %lu", file_index, file_offset, *size);
rc = asprintf (&path, "%s/data/data.%x", posix_dataset->base_path, file_index);
if (0 > rc) {
return HIO_ERR_OUT_OF_RESOURCE;
}
if (access (path, R_OK)) {
free (path);
rc = asprintf (&path, "%s/data.%x", posix_dataset->base_path, file_index);
if (0 > rc) {
return HIO_ERR_OUT_OF_RESOURCE;
}
}
}
/* use crc as a hash to pick a file index to use */
int internal_index = file_index % HIO_POSIX_MAX_OPEN_FILES;
file = posix_dataset->files + internal_index;
if (file_index != file->f_bid) {
if (file->f_bid >= 0) {
POSIX_TRACE_CALL(posix_dataset, hioi_file_close (file), "file_close", file->f_bid, 0);
}
file->f_bid = -1;
POSIX_TRACE_CALL(posix_dataset, rc = builtin_posix_open_file (posix_module, posix_dataset, path, file),
"file_open", file_index, 0);
if (HIO_SUCCESS != rc) {
free (path);
return rc;
}
file->f_bid = file_index;
}
free (path);
POSIX_TRACE_CALL(posix_dataset, hioi_file_seek (file, file_offset, SEEK_SET), "file_seek", file->f_bid, file_offset);
*file_out = file;
return HIO_SUCCESS;
}