static void* getBuffer(JNIEnv *env, ompi_java_buffer_t **item, int size)
{
if(size > ompi_mpi_java_eager)
{
*item = NULL;
return malloc(size);
}
else
{
int rc;
opal_free_list_item_t *freeListItem;
OPAL_FREE_LIST_GET(&ompi_java_buffers, freeListItem, rc);
ompi_java_exceptionCheck(env,
rc==OPAL_SUCCESS ? OMPI_SUCCESS : OMPI_ERROR);
*item = (ompi_java_buffer_t*)freeListItem;
return (*item)->buffer;
}
}
void* shmem_java_getBuffer(shmem_java_buffer_t **item, JNIEnv *env, int size)
{
if(size > oshmem_shmem_java_eager)
{
*item = NULL;
return malloc(size);
}
else
{
opal_free_list_item_t *freeListItem;
OPAL_FREE_LIST_GET(&shmem_java_buffers, freeListItem);
if (NULL == item) {
(*env)->ThrowNew(env, shmem_java.ExceptionClass,
"Error in OPAL_FREE_LIST_GET");
}
*item = (shmem_java_buffer_t*)freeListItem;
return (*item)->buffer;
}
}
int
ompi_osc_pt2pt_control_send(ompi_osc_pt2pt_module_t *module,
ompi_proc_t *proc,
uint8_t type, int32_t value0, int32_t value1)
{
int ret = OMPI_SUCCESS;
opal_free_list_item_t *item;
ompi_osc_pt2pt_buffer_t *buffer = NULL;
ompi_osc_pt2pt_control_header_t *header = NULL;
int rank = -1, i;
/* find the rank */
for (i = 0 ; i < module->p2p_comm->c_remote_group->grp_proc_count ; ++i) {
if (proc == module->p2p_comm->c_remote_group->grp_proc_pointers[i]) {
rank = i;
}
}
/* Get a buffer */
OPAL_FREE_LIST_GET(&mca_osc_pt2pt_component.p2p_c_buffers,
item, ret);
if (NULL == item) {
ret = OMPI_ERR_TEMP_OUT_OF_RESOURCE;
goto cleanup;
}
buffer = (ompi_osc_pt2pt_buffer_t*) item;
/* verify at least enough space for header */
if (mca_osc_pt2pt_component.p2p_c_eager_size < sizeof(ompi_osc_pt2pt_control_header_t)) {
ret = OMPI_ERR_OUT_OF_RESOURCE;
goto cleanup;
}
/* setup buffer */
buffer->cbfunc = ompi_osc_pt2pt_control_send_cb;
buffer->cbdata = NULL;
buffer->len = sizeof(ompi_osc_pt2pt_control_header_t);
/* pack header */
header = (ompi_osc_pt2pt_control_header_t*) buffer->payload;
header->hdr_base.hdr_type = type;
header->hdr_base.hdr_flags = 0;
header->hdr_value[0] = value0;
header->hdr_value[1] = value1;
header->hdr_windx = module->p2p_comm->c_contextid;
#ifdef WORDS_BIGENDIAN
header->hdr_base.hdr_flags |= OMPI_OSC_PT2PT_HDR_FLAG_NBO;
#elif OMPI_ENABLE_HETEROGENEOUS_SUPPORT
if (proc->proc_arch & OMPI_ARCH_ISBIGENDIAN) {
header->hdr_base.hdr_flags |= OMPI_OSC_PT2PT_HDR_FLAG_NBO;
OMPI_OSC_PT2PT_CONTROL_HDR_HTON(*header);
}
#endif
/* send fragment */
ret = MCA_PML_CALL(isend(buffer->payload,
buffer->len,
MPI_BYTE,
rank,
-200,
MCA_PML_BASE_SEND_STANDARD,
module->p2p_comm,
&buffer->request));
opal_list_append(&module->p2p_pending_control_sends,
&buffer->super.super);
goto done;
cleanup:
if (item != NULL) {
OPAL_FREE_LIST_RETURN(&mca_osc_pt2pt_component.p2p_c_buffers,
item);
}
done:
return ret;
}
int
ompi_osc_pt2pt_replyreq_send(ompi_osc_pt2pt_module_t *module,
ompi_osc_pt2pt_replyreq_t *replyreq)
{
int ret = OMPI_SUCCESS;
opal_free_list_item_t *item;
ompi_osc_pt2pt_buffer_t *buffer = NULL;
ompi_osc_pt2pt_reply_header_t *header = NULL;
size_t written_data = 0;
/* Get a buffer */
OPAL_FREE_LIST_GET(&mca_osc_pt2pt_component.p2p_c_buffers,
item, ret);
if (NULL == item) {
ret = OMPI_ERR_TEMP_OUT_OF_RESOURCE;
goto cleanup;
}
buffer = (ompi_osc_pt2pt_buffer_t*) item;
/* verify at least enough space for header */
if (mca_osc_pt2pt_component.p2p_c_eager_size < sizeof(ompi_osc_pt2pt_reply_header_t)) {
ret = OMPI_ERR_OUT_OF_RESOURCE;
goto cleanup;
}
/* setup buffer */
buffer->cbfunc = ompi_osc_pt2pt_replyreq_send_cb;
buffer->cbdata = (void*) replyreq;
/* pack header */
header = (ompi_osc_pt2pt_reply_header_t*) buffer->payload;
written_data += sizeof(ompi_osc_pt2pt_reply_header_t);
header->hdr_base.hdr_type = OMPI_OSC_PT2PT_HDR_REPLY;
header->hdr_base.hdr_flags = 0;
header->hdr_origin_sendreq = replyreq->rep_origin_sendreq;
header->hdr_target_tag = 0;
/* if sending data fits, pack payload */
if (mca_osc_pt2pt_component.p2p_c_eager_size >=
written_data + replyreq->rep_target_bytes_packed) {
struct iovec iov;
uint32_t iov_count = 1;
size_t max_data = replyreq->rep_target_bytes_packed;
iov.iov_len = max_data;
iov.iov_base = (IOVBASE_TYPE*)((unsigned char*) buffer->payload + written_data);
ret = ompi_convertor_pack(&replyreq->rep_target_convertor, &iov, &iov_count,
&max_data );
if (ret < 0) {
ret = OMPI_ERR_FATAL;
goto cleanup;
}
assert(max_data == replyreq->rep_target_bytes_packed);
written_data += max_data;
header->hdr_msg_length = replyreq->rep_target_bytes_packed;
} else {
header->hdr_msg_length = 0;
header->hdr_target_tag = create_send_tag(module);
}
buffer->len = written_data;
#ifdef WORDS_BIGENDIAN
header->hdr_base.hdr_flags |= OMPI_OSC_PT2PT_HDR_FLAG_NBO;
#elif OMPI_ENABLE_HETEROGENEOUS_SUPPORT
if (replyreq->rep_origin_proc->proc_arch & OMPI_ARCH_ISBIGENDIAN) {
header->hdr_base.hdr_flags |= OMPI_OSC_PT2PT_HDR_FLAG_NBO;
OMPI_OSC_PT2PT_REPLY_HDR_HTON(*header);
}
#endif
/* send fragment */
ret = MCA_PML_CALL(isend(buffer->payload,
buffer->len,
MPI_BYTE,
replyreq->rep_origin_rank,
-200,
MCA_PML_BASE_SEND_STANDARD,
module->p2p_comm,
&buffer->request));
opal_list_append(&module->p2p_pending_control_sends,
&buffer->super.super);
goto done;
cleanup:
if (item != NULL) {
OPAL_FREE_LIST_RETURN(&mca_osc_pt2pt_component.p2p_c_buffers,
item);
}
done:
return ret;
}
/* create the initial fragment, pack header, datatype, and payload (if
size fits) and send */
int
ompi_osc_pt2pt_sendreq_send(ompi_osc_pt2pt_module_t *module,
ompi_osc_pt2pt_sendreq_t *sendreq)
{
int ret = OMPI_SUCCESS;
opal_free_list_item_t *item;
ompi_osc_pt2pt_send_header_t *header = NULL;
ompi_osc_pt2pt_buffer_t *buffer = NULL;
size_t written_data = 0;
size_t needed_len = sizeof(ompi_osc_pt2pt_send_header_t);
const void *packed_ddt;
size_t packed_ddt_len = ompi_ddt_pack_description_length(sendreq->req_target_datatype);
/* we always need to send the ddt */
needed_len += packed_ddt_len;
if (OMPI_OSC_PT2PT_GET != sendreq->req_type) {
needed_len += sendreq->req_origin_bytes_packed;
}
/* Get a buffer */
OPAL_FREE_LIST_GET(&mca_osc_pt2pt_component.p2p_c_buffers,
item, ret);
if (NULL == item) {
ret = OMPI_ERR_TEMP_OUT_OF_RESOURCE;
goto cleanup;
}
buffer = (ompi_osc_pt2pt_buffer_t*) item;
/* verify at least enough space for header */
if (mca_osc_pt2pt_component.p2p_c_eager_size < sizeof(ompi_osc_pt2pt_send_header_t)) {
ret = OMPI_ERR_OUT_OF_RESOURCE;
goto cleanup;
}
/* setup buffer */
buffer->cbfunc = ompi_osc_pt2pt_sendreq_send_cb;
buffer->cbdata = (void*) sendreq;
/* pack header */
header = (ompi_osc_pt2pt_send_header_t*) buffer->payload;
written_data += sizeof(ompi_osc_pt2pt_send_header_t);
header->hdr_base.hdr_flags = 0;
header->hdr_windx = sendreq->req_module->p2p_comm->c_contextid;
header->hdr_origin = sendreq->req_module->p2p_comm->c_my_rank;
header->hdr_origin_sendreq.pval = (void*) sendreq;
header->hdr_origin_tag = 0;
header->hdr_target_disp = sendreq->req_target_disp;
header->hdr_target_count = sendreq->req_target_count;
switch (sendreq->req_type) {
case OMPI_OSC_PT2PT_PUT:
header->hdr_base.hdr_type = OMPI_OSC_PT2PT_HDR_PUT;
#if OMPI_ENABLE_MEM_DEBUG
header->hdr_target_op = 0;
#endif
break;
case OMPI_OSC_PT2PT_ACC:
header->hdr_base.hdr_type = OMPI_OSC_PT2PT_HDR_ACC;
header->hdr_target_op = sendreq->req_op_id;
break;
case OMPI_OSC_PT2PT_GET:
header->hdr_base.hdr_type = OMPI_OSC_PT2PT_HDR_GET;
#if OMPI_ENABLE_MEM_DEBUG
header->hdr_target_op = 0;
#endif
break;
}
/* Set datatype id and / or pack datatype */
ret = ompi_ddt_get_pack_description(sendreq->req_target_datatype, &packed_ddt);
if (OMPI_SUCCESS != ret) goto cleanup;
memcpy((unsigned char*) buffer->payload + written_data,
packed_ddt, packed_ddt_len);
written_data += packed_ddt_len;
if (OMPI_OSC_PT2PT_GET != sendreq->req_type) {
/* if sending data and it fits, pack payload */
if (mca_osc_pt2pt_component.p2p_c_eager_size >=
written_data + sendreq->req_origin_bytes_packed) {
struct iovec iov;
uint32_t iov_count = 1;
size_t max_data = sendreq->req_origin_bytes_packed;
iov.iov_len = max_data;
iov.iov_base = (IOVBASE_TYPE*)((unsigned char*) buffer->payload + written_data);
ret = ompi_convertor_pack(&sendreq->req_origin_convertor, &iov, &iov_count,
&max_data );
if (ret < 0) {
ret = OMPI_ERR_FATAL;
goto cleanup;
}
assert(max_data == sendreq->req_origin_bytes_packed);
written_data += max_data;
header->hdr_msg_length = sendreq->req_origin_bytes_packed;
} else {
header->hdr_msg_length = 0;
header->hdr_origin_tag = create_send_tag(module);
}
} else {
header->hdr_msg_length = 0;
}
buffer->len = written_data;
#ifdef WORDS_BIGENDIAN
header->hdr_base.hdr_flags |= OMPI_OSC_PT2PT_HDR_FLAG_NBO;
#elif OMPI_ENABLE_HETEROGENEOUS_SUPPORT
if (sendreq->req_target_proc->proc_arch & OMPI_ARCH_ISBIGENDIAN) {
header->hdr_base.hdr_flags |= OMPI_OSC_PT2PT_HDR_FLAG_NBO;
OMPI_OSC_PT2PT_SEND_HDR_HTON(*header);
}
#endif
/* send fragment */
opal_output_verbose(51, ompi_osc_base_output,
"%d sending sendreq to %d",
sendreq->req_module->p2p_comm->c_my_rank,
sendreq->req_target_rank);
ret = MCA_PML_CALL(isend(buffer->payload,
buffer->len,
MPI_BYTE,
sendreq->req_target_rank,
-200,
MCA_PML_BASE_SEND_STANDARD,
module->p2p_comm,
&buffer->request));
opal_list_append(&module->p2p_pending_control_sends,
&buffer->super.super);
goto done;
cleanup:
if (item != NULL) {
OPAL_FREE_LIST_RETURN(&mca_osc_pt2pt_component.p2p_c_buffers,
item);
}
done:
return ret;
}
static inline int
ompi_mtl_portals4_send_start(struct mca_mtl_base_module_t* mtl,
struct ompi_communicator_t* comm,
int dest,
int tag,
struct opal_convertor_t *convertor,
mca_pml_base_send_mode_t mode,
ompi_mtl_portals4_isend_request_t* ptl_request)
{
int ret= OMPI_SUCCESS;
void *start;
size_t length;
bool free_after;
ompi_proc_t *ompi_proc = ompi_comm_peer_lookup(comm, dest);
ptl_process_t *proc = (ptl_process_t*) ompi_proc->proc_endpoints[OMPI_PROC_ENDPOINT_TAG_PORTALS4];
#if OMPI_MTL_PORTALS4_FLOW_CONTROL
opal_free_list_item_t *item;
ompi_mtl_portals4_pending_request_t *pending;
#endif
ret = ompi_mtl_datatype_pack(convertor, &start, &length, &free_after);
if (OMPI_SUCCESS != ret) return ret;
ptl_request->opcount = OPAL_THREAD_ADD64((int64_t*)&ompi_mtl_portals4.opcount, 1);
ptl_request->buffer_ptr = (free_after) ? start : NULL;
ptl_request->event_count = 0;
OPAL_OUTPUT_VERBOSE((50, ompi_mtl_base_framework.framework_output,
"Send %lu to %x,%x of length %d\n",
ptl_request->opcount,
proc->phys.nid,
proc->phys.pid,
(int)length));
#if OMPI_MTL_PORTALS4_FLOW_CONTROL
OPAL_FREE_LIST_GET(&ompi_mtl_portals4.flowctl.pending_fl, item, ret);
if (NULL == item) return OMPI_ERR_OUT_OF_RESOURCE;
pending = (ompi_mtl_portals4_pending_request_t*) item;
ptl_request->pending = pending;
pending->mode = mode;
pending->start = start;
pending->length = length;
pending->contextid = comm->c_contextid;
pending->tag = tag;
pending->my_rank = comm->c_my_rank;
pending->fc_notified = 0;
pending->proc = proc;
pending->ptl_request = ptl_request;
if (OPAL_UNLIKELY(OPAL_THREAD_ADD32(&ompi_mtl_portals4.flowctl.send_slots, -1) < 0)) {
OPAL_THREAD_ADD32(&ompi_mtl_portals4.flowctl.send_slots, 1);
opal_list_append(&ompi_mtl_portals4.flowctl.pending_sends,
&pending->super.super);
return OMPI_SUCCESS;
}
if (OPAL_UNLIKELY(0 != opal_list_get_size(&ompi_mtl_portals4.flowctl.pending_sends))) {
OPAL_THREAD_ADD32(&ompi_mtl_portals4.flowctl.send_slots, 1);
opal_list_append(&ompi_mtl_portals4.flowctl.pending_sends,
&pending->super.super);
ompi_mtl_portals4_pending_list_progress();
return OMPI_SUCCESS;
}
if (OPAL_UNLIKELY(ompi_mtl_portals4.flowctl.flowctl_active)) {
OPAL_THREAD_ADD32(&ompi_mtl_portals4.flowctl.send_slots, 1);
opal_list_append(&ompi_mtl_portals4.flowctl.pending_sends,
&pending->super.super);
return OMPI_SUCCESS;
}
#endif
if (length <= ompi_mtl_portals4.eager_limit) {
ret = ompi_mtl_portals4_short_isend(mode,
start,
length,
comm->c_contextid,
tag,
comm->c_my_rank,
proc,
ptl_request);
} else {
ret = ompi_mtl_portals4_long_isend(start,
length,
comm->c_contextid,
tag,
comm->c_my_rank,
proc,
ptl_request);
}
return ret;
}
int
ompi_osc_pt2pt_component_select(ompi_win_t *win,
ompi_info_t *info,
ompi_communicator_t *comm)
{
ompi_osc_pt2pt_module_t *module = NULL;
int ret, i;
ompi_osc_pt2pt_buffer_t *buffer = NULL;
opal_free_list_item_t *item = NULL;
char *tmp = NULL;
/* create module structure */
module = (ompi_osc_pt2pt_module_t*)
calloc(1, sizeof(ompi_osc_pt2pt_module_t));
if (NULL == module) return OMPI_ERR_TEMP_OUT_OF_RESOURCE;
/* fill in the function pointer part */
memcpy(module, &ompi_osc_pt2pt_module_template,
sizeof(ompi_osc_base_module_t));
/* initialize the p2p part */
OBJ_CONSTRUCT(&(module->p2p_lock), opal_mutex_t);
OBJ_CONSTRUCT(&(module->p2p_cond), opal_condition_t);
OBJ_CONSTRUCT(&(module->p2p_acc_lock), opal_mutex_t);
OBJ_CONSTRUCT(&module->p2p_pending_sendreqs, opal_list_t);
OBJ_CONSTRUCT(&(module->p2p_copy_pending_sendreqs), opal_list_t);
OBJ_CONSTRUCT(&(module->p2p_locks_pending), opal_list_t);
OBJ_CONSTRUCT(&(module->p2p_unlocks_pending), opal_list_t);
module->p2p_win = win;
ret = ompi_comm_dup(comm, &(module->p2p_comm));
if (ret != OMPI_SUCCESS) goto cleanup;
opal_output_verbose(1, ompi_osc_base_framework.framework_output,
"pt2pt component creating window with id %d",
ompi_comm_get_cid(module->p2p_comm));
asprintf(&tmp, "%d", ompi_comm_get_cid(module->p2p_comm));
ompi_win_set_name(win, tmp);
free(tmp);
module->p2p_num_pending_sendreqs = (unsigned int*)
malloc(sizeof(unsigned int) * ompi_comm_size(module->p2p_comm));
if (NULL == module->p2p_num_pending_sendreqs) {
ret = OMPI_ERR_TEMP_OUT_OF_RESOURCE;
goto cleanup;
}
memset(module->p2p_num_pending_sendreqs, 0,
sizeof(unsigned int) * ompi_comm_size(module->p2p_comm));
module->p2p_num_pending_out = 0;
module->p2p_num_pending_in = 0;
module->p2p_num_post_msgs = 0;
module->p2p_num_complete_msgs = 0;
module->p2p_tag_counter = 0;
module->p2p_copy_num_pending_sendreqs = (unsigned int*)
malloc(sizeof(unsigned int) * ompi_comm_size(module->p2p_comm));
if (NULL == module->p2p_copy_num_pending_sendreqs) {
ret = OMPI_ERR_TEMP_OUT_OF_RESOURCE;
goto cleanup;
}
memset(module->p2p_num_pending_sendreqs, 0,
sizeof(unsigned int) * ompi_comm_size(module->p2p_comm));
/* fence data */
module->p2p_fence_coll_counts = (int*)
malloc(sizeof(int) * ompi_comm_size(module->p2p_comm));
if (NULL == module->p2p_fence_coll_counts) {
ret = OMPI_ERR_TEMP_OUT_OF_RESOURCE;
goto cleanup;
}
for (i = 0 ; i < ompi_comm_size(module->p2p_comm) ; ++i) {
module->p2p_fence_coll_counts[i] = 1;
}
/* pwsc data */
module->p2p_pw_group = NULL;
module->p2p_sc_group = NULL;
module->p2p_sc_remote_active_ranks = (bool*)
malloc(sizeof(bool) * ompi_comm_size(module->p2p_comm));
if (NULL == module->p2p_sc_remote_active_ranks) {
ret = OMPI_ERR_TEMP_OUT_OF_RESOURCE;
goto cleanup;
}
module->p2p_sc_remote_ranks = (int*)
malloc(sizeof(int) * ompi_comm_size(module->p2p_comm));
if (NULL == module->p2p_sc_remote_ranks) {
ret = OMPI_ERR_TEMP_OUT_OF_RESOURCE;
goto cleanup;
}
/* lock data */
module->p2p_lock_status = 0;
module->p2p_shared_count = 0;
module->p2p_lock_received_ack = 0;
/* fill in window information */
win->w_osc_module = (ompi_osc_base_module_t*) module;
/* sync memory - make sure all initialization completed */
opal_atomic_mb();
/* start up receive for protocol headers */
OPAL_FREE_LIST_GET(&mca_osc_pt2pt_component.p2p_c_buffers,
item, ret);
if (OMPI_SUCCESS != ret) goto cleanup;
buffer = (ompi_osc_pt2pt_buffer_t*) item;
buffer->data = (void*) module;
ret = ompi_osc_pt2pt_component_irecv(buffer->payload,
mca_osc_pt2pt_component.p2p_c_eager_size,
MPI_BYTE,
MPI_ANY_SOURCE,
CONTROL_MSG_TAG,
module->p2p_comm,
&(buffer->request),
component_fragment_cb,
buffer);
if (OMPI_SUCCESS != ret) goto cleanup;
return OMPI_SUCCESS;
cleanup:
OBJ_DESTRUCT(&module->p2p_unlocks_pending);
OBJ_DESTRUCT(&module->p2p_locks_pending);
OBJ_DESTRUCT(&module->p2p_copy_pending_sendreqs);
OBJ_DESTRUCT(&module->p2p_pending_sendreqs);
OBJ_DESTRUCT(&module->p2p_acc_lock);
OBJ_DESTRUCT(&module->p2p_cond);
OBJ_DESTRUCT(&module->p2p_lock);
if (NULL != buffer) {
OPAL_FREE_LIST_RETURN(&mca_osc_pt2pt_component.p2p_c_buffers, item);
}
if (NULL != module->p2p_sc_remote_ranks) {
free(module->p2p_sc_remote_ranks);
}
if (NULL != module->p2p_sc_remote_active_ranks) {
free(module->p2p_sc_remote_active_ranks);
}
if (NULL != module->p2p_fence_coll_counts) {
free(module->p2p_fence_coll_counts);
}
if (NULL != module->p2p_copy_num_pending_sendreqs) {
free(module->p2p_copy_num_pending_sendreqs);
}
if (NULL != module->p2p_num_pending_sendreqs) {
free(module->p2p_num_pending_sendreqs);
}
if (NULL != module->p2p_comm) ompi_comm_free(&module->p2p_comm);
#if OPAL_ENABLE_DEBUG
memset(module, 0, sizeof(ompi_osc_base_module_t));
#endif
if (NULL != module) free(module);
return ret;
}
