int MPIDI_CH3I_RMA_Cleanup_ops_aggressive(MPID_Win * win_ptr)
{
int i, local_completed = 0, remote_completed = 0;
int mpi_errno = MPI_SUCCESS;
MPIDI_RMA_Target_t *curr_target = NULL;
int made_progress = 0;
/* If we are in an aggressive cleanup, the window must be holding
* up resources. If it isn't, we are in the wrong window and
* incorrectly entered this function. */
MPIU_ERR_CHKANDJUMP(win_ptr->non_empty_slots == 0, mpi_errno, MPI_ERR_OTHER, "**rmanoop");
/* find the first target that has something to issue */
for (i = 0; i < win_ptr->num_slots; i++) {
if (win_ptr->slots[i].target_list != NULL) {
curr_target = win_ptr->slots[i].target_list;
while (curr_target != NULL && curr_target->pending_op_list == NULL)
curr_target = curr_target->next;
if (curr_target != NULL)
break;
}
}
if (curr_target == NULL)
goto fn_exit;
if (curr_target->sync.sync_flag < MPIDI_RMA_SYNC_FLUSH_LOCAL)
curr_target->sync.sync_flag = MPIDI_RMA_SYNC_FLUSH_LOCAL;
/* Issue out all operations. */
mpi_errno = MPIDI_CH3I_RMA_Make_progress_target(win_ptr, curr_target->target_rank,
&made_progress);
if (mpi_errno != MPI_SUCCESS)
MPIU_ERR_POP(mpi_errno);
/* Wait for local completion. */
do {
mpi_errno = MPIDI_CH3I_RMA_Cleanup_ops_target(win_ptr, curr_target,
&local_completed, &remote_completed);
if (mpi_errno != MPI_SUCCESS)
MPIU_ERR_POP(mpi_errno);
if (!local_completed) {
mpi_errno = wait_progress_engine();
if (mpi_errno != MPI_SUCCESS)
MPIU_ERR_POP(mpi_errno);
}
} while (!local_completed);
fn_exit:
return mpi_errno;
fn_fail:
goto fn_exit;
}
int MPIDI_CH3I_RMA_Cleanup_ops_aggressive(MPIR_Win * win_ptr)
{
int i, local_completed = 0, remote_completed ATTRIBUTE((unused)) = 0;
int mpi_errno = MPI_SUCCESS;
MPIDI_RMA_Target_t *curr_target = NULL;
int made_progress = 0;
/* find the first target that has something to issue */
for (i = 0; i < win_ptr->num_slots; i++) {
if (win_ptr->slots[i].target_list_head != NULL) {
curr_target = win_ptr->slots[i].target_list_head;
while (curr_target != NULL && curr_target->pending_net_ops_list_head == NULL &&
curr_target->pending_user_ops_list_head == NULL)
curr_target = curr_target->next;
if (curr_target != NULL)
break;
}
}
if (curr_target == NULL)
goto fn_exit;
if (curr_target->sync.sync_flag < MPIDI_RMA_SYNC_FLUSH_LOCAL)
curr_target->sync.sync_flag = MPIDI_RMA_SYNC_FLUSH_LOCAL;
/* Issue out all operations. */
mpi_errno = MPIDI_CH3I_RMA_Make_progress_target(win_ptr, curr_target->target_rank,
&made_progress);
if (mpi_errno != MPI_SUCCESS)
MPIR_ERR_POP(mpi_errno);
/* Wait for local completion. */
do {
MPIDI_CH3I_RMA_ops_completion(win_ptr, curr_target, local_completed, remote_completed);
if (!local_completed) {
mpi_errno = wait_progress_engine();
if (mpi_errno != MPI_SUCCESS)
MPIR_ERR_POP(mpi_errno);
}
} while (!local_completed);
fn_exit:
return mpi_errno;
fn_fail:
goto fn_exit;
}
int MPIDI_CH3I_RMA_Cleanup_target_aggressive(MPIR_Win * win_ptr, MPIDI_RMA_Target_t ** target)
{
int i, local_completed ATTRIBUTE((unused)) = 0, remote_completed = 0;
int made_progress = 0;
MPIDI_RMA_Target_t *curr_target = NULL;
int mpi_errno = MPI_SUCCESS;
(*target) = NULL;
if (win_ptr->states.access_state == MPIDI_RMA_LOCK_ALL_CALLED) {
/* switch to window-wide protocol */
MPIDI_VC_t *orig_vc = NULL, *target_vc = NULL;
MPIDI_Comm_get_vc(win_ptr->comm_ptr, win_ptr->comm_ptr->rank, &orig_vc);
for (i = 0; i < win_ptr->comm_ptr->local_size; i++) {
if (i == win_ptr->comm_ptr->rank)
continue;
MPIDI_Comm_get_vc(win_ptr->comm_ptr, i, &target_vc);
if (orig_vc->node_id != target_vc->node_id) {
mpi_errno = MPIDI_CH3I_Win_find_target(win_ptr, i, &curr_target);
if (mpi_errno)
MPIR_ERR_POP(mpi_errno);
if (curr_target == NULL) {
win_ptr->outstanding_locks++;
mpi_errno = send_lock_msg(i, MPI_LOCK_SHARED, win_ptr);
if (mpi_errno != MPI_SUCCESS)
MPIR_ERR_POP(mpi_errno);
}
}
}
win_ptr->states.access_state = MPIDI_RMA_LOCK_ALL_ISSUED;
}
do {
/* find a non-empty slot and set the FLUSH flag on the first
* target */
/* TODO: we should think about better strategies on selecting the target */
for (i = 0; i < win_ptr->num_slots; i++)
if (win_ptr->slots[i].target_list_head != NULL)
break;
curr_target = win_ptr->slots[i].target_list_head;
if (curr_target->sync.sync_flag < MPIDI_RMA_SYNC_FLUSH) {
curr_target->sync.sync_flag = MPIDI_RMA_SYNC_FLUSH;
}
/* Issue out all operations. */
mpi_errno = MPIDI_CH3I_RMA_Make_progress_target(win_ptr, curr_target->target_rank,
&made_progress);
if (mpi_errno != MPI_SUCCESS)
MPIR_ERR_POP(mpi_errno);
/* Wait for remote completion. */
do {
MPIDI_CH3I_RMA_ops_completion(win_ptr, curr_target, local_completed, remote_completed);
if (!remote_completed) {
mpi_errno = wait_progress_engine();
if (mpi_errno != MPI_SUCCESS)
MPIR_ERR_POP(mpi_errno);
}
} while (!remote_completed);
/* Cleanup the target. */
mpi_errno = MPIDI_CH3I_Win_target_dequeue_and_free(win_ptr, curr_target);
if (mpi_errno != MPI_SUCCESS)
MPIR_ERR_POP(mpi_errno);
/* check if we got a target */
(*target) = MPIDI_CH3I_Win_target_alloc(win_ptr);
} while ((*target) == NULL);
fn_exit:
return mpi_errno;
fn_fail:
goto fn_exit;
}
int MPID_Win_free(MPIR_Win ** win_ptr)
{
int mpi_errno = MPI_SUCCESS;
int in_use;
MPIR_Comm *comm_ptr;
MPIR_Errflag_t errflag = MPIR_ERR_NONE;
MPIR_FUNC_VERBOSE_STATE_DECL(MPID_STATE_MPID_WIN_FREE);
MPIR_FUNC_VERBOSE_RMA_ENTER(MPID_STATE_MPID_WIN_FREE);
MPIR_ERR_CHKANDJUMP(((*win_ptr)->states.access_state != MPIDI_RMA_NONE &&
(*win_ptr)->states.access_state != MPIDI_RMA_FENCE_ISSUED &&
(*win_ptr)->states.access_state != MPIDI_RMA_FENCE_GRANTED) ||
((*win_ptr)->states.exposure_state != MPIDI_RMA_NONE),
mpi_errno, MPI_ERR_RMA_SYNC, "**rmasync");
/* 1. Here we must wait until all passive locks are released on this target,
* because for some UNLOCK messages, we do not send ACK back to origin,
* we must wait until lock is released so that we can free window.
* 2. We also need to wait until AT completion counter being zero, because
* this counter is increment everytime we meet a GET-like operation, it is
* possible that when target entering Win_free, passive epoch is not finished
* yet and there are still GETs doing on this target.
* 3. We also need to wait until lock queue becomes empty. It is possible
* that some lock requests is still waiting in the queue when target is
* entering Win_free. */
while ((*win_ptr)->current_lock_type != MPID_LOCK_NONE ||
(*win_ptr)->at_completion_counter != 0 ||
(*win_ptr)->target_lock_queue_head != NULL ||
(*win_ptr)->current_target_lock_data_bytes != 0 || (*win_ptr)->sync_request_cnt != 0) {
mpi_errno = wait_progress_engine();
if (mpi_errno != MPI_SUCCESS)
MPIR_ERR_POP(mpi_errno);
}
mpi_errno = MPID_Barrier((*win_ptr)->comm_ptr, &errflag);
if (mpi_errno)
MPIR_ERR_POP(mpi_errno);
/* Free window resources in lower layer. */
if (MPIDI_CH3U_Win_hooks.win_free != NULL) {
mpi_errno = MPIDI_CH3U_Win_hooks.win_free(win_ptr);
if (mpi_errno != MPI_SUCCESS)
MPIR_ERR_POP(mpi_errno);
}
/* dequeue window from the global list */
MPIR_Assert((*win_ptr)->active == FALSE);
DL_DELETE(MPIDI_RMA_Win_inactive_list_head, (*win_ptr));
if (MPIDI_RMA_Win_inactive_list_head == NULL && MPIDI_RMA_Win_active_list_head == NULL) {
/* this is the last window, de-register RMA progress hook */
mpi_errno = MPID_Progress_deregister_hook(MPIDI_CH3I_RMA_Progress_hook_id);
if (mpi_errno != MPI_SUCCESS) {
MPIR_ERR_POP(mpi_errno);
}
}
comm_ptr = (*win_ptr)->comm_ptr;
mpi_errno = MPIR_Comm_free_impl(comm_ptr);
if (mpi_errno)
MPIR_ERR_POP(mpi_errno);
if ((*win_ptr)->basic_info_table != NULL)
MPL_free((*win_ptr)->basic_info_table);
MPL_free((*win_ptr)->op_pool_start);
MPL_free((*win_ptr)->target_pool_start);
MPL_free((*win_ptr)->slots);
MPL_free((*win_ptr)->target_lock_entry_pool_start);
MPIR_Assert((*win_ptr)->current_target_lock_data_bytes == 0);
/* Free the attached buffer for windows created with MPI_Win_allocate() */
if ((*win_ptr)->create_flavor == MPI_WIN_FLAVOR_ALLOCATE ||
(*win_ptr)->create_flavor == MPI_WIN_FLAVOR_SHARED) {
if ((*win_ptr)->shm_allocated == FALSE && (*win_ptr)->size > 0) {
MPL_free((*win_ptr)->base);
}
}
MPIR_Object_release_ref(*win_ptr, &in_use);
/* MPI windows don't have reference count semantics, so this should always be true */
MPIR_Assert(!in_use);
MPIR_Handle_obj_free(&MPIR_Win_mem, *win_ptr);
fn_exit:
MPIR_FUNC_VERBOSE_RMA_EXIT(MPID_STATE_MPID_WIN_FREE);
return mpi_errno;
fn_fail:
goto fn_exit;
}
int MPID_Compare_and_swap(const void *origin_addr, const void *compare_addr,
void *result_addr, MPI_Datatype datatype, int target_rank,
MPI_Aint target_disp, MPID_Win * win_ptr)
{
int mpi_errno = MPI_SUCCESS;
int rank;
MPIDI_VC_t *orig_vc = NULL, *target_vc = NULL;
int made_progress = 0;
MPIDI_STATE_DECL(MPID_STATE_MPID_COMPARE_AND_SWAP);
MPIDI_RMA_FUNC_ENTER(MPID_STATE_MPID_COMPARE_AND_SWAP);
MPIR_ERR_CHKANDJUMP(win_ptr->states.access_state == MPIDI_RMA_NONE,
mpi_errno, MPI_ERR_RMA_SYNC, "**rmasync");
if (target_rank == MPI_PROC_NULL) {
goto fn_exit;
}
rank = win_ptr->comm_ptr->rank;
if (win_ptr->shm_allocated == TRUE && target_rank != rank &&
win_ptr->create_flavor != MPI_WIN_FLAVOR_SHARED) {
/* check if target is local and shared memory is allocated on window,
* if so, we directly perform this operation on shared memory region. */
/* FIXME: Here we decide whether to perform SHM operations by checking if origin and target are on
* the same node. However, in ch3:sock, even if origin and target are on the same node, they do
* not within the same SHM region. Here we filter out ch3:sock by checking shm_allocated flag first,
* which is only set to TRUE when SHM region is allocated in nemesis.
* In future we need to figure out a way to check if origin and target are in the same "SHM comm".
*/
MPIDI_Comm_get_vc(win_ptr->comm_ptr, rank, &orig_vc);
MPIDI_Comm_get_vc(win_ptr->comm_ptr, target_rank, &target_vc);
}
/* The datatype must be predefined, and one of: C integer, Fortran integer,
* Logical, Multi-language types, or Byte. This is checked above the ADI,
* so there's no need to check it again here. */
/* FIXME: For shared memory windows, we should provide an implementation
* that uses a processor atomic operation. */
if (target_rank == rank || win_ptr->create_flavor == MPI_WIN_FLAVOR_SHARED ||
(win_ptr->shm_allocated == TRUE && orig_vc->node_id == target_vc->node_id)) {
mpi_errno = MPIDI_CH3I_Shm_cas_op(origin_addr, compare_addr, result_addr,
datatype, target_rank, target_disp, win_ptr);
if (mpi_errno)
MPIR_ERR_POP(mpi_errno);
}
else {
MPIDI_RMA_Op_t *op_ptr = NULL;
MPIDI_CH3_Pkt_cas_t *cas_pkt = NULL;
MPI_Aint type_size;
void *src = NULL, *dest = NULL;
/* Append this operation to the RMA ops queue */
mpi_errno = MPIDI_CH3I_Win_get_op(win_ptr, &op_ptr);
if (mpi_errno != MPI_SUCCESS)
MPIR_ERR_POP(mpi_errno);
MPIR_T_PVAR_TIMER_START(RMA, rma_rmaqueue_set);
/******************** Setting operation struct areas ***********************/
op_ptr->origin_addr = (void *) origin_addr;
op_ptr->origin_count = 1;
op_ptr->origin_datatype = datatype;
op_ptr->result_addr = result_addr;
op_ptr->result_datatype = datatype;
op_ptr->compare_addr = (void *) compare_addr;
op_ptr->compare_datatype = datatype;
op_ptr->target_rank = target_rank;
op_ptr->piggyback_lock_candidate = 1; /* CAS is always able to piggyback LOCK */
/************** Setting packet struct areas in operation ****************/
cas_pkt = &(op_ptr->pkt.cas);
MPIDI_Pkt_init(cas_pkt, MPIDI_CH3_PKT_CAS_IMMED);
cas_pkt->addr = (char *) win_ptr->basic_info_table[target_rank].base_addr +
win_ptr->basic_info_table[target_rank].disp_unit * target_disp;
cas_pkt->datatype = datatype;
cas_pkt->target_win_handle = win_ptr->basic_info_table[target_rank].win_handle;
cas_pkt->flags = MPIDI_CH3_PKT_FLAG_NONE;
/* REQUIRE: All datatype arguments must be of the same, builtin
* type and counts must be 1. */
MPID_Datatype_get_size_macro(datatype, type_size);
MPIU_Assert(type_size <= sizeof(MPIDI_CH3_CAS_Immed_u));
src = (void *) origin_addr, dest = (void *) (&(cas_pkt->origin_data));
mpi_errno = immed_copy(src, dest, type_size);
if (mpi_errno != MPI_SUCCESS)
MPIR_ERR_POP(mpi_errno);
src = (void *) compare_addr, dest = (void *) (&(cas_pkt->compare_data));
mpi_errno = immed_copy(src, dest, type_size);
if (mpi_errno != MPI_SUCCESS)
MPIR_ERR_POP(mpi_errno);
MPIR_T_PVAR_TIMER_END(RMA, rma_rmaqueue_set);
mpi_errno = MPIDI_CH3I_Win_enqueue_op(win_ptr, op_ptr);
if (mpi_errno)
MPIR_ERR_POP(mpi_errno);
mpi_errno = MPIDI_CH3I_RMA_Make_progress_target(win_ptr, target_rank, &made_progress);
if (mpi_errno != MPI_SUCCESS)
MPIR_ERR_POP(mpi_errno);
if (MPIR_CVAR_CH3_RMA_ACTIVE_REQ_THRESHOLD >= 0 &&
MPIDI_CH3I_RMA_Active_req_cnt >= MPIR_CVAR_CH3_RMA_ACTIVE_REQ_THRESHOLD) {
while (MPIDI_CH3I_RMA_Active_req_cnt >= MPIR_CVAR_CH3_RMA_ACTIVE_REQ_THRESHOLD) {
mpi_errno = wait_progress_engine();
if (mpi_errno != MPI_SUCCESS)
MPIR_ERR_POP(mpi_errno);
}
}
}
fn_exit:
MPIDI_RMA_FUNC_EXIT(MPID_STATE_MPID_COMPARE_AND_SWAP);
return mpi_errno;
/* --BEGIN ERROR HANDLING-- */
fn_fail:
goto fn_exit;
/* --END ERROR HANDLING-- */
}
int MPIDI_CH3I_Get_accumulate(const void *origin_addr, int origin_count,
MPI_Datatype origin_datatype, void *result_addr, int result_count,
MPI_Datatype result_datatype, int target_rank, MPI_Aint target_disp,
int target_count, MPI_Datatype target_datatype, MPI_Op op,
MPID_Win * win_ptr, MPID_Request * ureq)
{
int mpi_errno = MPI_SUCCESS;
MPIDI_msg_sz_t orig_data_sz, target_data_sz;
int rank;
int dt_contig ATTRIBUTE((unused));
MPI_Aint dt_true_lb ATTRIBUTE((unused));
MPID_Datatype *dtp;
MPIDI_VC_t *orig_vc = NULL, *target_vc = NULL;
int made_progress = 0;
MPIDI_STATE_DECL(MPID_STATE_MPIDI_CH3I_GET_ACCUMULATE);
MPIDI_RMA_FUNC_ENTER(MPID_STATE_MPIDI_CH3I_GET_ACCUMULATE);
MPIR_ERR_CHKANDJUMP(win_ptr->states.access_state == MPIDI_RMA_NONE,
mpi_errno, MPI_ERR_RMA_SYNC, "**rmasync");
if (target_rank == MPI_PROC_NULL) {
goto fn_exit;
}
MPIDI_Datatype_get_info(target_count, target_datatype, dt_contig, target_data_sz, dtp,
dt_true_lb);
if (target_data_sz == 0) {
goto fn_exit;
}
rank = win_ptr->comm_ptr->rank;
if (win_ptr->shm_allocated == TRUE && target_rank != rank &&
win_ptr->create_flavor != MPI_WIN_FLAVOR_SHARED) {
/* check if target is local and shared memory is allocated on window,
* if so, we directly perform this operation on shared memory region. */
/* FIXME: Here we decide whether to perform SHM operations by checking if origin and target are on
* the same node. However, in ch3:sock, even if origin and target are on the same node, they do
* not within the same SHM region. Here we filter out ch3:sock by checking shm_allocated flag first,
* which is only set to TRUE when SHM region is allocated in nemesis.
* In future we need to figure out a way to check if origin and target are in the same "SHM comm".
*/
MPIDI_Comm_get_vc(win_ptr->comm_ptr, rank, &orig_vc);
MPIDI_Comm_get_vc(win_ptr->comm_ptr, target_rank, &target_vc);
}
/* Do =! rank first (most likely branch?) */
if (target_rank == rank || win_ptr->create_flavor == MPI_WIN_FLAVOR_SHARED ||
(win_ptr->shm_allocated == TRUE && orig_vc->node_id == target_vc->node_id)) {
mpi_errno = MPIDI_CH3I_Shm_get_acc_op(origin_addr, origin_count, origin_datatype,
result_addr, result_count, result_datatype,
target_rank, target_disp, target_count,
target_datatype, op, win_ptr);
if (mpi_errno)
MPIR_ERR_POP(mpi_errno);
if (ureq) {
/* Complete user request and release the ch3 ref */
mpi_errno = MPID_Request_complete(ureq);
if (mpi_errno != MPI_SUCCESS) {
MPIR_ERR_POP(mpi_errno);
}
}
}
else {
MPIDI_RMA_Op_t *op_ptr = NULL;
MPIDI_CH3_Pkt_get_accum_t *get_accum_pkt;
MPI_Aint origin_type_size;
MPI_Aint target_type_size;
int use_immed_pkt = FALSE, i;
int is_origin_contig, is_target_contig, is_result_contig;
MPI_Aint stream_elem_count, stream_unit_count;
MPI_Aint predefined_dtp_size, predefined_dtp_count, predefined_dtp_extent;
MPID_Datatype *origin_dtp = NULL, *target_dtp = NULL, *result_dtp = NULL;
int is_empty_origin = FALSE;
/* Judge if origin buffer is empty */
if (op == MPI_NO_OP)
is_empty_origin = TRUE;
/* Append the operation to the window's RMA ops queue */
mpi_errno = MPIDI_CH3I_Win_get_op(win_ptr, &op_ptr);
if (mpi_errno != MPI_SUCCESS)
MPIR_ERR_POP(mpi_errno);
/* TODO: Can we use the MPIDI_RMA_ACC_CONTIG optimization? */
MPIR_T_PVAR_TIMER_START(RMA, rma_rmaqueue_set);
/******************** Setting operation struct areas ***********************/
op_ptr->origin_addr = (void *) origin_addr;
op_ptr->origin_count = origin_count;
op_ptr->origin_datatype = origin_datatype;
op_ptr->result_addr = result_addr;
op_ptr->result_count = result_count;
op_ptr->result_datatype = result_datatype;
op_ptr->target_rank = target_rank;
/* Remember user request */
op_ptr->ureq = ureq;
/* if source or target datatypes are derived, increment their
* reference counts */
if (is_empty_origin == FALSE && !MPIR_DATATYPE_IS_PREDEFINED(origin_datatype)) {
MPID_Datatype_get_ptr(origin_datatype, origin_dtp);
}
if (!MPIR_DATATYPE_IS_PREDEFINED(result_datatype)) {
MPID_Datatype_get_ptr(result_datatype, result_dtp);
}
if (!MPIR_DATATYPE_IS_PREDEFINED(target_datatype)) {
MPID_Datatype_get_ptr(target_datatype, target_dtp);
}
if (is_empty_origin == FALSE) {
MPID_Datatype_get_size_macro(origin_datatype, origin_type_size);
MPIU_Assign_trunc(orig_data_sz, origin_count * origin_type_size, MPIDI_msg_sz_t);
}
else {
/* If origin buffer is empty, set origin data size to 0 */
orig_data_sz = 0;
}
MPID_Datatype_get_size_macro(target_datatype, target_type_size);
/* Get size and count for predefined datatype elements */
if (MPIR_DATATYPE_IS_PREDEFINED(target_datatype)) {
predefined_dtp_size = target_type_size;
predefined_dtp_count = target_count;
MPID_Datatype_get_extent_macro(target_datatype, predefined_dtp_extent);
}
else {
MPIU_Assert(target_dtp->basic_type != MPI_DATATYPE_NULL);
MPID_Datatype_get_size_macro(target_dtp->basic_type, predefined_dtp_size);
predefined_dtp_count = target_data_sz / predefined_dtp_size;
MPID_Datatype_get_extent_macro(target_dtp->basic_type, predefined_dtp_extent);
}
MPIU_Assert(predefined_dtp_count > 0 && predefined_dtp_size > 0 &&
predefined_dtp_extent > 0);
/* Calculate number of predefined elements in each stream unit, and
* total number of stream units. */
stream_elem_count = MPIDI_CH3U_Acc_stream_size / predefined_dtp_extent;
stream_unit_count = (predefined_dtp_count - 1) / stream_elem_count + 1;
MPIU_Assert(stream_elem_count > 0 && stream_unit_count > 0);
for (i = 0; i < stream_unit_count; i++) {
if (origin_dtp != NULL) {
MPID_Datatype_add_ref(origin_dtp);
}
if (target_dtp != NULL) {
MPID_Datatype_add_ref(target_dtp);
}
if (result_dtp != NULL) {
MPID_Datatype_add_ref(result_dtp);
}
}
if (is_empty_origin == FALSE) {
MPID_Datatype_is_contig(origin_datatype, &is_origin_contig);
}
else {
/* If origin buffer is empty, mark origin data as contig data */
is_origin_contig = 1;
}
MPID_Datatype_is_contig(target_datatype, &is_target_contig);
MPID_Datatype_is_contig(result_datatype, &is_result_contig);
/* Judge if we can use IMMED data packet */
if ((is_empty_origin == TRUE || MPIR_DATATYPE_IS_PREDEFINED(origin_datatype)) &&
MPIR_DATATYPE_IS_PREDEFINED(result_datatype) &&
MPIR_DATATYPE_IS_PREDEFINED(target_datatype) &&
is_origin_contig && is_target_contig && is_result_contig) {
if (target_data_sz <= MPIDI_RMA_IMMED_BYTES)
use_immed_pkt = TRUE;
}
/* Judge if this operation is a piggyback candidate */
if ((is_empty_origin == TRUE || MPIR_DATATYPE_IS_PREDEFINED(origin_datatype)) &&
MPIR_DATATYPE_IS_PREDEFINED(result_datatype) &&
MPIR_DATATYPE_IS_PREDEFINED(target_datatype)) {
/* FIXME: currently we only piggyback LOCK flag with op using predefined datatypes
* for origin, target and result data. We should extend this optimization to derived
* datatypes as well. */
if (orig_data_sz <= MPIR_CVAR_CH3_RMA_OP_PIGGYBACK_LOCK_DATA_SIZE)
op_ptr->piggyback_lock_candidate = 1;
}
/************** Setting packet struct areas in operation ****************/
get_accum_pkt = &(op_ptr->pkt.get_accum);
if (use_immed_pkt) {
MPIDI_Pkt_init(get_accum_pkt, MPIDI_CH3_PKT_GET_ACCUM_IMMED);
}
else {
MPIDI_Pkt_init(get_accum_pkt, MPIDI_CH3_PKT_GET_ACCUM);
}
get_accum_pkt->addr = (char *) win_ptr->basic_info_table[target_rank].base_addr +
win_ptr->basic_info_table[target_rank].disp_unit * target_disp;
get_accum_pkt->count = target_count;
get_accum_pkt->datatype = target_datatype;
get_accum_pkt->info.dataloop_size = 0;
get_accum_pkt->op = op;
get_accum_pkt->target_win_handle = win_ptr->basic_info_table[target_rank].win_handle;
get_accum_pkt->flags = MPIDI_CH3_PKT_FLAG_NONE;
if (use_immed_pkt) {
void *src = (void *) origin_addr, *dest = (void *) (get_accum_pkt->info.data);
mpi_errno = immed_copy(src, dest, orig_data_sz);
if (mpi_errno != MPI_SUCCESS)
MPIR_ERR_POP(mpi_errno);
}
MPIR_T_PVAR_TIMER_END(RMA, rma_rmaqueue_set);
mpi_errno = MPIDI_CH3I_Win_enqueue_op(win_ptr, op_ptr);
if (mpi_errno)
MPIR_ERR_POP(mpi_errno);
mpi_errno = MPIDI_CH3I_RMA_Make_progress_target(win_ptr, target_rank, &made_progress);
if (mpi_errno != MPI_SUCCESS)
MPIR_ERR_POP(mpi_errno);
if (MPIR_CVAR_CH3_RMA_ACTIVE_REQ_THRESHOLD >= 0 &&
MPIDI_CH3I_RMA_Active_req_cnt >= MPIR_CVAR_CH3_RMA_ACTIVE_REQ_THRESHOLD) {
while (MPIDI_CH3I_RMA_Active_req_cnt >= MPIR_CVAR_CH3_RMA_ACTIVE_REQ_THRESHOLD) {
mpi_errno = wait_progress_engine();
if (mpi_errno != MPI_SUCCESS)
MPIR_ERR_POP(mpi_errno);
}
}
}
fn_exit:
MPIDI_RMA_FUNC_EXIT(MPID_STATE_MPIDI_CH3I_GET_ACCUMULATE);
return mpi_errno;
/* --BEGIN ERROR HANDLING-- */
fn_fail:
goto fn_exit;
/* --END ERROR HANDLING-- */
}
int MPIDI_CH3I_Put(const void *origin_addr, int origin_count, MPI_Datatype
origin_datatype, int target_rank, MPI_Aint target_disp,
int target_count, MPI_Datatype target_datatype, MPID_Win * win_ptr,
MPID_Request * ureq)
{
int mpi_errno = MPI_SUCCESS;
int dt_contig ATTRIBUTE((unused)), rank;
MPID_Datatype *dtp;
MPI_Aint dt_true_lb ATTRIBUTE((unused));
MPIDI_msg_sz_t data_sz;
MPIDI_VC_t *orig_vc = NULL, *target_vc = NULL;
int made_progress = 0;
MPIDI_STATE_DECL(MPID_STATE_MPIDI_CH3I_PUT);
MPIDI_RMA_FUNC_ENTER(MPID_STATE_MPIDI_CH3I_PUT);
MPIR_ERR_CHKANDJUMP(win_ptr->states.access_state == MPIDI_RMA_NONE,
mpi_errno, MPI_ERR_RMA_SYNC, "**rmasync");
if (target_rank == MPI_PROC_NULL) {
goto fn_exit;
}
MPIDI_Datatype_get_info(origin_count, origin_datatype, dt_contig, data_sz, dtp, dt_true_lb);
if (data_sz == 0) {
goto fn_exit;
}
rank = win_ptr->comm_ptr->rank;
if (win_ptr->shm_allocated == TRUE && target_rank != rank &&
win_ptr->create_flavor != MPI_WIN_FLAVOR_SHARED) {
/* check if target is local and shared memory is allocated on window,
* if so, we directly perform this operation on shared memory region. */
/* FIXME: Here we decide whether to perform SHM operations by checking if origin and target are on
* the same node. However, in ch3:sock, even if origin and target are on the same node, they do
* not within the same SHM region. Here we filter out ch3:sock by checking shm_allocated flag first,
* which is only set to TRUE when SHM region is allocated in nemesis.
* In future we need to figure out a way to check if origin and target are in the same "SHM comm".
*/
MPIDI_Comm_get_vc(win_ptr->comm_ptr, rank, &orig_vc);
MPIDI_Comm_get_vc(win_ptr->comm_ptr, target_rank, &target_vc);
}
/* If the put is a local operation, do it here */
if (target_rank == rank || win_ptr->create_flavor == MPI_WIN_FLAVOR_SHARED ||
(win_ptr->shm_allocated == TRUE && orig_vc->node_id == target_vc->node_id)) {
mpi_errno = MPIDI_CH3I_Shm_put_op(origin_addr, origin_count, origin_datatype, target_rank,
target_disp, target_count, target_datatype, win_ptr);
if (mpi_errno)
MPIR_ERR_POP(mpi_errno);
if (ureq) {
/* Complete user request and release the ch3 ref */
mpi_errno = MPID_Request_complete(ureq);
if (mpi_errno != MPI_SUCCESS) {
MPIR_ERR_POP(mpi_errno);
}
}
}
else {
MPIDI_RMA_Op_t *op_ptr = NULL;
MPIDI_CH3_Pkt_put_t *put_pkt = NULL;
int use_immed_pkt = FALSE;
int is_origin_contig, is_target_contig;
/* queue it up */
mpi_errno = MPIDI_CH3I_Win_get_op(win_ptr, &op_ptr);
if (mpi_errno != MPI_SUCCESS)
MPIR_ERR_POP(mpi_errno);
MPIR_T_PVAR_TIMER_START(RMA, rma_rmaqueue_set);
/******************** Setting operation struct areas ***********************/
/* FIXME: For contig and very short operations, use a streamlined op */
op_ptr->origin_addr = (void *) origin_addr;
op_ptr->origin_count = origin_count;
op_ptr->origin_datatype = origin_datatype;
op_ptr->target_rank = target_rank;
/* Remember user request */
op_ptr->ureq = ureq;
/* if source or target datatypes are derived, increment their
* reference counts */
if (!MPIR_DATATYPE_IS_PREDEFINED(origin_datatype)) {
MPID_Datatype_get_ptr(origin_datatype, dtp);
MPID_Datatype_add_ref(dtp);
}
if (!MPIR_DATATYPE_IS_PREDEFINED(target_datatype)) {
MPID_Datatype_get_ptr(target_datatype, dtp);
MPID_Datatype_add_ref(dtp);
}
MPID_Datatype_is_contig(origin_datatype, &is_origin_contig);
MPID_Datatype_is_contig(target_datatype, &is_target_contig);
/* Judge if we can use IMMED data packet */
if (MPIR_DATATYPE_IS_PREDEFINED(origin_datatype) &&
MPIR_DATATYPE_IS_PREDEFINED(target_datatype) && is_origin_contig && is_target_contig) {
if (data_sz <= MPIDI_RMA_IMMED_BYTES)
use_immed_pkt = TRUE;
}
/* Judge if this operation is an piggyback candidate */
if (MPIR_DATATYPE_IS_PREDEFINED(origin_datatype) &&
MPIR_DATATYPE_IS_PREDEFINED(target_datatype)) {
/* FIXME: currently we only piggyback LOCK flag with op using predefined datatypes
* for both origin and target data. We should extend this optimization to derived
* datatypes as well. */
if (data_sz <= MPIR_CVAR_CH3_RMA_OP_PIGGYBACK_LOCK_DATA_SIZE)
op_ptr->piggyback_lock_candidate = 1;
}
/************** Setting packet struct areas in operation ****************/
put_pkt = &(op_ptr->pkt.put);
if (use_immed_pkt) {
MPIDI_Pkt_init(put_pkt, MPIDI_CH3_PKT_PUT_IMMED);
}
else {
MPIDI_Pkt_init(put_pkt, MPIDI_CH3_PKT_PUT);
}
put_pkt->addr = (char *) win_ptr->basic_info_table[target_rank].base_addr +
win_ptr->basic_info_table[target_rank].disp_unit * target_disp;
put_pkt->count = target_count;
put_pkt->datatype = target_datatype;
put_pkt->info.dataloop_size = 0;
put_pkt->target_win_handle = win_ptr->basic_info_table[target_rank].win_handle;
put_pkt->source_win_handle = win_ptr->handle;
put_pkt->flags = MPIDI_CH3_PKT_FLAG_NONE;
if (use_immed_pkt) {
void *src = (void *) origin_addr, *dest = (void *) (put_pkt->info.data);
mpi_errno = immed_copy(src, dest, data_sz);
if (mpi_errno != MPI_SUCCESS)
MPIR_ERR_POP(mpi_errno);
}
MPIR_T_PVAR_TIMER_END(RMA, rma_rmaqueue_set);
mpi_errno = MPIDI_CH3I_Win_enqueue_op(win_ptr, op_ptr);
if (mpi_errno)
MPIR_ERR_POP(mpi_errno);
mpi_errno = MPIDI_CH3I_RMA_Make_progress_target(win_ptr, target_rank, &made_progress);
if (mpi_errno != MPI_SUCCESS)
MPIR_ERR_POP(mpi_errno);
if (MPIR_CVAR_CH3_RMA_ACTIVE_REQ_THRESHOLD >= 0 &&
MPIDI_CH3I_RMA_Active_req_cnt >= MPIR_CVAR_CH3_RMA_ACTIVE_REQ_THRESHOLD) {
while (MPIDI_CH3I_RMA_Active_req_cnt >= MPIR_CVAR_CH3_RMA_ACTIVE_REQ_THRESHOLD) {
mpi_errno = wait_progress_engine();
if (mpi_errno != MPI_SUCCESS)
MPIR_ERR_POP(mpi_errno);
}
}
}
fn_exit:
MPIDI_RMA_FUNC_EXIT(MPID_STATE_MPIDI_CH3I_PUT);
return mpi_errno;
/* --BEGIN ERROR HANDLING-- */
fn_fail:
goto fn_exit;
/* --END ERROR HANDLING-- */
}
int MPIDI_CH3I_RMA_Cleanup_target_aggressive(MPID_Win * win_ptr, MPIDI_RMA_Target_t ** target)
{
int i, local_completed = 0, remote_completed = 0;
int made_progress = 0;
MPIDI_RMA_Target_t *curr_target = NULL;
int mpi_errno = MPI_SUCCESS;
(*target) = NULL;
/* If we are in an aggressive cleanup, the window must be holding
* up resources. If it isn't, we are in the wrong window and
* incorrectly entered this function. */
MPIU_ERR_CHKANDJUMP(win_ptr->non_empty_slots == 0, mpi_errno, MPI_ERR_OTHER, "**rmanotarget");
if (win_ptr->states.access_state == MPIDI_RMA_LOCK_ALL_CALLED) {
/* switch to window-wide protocol */
MPIDI_VC_t *orig_vc = NULL, *target_vc = NULL;
MPIDI_Comm_get_vc(win_ptr->comm_ptr, win_ptr->comm_ptr->rank, &orig_vc);
for (i = 0; i < win_ptr->comm_ptr->local_size; i++) {
if (i == win_ptr->comm_ptr->rank)
continue;
MPIDI_Comm_get_vc(win_ptr->comm_ptr, i, &target_vc);
if (orig_vc->node_id != target_vc->node_id) {
mpi_errno = MPIDI_CH3I_Win_find_target(win_ptr, i, &curr_target);
if (mpi_errno)
MPIU_ERR_POP(mpi_errno);
if (curr_target == NULL) {
win_ptr->outstanding_locks++;
mpi_errno = send_lock_msg(i, MPI_LOCK_SHARED, win_ptr);
if (mpi_errno != MPI_SUCCESS)
MPIU_ERR_POP(mpi_errno);
}
}
}
win_ptr->states.access_state = MPIDI_RMA_LOCK_ALL_ISSUED;
}
do {
/* find a non-empty slot and set the FLUSH flag on the first
* target */
/* TODO: we should think about better strategies on selecting the target */
for (i = 0; i < win_ptr->num_slots; i++)
if (win_ptr->slots[i].target_list != NULL)
break;
curr_target = win_ptr->slots[i].target_list;
if (curr_target->sync.sync_flag < MPIDI_RMA_SYNC_FLUSH) {
curr_target->sync.sync_flag = MPIDI_RMA_SYNC_FLUSH;
curr_target->sync.have_remote_incomplete_ops = 0;
curr_target->sync.outstanding_acks++;
}
/* Issue out all operations. */
mpi_errno = MPIDI_CH3I_RMA_Make_progress_target(win_ptr, curr_target->target_rank,
&made_progress);
if (mpi_errno != MPI_SUCCESS)
MPIU_ERR_POP(mpi_errno);
/* Wait for remote completion. */
do {
mpi_errno = MPIDI_CH3I_RMA_Cleanup_ops_target(win_ptr, curr_target,
&local_completed, &remote_completed);
if (mpi_errno != MPI_SUCCESS)
MPIU_ERR_POP(mpi_errno);
if (!remote_completed) {
mpi_errno = wait_progress_engine();
if (mpi_errno != MPI_SUCCESS)
MPIU_ERR_POP(mpi_errno);
}
} while (!remote_completed);
/* Cleanup the target. */
mpi_errno = MPIDI_CH3I_RMA_Cleanup_single_target(win_ptr, curr_target);
if (mpi_errno != MPI_SUCCESS)
MPIU_ERR_POP(mpi_errno);
/* check if we got a target */
(*target) = MPIDI_CH3I_Win_target_alloc(win_ptr);
} while ((*target) == NULL);
fn_exit:
return mpi_errno;
fn_fail:
goto fn_exit;
}
