1
1
openmpi/ompi/mca/osc/rdma/osc_rdma_sync.c

642 строки
22 KiB
C
Исходник Обычный вид История

/*
* Copyright (c) 2004-2005 The Trustees of Indiana University.
* All rights reserved.
* Copyright (c) 2004-2005 The Trustees of the University of Tennessee.
* All rights reserved.
* Copyright (c) 2004-2005 High Performance Computing Center Stuttgart,
* University of Stuttgart. All rights reserved.
* Copyright (c) 2004-2005 The Regents of the University of California.
* All rights reserved.
* $COPYRIGHT$
*
* Additional copyrights may follow
*
* $HEADER$
*/
#include "ompi_config.h"
#include "osc_rdma.h"
#include "osc_rdma_sendreq.h"
#include "osc_rdma_longreq.h"
#include "osc_rdma_header.h"
#include "osc_rdma_data_move.h"
#include "mpi.h"
#include "opal/runtime/opal_progress.h"
#include "opal/threads/mutex.h"
#include "ompi/communicator/communicator.h"
#include "ompi/mca/osc/base/base.h"
/* should have p2p_lock before calling */
static inline void
ompi_osc_rdma_progress(ompi_osc_rdma_module_t *module)
{
if (0 != opal_list_get_size(&(module->p2p_long_msgs))) {
opal_list_item_t *item, *next;
OPAL_THREAD_LOCK(&(module->p2p_lock));
/* Have to go the convoluted while() route instead of a for()
loop because the callback will likely remove the request
from the list and free it, and that would lead to much
badness. */
next = opal_list_get_first(&(module->p2p_long_msgs));
while (opal_list_get_end(&(module->p2p_long_msgs)) != (item = next)) {
ompi_osc_rdma_longreq_t *longreq =
(ompi_osc_rdma_longreq_t*) item;
int ret, completed;
next = opal_list_get_next(item);
ret = ompi_request_test(&(longreq->req_pml_req), &completed, NULL);
/* BWB - FIX ME - error handling */
if (completed > 0) {
longreq->req_comp_cb(longreq);
}
}
OPAL_THREAD_UNLOCK(&(module->p2p_lock));
}
opal_progress();
}
static inline void
ompi_osc_rdma_flip_sendreqs(ompi_osc_rdma_module_t *module)
{
short *tmp;
OPAL_THREAD_LOCK(&(module->p2p_lock));
tmp = module->p2p_copy_num_pending_sendreqs;
module->p2p_copy_num_pending_sendreqs =
module->p2p_num_pending_sendreqs;
module->p2p_num_pending_sendreqs = tmp;
memset(module->p2p_num_pending_sendreqs, 0,
sizeof(short) * ompi_comm_size(module->p2p_comm));
/* Copy in all the pending requests */
opal_list_join(&module->p2p_copy_pending_sendreqs,
opal_list_get_end(&module->p2p_copy_pending_sendreqs),
&module->p2p_pending_sendreqs);
OPAL_THREAD_UNLOCK(&(module->p2p_lock));
}
int
ompi_osc_rdma_module_fence(int assert, ompi_win_t *win)
{
short incoming_reqs;
int ret = OMPI_SUCCESS, i;
if (0 != (assert & MPI_MODE_NOPRECEDE)) {
int num_pending;
/* check that the user didn't lie to us - since NOPRECEDED
must be specified by all processes if it is specified by
any process, if we see this it is safe to assume that there
are no pending operations anywhere needed to close out this
epoch. */
OPAL_THREAD_LOCK(&(P2P_MODULE(win)->p2p_lock));
num_pending = opal_list_get_size(&(P2P_MODULE(win)->p2p_pending_sendreqs));
OPAL_THREAD_UNLOCK(&(P2P_MODULE(win)->p2p_lock));
if (0 != num_pending) {
return MPI_ERR_RMA_SYNC;
}
} else {
opal_list_item_t *item;
ompi_osc_rdma_flip_sendreqs(P2P_MODULE(win));
switch (P2P_MODULE(win)->p2p_fence_sync_type) {
/* find out how much data everyone is going to send us. Need
to have the lock during this period so that we have a sane
view of the number of sendreqs */
case OSC_SYNC_REDUCE_SCATTER:
ret = P2P_MODULE(win)->p2p_comm->
c_coll.coll_reduce_scatter(P2P_MODULE(win)->p2p_copy_num_pending_sendreqs,
&incoming_reqs,
P2P_MODULE(win)->p2p_fence_coll_counts,
MPI_SHORT,
MPI_SUM,
P2P_MODULE(win)->p2p_comm);
break;
case OSC_SYNC_ALLREDUCE:
ret = P2P_MODULE(win)->p2p_comm->
c_coll.coll_allreduce(P2P_MODULE(win)->p2p_copy_num_pending_sendreqs,
P2P_MODULE(win)->p2p_fence_coll_results,
ompi_comm_size(P2P_MODULE(win)->p2p_comm),
MPI_SHORT,
MPI_SUM,
P2P_MODULE(win)->p2p_comm);
incoming_reqs = P2P_MODULE(win)->
p2p_fence_coll_results[P2P_MODULE(win)->p2p_comm->c_my_rank];
break;
case OSC_SYNC_ALLTOALL:
ret = P2P_MODULE(win)->p2p_comm->
c_coll.coll_alltoall(P2P_MODULE(win)->p2p_copy_num_pending_sendreqs,
1,
MPI_SHORT,
P2P_MODULE(win)->p2p_fence_coll_results,
1,
MPI_SHORT,
P2P_MODULE(win)->p2p_comm);
incoming_reqs = 0;
for (i = 0 ; i < ompi_comm_size(P2P_MODULE(win)->p2p_comm) ; ++i) {
incoming_reqs += P2P_MODULE(win)->p2p_fence_coll_results[i];
}
break;
default:
assert(0 == 1);
}
if (OMPI_SUCCESS != ret) {
/* put the stupid data back for the user. This is not
cheap, but the user lost his data if we don't. */
OPAL_THREAD_LOCK(&(P2P_MODULE(win)->p2p_lock));
opal_list_join(&P2P_MODULE(win)->p2p_pending_sendreqs,
opal_list_get_end(&P2P_MODULE(win)->p2p_pending_sendreqs),
&P2P_MODULE(win)->p2p_copy_pending_sendreqs);
for (i = 0 ; i < ompi_comm_size(P2P_MODULE(win)->p2p_comm) ; ++i) {
P2P_MODULE(win)->p2p_num_pending_sendreqs[i] +=
P2P_MODULE(win)->p2p_copy_num_pending_sendreqs[i];
}
OPAL_THREAD_UNLOCK(&(P2P_MODULE(win)->p2p_lock));
return ret;
}
/* possible we've already received a couple in messages, so
atomicall add however many we're going to wait for */
OPAL_THREAD_ADD32(&(P2P_MODULE(win)->p2p_num_pending_in), incoming_reqs);
OPAL_THREAD_ADD32(&(P2P_MODULE(win)->p2p_num_pending_out),
opal_list_get_size(&(P2P_MODULE(win)->p2p_copy_pending_sendreqs)));
opal_output_verbose(50, ompi_osc_base_output,
"fence: waiting on %d in and %d out",
P2P_MODULE(win)->p2p_num_pending_in,
P2P_MODULE(win)->p2p_num_pending_out);
/* try to start all the requests. We've copied everything we
need out of pending_sendreqs, so don't need the lock
here */
while (NULL !=
(item = opal_list_remove_first(&(P2P_MODULE(win)->p2p_copy_pending_sendreqs)))) {
ompi_osc_rdma_sendreq_t *req =
(ompi_osc_rdma_sendreq_t*) item;
ret = ompi_osc_rdma_sendreq_send(P2P_MODULE(win), req);
if (OMPI_SUCCESS != ret) {
opal_output_verbose(5, ompi_osc_base_output,
"fence: failure in starting sendreq (%d). Will try later.",
ret);
opal_list_append(&(P2P_MODULE(win)->p2p_copy_pending_sendreqs), item);
}
}
/* now we know how many things we're waiting for - wait for them... */
while (P2P_MODULE(win)->p2p_num_pending_in > 0 ||
0 != P2P_MODULE(win)->p2p_num_pending_out) {
ompi_osc_rdma_progress(P2P_MODULE(win));
}
}
/* all transfers are done - back to the real world we go */
if (0 == (assert & MPI_MODE_NOSUCCEED)) {
ompi_win_set_mode(win, OMPI_WIN_FENCE);
} else {
ompi_win_set_mode(win, 0);
}
return OMPI_SUCCESS;
}
int
ompi_osc_rdma_module_start(ompi_group_t *group,
int assert,
ompi_win_t *win)
{
int i;
OBJ_RETAIN(group);
/* BWB - do I need this? */
ompi_group_increment_proc_count(group);
OPAL_THREAD_LOCK(&(P2P_MODULE(win)->p2p_lock));
assert(NULL == P2P_MODULE(win)->p2p_sc_group);
P2P_MODULE(win)->p2p_sc_group = group;
OPAL_THREAD_UNLOCK(&(P2P_MODULE(win)->p2p_lock));
memset(P2P_MODULE(win)->p2p_sc_remote_active_ranks, 0,
sizeof(bool) * ompi_comm_size(P2P_MODULE(win)->p2p_comm));
/* for each process in the specified group, find it's rank in our
communicator, store those indexes, and set the true / false in
the active ranks table */
for (i = 0 ; i < ompi_group_size(group) ; i++) {
int comm_rank = -1, j;
/* no need to increment ref count - the communicator isn't
going anywhere while we're here */
ompi_group_t *comm_group = P2P_MODULE(win)->p2p_comm->c_local_group;
/* find the rank in the communicator associated with this windows */
for (j = 0 ;
j < ompi_group_size(comm_group) ;
++j) {
if (P2P_MODULE(win)->p2p_sc_group->grp_proc_pointers[i] ==
comm_group->grp_proc_pointers[j]) {
comm_rank = j;
break;
}
}
if (comm_rank == -1) {
return MPI_ERR_RMA_SYNC;
}
P2P_MODULE(win)->p2p_sc_remote_active_ranks[comm_rank] = true;
P2P_MODULE(win)->p2p_sc_remote_ranks[i] = comm_rank;
}
/* Set our mode to access w/ start */
ompi_win_remove_mode(win, OMPI_WIN_FENCE);
ompi_win_append_mode(win, OMPI_WIN_ACCESS_EPOCH | OMPI_WIN_STARTED);
/* possible we've already received a couple in messages, so
atomicall add however many we're going to wait for */
OPAL_THREAD_ADD32(&(P2P_MODULE(win)->p2p_num_post_msgs),
ompi_group_size(P2P_MODULE(win)->p2p_sc_group));
return OMPI_SUCCESS;
}
int
ompi_osc_rdma_module_complete(ompi_win_t *win)
{
int i;
int ret = OMPI_SUCCESS;
ompi_group_t *group;
opal_list_item_t *item;
/* wait for all the post messages */
while (0 != P2P_MODULE(win)->p2p_num_post_msgs) {
ompi_osc_rdma_progress(P2P_MODULE(win));
}
ompi_osc_rdma_flip_sendreqs(P2P_MODULE(win));
/* for each process in group, send a control message with number
of updates coming, then start all the requests */
for (i = 0 ; i < ompi_group_size(P2P_MODULE(win)->p2p_sc_group) ; ++i) {
int comm_rank = P2P_MODULE(win)->p2p_sc_remote_ranks[i];
OPAL_THREAD_ADD32(&(P2P_MODULE(win)->p2p_num_pending_out),
P2P_MODULE(win)->p2p_copy_num_pending_sendreqs[comm_rank]);
ompi_osc_rdma_control_send(P2P_MODULE(win),
P2P_MODULE(win)->p2p_sc_group->grp_proc_pointers[i],
OMPI_OSC_RDMA_HDR_COMPLETE,
P2P_MODULE(win)->p2p_copy_num_pending_sendreqs[comm_rank],
0);
}
/* try to start all the requests. We've copied everything we
need out of pending_sendreqs, so don't need the lock
here */
while (NULL !=
(item = opal_list_remove_first(&(P2P_MODULE(win)->p2p_copy_pending_sendreqs)))) {
ompi_osc_rdma_sendreq_t *req =
(ompi_osc_rdma_sendreq_t*) item;
ret = ompi_osc_rdma_sendreq_send(P2P_MODULE(win), req);
if (OMPI_SUCCESS != ret) {
opal_output_verbose(5, ompi_osc_base_output,
"complete: failure in starting sendreq (%d). Will try later.",
ret);
opal_list_append(&(P2P_MODULE(win)->p2p_copy_pending_sendreqs), item);
}
}
/* wait for all the requests */
while (0 != P2P_MODULE(win)->p2p_num_pending_out) {
ompi_osc_rdma_progress(P2P_MODULE(win));
}
/* remove WIN_POSTED from our mode */
ompi_win_remove_mode(win, OMPI_WIN_ACCESS_EPOCH | OMPI_WIN_STARTED);
OPAL_THREAD_LOCK(&(P2P_MODULE(win)->p2p_lock));
group = P2P_MODULE(win)->p2p_sc_group;
P2P_MODULE(win)->p2p_sc_group = NULL;
OPAL_THREAD_UNLOCK(&(P2P_MODULE(win)->p2p_lock));
/* BWB - do I need this? */
ompi_group_decrement_proc_count(group);
OBJ_RELEASE(group);
return ret;
}
int
ompi_osc_rdma_module_post(ompi_group_t *group,
int assert,
ompi_win_t *win)
{
int i;
OBJ_RETAIN(group);
/* BWB - do I need this? */
ompi_group_increment_proc_count(group);
OPAL_THREAD_LOCK(&(P2P_MODULE(win)->p2p_lock));
assert(NULL == P2P_MODULE(win)->p2p_pw_group);
P2P_MODULE(win)->p2p_pw_group = group;
OPAL_THREAD_UNLOCK(&(P2P_MODULE(win)->p2p_lock));
/* Set our mode to expose w/ post */
ompi_win_remove_mode(win, OMPI_WIN_FENCE);
ompi_win_set_mode(win, OMPI_WIN_EXPOSE_EPOCH | OMPI_WIN_POSTED);
/* list how many complete counters we're still waiting on */
OPAL_THREAD_ADD32(&(P2P_MODULE(win)->p2p_num_complete_msgs),
ompi_group_size(P2P_MODULE(win)->p2p_pw_group));
/* send a hello counter to everyone in group */
for (i = 0 ; i < ompi_group_size(P2P_MODULE(win)->p2p_pw_group) ; ++i) {
ompi_osc_rdma_control_send(P2P_MODULE(win),
group->grp_proc_pointers[i],
OMPI_OSC_RDMA_HDR_POST, 1, 0);
}
return OMPI_SUCCESS;
}
int
ompi_osc_rdma_module_wait(ompi_win_t *win)
{
ompi_group_t *group;
while (0 != (P2P_MODULE(win)->p2p_num_pending_in) ||
0 != (P2P_MODULE(win)->p2p_num_complete_msgs)) {
ompi_osc_rdma_progress(P2P_MODULE(win));
}
ompi_win_remove_mode(win, OMPI_WIN_EXPOSE_EPOCH | OMPI_WIN_POSTED);
OPAL_THREAD_LOCK(&(P2P_MODULE(win)->p2p_lock));
group = P2P_MODULE(win)->p2p_pw_group;
P2P_MODULE(win)->p2p_pw_group = NULL;
OPAL_THREAD_UNLOCK(&(P2P_MODULE(win)->p2p_lock));
/* BWB - do I need this? */
ompi_group_decrement_proc_count(group);
OBJ_RELEASE(group);
return OMPI_SUCCESS;
}
int
ompi_osc_rdma_module_test(ompi_win_t *win,
int *flag)
{
ompi_group_t *group;
if (0 != (P2P_MODULE(win)->p2p_num_pending_in) ||
0 != (P2P_MODULE(win)->p2p_num_complete_msgs)) {
ompi_osc_rdma_progress(P2P_MODULE(win));
if (0 != (P2P_MODULE(win)->p2p_num_pending_in) ||
0 != (P2P_MODULE(win)->p2p_num_complete_msgs)) {
*flag = 0;
return OMPI_SUCCESS;
}
}
*flag = 1;
ompi_win_remove_mode(win, OMPI_WIN_EXPOSE_EPOCH | OMPI_WIN_POSTED);
OPAL_THREAD_LOCK(&(P2P_MODULE(win)->p2p_lock));
group = P2P_MODULE(win)->p2p_pw_group;
P2P_MODULE(win)->p2p_pw_group = NULL;
OPAL_THREAD_UNLOCK(&(P2P_MODULE(win)->p2p_lock));
/* BWB - do I need this? */
ompi_group_decrement_proc_count(group);
OBJ_RELEASE(group);
return OMPI_SUCCESS;
}
struct ompi_osc_rdma_pending_lock_t {
opal_list_item_t super;
ompi_proc_t *proc;
int32_t lock_type;
};
typedef struct ompi_osc_rdma_pending_lock_t ompi_osc_rdma_pending_lock_t;
OBJ_CLASS_INSTANCE(ompi_osc_rdma_pending_lock_t, opal_list_item_t,
NULL, NULL);
int
ompi_osc_rdma_module_lock(int lock_type,
int target,
int assert,
ompi_win_t *win)
{
ompi_proc_t *proc = ompi_comm_peer_lookup( P2P_MODULE(win)->p2p_comm, target );
assert(lock_type != 0);
/* set our mode on the window */
ompi_win_set_mode(win, OMPI_WIN_ACCESS_EPOCH | OMPI_WIN_LOCK_ACCESS);
opal_output_verbose(50, ompi_osc_base_output,
"%d sending lock request to %d",
P2P_MODULE(win)->p2p_comm->c_my_rank, target);
/* generate a lock request */
ompi_osc_rdma_control_send(P2P_MODULE(win),
proc,
OMPI_OSC_RDMA_HDR_LOCK_REQ,
P2P_MODULE(win)->p2p_comm->c_my_rank,
lock_type);
/* return */
return OMPI_SUCCESS;
}
int
ompi_osc_rdma_module_unlock(int target,
ompi_win_t *win)
{
int32_t out_count;
opal_list_item_t *item;
int ret;
ompi_proc_t *proc = ompi_comm_peer_lookup( P2P_MODULE(win)->p2p_comm, target );
while (0 == P2P_MODULE(win)->p2p_lock_received_ack) {
ompi_osc_rdma_progress(P2P_MODULE(win));
}
P2P_MODULE(win)->p2p_lock_received_ack = 0;
/* start all the requests */
ompi_osc_rdma_flip_sendreqs(P2P_MODULE(win));
/* try to start all the requests. We've copied everything we need
out of pending_sendreqs, so don't need the lock here */
out_count = opal_list_get_size(&(P2P_MODULE(win)->p2p_copy_pending_sendreqs));
OPAL_THREAD_ADD32(&(P2P_MODULE(win)->p2p_num_pending_out), out_count);
while (NULL !=
(item = opal_list_remove_first(&(P2P_MODULE(win)->p2p_copy_pending_sendreqs)))) {
ompi_osc_rdma_sendreq_t *req =
(ompi_osc_rdma_sendreq_t*) item;
ret = ompi_osc_rdma_sendreq_send(P2P_MODULE(win), req);
if (OMPI_SUCCESS != ret) {
opal_output_verbose(5, ompi_osc_base_output,
"unlock: failure in starting sendreq (%d). Will try later.",
ret);
opal_list_append(&(P2P_MODULE(win)->p2p_copy_pending_sendreqs), item);
}
}
/* wait for all the requests */
while (0 != P2P_MODULE(win)->p2p_num_pending_out) {
ompi_osc_rdma_progress(P2P_MODULE(win));
}
/* send the unlock request */
opal_output_verbose(50, ompi_osc_base_output,
"%d sending unlock request to %d",
P2P_MODULE(win)->p2p_comm->c_my_rank, target);
ompi_osc_rdma_control_send(P2P_MODULE(win),
proc,
OMPI_OSC_RDMA_HDR_UNLOCK_REQ,
P2P_MODULE(win)->p2p_comm->c_my_rank,
out_count);
/* set our mode on the window */
ompi_win_set_mode(win, 0);
return OMPI_SUCCESS;
}
int
ompi_osc_rdma_passive_lock(ompi_osc_rdma_module_t *module,
int32_t origin,
int32_t lock_type)
{
bool send_ack = false;
int ret = OMPI_SUCCESS;
ompi_proc_t *proc = ompi_comm_peer_lookup( module->p2p_comm, origin );
ompi_osc_rdma_pending_lock_t *new_pending;
OPAL_THREAD_LOCK(&(module->p2p_lock));
if (lock_type == MPI_LOCK_EXCLUSIVE) {
if (module->p2p_lock_status == 0) {
module->p2p_lock_status = MPI_LOCK_EXCLUSIVE;
send_ack = true;
} else {
opal_output_verbose(50, ompi_osc_base_output,
"%d queuing lock request from %d (%d)",
module->p2p_comm->c_my_rank, origin, lock_type);
new_pending = OBJ_NEW(ompi_osc_rdma_pending_lock_t);
new_pending->proc = proc;
new_pending->lock_type = lock_type;
opal_list_append(&(module->p2p_locks_pending), &(new_pending->super));
}
} else if (lock_type == MPI_LOCK_SHARED) {
if (module->p2p_lock_status != MPI_LOCK_EXCLUSIVE) {
module->p2p_lock_status = MPI_LOCK_SHARED;
module->p2p_shared_count++;
send_ack = true;
} else {
opal_output_verbose(50, ompi_osc_base_output,
"queuing lock request from %d (%d)",
module->p2p_comm->c_my_rank, origin, lock_type);
new_pending = OBJ_NEW(ompi_osc_rdma_pending_lock_t);
new_pending->proc = proc;
new_pending->lock_type = lock_type;
opal_list_append(&(module->p2p_locks_pending), &(new_pending->super));
}
} else {
ret = OMPI_ERROR;
}
OPAL_THREAD_UNLOCK(&(module->p2p_lock));
if (send_ack) {
opal_output_verbose(50, ompi_osc_base_output,
"%d sending lock ack to %d",
module->p2p_comm->c_my_rank, origin);
ompi_osc_rdma_control_send(module, proc,
OMPI_OSC_RDMA_HDR_LOCK_REQ,
module->p2p_comm->c_my_rank,
OMPI_SUCCESS);
}
return OMPI_SUCCESS;
}
int
ompi_osc_rdma_passive_unlock(ompi_osc_rdma_module_t *module,
int32_t origin,
int32_t count)
{
ompi_osc_rdma_pending_lock_t *new_pending = NULL;
assert(module->p2p_lock_status != 0);
OPAL_THREAD_ADD32(&(module->p2p_num_pending_in), count);
while (0 != module->p2p_num_pending_in) {
ompi_osc_rdma_progress(module);
}
OPAL_THREAD_LOCK(&(module->p2p_lock));
if (module->p2p_lock_status == MPI_LOCK_EXCLUSIVE) {
module->p2p_lock_status = 0;
} else {
module->p2p_shared_count--;
if (module->p2p_shared_count == 0) {
module->p2p_lock_status = 0;
}
}
/* if we were really unlocked, see if we have more to process */
new_pending = (ompi_osc_rdma_pending_lock_t*)
opal_list_remove_first(&(module->p2p_locks_pending));
OPAL_THREAD_UNLOCK(&(module->p2p_lock));
if (NULL != new_pending) {
opal_output_verbose(50, ompi_osc_base_output,
"sending lock request to proc");
/* set lock state and generate a lock request */
module->p2p_lock_status = new_pending->lock_type;
ompi_osc_rdma_control_send(module,
new_pending->proc,
OMPI_OSC_RDMA_HDR_LOCK_REQ,
module->p2p_comm->c_my_rank,
OMPI_SUCCESS);
OBJ_DESTRUCT(new_pending);
}
return OMPI_SUCCESS;
}