static int ompi_comm_allreduce_getnextcid (ompi_comm_request_t *request)
{
ompi_comm_cid_context_t *context = (ompi_comm_cid_context_t *) request->context;
int64_t my_id = ((int64_t) ompi_comm_get_cid (context->comm) << 32 | context->pml_tag);
ompi_request_t *subreq;
bool flag;
int ret;
if (OPAL_THREAD_TRYLOCK(&ompi_cid_lock)) {
return ompi_comm_request_schedule_append (request, ompi_comm_allreduce_getnextcid, NULL, 0);
}
if (ompi_comm_cid_lowest_id < my_id) {
OPAL_THREAD_UNLOCK(&ompi_cid_lock);
return ompi_comm_request_schedule_append (request, ompi_comm_allreduce_getnextcid, NULL, 0);
}
ompi_comm_cid_lowest_id = my_id;
/**
* This is the real algorithm described in the doc
*/
flag = false;
context->nextlocal_cid = mca_pml.pml_max_contextid;
for (unsigned int i = context->start ; i < mca_pml.pml_max_contextid ; ++i) {
flag = opal_pointer_array_test_and_set_item (&ompi_mpi_communicators, i,
context->comm);
if (true == flag) {
context->nextlocal_cid = i;
break;
}
}
ret = context->allreduce_fn (&context->nextlocal_cid, &context->nextcid, 1, MPI_MAX,
context, &subreq);
if (OMPI_SUCCESS != ret) {
ompi_comm_cid_lowest_id = INT64_MAX;
OPAL_THREAD_UNLOCK(&ompi_cid_lock);
return ret;
}
if ((unsigned int) context->nextlocal_cid == mca_pml.pml_max_contextid) {
/* at least one peer ran out of CIDs */
if (flag) {
opal_pointer_array_test_and_set_item(&ompi_mpi_communicators, context->nextlocal_cid, NULL);
}
ompi_comm_cid_lowest_id = INT64_MAX;
OPAL_THREAD_UNLOCK(&ompi_cid_lock);
return OMPI_ERR_OUT_OF_RESOURCE;
}
OPAL_THREAD_UNLOCK(&ompi_cid_lock);
/* next we want to verify that the resulting commid is ok */
return ompi_comm_request_schedule_append (request, ompi_comm_checkcid, &subreq, 1);
}
static int ompi_osc_pt2pt_dt_send_complete (ompi_request_t *request)
{
ompi_datatype_t *datatype = (ompi_datatype_t *) request->req_complete_cb_data;
ompi_osc_pt2pt_module_t *module = NULL;
OMPI_DATATYPE_RELEASE(datatype);
OPAL_THREAD_LOCK(&mca_osc_pt2pt_component.lock);
(void) opal_hash_table_get_value_uint32(&mca_osc_pt2pt_component.modules,
ompi_comm_get_cid(request->req_mpi_object.comm),
(void **) &module);
OPAL_THREAD_UNLOCK(&mca_osc_pt2pt_component.lock);
assert (NULL != module);
ompi_request_free (&request);
return 1;
}
static int ompi_osc_pt2pt_dt_send_complete (ompi_request_t *request)
{
ompi_datatype_t *datatype = (ompi_datatype_t *) request->req_complete_cb_data;
ompi_osc_pt2pt_module_t *module = NULL;
OBJ_RELEASE(datatype);
OPAL_THREAD_LOCK(&mca_osc_pt2pt_component.lock);
(void) opal_hash_table_get_value_uint32(&mca_osc_pt2pt_component.modules,
ompi_comm_get_cid(request->req_mpi_object.comm),
(void **) &module);
OPAL_THREAD_UNLOCK(&mca_osc_pt2pt_component.lock);
assert (NULL != module);
/* put this request on the garbage colletion list */
osc_pt2pt_gc_add_request (module, request);
return OMPI_SUCCESS;
}
static int
component_select(struct ompi_win_t *win, void **base, size_t size, int disp_unit,
struct ompi_communicator_t *comm, struct ompi_info_t *info,
int flavor, int *model)
{
ompi_osc_portals4_module_t *module = NULL;
int ret = OMPI_ERROR;
int tmp;
ptl_md_t md;
ptl_me_t me;
char *name;
if (MPI_WIN_FLAVOR_SHARED == flavor) return OMPI_ERR_NOT_SUPPORTED;
/* create module structure */
module = (ompi_osc_portals4_module_t*)
calloc(1, sizeof(ompi_osc_portals4_module_t));
if (NULL == module) return OMPI_ERR_TEMP_OUT_OF_RESOURCE;
/* fill in the function pointer part */
memcpy(module, &ompi_osc_portals4_module_template,
sizeof(ompi_osc_base_module_t));
/* fill in our part */
if (MPI_WIN_FLAVOR_ALLOCATE == flavor) {
module->free_after = *base = malloc(size);
if (NULL == *base) goto error;
} else {
module->free_after = NULL;
}
ret = ompi_comm_dup(comm, &module->comm);
if (OMPI_SUCCESS != ret) goto error;
opal_output_verbose(1, ompi_osc_base_framework.framework_output,
"portals4 component creating window with id %d",
ompi_comm_get_cid(module->comm));
asprintf(&name, "portals4 window %d", ompi_comm_get_cid(module->comm));
ompi_win_set_name(win, name);
free(name);
/* share everyone's displacement units. Only do an allgather if
strictly necessary, since it requires O(p) state. */
tmp = disp_unit;
ret = module->comm->c_coll.coll_bcast(&tmp, 1, MPI_INT, 0,
module->comm,
module->comm->c_coll.coll_bcast_module);
if (OMPI_SUCCESS != ret) {
opal_output_verbose(1, ompi_osc_base_framework.framework_output,
"%s:%d: MPI_Bcast failed: %d\n",
__FILE__, __LINE__, ret);
goto error;
}
tmp = (tmp == disp_unit) ? 1 : 0;
ret = module->comm->c_coll.coll_allreduce(MPI_IN_PLACE, &tmp, 1, MPI_INT, MPI_LAND,
module->comm, module->comm->c_coll.coll_allreduce_module);
if (OMPI_SUCCESS != ret) goto error;
if (tmp == 1) {
module->disp_unit = disp_unit;
module->disp_units = NULL;
} else {
module->disp_unit = -1;
module->disp_units = malloc(sizeof(int) * ompi_comm_size(module->comm));
ret = module->comm->c_coll.coll_allgather(&disp_unit, 1, MPI_INT,
module->disp_units, 1, MPI_INT,
module->comm,
module->comm->c_coll.coll_allgather_module);
if (OMPI_SUCCESS != ret) goto error;
}
module->ni_h = mca_osc_portals4_component.matching_ni_h;
module->pt_idx = mca_osc_portals4_component.matching_pt_idx;
ret = PtlCTAlloc(module->ni_h, &(module->ct_h));
if (PTL_OK != ret) {
opal_output_verbose(1, ompi_osc_base_framework.framework_output,
"%s:%d: PtlCTAlloc failed: %d\n",
__FILE__, __LINE__, ret);
goto error;
}
md.start = 0;
md.length = PTL_SIZE_MAX;
md.options = PTL_MD_EVENT_SUCCESS_DISABLE | PTL_MD_EVENT_CT_REPLY | PTL_MD_EVENT_CT_ACK;
md.eq_handle = mca_osc_portals4_component.matching_eq_h;
md.ct_handle = module->ct_h;
ret = PtlMDBind(module->ni_h, &md, &module->md_h);
if (PTL_OK != ret) {
opal_output_verbose(1, ompi_osc_base_framework.framework_output,
"%s:%d: PtlMDBind failed: %d\n",
__FILE__, __LINE__, ret);
goto error;
}
md.start = 0;
md.length = PTL_SIZE_MAX;
md.options = PTL_MD_EVENT_CT_REPLY | PTL_MD_EVENT_CT_ACK;
md.eq_handle = mca_osc_portals4_component.matching_eq_h;
md.ct_handle = module->ct_h;
ret = PtlMDBind(module->ni_h, &md, &module->req_md_h);
if (PTL_OK != ret) {
opal_output_verbose(1, ompi_osc_base_framework.framework_output,
"%s:%d: PtlMDBind failed: %d\n",
__FILE__, __LINE__, ret);
goto error;
}
if (MPI_WIN_FLAVOR_DYNAMIC == flavor) {
me.start = 0;
me.length = PTL_SIZE_MAX;
} else {
me.start = *base;
me.length = size;
}
me.ct_handle = PTL_CT_NONE;
me.uid = PTL_UID_ANY;
me.options = PTL_ME_OP_PUT | PTL_ME_OP_GET | PTL_ME_NO_TRUNCATE | PTL_ME_EVENT_SUCCESS_DISABLE;
me.match_id.phys.nid = PTL_NID_ANY;
me.match_id.phys.pid = PTL_PID_ANY;
me.match_bits = module->comm->c_contextid;
me.ignore_bits = 0;
ret = PtlMEAppend(module->ni_h,
module->pt_idx,
&me,
PTL_PRIORITY_LIST,
NULL,
&module->data_me_h);
if (PTL_OK != ret) {
opal_output_verbose(1, ompi_osc_base_framework.framework_output,
"%s:%d: PtlMEAppend failed: %d\n",
__FILE__, __LINE__, ret);
goto error;
}
me.start = &module->state;
me.length = sizeof(module->state);
me.ct_handle = PTL_CT_NONE;
me.uid = PTL_UID_ANY;
me.options = PTL_ME_OP_PUT | PTL_ME_OP_GET | PTL_ME_NO_TRUNCATE | PTL_ME_EVENT_SUCCESS_DISABLE;
me.match_id.phys.nid = PTL_NID_ANY;
me.match_id.phys.pid = PTL_PID_ANY;
me.match_bits = module->comm->c_contextid | OSC_PORTALS4_MB_CONTROL;
me.ignore_bits = 0;
ret = PtlMEAppend(module->ni_h,
module->pt_idx,
&me,
PTL_PRIORITY_LIST,
NULL,
&module->control_me_h);
if (PTL_OK != ret) {
opal_output_verbose(1, ompi_osc_base_framework.framework_output,
"%s:%d: PtlMEAppend failed: %d\n",
__FILE__, __LINE__, ret);
goto error;
}
module->opcount = 0;
module->match_bits = module->comm->c_contextid;
module->atomic_max = (check_config_value_equal("accumulate_ordering", info, "none")) ?
mca_osc_portals4_component.matching_atomic_max :
MIN(mca_osc_portals4_component.matching_atomic_max,
mca_osc_portals4_component.matching_atomic_ordered_size);
module->fetch_atomic_max = (check_config_value_equal("accumulate_ordering", info, "none")) ?
mca_osc_portals4_component.matching_fetch_atomic_max :
MIN(mca_osc_portals4_component.matching_fetch_atomic_max,
mca_osc_portals4_component.matching_atomic_ordered_size);
module->zero = 0;
module->one = 1;
module->start_group = NULL;
module->post_group = NULL;
module->state.post_count = 0;
module->state.complete_count = 0;
if (check_config_value_bool("no_locks", info)) {
module->state.lock = LOCK_ILLEGAL;
} else {
module->state.lock = LOCK_UNLOCKED;
}
OBJ_CONSTRUCT(&module->outstanding_locks, opal_list_t);
module->passive_target_access_epoch = false;
#if OPAL_ASSEMBLY_ARCH == OPAL_AMD64 || OPAL_ASSEMBLY_ARCH == OPAL_IA32
*model = MPI_WIN_UNIFIED;
#else
*model = MPI_WIN_SEPARATE;
#endif
win->w_osc_module = &module->super;
PtlAtomicSync();
/* Make sure that everyone's ready to receive. */
module->comm->c_coll.coll_barrier(module->comm,
module->comm->c_coll.coll_barrier_module);
return OMPI_SUCCESS;
error:
/* BWB: FIX ME: This is all wrong... */
if (0 != module->ct_h) PtlCTFree(module->ct_h);
if (0 != module->data_me_h) PtlMEUnlink(module->data_me_h);
if (0 != module->req_md_h) PtlMDRelease(module->req_md_h);
if (0 != module->md_h) PtlMDRelease(module->md_h);
if (NULL != module->comm) ompi_comm_free(&module->comm);
if (NULL != module) free(module);
return ret;
}
int
ompi_osc_pt2pt_free(ompi_win_t *win)
{
int ret = OMPI_SUCCESS;
ompi_osc_pt2pt_module_t *module = GET_MODULE(win);
if (NULL == module) {
return OMPI_SUCCESS;
}
if (NULL != module->comm) {
opal_output_verbose(1, ompi_osc_base_framework.framework_output,
"pt2pt component destroying window with id %d",
ompi_comm_get_cid(module->comm));
/* finish with a barrier */
if (ompi_group_size(win->w_group) > 1) {
ret = module->comm->c_coll.coll_barrier(module->comm,
module->comm->c_coll.coll_barrier_module);
}
/* remove from component information */
OPAL_THREAD_SCOPED_LOCK(&mca_osc_pt2pt_component.lock,
opal_hash_table_remove_value_uint32(&mca_osc_pt2pt_component.modules,
ompi_comm_get_cid(module->comm)));
}
win->w_osc_module = NULL;
OBJ_DESTRUCT(&module->outstanding_locks);
OBJ_DESTRUCT(&module->locks_pending);
OBJ_DESTRUCT(&module->locks_pending_lock);
OBJ_DESTRUCT(&module->acc_lock);
OBJ_DESTRUCT(&module->cond);
OBJ_DESTRUCT(&module->lock);
/* it is erroneous to close a window with active operations on it so we should
* probably produce an error here instead of cleaning up */
OPAL_LIST_DESTRUCT(&module->pending_acc);
OPAL_LIST_DESTRUCT(&module->pending_posts);
osc_pt2pt_gc_clean (module);
OPAL_LIST_DESTRUCT(&module->request_gc);
OPAL_LIST_DESTRUCT(&module->buffer_gc);
OBJ_DESTRUCT(&module->gc_lock);
if (NULL != module->peers) {
for (int i = 0 ; i < ompi_comm_size (module->comm) ; ++i) {
OBJ_DESTRUCT(module->peers + i);
}
free(module->peers);
}
if (NULL != module->epoch_outgoing_frag_count) free(module->epoch_outgoing_frag_count);
if (NULL != module->frag_request) {
module->frag_request->req_complete_cb = NULL;
ompi_request_cancel (module->frag_request);
ompi_request_free (&module->frag_request);
}
if (NULL != module->comm) {
ompi_comm_free(&module->comm);
}
if (NULL != module->incoming_buffer) free (module->incoming_buffer);
if (NULL != module->free_after) free(module->free_after);
free (module);
return ret;
}
int
ompi_coll_portals4_barrier_intra(struct ompi_communicator_t *comm,
mca_coll_base_module_t *module)
{
mca_coll_portals4_module_t *portals4_module = (mca_coll_portals4_module_t*) module;
int ret, i, dim, hibit, mask, num_msgs;
int size = ompi_comm_size(comm);
int rank = ompi_comm_rank(comm);
ptl_ct_event_t ct;
ptl_handle_ct_t ct_h;
ptl_handle_me_t me_h;
ptl_me_t me;
size_t count;
ptl_match_bits_t match_bits;
ptl_handle_md_t md_h;
void *base;
ompi_coll_portals4_get_md(0, &md_h, &base);
count = opal_atomic_add_size_t(&portals4_module->barrier_count, 1);
ret = PtlCTAlloc(mca_coll_portals4_component.ni_h,
&ct_h);
if (PTL_OK != ret) {
opal_output_verbose(1, ompi_coll_base_framework.framework_output,
"%s:%d: PtlCTAlloc failed: %d\n",
__FILE__, __LINE__, ret);
return OMPI_ERR_TEMP_OUT_OF_RESOURCE;
}
COLL_PORTALS4_SET_BITS(match_bits, ompi_comm_get_cid(comm),
0, COLL_PORTALS4_BARRIER, count);
/* Build "tree" out of hypercube */
dim = comm->c_cube_dim;
hibit = opal_hibit(rank, dim);
--dim;
/* receive space */
me.start = NULL;
me.length = 0;
me.ct_handle = ct_h;
me.min_free = 0;
me.uid = mca_coll_portals4_component.uid;
me.options = PTL_ME_OP_PUT | PTL_ME_EVENT_SUCCESS_DISABLE |
PTL_ME_EVENT_LINK_DISABLE | PTL_ME_EVENT_UNLINK_DISABLE |
PTL_ME_EVENT_CT_COMM | PTL_ME_EVENT_CT_OVERFLOW;
me.match_id.phys.nid = PTL_NID_ANY;
me.match_id.phys.pid = PTL_PID_ANY;
me.match_bits = match_bits;
me.ignore_bits = 0;
ret = PtlMEAppend(mca_coll_portals4_component.ni_h,
mca_coll_portals4_component.pt_idx,
&me,
PTL_PRIORITY_LIST,
NULL,
&me_h);
if (PTL_OK != ret) {
opal_output_verbose(1, ompi_coll_base_framework.framework_output,
"%s:%d: PtlMEAppend failed: %d\n",
__FILE__, __LINE__, ret);
return OMPI_ERROR;
}
/* calculate number of children to receive from */
num_msgs = ompi_coll_portals4_get_nchildren(dim + 1, hibit, rank, size);
/* send to parent when children have sent to us */
if (rank > 0) {
int parent = rank & ~(1 << hibit);
ret = PtlTriggeredPut(md_h,
0,
0,
PTL_NO_ACK_REQ,
ompi_coll_portals4_get_peer(comm, parent),
mca_coll_portals4_component.pt_idx,
match_bits,
0,
NULL,
0,
ct_h,
num_msgs);
if (PTL_OK != ret) {
opal_output_verbose(1, ompi_coll_base_framework.framework_output,
"%s:%d: PtlTriggeredPut failed: %d\n",
__FILE__, __LINE__, ret);
return OMPI_ERROR;
}
/* we'll need to wait for the parent response before the next set of comms */
num_msgs++;
}
/* send to children when parent (or all children if root) has sent to us */
for (i = hibit + 1, mask = 1 << i; i <= dim; ++i, mask <<= 1) {
int peer = rank | mask;
if (peer < size) {
ret = PtlTriggeredPut(md_h,
0,
0,
PTL_NO_ACK_REQ,
ompi_coll_portals4_get_peer(comm, peer),
mca_coll_portals4_component.pt_idx,
match_bits,
0,
NULL,
0,
ct_h,
num_msgs);
if (PTL_OK != ret) {
opal_output_verbose(1, ompi_coll_base_framework.framework_output,
"%s:%d: PtlTriggeredPut failed: %d\n",
__FILE__, __LINE__, ret);
return OMPI_ERROR;
}
}
}
/* Wait for all incoming messages */
ret = PtlCTWait(ct_h, num_msgs, &ct);
if (PTL_OK != ret) {
opal_output_verbose(1, ompi_coll_base_framework.framework_output,
"%s:%d: PtlCTWait failed: %d\n",
__FILE__, __LINE__, ret);
return OMPI_ERROR;
}
/* cleanup */
ret = PtlMEUnlink(me_h);
if (PTL_OK != ret) {
opal_output_verbose(1, ompi_coll_base_framework.framework_output,
"%s:%d: PtlMEUnlink failed: %d\n",
__FILE__, __LINE__, ret);
return OMPI_ERROR;
}
ret = PtlCTFree(ct_h);
if (PTL_OK != ret) {
opal_output_verbose(1, ompi_coll_base_framework.framework_output,
"%s:%d: PtlCTFree failed: %d\n",
__FILE__, __LINE__, ret);
return OMPI_ERROR;
}
return OMPI_SUCCESS;
}
static int
setup_sync_handles(struct ompi_communicator_t *comm,
ompi_coll_portals4_request_t *request,
mca_coll_portals4_module_t *portals4_module)
{
int ret, line;
ptl_me_t me;
OPAL_OUTPUT((ompi_coll_base_framework.framework_output,
"coll:portals4:setup_sync_handles enter rank %d", request->u.scatter.my_rank));
/**********************************/
/* Setup Sync Handles */
/**********************************/
COLL_PORTALS4_SET_BITS(request->u.scatter.sync_match_bits, ompi_comm_get_cid(comm),
0, 1, COLL_PORTALS4_SCATTER, 0, request->u.scatter.coll_count);
OPAL_OUTPUT((ompi_coll_base_framework.framework_output,
"coll:portals4:setup_sync_handles rank(%d) sync_match_bits(0x%016lX)",
request->u.scatter.my_rank, request->u.scatter.sync_match_bits));
ret = PtlCTAlloc(mca_coll_portals4_component.ni_h,
&request->u.scatter.sync_cth);
if (PTL_OK != ret) {
ret = OMPI_ERR_TEMP_OUT_OF_RESOURCE;
line = __LINE__;
goto err_hdlr;
}
request->u.scatter.sync_mdh = mca_coll_portals4_component.zero_md_h;
me.start = NULL;
me.length = 0;
me.ct_handle = request->u.scatter.sync_cth;
me.min_free = 0;
me.uid = mca_coll_portals4_component.uid;
me.options = PTL_ME_OP_PUT | PTL_ME_EVENT_SUCCESS_DISABLE |
PTL_ME_EVENT_LINK_DISABLE | PTL_ME_EVENT_UNLINK_DISABLE |
PTL_ME_EVENT_CT_COMM | PTL_ME_EVENT_CT_OVERFLOW;
me.match_id.phys.nid = PTL_NID_ANY;
me.match_id.phys.pid = PTL_PID_ANY;
me.match_bits = request->u.scatter.sync_match_bits;
me.ignore_bits = 0;
ret = PtlMEAppend(mca_coll_portals4_component.ni_h,
mca_coll_portals4_component.pt_idx,
&me,
PTL_PRIORITY_LIST,
NULL,
&request->u.scatter.sync_meh);
if (PTL_OK != ret) {
ret = OMPI_ERROR;
line = __LINE__;
goto err_hdlr;
}
OPAL_OUTPUT((ompi_coll_base_framework.framework_output,
"coll:portals4:setup_sync_handles exit rank %d", request->u.scatter.my_rank));
return OMPI_SUCCESS;
err_hdlr:
opal_output(ompi_coll_base_framework.framework_output,
"%s:%4d:%4d\tError occurred ret=%d, rank %2d",
__FILE__, __LINE__, line, ret, request->u.scatter.my_rank);
return ret;
}
int
ompi_osc_rdma_module_free(ompi_win_t *win)
{
int ret = OMPI_SUCCESS;
int tmp, i;
ompi_osc_rdma_module_t *module = GET_MODULE(win);
opal_output_verbose(1, ompi_osc_base_output,
"rdma component destroying window with id %d",
ompi_comm_get_cid(module->m_comm));
/* finish with a barrier */
if (ompi_group_size(win->w_group) > 1) {
ret = module->m_comm->c_coll.coll_barrier(module->m_comm,
module->m_comm->c_coll.coll_barrier_module);
}
/* remove from component information */
OPAL_THREAD_LOCK(&mca_osc_rdma_component.c_lock);
tmp = opal_hash_table_remove_value_uint32(&mca_osc_rdma_component.c_modules,
ompi_comm_get_cid(module->m_comm));
/* only take the output of hast_table_remove if there wasn't already an error */
ret = (ret != OMPI_SUCCESS) ? ret : tmp;
if (0 == opal_hash_table_get_size(&mca_osc_rdma_component.c_modules)) {
#if OPAL_ENABLE_PROGRESS_THREADS
void *foo;
mca_osc_rdma_component.c_thread_run = false;
opal_condition_broadcast(&ompi_request_cond);
opal_thread_join(&mca_osc_rdma_component.c_thread, &foo);
#else
opal_progress_unregister(ompi_osc_rdma_component_progress);
#endif
}
OPAL_THREAD_UNLOCK(&mca_osc_rdma_component.c_lock);
win->w_osc_module = NULL;
OBJ_DESTRUCT(&module->m_unlocks_pending);
OBJ_DESTRUCT(&module->m_locks_pending);
OBJ_DESTRUCT(&module->m_queued_sendreqs);
OBJ_DESTRUCT(&module->m_copy_pending_sendreqs);
OBJ_DESTRUCT(&module->m_pending_sendreqs);
OBJ_DESTRUCT(&module->m_acc_lock);
OBJ_DESTRUCT(&module->m_cond);
OBJ_DESTRUCT(&module->m_lock);
if (NULL != module->m_sc_remote_ranks) {
free(module->m_sc_remote_ranks);
}
if (NULL != module->m_sc_remote_active_ranks) {
free(module->m_sc_remote_active_ranks);
}
if (NULL != module->m_pending_buffers) {
free(module->m_pending_buffers);
}
if (NULL != module->m_fence_coll_counts) {
free(module->m_fence_coll_counts);
}
if (NULL != module->m_copy_num_pending_sendreqs) {
free(module->m_copy_num_pending_sendreqs);
}
if (NULL != module->m_num_pending_sendreqs) {
free(module->m_num_pending_sendreqs);
}
if (NULL != module->m_peer_info) {
for (i = 0 ; i < ompi_comm_size(module->m_comm) ; ++i) {
ompi_osc_rdma_peer_info_free(&module->m_peer_info[i]);
}
free(module->m_peer_info);
}
if (NULL != module->m_comm) ompi_comm_free(&module->m_comm);
if (NULL != module) free(module);
return ret;
}
static int component_select(struct ompi_win_t *win, void **base, size_t size, int disp_unit,
struct ompi_communicator_t *comm, struct opal_info_t *info,
int flavor, int *model) {
ompi_osc_ucx_module_t *module = NULL;
char *name = NULL;
long values[2];
int ret = OMPI_SUCCESS;
ucs_status_t status;
int i, comm_size = ompi_comm_size(comm);
int is_eps_ready;
bool eps_created = false, worker_created = false;
ucp_address_t *my_addr = NULL;
size_t my_addr_len;
char *recv_buf = NULL;
void *rkey_buffer = NULL, *state_rkey_buffer = NULL;
size_t rkey_buffer_size, state_rkey_buffer_size;
void *state_base = NULL;
void * my_info = NULL;
size_t my_info_len;
int disps[comm_size];
int rkey_sizes[comm_size];
/* the osc/sm component is the exclusive provider for support for
* shared memory windows */
if (flavor == MPI_WIN_FLAVOR_SHARED) {
return OMPI_ERR_NOT_SUPPORTED;
}
/* if UCP worker has never been initialized before, init it first */
if (mca_osc_ucx_component.ucp_worker == NULL) {
ucp_worker_params_t worker_params;
ucp_worker_attr_t worker_attr;
memset(&worker_params, 0, sizeof(ucp_worker_h));
worker_params.field_mask = UCP_WORKER_PARAM_FIELD_THREAD_MODE;
worker_params.thread_mode = (mca_osc_ucx_component.enable_mpi_threads == true)
? UCS_THREAD_MODE_MULTI : UCS_THREAD_MODE_SINGLE;
status = ucp_worker_create(mca_osc_ucx_component.ucp_context, &worker_params,
&(mca_osc_ucx_component.ucp_worker));
if (UCS_OK != status) {
opal_output_verbose(1, ompi_osc_base_framework.framework_output,
"%s:%d: ucp_worker_create failed: %d\n",
__FILE__, __LINE__, status);
ret = OMPI_ERROR;
goto error;
}
/* query UCP worker attributes */
worker_attr.field_mask = UCP_WORKER_ATTR_FIELD_THREAD_MODE;
status = ucp_worker_query(mca_osc_ucx_component.ucp_worker, &worker_attr);
if (UCS_OK != status) {
opal_output_verbose(1, ompi_osc_base_framework.framework_output,
"%s:%d: ucp_worker_query failed: %d\n",
__FILE__, __LINE__, status);
ret = OMPI_ERROR;
goto error;
}
if (mca_osc_ucx_component.enable_mpi_threads == true &&
worker_attr.thread_mode != UCS_THREAD_MODE_MULTI) {
opal_output_verbose(1, ompi_osc_base_framework.framework_output,
"%s:%d: ucx does not support multithreading\n",
__FILE__, __LINE__);
ret = OMPI_ERROR;
goto error;
}
worker_created = true;
}
/* create module structure */
module = (ompi_osc_ucx_module_t *)calloc(1, sizeof(ompi_osc_ucx_module_t));
if (module == NULL) {
ret = OMPI_ERR_TEMP_OUT_OF_RESOURCE;
goto error;
}
/* fill in the function pointer part */
memcpy(module, &ompi_osc_ucx_module_template, sizeof(ompi_osc_base_module_t));
ret = ompi_comm_dup(comm, &module->comm);
if (ret != OMPI_SUCCESS) {
goto error;
}
asprintf(&name, "ucx window %d", ompi_comm_get_cid(module->comm));
ompi_win_set_name(win, name);
free(name);
/* share everyone's displacement units. Only do an allgather if
strictly necessary, since it requires O(p) state. */
values[0] = disp_unit;
values[1] = -disp_unit;
ret = module->comm->c_coll->coll_allreduce(MPI_IN_PLACE, values, 2, MPI_LONG,
MPI_MIN, module->comm,
module->comm->c_coll->coll_allreduce_module);
if (OMPI_SUCCESS != ret) {
goto error;
}
if (values[0] == -values[1]) { /* everyone has the same disp_unit, we do not need O(p) space */
module->disp_unit = disp_unit;
} else { /* different disp_unit sizes, allocate O(p) space to store them */
module->disp_unit = -1;
module->disp_units = calloc(comm_size, sizeof(int));
if (module->disp_units == NULL) {
ret = OMPI_ERR_TEMP_OUT_OF_RESOURCE;
goto error;
}
ret = module->comm->c_coll->coll_allgather(&disp_unit, 1, MPI_INT,
module->disp_units, 1, MPI_INT,
module->comm,
module->comm->c_coll->coll_allgather_module);
if (OMPI_SUCCESS != ret) {
goto error;
}
}
/* exchange endpoints if necessary */
is_eps_ready = 1;
for (i = 0; i < comm_size; i++) {
if (OSC_UCX_GET_EP(module->comm, i) == NULL) {
is_eps_ready = 0;
break;
}
}
ret = module->comm->c_coll->coll_allreduce(MPI_IN_PLACE, &is_eps_ready, 1, MPI_INT,
MPI_LAND,
module->comm,
module->comm->c_coll->coll_allreduce_module);
if (OMPI_SUCCESS != ret) {
goto error;
}
if (!is_eps_ready) {
status = ucp_worker_get_address(mca_osc_ucx_component.ucp_worker,
&my_addr, &my_addr_len);
if (status != UCS_OK) {
opal_output_verbose(1, ompi_osc_base_framework.framework_output,
"%s:%d: ucp_worker_get_address failed: %d\n",
__FILE__, __LINE__, status);
ret = OMPI_ERROR;
goto error;
}
ret = allgather_len_and_info(my_addr, (int)my_addr_len,
&recv_buf, disps, module->comm);
if (ret != OMPI_SUCCESS) {
goto error;
}
for (i = 0; i < comm_size; i++) {
if (OSC_UCX_GET_EP(module->comm, i) == NULL) {
ucp_ep_params_t ep_params;
ucp_ep_h ep;
memset(&ep_params, 0, sizeof(ucp_ep_params_t));
ep_params.field_mask = UCP_EP_PARAM_FIELD_REMOTE_ADDRESS;
ep_params.address = (ucp_address_t *)&(recv_buf[disps[i]]);
status = ucp_ep_create(mca_osc_ucx_component.ucp_worker, &ep_params, &ep);
if (status != UCS_OK) {
opal_output_verbose(1, ompi_osc_base_framework.framework_output,
"%s:%d: ucp_ep_create failed: %d\n",
__FILE__, __LINE__, status);
ret = OMPI_ERROR;
goto error;
}
ompi_comm_peer_lookup(module->comm, i)->proc_endpoints[OMPI_PROC_ENDPOINT_TAG_UCX] = ep;
}
}
ucp_worker_release_address(mca_osc_ucx_component.ucp_worker, my_addr);
my_addr = NULL;
free(recv_buf);
recv_buf = NULL;
eps_created = true;
}
ret = mem_map(base, size, &(module->memh), module, flavor);
if (ret != OMPI_SUCCESS) {
goto error;
}
state_base = (void *)&(module->state);
ret = mem_map(&state_base, sizeof(ompi_osc_ucx_state_t), &(module->state_memh),
module, MPI_WIN_FLAVOR_CREATE);
if (ret != OMPI_SUCCESS) {
goto error;
}
module->win_info_array = calloc(comm_size, sizeof(ompi_osc_ucx_win_info_t));
if (module->win_info_array == NULL) {
ret = OMPI_ERR_TEMP_OUT_OF_RESOURCE;
goto error;
}
module->state_info_array = calloc(comm_size, sizeof(ompi_osc_ucx_win_info_t));
if (module->state_info_array == NULL) {
ret = OMPI_ERR_TEMP_OUT_OF_RESOURCE;
goto error;
}
status = ucp_rkey_pack(mca_osc_ucx_component.ucp_context, module->memh,
&rkey_buffer, &rkey_buffer_size);
if (status != UCS_OK) {
opal_output_verbose(1, ompi_osc_base_framework.framework_output,
"%s:%d: ucp_rkey_pack failed: %d\n",
__FILE__, __LINE__, status);
ret = OMPI_ERROR;
goto error;
}
status = ucp_rkey_pack(mca_osc_ucx_component.ucp_context, module->state_memh,
&state_rkey_buffer, &state_rkey_buffer_size);
if (status != UCS_OK) {
opal_output_verbose(1, ompi_osc_base_framework.framework_output,
"%s:%d: ucp_rkey_pack failed: %d\n",
__FILE__, __LINE__, status);
ret = OMPI_ERROR;
goto error;
}
my_info_len = 2 * sizeof(uint64_t) + rkey_buffer_size + state_rkey_buffer_size;
my_info = malloc(my_info_len);
if (my_info == NULL) {
ret = OMPI_ERR_TEMP_OUT_OF_RESOURCE;
goto error;
}
memcpy(my_info, base, sizeof(uint64_t));
memcpy((void *)((char *)my_info + sizeof(uint64_t)), &state_base, sizeof(uint64_t));
memcpy((void *)((char *)my_info + 2 * sizeof(uint64_t)), rkey_buffer, rkey_buffer_size);
memcpy((void *)((char *)my_info + 2 * sizeof(uint64_t) + rkey_buffer_size),
state_rkey_buffer, state_rkey_buffer_size);
ret = allgather_len_and_info(my_info, (int)my_info_len, &recv_buf, disps, module->comm);
if (ret != OMPI_SUCCESS) {
goto error;
}
ret = comm->c_coll->coll_allgather((void *)&rkey_buffer_size, 1, MPI_INT,
rkey_sizes, 1, MPI_INT, comm,
comm->c_coll->coll_allgather_module);
if (OMPI_SUCCESS != ret) {
goto error;
}
for (i = 0; i < comm_size; i++) {
ucp_ep_h ep = OSC_UCX_GET_EP(module->comm, i);
assert(ep != NULL);
memcpy(&(module->win_info_array[i]).addr, &recv_buf[disps[i]], sizeof(uint64_t));
memcpy(&(module->state_info_array[i]).addr, &recv_buf[disps[i] + sizeof(uint64_t)],
sizeof(uint64_t));
status = ucp_ep_rkey_unpack(ep, &(recv_buf[disps[i] + 2 * sizeof(uint64_t)]),
&((module->win_info_array[i]).rkey));
if (status != UCS_OK) {
opal_output_verbose(1, ompi_osc_base_framework.framework_output,
"%s:%d: ucp_ep_rkey_unpack failed: %d\n",
__FILE__, __LINE__, status);
ret = OMPI_ERROR;
goto error;
}
status = ucp_ep_rkey_unpack(ep, &(recv_buf[disps[i] + 2 * sizeof(uint64_t) + rkey_sizes[i]]),
&((module->state_info_array[i]).rkey));
if (status != UCS_OK) {
opal_output_verbose(1, ompi_osc_base_framework.framework_output,
"%s:%d: ucp_ep_rkey_unpack failed: %d\n",
__FILE__, __LINE__, status);
ret = OMPI_ERROR;
goto error;
}
}
free(my_info);
free(recv_buf);
ucp_rkey_buffer_release(rkey_buffer);
ucp_rkey_buffer_release(state_rkey_buffer);
module->state.lock = TARGET_LOCK_UNLOCKED;
module->state.post_index = 0;
memset((void *)module->state.post_state, 0, sizeof(uint64_t) * OMPI_OSC_UCX_POST_PEER_MAX);
module->state.complete_count = 0;
module->state.req_flag = 0;
module->state.acc_lock = TARGET_LOCK_UNLOCKED;
module->epoch_type.access = NONE_EPOCH;
module->epoch_type.exposure = NONE_EPOCH;
module->lock_count = 0;
module->post_count = 0;
module->start_group = NULL;
module->post_group = NULL;
OBJ_CONSTRUCT(&module->outstanding_locks, opal_hash_table_t);
OBJ_CONSTRUCT(&module->pending_posts, opal_list_t);
module->global_ops_num = 0;
module->per_target_ops_nums = calloc(comm_size, sizeof(int));
module->start_grp_ranks = NULL;
module->lock_all_is_nocheck = false;
ret = opal_hash_table_init(&module->outstanding_locks, comm_size);
if (ret != OPAL_SUCCESS) {
goto error;
}
win->w_osc_module = &module->super;
/* sync with everyone */
ret = module->comm->c_coll->coll_barrier(module->comm,
module->comm->c_coll->coll_barrier_module);
if (ret != OMPI_SUCCESS) {
goto error;
}
return ret;
error:
if (my_addr) ucp_worker_release_address(mca_osc_ucx_component.ucp_worker, my_addr);
if (recv_buf) free(recv_buf);
if (my_info) free(my_info);
for (i = 0; i < comm_size; i++) {
if ((module->win_info_array[i]).rkey != NULL) {
ucp_rkey_destroy((module->win_info_array[i]).rkey);
}
if ((module->state_info_array[i]).rkey != NULL) {
ucp_rkey_destroy((module->state_info_array[i]).rkey);
}
}
if (rkey_buffer) ucp_rkey_buffer_release(rkey_buffer);
if (state_rkey_buffer) ucp_rkey_buffer_release(state_rkey_buffer);
if (module->win_info_array) free(module->win_info_array);
if (module->state_info_array) free(module->state_info_array);
if (module->disp_units) free(module->disp_units);
if (module->comm) ompi_comm_free(&module->comm);
if (module->per_target_ops_nums) free(module->per_target_ops_nums);
if (eps_created) {
for (i = 0; i < comm_size; i++) {
ucp_ep_h ep = OSC_UCX_GET_EP(module->comm, i);
ucp_ep_destroy(ep);
}
}
if (worker_created) ucp_worker_destroy(mca_osc_ucx_component.ucp_worker);
if (module) free(module);
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;
}
