ucs_status_t ucp_rma_get(ucp_ep_h ep, void *buffer, size_t length,
uint64_t remote_addr, ucp_rkey_h rkey)
{
uct_completion_t comp;
ucs_status_t status;
uct_rkey_t uct_rkey;
size_t frag_length;
if (length == 0) {
return UCS_OK;
}
uct_rkey = UCP_RMA_RKEY_LOOKUP(ep, rkey);
comp.count = 1;
for (;;) {
/* Push out all fragments, and request completion only for the last
* fragment.
*/
frag_length = ucs_min(ep->config.max_bcopy_get, length);
status = uct_ep_get_bcopy(ep->uct.ep, (uct_unpack_callback_t)memcpy,
(void*)buffer, frag_length, remote_addr,
uct_rkey, &comp);
if (ucs_likely(status == UCS_OK)) {
goto posted;
} else if (status == UCS_INPROGRESS) {
++comp.count;
goto posted;
} else if (status == UCS_ERR_NO_RESOURCE) {
goto retry;
} else {
return status;
}
posted:
length -= frag_length;
if (length == 0) {
break;
}
buffer += frag_length;
remote_addr += frag_length;
retry:
ucp_worker_progress(ep->worker);
}
/* coverity[loop_condition] */
while (comp.count > 1) {
ucp_worker_progress(ep->worker);
};
return UCS_OK;
}
static int progress_callback(void) {
if (mca_osc_ucx_component.ucp_worker != NULL &&
mca_osc_ucx_component.num_incomplete_req_ops > 0) {
ucp_worker_progress(mca_osc_ucx_component.ucp_worker);
}
return 0;
}
static UCS_F_NOINLINE
ucs_status_ptr_t ucp_tag_send_slow(ucp_ep_h ep, const void *buffer, size_t count,
uintptr_t datatype, ucp_tag_t tag,
ucp_send_callback_t cb)
{
ucp_request_t *req;
ucs_status_t status;
req = ucs_mpool_get_inline(&ep->worker->req_mp);
if (req == NULL) {
return UCS_STATUS_PTR(UCS_ERR_NO_MEMORY);
}
VALGRIND_MAKE_MEM_DEFINED(req + 1, ep->worker->context->config.request.size);
req->flags = 0;
req->cb.send = cb;
status = ucp_tag_send_start_req(ep, buffer, count, datatype, tag, req);
if (status != UCS_OK) {
return UCS_STATUS_PTR(status); /* UCS_OK also goes here */
}
if (!(req->flags & UCP_REQUEST_FLAG_COMPLETED)) {
ucp_ep_add_pending(ep, ep->uct_ep, req);
ucp_worker_progress(ep->worker);
}
ucs_trace_req("send_nb returning request %p", req);
return req + 1;
}
static void ucp_perf_test_destroy_eps(ucx_perf_context_t* perf,
unsigned group_size)
{
ucs_status_ptr_t *reqs;
ucp_tag_recv_info_t info;
ucs_status_t status;
unsigned i;
reqs = calloc(sizeof(*reqs), group_size);
for (i = 0; i < group_size; ++i) {
if (perf->ucp.peers[i].rkey != NULL) {
ucp_rkey_destroy(perf->ucp.peers[i].rkey);
}
if (perf->ucp.peers[i].ep != NULL) {
reqs[i] = ucp_disconnect_nb(perf->ucp.peers[i].ep);
}
}
for (i = 0; i < group_size; ++i) {
if (!UCS_PTR_IS_PTR(reqs[i])) {
continue;
}
do {
ucp_worker_progress(perf->ucp.worker);
status = ucp_request_test(reqs[i], &info);
} while (status == UCS_INPROGRESS);
ucp_request_release(reqs[i]);
}
free(reqs);
free(perf->ucp.peers);
}
int mca_pml_ucx_send(const void *buf, size_t count, ompi_datatype_t *datatype, int dst,
int tag, mca_pml_base_send_mode_t mode,
struct ompi_communicator_t* comm)
{
ompi_request_t *req;
ucp_ep_h ep;
PML_UCX_TRACE_SEND("%s", buf, count, datatype, dst, tag, mode, comm, "send");
/* TODO special care to sync/buffered send */
ep = mca_pml_ucx_get_ep(comm, dst);
if (OPAL_UNLIKELY(NULL == ep)) {
PML_UCX_ERROR("Failed to get ep for rank %d", dst);
return OMPI_ERROR;
}
req = (ompi_request_t*)ucp_tag_send_nb(ep, buf, count,
mca_pml_ucx_get_datatype(datatype),
PML_UCX_MAKE_SEND_TAG(tag, comm),
mca_pml_ucx_send_completion);
if (OPAL_LIKELY(req == NULL)) {
return OMPI_SUCCESS;
} else if (!UCS_PTR_IS_ERR(req)) {
PML_UCX_VERBOSE(8, "got request %p", (void*)req);
ucp_worker_progress(ompi_pml_ucx.ucp_worker);
ompi_request_wait(&req, MPI_STATUS_IGNORE);
return OMPI_SUCCESS;
} else {
PML_UCX_ERROR("ucx send failed: %s", ucs_status_string(UCS_PTR_STATUS(req)));
return OMPI_ERROR;
}
}
int mca_pml_ucx_recv(void *buf, size_t count, ompi_datatype_t *datatype, int src,
int tag, struct ompi_communicator_t* comm,
ompi_status_public_t* mpi_status)
{
ucp_tag_t ucp_tag, ucp_tag_mask;
ompi_request_t *req;
PML_UCX_TRACE_RECV("%s", buf, count, datatype, src, tag, comm, "recv");
PML_UCX_MAKE_RECV_TAG(ucp_tag, ucp_tag_mask, tag, src, comm);
req = (ompi_request_t*)ucp_tag_recv_nb(ompi_pml_ucx.ucp_worker, buf, count,
mca_pml_ucx_get_datatype(datatype),
ucp_tag, ucp_tag_mask,
mca_pml_ucx_blocking_recv_completion);
if (UCS_PTR_IS_ERR(req)) {
PML_UCX_ERROR("ucx recv failed: %s", ucs_status_string(UCS_PTR_STATUS(req)));
return OMPI_ERROR;
}
ucp_worker_progress(ompi_pml_ucx.ucp_worker);
while ( !REQUEST_COMPLETE(req) ) {
opal_progress();
}
if (mpi_status != MPI_STATUS_IGNORE) {
*mpi_status = req->req_status;
}
req->req_complete = REQUEST_PENDING;
ucp_request_release(req);
return OMPI_SUCCESS;
}
OPAL_DECLSPEC void opal_common_ucx_mca_pmix_fence(ucp_worker_h worker)
{
volatile int fenced = 0;
opal_pmix.fence_nb(NULL, 0, opal_common_ucx_mca_fence_complete_cb, (void*)&fenced);
while (!fenced) {
ucp_worker_progress(worker);
}
}
ucs_status_t ucp_put(ucp_ep_h ep, const void *buffer, size_t length,
uint64_t remote_addr, ucp_rkey_h rkey)
{
ucp_ep_rma_config_t *rma_config;
ucs_status_t status;
uct_rkey_t uct_rkey;
size_t frag_length;
ssize_t packed_len;
ucp_lane_index_t lane;
UCP_RMA_CHECK_PARAMS(buffer, length);
/* Loop until all message has been sent.
* We re-check the configuration on every iteration, because it can be
* changed by transport switch.
*/
for (;;) {
UCP_EP_RESOLVE_RKEY_RMA(ep, rkey, lane, uct_rkey, rma_config);
if (length <= rma_config->max_put_short) {
status = uct_ep_put_short(ep->uct_eps[lane], buffer, length,
remote_addr, uct_rkey);
if (ucs_likely(status != UCS_ERR_NO_RESOURCE)) {
break;
}
} else {
if (length <= ucp_ep_config(ep)->bcopy_thresh) {
frag_length = ucs_min(length, rma_config->max_put_short);
status = uct_ep_put_short(ep->uct_eps[lane], buffer, frag_length,
remote_addr, uct_rkey);
} else {
ucp_memcpy_pack_context_t pack_ctx;
pack_ctx.src = buffer;
pack_ctx.length = frag_length =
ucs_min(length, rma_config->max_put_bcopy);
packed_len = uct_ep_put_bcopy(ep->uct_eps[lane], ucp_memcpy_pack,
&pack_ctx, remote_addr, uct_rkey);
status = (packed_len > 0) ? UCS_OK : (ucs_status_t)packed_len;
}
if (ucs_likely(status == UCS_OK)) {
length -= frag_length;
if (length == 0) {
break;
}
buffer += frag_length;
remote_addr += frag_length;
} else if (status != UCS_ERR_NO_RESOURCE) {
break;
}
}
ucp_worker_progress(ep->worker);
}
return status;
}
ucs_status_t ucp_rma_put(ucp_ep_h ep, const void *buffer, size_t length,
uint64_t remote_addr, ucp_rkey_h rkey)
{
ucs_status_t status;
uct_rkey_t uct_rkey;
size_t frag_length;
if (length == 0) {
return UCS_OK;
}
if (ENABLE_PARAMS_CHECK && (buffer == NULL)) {
return UCS_ERR_INVALID_PARAM;
}
uct_rkey = UCP_RMA_RKEY_LOOKUP(ep, rkey);
/* Loop until all message has been sent.
* We re-check the configuration on every iteration, because it can be
* changed by transport switch.
*/
for (;;) {
if (length <= ep->config.max_short_put) {
status = uct_ep_put_short(ep->uct.ep, buffer, length, remote_addr,
uct_rkey);
if (ucs_likely(status != UCS_ERR_NO_RESOURCE)) {
break;
}
} else {
if (length <= ep->worker->context->config.bcopy_thresh) {
frag_length = ucs_min(length, ep->config.max_short_put);
status = uct_ep_put_short(ep->uct.ep, buffer, frag_length, remote_addr,
uct_rkey);
} else {
frag_length = ucs_min(length, ep->config.max_bcopy_put);
status = uct_ep_put_bcopy(ep->uct.ep, (uct_pack_callback_t)memcpy,
(void*)buffer, frag_length, remote_addr,
uct_rkey);
}
if (ucs_likely(status == UCS_OK)) {
length -= frag_length;
if (length == 0) {
break;
}
buffer += frag_length;
remote_addr += frag_length;
} else if (status != UCS_ERR_NO_RESOURCE) {
break;
}
}
ucp_worker_progress(ep->worker);
}
return status;
}
ucs_status_t ucp_worker_flush(ucp_worker_h worker)
{
unsigned rsc_index;
while (worker->stub_pend_count > 0) {
ucp_worker_progress(worker);
}
/* TODO flush in parallel */
for (rsc_index = 0; rsc_index < worker->context->num_tls; ++rsc_index) {
if (worker->ifaces[rsc_index] == NULL) {
continue;
}
while (uct_iface_flush(worker->ifaces[rsc_index], 0, NULL) != UCS_OK) {
ucp_worker_progress(worker);
}
}
return UCS_OK;
}
int mca_pml_ucx_progress(void)
{
static int inprogress = 0;
if (inprogress != 0) {
return 0;
}
++inprogress;
ucp_worker_progress(ompi_pml_ucx.ucp_worker);
--inprogress;
return OMPI_SUCCESS;
}
void ucp_ep_add_pending(ucp_ep_h ep, uct_ep_h uct_ep, ucp_request_t *req,
int progress)
{
ucs_status_t status;
req->send.ep = ep;
status = ucp_ep_add_pending_uct(ep, uct_ep, &req->send.uct);
while (status != UCS_OK) {
if (progress) {
ucp_worker_progress(ep->worker);
}
status = ucp_ep_add_pending_uct(ep, uct_ep, &req->send.uct);
}
}
void ucp_ep_wireup_stop(ucp_ep_h ep)
{
ucp_worker_h worker = ep->worker;
ucs_trace_func("ep=%p", ep);
if (ep->uct.next_ep != NULL) {
while (uct_ep_flush(ep->uct.next_ep) != UCS_OK) {
ucp_worker_progress(ep->worker);
}
uct_ep_destroy(ep->uct.next_ep);
}
if (ep->wireup_ep != NULL) {
while (uct_ep_flush(ep->wireup_ep) != UCS_OK) {
ucp_worker_progress(ep->worker);
}
uct_ep_destroy(ep->wireup_ep);
}
UCS_ASYNC_BLOCK(&worker->async);
sglib_hashed_ucp_ep_t_delete(worker->ep_hash, ep);
UCS_ASYNC_UNBLOCK(&worker->async);
}
int ompi_osc_ucx_free(struct ompi_win_t *win) {
ompi_osc_ucx_module_t *module = (ompi_osc_ucx_module_t*) win->w_osc_module;
int i, ret = OMPI_SUCCESS;
if ((module->epoch_type.access != NONE_EPOCH && module->epoch_type.access != FENCE_EPOCH)
|| module->epoch_type.exposure != NONE_EPOCH) {
ret = OMPI_ERR_RMA_SYNC;
}
if (module->start_group != NULL || module->post_group != NULL) {
ret = OMPI_ERR_RMA_SYNC;
}
assert(module->global_ops_num == 0);
assert(module->lock_count == 0);
assert(opal_list_is_empty(&module->pending_posts) == true);
OBJ_DESTRUCT(&module->outstanding_locks);
OBJ_DESTRUCT(&module->pending_posts);
while (module->state.lock != TARGET_LOCK_UNLOCKED) {
/* not sure if this is required */
ucp_worker_progress(mca_osc_ucx_component.ucp_worker);
}
ret = module->comm->c_coll->coll_barrier(module->comm,
module->comm->c_coll->coll_barrier_module);
for (i = 0; i < ompi_comm_size(module->comm); i++) {
ucp_rkey_destroy((module->win_info_array[i]).rkey);
ucp_rkey_destroy((module->state_info_array[i]).rkey);
}
free(module->win_info_array);
free(module->state_info_array);
free(module->per_target_ops_nums);
ucp_mem_unmap(mca_osc_ucx_component.ucp_context, module->memh);
ucp_mem_unmap(mca_osc_ucx_component.ucp_context, module->state_memh);
if (module->disp_units) free(module->disp_units);
ompi_comm_free(&module->comm);
free(module);
return ret;
}
ucs_status_t ucp_ep_flush(ucp_ep_h ep)
{
ucp_lane_index_t lane;
ucs_status_t status;
for (lane = 0; lane < ucp_ep_num_lanes(ep); ++lane) {
for (;;) {
status = uct_ep_flush(ep->uct_eps[lane], 0, NULL);
if (status == UCS_OK) {
break;
} else if ((status != UCS_INPROGRESS) && (status != UCS_ERR_NO_RESOURCE)) {
return status;
}
ucp_worker_progress(ep->worker);
}
}
return UCS_OK;
}
ucs_status_t ucp_tag_send(ucp_ep_h ep, const void *buffer, size_t length,
ucp_tag_t tag)
{
ucp_worker_h worker = ep->worker;
ucs_status_t status;
retry:
if (ucs_likely(length < ep->config.max_short_tag)) {
UCS_STATIC_ASSERT(sizeof(ucp_tag_t) == sizeof(uint64_t));
status = uct_ep_am_short(ep->uct_ep, UCP_AM_ID_EAGER_ONLY, tag,
buffer, length);
if (status == UCS_ERR_NO_RESOURCE) {
ucp_worker_progress(worker);
goto retry;
}
return status;
}
ucs_fatal("unsupported");
}
void ucp_ep_destroy(ucp_ep_h ep)
{
ucp_worker_h worker = ep->worker;
ucs_status_ptr_t *request;
ucs_status_t status;
request = ucp_disconnect_nb(ep);
if (request == NULL) {
return;
} else if (UCS_PTR_IS_ERR(request)) {
ucs_warn("disconnect failed: %s",
ucs_status_string(UCS_PTR_STATUS(request)));
return;
} else {
do {
ucp_worker_progress(worker);
status = ucp_request_test(request, NULL);
} while (status == UCS_INPROGRESS);
ucp_request_release(request);
}
}
int progress(int server, int msg_size)
{
int flag = 1;
int recvd = 0, sent = 0;
struct ucx_context *sreq = NULL, *rreq = NULL;
struct epoll_event events[2];
int ret;
if( server ){
sreq = launch_send(msg_size);
if( !sreq ){
/* inline send */
sent = 1;
}
}
progress_calls++;
while (flag) {
int i;
ucs_status_t status;
status = ucp_worker_arm(ucp_worker);
if (status == UCS_ERR_BUSY) { /* some events are arrived already */
struct ucx_context *tmp;
ucp_worker_progress(ucp_worker);
progress_count++;
tmp = launch_recv();
rreq = (rreq) ? rreq : tmp;
ret = 0;
goto progress;
} else if ( UCS_OK != status ){
abort();
}
ret = epoll_wait(epoll_fd_local, events, 2, -1);
if ( 0 > ret ) {
if (errno == EINTR) {
continue;
} else {
abort();
}
}
ucp_worker_progress(ucp_worker);
progress_count++;
for(i=0; i<ret; i++){
if( events[i].data.fd == epoll_fd){
struct ucx_context *tmp;
tmp = launch_recv();
rreq = (rreq) ? rreq : tmp;
continue;
}
if( events[i].data.fd == signal_pipe[0]){
char buf;
read(signal_pipe[0], &buf, sizeof(buf));
continue;
}
}
progress:
if( sreq ){
if( sreq->completed ){
if( !same_buf ){
free(sreq->buf);
}
sreq->completed = 0;
ucp_request_release(sreq);
sreq = NULL;
sent = 1;
}
}
if( rreq ){
if( rreq->completed ){
if( !same_buf ) {
free(rreq->buf);
}
rreq->completed = 0;
ucp_request_release(rreq);
rreq = NULL;
recvd = 1;
if( !server ){
sreq = launch_send(msg_size);
if( !sreq ){
/* inline send */
sent = 1;
}
}
}
}
if( recvd && sent ){
flag = 0;
}
if( sreq || rreq ){
activate_progress();
}
}
}
int spml_ucx_progress(void)
{
ucp_worker_progress(mca_spml_ucx.ucp_worker);
return 1;
}
void print_ucp_info(int print_opts, ucs_config_print_flags_t print_flags,
uint64_t ctx_features, const ucp_ep_params_t *base_ep_params,
size_t estimated_num_eps, unsigned dev_type_bitmap,
const char *mem_size)
{
ucp_config_t *config;
ucs_status_t status;
ucs_status_ptr_t status_ptr;
ucp_context_h context;
ucp_worker_h worker;
ucp_params_t params;
ucp_worker_params_t worker_params;
ucp_ep_params_t ep_params;
ucp_address_t *address;
size_t address_length;
resource_usage_t usage;
ucp_ep_h ep;
status = ucp_config_read(NULL, NULL, &config);
if (status != UCS_OK) {
return;
}
memset(¶ms, 0, sizeof(params));
params.field_mask = UCP_PARAM_FIELD_FEATURES |
UCP_PARAM_FIELD_ESTIMATED_NUM_EPS;
params.features = ctx_features;
params.estimated_num_eps = estimated_num_eps;
get_resource_usage(&usage);
if (!(dev_type_bitmap & UCS_BIT(UCT_DEVICE_TYPE_SELF))) {
ucp_config_modify(config, "SELF_DEVICES", "");
}
if (!(dev_type_bitmap & UCS_BIT(UCT_DEVICE_TYPE_SHM))) {
ucp_config_modify(config, "SHM_DEVICES", "");
}
if (!(dev_type_bitmap & UCS_BIT(UCT_DEVICE_TYPE_NET))) {
ucp_config_modify(config, "NET_DEVICES", "");
}
status = ucp_init(¶ms, config, &context);
if (status != UCS_OK) {
printf("<Failed to create UCP context>\n");
goto out_release_config;
}
if ((print_opts & PRINT_MEM_MAP) && (mem_size != NULL)) {
ucp_mem_print_info(mem_size, context, stdout);
}
if (print_opts & PRINT_UCP_CONTEXT) {
ucp_context_print_info(context, stdout);
print_resource_usage(&usage, "UCP context");
}
if (!(print_opts & (PRINT_UCP_WORKER|PRINT_UCP_EP))) {
goto out_cleanup_context;
}
worker_params.field_mask = UCP_WORKER_PARAM_FIELD_THREAD_MODE;
worker_params.thread_mode = UCS_THREAD_MODE_MULTI;
get_resource_usage(&usage);
status = ucp_worker_create(context, &worker_params, &worker);
if (status != UCS_OK) {
printf("<Failed to create UCP worker>\n");
goto out_cleanup_context;
}
if (print_opts & PRINT_UCP_WORKER) {
ucp_worker_print_info(worker, stdout);
print_resource_usage(&usage, "UCP worker");
}
if (print_opts & PRINT_UCP_EP) {
status = ucp_worker_get_address(worker, &address, &address_length);
if (status != UCS_OK) {
printf("<Failed to get UCP worker address>\n");
goto out_destroy_worker;
}
ep_params = *base_ep_params;
ep_params.field_mask |= UCP_EP_PARAM_FIELD_REMOTE_ADDRESS;
ep_params.address = address;
status = ucp_ep_create(worker, &ep_params, &ep);
ucp_worker_release_address(worker, address);
if (status != UCS_OK) {
printf("<Failed to create UCP endpoint>\n");
goto out_destroy_worker;
}
ucp_ep_print_info(ep, stdout);
status_ptr = ucp_disconnect_nb(ep);
if (UCS_PTR_IS_PTR(status_ptr)) {
do {
ucp_worker_progress(worker);
status = ucp_request_test(status_ptr, NULL);
} while (status == UCS_INPROGRESS);
ucp_request_release(status_ptr);
}
}
out_destroy_worker:
ucp_worker_destroy(worker);
out_cleanup_context:
ucp_cleanup(context);
out_release_config:
ucp_config_release(config);
}
static int run_ucx_server(ucp_worker_h ucp_worker)
{
ucp_tag_recv_info_t info_tag;
ucp_tag_message_h msg_tag;
ucs_status_t status;
ucp_ep_h client_ep;
struct msg *msg = 0;
struct ucx_context *request = 0;
size_t msg_len = 0;
int ret = -1;
/* Receive client UCX address */
do {
/* Following blocked methods used to polling internal file descriptor
* to make CPU idle and don't spin loop
*/
if (ucp_test_mode == TEST_MODE_WAIT) {
status = ucp_worker_wait(ucp_worker);
if (status != UCS_OK) {
goto err;
}
} else if (ucp_test_mode == TEST_MODE_EVENTFD) {
status = test_poll_wait(ucp_worker);
if (status != UCS_OK) {
goto err;
}
}
/* Progressing before probe to update the state */
ucp_worker_progress(ucp_worker);
/* Probing incoming events in non-block mode */
msg_tag = ucp_tag_probe_nb(ucp_worker, tag, tag_mask, 1, &info_tag);
} while (msg_tag == NULL);
msg = malloc(info_tag.length);
if (!msg) {
fprintf(stderr, "unable to allocate memory\n");
goto err;
}
request = ucp_tag_msg_recv_nb(ucp_worker, msg, info_tag.length,
ucp_dt_make_contig(1), msg_tag, recv_handle);
if (UCS_PTR_IS_ERR(request)) {
fprintf(stderr, "unable to receive UCX address message (%s)\n",
ucs_status_string(UCS_PTR_STATUS(request)));
free(msg);
goto err;
} else {
wait(ucp_worker, request);
ucp_request_release(request);
printf("UCX address message was received\n");
}
peer_addr = malloc(msg->data_len);
if (!peer_addr) {
fprintf(stderr, "unable to allocate memory for peer address\n");
free(msg);
goto err;
}
peer_addr_len = msg->data_len;
memcpy(peer_addr, msg->data, peer_addr_len);
free(msg);
/* Send test string to client */
status = ucp_ep_create(ucp_worker, peer_addr, &client_ep);
if (status != UCS_OK) {
goto err;
}
msg_len = sizeof(*msg) + strlen(test_str) + 1;
msg = calloc(1, msg_len);
if (!msg) {
printf("unable to allocate memory\n");
goto err_ep;
}
msg->data_len = msg_len - sizeof(*msg);
snprintf((char *)msg->data, msg->data_len, "%s", test_str);
request = ucp_tag_send_nb(client_ep, msg, msg_len,
ucp_dt_make_contig(1), tag,
send_handle);
if (UCS_PTR_IS_ERR(request)) {
fprintf(stderr, "unable to send UCX data message\n");
free(msg);
goto err_ep;
} else if (UCS_PTR_STATUS(request) != UCS_OK) {
printf("UCX data message was scheduled for send\n");
wait(ucp_worker, request);
ucp_request_release(request);
}
ret = 0;
free(msg);
err_ep:
ucp_ep_destroy(client_ep);
err:
return ret;
}
static int connect_server()
{
ucp_tag_recv_info_t info_tag;
ucp_tag_message_h msg_tag;
ucs_status_t status;
ucp_ep_params_t ep_params;
struct msg *msg = 0;
struct ucx_context *request = 0;
/* Receive client UCX address */
do {
/* Following blocked methods used to polling internal file descriptor
* to make CPU idle and don't spin loop
*/
/* Progressing before probe to update the state */
ucp_worker_progress(ucp_worker);
/* Probing incoming events in non-block mode */
msg_tag = ucp_tag_probe_nb(ucp_worker, tag, tag_mask, 1, &info_tag);
} while (msg_tag == NULL);
msg = malloc(info_tag.length);
if (!msg) {
fprintf(stderr, "unable to allocate memory\n");
abort();
}
request = ucp_tag_msg_recv_nb(ucp_worker, msg, info_tag.length,
ucp_dt_make_contig(1), msg_tag, recv_handle);
if (UCS_PTR_IS_ERR(request)) {
fprintf(stderr, "unable to receive UCX address message (%s)\n",
ucs_status_string(UCS_PTR_STATUS(request)));
free(msg);
abort();
} else {
wait(ucp_worker, request);
request->completed = 0;
ucp_request_release(request);
printf("UCX address message was received\n");
}
peer_addr = malloc(msg->data_len);
if (!peer_addr) {
fprintf(stderr, "unable to allocate memory for peer address\n");
free(msg);
abort();
}
peer_addr_len = msg->data_len;
memcpy(peer_addr, msg->data, peer_addr_len);
free(msg);
/* Send test string to client */
ep_params.field_mask = UCP_EP_PARAM_FIELD_REMOTE_ADDRESS;
ep_params.address = peer_addr;
status = ucp_ep_create(ucp_worker, &ep_params, &rem_ep);
if (status != UCS_OK) {
abort();
}
}
static void wait(ucp_worker_h ucp_worker, struct ucx_context *context)
{
while (context->completed == 0) {
ucp_worker_progress(ucp_worker);
}
}
static void mlx_ep_progress( struct util_ep *util_ep)
{
struct mlx_ep* ep;
ep = container_of(util_ep, struct mlx_ep, ep);
ucp_worker_progress(ep->worker);
}
