static int udapl_reg_mr(void *reg_data, void *base, size_t size, mca_mpool_base_registration_t *reg) { mca_btl_udapl_module_t *btl = (mca_btl_udapl_module_t*)reg_data; mca_btl_udapl_reg_t *udapl_reg = (mca_btl_udapl_reg_t*)reg; DAT_REGION_DESCRIPTION region; DAT_VLEN dat_size; DAT_VADDR dat_addr; int rc; DAT_MEM_TYPE lmr_mem_type = DAT_MEM_TYPE_VIRTUAL; region.for_va = base; udapl_reg->lmr_triplet.virtual_address = (DAT_VADDR)(uintptr_t)base; udapl_reg->lmr_triplet.segment_length = size; udapl_reg->lmr = NULL; #if HAVE_DAT_MEM_TYPE_SO_VIRTUAL if (reg->flags & MCA_MPOOL_FLAGS_SO_MEM) { lmr_mem_type = DAT_MEM_TYPE_SO_VIRTUAL; } #endif rc = dat_lmr_create(btl->udapl_ia, lmr_mem_type, region, size, btl->udapl_pz, DAT_MEM_PRIV_ALL_FLAG, &udapl_reg->lmr, &udapl_reg->lmr_triplet.lmr_context, &udapl_reg->rmr_context, &dat_size, &dat_addr); if(rc != DAT_SUCCESS) { BTL_UDAPL_VERBOSE_HELP(VERBOSE_SHOW_HELP, ("help-mpi-btl-udapl.txt", "dat_lmr_create DAT_INSUFFICIENT_RESOURCES", true)); return OMPI_ERR_OUT_OF_RESOURCE; } return OMPI_SUCCESS; }
/* return -1 on error. on error debug messages printed on stderr */ static int psdapl_bufpair_init_local(psdapl_bufpair_t *bufp, size_t size, DAT_IA_HANDLE ia_handle, DAT_PZ_HANDLE pz_handle) { DAT_RETURN res; bufp->lmr_mem = valloc(size); if (!bufp->lmr_mem) goto err_malloc; memset(bufp->lmr_mem, 0xee, size); /* touch the mem */ DAT_REGION_DESCRIPTION region; region.for_va = bufp->lmr_mem; DAT_VLEN registered_size = 0; DAT_VADDR registered_address = 0; res = dat_lmr_create( ia_handle, DAT_MEM_TYPE_VIRTUAL, region, size, pz_handle, DAT_MEM_PRIV_ALL_FLAG, &bufp->lmr_handle, &bufp->lmr_context, &bufp->lmr_rmr_context, ®istered_size, ®istered_address); if (res != DAT_SUCCESS) goto err_lmr_create; return 0; err_malloc: psdapl_dprint_errno(1, errno, "calloc(%lu, 1) failed", (long)size); return -1; err_lmr_create: psdapl_dprint_dat_err(1, res, "dat_lmr_create() failed"); return -1; }
int connect_ep(char *hostname) { DAT_IA_ADDRESS_PTR remote_addr = (DAT_IA_ADDRESS_PTR)&remote; DAT_EP_ATTR ep_attr; DAT_IA_ATTR ia_attr; DAT_RETURN status; DAT_REGION_DESCRIPTION region; DAT_EVENT event; DAT_COUNT nmore; DAT_LMR_TRIPLET iov; DAT_RMR_TRIPLET *r_iov; DAT_DTO_COOKIE cookie; DAT_CONN_QUAL conn_qual; DAT_BOOLEAN in, out; int i, ii, pdata, ctx; DAT_PROVIDER_ATTR prov_attrs; DAT_DTO_COMPLETION_EVENT_DATA *dto_event = &event.event_data.dto_completion_event_data; status = dat_ia_open(provider, 8, &async_evd, &ia); _OK(status, "dat_ia_open"); memset(&prov_attrs, 0, sizeof(prov_attrs)); status = dat_ia_query(ia, NULL, DAT_IA_FIELD_ALL, &ia_attr, DAT_PROVIDER_FIELD_ALL, &prov_attrs); _OK(status, "dat_ia_query"); print_ia_address(ia_attr.ia_address_ptr); if (ucm && ud_test) { printf("%d UD test over UCM provider not supported\n", getpid()); exit(1); } /* Print provider specific attributes */ for (i = 0; i < prov_attrs.num_provider_specific_attr; i++) { LOGPRINTF(" Provider Specific Attribute[%d] %s=%s\n", i, prov_attrs.provider_specific_attr[i].name, prov_attrs.provider_specific_attr[i].value); /* check for counter support */ status = strcmp(prov_attrs.provider_specific_attr[i].name, "DAT_COUNTERS"); if (!status) counters_ok = 1; } /* make sure provider supports counters */ if ((counters) && (!counters_ok)) { printf("Disable dat_query_counters:" " Provider not built with counters\n"); counters = 0; } status = dat_pz_create(ia, &pz); _OK(status, "dat_pz_create"); status = dat_evd_create(ia, eps * 2, DAT_HANDLE_NULL, DAT_EVD_CR_FLAG, &cr_evd); _OK(status, "dat_evd_create CR"); status = dat_evd_create(ia, eps * 2, DAT_HANDLE_NULL, DAT_EVD_CONNECTION_FLAG, &con_evd); _OK(status, "dat_evd_create CR"); status = dat_evd_create(ia, eps * 10, DAT_HANDLE_NULL, DAT_EVD_DTO_FLAG, &dto_evd); _OK(status, "dat_evd_create DTO"); memset(&ep_attr, 0, sizeof(ep_attr)); if (ud_test) { msg_size += 40; ep_attr.service_type = DAT_IB_SERVICE_TYPE_UD; ep_attr.max_message_size = buf_size; ep_attr.max_rdma_read_in = 0; ep_attr.max_rdma_read_out = 0; } else { ep_attr.service_type = DAT_SERVICE_TYPE_RC; ep_attr.max_rdma_size = 0x10000; ep_attr.max_rdma_read_in = 4; ep_attr.max_rdma_read_out = 4; } ep_attr.qos = 0; ep_attr.recv_completion_flags = 0; ep_attr.max_recv_dtos = eps * 10; ep_attr.max_request_dtos = eps * 10; ep_attr.max_recv_iov = 1; ep_attr.max_request_iov = 1; ep_attr.request_completion_flags = DAT_COMPLETION_DEFAULT_FLAG; ep_attr.ep_transport_specific_count = 0; ep_attr.ep_transport_specific = NULL; ep_attr.ep_provider_specific_count = 0; ep_attr.ep_provider_specific = NULL; for (i = 0; i < eps; i++) { status = dat_ep_create(ia, pz, dto_evd, dto_evd, con_evd, &ep_attr, &ep[i]); _OK(status, "dat_ep_create"); LOGPRINTF(" create_ep[%d]=%p\n", i, ep[i]); } for (i = 0; i < REG_MEM_COUNT * eps; i++) { buf[i] = (DAT_RMR_TRIPLET *) malloc(buf_size); region.for_va = buf[i]; status = dat_lmr_create(ia, DAT_MEM_TYPE_VIRTUAL, region, buf_size, pz, DAT_MEM_PRIV_ALL_FLAG | DAT_IB_MEM_PRIV_REMOTE_ATOMIC, DAT_VA_TYPE_VA, &lmr[i], &lmr_context[i], &rmr_context[i], ®_size[i], ®_addr[i]); _OK(status, "dat_lmr_create"); } /* register atomic return buffer for original data */ atomic_buf = (DAT_UINT64 *) malloc(BUF_SIZE_ATOMIC); region.for_va = atomic_buf; status = dat_lmr_create(ia, DAT_MEM_TYPE_VIRTUAL, region, BUF_SIZE_ATOMIC, pz, DAT_MEM_PRIV_ALL_FLAG | DAT_IB_MEM_PRIV_REMOTE_ATOMIC, DAT_VA_TYPE_VA, &lmr_atomic, &lmr_atomic_context, &rmr_atomic_context, ®_atomic_size, ®_atomic_addr); _OK(status, "dat_lmr_create atomic"); for (ii = 0; ii < eps; ii++) { for (i = RECV_BUF_INDEX; i < REG_MEM_COUNT; i++) { int ep_idx = 0; cookie.as_64 = (ii * REG_MEM_COUNT) + i; iov.lmr_context = lmr_context[(ii * REG_MEM_COUNT) + i]; iov.virtual_address = (DAT_VADDR) (uintptr_t) buf[(ii * REG_MEM_COUNT) + i]; iov.segment_length = buf_size; LOGPRINTF(" post_recv (%p) on ep[%d]=%p\n", buf[(ii * REG_MEM_COUNT) + i], ii, ep[ii]); /* ep[0], unless testing Server and multi EP's */ if (server && multi_eps) { ep_idx = ii; cookie.as_64 = i; } status = dat_ep_post_recv(ep[ep_idx], 1, &iov, cookie, DAT_COMPLETION_DEFAULT_FLAG); _OK(status, "dat_ep_post_recv"); } } /* setup receive buffer to initial string to be overwritten */ strcpy((char *)buf[RCV_RDMA_BUF_INDEX], "blah, blah, blah\n"); /* ud can resolve_ah and connect both ways, same EP */ if (server || (!server && ud_test)) { if (server) { conn_qual = SERVER_ID; strcpy((char *)buf[SND_RDMA_BUF_INDEX], "Server data"); } else { conn_qual = CLIENT_ID; strcpy((char *)buf[SND_RDMA_BUF_INDEX], "Client data"); } status = dat_psp_create(ia, conn_qual, cr_evd, DAT_PSP_CONSUMER_FLAG, &psp); _OK(status, "dat_psp_create"); /* Server always waits for first CR from Client */ if (server) process_cr(0); } /* ud can resolve_ah and connect both ways */ if (!server || (server && ud_test)) { struct addrinfo *target; if (ucm) goto no_resolution; if (getaddrinfo(hostname, NULL, NULL, &target) != 0) { printf("Error getting remote address.\n"); exit(1); } printf("Remote %s Name: %s \n", server ? "Client" : "Server", hostname); printf("Remote %s Net Address: %s\n", server ? "Client" : "Server", inet_ntoa(((struct sockaddr_in *) target->ai_addr)->sin_addr)); strcpy((char *)buf[SND_RDMA_BUF_INDEX], "Client written data"); remote_addr = (DAT_IA_ADDRESS_PTR)target->ai_addr; /* IP */ no_resolution: /* one Client EP, multiple Server EPs, same conn_qual * use private data to select EP on Server */ for (i = 0; i < eps; i++) { /* pdata selects Server EP, * support both muliple Server and single EP's */ if (multi_eps) pdata = hton32(i); else pdata = 0; /* just use first EP */ status = dat_ep_connect(ep[0], remote_addr, (server ? CLIENT_ID : SERVER_ID), CONN_TIMEOUT, 4, (DAT_PVOID) & pdata, 0, DAT_CONNECT_DEFAULT_FLAG); _OK(status, "dat_ep_connect"); } if (!ucm) freeaddrinfo(target); } /* UD: process CR's starting with 2nd on server, 1st for client */ if (ud_test) { for (i = (server ? 1 : 0); i < eps; i++) process_cr(i); } /* RC and UD: process CONN EST events */ for (i = 0; i < eps; i++) process_conn(i); /* UD: CONN EST events for CONN's and CR's */ if (ud_test) { for (i = 0; i < eps; i++) process_conn(i); } printf("Connected! %d endpoints\n", eps); /* * Setup our remote memory and tell the other side about it * Swap to network order. */ r_iov = (DAT_RMR_TRIPLET *) buf[SEND_BUF_INDEX]; r_iov->rmr_context = hton32(rmr_context[RCV_RDMA_BUF_INDEX]); r_iov->virtual_address = hton64((DAT_VADDR) (uintptr_t) buf[RCV_RDMA_BUF_INDEX]); r_iov->segment_length = hton32(buf_size); printf("Send RMR message: r_key_ctx=0x%x,va=" F64x ",len=0x%x\n", hton32(r_iov->rmr_context), hton64(r_iov->virtual_address), hton32(r_iov->segment_length)); send_msg(buf[SEND_BUF_INDEX], sizeof(DAT_RMR_TRIPLET), lmr_context[SEND_BUF_INDEX], cookie, DAT_COMPLETION_SUPPRESS_FLAG); dat_ep_get_status(ep[0], NULL, &in, &out); printf("EP[0] status: posted buffers: Req=%d, Rcv=%d\n", in, out); /* * Wait for their RMR */ for (i = 0, ctx = 0; i < eps; i++, ctx++) { /* expected cookie, recv buf idx in every mem pool */ ctx = (ctx % REG_MEM_COUNT) ? ctx : ctx + RECV_BUF_INDEX; LOGPRINTF("Waiting for remote to send RMR data\n"); status = dat_evd_wait(dto_evd, DTO_TIMEOUT, 1, &event, &nmore); _OK(status, "dat_evd_wait after dat_ep_post_send"); if ((event.event_number != DAT_DTO_COMPLETION_EVENT) && (ud_test && event.event_number != DAT_IB_DTO_EVENT)) { printf("unexpected event waiting for RMR context " "- 0x%x\n", event.event_number); exit(1); } _OK(dto_event->status, "event status for post_recv"); /* careful when checking cookies: * Client - receiving multi messages on a single EP * Server - not receiving on multiple EP's */ if (!server || (server && !multi_eps)) { if (dto_event->transfered_length != msg_size || dto_event->user_cookie.as_64 != ctx) { printf("unexpected event data on recv: len=%d" " cookie=" F64x " expected %d/%d\n", (int)dto_event->transfered_length, dto_event->user_cookie.as_64, msg_size, ctx); exit(1); } /* Server - receiving one message each across many EP's */ } else { if (dto_event->transfered_length != msg_size || dto_event->user_cookie.as_64 != RECV_BUF_INDEX) { printf("unexpected event data on recv: len=%d" "cookie=" F64x " expected %d/%d\n", (int)dto_event->transfered_length, dto_event->user_cookie.as_64, msg_size, RECV_BUF_INDEX); exit(1); } } /* swap RMR,address info to host order */ if (!server || (server && !multi_eps)) r_iov = (DAT_RMR_TRIPLET *) buf[ctx]; else r_iov = (DAT_RMR_TRIPLET *) buf[(i * REG_MEM_COUNT) + RECV_BUF_INDEX]; if (ud_test) r_iov = (DAT_RMR_TRIPLET *) ((char *)r_iov + 40); r_iov->rmr_context = ntoh32(r_iov->rmr_context); r_iov->virtual_address = ntoh64(r_iov->virtual_address); r_iov->segment_length = ntoh32(r_iov->segment_length); printf("Recv RMR message: r_iov(%p):" " r_key_ctx=%x,va=" F64x ",len=0x%x on EP=%p\n", r_iov, r_iov->rmr_context, r_iov->virtual_address, r_iov->segment_length, dto_event->ep_handle); } return (0); }