void MV_Generate_Address_Handles() {
int i;
/*Create address handles */
for (i = 0; i < mvdev.np; i++) {
mvdev_connection_t * c = &(mvdev.connections[i]);
int j;
int total_count = 0;
int k;
for(k = 0; k < mvparams.max_sl; k++) {
for(j = 0; j < mvparams.max_lmc_total; j++) {
struct ibv_ah_attr ah_attr;
memset(&ah_attr, 0, sizeof(ah_attr));
ah_attr.is_global = 0;
ah_attr.dlid = mvdev.lids[i] + j;
ah_attr.sl = k;
ah_attr.src_path_bits = 0;
ah_attr.port_num = mvparams.default_port;
c->data_ud_ah[total_count] = ibv_create_ah(mvdev.default_hca->pd, &ah_attr);
if (!c->data_ud_ah[total_count]) {
error_abort_all(IBV_RETURN_ERR, "Failed to create AH");
}
total_count++;
}
}
}
mvparams.max_ah_total = mvparams.max_lmc_total * mvparams.max_sl;
}
int mca_oob_ud_peer_update_with_uri (mca_oob_ud_peer_t *peer, const char *uri)
{
opal_list_item_t *item;
struct ibv_ah_attr ah_attr;
mca_oob_ud_device_t *device;
uint32_t qp_num;
/* NTH: port is 16-bit here because C90 does not support hh in sscanf */
uint16_t lid, port_num;
int rc;
rc = mca_oob_ud_parse_uri (uri, &qp_num, &lid, &port_num);
if (ORTE_SUCCESS != rc) {
return rc;
}
if (peer->peer_lid != lid || peer->peer_port != port_num) {
if (NULL != peer->peer_ah) {
(void) ibv_destroy_ah (peer->peer_ah);
peer->peer_ah = NULL;
}
}
peer->peer_qpn = qp_num;
peer->peer_qkey = 0; /* NTH: todo -- add qkey support if needed */
peer->peer_lid = lid;
peer->peer_port = port_num;
if (NULL == peer->peer_ah) {
memset (&ah_attr, 0, sizeof (ah_attr));
ah_attr.dlid = lid;
ah_attr.port_num = port_num;
for (item = opal_list_get_first (&mca_oob_ud_component.ud_devices);
item != opal_list_get_end (&mca_oob_ud_component.ud_devices);
item = opal_list_get_next (item)) {
device = (mca_oob_ud_device_t *)item;
/* try to create an address handle using this device */
peer->peer_ah = ibv_create_ah (device->ib_pd, &ah_attr);
if (NULL != peer->peer_ah) {
peer->peer_context = (void *) item;
break;
}
}
if (NULL == peer->peer_ah) {
free (peer);
return ORTE_ERROR;
}
}
return ORTE_SUCCESS;
}
struct ibv_ah *uct_ib_create_ah(uct_ib_iface_t *iface, uint16_t dlid)
{
uct_ib_pd_t *ib_pd = ucs_derived_of(iface->super.pd, uct_ib_pd_t);
struct ibv_ah_attr ah_attr;
memset(&ah_attr, 0, sizeof(ah_attr));
ah_attr.port_num = iface->port_num;
ah_attr.sl = iface->sl;
ah_attr.is_global = 0;
ah_attr.dlid = dlid;
return ibv_create_ah(ib_pd->pd, &ah_attr);
}
struct ibv_ah *uct_ib_create_ah(uct_ib_iface_t *iface, uint16_t dlid)
{
struct ibv_ah_attr ah_attr;
uct_ib_device_t *dev = uct_ib_iface_device(iface);
memset(&ah_attr, 0, sizeof(ah_attr));
ah_attr.port_num = iface->port_num;
ah_attr.sl = iface->sl;
ah_attr.is_global = 0;
ah_attr.dlid = dlid;
return ibv_create_ah(dev->pd, &ah_attr);
}
mca_oob_ud_peer_t *mca_oob_ud_get_peer (struct mca_oob_ud_port_t *port,
orte_process_name_t *name,
uint32_t qpn, uint32_t qkey,
uint16_t lid, uint8_t port_num)
{
struct ibv_ah_attr ah_attr;
mca_oob_ud_peer_t *peer;
int rc;
rc = mca_oob_ud_peer_lookup (name, &peer);
if (ORTE_SUCCESS == rc) {
OPAL_OUTPUT_VERBOSE((20, mca_oob_base_output, "%s oob:ud:peer_from_msg_hdr using "
"cached peer", ORTE_NAME_PRINT(ORTE_PROC_MY_NAME)));
return peer;
}
OPAL_OUTPUT_VERBOSE((10, mca_oob_base_output, "%s oob:ud:peer_from_msg_hdr creating "
"peer from return address", ORTE_NAME_PRINT(ORTE_PROC_MY_NAME)));
peer = OBJ_NEW(mca_oob_ud_peer_t);
if (NULL == peer) {
return NULL;
}
peer->peer_qpn = qpn;
peer->peer_qkey = qkey;
peer->peer_name = *name;
peer->peer_lid = lid;
peer->peer_port = port_num;
memset (&ah_attr, 0, sizeof (ah_attr));
ah_attr.dlid = peer->peer_lid;
ah_attr.port_num = peer->peer_port;
peer->peer_ah = ibv_create_ah (port->device->ib_pd, &ah_attr);
if (NULL == peer->peer_ah) {
free (peer);
return NULL;
}
peer->peer_context = port->device;
OPAL_THREAD_LOCK(&mca_oob_ud_component.ud_lock);
opal_hash_table_set_value_uint64(&mca_oob_ud_component.ud_peers,
orte_util_hash_name(name),
(void *) peer);
OPAL_THREAD_UNLOCK(&mca_oob_ud_component.ud_lock);
return peer;
}
static int resolved_handler(struct cmatest_node *node,
struct rdma_cm_event *event)
{
node->remote_qpn = event->param.ud.qp_num;
node->remote_qkey = event->param.ud.qkey;
node->ah = ibv_create_ah(node->pd, &event->param.ud.ah_attr);
if (!node->ah) {
printf("udaddy: failure creating address handle\n");
goto err;
}
node->connected = 1;
test.connects_left--;
return 0;
err:
connect_error();
return -1;
}
static struct ibv_ah *create_ah(RdmaBackendDev *backend_dev, struct ibv_pd *pd,
uint8_t sgid_idx, union ibv_gid *dgid)
{
GBytes *ah_key = g_bytes_new(dgid, sizeof(*dgid));
struct ibv_ah *ah = g_hash_table_lookup(ah_hash, ah_key);
if (ah) {
trace_rdma_create_ah_cache_hit(be64_to_cpu(dgid->global.subnet_prefix),
be64_to_cpu(dgid->global.interface_id));
g_bytes_unref(ah_key);
} else {
struct ibv_ah_attr ah_attr = {
.is_global = 1,
.port_num = backend_dev->port_num,
.grh.hop_limit = 1,
};
ah_attr.grh.dgid = *dgid;
ah_attr.grh.sgid_index = sgid_idx;
ah = ibv_create_ah(pd, &ah_attr);
if (ah) {
g_hash_table_insert(ah_hash, ah_key, ah);
} else {
g_bytes_unref(ah_key);
rdma_error_report("Failed to create AH for gid <0x%" PRIx64", 0x%"PRIx64">",
be64_to_cpu(dgid->global.subnet_prefix),
be64_to_cpu(dgid->global.interface_id));
}
trace_rdma_create_ah_cache_miss(be64_to_cpu(dgid->global.subnet_prefix),
be64_to_cpu(dgid->global.interface_id));
}
return ah;
}
static void destroy_ah_hash_key(gpointer data)
{
g_bytes_unref(data);
}
struct ibv_ah_1_0 *__ibv_create_ah_1_0(struct ibv_pd_1_0 *pd,
struct ibv_ah_attr *attr)
{
struct ibv_ah *real_ah;
struct ibv_ah_1_0 *ah;
ah = malloc(sizeof *ah);
if (!ah)
return NULL;
real_ah = ibv_create_ah(pd->real_pd, attr);
if (!real_ah) {
free(ah);
return NULL;
}
ah->context = pd->context;
ah->pd = pd;
ah->real_ah = real_ah;
return ah;
}
/* create ud vc */
int mv2_ud_set_vc_info (mv2_ud_vc_info_t *ud_vc_info, mv2_ud_exch_info_t *rem_info, struct ibv_pd *pd, int rdma_default_port)
{
struct ibv_ah_attr ah_attr;
PRINT_DEBUG(DEBUG_UD_verbose>0,"lid:%d\n", rem_info->lid );
memset(&ah_attr, 0, sizeof(ah_attr));
ah_attr.is_global = 0;
ah_attr.dlid = rem_info->lid;
ah_attr.sl = rdma_default_service_level;
ah_attr.src_path_bits = 0;
ah_attr.port_num = rdma_default_port;
ud_vc_info->ah = ibv_create_ah(pd, &ah_attr);
if(!(ud_vc_info->ah)){
fprintf(stderr, "Error in creating address handle\n");
return -1;
}
ud_vc_info->lid = rem_info->lid;
ud_vc_info->qpn = rem_info->qpn;
return 0;
}
struct ibv_ah_1_0 *__ibv_create_ah_1_0(struct ibv_pd_1_0 *pd,
struct ibv_ah_attr *attr)
{ fprintf(stderr, "%s:%s:%d \n", __func__, __FILE__, __LINE__);
struct ibv_ah *real_ah;
struct ibv_ah_1_0 *ah;
ah = malloc(sizeof *ah);
if (!ah)
return NULL;
real_ah = ibv_create_ah(pd->real_pd, attr);
if (!real_ah) {
free(ah);
return NULL;
}
ah->context = pd->context;
ah->pd = pd;
ah->real_ah = real_ah;
return ah;
}
ucs_status_t uct_ib_iface_create_ah(uct_ib_iface_t *iface,
const uct_ib_address_t *ib_addr,
uint8_t path_bits,
struct ibv_ah **ah_p,
int *is_global_p)
{
struct ibv_ah_attr ah_attr;
struct ibv_ah *ah;
char buf[128];
char *p, *endp;
uct_ib_iface_fill_ah_attr(iface, ib_addr, path_bits, &ah_attr);
ah = ibv_create_ah(uct_ib_iface_md(iface)->pd, &ah_attr);
if (ah == NULL) {
p = buf;
endp = buf + sizeof(buf);
snprintf(p, endp - p, "dlid=%d sl=%d port=%d src_path_bits=%d",
ah_attr.dlid, ah_attr.sl, ah_attr.port_num, ah_attr.src_path_bits);
p += strlen(p);
if (ah_attr.is_global) {
snprintf(p, endp - p, " dgid=");
p += strlen(p);
inet_ntop(AF_INET6, &ah_attr.grh.dgid, p, endp - p);
p += strlen(p);
snprintf(p, endp - p, " sgid_index=%d", ah_attr.grh.sgid_index);
}
ucs_error("ibv_create_ah(%s) on "UCT_IB_IFACE_FMT" failed: %m", buf,
UCT_IB_IFACE_ARG(iface));
return UCS_ERR_INVALID_ADDR;
}
*ah_p = ah;
*is_global_p = ah_attr.is_global;
return UCS_OK;
}
static int join_handler(struct cmatest_node *node,
struct rdma_ud_param *param)
{
char buf[40];
inet_ntop(AF_INET6, param->ah_attr.grh.dgid.raw, buf, 40);
printf("mckey: joined dgid: %s mlid 0x%x sl %d\n", buf,
param->ah_attr.dlid, param->ah_attr.sl);
node->remote_qpn = param->qp_num;
node->remote_qkey = param->qkey;
node->ah = ibv_create_ah(node->pd, ¶m->ah_attr);
if (!node->ah) {
printf("mckey: failure creating address handle\n");
goto err;
}
node->connected = 1;
test.connects_left--;
return 0;
err:
connect_error();
return -1;
}
void *run_client(void *arg)
{
int ud_qp_i = 0;
struct thread_params params = *(struct thread_params *) arg;
/*
* The local HID of a control block should be <= 64 to keep the SHM key low.
* But the number of clients over all machines can be larger.
*/
int clt_gid = params.id; /* Global ID of this client thread */
int clt_local_hid = clt_gid % params.num_threads;
int srv_gid = clt_gid % NUM_SERVER_THREADS;
int ib_port_index = params.dual_port == 0 ? 0 : srv_gid % 2;
struct hrd_ctrl_blk *cb = hrd_ctrl_blk_init(clt_local_hid,
ib_port_index, -1, /* port_index, numa_node_id */
0, 0, /* conn qps, use uc */
NULL, 0, -1, /* prealloc conn buf, conn buf size, key */
1, BUF_SIZE, -1); /* num_dgram_qps, dgram_buf_size, key */
/* Buffer to receive responses into */
memset((void *) cb->dgram_buf, 0, BUF_SIZE);
/* Buffer to send requests from */
uint8_t *req_buf = malloc(params.size);
assert(req_buf != 0);
memset(req_buf, clt_gid, params.size);
printf("main: Client %d waiting for server %d\n", clt_gid, srv_gid);
struct hrd_qp_attr *srv_qp[NUM_UD_QPS] = {NULL};
for(ud_qp_i = 0; ud_qp_i < NUM_UD_QPS; ud_qp_i++) {
char srv_name[HRD_QP_NAME_SIZE];
sprintf(srv_name, "server-%d-%d", srv_gid, ud_qp_i);
while(srv_qp[ud_qp_i] == NULL) {
srv_qp[ud_qp_i] = hrd_get_published_qp(srv_name);
if(srv_qp[ud_qp_i] == NULL) {
usleep(200000);
}
}
}
ud_qp_i = 0;
printf("main: Client %d found server! Now posting SENDs.\n", clt_gid);
/* We need only 1 ah because a client contacts only 1 server */
struct ibv_ah_attr ah_attr = {
.is_global = 0,
.dlid = srv_qp[0]->lid, /* All srv_qp have same LID */
.sl = 0,
.src_path_bits = 0,
.port_num = cb->dev_port_id,
};
struct ibv_ah *ah = ibv_create_ah(cb->pd, &ah_attr);
assert(ah != NULL);
struct ibv_send_wr wr[MAX_POSTLIST], *bad_send_wr;
struct ibv_wc wc[MAX_POSTLIST];
struct ibv_sge sgl[MAX_POSTLIST];
long long rolling_iter = 0; /* For throughput measurement */
long long nb_tx = 0;
int w_i = 0; /* Window index */
int ret;
struct timespec start, end;
clock_gettime(CLOCK_REALTIME, &start);
while(1) {
if(rolling_iter >= M_2) {
clock_gettime(CLOCK_REALTIME, &end);
double seconds = (end.tv_sec - start.tv_sec) +
(double) (end.tv_nsec - start.tv_nsec) / 1000000000;
printf("main: Client %d: %.2f Mops\n",
clt_gid, rolling_iter / seconds);
rolling_iter = 0;
clock_gettime(CLOCK_REALTIME, &start);
}
for(w_i = 0; w_i < params.postlist; w_i++) {
wr[w_i].wr.ud.ah = ah;
wr[w_i].wr.ud.remote_qpn = srv_qp[ud_qp_i]->qpn;
wr[w_i].wr.ud.remote_qkey = HRD_DEFAULT_QKEY;
wr[w_i].opcode = IBV_WR_SEND_WITH_IMM;
wr[w_i].num_sge = 1;
wr[w_i].next = (w_i == params.postlist - 1) ? NULL : &wr[w_i + 1];
wr[w_i].imm_data = 3185;
wr[w_i].sg_list = &sgl[w_i];
/*
* UNSIG_BATCH >= 2 * postlist ensures that we poll for a
* completed send() only after we have performed a signaled send().
*/
wr[w_i].send_flags = (nb_tx & UNSIG_BATCH_) == 0 ?
IBV_SEND_SIGNALED : 0;
if((nb_tx & UNSIG_BATCH_) == UNSIG_BATCH_) {
hrd_poll_cq(cb->dgram_send_cq[0], 1, wc);
}
wr[w_i].send_flags |= IBV_SEND_INLINE;
sgl[w_i].addr = (uint64_t) (uintptr_t) req_buf;
sgl[w_i].length = params.size;
rolling_iter++;
nb_tx++;
}
ret = ibv_post_send(cb->dgram_qp[0], &wr[0], &bad_send_wr);
CPE(ret, "ibv_post_send error", ret);
HRD_MOD_ADD(ud_qp_i, NUM_UD_QPS);
}
return NULL;
}
static int pp_connect_ctx(struct pingpong_context *ctx, int port, int my_psn,
struct pingpong_dest *dest,struct user_parameters *user_parm)
{
struct ibv_qp_attr attr;
memset(&attr, 0, sizeof(struct ibv_qp_attr));
attr.qp_state = IBV_QPS_RTR;
if (user_parm->connection_type != UD) {
switch (user_parm->mtu) {
case 256 :
attr.path_mtu = IBV_MTU_256;
break;
case 512 :
attr.path_mtu = IBV_MTU_512;
break;
case 1024 :
attr.path_mtu = IBV_MTU_1024;
break;
case 2048 :
attr.path_mtu = IBV_MTU_2048;
break;
case 4096 :
attr.path_mtu = IBV_MTU_4096;
break;
}
printf("Mtu : %d\n", user_parm->mtu);
attr.dest_qp_num = dest->qpn;
attr.rq_psn = dest->psn;
}
if (user_parm->connection_type==RC) {
attr.max_dest_rd_atomic = 1;
attr.min_rnr_timer = 12;
}
if (user_parm->gid_index < 0) {
attr.ah_attr.is_global = 0;
attr.ah_attr.dlid = dest->lid;
attr.ah_attr.sl = sl;
} else {
attr.ah_attr.is_global = 1;
attr.ah_attr.grh.dgid = dest->dgid;
attr.ah_attr.grh.hop_limit = 1;
attr.ah_attr.sl = 0;
}
attr.ah_attr.src_path_bits = 0;
attr.ah_attr.port_num = port;
if ((user_parm->connection_type==UD) && (user_parm->use_mcg)) {
uint8_t mcg_gid[16] = MCG_GID;
/* send the message to the mcg of the other side */
mcg_gid[11] = (user_parm->servername) ? 1 : 0;
*(uint32_t *)(&mcg_gid[12]) = dest->qpn;
attr.ah_attr.dlid = MCG_LID;
attr.ah_attr.is_global = 1;
attr.ah_attr.grh.sgid_index = 0;
memcpy(attr.ah_attr.grh.dgid.raw, mcg_gid, 16);
}
if (user_parm->connection_type==RC) {
if (ibv_modify_qp(ctx->qp, &attr,
IBV_QP_STATE |
IBV_QP_AV |
IBV_QP_PATH_MTU |
IBV_QP_DEST_QPN |
IBV_QP_RQ_PSN |
IBV_QP_MIN_RNR_TIMER |
IBV_QP_MAX_DEST_RD_ATOMIC)) {
fprintf(stderr, "Failed to modify RC QP to RTR\n");
return 1;
}
attr.timeout = user_parm->qp_timeout;
attr.retry_cnt = 7;
attr.rnr_retry = 7;
} else if (user_parm->connection_type==UC) {
if (ibv_modify_qp(ctx->qp, &attr,
IBV_QP_STATE |
IBV_QP_AV |
IBV_QP_PATH_MTU |
IBV_QP_DEST_QPN |
IBV_QP_RQ_PSN)) {
fprintf(stderr, "Failed to modify UC QP to RTR\n");
return 1;
}
} else {
if (ibv_modify_qp(ctx->qp, &attr,
IBV_QP_STATE )) {
fprintf(stderr, "Failed to modify UC QP to RTR\n");
return 1;
}
}
attr.qp_state = IBV_QPS_RTS;
attr.sq_psn = my_psn;
if (user_parm->connection_type==RC) {
attr.max_rd_atomic = 1;
if (ibv_modify_qp(ctx->qp, &attr,
IBV_QP_STATE |
IBV_QP_SQ_PSN |
IBV_QP_TIMEOUT |
IBV_QP_RETRY_CNT |
IBV_QP_RNR_RETRY |
IBV_QP_MAX_QP_RD_ATOMIC)) {
fprintf(stderr, "Failed to modify RC QP to RTS\n");
return 1;
}
} else { /*both UC and UD */
if (ibv_modify_qp(ctx->qp, &attr,
IBV_QP_STATE |
IBV_QP_SQ_PSN)) {
fprintf(stderr, "Failed to modify UC/UD QP to RTS\n");
return 1;
}
}
if (user_parm->connection_type==UD) {
ctx->ah = ibv_create_ah(ctx->pd, &attr.ah_attr);
if (!ctx->ah) {
fprintf(stderr, "Failed to create AH for UD\n");
return 1;
}
}
/* post recieve max msg size*/
{
int i;
struct ibv_recv_wr *bad_wr_recv;
ctx->recv_list.addr = (uintptr_t) ctx->buf;
if (user_parm->connection_type==UD) {
ctx->recv_list.length = ctx->size + 40;
} else {
ctx->recv_list.length = ctx->size;
}
ctx->recv_list.lkey = ctx->mr->lkey;
for (i = 0; i < user_parm->tx_depth / 2; ++i) {
if (ibv_post_recv(ctx->qp, &ctx->rwr, &bad_wr_recv)) {
fprintf(stderr, "Couldn't post recv: counter=%d\n",
i);
return 14;
}
}
}
return 0;
}
int rdma_client_connect(struct pingpong_context *ctx,struct perftest_parameters *user_param)
{
char *service;
int temp,num_of_retry= NUM_OF_RETRIES;
struct sockaddr_in sin;
struct addrinfo *res;
struct rdma_cm_event *event;
struct rdma_conn_param conn_param;
struct addrinfo hints;
memset(&hints, 0, sizeof hints);
hints.ai_family = AF_UNSPEC;
hints.ai_socktype = SOCK_STREAM;
if (check_add_port(&service,user_param->port,user_param->servername,&hints,&res)) {
fprintf(stderr, "Problem in resolving basic adress and port\n");
return FAILURE;
}
sin.sin_addr.s_addr = ((struct sockaddr_in*)res->ai_addr)->sin_addr.s_addr;
sin.sin_family = PF_INET;
sin.sin_port = htons((unsigned short)user_param->port);
while (1) {
if (num_of_retry == 0) {
fprintf(stderr, "Received %d times ADDR_ERROR\n",NUM_OF_RETRIES);
return FAILURE;
}
if (rdma_resolve_addr(ctx->cm_id, NULL,(struct sockaddr *)&sin,2000)) {
fprintf(stderr, "rdma_resolve_addr failed\n");
return FAILURE;
}
if (rdma_get_cm_event(ctx->cm_channel,&event)) {
fprintf(stderr, "rdma_get_cm_events failed\n");
return FAILURE;
}
if (event->event == RDMA_CM_EVENT_ADDR_ERROR) {
num_of_retry--;
rdma_ack_cm_event(event);
continue;
}
if (event->event != RDMA_CM_EVENT_ADDR_RESOLVED) {
fprintf(stderr, "unexpected CM event %d\n",event->event);
rdma_ack_cm_event(event);
return FAILURE;
}
rdma_ack_cm_event(event);
break;
}
if (user_param->tos != DEF_TOS) {
if (rdma_set_option(ctx->cm_id,RDMA_OPTION_ID,RDMA_OPTION_ID_TOS,&user_param->tos,sizeof(uint8_t))) {
fprintf(stderr, " Set TOS option failed: %d\n",event->event);
return FAILURE;
}
}
while (1) {
if (num_of_retry <= 0) {
fprintf(stderr, "Received %d times ADDR_ERROR - aborting\n",NUM_OF_RETRIES);
return FAILURE;
}
if (rdma_resolve_route(ctx->cm_id,2000)) {
fprintf(stderr, "rdma_resolve_route failed\n");
return FAILURE;
}
if (rdma_get_cm_event(ctx->cm_channel,&event)) {
fprintf(stderr, "rdma_get_cm_events failed\n");
return FAILURE;
}
if (event->event == RDMA_CM_EVENT_ROUTE_ERROR) {
num_of_retry--;
rdma_ack_cm_event(event);
continue;
}
if (event->event != RDMA_CM_EVENT_ROUTE_RESOLVED) {
fprintf(stderr, "unexpected CM event %d\n",event->event);
rdma_ack_cm_event(event);
return FAILURE;
}
rdma_ack_cm_event(event);
break;
}
ctx->context = ctx->cm_id->verbs;
temp = user_param->work_rdma_cm;
user_param->work_rdma_cm = ON;
if (ctx_init(ctx,user_param)) {
fprintf(stderr," Unable to create the resources needed by comm struct\n");
return FAILURE;
}
memset(&conn_param, 0, sizeof conn_param);
if (user_param->verb == READ || user_param->verb == ATOMIC) {
conn_param.responder_resources = user_param->out_reads;
conn_param.initiator_depth = user_param->out_reads;
}
user_param->work_rdma_cm = temp;
conn_param.retry_count = user_param->retry_count;
conn_param.rnr_retry_count = 7;
if (user_param->work_rdma_cm == OFF) {
if (post_one_recv_wqe(ctx)) {
fprintf(stderr, "Couldn't post send \n");
return 1;
}
}
if (rdma_connect(ctx->cm_id,&conn_param)) {
fprintf(stderr, "Function rdma_connect failed\n");
return FAILURE;
}
if (rdma_get_cm_event(ctx->cm_channel,&event)) {
fprintf(stderr, "rdma_get_cm_events failed\n");
return FAILURE;
}
if (event->event != RDMA_CM_EVENT_ESTABLISHED) {
rdma_ack_cm_event(event);
fprintf(stderr, "Unexpected CM event bl blka %d\n", event->event);
return FAILURE;
}
if (user_param->connection_type == UD) {
user_param->rem_ud_qpn = event->param.ud.qp_num;
user_param->rem_ud_qkey = event->param.ud.qkey;
ctx->ah[0] = ibv_create_ah(ctx->pd,&event->param.ud.ah_attr);
if (!ctx->ah) {
printf(" Unable to create address handler for UD QP\n");
return FAILURE;
}
if (user_param->tst == LAT || (user_param->tst == BW && user_param->duplex)) {
if (send_qp_num_for_ah(ctx,user_param)) {
printf(" Unable to send my QP number\n");
return FAILURE;
}
}
}
rdma_ack_cm_event(event);
return SUCCESS;
}
void *run_server(void *arg)
{
int i;
struct thread_params params = *(struct thread_params *) arg;
int srv_gid = params.id; /* Global ID of this server thread */
int ib_port_index = params.dual_port == 0 ? 0 : srv_gid % 2;
struct hrd_ctrl_blk *cb = hrd_ctrl_blk_init(srv_gid, /* local_hid */
ib_port_index, -1, /* port_index, numa_node_id */
0, 0, /* conn qps, use uc */
NULL, 0, -1, /* prealloc conn buf, conn buf size, key */
NUM_UD_QPS, BUF_SIZE, -1); /* num_dgram_qps, dgram_buf_size, key */
/* Buffer to receive requests into */
memset((void *) cb->dgram_buf, 0, BUF_SIZE);
/* Buffer to send responses from */
uint8_t *resp_buf = malloc(params.size);
assert(resp_buf != 0);
memset(resp_buf, 1, params.size);
/* Create an address handle for each client */
struct ibv_ah *ah[NUM_CLIENTS];
memset(ah, 0, NUM_CLIENTS * sizeof(uintptr_t));
struct hrd_qp_attr *clt_qp[NUM_CLIENTS];
/*
* Connect this server to NUM_CLIENTS clients whose global IDs are the
* same as this server's modulo 2. This ensures that the connected
* clients are on the same port as the server.
*/
for(i = 0; i < NUM_CLIENTS; i++) {
char clt_name[HRD_QP_NAME_SIZE];
/* ah[i] maps to client clt_id */
int clt_id = params.dual_port == 0 ? i : 2 * i + (srv_gid % 2);
sprintf(clt_name, "client-%d", clt_id);
/* Get the UD queue pair for the ith client */
clt_qp[i] = NULL;
while(clt_qp[i] == NULL) {
clt_qp[i] = hrd_get_published_qp(clt_name);
if(clt_qp[i] == NULL) {
usleep(200000);
}
}
printf("main: Server %d got client %d (clt_id = %d) of %d clients.\n",
srv_gid, i, clt_id, NUM_CLIENTS);
struct ibv_ah_attr ah_attr = {
.is_global = 0,
.dlid = clt_qp[i]->lid,
.sl = 0,
.src_path_bits = 0,
.port_num = cb->dev_port_id,
};
ah[i]= ibv_create_ah(cb->pd, &ah_attr);
assert(ah[i] != NULL);
}
struct ibv_send_wr wr[MAX_POSTLIST], *bad_send_wr;
struct ibv_wc wc[MAX_POSTLIST];
struct ibv_sge sgl[MAX_POSTLIST];
long long rolling_iter = 0; /* For throughput measurement */
long long nb_tx[NUM_UD_QPS] = {0}; /* For selective signaling */
int ud_qp_i = 0; /* Round-robin between QPs across postlists */
int w_i = 0; /* Window index */
int ret;
struct timespec start, end;
clock_gettime(CLOCK_REALTIME, &start);
while(1) {
if(rolling_iter >= M_4) {
clock_gettime(CLOCK_REALTIME, &end);
double seconds = (end.tv_sec - start.tv_sec) +
(double) (end.tv_nsec - start.tv_nsec) / 1000000000;
double my_tput = M_4 / seconds;
printf("main: Server %d: %.2f Mops. \n", srv_gid, my_tput);
params.tput[srv_gid] = my_tput;
if(srv_gid == 0) {
double total_tput = 0;
for(i = 0; i < params.num_threads; i++) {
total_tput += params.tput[i];
}
hrd_red_printf("main: Total tput = %.2f Mops.\n", total_tput);
}
rolling_iter = 0;
clock_gettime(CLOCK_REALTIME, &start);
}
for(w_i = 0; w_i < params.postlist; w_i++) {
int cn = nb_tx[ud_qp_i] & NUM_CLIENTS_;
wr[w_i].wr.ud.ah = ah[cn];
wr[w_i].wr.ud.remote_qpn = clt_qp[cn]->qpn;
wr[w_i].wr.ud.remote_qkey = HRD_DEFAULT_QKEY;
wr[w_i].opcode = IBV_WR_SEND;
wr[w_i].num_sge = 1;
wr[w_i].next = (w_i == params.postlist - 1) ? NULL : &wr[w_i + 1];
wr[w_i].sg_list = &sgl[w_i];
wr[w_i].send_flags = ((nb_tx[ud_qp_i] & UNSIG_BATCH_) == 0) ?
IBV_SEND_SIGNALED : 0;
if((nb_tx[ud_qp_i] & UNSIG_BATCH_) == 0 && nb_tx[ud_qp_i] > 0) {
hrd_poll_cq(cb->dgram_send_cq[ud_qp_i], 1, wc);
}
wr[w_i].send_flags |= IBV_SEND_INLINE;
sgl[w_i].addr = (uint64_t) (uintptr_t) resp_buf;
sgl[w_i].length = params.size;
nb_tx[ud_qp_i]++;
rolling_iter++;
}
ret = ibv_post_send(cb->dgram_qp[ud_qp_i], &wr[0], &bad_send_wr);
CPE(ret, "ibv_post_send error", ret);
/* Use a different QP for the next postlist */
ud_qp_i++;
if(ud_qp_i == NUM_UD_QPS) {
ud_qp_i = 0;
}
}
return NULL;
}
static int init_device(struct ibv_context *context_arg,
struct mca_btl_openib_sa_qp_cache *cache,
uint32_t port_num)
{
struct ibv_ah_attr aattr;
struct ibv_port_attr pattr;
int rc;
cache->context = ibv_open_device(context_arg->device);
if (NULL == cache->context) {
BTL_ERROR(("error obtaining device context for %s errno says %s",
ibv_get_device_name(context_arg->device), strerror(errno)));
return OPAL_ERROR;
}
cache->device_name = strdup(ibv_get_device_name(cache->context->device));
cache->port_num = port_num;
/* init all sl_values to be SL_NOT_PRESENT */
memset(&cache->sl_values, SL_NOT_PRESENT, sizeof(cache->sl_values));
cache->next = sa_qp_cache;
sa_qp_cache = cache;
/* allocate the protection domain for the device */
cache->pd = ibv_alloc_pd(cache->context);
if (NULL == cache->pd) {
BTL_ERROR(("error allocating protection domain for %s errno says %s",
ibv_get_device_name(context_arg->device), strerror(errno)));
return OPAL_ERROR;
}
/* register memory region */
cache->mr = ibv_reg_mr(cache->pd, cache->send_recv_buffer,
sizeof(cache->send_recv_buffer),
IBV_ACCESS_REMOTE_WRITE | IBV_ACCESS_LOCAL_WRITE);
if (NULL == cache->mr) {
BTL_ERROR(("error registering memory region, errno says %s", strerror(errno)));
return OPAL_ERROR;
}
/* init the ud qp */
rc = init_ud_qp(context_arg, cache);
if (OPAL_ERROR == rc) {
return OPAL_ERROR;
}
rc = ibv_query_port(cache->context, cache->port_num, &pattr);
if (rc) {
BTL_ERROR(("error getting port attributes for device %s "
"port number %d errno says %s",
ibv_get_device_name(context_arg->device),
cache->port_num, strerror(errno)));
return OPAL_ERROR;
}
/* create address handle */
memset(&aattr, 0, sizeof(aattr));
aattr.dlid = pattr.sm_lid;
aattr.sl = pattr.sm_sl;
aattr.port_num = cache->port_num;
cache->ah = ibv_create_ah(cache->pd, &aattr);
if (NULL == cache->ah) {
BTL_ERROR(("error creating address handle: %s", strerror(errno)));
return OPAL_ERROR;
}
memset(&(cache->rwr), 0, sizeof(cache->rwr));
cache->rwr.num_sge = 1;
cache->rwr.sg_list = &(cache->rsge);
memset(&(cache->rsge), 0, sizeof(cache->rsge));
cache->rsge.addr = (uint64_t)(void *)
(cache->send_recv_buffer + MAD_BLOCK_SIZE);
cache->rsge.length = MAD_BLOCK_SIZE + 40;
cache->rsge.lkey = cache->mr->lkey;
return 0;
}