int main(int argc, char *argv[]) {
int i = 0;
int size_max_pow = 24;
int ret_val;
struct report_options report;
struct pingpong_context ctx;
struct pingpong_dest *my_dest = NULL;
struct pingpong_dest *rem_dest = NULL;
struct mcast_parameters mcg_params;
struct ibv_device *ib_dev = NULL;
struct perftest_parameters user_param;
struct perftest_comm user_comm;
/* init default values to user's parameters */
memset(&ctx, 0, sizeof(struct pingpong_context));
memset(&user_param, 0, sizeof(struct perftest_parameters));
memset(&user_comm , 0, sizeof(struct perftest_comm));
memset(&mcg_params, 0, sizeof(struct mcast_parameters));
user_param.verb = SEND;
user_param.tst = LAT;
strncpy(user_param.version, VERSION, sizeof(user_param.version));
user_param.r_flag = &report;
// Configure the parameters values according to user arguments or defalut values.
ret_val = parser(&user_param,argv,argc);
if (ret_val) {
if (ret_val != VERSION_EXIT && ret_val != HELP_EXIT)
fprintf(stderr," Parser function exited with Error\n");
return 1;
}
if(user_param.use_xrc || user_param.connection_type == DC) {
user_param.num_of_qps *= 2;
}
//Checking that the user did not run with RawEth. for this we have raw_etherent_bw test.
if (user_param.connection_type == RawEth) {
fprintf(stderr," This test cannot run Raw Ethernet QPs (you have chosen RawEth as connection type\n");
return FAILURE;
}
// Finding the IB device selected (or defalut if no selected).
ib_dev = ctx_find_dev(user_param.ib_devname);
if (!ib_dev) {
fprintf(stderr," Unable to find the Infiniband/RoCE device\n");
return 1;
}
if (user_param.use_mcg)
GET_STRING(mcg_params.ib_devname,ibv_get_device_name(ib_dev));
// Getting the relevant context from the device
ctx.context = ibv_open_device(ib_dev);
if (!ctx.context) {
fprintf(stderr, " Couldn't get context for the device\n");
return 1;
}
// See if MTU and link type are valid and supported.
if (check_link(ctx.context,&user_param)) {
fprintf(stderr, " Couldn't get context for the device\n");
return FAILURE;
}
// copy the relevant user parameters to the comm struct + creating rdma_cm resources.
if (create_comm_struct(&user_comm,&user_param)) {
fprintf(stderr," Unable to create RDMA_CM resources\n");
return 1;
}
if (user_param.output == FULL_VERBOSITY && user_param.machine == SERVER) {
printf("\n************************************\n");
printf("* Waiting for client to connect... *\n");
printf("************************************\n");
}
// Initialize the connection and print the local data.
if (establish_connection(&user_comm)) {
fprintf(stderr," Unable to init the socket connection\n");
return FAILURE;
}
exchange_versions(&user_comm, &user_param);
check_sys_data(&user_comm, &user_param);
// See if MTU and link type are valid and supported.
if (check_mtu(ctx.context,&user_param, &user_comm)) {
fprintf(stderr, " Couldn't get context for the device\n");
return FAILURE;
}
// Print basic test information.
ctx_print_test_info(&user_param);
ALLOCATE(my_dest , struct pingpong_dest , user_param.num_of_qps);
memset(my_dest, 0, sizeof(struct pingpong_dest)*user_param.num_of_qps);
ALLOCATE(rem_dest , struct pingpong_dest , user_param.num_of_qps);
memset(rem_dest, 0, sizeof(struct pingpong_dest)*user_param.num_of_qps);
// Allocating arrays needed for the test.
alloc_ctx(&ctx,&user_param);
// Create (if nessacery) the rdma_cm ids and channel.
if (user_param.work_rdma_cm == ON) {
if (user_param.machine == CLIENT) {
if (retry_rdma_connect(&ctx,&user_param)) {
fprintf(stderr,"Unable to perform rdma_client function\n");
return FAILURE;
}
} else {
if (create_rdma_resources(&ctx,&user_param)) {
fprintf(stderr," Unable to create the rdma_resources\n");
return FAILURE;
}
if (rdma_server_connect(&ctx,&user_param)) {
fprintf(stderr,"Unable to perform rdma_client function\n");
return FAILURE;
}
}
} else {
// create all the basic IB resources (data buffer, PD, MR, CQ and events channel)
if (ctx_init(&ctx,&user_param)) {
fprintf(stderr, " Couldn't create IB resources\n");
return FAILURE;
}
}
// Set up the Connection.
if (send_set_up_connection(&ctx,&user_param,my_dest,&mcg_params,&user_comm)) {
fprintf(stderr," Unable to set up socket connection\n");
return 1;
}
for (i=0; i < user_param.num_of_qps; i++)
ctx_print_pingpong_data(&my_dest[i],&user_comm);
user_comm.rdma_params->side = REMOTE;
for (i=0; i < user_param.num_of_qps; i++) {
// shaking hands and gather the other side info.
if (ctx_hand_shake(&user_comm,&my_dest[i],&rem_dest[i])) {
fprintf(stderr,"Failed to exchange data between server and clients\n");
return 1;
}
ctx_print_pingpong_data(&rem_dest[i],&user_comm);
}
if (user_param.work_rdma_cm == OFF)
{
if (ctx_check_gid_compatibility(&my_dest[0], &rem_dest[0]))
{
fprintf(stderr,"\n Found Incompatibility issue with GID types.\n");
fprintf(stderr," Please Try to use a different IP version.\n\n");
return 1;
}
}
if (user_param.use_mcg) {
memcpy(mcg_params.base_mgid.raw,mcg_params.mgid.raw,16);
memcpy(mcg_params.mgid.raw,rem_dest[0].gid.raw,16);
mcg_params.base_mlid = mcg_params.mlid;
mcg_params.is_2nd_mgid_used = ON;
if (!strcmp(link_layer_str(user_param.link_type),"IB")) {
// Request for Mcast group create registery in SM.
if (join_multicast_group(SUBN_ADM_METHOD_SET,&mcg_params)) {
fprintf(stderr," Failed to Join Mcast request\n");
return 1;
}
}
/*
* The next stall in code (50 ms sleep) is a work around for fixing the
* the bug this test had in Multicast for the past 1 year.
* It appears, that when a switch involved, it takes ~ 10 ms for the join
* request to propogate on the IB fabric, thus we need to wait for it.
* what happened before this fix was reaching the post_send
* code segment in about 350 ns from here, and the switch(es) dropped
* the packet because join request wasn't finished.
*/
usleep(50000);
}
if (user_param.work_rdma_cm == OFF) {
// Prepare IB resources for rtr/rts.
if (ctx_connect(&ctx,rem_dest,&user_param,my_dest)) {
fprintf(stderr," Unable to Connect the HCA's through the link\n");
return 1;
}
}
// shaking hands and gather the other side info.
if (ctx_hand_shake(&user_comm,&my_dest[0],&rem_dest[0])) {
fprintf(stderr,"Failed to exchange data between server and clients\n");
return 1;
}
if (user_param.use_event) {
if (ibv_req_notify_cq(ctx.send_cq, 0)) {
fprintf(stderr, "Couldn't request RCQ notification\n");
return 1;
}
if (ibv_req_notify_cq(ctx.recv_cq, 0)) {
fprintf(stderr, "Couldn't request RCQ notification\n");
return 1;
}
}
if (user_param.output == FULL_VERBOSITY) {
printf(RESULT_LINE);
printf("%s",(user_param.test_type == ITERATIONS) ? RESULT_FMT_LAT : RESULT_FMT_LAT_DUR);
printf((user_param.cpu_util_data.enable ? RESULT_EXT_CPU_UTIL : RESULT_EXT));
}
ctx_set_send_wqes(&ctx,&user_param,rem_dest);
if (user_param.test_method == RUN_ALL) {
if (user_param.connection_type == UD)
size_max_pow = (int)UD_MSG_2_EXP(MTU_SIZE(user_param.curr_mtu)) + 1;
for (i = 1; i < size_max_pow ; ++i) {
user_param.size = (uint64_t)1 << i;
// Post recevie recv_wqes fo current message size
if (ctx_set_recv_wqes(&ctx,&user_param)) {
fprintf(stderr," Failed to post receive recv_wqes\n");
return 1;
}
// Sync between the client and server so the client won't send packets
// Before the server has posted his receive wqes (in UC/UD it will result in a deadlock).
if (ctx_hand_shake(&user_comm,&my_dest[0],&rem_dest[0])) {
fprintf(stderr,"Failed to exchange data between server and clients\n");
return 1;
}
if(run_iter_lat_send(&ctx, &user_param))
return 17;
user_param.test_type == ITERATIONS ? print_report_lat(&user_param) : print_report_lat_duration(&user_param);
}
} else {
// Post recevie recv_wqes fo current message size
if (ctx_set_recv_wqes(&ctx,&user_param)) {
fprintf(stderr," Failed to post receive recv_wqes\n");
return 1;
}
// Sync between the client and server so the client won't send packets
// Before the server has posted his receive wqes (in UC/UD it will result in a deadlock).
if (ctx_hand_shake(&user_comm,my_dest,rem_dest)) {
fprintf(stderr,"Failed to exchange data between server and clients\n");
return 1;
}
if(run_iter_lat_send(&ctx, &user_param))
return 17;
user_param.test_type == ITERATIONS ? print_report_lat(&user_param) : print_report_lat_duration(&user_param);
}
if (user_param.output == FULL_VERBOSITY) {
printf(RESULT_LINE);
}
if (ctx_close_connection(&user_comm,my_dest,rem_dest)) {
fprintf(stderr,"Failed to close connection between server and client\n");
fprintf(stderr," Trying to close this side resources\n");
}
return send_destroy_ctx(&ctx,&user_param,&mcg_params);
}
int main(int argc, char *argv[])
{
struct ibv_device *ib_dev = NULL;
struct pingpong_context ctx;
struct pingpong_dest *my_dest = NULL;
struct pingpong_dest *rem_dest = NULL;
struct perftest_parameters user_param;
struct perftest_comm user_comm;
struct mcast_parameters mcg_params;
struct bw_report_data my_bw_rep, rem_bw_rep;
int ret_parser,i = 0;
int size_max_pow = 24;
/* init default values to user's parameters */
memset(&ctx, 0,sizeof(struct pingpong_context));
memset(&user_param, 0 , sizeof(struct perftest_parameters));
memset(&mcg_params, 0 , sizeof(struct mcast_parameters));
memset(&user_comm, 0,sizeof(struct perftest_comm));
user_param.verb = SEND;
user_param.tst = BW;
strncpy(user_param.version, VERSION, sizeof(user_param.version));
/* Configure the parameters values according to user arguments or defalut values. */
ret_parser = parser(&user_param,argv,argc);
if (ret_parser) {
if (ret_parser != VERSION_EXIT && ret_parser != HELP_EXIT)
fprintf(stderr," Parser function exited with Error\n");
return 1;
}
if((user_param.connection_type == DC || user_param.use_xrc) && user_param.duplex) {
user_param.num_of_qps *= 2;
}
/* Checking that the user did not run with RawEth. for this we have raw_etherent_bw test. */
if (user_param.connection_type == RawEth) {
fprintf(stderr," This test cannot run Raw Ethernet QPs (you have chosen RawEth as connection type\n");
fprintf(stderr," For this we have raw_ethernet_bw test in this package.\n");
return FAILURE;
}
/* Finding the IB device selected (or defalut if no selected). */
ib_dev = ctx_find_dev(user_param.ib_devname);
if (!ib_dev) {
fprintf(stderr," Unable to find the Infiniband/RoCE device\n");
return 1;
}
if (user_param.use_mcg)
GET_STRING(mcg_params.ib_devname,ibv_get_device_name(ib_dev));
/* Getting the relevant context from the device */
ctx.context = ibv_open_device(ib_dev);
if (!ctx.context) {
fprintf(stderr, " Couldn't get context for the device\n");
return 1;
}
/* See if MTU and link type are valid and supported. */
if (check_link(ctx.context,&user_param)) {
fprintf(stderr, " Couldn't get context for the device\n");
return FAILURE;
}
/* copy the relevant user parameters to the comm struct + creating rdma_cm resources. */
if (create_comm_struct(&user_comm,&user_param)) {
fprintf(stderr," Unable to create RDMA_CM resources\n");
return 1;
}
if (user_param.output == FULL_VERBOSITY && user_param.machine == SERVER) {
printf("\n************************************\n");
printf("* Waiting for client to connect... *\n");
printf("************************************\n");
}
/* Initialize the connection and print the local data. */
if (establish_connection(&user_comm)) {
fprintf(stderr," Unable to init the socket connection\n");
return FAILURE;
}
exchange_versions(&user_comm, &user_param);
check_sys_data(&user_comm, &user_param);
/* See if MTU and link type are valid and supported. */
if (check_mtu(ctx.context,&user_param, &user_comm)) {
fprintf(stderr, " Couldn't get context for the device\n");
return FAILURE;
}
/* Print basic test information. */
ctx_print_test_info(&user_param);
ALLOCATE(my_dest , struct pingpong_dest , user_param.num_of_qps);
memset(my_dest, 0, sizeof(struct pingpong_dest)*user_param.num_of_qps);
ALLOCATE(rem_dest , struct pingpong_dest , user_param.num_of_qps);
memset(rem_dest, 0, sizeof(struct pingpong_dest)*user_param.num_of_qps);
if (user_param.transport_type == IBV_TRANSPORT_IWARP)
ctx.send_rcredit = 1;
/* Allocating arrays needed for the test. */
alloc_ctx(&ctx,&user_param);
/* Create (if nessacery) the rdma_cm ids and channel. */
if (user_param.work_rdma_cm == ON) {
if (user_param.machine == CLIENT) {
if (retry_rdma_connect(&ctx,&user_param)) {
fprintf(stderr,"Unable to perform rdma_client function\n");
return FAILURE;
}
} else {
if (create_rdma_resources(&ctx,&user_param)) {
fprintf(stderr," Unable to create the rdma_resources\n");
return FAILURE;
}
if (rdma_server_connect(&ctx,&user_param)) {
fprintf(stderr,"Unable to perform rdma_client function\n");
return FAILURE;
}
}
} else {
/* create all the basic IB resources (data buffer, PD, MR, CQ and events channel) */
if (ctx_init(&ctx,&user_param)) {
fprintf(stderr, " Couldn't create IB resources\n");
return FAILURE;
}
}
/* Set up the Connection. */
if (send_set_up_connection(&ctx,&user_param,my_dest,&mcg_params,&user_comm)) {
fprintf(stderr," Unable to set up socket connection\n");
return 1;
}
if (ctx.send_rcredit)
ctx_alloc_credit(&ctx,&user_param,my_dest);
for (i=0; i < user_param.num_of_qps; i++)
ctx_print_pingpong_data(&my_dest[i],&user_comm);
user_comm.rdma_params->side = REMOTE;
for (i=0; i < user_param.num_of_qps; i++) {
/* shaking hands and gather the other side info. */
if (ctx_hand_shake(&user_comm,&my_dest[i],&rem_dest[i])) {
fprintf(stderr,"Failed to exchange data between server and clients\n");
return 1;
}
ctx_print_pingpong_data(&rem_dest[i],&user_comm);
}
if (user_param.work_rdma_cm == OFF) {
if (ctx_check_gid_compatibility(&my_dest[0], &rem_dest[0])) {
fprintf(stderr,"\n Found Incompatibility issue with GID types.\n");
fprintf(stderr," Please Try to use a different IP version.\n\n");
return 1;
}
}
/* If credit for available recieve buffers is necessary,
* the credit sending is done via RDMA WRITE ops and the ctx_hand_shake above
* is used to exchange the rkeys and buf addresses for the RDMA WRITEs
*/
if (ctx.send_rcredit)
ctx_set_credit_wqes(&ctx,&user_param,rem_dest);
/* Joining the Send side port the Mcast gid */
if (user_param.use_mcg && (user_param.machine == CLIENT || user_param.duplex)) {
memcpy(mcg_params.mgid.raw, rem_dest[0].gid.raw, 16);
if (set_mcast_group(&ctx,&user_param,&mcg_params)) {
fprintf(stderr," Unable to Join Sender to Mcast gid\n");
return 1;
}
/*
* The next stall in code (50 ms sleep) is a work around for fixing the
* the bug this test had in Multicast for the past 1 year.
* It appears, that when a switch involved, it takes ~ 10 ms for the join
* request to propogate on the IB fabric, thus we need to wait for it.
* what happened before this fix was client reaching the post_send
* code segment in about 350 ns from here, and the switch(es) dropped
* the packet because join request wasn't finished.
*/
usleep(50000);
}
if (user_param.work_rdma_cm == OFF) {
/* Prepare IB resources for rtr/rts. */
if (ctx_connect(&ctx,rem_dest,&user_param,my_dest)) {
fprintf(stderr," Unable to Connect the HCA's through the link\n");
return 1;
}
}
/* shaking hands and gather the other side info. */
if (ctx_hand_shake(&user_comm,&my_dest[0],&rem_dest[0])) {
fprintf(stderr,"Failed to exchange data between server and clients\n");
return 1;
}
if (user_param.use_event) {
if (ibv_req_notify_cq(ctx.send_cq, 0)) {
fprintf(stderr, " Couldn't request CQ notification\n");
return 1;
}
if (ibv_req_notify_cq(ctx.recv_cq, 0)) {
fprintf(stderr, " Couldn't request CQ notification\n");
return 1;
}
}
if (user_param.output == FULL_VERBOSITY) {
if (user_param.report_per_port) {
printf(RESULT_LINE_PER_PORT);
printf((user_param.report_fmt == MBS ? RESULT_FMT_PER_PORT : RESULT_FMT_G_PER_PORT));
}
else {
printf(RESULT_LINE);
printf((user_param.report_fmt == MBS ? RESULT_FMT : RESULT_FMT_G));
}
printf((user_param.cpu_util_data.enable ? RESULT_EXT_CPU_UTIL : RESULT_EXT));
}
if (user_param.test_method == RUN_ALL) {
if (user_param.connection_type == UD)
size_max_pow = (int)UD_MSG_2_EXP(MTU_SIZE(user_param.curr_mtu)) + 1;
for (i = 1; i < size_max_pow ; ++i) {
user_param.size = (uint64_t)1 << i;
if (user_param.machine == CLIENT || user_param.duplex)
ctx_set_send_wqes(&ctx,&user_param,rem_dest);
if (user_param.machine == SERVER || user_param.duplex) {
if (ctx_set_recv_wqes(&ctx,&user_param)) {
fprintf(stderr," Failed to post receive recv_wqes\n");
return 1;
}
}
if (ctx_hand_shake(&user_comm,&my_dest[0],&rem_dest[0])) {
fprintf(stderr,"Failed to exchange data between server and clients\n");
return 1;
}
if (ctx.send_rcredit) {
int j;
for (j = 0; j < user_param.num_of_qps; j++)
ctx.credit_buf[j] = 0;
}
if (user_param.duplex) {
if(run_iter_bi(&ctx,&user_param))
return 17;
} else if (user_param.machine == CLIENT) {
if(run_iter_bw(&ctx,&user_param)) {
return 17;
}
} else {
if(run_iter_bw_server(&ctx,&user_param)) {
return 17;
}
}
print_report_bw(&user_param,&my_bw_rep);
if (user_param.duplex && user_param.test_type != DURATION) {
xchg_bw_reports(&user_comm, &my_bw_rep,&rem_bw_rep,atof(user_param.rem_version));
print_full_bw_report(&user_param, &my_bw_rep, &rem_bw_rep);
}
if (ctx_hand_shake(&user_comm,&my_dest[0],&rem_dest[0])) {
fprintf(stderr,"Failed to exchange data between server and clients\n");
return 1;
}
/* Check if last iteration ended well in UC/UD */
if (user_param.check_alive_exited) {
break;
}
}
} else if (user_param.test_method == RUN_REGULAR) {
if (user_param.machine == CLIENT || user_param.duplex)
ctx_set_send_wqes(&ctx,&user_param,rem_dest);
if (user_param.machine == SERVER || user_param.duplex) {
if (ctx_set_recv_wqes(&ctx,&user_param)) {
fprintf(stderr," Failed to post receive recv_wqes\n");
return 1;
}
}
if (ctx_hand_shake(&user_comm,&my_dest[0],&rem_dest[0])) {
fprintf(stderr,"Failed to exchange data between server and clients\n");
return 1;
}
if (user_param.duplex) {
if(run_iter_bi(&ctx,&user_param))
return 17;
} else if (user_param.machine == CLIENT) {
if(run_iter_bw(&ctx,&user_param)) {
return 17;
}
} else if(run_iter_bw_server(&ctx,&user_param)) {
return 17;
}
print_report_bw(&user_param,&my_bw_rep);
if (user_param.duplex && user_param.test_type != DURATION) {
xchg_bw_reports(&user_comm, &my_bw_rep,&rem_bw_rep,atof(user_param.rem_version));
print_full_bw_report(&user_param, &my_bw_rep, &rem_bw_rep);
}
if (user_param.report_both && user_param.duplex) {
printf(RESULT_LINE);
printf("\n Local results: \n");
printf(RESULT_LINE);
printf((user_param.report_fmt == MBS ? RESULT_FMT : RESULT_FMT_G));
printf((user_param.cpu_util_data.enable ? RESULT_EXT_CPU_UTIL : RESULT_EXT));
print_full_bw_report(&user_param, &my_bw_rep, NULL);
printf(RESULT_LINE);
printf("\n Remote results: \n");
printf(RESULT_LINE);
printf((user_param.report_fmt == MBS ? RESULT_FMT : RESULT_FMT_G));
printf((user_param.cpu_util_data.enable ? RESULT_EXT_CPU_UTIL : RESULT_EXT));
print_full_bw_report(&user_param, &rem_bw_rep, NULL);
}
} else if (user_param.test_method == RUN_INFINITELY) {
if (user_param.machine == CLIENT)
ctx_set_send_wqes(&ctx,&user_param,rem_dest);
else if (user_param.machine == SERVER) {
if (ctx_set_recv_wqes(&ctx,&user_param)) {
fprintf(stderr," Failed to post receive recv_wqes\n");
return 1;
}
}
if (ctx_hand_shake(&user_comm,&my_dest[0],&rem_dest[0])) {
fprintf(stderr,"Failed to exchange data between server and clients\n");
return 1;
}
if (user_param.machine == CLIENT) {
if(run_iter_bw_infinitely(&ctx,&user_param)) {
fprintf(stderr," Error occured while running infinitely! aborting ...\n");
return 1;
}
} else if (user_param.machine == SERVER) {
if(run_iter_bw_infinitely_server(&ctx,&user_param)) {
fprintf(stderr," Error occured while running infinitely on server! aborting ...\n");
return 1;
}
}
}
if (user_param.output == FULL_VERBOSITY) {
if (user_param.report_per_port)
printf(RESULT_LINE_PER_PORT);
else
printf(RESULT_LINE);
}
if (ctx_close_connection(&user_comm,&my_dest[0],&rem_dest[0])) {
fprintf(stderr," Failed to close connection between server and client\n");
fprintf(stderr," Trying to close this side resources\n");
}
/* Destory all test resources, including Mcast if exists */
if (send_destroy_ctx(&ctx,&user_param,&mcg_params)) {
fprintf(stderr,"Couldn't Destory all SEND resources\n");
return FAILURE;
}
if (user_param.work_rdma_cm == ON) {
user_comm.rdma_params->work_rdma_cm = ON;
if (destroy_ctx(user_comm.rdma_ctx,user_comm.rdma_params)) {
fprintf(stderr,"Failed to destroy resources\n");
return 1;
}
}
if (!user_param.is_bw_limit_passed && (user_param.is_limit_bw == ON ) ) {
fprintf(stderr,"Error: BW result is below bw limit\n");
return 1;
}
if (!user_param.is_msgrate_limit_passed && (user_param.is_limit_bw == ON )) {
fprintf(stderr,"Error: Msg rate is below msg_rate limit\n");
return 1;
}
return 0;
}
/*
* Main function. implements raw_ethernet_send_lat
*/
int main(int argc, char *argv[])
{
struct ibv_device *ib_dev = NULL;
struct pingpong_context ctx;
struct raw_ethernet_info my_dest_info,rem_dest_info;
int ret_parser;
struct perftest_parameters user_param;
#ifdef HAVE_RAW_ETH_EXP
struct ibv_exp_flow *flow_create_result = NULL;
struct ibv_exp_flow_attr *flow_rules = NULL;
struct ibv_exp_flow *flow_promisc = NULL;
#else
struct ibv_flow *flow_create_result = NULL;
struct ibv_flow_attr *flow_rules = NULL;
#endif
struct report_options report;
//allocate memory space for user parameters
memset(&ctx, 0, sizeof(struct pingpong_context));
memset(&user_param, 0, sizeof(struct perftest_parameters));
memset(&my_dest_info, 0 , sizeof(struct raw_ethernet_info));
memset(&rem_dest_info, 0 , sizeof(struct raw_ethernet_info));
/* init default values to user's parameters that's relvant for this test:
* Raw Ethernet Send Latency Test
*/
user_param.verb = SEND;
user_param.tst = LAT;
strncpy(user_param.version, VERSION, sizeof(user_param.version));
user_param.connection_type = RawEth;
user_param.r_flag = &report;
if (check_flow_steering_support()) {
return 1;
}
/* Configure the parameters values according to user
arguments or default values. */
ret_parser = parser(&user_param, argv,argc);
//check for parsing errors
if (ret_parser) {
if (ret_parser != VERSION_EXIT && ret_parser != HELP_EXIT)
fprintf(stderr," Parser function exited with Error\n");
DEBUG_LOG(TRACE,"<<<<<<%s",__FUNCTION__);
return 1;
}
//this is a bidirectional test, so we need to let the init functions
//to think we are in duplex mode
//TODO: ask Ido if that's ok, or should I add another field in user_param
user_param.duplex = 1;
// Find the selected IB device (or default if the user didn't select one).
ib_dev = ctx_find_dev(user_param.ib_devname);
if (!ib_dev) {
fprintf(stderr," Unable to find the Infiniband/RoCE device\n");
DEBUG_LOG(TRACE,"<<<<<<%s",__FUNCTION__);
return 1;
}
// Getting the relevant context from the device
ctx.context = ibv_open_device(ib_dev);
if (!ctx.context) {
fprintf(stderr, " Couldn't get context for the device\n");
DEBUG_LOG(TRACE,"<<<<<<%s",__FUNCTION__);
return 1;
}
// See if MTU and link type are valid and supported.
if (check_link_and_mtu(ctx.context, &user_param)) {
fprintf(stderr, " Couldn't get context for the device\n");
DEBUG_LOG(TRACE,"<<<<<<%s",__FUNCTION__);
return FAILURE;
}
// Allocating arrays needed for the test.
alloc_ctx(&ctx, &user_param);
// Print basic test information.
ctx_print_test_info(&user_param);
//set up the connection, return the required flow rules (notice that user_param->duplex == TRUE)
//so the function will setup like it's a bidirectional test
if (send_set_up_connection(&flow_rules, &ctx, &user_param, &my_dest_info, &rem_dest_info)) {
fprintf(stderr," Unable to set up socket connection\n");
return 1;
}
//print specifications of the test
print_spec(flow_rules,&user_param);
// Create (if necessary) the rdma_cm ids and channel.
if (user_param.work_rdma_cm == ON) {
//create resources
if (create_rdma_resources(&ctx, &user_param)) {
fprintf(stderr," Unable to create the rdma_resources\n");
return FAILURE;
}
if (user_param.machine == CLIENT) {
//Connects the client to a QP on the other machine with rdma_cm
if (rdma_client_connect(&ctx, &user_param)) {
fprintf(stderr,"Unable to perform rdma_client function\n");
return FAILURE;
}
} else if (rdma_server_connect(&ctx, &user_param)) {
//Assigning a server to listen on rdma_cm port and connect to it.
fprintf(stderr,"Unable to perform rdma_server function\n");
return FAILURE;
}
} else {
// initalize IB resources (data buffer, PD, MR, CQ and events channel)
if (ctx_init(&ctx, &user_param)) {
fprintf(stderr, " Couldn't create IB resources\n");
return FAILURE;
}
}
//attaching the qp to the spec
#ifdef HAVE_RAW_ETH_EXP
flow_create_result = ibv_exp_create_flow(ctx.qp[0], flow_rules);
#else
flow_create_result = ibv_create_flow(ctx.qp[0], flow_rules);
#endif
if (!flow_create_result){
perror("error");
fprintf(stderr, "Couldn't attach QP\n");
return FAILURE;
}
#ifdef HAVE_RAW_ETH_EXP
if (user_param.use_promiscuous) {
struct ibv_exp_flow_attr attr = {
.type = IBV_EXP_FLOW_ATTR_ALL_DEFAULT,
.num_of_specs = 0,
.port = user_param.ib_port,
.flags = 0
};
if ((flow_promisc = ibv_exp_create_flow(ctx.qp[0], &attr)) == NULL) {
perror("error");
fprintf(stderr, "Couldn't attach promiscous rule QP\n");
}
}
#endif
//build ONE Raw Ethernet packets on ctx buffer
create_raw_eth_pkt(&user_param,&ctx, &my_dest_info , &rem_dest_info);
if (user_param.output == FULL_VERBOSITY) {
printf(RESULT_LINE);
printf("%s",(user_param.test_type == ITERATIONS) ? RESULT_FMT_LAT : RESULT_FMT_LAT_DUR);
printf((user_param.cpu_util_data.enable ? RESULT_EXT_CPU_UTIL : RESULT_EXT));
}
// Prepare IB resources for rtr(ready to read)/rts(ready to send)
if (user_param.work_rdma_cm == OFF) {
if (ctx_connect(&ctx, NULL, &user_param, NULL)) {
fprintf(stderr," Unable to Connect the HCA's through the link\n");
DEBUG_LOG(TRACE,"<<<<<<%s",__FUNCTION__);
return 1;
}
}
//Post Send send_wqes for current message size
ctx_set_send_wqes(&ctx,&user_param,NULL);
// Post receive recv_wqes for current message size
if (ctx_set_recv_wqes(&ctx,&user_param)) {
fprintf(stderr," Failed to post receive recv_wqes\n");
return 1;
}
//latency test function for SEND verb latency test.
if (run_iter_lat_send(&ctx, &user_param))
{
return 17;
}
//print report (like print_report_bw) in the correct format
// (as set before: FMT_LAT or FMT_LAT_DUR)
user_param.test_type == ITERATIONS ? print_report_lat(&user_param) :
print_report_lat_duration(&user_param);
//destory promisc flow
#ifdef HAVE_RAW_ETH_EXP
if (user_param.use_promiscuous) {
if (ibv_exp_destroy_flow(flow_promisc)) {
perror("error");
fprintf(stderr, "Couldn't Destory promisc flow\n");
return FAILURE;
}
}
#endif
//destroy flow
#ifdef HAVE_RAW_ETH_EXP
if (ibv_exp_destroy_flow(flow_create_result)) {
#else
if (ibv_destroy_flow(flow_create_result)) {
#endif
perror("error");
fprintf(stderr, "Couldn't Destory flow\n");
return FAILURE;
}
free(flow_rules);
//Deallocate all perftest resources.
if (destroy_ctx(&ctx, &user_param)) {
fprintf(stderr,"Failed to destroy_ctx\n");
DEBUG_LOG(TRACE,"<<<<<<%s",__FUNCTION__);
return 1;
}
if (user_param.output == FULL_VERBOSITY)
printf(RESULT_LINE);
DEBUG_LOG(TRACE,"<<<<<<%s",__FUNCTION__);
return 0;
}
/******************************************************************************
+ *
+ ******************************************************************************/
int rdma_server_connect(struct pingpong_context *ctx,
struct perftest_parameters *user_param)
{
int temp;
struct addrinfo *res;
struct rdma_cm_event *event;
struct rdma_conn_param conn_param;
struct addrinfo hints;
char *service;
struct sockaddr_in sin;
memset(&hints, 0, sizeof hints);
hints.ai_flags = AI_PASSIVE;
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 = 0;
sin.sin_family = PF_INET;
sin.sin_port = htons((unsigned short)user_param->port);
if (rdma_bind_addr(ctx->cm_id_control,(struct sockaddr *)&sin)) {
fprintf(stderr," rdma_bind_addr failed\n");
return 1;
}
if (rdma_listen(ctx->cm_id_control,0)) {
fprintf(stderr, "rdma_listen failed\n");
return 1;
}
if (rdma_get_cm_event(ctx->cm_channel,&event)) {
fprintf(stderr, "rdma_get_cm_events failed\n");
return 1;
}
if (event->event != RDMA_CM_EVENT_CONNECT_REQUEST) {
fprintf(stderr, "bad event waiting for connect request %d\n",event->event);
return 1;
}
ctx->cm_id = (struct rdma_cm_id*)event->id;
ctx->context = ctx->cm_id->verbs;
if (user_param->work_rdma_cm == ON)
alloc_ctx(ctx,user_param);
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;
}
if (user_param->connection_type == UD)
conn_param.qp_num = ctx->qp[0]->qp_num;
conn_param.retry_count = user_param->retry_count;
conn_param.rnr_retry_count = 7;
user_param->work_rdma_cm = temp;
if (user_param->work_rdma_cm == OFF) {
if (post_one_recv_wqe(ctx)) {
fprintf(stderr, "Couldn't post send \n");
return 1;
}
} else if (user_param->connection_type == UD) {
if (user_param->tst == LAT || (user_param->tst == BW && user_param->duplex)) {
if (post_recv_to_get_ah(ctx)) {
fprintf(stderr, "Couldn't post send \n");
return 1;
}
}
}
if (rdma_accept(ctx->cm_id, &conn_param)) {
fprintf(stderr, "Function rdma_accept failed\n");
return 1;
}
if (user_param->work_rdma_cm && user_param->connection_type == UD) {
if (user_param->tst == LAT || (user_param->tst == BW && user_param->duplex)) {
if (create_ah_from_wc_recv(ctx,user_param)) {
fprintf(stderr, "Unable to create AH from WC\n");
return 1;
}
}
}
rdma_ack_cm_event(event);
rdma_destroy_id(ctx->cm_id_control);
freeaddrinfo(res);
return 0;
}
int main(int argc, char *argv[])
{
int ret_parser,i = 0;
struct ibv_device *ib_dev = NULL;
struct pingpong_context ctx;
struct pingpong_dest *my_dest = NULL;
struct pingpong_dest *rem_dest = NULL;
struct perftest_parameters user_param;
struct perftest_comm user_comm;
struct bw_report_data my_bw_rep, rem_bw_rep;
/* init default values to user's parameters */
memset(&ctx,0,sizeof(struct pingpong_context));
memset(&user_param , 0 , sizeof(struct perftest_parameters));
memset(&user_comm,0,sizeof(struct perftest_comm));
user_param.verb = READ;
user_param.tst = BW;
strncpy(user_param.version, VERSION, sizeof(user_param.version));
ret_parser = parser(&user_param,argv,argc);
if (ret_parser) {
if (ret_parser != VERSION_EXIT && ret_parser != HELP_EXIT)
fprintf(stderr," Parser function exited with Error\n");
return 1;
}
if((user_param.connection_type == DC || user_param.use_xrc) && user_param.duplex) {
user_param.num_of_qps *= 2;
}
ib_dev =ctx_find_dev(user_param.ib_devname);
if (!ib_dev)
return 7;
/* Getting the relevant context from the device */
ctx.context = ibv_open_device(ib_dev);
if (!ctx.context) {
fprintf(stderr, " Couldn't get context for the device\n");
return 1;
}
/* See if MTU and link type are valid and supported. */
if (check_link(ctx.context,&user_param)) {
fprintf(stderr, " Couldn't get context for the device\n");
return FAILURE;
}
/* copy the relevant user parameters to the comm struct + creating rdma_cm resources. */
if (create_comm_struct(&user_comm,&user_param)) {
fprintf(stderr," Unable to create RDMA_CM resources\n");
return 1;
}
if (user_param.output == FULL_VERBOSITY && user_param.machine == SERVER) {
printf("\n************************************\n");
printf("* Waiting for client to connect... *\n");
printf("************************************\n");
}
/* Initialize the connection and print the local data. */
if (establish_connection(&user_comm)) {
fprintf(stderr," Unable to init the socket connection\n");
return FAILURE;
}
exchange_versions(&user_comm, &user_param);
check_sys_data(&user_comm, &user_param);
/* See if MTU and link type are valid and supported. */
if (check_mtu(ctx.context,&user_param, &user_comm)) {
fprintf(stderr, " Couldn't get context for the device\n");
return FAILURE;
}
/* Print basic test information. */
ctx_print_test_info(&user_param);
ALLOCATE(my_dest , struct pingpong_dest , user_param.num_of_qps);
memset(my_dest, 0, sizeof(struct pingpong_dest)*user_param.num_of_qps);
ALLOCATE(rem_dest , struct pingpong_dest , user_param.num_of_qps);
memset(rem_dest, 0, sizeof(struct pingpong_dest)*user_param.num_of_qps);
/* Allocating arrays needed for the test. */
alloc_ctx(&ctx,&user_param);
/* Create (if nessacery) the rdma_cm ids and channel. */
if (user_param.work_rdma_cm == ON) {
if (user_param.machine == CLIENT) {
if (retry_rdma_connect(&ctx,&user_param)) {
fprintf(stderr,"Unable to perform rdma_client function\n");
return FAILURE;
}
} else {
if (create_rdma_resources(&ctx,&user_param)) {
fprintf(stderr," Unable to create the rdma_resources\n");
return FAILURE;
}
if (rdma_server_connect(&ctx,&user_param)) {
fprintf(stderr,"Unable to perform rdma_client function\n");
return FAILURE;
}
}
} else {
/* create all the basic IB resources. */
if (ctx_init(&ctx,&user_param)) {
fprintf(stderr, " Couldn't create IB resources\n");
return FAILURE;
}
}
/* Set up the Connection. */
if (set_up_connection(&ctx,&user_param,my_dest)) {
fprintf(stderr," Unable to set up socket connection\n");
return FAILURE;
}
/* Print this machine QP information */
for (i=0; i < user_param.num_of_qps; i++)
ctx_print_pingpong_data(&my_dest[i],&user_comm);
user_comm.rdma_params->side = REMOTE;
for (i=0; i < user_param.num_of_qps; i++) {
/* shaking hands and gather the other side info. */
if (ctx_hand_shake(&user_comm,&my_dest[i],&rem_dest[i])) {
fprintf(stderr,"Failed to exchange data between server and clients\n");
return 1;
}
ctx_print_pingpong_data(&rem_dest[i],&user_comm);
}
if (user_param.work_rdma_cm == OFF) {
if (ctx_check_gid_compatibility(&my_dest[0], &rem_dest[0])) {
fprintf(stderr,"\n Found Incompatibility issue with GID types.\n");
fprintf(stderr," Please Try to use a different IP version.\n\n");
return 1;
}
}
if (user_param.work_rdma_cm == OFF) {
if (ctx_connect(&ctx,rem_dest,&user_param,my_dest)) {
fprintf(stderr," Unable to Connect the HCA's through the link\n");
return 1;
}
}
/* An additional handshake is required after moving qp to RTR. */
if (ctx_hand_shake(&user_comm,&my_dest[0],&rem_dest[0])) {
fprintf(stderr,"Failed to exchange data between server and clients\n");
return 1;
}
if (user_param.output == FULL_VERBOSITY) {
if (user_param.report_per_port) {
printf(RESULT_LINE_PER_PORT);
printf((user_param.report_fmt == MBS ? RESULT_FMT_PER_PORT : RESULT_FMT_G_PER_PORT));
}
else {
printf(RESULT_LINE);
printf((user_param.report_fmt == MBS ? RESULT_FMT : RESULT_FMT_G));
}
printf((user_param.cpu_util_data.enable ? RESULT_EXT_CPU_UTIL : RESULT_EXT));
}
/* For half duplex tests, server just waits for client to exit */
if (user_param.machine == SERVER && !user_param.duplex) {
if (ctx_hand_shake(&user_comm,&my_dest[0],&rem_dest[0])) {
fprintf(stderr," Failed to exchange data between server and clients\n");
return FAILURE;
}
xchg_bw_reports(&user_comm, &my_bw_rep,&rem_bw_rep,atof(user_param.rem_version));
print_full_bw_report(&user_param, &rem_bw_rep, NULL);
if (ctx_close_connection(&user_comm,&my_dest[0],&rem_dest[0])) {
fprintf(stderr,"Failed to close connection between server and client\n");
return 1;
}
if (user_param.output == FULL_VERBOSITY) {
if (user_param.report_per_port)
printf(RESULT_LINE_PER_PORT);
else
printf(RESULT_LINE);
}
if (user_param.work_rdma_cm == ON) {
if (destroy_ctx(&ctx,&user_param)) {
fprintf(stderr, "Failed to destroy resources\n");
return 1;
}
user_comm.rdma_params->work_rdma_cm = ON;
return destroy_ctx(user_comm.rdma_ctx,user_comm.rdma_params);
}
return destroy_ctx(&ctx,&user_param);
}
if (user_param.use_event) {
if (ibv_req_notify_cq(ctx.send_cq, 0)) {
fprintf(stderr, "Couldn't request CQ notification\n");
return 1;
}
}
if (user_param.test_method == RUN_ALL) {
for (i = 1; i < 24 ; ++i) {
user_param.size = (uint64_t)1 << i;
ctx_set_send_wqes(&ctx,&user_param,rem_dest);
if(perform_warm_up(&ctx,&user_param)) {
fprintf(stderr,"Problems with warm up\n");
return 1;
}
if(user_param.duplex) {
if (ctx_hand_shake(&user_comm,&my_dest[0],&rem_dest[0])) {
fprintf(stderr,"Failed to sync between server and client between different msg sizes\n");
return 1;
}
}
if(run_iter_bw(&ctx,&user_param))
return 17;
if (user_param.duplex && (atof(user_param.version) >= 4.6)) {
if (ctx_hand_shake(&user_comm,&my_dest[0],&rem_dest[0])) {
fprintf(stderr,"Failed to sync between server and client between different msg sizes\n");
return 1;
}
}
print_report_bw(&user_param,&my_bw_rep);
if (user_param.duplex) {
xchg_bw_reports(&user_comm, &my_bw_rep,&rem_bw_rep,atof(user_param.rem_version));
print_full_bw_report(&user_param, &my_bw_rep, &rem_bw_rep);
}
}
} else if (user_param.test_method == RUN_REGULAR) {
ctx_set_send_wqes(&ctx,&user_param,rem_dest);
if(perform_warm_up(&ctx,&user_param)) {
fprintf(stderr,"Problems with warm up\n");
return 1;
}
if(user_param.duplex) {
if (ctx_hand_shake(&user_comm,&my_dest[0],&rem_dest[0])) {
fprintf(stderr,"Failed to sync between server and client between different msg sizes\n");
return 1;
}
}
if(run_iter_bw(&ctx,&user_param)) {
fprintf(stderr," Failed to complete run_iter_bw function successfully\n");
return 1;
}
print_report_bw(&user_param,&my_bw_rep);
if (user_param.duplex) {
xchg_bw_reports(&user_comm, &my_bw_rep,&rem_bw_rep,atof(user_param.rem_version));
print_full_bw_report(&user_param, &my_bw_rep, &rem_bw_rep);
}
if (user_param.report_both && user_param.duplex) {
printf(RESULT_LINE);
printf("\n Local results: \n");
printf(RESULT_LINE);
printf((user_param.report_fmt == MBS ? RESULT_FMT : RESULT_FMT_G));
printf((user_param.cpu_util_data.enable ? RESULT_EXT_CPU_UTIL : RESULT_EXT));
print_full_bw_report(&user_param, &my_bw_rep, NULL);
printf(RESULT_LINE);
printf("\n Remote results: \n");
printf(RESULT_LINE);
printf((user_param.report_fmt == MBS ? RESULT_FMT : RESULT_FMT_G));
printf((user_param.cpu_util_data.enable ? RESULT_EXT_CPU_UTIL : RESULT_EXT));
print_full_bw_report(&user_param, &rem_bw_rep, NULL);
}
} else if (user_param.test_method == RUN_INFINITELY) {
ctx_set_send_wqes(&ctx,&user_param,rem_dest);
if(run_iter_bw_infinitely(&ctx,&user_param)) {
fprintf(stderr," Error occured while running! aborting ...\n");
return 1;
}
}
if (user_param.output == FULL_VERBOSITY) {
if (user_param.report_per_port)
printf(RESULT_LINE_PER_PORT);
else
printf(RESULT_LINE);
}
/* For half duplex tests, server just waits for client to exit */
if (user_param.machine == CLIENT && !user_param.duplex) {
if (ctx_hand_shake(&user_comm,&my_dest[0],&rem_dest[0])) {
fprintf(stderr," Failed to exchange data between server and clients\n");
return FAILURE;
}
xchg_bw_reports(&user_comm, &my_bw_rep,&rem_bw_rep,atof(user_param.rem_version));
}
if (ctx_close_connection(&user_comm,&my_dest[0],&rem_dest[0])) {
fprintf(stderr,"Failed to close connection between server and client\n");
return 1;
}
if (!user_param.is_bw_limit_passed && (user_param.is_limit_bw == ON ) ) {
fprintf(stderr,"Error: BW result is below bw limit\n");
return 1;
}
if (!user_param.is_msgrate_limit_passed && (user_param.is_limit_bw == ON )) {
fprintf(stderr,"Error: Msg rate is below msg_rate limit\n");
return 1;
}
if (user_param.work_rdma_cm == ON) {
if (destroy_ctx(&ctx,&user_param)) {
fprintf(stderr, "Failed to destroy resources\n");
return 1;
}
user_comm.rdma_params->work_rdma_cm = ON;
return destroy_ctx(user_comm.rdma_ctx,user_comm.rdma_params);
}
return destroy_ctx(&ctx,&user_param);
}
