int main(int argc, char *argv[]) {
struct ibv_device *ib_dev;
struct pingpong_context ctx;
struct pingpong_dest *my_dest,*rem_dest;
struct perftest_parameters user_param;
struct perftest_comm user_comm;
int i = 0;
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 = WRITE;
user_param.tst = BW;
user_param.spec = PL;
user_param.version = VERSION;
// Configure the parameters values according to user arguments or defalut values.
if (parser(&user_param,argv,argc)) {
fprintf(stderr," Parser function exited with Error\n");
return 1;
}
// 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 deivce\n");
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");
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");
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);
// copy the rellevant 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;
}
// Create (if nessacery) the rdma_cm ids and channel.
if (user_param.work_rdma_cm == ON) {
if (create_rdma_resources(&ctx,&user_param)) {
fprintf(stderr," Unable to create the rdma_resources\n");
return FAILURE;
}
if (user_param.machine == CLIENT) {
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)) {
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 (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);
// Init the connection and print the local data.
if (establish_connection(&user_comm)) {
fprintf(stderr," Unable to init the socket connection\n");
return FAILURE;
}
// shaking hands and gather the other side info.
for (i=0; i < user_param.num_of_qps; i++) {
if (ctx_hand_shake(&user_comm,&my_dest[i],&rem_dest[i])) {
fprintf(stderr," Failed to exchange date between server and clients\n");
return 1;
}
// Print remote machine QP information
user_comm.rdma_params->side = REMOTE;
ctx_print_pingpong_data(&rem_dest[i],&user_comm);
if (user_param.work_rdma_cm == OFF) {
if (pp_connect_ctx(&ctx,my_dest[i].psn,&rem_dest[i],&user_param,i)) {
fprintf(stderr," Unable to Connect the HCA's through the link\n");
return FAILURE;
}
}
// An additional handshake is required after moving qp to RTR.
if (ctx_hand_shake(&user_comm,&my_dest[i],&rem_dest[i])) {
fprintf(stderr," Failed to exchange date between server and clients\n");
return FAILURE;
}
}
printf(RESULT_LINE);
printf(RESULT_FMT);
// For half duplex tests, server just waits for client to exit
if (user_param.machine == SERVER && !user_param.duplex) {
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;
}
printf(RESULT_LINE);
return 0;
}
ALLOCATE(tposted,cycles_t,user_param.iters*user_param.num_of_qps);
ALLOCATE(tcompleted,cycles_t,user_param.iters*user_param.num_of_qps);
if (user_param.all == ON) {
for (i = 1; i < 24 ; ++i) {
user_param.size = 1 << i;
if(run_iter(&ctx,&user_param,rem_dest))
return 17;
print_report(&user_param);
}
} else {
if(run_iter(&ctx,&user_param,rem_dest))
return 18;
print_report(&user_param);
}
free(tposted);
free(tcompleted);
// Closing connection.
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;
}
free(my_dest);
free(rem_dest);
printf(RESULT_LINE);
return 0;
}
static int pp_open_port(struct pingpong_context *ctx, const char * servername,
int ib_port, int port, struct pingpong_dest *rem_dest,struct user_parameters *user_parm)
{
char addr_fmt[] = "%8s address: LID %#04x QPN %#06x PSN %#06x\n";
struct pingpong_dest my_dest;
int sockfd;
int rc;
union ibv_gid gid;
/* Create connection between client and server.
* We do it by exchanging data over a TCP socket connection. */
my_dest.lid = pp_get_local_lid(ctx, ib_port);
my_dest.qpn = ctx->qp->qp_num;
my_dest.psn = lrand48() & 0xffffff;
if (user_parm->gid_index < 0) {/*We do not fail test upon lid in RDMA0E/Eth conf*/
if (!my_dest.lid) {
fprintf(stderr, "Local lid 0x0 detected. Is an SM running?\n");
return -1;
}
}
if (user_parm->gid_index != -1) {
int err=0;
err = ibv_query_gid (ctx->context, ib_port, user_parm->gid_index, &gid);
if (err) {
return -1;
}
ctx->dgid=gid;
}
my_dest.dgid = gid;
my_dest.rkey = ctx->mr->rkey;
my_dest.vaddr = (uintptr_t)ctx->buf + ctx->size;
printf(addr_fmt, "local", my_dest.lid, my_dest.qpn, my_dest.psn);
if (user_parm->gid_index > -1) {
printf(" GID: %02x:%02x:%02x:%02x:%02x:%02x:%02x:%02x:%02x:%02x:%02x:%02x:%02x:%02x:%02x:%02x\n",
my_dest.dgid.raw[0],my_dest.dgid.raw[1],
my_dest.dgid.raw[2], my_dest.dgid.raw[3], my_dest.dgid.raw[4],
my_dest.dgid.raw[5], my_dest.dgid.raw[6], my_dest.dgid.raw[7],
my_dest.dgid.raw[8], my_dest.dgid.raw[9], my_dest.dgid.raw[10],
my_dest.dgid.raw[11], my_dest.dgid.raw[12], my_dest.dgid.raw[13],
my_dest.dgid.raw[14], my_dest.dgid.raw[15]);
}
sockfd = servername ? pp_client_connect(servername, port) :
pp_server_connect(port);
if (sockfd < 0) {
printf("pp_connect_sock(%s,%d) failed (%d)!\n",
servername, port, sockfd);
return sockfd;
}
rc = servername ? pp_client_exch_dest(sockfd, &my_dest, rem_dest, user_parm) :
pp_server_exch_dest(sockfd, &my_dest, rem_dest, user_parm);
if (rc)
return rc;
printf(addr_fmt, "remote", rem_dest->lid, rem_dest->qpn, rem_dest->psn,
rem_dest->rkey, rem_dest->vaddr);
if (user_parm->gid_index > -1) {
printf(" GID: %02x:%02x:%02x:%02x:%02x:%02x:%02x:%02x:%02x:%02x:%02x:%02x:%02x:%02x:%02x:%02x\n",
rem_dest->dgid.raw[0],rem_dest->dgid.raw[1],
rem_dest->dgid.raw[2], rem_dest->dgid.raw[3], rem_dest->dgid.raw[4],
rem_dest->dgid.raw[5], rem_dest->dgid.raw[6], rem_dest->dgid.raw[7],
rem_dest->dgid.raw[8], rem_dest->dgid.raw[9], rem_dest->dgid.raw[10],
rem_dest->dgid.raw[11], rem_dest->dgid.raw[12], rem_dest->dgid.raw[13],
rem_dest->dgid.raw[14], rem_dest->dgid.raw[15]);
}
if ((rc = pp_connect_ctx(ctx, ib_port, my_dest.psn, rem_dest,user_parm)))
return rc;
/* An additional handshake is required *after* moving qp to RTR.
* Arbitrarily reuse exch_dest for this purpose.
*/
rc = servername ? pp_client_exch_dest(sockfd, &my_dest, rem_dest, user_parm) :
pp_server_exch_dest(sockfd, &my_dest, rem_dest, user_parm);
if (rc)
return rc;
if (write(sockfd, "done", sizeof "done") != sizeof "done"){
perror("write");
fprintf(stderr, "Couldn't write to socket\n");
return 1;
}
close(sockfd);
return 0;
}
static struct pingpong_dest *pp_server_exch_dest(struct pingpong_context *ctx,
int ib_port,
enum ibv_mtu mtu,
int port,
int sl,
const struct pingpong_dest *my_dest)
{
struct addrinfo *res, *t;
struct addrinfo hints = {
.ai_flags = AI_PASSIVE,
.ai_family = AF_UNSPEC,
.ai_socktype = SOCK_STREAM
};
char *service;
char msg[sizeof "0000:000000:000000"];
int n;
int sockfd = -1, connfd;
struct pingpong_dest *rem_dest = NULL;
if (asprintf(&service, "%d", port) < 0)
return NULL;
n = getaddrinfo(NULL, service, &hints, &res);
if (n < 0) {
fprintf(stderr, "%s for port %d\n", gai_strerror(n), port);
free(service);
return NULL;
}
for (t = res; t; t = t->ai_next) {
sockfd = socket(t->ai_family, t->ai_socktype, t->ai_protocol);
if (sockfd >= 0) {
n = 1;
setsockopt(sockfd, SOL_SOCKET, SO_REUSEADDR, &n, sizeof n);
if (!bind(sockfd, t->ai_addr, t->ai_addrlen))
break;
close(sockfd);
sockfd = -1;
}
}
freeaddrinfo(res);
free(service);
if (sockfd < 0) {
fprintf(stderr, "Couldn't listen to port %d\n", port);
return NULL;
}
listen(sockfd, 1);
connfd = accept(sockfd, NULL, 0);
close(sockfd);
if (connfd < 0) {
fprintf(stderr, "accept() failed\n");
return NULL;
}
n = read(connfd, msg, sizeof msg);
if (n != sizeof msg) {
perror("server read");
fprintf(stderr, "%d/%d: Couldn't read remote address\n",
n, (int) sizeof msg);
goto out;
}
rem_dest = malloc(sizeof *rem_dest);
if (!rem_dest)
goto out;
sscanf(msg, "%x:%x:%x", &rem_dest->lid, &rem_dest->qpn, &rem_dest->psn);
if (pp_connect_ctx(ctx, ctx->qp, ib_port, my_dest->psn, mtu,
sl, rem_dest)) {
fprintf(stderr, "Couldn't connect to remote QP\n");
free(rem_dest);
rem_dest = NULL;
goto out;
}
sprintf(msg, "%04x:%06x:%06x", my_dest->lid, my_dest->qpn,
my_dest->psn);
if (write(connfd, msg, sizeof msg) != sizeof msg) {
fprintf(stderr, "Couldn't send local address\n");
free(rem_dest);
rem_dest = NULL;
goto out;
}
/* expecting "done" msg */
if (read(connfd, msg, sizeof(msg)) <= 0) {
fprintf(stderr, "Couldn't read \"done\" msg\n");
free(rem_dest);
rem_dest = NULL;
goto out;
}
out:
close(connfd);
return rem_dest;
}
int main(int argc, char *argv[]) {
int i = 0;
struct report_options report = {};
struct pingpong_context ctx;
struct ibv_device *ib_dev;
struct perftest_parameters user_param;
struct pingpong_dest my_dest,rem_dest;
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(&my_dest,0,sizeof(struct pingpong_dest));
memset(&rem_dest,0,sizeof(struct pingpong_dest));
user_param.verb = READ;
user_param.tst = LAT;
user_param.r_flag = &report;
user_param.version = VERSION;
// Configure the parameters values according to user arguments or defalut values.
if (parser(&user_param,argv,argc)) {
fprintf(stderr," Parser function exited with Error\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 deivce\n");
return FAILURE;
}
// 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_and_mtu(ctx.context,&user_param)) {
fprintf(stderr, " Couldn't get context for the device\n");
return FAILURE;
}
// Print basic test information.
ctx_print_test_info(&user_param);
// copy the rellevant 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;
}
// Create (if nessacery) the rdma_cm ids and channel.
if (user_param.work_rdma_cm == ON) {
if (create_rdma_resources(&ctx,&user_param)) {
fprintf(stderr," Unable to create the rdma_resources\n");
return FAILURE;
}
if (user_param.machine == CLIENT) {
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)) {
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 (set_up_connection(&ctx,&user_param,&my_dest)) {
fprintf(stderr," Unable to set up socket connection\n");
return 1;
}
ctx_print_pingpong_data(&my_dest,&user_comm);
// Init the connection and print the local data.
if (establish_connection(&user_comm)) {
fprintf(stderr," Unable to init the socket connection\n");
return 1;
}
// shaking hands and gather the other side info.
if (ctx_hand_shake(&user_comm,&my_dest,&rem_dest)) {
fprintf(stderr,"Failed to exchange date between server and clients\n");
return 1;
}
user_comm.rdma_params->side = REMOTE;
ctx_print_pingpong_data(&rem_dest,&user_comm);
if (user_param.work_rdma_cm == OFF) {
if (pp_connect_ctx(&ctx,my_dest.psn,&rem_dest,my_dest.out_reads,&user_param)) {
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,&rem_dest)) {
fprintf(stderr,"Failed to exchange date between server and clients\n");
return 1;
}
ALLOCATE(tstamp,cycles_t,user_param.iters);
// Only Client post read request.
if (user_param.machine == SERVER) {
if (ctx_close_connection(&user_comm,&my_dest,&rem_dest)) {
fprintf(stderr,"Failed to close connection between server and client\n");
return 1;
}
printf(RESULT_LINE);
return 0; // 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;
}
}
printf(RESULT_LINE);
printf(RESULT_FMT_LAT);
if (user_param.all == ON) {
for (i = 1; i < 24 ; ++i) {
user_param.size = 1 << i;
if(run_iter(&ctx,&user_param,&rem_dest))
return 17;
print_report(&user_param);
}
} else {
if(run_iter(&ctx,&user_param,&rem_dest))
return 18;
print_report(&user_param);
}
if (ctx_close_connection(&user_comm,&my_dest,&rem_dest)) {
fprintf(stderr,"Failed to close connection between server and client\n");
return 1;
}
printf(RESULT_LINE);
return 0; // destroy_ctx(&ctx,&user_param);
}
int main(int argc, char *argv[])
{
uint64_t flags = 0;
char *service = NULL;
char *node = NULL;
struct pingpong_context *ctx;
struct timeval start, end;
unsigned long size = 4096;
// No provider support yet
//enum ibv_mtu mtu = IBV_MTU_1024;
//size_t mtu = 1024;
int rx_depth_default = 500;
int rx_depth = 0;
int iters = 1000;
int use_event = 0;
int rcnt, scnt;
int ret, rc = 0;
char * ptr;
srand48(getpid() * time(NULL));
opts = INIT_OPTS;
hints = fi_allocinfo();
if (!hints)
return 1;
while (1) {
int c;
c = getopt(argc, argv, "S:m:r:n:eh" ADDR_OPTS INFO_OPTS);
if (c == -1)
break;
switch (c) {
case 'S':
errno = 0;
size = strtol(optarg, &ptr, 10);
if (ptr == optarg || *ptr != '\0' ||
((size == LONG_MIN || size == LONG_MAX) && errno == ERANGE)) {
fprintf(stderr, "Cannot convert from string to long\n");
rc = 1;
goto err1;
}
break;
// No provider support yet
/*case 'm':
mtu = strtol(optarg, NULL, 0);
mtu = pp_mtu_to_enum(strtol(optarg, NULL, 0));
if (mtu < 0) {
usage(argv[0]);
return 1;
}
break;
*/
case 'r':
rx_depth = strtol(optarg, NULL, 0);
break;
case 'n':
iters = strtol(optarg, NULL, 0);
break;
case 'e':
++use_event;
break;
default:
ft_parse_addr_opts(c, optarg, &opts);
ft_parseinfo(c, optarg, hints);
break;
case '?':
case 'h':
usage(argv[0]);
return 1;
}
}
if (optind == argc - 1)
opts.dst_addr = argv[optind];
else if (optind < argc) {
usage(argv[0]);
return 1;
}
page_size = sysconf(_SC_PAGESIZE);
hints->ep_attr->type = FI_EP_MSG;
hints->caps = FI_MSG;
hints->mode = FI_LOCAL_MR;
rc = ft_read_addr_opts(&node, &service, hints, &flags, &opts);
if (rc)
return -rc;
rc = fi_getinfo(FT_FIVERSION, node, service, flags, hints, &fi);
if (rc) {
FT_PRINTERR("fi_getinfo", rc);
return -rc;
}
fi_freeinfo(hints);
if (rx_depth) {
if (rx_depth > fi->rx_attr->size) {
fprintf(stderr, "rx_depth requested: %d, "
"rx_depth supported: %zd\n", rx_depth, fi->rx_attr->size);
rc = 1;
goto err1;
}
} else {
rx_depth = (rx_depth_default > fi->rx_attr->size) ?
fi->rx_attr->size : rx_depth_default;
}
ctx = pp_init_ctx(fi, size, rx_depth, use_event);
if (!ctx) {
rc = 1;
goto err1;
}
if (opts.dst_addr) {
/* client connect */
if (pp_connect_ctx(ctx)) {
rc = 1;
goto err2;
}
} else {
/* server listen and accept */
pp_listen_ctx(ctx);
pp_accept_ctx(ctx);
}
ctx->pending = PINGPONG_RECV_WCID;
if (opts.dst_addr) {
if (pp_post_send(ctx)) {
fprintf(stderr, "Couldn't post send\n");
rc = 1;
goto err3;
}
ctx->pending |= PINGPONG_SEND_WCID;
}
if (gettimeofday(&start, NULL)) {
perror("gettimeofday");
rc = 1;
goto err3;
}
rcnt = scnt = 0;
while (rcnt < iters || scnt < iters) {
struct fi_cq_entry wc;
struct fi_cq_err_entry cq_err;
int rd;
if (use_event) {
/* Blocking read */
rd = fi_cq_sread(ctx->cq, &wc, 1, NULL, -1);
} else {
do {
rd = fi_cq_read(ctx->cq, &wc, 1);
} while (rd == -FI_EAGAIN);
}
if (rd < 0) {
fi_cq_readerr(ctx->cq, &cq_err, 0);
fprintf(stderr, "cq fi_cq_readerr() %s (%d)\n",
fi_cq_strerror(ctx->cq, cq_err.err, cq_err.err_data, NULL, 0),
cq_err.err);
rc = rd;
goto err3;
}
switch ((int) (uintptr_t) wc.op_context) {
case PINGPONG_SEND_WCID:
++scnt;
break;
case PINGPONG_RECV_WCID:
if (--ctx->routs <= 1) {
ctx->routs += pp_post_recv(ctx, ctx->rx_depth - ctx->routs);
if (ctx->routs < ctx->rx_depth) {
fprintf(stderr,
"Couldn't post receive (%d)\n",
ctx->routs);
rc = 1;
goto err3;
}
}
++rcnt;
break;
default:
fprintf(stderr, "Completion for unknown wc_id %d\n",
(int) (uintptr_t) wc.op_context);
rc = 1;
goto err3;
}
ctx->pending &= ~(int) (uintptr_t) wc.op_context;
if (scnt < iters && !ctx->pending) {
if (pp_post_send(ctx)) {
fprintf(stderr, "Couldn't post send\n");
rc = 1;
goto err3;
}
ctx->pending = PINGPONG_RECV_WCID | PINGPONG_SEND_WCID;
}
}
if (gettimeofday(&end, NULL)) {
perror("gettimeofday");
rc = 1;
goto err3;
}
{
float usec = (end.tv_sec - start.tv_sec) * 1000000 +
(end.tv_usec - start.tv_usec);
long long bytes = (long long) size * iters * 2;
printf("%lld bytes in %.2f seconds = %.2f Mbit/sec\n",
bytes, usec / 1000000., bytes * 8. / usec);
printf("%d iters in %.2f seconds = %.2f usec/iter\n",
iters, usec / 1000000., usec / iters);
}
err3:
fi_shutdown(ctx->ep, 0);
err2:
ret = pp_close_ctx(ctx);
if (!rc)
rc = ret;
err1:
fi_freeinfo(fi);
return rc;
}