static int receive_loop(size_t size)
{
int ret;
int q_opts = 0;
struct fi_context recv_ctx[max_opts];
while (q_opts < max_opts) {
do {
ft_tag = q_opts + 1;
if (tagged)
ret = fi_trecv(ep, rx_buf, size, NULL, remote_fi_addr,
ft_tag, 0x0, (void *) &recv_ctx[q_opts]);
else
ret = fi_recv(ep, rx_buf, size, NULL, remote_fi_addr,
(void *) &recv_ctx[q_opts]);
if (ret == FI_SUCCESS) {
rx_seq++;
q_opts++;
}
} while (!ret && (q_opts != max_opts));
if (ret < 0) {
if (ret == -FI_EAGAIN) {
if (delay > 0)
sleep(delay);
ret = ft_get_rx_comp(rx_seq);
if (ret)
return ret;
} else {
FT_PRINTERR("Recv OP", ret);
return ret;
}
}
}
if (delay > 0)
sleep(delay);
ret = ft_get_rx_comp(rx_seq);
if (ret)
return ret;
if (opts.verbose)
printf("Success: Completed %d queued ops\n", q_opts);
return 0;
}
int send_recv_greeting(struct fid_ep *ep)
{
int ret;
const char *message = "Hello from Client!";
size_t message_len = strlen(message) + 1;
if (opts.dst_addr) {
fprintf(stdout, "Sending message...\n");
if (snprintf(tx_buf, tx_size, "%s", message) >= tx_size) {
fprintf(stderr, "Transmit buffer too small.\n");
return -FI_ETOOSMALL;
}
ret = ft_tx(ep, remote_fi_addr, message_len, &tx_ctx);
if (ret)
return ret;
fprintf(stdout, "Send completion received\n");
} else {
fprintf(stdout, "Waiting for message from client...\n");
ret = ft_get_rx_comp(rx_seq);
if (ret)
return ret;
ret = check_recv_msg(message);
if (ret)
return ret;
fprintf(stdout, "Received data from client: %s\n",
(char *) rx_buf);
}
return 0;
}
static int run_test()
{
int ret, i;
/* Post recvs */
for (i = 0; i < ep_cnt; i++) {
fprintf(stdout, "Posting recv for ctx: %d\n", i);
ret = fi_recv(srx_ctx, rx_buf, rx_size, fi_mr_desc(mr),
FI_ADDR_UNSPEC, NULL);
if (ret) {
FT_PRINTERR("fi_recv", ret);
return ret;
}
rx_seq++;
}
if (opts.dst_addr) {
/* Post sends addressed to remote EPs */
for (i = 0; i < ep_cnt; i++) {
fprintf(stdout, "Posting send to remote ctx: %d\n", i);
ret = fi_send(ep_array[i], tx_buf, tx_size, fi_mr_desc(mr),
addr_array[i], NULL);
if (ret) {
FT_PRINTERR("fi_send", ret);
return ret;
}
ret = ft_get_tx_comp(++tx_seq);
if (ret)
return ret;
}
}
/* Wait for recv completions */
ret = ft_get_rx_comp(rx_seq);
if (ret)
return ret;
if (!opts.dst_addr) {
/* Post sends addressed to remote EPs */
for (i = 0; i < ep_cnt; i++) {
fprintf(stdout, "Posting send to remote ctx: %d\n", i);
ret = fi_send(ep_array[i], tx_buf, tx_size, fi_mr_desc(mr),
addr_array[i], NULL);
if (ret) {
FT_PRINTERR("fi_send", ret);
return ret;
}
ret = ft_get_tx_comp(++tx_seq);
if (ret)
return ret;
}
}
return 0;
}
static int run_test()
{
int ret, i;
if (!(tx_ctx_arr = calloc(ep_cnt, sizeof *tx_ctx_arr)))
return -FI_ENOMEM;
if (!(rx_ctx_arr = calloc(ep_cnt, sizeof *rx_ctx_arr)))
return -FI_ENOMEM;
/* Post recvs */
for (i = 0; i < ep_cnt; i++) {
if (rx_shared_ctx) {
fprintf(stdout, "Posting recv #%d for shared rx ctx\n", i);
ret = ft_post_rx(srx_ctx, rx_size, &rx_ctx_arr[i]);
} else {
fprintf(stdout, "Posting recv for endpoint #%d\n", i);
ret = ft_post_rx(ep_array[i], rx_size, &rx_ctx_arr[i]);
}
if (ret)
return ret;
}
if (opts.dst_addr) {
/* Post sends addressed to remote EPs */
for (i = 0; i < ep_cnt; i++) {
if (tx_shared_ctx)
fprintf(stdout, "Posting send #%d to shared tx ctx\n", i);
else
fprintf(stdout, "Posting send to endpoint #%d\n", i);
ret = ft_tx(ep_array[i], addr_array[i], tx_size, &tx_ctx_arr[i]);
if (ret)
return ret;
}
}
/* Wait for recv completions */
ret = ft_get_rx_comp(rx_seq - 1);
if (ret)
return ret;
if (!opts.dst_addr) {
/* Post sends addressed to remote EPs */
for (i = 0; i < ep_cnt; i++) {
if (tx_shared_ctx)
fprintf(stdout, "Posting send #%d to shared tx ctx\n", i);
else
fprintf(stdout, "Posting send to endpoint #%d\n", i);
ret = ft_tx(ep_array[i], addr_array[i], tx_size, &tx_ctx_arr[i]);
if (ret)
return ret;
}
}
return 0;
}
int ft_exchange_keys(struct fi_rma_iov *peer_iov)
{
struct fi_rma_iov *rma_iov;
int ret;
if (opts.dst_addr) {
rma_iov = tx_buf + ft_tx_prefix_size();
rma_iov->addr = fi->domain_attr->mr_mode == FI_MR_SCALABLE ?
0 : (uintptr_t) rx_buf + ft_rx_prefix_size();
rma_iov->key = fi_mr_key(mr);
ret = ft_tx(sizeof *rma_iov);
if (ret)
return ret;
ret = ft_get_rx_comp(rx_seq);
if (ret)
return ret;
rma_iov = rx_buf + ft_rx_prefix_size();
*peer_iov = *rma_iov;
ret = ft_post_rx(rx_size);
} else {
ret = ft_get_rx_comp(rx_seq);
if (ret)
return ret;
rma_iov = rx_buf + ft_rx_prefix_size();
*peer_iov = *rma_iov;
ret = ft_post_rx(rx_size);
if (ret)
return ret;
rma_iov = tx_buf + ft_tx_prefix_size();
rma_iov->addr = fi->domain_attr->mr_mode == FI_MR_SCALABLE ?
0 : (uintptr_t) rx_buf + ft_rx_prefix_size();
rma_iov->key = fi_mr_key(mr);
ret = ft_tx(sizeof *rma_iov);
}
return ret;
}
int ft_finalize(void)
{
struct iovec iov;
int ret;
struct fi_context ctx;
void *desc = fi_mr_desc(mr);
strcpy(tx_buf + ft_tx_prefix_size(), "fin");
iov.iov_base = tx_buf;
iov.iov_len = 4 + ft_tx_prefix_size();
if (hints->caps & FI_TAGGED) {
struct fi_msg_tagged tmsg;
memset(&tmsg, 0, sizeof tmsg);
tmsg.msg_iov = &iov;
tmsg.desc = &desc;
tmsg.iov_count = 1;
tmsg.addr = remote_fi_addr;
tmsg.tag = tx_seq;
tmsg.ignore = 0;
tmsg.context = &ctx;
ret = fi_tsendmsg(ep, &tmsg, FI_INJECT | FI_TRANSMIT_COMPLETE);
} else {
struct fi_msg msg;
memset(&msg, 0, sizeof msg);
msg.msg_iov = &iov;
msg.desc = &desc;
msg.iov_count = 1;
msg.addr = remote_fi_addr;
msg.context = &ctx;
ret = fi_sendmsg(ep, &msg, FI_INJECT | FI_TRANSMIT_COMPLETE);
}
if (ret) {
FT_PRINTERR("transmit", ret);
return ret;
}
ret = ft_get_tx_comp(++tx_seq);
if (ret)
return ret;
ret = ft_get_rx_comp(rx_seq);
if (ret)
return ret;
return 0;
}
static int recv_msg(void)
{
int ret;
ret = fi_recv(srx_ctx, buf, rx_size, fi_mr_desc(mr), 0, &rx_ctx);
if (ret) {
FT_PRINTERR("fi_recv", ret);
return ret;
}
ret = ft_get_rx_comp(++rx_seq);
return ret;
}
int send_recv_greeting(void)
{
int ret;
const char *message = "Hello from Client!";
/* strlen doesn't include null terminated byte. snprintf size includes
* null terminated byte.
*/
size_t message_len = strlen(message) + 1;
size_t recv_len;
if (opts.dst_addr) {
fprintf(stdout, "Sending message...\n");
if (snprintf(tx_buf, tx_size, "%s", message) >= tx_size) {
fprintf(stderr, "Transmit buffer too small.\n");
return -FI_ETOOSMALL;
}
ret = ft_tx(message_len);
if (ret)
return ret;
fprintf(stdout, "Send completion received\n");
} else {
fprintf(stdout, "Waiting for message from client...\n");
ret = ft_get_rx_comp(rx_seq);
if (ret)
return ret;
/* Account for null terminated byte. */
recv_len = strlen(rx_buf) + 1;
if (recv_len != message_len) {
fprintf(stderr,
"Received length does not match expected length.\n");
return -1;
}
if (strncmp(rx_buf, message, MIN(recv_len, message_len))) {
fprintf(stderr,
"Received message does not match expected message.\n");
return -1;
}
fprintf(stdout, "Received data from client: %s\n",
(char *) rx_buf);
}
return 0;
}
static int run_test(void)
{
int ret = 0;
const char *message = "Hello from Client!";
size_t message_len = strlen(message) + 1;
ret = ft_init_fabric();
if (ret)
return ret;
ret = ft_exchange_keys(&remote);
if (ret)
return ret;
if (opts.dst_addr) {
fprintf(stdout, "RMA write to server\n");
if (snprintf(tx_buf, tx_size, "%s", message) >= tx_size) {
fprintf(stderr, "Transmit buffer too small.\n");
return -FI_ETOOSMALL;
}
ret = fi_write(ep, tx_buf, message_len, mr_desc,
remote_fi_addr, remote.addr, remote.key,
&fi_ctx_write);
if (ret)
return ret;
ret = ft_get_tx_comp(++tx_seq);
if (ret)
return ret;
fprintf(stdout, "Received a completion event for RMA write\n");
} else {
ret = ft_get_rx_comp(rx_seq);
if (ret)
return ret;
ret = check_recv_msg(message);
if (ret)
return ret;
fprintf(stdout, "Received data from Client: %s\n", (char *) rx_buf);
}
/* TODO: need support for finalize operation to sync test */
return 0;
}
ssize_t ft_rx(size_t size)
{
ssize_t ret;
ret = ft_get_rx_comp(rx_seq);
if (ret)
return ret;
if (ft_check_opts(FT_OPT_VERIFY_DATA | FT_OPT_ACTIVE)) {
ret = ft_check_buf((char *) rx_buf + ft_rx_prefix_size(), size);
if (ret)
return ret;
}
/* TODO: verify CQ data, if available */
ret = ft_post_rx(rx_size);
return ret;
}
static int do_transfers(void)
{
int i, ret;
for (i = 0; i < num_eps; i++) {
rx_buf = recv_bufs[i];
ret = ft_post_rx(eps[i], opts.transfer_size, &recv_ctx[i]);
if (ret)
return ret;
}
for (i = 0; i < num_eps; i++) {
if (ft_check_opts(FT_OPT_VERIFY_DATA))
ft_fill_buf(send_bufs[i], opts.transfer_size);
tx_buf = send_bufs[i];
ret = ft_post_tx(eps[i], remote_addr[i], opts.transfer_size, &send_ctx[i]);
if (ret)
return ret;
}
ret = ft_get_tx_comp(num_eps);
if (ret < 0)
return ret;
ret = ft_get_rx_comp(num_eps);
if (ret < 0)
return ret;
if (ft_check_opts(FT_OPT_VERIFY_DATA)) {
for (i = 0; i < num_eps; i++) {
ret = ft_check_buf(recv_bufs[i], opts.transfer_size);
if (ret)
return ret;
}
}
for (i = 0; i < num_eps; i++)
ft_finalize_ep(eps[i]);
printf("PASSED multi ep\n");
return 0;
}
ssize_t ft_rx(struct fid_ep *ep, size_t size)
{
ssize_t ret;
ret = ft_get_rx_comp(rx_seq);
if (ret)
return ret;
if (ft_check_opts(FT_OPT_VERIFY_DATA | FT_OPT_ACTIVE)) {
ret = ft_check_buf((char *) rx_buf + ft_rx_prefix_size(), size);
if (ret)
return ret;
}
/* TODO: verify CQ data, if available */
/* Ignore the size arg. Post a buffer large enough to handle all message
* sizes. ft_sync() makes use of ft_rx() and gets called in tests just before
* message size is updated. The recvs posted are always for the next incoming
* message */
ret = ft_post_rx(ep, rx_size, &rx_ctx);
return ret;
}
static int send_recv()
{
int ret;
if (opts.dst_addr) {
/* Client */
fprintf(stdout, "Sending message to server...\n");
sprintf(tx_buf, "Hello World!");
ret = ft_tx(sizeof("Hello World!"));
if (ret)
return ret;
fprintf(stdout, "Send completed\n");
} else {
fprintf(stdout, "Waiting for client...\n");
ret = ft_get_rx_comp(rx_seq);
if (ret)
return ret;
fprintf(stdout, "Received data from client: %s\n", (char *) rx_buf);
}
return 0;
}
int bandwidth(void)
{
int ret, i, j;
ret = ft_sync();
if (ret)
return ret;
/* The loop structured allows for the possibility that the sender
* immediately overruns the receiving side on the first transfer (or
* the entire window). This could result in exercising parts of the
* provider's implementation of FI_RM_ENABLED. For better or worse,
* some MPI-level benchmarks tend to use this type of loop for measuring
* bandwidth. */
if (opts.dst_addr) {
for (i = 0; i < opts.iterations + opts.warmup_iterations; i++) {
if (i == opts.warmup_iterations)
ft_start();
for(j = 0; j < opts.window_size; j++) {
if (opts.transfer_size < fi->tx_attr->inject_size)
ret = ft_inject(opts.transfer_size);
else
ret = ft_post_tx(opts.transfer_size);
if (ret)
return ret;
}
ret = ft_get_tx_comp(tx_seq);
if (ret)
return ret;
ret = ft_rx(4);
if (ret)
return ret;
}
} else {
for (i = 0; i < opts.iterations + opts.warmup_iterations; i++) {
if (i == opts.warmup_iterations)
ft_start();
for(j = 0; j < opts.window_size; j++) {
ret = ft_post_rx(opts.transfer_size);
if (ret)
return ret;
}
ret = ft_get_rx_comp(rx_seq-1); /* rx_seq is always one ahead */
if (ret)
return ret;
ret = ft_tx(4);
if (ret)
return ret;
}
}
ft_stop();
if (opts.machr)
show_perf_mr(opts.transfer_size, opts.iterations, &start, &end,
opts.window_size, opts.argc, opts.argv);
else
show_perf(NULL, opts.transfer_size, opts.iterations, &start, &end,
opts.window_size);
return 0;
}
static int run(void)
{
int ret;
ret = init_fabric();
if (ret)
return ret;
ret = ft_init_av();
if (ret)
return ret;
if (opts.dst_addr) {
printf("Searching for a bad msg\n");
ret = tagged_peek(0xbad);
if (ret != -FI_ENOMSG) {
FT_PRINTERR("FI_PEEK", ret);
return ret;
}
printf("Synchronizing with sender..\n");
ret = ft_sync();
if (ret)
return ret;
printf("Searching for a good msg\n");
ret = tagged_peek(0x900d);
if (ret != 1) {
FT_PRINTERR("FI_PEEK", ret);
return ret;
}
printf("Receiving msg\n");
ret = fi_trecv(ep, buf, rx_size, fi_mr_desc(mr), remote_fi_addr,
0x900d, 0, &rx_ctx);
if (ret) {
FT_PRINTERR("fi_trecv", ret);
return ret;
}
printf("Completing recv\n");
ret = ft_get_rx_comp(++rx_seq);
if (ret)
return ret;
} else {
printf("Sending tagged message\n");
ret = fi_tsend(ep, tx_buf, tx_size, fi_mr_desc(mr),
remote_fi_addr, 0x900d, &tx_ctx);
if (ret)
return ret;
printf("Synchronizing with receiver..\n");
ret = ft_sync();
if (ret)
return ret;
printf("Getting send completion\n");
ret = ft_get_tx_comp(tx_seq + 1);
if (ret)
return ret;
}
ft_finalize();
return 0;
}
