2 * Copyright (c) 2005 Topspin Communications. All rights reserved.
4 * This software is available to you under a choice of one of two
5 * licenses. You may choose to be licensed under the terms of the GNU
6 * General Public License (GPL) Version 2, available from the file
7 * COPYING in the main directory of this source tree, or the
8 * OpenIB.org BSD license below:
10 * Redistribution and use in source and binary forms, with or
11 * without modification, are permitted provided that the following
14 * - Redistributions of source code must retain the above
15 * copyright notice, this list of conditions and the following
18 * - Redistributions in binary form must reproduce the above
19 * copyright notice, this list of conditions and the following
20 * disclaimer in the documentation and/or other materials
21 * provided with the distribution.
23 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
24 * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
25 * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
26 * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
27 * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
28 * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
29 * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
39 #include <sys/types.h>
40 #include <sys/socket.h>
45 #include <arpa/inet.h>
51 PINGPONG_RECV_WRID = 1,
52 PINGPONG_SEND_WRID = 2,
57 struct pingpong_context {
58 struct ibv_context *context;
59 struct ibv_comp_channel *channel;
69 struct ibv_port_attr portinfo;
72 struct pingpong_dest {
79 static int pp_connect_ctx(struct pingpong_context *ctx, int port, int my_psn,
80 enum ibv_mtu mtu, int sl,
81 struct pingpong_dest *dest, int sgid_idx)
83 struct ibv_qp_attr attr = {
84 .qp_state = IBV_QPS_RTR,
86 .dest_qp_num = dest->qpn,
97 if (dest->gid.global.interface_id) {
98 attr.ah_attr.is_global = 1;
99 attr.ah_attr.grh.hop_limit = 1;
100 attr.ah_attr.grh.dgid = dest->gid;
101 attr.ah_attr.grh.sgid_index = sgid_idx;
103 if (ibv_modify_qp(ctx->qp, &attr,
109 fprintf(stderr, "Failed to modify QP to RTR\n");
113 attr.qp_state = IBV_QPS_RTS;
114 attr.sq_psn = my_psn;
115 if (ibv_modify_qp(ctx->qp, &attr,
118 fprintf(stderr, "Failed to modify QP to RTS\n");
125 static struct pingpong_dest *pp_client_exch_dest(const char *servername, int port,
126 const struct pingpong_dest *my_dest)
128 struct addrinfo *res, *t;
129 struct addrinfo hints = {
130 .ai_family = AF_UNSPEC,
131 .ai_socktype = SOCK_STREAM
134 char msg[sizeof "0000:000000:000000:00000000000000000000000000000000"];
137 struct pingpong_dest *rem_dest = NULL;
140 if (asprintf(&service, "%d", port) < 0)
143 n = getaddrinfo(servername, service, &hints, &res);
146 fprintf(stderr, "%s for %s:%d\n", gai_strerror(n), servername, port);
151 for (t = res; t; t = t->ai_next) {
152 sockfd = socket(t->ai_family, t->ai_socktype, t->ai_protocol);
154 if (!connect(sockfd, t->ai_addr, t->ai_addrlen))
161 freeaddrinfo_null(res);
165 fprintf(stderr, "Couldn't connect to %s:%d\n", servername, port);
169 gid_to_wire_gid(&my_dest->gid, gid);
170 sprintf(msg, "%04x:%06x:%06x:%s", my_dest->lid, my_dest->qpn,
172 if (write(sockfd, msg, sizeof msg) != sizeof msg) {
173 fprintf(stderr, "Couldn't send local address\n");
177 if (read(sockfd, msg, sizeof msg) != sizeof msg ||
178 write(sockfd, "done", sizeof "done") != sizeof "done") {
179 perror("client read/write");
180 fprintf(stderr, "Couldn't read/write remote address\n");
185 rem_dest = malloc(sizeof *rem_dest);
189 sscanf(msg, "%x:%x:%x:%s", &rem_dest->lid, &rem_dest->qpn,
190 &rem_dest->psn, gid);
191 wire_gid_to_gid(gid, &rem_dest->gid);
198 static struct pingpong_dest *pp_server_exch_dest(struct pingpong_context *ctx,
199 int ib_port, enum ibv_mtu mtu,
201 const struct pingpong_dest *my_dest,
204 struct addrinfo *res, *t;
205 struct addrinfo hints = {
206 .ai_flags = AI_PASSIVE,
207 .ai_family = AF_INET,
208 .ai_socktype = SOCK_STREAM
211 char msg[sizeof "0000:000000:000000:00000000000000000000000000000000"];
213 int sockfd = -1, connfd;
214 struct pingpong_dest *rem_dest = NULL;
217 if (asprintf(&service, "%d", port) < 0)
220 n = getaddrinfo(NULL, service, &hints, &res);
223 fprintf(stderr, "%s for port %d\n", gai_strerror(n), port);
228 for (t = res; t; t = t->ai_next) {
229 sockfd = socket(t->ai_family, t->ai_socktype, t->ai_protocol);
233 setsockopt(sockfd, SOL_SOCKET, SO_REUSEADDR, &n, sizeof n);
235 if (!bind(sockfd, t->ai_addr, t->ai_addrlen))
242 freeaddrinfo_null(res);
246 fprintf(stderr, "Couldn't listen to port %d\n", port);
251 connfd = accept(sockfd, NULL, NULL);
254 fprintf(stderr, "accept() failed\n");
258 n = read(connfd, msg, sizeof msg);
259 if (n != sizeof msg) {
260 perror("server read");
261 fprintf(stderr, "%d/%d: Couldn't read remote address\n", n, (int) sizeof msg);
265 rem_dest = malloc(sizeof *rem_dest);
269 sscanf(msg, "%x:%x:%x:%s", &rem_dest->lid, &rem_dest->qpn,
270 &rem_dest->psn, gid);
271 wire_gid_to_gid(gid, &rem_dest->gid);
273 if (pp_connect_ctx(ctx, ib_port, my_dest->psn, mtu, sl, rem_dest,
275 fprintf(stderr, "Couldn't connect to remote QP\n");
282 gid_to_wire_gid(&my_dest->gid, gid);
283 sprintf(msg, "%04x:%06x:%06x:%s", my_dest->lid, my_dest->qpn,
285 if (write(connfd, msg, sizeof msg) != sizeof msg ||
286 read(connfd, msg, sizeof msg) != sizeof "done") {
287 fprintf(stderr, "Couldn't send/recv local address\n");
298 static struct pingpong_context *pp_init_ctx(struct ibv_device *ib_dev, int size,
299 int rx_depth, int port,
302 struct pingpong_context *ctx;
304 ctx = calloc(1, sizeof *ctx);
309 ctx->send_flags = IBV_SEND_SIGNALED;
310 ctx->rx_depth = rx_depth;
312 ctx->buf = memalign(page_size, size);
314 fprintf(stderr, "Couldn't allocate work buf.\n");
318 /* FIXME memset(ctx->buf, 0, size); */
319 memset(ctx->buf, 0x7b, size);
321 ctx->context = ibv_open_device(ib_dev);
323 fprintf(stderr, "Couldn't get context for %s\n",
324 ibv_get_device_name(ib_dev));
329 ctx->channel = ibv_create_comp_channel(ctx->context);
331 fprintf(stderr, "Couldn't create completion channel\n");
337 ctx->pd = ibv_alloc_pd(ctx->context);
339 fprintf(stderr, "Couldn't allocate PD\n");
340 goto clean_comp_channel;
343 ctx->mr = ibv_reg_mr(ctx->pd, ctx->buf, size, IBV_ACCESS_LOCAL_WRITE);
345 fprintf(stderr, "Couldn't register MR\n");
349 ctx->cq = ibv_create_cq(ctx->context, rx_depth + 1, NULL,
352 fprintf(stderr, "Couldn't create CQ\n");
357 struct ibv_qp_attr attr;
358 struct ibv_qp_init_attr init_attr = {
363 .max_recv_wr = rx_depth,
367 .qp_type = IBV_QPT_UC
370 ctx->qp = ibv_create_qp(ctx->pd, &init_attr);
372 fprintf(stderr, "Couldn't create QP\n");
375 ibv_query_qp(ctx->qp, &attr, IBV_QP_CAP, &init_attr);
376 if (init_attr.cap.max_inline_data >= size) {
377 ctx->send_flags |= IBV_SEND_INLINE;
382 struct ibv_qp_attr attr = {
383 .qp_state = IBV_QPS_INIT,
389 if (ibv_modify_qp(ctx->qp, &attr,
393 IBV_QP_ACCESS_FLAGS)) {
394 fprintf(stderr, "Failed to modify QP to INIT\n");
402 ibv_destroy_qp(ctx->qp);
405 ibv_destroy_cq(ctx->cq);
408 ibv_dereg_mr(ctx->mr);
411 ibv_dealloc_pd(ctx->pd);
415 ibv_destroy_comp_channel(ctx->channel);
418 ibv_close_device(ctx->context);
429 static int pp_close_ctx(struct pingpong_context *ctx)
431 if (ibv_destroy_qp(ctx->qp)) {
432 fprintf(stderr, "Couldn't destroy QP\n");
436 if (ibv_destroy_cq(ctx->cq)) {
437 fprintf(stderr, "Couldn't destroy CQ\n");
441 if (ibv_dereg_mr(ctx->mr)) {
442 fprintf(stderr, "Couldn't deregister MR\n");
446 if (ibv_dealloc_pd(ctx->pd)) {
447 fprintf(stderr, "Couldn't deallocate PD\n");
452 if (ibv_destroy_comp_channel(ctx->channel)) {
453 fprintf(stderr, "Couldn't destroy completion channel\n");
458 if (ibv_close_device(ctx->context)) {
459 fprintf(stderr, "Couldn't release context\n");
469 static int pp_post_recv(struct pingpong_context *ctx, int n)
471 struct ibv_sge list = {
472 .addr = (uintptr_t) ctx->buf,
474 .lkey = ctx->mr->lkey
476 struct ibv_recv_wr wr = {
477 .wr_id = PINGPONG_RECV_WRID,
481 struct ibv_recv_wr *bad_wr;
484 for (i = 0; i < n; ++i)
485 if (ibv_post_recv(ctx->qp, &wr, &bad_wr))
491 static int pp_post_send(struct pingpong_context *ctx)
493 struct ibv_sge list = {
494 .addr = (uintptr_t) ctx->buf,
496 .lkey = ctx->mr->lkey
498 struct ibv_send_wr wr = {
499 .wr_id = PINGPONG_SEND_WRID,
502 .opcode = IBV_WR_SEND,
503 .send_flags = ctx->send_flags,
505 struct ibv_send_wr *bad_wr;
507 return ibv_post_send(ctx->qp, &wr, &bad_wr);
510 static void usage(const char *argv0)
513 printf(" %s start a server and wait for connection\n", argv0);
514 printf(" %s <host> connect to server at <host>\n", argv0);
516 printf("Options:\n");
517 printf(" -p, --port=<port> listen on/connect to port <port> (default 18515)\n");
518 printf(" -d, --ib-dev=<dev> use IB device <dev> (default first device found)\n");
519 printf(" -i, --ib-port=<port> use port <port> of IB device (default 1)\n");
520 printf(" -s, --size=<size> size of message to exchange (default 4096)\n");
521 printf(" -m, --mtu=<size> path MTU (default 1024)\n");
522 printf(" -r, --rx-depth=<dep> number of receives to post at a time (default 500)\n");
523 printf(" -n, --iters=<iters> number of exchanges (default 1000)\n");
524 printf(" -l, --sl=<sl> service level value\n");
525 printf(" -e, --events sleep on CQ events (default poll)\n");
526 printf(" -g, --gid-idx=<gid index> local port gid index\n");
529 int main(int argc, char *argv[])
531 struct ibv_device **dev_list;
532 struct ibv_device *ib_dev;
533 struct pingpong_context *ctx;
534 struct pingpong_dest my_dest;
535 struct pingpong_dest *rem_dest;
536 struct timeval start, end;
537 char *ib_devname = NULL;
538 char *servername = NULL;
539 unsigned int port = 18515;
541 unsigned int size = 4096;
542 enum ibv_mtu mtu = IBV_MTU_1024;
543 unsigned int rx_depth = 500;
544 unsigned int iters = 1000;
548 int num_cq_events = 0;
553 srand48(getpid() * time(NULL));
558 static struct option long_options[] = {
559 { .name = "port", .has_arg = 1, .val = 'p' },
560 { .name = "ib-dev", .has_arg = 1, .val = 'd' },
561 { .name = "ib-port", .has_arg = 1, .val = 'i' },
562 { .name = "size", .has_arg = 1, .val = 's' },
563 { .name = "mtu", .has_arg = 1, .val = 'm' },
564 { .name = "rx-depth", .has_arg = 1, .val = 'r' },
565 { .name = "iters", .has_arg = 1, .val = 'n' },
566 { .name = "sl", .has_arg = 1, .val = 'l' },
567 { .name = "events", .has_arg = 0, .val = 'e' },
568 { .name = "gid-idx", .has_arg = 1, .val = 'g' },
572 c = getopt_long(argc, argv, "p:d:i:s:m:r:n:l:eg:",
579 port = strtoul(optarg, NULL, 0);
587 ib_devname = strdupa(optarg);
591 ib_port = strtol(optarg, NULL, 0);
599 size = strtoul(optarg, NULL, 0);
603 mtu = pp_mtu_to_enum(strtol(optarg, NULL, 0));
611 rx_depth = strtoul(optarg, NULL, 0);
615 iters = strtoul(optarg, NULL, 0);
619 sl = strtol(optarg, NULL, 0);
627 gidx = strtol(optarg, NULL, 0);
636 if (optind == argc - 1)
637 servername = strdupa(argv[optind]);
638 else if (optind < argc) {
643 page_size = sysconf(_SC_PAGESIZE);
645 dev_list = ibv_get_device_list(NULL);
647 perror("Failed to get IB devices list");
654 fprintf(stderr, "No IB devices found\n");
659 for (i = 0; dev_list[i]; ++i)
660 if (!strcmp(ibv_get_device_name(dev_list[i]), ib_devname))
662 ib_dev = dev_list[i];
664 fprintf(stderr, "IB device %s not found\n", ib_devname);
669 ctx = pp_init_ctx(ib_dev, size, rx_depth, ib_port, use_event);
673 routs = pp_post_recv(ctx, ctx->rx_depth);
674 if (routs < ctx->rx_depth) {
675 fprintf(stderr, "Couldn't post receive (%d)\n", routs);
680 if (ibv_req_notify_cq(ctx->cq, 0)) {
681 fprintf(stderr, "Couldn't request CQ notification\n");
686 if (pp_get_port_info(ctx->context, ib_port, &ctx->portinfo)) {
687 fprintf(stderr, "Couldn't get port info\n");
691 my_dest.lid = ctx->portinfo.lid;
692 if (ctx->portinfo.link_layer != IBV_LINK_LAYER_ETHERNET &&
694 fprintf(stderr, "Couldn't get local LID\n");
699 if (ibv_query_gid(ctx->context, ib_port, gidx, &my_dest.gid)) {
700 fprintf(stderr, "can't read sgid of index %d\n", gidx);
704 memset(&my_dest.gid, 0, sizeof my_dest.gid);
706 my_dest.qpn = ctx->qp->qp_num;
707 my_dest.psn = lrand48() & 0xffffff;
708 inet_ntop(AF_INET6, &my_dest.gid, gid, sizeof gid);
709 printf(" local address: LID 0x%04x, QPN 0x%06x, PSN 0x%06x, GID %s\n",
710 my_dest.lid, my_dest.qpn, my_dest.psn, gid);
714 rem_dest = pp_client_exch_dest(servername, port, &my_dest);
716 rem_dest = pp_server_exch_dest(ctx, ib_port, mtu, port, sl,
722 inet_ntop(AF_INET6, &rem_dest->gid, gid, sizeof gid);
723 printf(" remote address: LID 0x%04x, QPN 0x%06x, PSN 0x%06x, GID %s\n",
724 rem_dest->lid, rem_dest->qpn, rem_dest->psn, gid);
727 if (pp_connect_ctx(ctx, ib_port, my_dest.psn, mtu, sl, rem_dest,
731 ctx->pending = PINGPONG_RECV_WRID;
734 if (pp_post_send(ctx)) {
735 fprintf(stderr, "Couldn't post send\n");
738 ctx->pending |= PINGPONG_SEND_WRID;
741 if (gettimeofday(&start, NULL)) {
742 perror("gettimeofday");
747 while (rcnt < iters || scnt < iters) {
749 struct ibv_cq *ev_cq;
752 if (ibv_get_cq_event(ctx->channel, &ev_cq, &ev_ctx)) {
753 fprintf(stderr, "Failed to get cq_event\n");
759 if (ev_cq != ctx->cq) {
760 fprintf(stderr, "CQ event for unknown CQ %p\n", ev_cq);
764 if (ibv_req_notify_cq(ctx->cq, 0)) {
765 fprintf(stderr, "Couldn't request CQ notification\n");
775 ne = ibv_poll_cq(ctx->cq, 2, wc);
777 fprintf(stderr, "poll CQ failed %d\n", ne);
781 } while (!use_event && ne < 1);
783 for (i = 0; i < ne; ++i) {
784 if (wc[i].status != IBV_WC_SUCCESS) {
785 fprintf(stderr, "Failed status %s (%d) for wr_id %d\n",
786 ibv_wc_status_str(wc[i].status),
787 wc[i].status, (int) wc[i].wr_id);
791 switch ((int) wc[i].wr_id) {
792 case PINGPONG_SEND_WRID:
796 case PINGPONG_RECV_WRID:
798 routs += pp_post_recv(ctx, ctx->rx_depth - routs);
799 if (routs < ctx->rx_depth) {
801 "Couldn't post receive (%d)\n",
811 fprintf(stderr, "Completion for unknown wr_id %d\n",
816 ctx->pending &= ~(int) wc[i].wr_id;
817 if (scnt < iters && !ctx->pending) {
818 if (pp_post_send(ctx)) {
819 fprintf(stderr, "Couldn't post send\n");
822 ctx->pending = PINGPONG_RECV_WRID |
829 if (gettimeofday(&end, NULL)) {
830 perror("gettimeofday");
835 float usec = (end.tv_sec - start.tv_sec) * 1000000 +
836 (end.tv_usec - start.tv_usec);
837 long long bytes = (long long) size * iters * 2;
839 printf("%lld bytes in %.2f seconds = %.2f Mbit/sec\n",
840 bytes, usec / 1000000., bytes * 8. / usec);
841 printf("%d iters in %.2f seconds = %.2f usec/iter\n",
842 iters, usec / 1000000., usec / iters);
845 ibv_ack_cq_events(ctx->cq, num_cq_events);
847 if (pp_close_ctx(ctx))
850 ibv_free_device_list(dev_list);