2 * Copyright (c) 2005 Ammasso, Inc. All rights reserved.
3 * Copyright (c) 2006 Open Grid Computing, Inc. All rights reserved.
5 * This software is available to you under a choice of one of two
6 * licenses. You may choose to be licensed under the terms of the GNU
7 * General Public License (GPL) Version 2, available from the file
8 * COPYING in the main directory of this source tree, or the
9 * OpenIB.org BSD license below:
11 * Redistribution and use in source and binary forms, with or
12 * without modification, are permitted provided that the following
15 * - Redistributions of source code must retain the above
16 * copyright notice, this list of conditions and the following
19 * - Redistributions in binary form must reproduce the above
20 * copyright notice, this list of conditions and the following
21 * disclaimer in the documentation and/or other materials
22 * provided with the distribution.
24 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
25 * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
26 * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
27 * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
28 * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
29 * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
30 * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
34 #include <sys/cdefs.h>
35 __FBSDID("$FreeBSD$");
37 #include <sys/ctype.h>
39 #include <sys/param.h>
40 #include <sys/condvar.h>
41 #include <sys/systm.h>
42 #include <sys/kernel.h>
43 #include <sys/socket.h>
44 #include <sys/module.h>
45 #include <sys/endian.h>
46 #include <sys/limits.h>
48 #include <sys/signalvar.h>
51 #include <sys/mutex.h>
52 #include <sys/rwlock.h>
53 #include <sys/queue.h>
54 #include <sys/taskqueue.h>
55 #include <sys/syslog.h>
60 #include <contrib/rdma/rdma_cm.h>
65 #define PFX "krping: "
68 #define DEBUG_LOG if (debug) printf
70 static const struct krping_option krping_opts[] = {
71 {"count", OPT_INT, 'C'},
72 {"size", OPT_INT, 'S'},
73 {"addr", OPT_STRING, 'a'},
74 {"port", OPT_INT, 'p'},
75 {"verbose", OPT_NOPARAM, 'v'},
76 {"validate", OPT_NOPARAM, 'V'},
77 {"server", OPT_NOPARAM, 's'},
78 {"client", OPT_NOPARAM, 'c'},
79 {"dmamr", OPT_NOPARAM, 'D'},
80 {"debug", OPT_NOPARAM, 'd'},
81 {"wlat", OPT_NOPARAM, 'l'},
82 {"rlat", OPT_NOPARAM, 'L'},
83 {"bw", OPT_NOPARAM, 'B'},
84 {"tx-depth", OPT_INT, 't'},
85 {"poll", OPT_NOPARAM, 'P'},
89 struct mtx krping_mutex;
92 * List of running krping threads.
94 struct krping_cb_list krping_cbs;
97 * krping "ping/pong" loop:
98 * client sends source rkey/addr/len
99 * server receives source rkey/add/len
100 * server rdma reads "ping" data from source
101 * server sends "go ahead" on rdma read completion
102 * client sends sink rkey/addr/len
103 * server receives sink rkey/addr/len
104 * server rdma writes "pong" data to sink
105 * server sends "go ahead" on rdma write completion
110 * Default max buffer size for IO...
112 #define RPING_BUFSIZE 128*1024
113 #define RPING_SQ_DEPTH 32
116 /* lifted from netinet/libalias/alias_proxy.c */
117 static int inet_aton(const char *cp, struct in_addr *addr);
121 struct in_addr *addr;
129 c = (const char *)cp;
132 * Run through the string, grabbing numbers until
133 * the end of the string, or some error
139 l = strtoul(c, &endptr, 0);
141 if (l == ULONG_MAX || (l == 0 && endptr == c))
146 * If the whole string is invalid, endptr will equal
147 * c.. this way we can make sure someone hasn't
148 * gone '.12' or something which would get past
156 /* Check the next character past the previous number's end */
159 /* Make sure we only do 3 dots .. */
160 if (n == 3) /* Whoops. Quit. */
171 if (isspace((unsigned char)*c)) {
175 return (0); /* Invalid character, so fail */
181 * Concoct the address according to
182 * the number of parts specified.
186 case 0: /* a -- 32 bits */
188 * Nothing is necessary here. Overflow checking was
189 * already done in strtoul().
192 case 1: /* a.b -- 8.24 bits */
193 if (val > 0xffffff || parts[0] > 0xff)
195 val |= parts[0] << 24;
198 case 2: /* a.b.c -- 8.8.16 bits */
199 if (val > 0xffff || parts[0] > 0xff || parts[1] > 0xff)
201 val |= (parts[0] << 24) | (parts[1] << 16);
204 case 3: /* a.b.c.d -- 8.8.8.8 bits */
205 if (val > 0xff || parts[0] > 0xff || parts[1] > 0xff ||
208 val |= (parts[0] << 24) | (parts[1] << 16) | (parts[2] << 8);
213 addr->s_addr = htonl(val);
218 static void krping_wait(struct krping_cb *cb, int state)
222 while (cb->state < state) {
223 rc = msleep(cb, &cb->lock, 0, "krping", 0);
224 if (rc && rc != ERESTART) {
229 mtx_unlock(&cb->lock);
232 static int krping_cma_event_handler(struct rdma_cm_id *cma_id,
233 struct rdma_cm_event *event)
236 struct krping_cb *cb = cma_id->context;
238 DEBUG_LOG(PFX "cma_event type %d cma_id %p (%s)\n", event->event, cma_id,
239 (cma_id == cb->cm_id) ? "parent" : "child");
242 switch (event->event) {
243 case RDMA_CM_EVENT_ADDR_RESOLVED:
244 cb->state = ADDR_RESOLVED;
245 ret = rdma_resolve_route(cma_id, 2000);
247 log(LOG_ERR, "rdma_resolve_route error %d\n",
253 case RDMA_CM_EVENT_ROUTE_RESOLVED:
254 cb->state = ROUTE_RESOLVED;
258 case RDMA_CM_EVENT_CONNECT_REQUEST:
259 cb->state = CONNECT_REQUEST;
260 cb->child_cm_id = cma_id;
261 DEBUG_LOG(PFX "child cma %p\n", cb->child_cm_id);
265 case RDMA_CM_EVENT_ESTABLISHED:
266 DEBUG_LOG(PFX "ESTABLISHED\n");
268 cb->state = CONNECTED;
273 case RDMA_CM_EVENT_ADDR_ERROR:
274 case RDMA_CM_EVENT_ROUTE_ERROR:
275 case RDMA_CM_EVENT_CONNECT_ERROR:
276 case RDMA_CM_EVENT_UNREACHABLE:
277 case RDMA_CM_EVENT_REJECTED:
278 log(LOG_ERR, "cma event %d, error %d\n", event->event,
284 case RDMA_CM_EVENT_DISCONNECTED:
285 DEBUG_LOG(PFX "DISCONNECT EVENT...\n");
290 case RDMA_CM_EVENT_DEVICE_REMOVAL:
291 DEBUG_LOG(PFX "cma detected device removal!!!!\n");
295 log(LOG_ERR, "oof bad type!\n");
299 mtx_unlock(&cb->lock);
303 static int server_recv(struct krping_cb *cb, struct ib_wc *wc)
305 if (wc->byte_len != sizeof(cb->recv_buf)) {
306 log(LOG_ERR, "Received bogus data, size %d\n",
311 cb->remote_rkey = ntohl(cb->recv_buf.rkey);
312 cb->remote_addr = ntohll(cb->recv_buf.buf);
313 cb->remote_len = ntohl(cb->recv_buf.size);
314 DEBUG_LOG(PFX "Received rkey %x addr %llx len %d from peer\n",
315 cb->remote_rkey, (unsigned long long)cb->remote_addr,
318 if (cb->state <= CONNECTED || cb->state == RDMA_WRITE_COMPLETE)
319 cb->state = RDMA_READ_ADV;
321 cb->state = RDMA_WRITE_ADV;
326 static int client_recv(struct krping_cb *cb, struct ib_wc *wc)
328 if (wc->byte_len != sizeof(cb->recv_buf)) {
329 log(LOG_ERR, "Received bogus data, size %d\n",
334 if (cb->state == RDMA_READ_ADV)
335 cb->state = RDMA_WRITE_ADV;
337 cb->state = RDMA_WRITE_COMPLETE;
342 static void krping_cq_event_handler(struct ib_cq *cq, void *ctx)
344 struct krping_cb *cb = ctx;
346 struct ib_recv_wr *bad_wr;
350 KASSERT(cb->cq == cq, ("bad condition"));
351 if (cb->state == ERROR) {
352 log(LOG_ERR, "cq completion in ERROR state\n");
353 mtx_unlock(&cb->lock);
356 if (!cb->wlat && !cb->rlat && !cb->bw)
357 ib_req_notify_cq(cb->cq, IB_CQ_NEXT_COMP);
358 while ((ret = ib_poll_cq(cb->cq, 1, &wc)) == 1) {
360 if (wc.status != IB_WC_WR_FLUSH_ERR)
361 log(LOG_ERR, "cq completion failed status %d\n",
368 DEBUG_LOG(PFX "send completion\n");
369 cb->stats.send_bytes += cb->send_sgl.length;
370 cb->stats.send_msgs++;
373 case IB_WC_RDMA_WRITE:
374 DEBUG_LOG(PFX "rdma write completion\n");
375 cb->stats.write_bytes += cb->rdma_sq_wr.sg_list->length;
376 cb->stats.write_msgs++;
377 cb->state = RDMA_WRITE_COMPLETE;
381 case IB_WC_RDMA_READ:
382 DEBUG_LOG(PFX "rdma read completion\n");
383 cb->stats.read_bytes += cb->rdma_sq_wr.sg_list->length;
384 cb->stats.read_msgs++;
385 cb->state = RDMA_READ_COMPLETE;
390 DEBUG_LOG(PFX "recv completion\n");
391 cb->stats.recv_bytes += sizeof(cb->recv_buf);
392 cb->stats.recv_msgs++;
393 if (cb->wlat || cb->rlat || cb->bw)
394 ret = server_recv(cb, &wc);
396 ret = cb->server ? server_recv(cb, &wc) :
397 client_recv(cb, &wc);
399 log(LOG_ERR, "recv wc error: %d\n", ret);
403 ret = ib_post_recv(cb->qp, &cb->rq_wr, &bad_wr);
405 log(LOG_ERR, "post recv error: %d\n",
413 log(LOG_ERR, "unknown!!!!! completion\n");
418 log(LOG_ERR, "poll error %d\n", ret);
421 mtx_unlock(&cb->lock);
426 mtx_unlock(&cb->lock);
429 static int krping_accept(struct krping_cb *cb)
431 struct rdma_conn_param conn_param;
434 DEBUG_LOG(PFX "accepting client connection request\n");
436 memset(&conn_param, 0, sizeof conn_param);
437 conn_param.responder_resources = 1;
438 conn_param.initiator_depth = 1;
440 ret = rdma_accept(cb->child_cm_id, &conn_param);
442 log(LOG_ERR, "rdma_accept error: %d\n", ret);
446 if (!cb->wlat && !cb->rlat && !cb->bw) {
447 krping_wait(cb, CONNECTED);
448 if (cb->state == ERROR) {
449 log(LOG_ERR, "wait for CONNECTED state %d\n", cb->state);
456 static void krping_setup_wr(struct krping_cb *cb)
458 /* XXX X86 only here... not mapping for dma! */
459 cb->recv_sgl.addr = vtophys(&cb->recv_buf);
460 cb->recv_sgl.length = sizeof cb->recv_buf;
462 cb->recv_sgl.lkey = cb->dma_mr->lkey;
464 cb->recv_sgl.lkey = cb->recv_mr->lkey;
465 cb->rq_wr.sg_list = &cb->recv_sgl;
466 cb->rq_wr.num_sge = 1;
468 cb->send_sgl.addr = vtophys(&cb->send_buf);
469 cb->send_sgl.length = sizeof cb->send_buf;
471 cb->send_sgl.lkey = cb->dma_mr->lkey;
473 cb->send_sgl.lkey = cb->send_mr->lkey;
475 cb->sq_wr.opcode = IB_WR_SEND;
476 cb->sq_wr.send_flags = IB_SEND_SIGNALED;
477 cb->sq_wr.sg_list = &cb->send_sgl;
478 cb->sq_wr.num_sge = 1;
480 cb->rdma_addr = vtophys(cb->rdma_buf);
481 cb->rdma_sgl.addr = cb->rdma_addr;
483 cb->rdma_sgl.lkey = cb->dma_mr->lkey;
485 cb->rdma_sgl.lkey = cb->rdma_mr->lkey;
486 cb->rdma_sq_wr.send_flags = IB_SEND_SIGNALED;
487 cb->rdma_sq_wr.sg_list = &cb->rdma_sgl;
488 cb->rdma_sq_wr.num_sge = 1;
490 if (!cb->server || cb->wlat || cb->rlat || cb->bw) {
491 cb->start_addr = vtophys(cb->start_buf);
495 static int krping_setup_buffers(struct krping_cb *cb)
498 struct ib_phys_buf buf;
501 DEBUG_LOG(PFX "krping_setup_buffers called on cb %p\n", cb);
504 cb->dma_mr = ib_get_dma_mr(cb->pd, IB_ACCESS_LOCAL_WRITE|
505 IB_ACCESS_REMOTE_READ|
506 IB_ACCESS_REMOTE_WRITE);
507 if (IS_ERR(cb->dma_mr)) {
508 log(LOG_ERR, "reg_dmamr failed\n");
509 return PTR_ERR(cb->dma_mr);
513 buf.addr = vtophys(&cb->recv_buf);
514 buf.size = sizeof cb->recv_buf;
515 iovbase = vtophys(&cb->recv_buf);
516 cb->recv_mr = ib_reg_phys_mr(cb->pd, &buf, 1,
517 IB_ACCESS_LOCAL_WRITE,
520 if (IS_ERR(cb->recv_mr)) {
521 log(LOG_ERR, "recv_buf reg_mr failed\n");
522 return PTR_ERR(cb->recv_mr);
525 buf.addr = vtophys(&cb->send_buf);
526 buf.size = sizeof cb->send_buf;
527 iovbase = vtophys(&cb->send_buf);
528 cb->send_mr = ib_reg_phys_mr(cb->pd, &buf, 1,
531 if (IS_ERR(cb->send_mr)) {
532 log(LOG_ERR, "send_buf reg_mr failed\n");
533 ib_dereg_mr(cb->recv_mr);
534 return PTR_ERR(cb->send_mr);
538 cb->rdma_buf = contigmalloc(cb->size, M_DEVBUF, M_WAITOK, 0, -1UL,
542 log(LOG_ERR, "rdma_buf malloc failed\n");
546 if (!cb->use_dmamr) {
548 buf.addr = vtophys(cb->rdma_buf);
550 iovbase = vtophys(cb->rdma_buf);
551 cb->rdma_mr = ib_reg_phys_mr(cb->pd, &buf, 1,
552 IB_ACCESS_REMOTE_READ|
553 IB_ACCESS_REMOTE_WRITE,
556 if (IS_ERR(cb->rdma_mr)) {
557 log(LOG_ERR, "rdma_buf reg_mr failed\n");
558 ret = PTR_ERR(cb->rdma_mr);
563 if (!cb->server || cb->wlat || cb->rlat || cb->bw) {
564 cb->start_buf = contigmalloc(cb->size, M_DEVBUF, M_WAITOK,
565 0, -1UL, PAGE_SIZE, 0);
566 if (!cb->start_buf) {
567 log(LOG_ERR, "start_buf malloc failed\n");
571 if (!cb->use_dmamr) {
572 unsigned flags = IB_ACCESS_REMOTE_READ;
574 if (cb->wlat || cb->rlat || cb->bw)
575 flags |= IB_ACCESS_REMOTE_WRITE;
576 buf.addr = vtophys(cb->start_buf);
578 iovbase = vtophys(cb->start_buf);
579 cb->start_mr = ib_reg_phys_mr(cb->pd, &buf, 1,
583 if (IS_ERR(cb->start_mr)) {
584 log(LOG_ERR, "start_buf reg_mr failed\n");
585 ret = PTR_ERR(cb->start_mr);
592 DEBUG_LOG(PFX "allocated & registered buffers...\n");
595 contigfree(cb->start_buf, cb->size, M_DEVBUF);
598 ib_dereg_mr(cb->rdma_mr);
600 contigfree(cb->rdma_buf, cb->size, M_DEVBUF);
603 ib_dereg_mr(cb->dma_mr);
605 ib_dereg_mr(cb->recv_mr);
606 ib_dereg_mr(cb->send_mr);
611 static void krping_free_buffers(struct krping_cb *cb)
613 DEBUG_LOG(PFX "krping_free_buffers called on cb %p\n", cb);
616 dma_unmap_single(cb->pd->device->dma_device,
617 pci_unmap_addr(cb, recv_mapping),
618 sizeof(cb->recv_buf), DMA_BIDIRECTIONAL);
619 dma_unmap_single(cb->pd->device->dma_device,
620 pci_unmap_addr(cb, send_mapping),
621 sizeof(cb->send_buf), DMA_BIDIRECTIONAL);
622 dma_unmap_single(cb->pd->device->dma_device,
623 pci_unmap_addr(cb, rdma_mapping),
624 cb->size, DMA_BIDIRECTIONAL);
626 contigfree(cb->rdma_buf, cb->size, M_DEVBUF);
627 if (!cb->server || cb->wlat || cb->rlat || cb->bw) {
629 dma_unmap_single(cb->pd->device->dma_device,
630 pci_unmap_addr(cb, start_mapping),
631 cb->size, DMA_BIDIRECTIONAL);
633 contigfree(cb->start_buf, cb->size, M_DEVBUF);
636 ib_dereg_mr(cb->dma_mr);
638 ib_dereg_mr(cb->send_mr);
639 ib_dereg_mr(cb->recv_mr);
640 ib_dereg_mr(cb->rdma_mr);
642 ib_dereg_mr(cb->start_mr);
646 static int krping_create_qp(struct krping_cb *cb)
648 struct ib_qp_init_attr init_attr;
651 memset(&init_attr, 0, sizeof(init_attr));
652 init_attr.cap.max_send_wr = cb->txdepth;
653 init_attr.cap.max_recv_wr = 2;
654 init_attr.cap.max_recv_sge = 1;
655 init_attr.cap.max_send_sge = 1;
656 init_attr.qp_type = IB_QPT_RC;
657 init_attr.send_cq = cb->cq;
658 init_attr.recv_cq = cb->cq;
661 ret = rdma_create_qp(cb->child_cm_id, cb->pd, &init_attr);
663 cb->qp = cb->child_cm_id->qp;
665 ret = rdma_create_qp(cb->cm_id, cb->pd, &init_attr);
667 cb->qp = cb->cm_id->qp;
673 static void krping_free_qp(struct krping_cb *cb)
675 ib_destroy_qp(cb->qp);
676 ib_destroy_cq(cb->cq);
677 ib_dealloc_pd(cb->pd);
680 static int krping_setup_qp(struct krping_cb *cb, struct rdma_cm_id *cm_id)
683 cb->pd = ib_alloc_pd(cm_id->device);
684 if (IS_ERR(cb->pd)) {
685 log(LOG_ERR, "ib_alloc_pd failed\n");
686 return PTR_ERR(cb->pd);
688 DEBUG_LOG(PFX "created pd %p\n", cb->pd);
690 cb->cq = ib_create_cq(cm_id->device, krping_cq_event_handler, NULL,
691 cb, cb->txdepth * 2, 0);
692 if (IS_ERR(cb->cq)) {
693 log(LOG_ERR, "ib_create_cq failed\n");
694 ret = PTR_ERR(cb->cq);
697 DEBUG_LOG(PFX "created cq %p\n", cb->cq);
699 if (!cb->wlat && !cb->rlat && !cb->bw) {
700 ret = ib_req_notify_cq(cb->cq, IB_CQ_NEXT_COMP);
702 log(LOG_ERR, "ib_create_cq failed\n");
707 ret = krping_create_qp(cb);
709 log(LOG_ERR, "krping_create_qp failed: %d\n", ret);
712 DEBUG_LOG(PFX "created qp %p\n", cb->qp);
715 ib_destroy_cq(cb->cq);
717 ib_dealloc_pd(cb->pd);
721 static void krping_format_send(struct krping_cb *cb, u64 buf,
724 struct krping_rdma_info *info = &cb->send_buf;
726 info->buf = htonll(buf);
727 info->rkey = htonl(mr->rkey);
728 info->size = htonl(cb->size);
730 DEBUG_LOG(PFX "RDMA addr %llx rkey %x len %d\n",
731 (unsigned long long)buf, mr->rkey, cb->size);
734 static void krping_test_server(struct krping_cb *cb)
736 struct ib_send_wr *bad_wr;
740 /* Wait for client's Start STAG/TO/Len */
741 krping_wait(cb, RDMA_READ_ADV);
742 if (cb->state != RDMA_READ_ADV) {
743 DEBUG_LOG(PFX "wait for RDMA_READ_ADV state %d\n",
748 DEBUG_LOG(PFX "server received sink adv\n");
750 /* Issue RDMA Read. */
751 cb->rdma_sq_wr.opcode = IB_WR_RDMA_READ;
752 cb->rdma_sq_wr.wr.rdma.rkey = cb->remote_rkey;
753 cb->rdma_sq_wr.wr.rdma.remote_addr = cb->remote_addr;
754 cb->rdma_sq_wr.sg_list->length = cb->remote_len;
756 ret = ib_post_send(cb->qp, &cb->rdma_sq_wr, &bad_wr);
758 log(LOG_ERR, "post send error %d\n", ret);
761 DEBUG_LOG(PFX "server posted rdma read req \n");
763 /* Wait for read completion */
764 krping_wait(cb, RDMA_READ_COMPLETE);
765 if (cb->state != RDMA_READ_COMPLETE) {
767 "wait for RDMA_READ_COMPLETE state %d\n",
771 DEBUG_LOG(PFX "server received read complete\n");
773 /* Display data in recv buf */
775 DEBUG_LOG("server ping data: %s\n", cb->rdma_buf);
777 /* Tell client to continue */
778 ret = ib_post_send(cb->qp, &cb->sq_wr, &bad_wr);
780 log(LOG_ERR, "post send error %d\n", ret);
783 DEBUG_LOG(PFX "server posted go ahead\n");
785 /* Wait for client's RDMA STAG/TO/Len */
786 krping_wait(cb, RDMA_WRITE_ADV);
787 if (cb->state != RDMA_WRITE_ADV) {
789 "wait for RDMA_WRITE_ADV state %d\n",
793 DEBUG_LOG(PFX "server received sink adv\n");
795 /* RDMA Write echo data */
796 cb->rdma_sq_wr.opcode = IB_WR_RDMA_WRITE;
797 cb->rdma_sq_wr.wr.rdma.rkey = cb->remote_rkey;
798 cb->rdma_sq_wr.wr.rdma.remote_addr = cb->remote_addr;
799 cb->rdma_sq_wr.sg_list->length = strlen(cb->rdma_buf) + 1;
800 DEBUG_LOG(PFX "rdma write from lkey %x laddr %llx len %d\n",
801 cb->rdma_sq_wr.sg_list->lkey,
802 (unsigned long long)cb->rdma_sq_wr.sg_list->addr,
803 cb->rdma_sq_wr.sg_list->length);
805 ret = ib_post_send(cb->qp, &cb->rdma_sq_wr, &bad_wr);
807 log(LOG_ERR, "post send error %d\n", ret);
811 /* Wait for completion */
812 krping_wait(cb, RDMA_WRITE_COMPLETE);
813 if (cb->state != RDMA_WRITE_COMPLETE) {
815 "wait for RDMA_WRITE_COMPLETE state %d\n",
819 DEBUG_LOG(PFX "server rdma write complete \n");
821 cb->state = CONNECTED;
823 /* Tell client to begin again */
824 ret = ib_post_send(cb->qp, &cb->sq_wr, &bad_wr);
826 log(LOG_ERR, "post send error %d\n", ret);
829 DEBUG_LOG(PFX "server posted go ahead\n");
833 static void rlat_test(struct krping_cb *cb)
836 int iters = cb->count;
837 struct timeval start_tv, stop_tv;
840 struct ib_send_wr *bad_wr;
844 cb->rdma_sq_wr.opcode = IB_WR_RDMA_READ;
845 cb->rdma_sq_wr.wr.rdma.rkey = cb->remote_rkey;
846 cb->rdma_sq_wr.wr.rdma.remote_addr = cb->remote_addr;
847 cb->rdma_sq_wr.sg_list->length = cb->size;
849 microtime(&start_tv);
851 cb->state = RDMA_READ_ADV;
852 ib_req_notify_cq(cb->cq, IB_CQ_NEXT_COMP);
854 while (scnt < iters) {
856 cb->state = RDMA_READ_ADV;
857 ret = ib_post_send(cb->qp, &cb->rdma_sq_wr, &bad_wr);
860 "Couldn't post send: ret=%d scnt %d\n",
867 krping_wait(cb, RDMA_READ_COMPLETE);
868 if (cb->state == RDMA_READ_COMPLETE) {
870 ib_req_notify_cq(cb->cq, IB_CQ_NEXT_COMP);
875 ne = ib_poll_cq(cb->cq, 1, &wc);
876 if (cb->state == ERROR) {
878 "state == ERROR...bailing scnt %d\n", scnt);
884 log(LOG_ERR, "poll CQ failed %d\n", ne);
887 if (cb->poll && wc.status != IB_WC_SUCCESS) {
888 log(LOG_ERR, "Completion wth error at %s:\n",
889 cb->server ? "server" : "client");
890 log(LOG_ERR, "Failed status %d: wr_id %d\n",
891 wc.status, (int) wc.wr_id);
898 if (stop_tv.tv_usec < start_tv.tv_usec) {
899 stop_tv.tv_usec += 1000000;
903 log(LOG_ERR, "delta sec %zu delta usec %lu iter %d size %d\n",
904 stop_tv.tv_sec - start_tv.tv_sec,
905 stop_tv.tv_usec - start_tv.tv_usec,
909 static int alloc_cycle_mem(int cycle_iters,
910 cycles_t **post_cycles_start,
911 cycles_t **post_cycles_stop,
912 cycles_t **poll_cycles_start,
913 cycles_t **poll_cycles_stop,
914 cycles_t **last_poll_cycles_start)
916 *post_cycles_start = malloc(cycle_iters * sizeof(cycles_t), M_DEVBUF, M_WAITOK);
917 if (!*post_cycles_start) {
920 *post_cycles_stop = malloc(cycle_iters * sizeof(cycles_t), M_DEVBUF, M_WAITOK);
921 if (!*post_cycles_stop) {
924 *poll_cycles_start = malloc(cycle_iters * sizeof(cycles_t), M_DEVBUF, M_WAITOK);
925 if (!*poll_cycles_start) {
928 *poll_cycles_stop = malloc(cycle_iters * sizeof(cycles_t), M_DEVBUF, M_WAITOK);
929 if (!*poll_cycles_stop) {
932 *last_poll_cycles_start = malloc(cycle_iters * sizeof(cycles_t), M_DEVBUF, M_WAITOK);
933 if (!*last_poll_cycles_start) {
938 free(*poll_cycles_stop, M_DEVBUF);
940 free(*poll_cycles_start, M_DEVBUF);
942 free(*post_cycles_stop, M_DEVBUF);
944 free(*post_cycles_start, M_DEVBUF);
946 log(LOG_ERR, "%s malloc failed\n", __FUNCTION__);
950 static void free_cycle_mem(cycles_t *post_cycles_start,
951 cycles_t *post_cycles_stop,
952 cycles_t *poll_cycles_start,
953 cycles_t *poll_cycles_stop,
954 cycles_t *last_poll_cycles_start)
956 free(last_poll_cycles_start, M_DEVBUF);
957 free(poll_cycles_stop, M_DEVBUF);
958 free(poll_cycles_start, M_DEVBUF);
959 free(post_cycles_stop, M_DEVBUF);
960 free(post_cycles_start, M_DEVBUF);
963 static void wlat_test(struct krping_cb *cb)
965 int ccnt, scnt, rcnt;
967 volatile char *poll_buf = (char *) cb->start_buf;
968 char *buf = (char *)cb->rdma_buf;
972 struct timeval start_tv, stop_tv;
973 cycles_t *post_cycles_start, *post_cycles_stop;
974 cycles_t *poll_cycles_start, *poll_cycles_stop;
975 cycles_t *last_poll_cycles_start;
976 cycles_t sum_poll = 0, sum_post = 0, sum_last_poll = 0;
978 int cycle_iters = 1000;
981 err = alloc_cycle_mem(cycle_iters, &post_cycles_start, &post_cycles_stop,
982 &poll_cycles_start, &poll_cycles_stop, &last_poll_cycles_start);
985 log(LOG_ERR, "%s malloc failed\n", __FUNCTION__);
989 cb->rdma_sq_wr.opcode = IB_WR_RDMA_WRITE;
990 cb->rdma_sq_wr.wr.rdma.rkey = cb->remote_rkey;
991 cb->rdma_sq_wr.wr.rdma.remote_addr = cb->remote_addr;
992 cb->rdma_sq_wr.sg_list->length = cb->size;
994 if (cycle_iters > iters)
996 microtime(&start_tv);
997 while (scnt < iters || ccnt < iters || rcnt < iters) {
999 /* Wait till buffer changes. */
1000 if (rcnt < iters && !(scnt < 1 && !cb->server)) {
1002 while (*poll_buf != (char)rcnt) {
1003 if (cb->state == ERROR) {
1004 log(LOG_ERR, "state = ERROR, bailing\n");
1011 struct ib_send_wr *bad_wr;
1013 *buf = (char)scnt+1;
1014 if (scnt < cycle_iters)
1015 post_cycles_start[scnt] = get_cycles();
1016 if (ib_post_send(cb->qp, &cb->rdma_sq_wr, &bad_wr)) {
1017 log(LOG_ERR, "Couldn't post send: scnt=%d\n",
1021 if (scnt < cycle_iters)
1022 post_cycles_stop[scnt] = get_cycles();
1030 if (ccnt < cycle_iters)
1031 poll_cycles_start[ccnt] = get_cycles();
1033 if (ccnt < cycle_iters)
1034 last_poll_cycles_start[ccnt] = get_cycles();
1035 ne = ib_poll_cq(cb->cq, 1, &wc);
1037 if (ccnt < cycle_iters)
1038 poll_cycles_stop[ccnt] = get_cycles();
1042 log(LOG_ERR, "poll CQ failed %d\n", ne);
1045 if (wc.status != IB_WC_SUCCESS) {
1046 log(LOG_ERR, "Completion wth error at %s:\n",
1047 cb->server ? "server" : "client");
1048 log(LOG_ERR, "Failed status %d: wr_id %d\n",
1049 wc.status, (int) wc.wr_id);
1050 log(LOG_ERR, "scnt=%d, rcnt=%d, ccnt=%d\n",
1056 microtime(&stop_tv);
1058 if (stop_tv.tv_usec < start_tv.tv_usec) {
1059 stop_tv.tv_usec += 1000000;
1060 stop_tv.tv_sec -= 1;
1063 for (i=0; i < cycle_iters; i++) {
1064 sum_post += post_cycles_stop[i] - post_cycles_start[i];
1065 sum_poll += poll_cycles_stop[i] - poll_cycles_start[i];
1066 sum_last_poll += poll_cycles_stop[i] - last_poll_cycles_start[i];
1069 log(LOG_ERR, "delta sec %zu delta usec %lu iter %d size %d cycle_iters %d sum_post %llu sum_poll %llu sum_last_poll %llu\n",
1070 stop_tv.tv_sec - start_tv.tv_sec,
1071 stop_tv.tv_usec - start_tv.tv_usec,
1072 scnt, cb->size, cycle_iters,
1073 (unsigned long long)sum_post, (unsigned long long)sum_poll,
1074 (unsigned long long)sum_last_poll);
1076 free_cycle_mem(post_cycles_start, post_cycles_stop, poll_cycles_start,
1077 poll_cycles_stop, last_poll_cycles_start);
1080 static void bw_test(struct krping_cb *cb)
1082 int ccnt, scnt, rcnt;
1083 int iters=cb->count;
1087 struct timeval start_tv, stop_tv;
1088 cycles_t *post_cycles_start, *post_cycles_stop;
1089 cycles_t *poll_cycles_start, *poll_cycles_stop;
1090 cycles_t *last_poll_cycles_start;
1091 cycles_t sum_poll = 0, sum_post = 0, sum_last_poll = 0;
1093 int cycle_iters = 1000;
1096 err = alloc_cycle_mem(cycle_iters, &post_cycles_start, &post_cycles_stop,
1097 &poll_cycles_start, &poll_cycles_stop, &last_poll_cycles_start);
1100 log(LOG_ERR, "%s kmalloc failed\n", __FUNCTION__);
1104 cb->rdma_sq_wr.opcode = IB_WR_RDMA_WRITE;
1105 cb->rdma_sq_wr.wr.rdma.rkey = cb->remote_rkey;
1106 cb->rdma_sq_wr.wr.rdma.remote_addr = cb->remote_addr;
1107 cb->rdma_sq_wr.sg_list->length = cb->size;
1109 if (cycle_iters > iters)
1110 cycle_iters = iters;
1111 microtime(&start_tv);
1112 while (scnt < iters || ccnt < iters) {
1114 while (scnt < iters && scnt - ccnt < cb->txdepth) {
1115 struct ib_send_wr *bad_wr;
1117 if (scnt < cycle_iters)
1118 post_cycles_start[scnt] = get_cycles();
1119 if (ib_post_send(cb->qp, &cb->rdma_sq_wr, &bad_wr)) {
1120 log(LOG_ERR, "Couldn't post send: scnt=%d\n",
1124 if (scnt < cycle_iters)
1125 post_cycles_stop[scnt] = get_cycles();
1133 if (ccnt < cycle_iters)
1134 poll_cycles_start[ccnt] = get_cycles();
1136 if (ccnt < cycle_iters)
1137 last_poll_cycles_start[ccnt] = get_cycles();
1138 ne = ib_poll_cq(cb->cq, 1, &wc);
1140 if (ccnt < cycle_iters)
1141 poll_cycles_stop[ccnt] = get_cycles();
1145 log(LOG_ERR, "poll CQ failed %d\n", ne);
1148 if (wc.status != IB_WC_SUCCESS) {
1149 log(LOG_ERR, "Completion wth error at %s:\n",
1150 cb->server ? "server" : "client");
1151 log(LOG_ERR, "Failed status %d: wr_id %d\n",
1152 wc.status, (int) wc.wr_id);
1157 microtime(&stop_tv);
1159 if (stop_tv.tv_usec < start_tv.tv_usec) {
1160 stop_tv.tv_usec += 1000000;
1161 stop_tv.tv_sec -= 1;
1164 for (i=0; i < cycle_iters; i++) {
1165 sum_post += post_cycles_stop[i] - post_cycles_start[i];
1166 sum_poll += poll_cycles_stop[i] - poll_cycles_start[i];
1167 sum_last_poll += poll_cycles_stop[i] - last_poll_cycles_start[i];
1170 log(LOG_ERR, "delta sec %zu delta usec %lu iter %d size %d cycle_iters %d sum_post %llu sum_poll %llu sum_last_poll %llu\n",
1171 stop_tv.tv_sec - start_tv.tv_sec,
1172 stop_tv.tv_usec - start_tv.tv_usec,
1173 scnt, cb->size, cycle_iters,
1174 (unsigned long long)sum_post, (unsigned long long)sum_poll,
1175 (unsigned long long)sum_last_poll);
1177 free_cycle_mem(post_cycles_start, post_cycles_stop, poll_cycles_start,
1178 poll_cycles_stop, last_poll_cycles_start);
1181 static void krping_rlat_test_server(struct krping_cb *cb)
1183 struct ib_send_wr *bad_wr;
1187 /* Spin waiting for client's Start STAG/TO/Len */
1188 while (cb->state < RDMA_READ_ADV) {
1189 krping_cq_event_handler(cb->cq, cb);
1192 /* Send STAG/TO/Len to client */
1194 krping_format_send(cb, cb->start_addr, cb->dma_mr);
1196 krping_format_send(cb, cb->start_addr, cb->start_mr);
1197 ret = ib_post_send(cb->qp, &cb->sq_wr, &bad_wr);
1199 log(LOG_ERR, "post send error %d\n", ret);
1203 /* Spin waiting for send completion */
1204 while ((ret = ib_poll_cq(cb->cq, 1, &wc) == 0));
1206 log(LOG_ERR, "poll error %d\n", ret);
1210 log(LOG_ERR, "send completiong error %d\n", wc.status);
1214 krping_wait(cb, ERROR);
1217 static void krping_wlat_test_server(struct krping_cb *cb)
1219 struct ib_send_wr *bad_wr;
1223 /* Spin waiting for client's Start STAG/TO/Len */
1224 while (cb->state < RDMA_READ_ADV) {
1225 krping_cq_event_handler(cb->cq, cb);
1228 /* Send STAG/TO/Len to client */
1230 krping_format_send(cb, cb->start_addr, cb->dma_mr);
1232 krping_format_send(cb, cb->start_addr, cb->start_mr);
1233 ret = ib_post_send(cb->qp, &cb->sq_wr, &bad_wr);
1235 log(LOG_ERR, "post send error %d\n", ret);
1239 /* Spin waiting for send completion */
1240 while ((ret = ib_poll_cq(cb->cq, 1, &wc) == 0));
1242 log(LOG_ERR, "poll error %d\n", ret);
1246 log(LOG_ERR, "send completiong error %d\n", wc.status);
1254 static void krping_bw_test_server(struct krping_cb *cb)
1256 struct ib_send_wr *bad_wr;
1260 /* Spin waiting for client's Start STAG/TO/Len */
1261 while (cb->state < RDMA_READ_ADV) {
1262 krping_cq_event_handler(cb->cq, cb);
1265 /* Send STAG/TO/Len to client */
1267 krping_format_send(cb, cb->start_addr, cb->dma_mr);
1269 krping_format_send(cb, cb->start_addr, cb->start_mr);
1270 ret = ib_post_send(cb->qp, &cb->sq_wr, &bad_wr);
1272 log(LOG_ERR, "post send error %d\n", ret);
1276 /* Spin waiting for send completion */
1277 while ((ret = ib_poll_cq(cb->cq, 1, &wc) == 0));
1279 log(LOG_ERR, "poll error %d\n", ret);
1283 log(LOG_ERR, "send completiong error %d\n", wc.status);
1289 krping_wait(cb, ERROR);
1292 static int krping_bind_server(struct krping_cb *cb)
1294 struct sockaddr_in sin;
1297 memset(&sin, 0, sizeof(sin));
1298 sin.sin_len = sizeof sin;
1299 sin.sin_family = AF_INET;
1300 sin.sin_addr.s_addr = cb->addr.s_addr;
1301 sin.sin_port = cb->port;
1303 ret = rdma_bind_addr(cb->cm_id, (struct sockaddr *) &sin);
1305 log(LOG_ERR, "rdma_bind_addr error %d\n", ret);
1308 DEBUG_LOG(PFX "rdma_bind_addr successful\n");
1310 DEBUG_LOG(PFX "rdma_listen\n");
1311 ret = rdma_listen(cb->cm_id, 3);
1313 log(LOG_ERR, "rdma_listen failed: %d\n", ret);
1317 krping_wait(cb, CONNECT_REQUEST);
1318 if (cb->state != CONNECT_REQUEST) {
1319 log(LOG_ERR, "wait for CONNECT_REQUEST state %d\n",
1327 static void krping_run_server(struct krping_cb *cb)
1329 struct ib_recv_wr *bad_wr;
1332 ret = krping_bind_server(cb);
1336 ret = krping_setup_qp(cb, cb->child_cm_id);
1338 log(LOG_ERR, "setup_qp failed: %d\n", ret);
1342 ret = krping_setup_buffers(cb);
1344 log(LOG_ERR, "krping_setup_buffers failed: %d\n", ret);
1348 ret = ib_post_recv(cb->qp, &cb->rq_wr, &bad_wr);
1350 log(LOG_ERR, "ib_post_recv failed: %d\n", ret);
1354 ret = krping_accept(cb);
1356 log(LOG_ERR, "connect error %d\n", ret);
1361 krping_wlat_test_server(cb);
1363 krping_rlat_test_server(cb);
1365 krping_bw_test_server(cb);
1367 krping_test_server(cb);
1369 rdma_disconnect(cb->child_cm_id);
1370 rdma_destroy_id(cb->child_cm_id);
1372 krping_free_buffers(cb);
1377 static void krping_test_client(struct krping_cb *cb)
1379 int ping, start, cc, i, ret;
1380 struct ib_send_wr *bad_wr;
1384 for (ping = 0; !cb->count || ping < cb->count; ping++) {
1385 cb->state = RDMA_READ_ADV;
1387 /* Put some ascii text in the buffer. */
1388 cc = sprintf(cb->start_buf, "rdma-ping-%d: ", ping);
1389 for (i = cc, c = start; i < cb->size; i++) {
1390 cb->start_buf[i] = c;
1398 cb->start_buf[cb->size - 1] = 0;
1401 krping_format_send(cb, cb->start_addr, cb->dma_mr);
1403 krping_format_send(cb, cb->start_addr, cb->start_mr);
1405 ret = ib_post_send(cb->qp, &cb->sq_wr, &bad_wr);
1407 log(LOG_ERR, "post send error %d\n", ret);
1411 /* Wait for server to ACK */
1412 krping_wait(cb, RDMA_WRITE_ADV);
1413 if (cb->state != RDMA_WRITE_ADV) {
1415 "wait for RDMA_WRITE_ADV state %d\n",
1421 krping_format_send(cb, cb->rdma_addr, cb->dma_mr);
1423 krping_format_send(cb, cb->rdma_addr, cb->rdma_mr);
1425 ret = ib_post_send(cb->qp, &cb->sq_wr, &bad_wr);
1427 log(LOG_ERR, "post send error %d\n", ret);
1431 /* Wait for the server to say the RDMA Write is complete. */
1432 krping_wait(cb, RDMA_WRITE_COMPLETE);
1433 if (cb->state != RDMA_WRITE_COMPLETE) {
1435 "wait for RDMA_WRITE_COMPLETE state %d\n",
1441 if (memcmp(cb->start_buf, cb->rdma_buf, cb->size)) {
1442 log(LOG_ERR, "data mismatch!\n");
1447 DEBUG_LOG("ping data: %s\n", cb->rdma_buf);
1451 static void krping_rlat_test_client(struct krping_cb *cb)
1453 struct ib_send_wr *bad_wr;
1457 cb->state = RDMA_READ_ADV;
1459 /* Send STAG/TO/Len to client */
1461 krping_format_send(cb, cb->start_addr, cb->dma_mr);
1463 krping_format_send(cb, cb->start_addr, cb->rdma_mr);
1464 ret = ib_post_send(cb->qp, &cb->sq_wr, &bad_wr);
1466 log(LOG_ERR, "post send error %d\n", ret);
1470 /* Spin waiting for send completion */
1471 while ((ret = ib_poll_cq(cb->cq, 1, &wc) == 0));
1473 log(LOG_ERR, "poll error %d\n", ret);
1477 log(LOG_ERR, "send completion error %d\n", wc.status);
1481 /* Spin waiting for server's Start STAG/TO/Len */
1482 while (cb->state < RDMA_WRITE_ADV) {
1483 krping_cq_event_handler(cb->cq, cb);
1489 struct timeval start, stop;
1492 unsigned long long elapsed;
1494 struct ib_send_wr *bad_wr;
1497 cb->rdma_sq_wr.opcode = IB_WR_RDMA_WRITE;
1498 cb->rdma_sq_wr.wr.rdma.rkey = cb->remote_rkey;
1499 cb->rdma_sq_wr.wr.rdma.remote_addr = cb->remote_addr;
1500 cb->rdma_sq_wr.sg_list->length = 0;
1501 cb->rdma_sq_wr.num_sge = 0;
1504 for (i=0; i < 100000; i++) {
1505 if (ib_post_send(cb->qp, &cb->rdma_sq_wr, &bad_wr)) {
1506 log(LOG_ERR, "Couldn't post send\n");
1510 ne = ib_poll_cq(cb->cq, 1, &wc);
1513 log(LOG_ERR, "poll CQ failed %d\n", ne);
1516 if (wc.status != IB_WC_SUCCESS) {
1517 log(LOG_ERR, "Completion wth error at %s:\n",
1518 cb->server ? "server" : "client");
1519 log(LOG_ERR, "Failed status %d: wr_id %d\n",
1520 wc.status, (int) wc.wr_id);
1526 if (stop.tv_usec < start.tv_usec) {
1527 stop.tv_usec += 1000000;
1530 sec = stop.tv_sec - start.tv_sec;
1531 usec = stop.tv_usec - start.tv_usec;
1532 elapsed = sec * 1000000 + usec;
1533 log(LOG_ERR, "0B-write-lat iters 100000 usec %llu\n", elapsed);
1540 static void krping_wlat_test_client(struct krping_cb *cb)
1542 struct ib_send_wr *bad_wr;
1546 cb->state = RDMA_READ_ADV;
1548 /* Send STAG/TO/Len to client */
1550 krping_format_send(cb, cb->start_addr, cb->dma_mr);
1552 krping_format_send(cb, cb->start_addr, cb->start_mr);
1553 ret = ib_post_send(cb->qp, &cb->sq_wr, &bad_wr);
1555 log(LOG_ERR, "post send error %d\n", ret);
1559 /* Spin waiting for send completion */
1560 while ((ret = ib_poll_cq(cb->cq, 1, &wc) == 0));
1562 log(LOG_ERR, "poll error %d\n", ret);
1566 log(LOG_ERR, "send completion error %d\n", wc.status);
1570 /* Spin waiting for server's Start STAG/TO/Len */
1571 while (cb->state < RDMA_WRITE_ADV) {
1572 krping_cq_event_handler(cb->cq, cb);
1578 static void krping_bw_test_client(struct krping_cb *cb)
1580 struct ib_send_wr *bad_wr;
1584 cb->state = RDMA_READ_ADV;
1586 /* Send STAG/TO/Len to client */
1588 krping_format_send(cb, cb->start_addr, cb->dma_mr);
1590 krping_format_send(cb, cb->start_addr, cb->start_mr);
1591 ret = ib_post_send(cb->qp, &cb->sq_wr, &bad_wr);
1593 log(LOG_ERR, "post send error %d\n", ret);
1597 /* Spin waiting for send completion */
1598 while ((ret = ib_poll_cq(cb->cq, 1, &wc) == 0));
1600 log(LOG_ERR, "poll error %d\n", ret);
1604 log(LOG_ERR, "send completion error %d\n", wc.status);
1608 /* Spin waiting for server's Start STAG/TO/Len */
1609 while (cb->state < RDMA_WRITE_ADV) {
1610 krping_cq_event_handler(cb->cq, cb);
1616 static int krping_connect_client(struct krping_cb *cb)
1618 struct rdma_conn_param conn_param;
1621 memset(&conn_param, 0, sizeof conn_param);
1622 conn_param.responder_resources = 1;
1623 conn_param.initiator_depth = 1;
1624 conn_param.retry_count = 10;
1626 ret = rdma_connect(cb->cm_id, &conn_param);
1628 log(LOG_ERR, "rdma_connect error %d\n", ret);
1632 krping_wait(cb, CONNECTED);
1633 if (cb->state == ERROR) {
1634 log(LOG_ERR, "wait for CONNECTED state %d\n", cb->state);
1638 DEBUG_LOG(PFX "rdma_connect successful\n");
1642 static int krping_bind_client(struct krping_cb *cb)
1644 struct sockaddr_in sin;
1647 memset(&sin, 0, sizeof(sin));
1648 sin.sin_len = sizeof sin;
1649 sin.sin_family = AF_INET;
1650 sin.sin_addr.s_addr = cb->addr.s_addr;
1651 sin.sin_port = cb->port;
1653 ret = rdma_resolve_addr(cb->cm_id, NULL, (struct sockaddr *) &sin,
1656 log(LOG_ERR, "rdma_resolve_addr error %d\n", ret);
1660 krping_wait(cb, ROUTE_RESOLVED);
1661 if (cb->state != ROUTE_RESOLVED) {
1663 "addr/route resolution did not resolve: state %d\n",
1668 DEBUG_LOG(PFX "rdma_resolve_addr - rdma_resolve_route successful\n");
1672 static void krping_run_client(struct krping_cb *cb)
1674 struct ib_recv_wr *bad_wr;
1677 ret = krping_bind_client(cb);
1681 ret = krping_setup_qp(cb, cb->cm_id);
1683 log(LOG_ERR, "setup_qp failed: %d\n", ret);
1687 ret = krping_setup_buffers(cb);
1689 log(LOG_ERR, "krping_setup_buffers failed: %d\n", ret);
1693 ret = ib_post_recv(cb->qp, &cb->rq_wr, &bad_wr);
1695 log(LOG_ERR, "ib_post_recv failed: %d\n", ret);
1699 ret = krping_connect_client(cb);
1701 log(LOG_ERR, "connect error %d\n", ret);
1706 krping_wlat_test_client(cb);
1708 krping_rlat_test_client(cb);
1710 krping_bw_test_client(cb);
1712 krping_test_client(cb);
1713 rdma_disconnect(cb->cm_id);
1715 krping_free_buffers(cb);
1720 int krping_doit(char *cmd)
1722 struct krping_cb *cb;
1726 unsigned long optint;
1729 cb = malloc(sizeof(*cb), M_DEVBUF, M_WAITOK);
1732 bzero(cb, sizeof *cb);
1734 mtx_lock(&krping_mutex);
1735 TAILQ_INSERT_TAIL(&krping_cbs, cb, list);
1736 mtx_unlock(&krping_mutex);
1741 cb->txdepth = RPING_SQ_DEPTH;
1742 mtx_init(&cb->lock, "krping mtx", NULL, MTX_DUPOK|MTX_DEF);
1744 while ((op = krping_getopt("krping", &cmd, krping_opts, NULL, &optarg,
1748 cb->addr_str = optarg;
1749 DEBUG_LOG(PFX "ipaddr (%s)\n", optarg);
1750 if (!inet_aton(optarg, &cb->addr)) {
1751 log(LOG_ERR, "bad addr string %s\n", optarg);
1757 DEBUG_LOG(PFX "using dma mr\n");
1760 cb->port = htons(optint);
1761 DEBUG_LOG(PFX "port %d\n", (int)optint);
1765 DEBUG_LOG("server\n");
1769 DEBUG_LOG(PFX "server\n");
1773 DEBUG_LOG(PFX "client\n");
1777 if ((cb->size < 1) ||
1778 (cb->size > RPING_BUFSIZE)) {
1779 log(LOG_ERR, "Invalid size %d "
1780 "(valid range is 1 to %d)\n",
1781 cb->size, RPING_BUFSIZE);
1784 DEBUG_LOG(PFX "size %d\n", (int)optint);
1788 if (cb->count < 0) {
1789 log(LOG_ERR, "Invalid count %d\n",
1793 DEBUG_LOG(PFX "count %d\n", (int) cb->count);
1797 DEBUG_LOG(PFX "verbose\n");
1801 DEBUG_LOG(PFX "validate data\n");
1813 cb->txdepth = optint;
1814 DEBUG_LOG(PFX "txdepth %d\n", cb->txdepth);
1820 log(LOG_ERR, "unknown opt %s\n", optarg);
1828 if (cb->server == -1) {
1829 log(LOG_ERR, "must be either client or server\n");
1833 if ((cb->bw + cb->rlat + cb->wlat) > 1) {
1834 log(LOG_ERR, "Pick only one test: bw, rlat, wlat\n");
1840 cb->cm_id = rdma_create_id(krping_cma_event_handler, cb, RDMA_PS_TCP);
1841 if (IS_ERR(cb->cm_id)) {
1842 ret = PTR_ERR(cb->cm_id);
1843 log(LOG_ERR, "rdma_create_id error %d\n", ret);
1846 DEBUG_LOG(PFX "created cm_id %p\n", cb->cm_id);
1848 krping_run_server(cb);
1850 krping_run_client(cb);
1851 DEBUG_LOG(PFX "destroy cm_id %p\n", cb->cm_id);
1852 rdma_destroy_id(cb->cm_id);
1854 mtx_lock(&krping_mutex);
1855 TAILQ_REMOVE(&krping_cbs, cb, list);
1856 mtx_unlock(&krping_mutex);
1861 void krping_init(void)
1863 mtx_init(&krping_mutex, "krping lock", NULL, MTX_DEF);
1864 TAILQ_INIT(&krping_cbs);