PERFORCE change 136780 for review
Steve Wise
swise at FreeBSD.org
Tue Mar 4 03:37:47 UTC 2008
http://perforce.freebsd.org/chv.cgi?CH=136780
Change 136780 by swise at swise:vic10:iwarp on 2008/03/04 03:37:25
Added wlat, rlat, and bw tests to krping.
Affected files ...
.. //depot/projects/iwarp/sys/contrib/rdma/krping/krping.c#11 edit
.. //depot/projects/iwarp/sys/contrib/rdma/krping/krping.h#5 edit
Differences ...
==== //depot/projects/iwarp/sys/contrib/rdma/krping/krping.c#11 (text+ko) ====
@@ -76,6 +76,10 @@
{"client", OPT_NOPARAM, 'c'},
{"dmamr", OPT_NOPARAM, 'D'},
{"debug", OPT_NOPARAM, 'd'},
+ {"wlat", OPT_NOPARAM, 'l'},
+ {"rlat", OPT_NOPARAM, 'L'},
+ {"bw", OPT_NOPARAM, 'B'},
+ {"tx-depth", OPT_INT, 't'},
{NULL, 0, 0}
};
@@ -102,8 +106,8 @@
/*
* Default max buffer size for IO...
*/
-#define RPING_BUFSIZE 64*1024
-#define RPING_SQ_DEPTH 16
+#define RPING_BUFSIZE 128*1024
+#define RPING_SQ_DEPTH 32
/* lifted from netinet/libalias/alias_proxy.c */
@@ -346,7 +350,8 @@
mtx_unlock_spin(&cb->lock);
return;
}
- ib_req_notify_cq(cb->cq, IB_CQ_NEXT_COMP);
+ if (!cb->wlat && !cb->rlat && !cb->bw)
+ ib_req_notify_cq(cb->cq, IB_CQ_NEXT_COMP);
while ((ret = ib_poll_cq(cb->cq, 1, &wc)) == 1) {
if (wc.status) {
DEBUG_LOG(PFX "cq completion failed status %d\n",
@@ -381,7 +386,10 @@
DEBUG_LOG(PFX "recv completion\n");
cb->stats.recv_bytes += sizeof(cb->recv_buf);
cb->stats.recv_msgs++;
- ret = cb->server ? server_recv(cb, &wc) :
+ if (cb->wlat || cb->rlat || cb->bw)
+ ret = server_recv(cb, &wc);
+ else
+ ret = cb->server ? server_recv(cb, &wc) :
client_recv(cb, &wc);
if (ret) {
DEBUG_LOG(PFX "recv wc error: %d\n", ret);
@@ -431,10 +439,12 @@
return ret;
}
- krping_wait(cb, CONNECTED);
- if (cb->state == ERROR) {
- DEBUG_LOG(PFX "wait for CONNECTED state %d\n", cb->state);
- return -1;
+ if (!cb->wlat && !cb->rlat && !cb->bw) {
+ krping_wait(cb, CONNECTED);
+ if (cb->state == ERROR) {
+ DEBUG_LOG(PFX "wait for CONNECTED state %d\n", cb->state);
+ return -1;
+ }
}
return 0;
}
@@ -473,7 +483,7 @@
cb->rdma_sq_wr.sg_list = &cb->rdma_sgl;
cb->rdma_sq_wr.num_sge = 1;
- if (!cb->server) {
+ if (!cb->server || cb->wlat || cb->rlat || cb->bw) {
cb->start_addr = virt_to_phys(cb->start_buf);
}
}
@@ -546,7 +556,7 @@
}
}
- if (!cb->server) {
+ if (!cb->server || cb->wlat || cb->rlat || cb->bw) {
cb->start_buf = contigmalloc(cb->size, M_DEVBUF, M_WAITOK,
0, (1L<<22), 4096, 1024*1024);
if (!cb->start_buf) {
@@ -555,12 +565,15 @@
goto err2;
}
if (!cb->use_dmamr) {
+ unsigned flags = IB_ACCESS_REMOTE_READ;
+ if (cb->wlat || cb->rlat || cb->bw)
+ flags |= IB_ACCESS_REMOTE_WRITE;
buf.addr = virt_to_phys(cb->start_buf);
buf.size = cb->size;
iovbase = virt_to_phys(cb->start_buf);
cb->start_mr = ib_reg_phys_mr(cb->pd, &buf, 1,
- IB_ACCESS_REMOTE_READ,
+ flags,
&iovbase);
if (IS_ERR(cb->start_mr)) {
@@ -607,7 +620,7 @@
cb->size, DMA_BIDIRECTIONAL);
#endif
contigfree(cb->rdma_buf, cb->size, M_DEVBUF);
- if (!cb->server) {
+ if (!cb->server || cb->wlat || cb->rlat || cb->bw) {
dma_unmap_single(cb->pd->device->dma_device,
pci_unmap_addr(cb, start_mapping),
cb->size, DMA_BIDIRECTIONAL);
@@ -630,7 +643,7 @@
int ret;
memset(&init_attr, 0, sizeof(init_attr));
- init_attr.cap.max_send_wr = RPING_SQ_DEPTH;
+ init_attr.cap.max_send_wr = cb->txdepth;
init_attr.cap.max_recv_wr = 2;
init_attr.cap.max_recv_sge = 1;
init_attr.cap.max_send_sge = 1;
@@ -669,7 +682,7 @@
DEBUG_LOG(PFX "created pd %p\n", cb->pd);
cb->cq = ib_create_cq(cm_id->device, krping_cq_event_handler, NULL,
- cb, RPING_SQ_DEPTH * 2, 0);
+ cb, cb->txdepth * 2, 0);
if (IS_ERR(cb->cq)) {
DEBUG_LOG(PFX "ib_create_cq failed\n");
ret = PTR_ERR(cb->cq);
@@ -677,10 +690,12 @@
}
DEBUG_LOG(PFX "created cq %p\n", cb->cq);
- ret = ib_req_notify_cq(cb->cq, IB_CQ_NEXT_COMP);
- if (ret) {
- DEBUG_LOG(PFX "ib_create_cq failed\n");
- goto err2;
+ if (!cb->wlat && !cb->rlat && !cb->bw) {
+ ret = ib_req_notify_cq(cb->cq, IB_CQ_NEXT_COMP);
+ if (ret) {
+ DEBUG_LOG(PFX "ib_create_cq failed\n");
+ goto err2;
+ }
}
ret = krping_create_qp(cb);
@@ -809,6 +824,451 @@
}
}
+static void rlat_test(struct krping_cb *cb)
+{
+ int scnt;
+ int iters = cb->count;
+ struct timeval start_tv, stop_tv;
+ int ret;
+ struct ib_wc wc;
+ struct ib_send_wr *bad_wr;
+ int ne;
+
+ scnt = 0;
+ cb->rdma_sq_wr.opcode = IB_WR_RDMA_READ;
+ cb->rdma_sq_wr.wr.rdma.rkey = cb->remote_rkey;
+ cb->rdma_sq_wr.wr.rdma.remote_addr = cb->remote_addr;
+ cb->rdma_sq_wr.sg_list->length = cb->size;
+
+ microtime(&start_tv);
+ while (scnt < iters) {
+
+ ret = ib_post_send(cb->qp, &cb->rdma_sq_wr, &bad_wr);
+ if (ret) {
+ printf(PFX
+ "Couldn't post send: ret=%d scnt %d\n",
+ ret, scnt);
+ return;
+ }
+
+ do {
+ ne = ib_poll_cq(cb->cq, 1, &wc);
+ if (cb->state == ERROR) {
+ printf(PFX
+ "state == ERROR...bailing scnt %d\n", scnt);
+ return;
+ }
+ } while (ne == 0);
+
+ if (ne < 0) {
+ printf(PFX "poll CQ failed %d\n", ne);
+ return;
+ }
+ if (wc.status != IB_WC_SUCCESS) {
+ printf(PFX "Completion wth error at %s:\n",
+ cb->server ? "server" : "client");
+ printf(PFX "Failed status %d: wr_id %d\n",
+ wc.status, (int) wc.wr_id);
+ return;
+ }
+ ++scnt;
+ }
+ microtime(&stop_tv);
+
+ if (stop_tv.tv_usec < start_tv.tv_usec) {
+ stop_tv.tv_usec += 1000000;
+ stop_tv.tv_sec -= 1;
+ }
+
+ printf(PFX "delta sec %lu delta usec %lu iter %d size %d\n",
+ stop_tv.tv_sec - start_tv.tv_sec,
+ stop_tv.tv_usec - start_tv.tv_usec,
+ scnt, cb->size);
+}
+
+static int alloc_cycle_mem(int cycle_iters,
+ cycles_t **post_cycles_start,
+ cycles_t **post_cycles_stop,
+ cycles_t **poll_cycles_start,
+ cycles_t **poll_cycles_stop,
+ cycles_t **last_poll_cycles_start)
+{
+ *post_cycles_start = malloc(cycle_iters * sizeof(cycles_t), M_DEVBUF, M_WAITOK);
+ if (!*post_cycles_start) {
+ goto fail1;
+ }
+ *post_cycles_stop = malloc(cycle_iters * sizeof(cycles_t), M_DEVBUF, M_WAITOK);
+ if (!*post_cycles_stop) {
+ goto fail2;
+ }
+ *poll_cycles_start = malloc(cycle_iters * sizeof(cycles_t), M_DEVBUF, M_WAITOK);
+ if (!*poll_cycles_start) {
+ goto fail3;
+ }
+ *poll_cycles_stop = malloc(cycle_iters * sizeof(cycles_t), M_DEVBUF, M_WAITOK);
+ if (!*poll_cycles_stop) {
+ goto fail4;
+ }
+ *last_poll_cycles_start = malloc(cycle_iters * sizeof(cycles_t), M_DEVBUF, M_WAITOK);
+ if (!*last_poll_cycles_start) {
+ goto fail5;
+ }
+ return 0;
+fail5:
+ free(*poll_cycles_stop, M_DEVBUF);
+fail4:
+ free(*poll_cycles_start, M_DEVBUF);
+fail3:
+ free(*post_cycles_stop, M_DEVBUF);
+fail2:
+ free(*post_cycles_start, M_DEVBUF);
+fail1:
+ printf(PFX "%s malloc failed\n", __FUNCTION__);
+ return ENOMEM;
+}
+
+static void free_cycle_mem(cycles_t *post_cycles_start,
+ cycles_t *post_cycles_stop,
+ cycles_t *poll_cycles_start,
+ cycles_t *poll_cycles_stop,
+ cycles_t *last_poll_cycles_start)
+{
+ free(last_poll_cycles_start, M_DEVBUF);
+ free(poll_cycles_stop, M_DEVBUF);
+ free(poll_cycles_start, M_DEVBUF);
+ free(post_cycles_stop, M_DEVBUF);
+ free(post_cycles_start, M_DEVBUF);
+}
+
+static void wlat_test(struct krping_cb *cb)
+{
+ int ccnt, scnt, rcnt;
+ int iters=cb->count;
+ volatile char *poll_buf = (char *) cb->start_buf;
+ char *buf = (char *)cb->rdma_buf;
+ ccnt = 0;
+ scnt = 0;
+ rcnt = 0;
+ struct timeval start_tv, stop_tv;
+ cycles_t *post_cycles_start, *post_cycles_stop;
+ cycles_t *poll_cycles_start, *poll_cycles_stop;
+ cycles_t *last_poll_cycles_start;
+ cycles_t sum_poll = 0, sum_post = 0, sum_last_poll = 0;
+ int i;
+ int cycle_iters = 1000;
+ int err;
+
+ err = alloc_cycle_mem(cycle_iters, &post_cycles_start, &post_cycles_stop,
+ &poll_cycles_start, &poll_cycles_stop, &last_poll_cycles_start);
+
+ if (err) {
+ printf(PFX "%s malloc failed\n", __FUNCTION__);
+ return;
+ }
+
+ cb->rdma_sq_wr.opcode = IB_WR_RDMA_WRITE;
+ cb->rdma_sq_wr.wr.rdma.rkey = cb->remote_rkey;
+ cb->rdma_sq_wr.wr.rdma.remote_addr = cb->remote_addr;
+ cb->rdma_sq_wr.sg_list->length = cb->size;
+
+ if (cycle_iters > iters)
+ cycle_iters = iters;
+ microtime(&start_tv);
+ while (scnt < iters || ccnt < iters || rcnt < iters) {
+
+ /* Wait till buffer changes. */
+ if (rcnt < iters && !(scnt < 1 && !cb->server)) {
+ ++rcnt;
+ while (*poll_buf != (char)rcnt) {
+ if (cb->state == ERROR) {
+ printf(PFX "state = ERROR, bailing\n");
+ return;
+ }
+ }
+ }
+
+ if (scnt < iters) {
+ struct ib_send_wr *bad_wr;
+
+ *buf = (char)scnt+1;
+ if (scnt < cycle_iters)
+ post_cycles_start[scnt] = get_cycles();
+ if (ib_post_send(cb->qp, &cb->rdma_sq_wr, &bad_wr)) {
+ printf(PFX "Couldn't post send: scnt=%d\n",
+ scnt);
+ return;
+ }
+ if (scnt < cycle_iters)
+ post_cycles_stop[scnt] = get_cycles();
+ scnt++;
+ }
+
+ if (ccnt < iters) {
+ struct ib_wc wc;
+ int ne;
+
+ if (ccnt < cycle_iters)
+ poll_cycles_start[ccnt] = get_cycles();
+ do {
+ if (ccnt < cycle_iters)
+ last_poll_cycles_start[ccnt] = get_cycles();
+ ne = ib_poll_cq(cb->cq, 1, &wc);
+ } while (ne == 0);
+ if (ccnt < cycle_iters)
+ poll_cycles_stop[ccnt] = get_cycles();
+ ++ccnt;
+
+ if (ne < 0) {
+ printf(PFX "poll CQ failed %d\n", ne);
+ return;
+ }
+ if (wc.status != IB_WC_SUCCESS) {
+ printf(PFX "Completion wth error at %s:\n",
+ cb->server ? "server" : "client");
+ printf(PFX "Failed status %d: wr_id %d\n",
+ wc.status, (int) wc.wr_id);
+ printf(PFX "scnt=%d, rcnt=%d, ccnt=%d\n",
+ scnt, rcnt, ccnt);
+ return;
+ }
+ }
+ }
+ microtime(&stop_tv);
+
+ if (stop_tv.tv_usec < start_tv.tv_usec) {
+ stop_tv.tv_usec += 1000000;
+ stop_tv.tv_sec -= 1;
+ }
+
+ for (i=0; i < cycle_iters; i++) {
+ sum_post += post_cycles_stop[i] - post_cycles_start[i];
+ sum_poll += poll_cycles_stop[i] - poll_cycles_start[i];
+ sum_last_poll += poll_cycles_stop[i] - last_poll_cycles_start[i];
+ }
+
+ printf(PFX "delta sec %lu delta usec %lu iter %d size %d cycle_iters %d sum_post %llu sum_poll %llu sum_last_poll %llu\n",
+ stop_tv.tv_sec - start_tv.tv_sec,
+ stop_tv.tv_usec - start_tv.tv_usec,
+ scnt, cb->size, cycle_iters,
+ (unsigned long long)sum_post, (unsigned long long)sum_poll,
+ (unsigned long long)sum_last_poll);
+
+ free_cycle_mem(post_cycles_start, post_cycles_stop, poll_cycles_start,
+ poll_cycles_stop, last_poll_cycles_start);
+}
+
+static void bw_test(struct krping_cb *cb)
+{
+ int ccnt, scnt, rcnt;
+ int iters=cb->count;
+ ccnt = 0;
+ scnt = 0;
+ rcnt = 0;
+ struct timeval start_tv, stop_tv;
+ cycles_t *post_cycles_start, *post_cycles_stop;
+ cycles_t *poll_cycles_start, *poll_cycles_stop;
+ cycles_t *last_poll_cycles_start;
+ cycles_t sum_poll = 0, sum_post = 0, sum_last_poll = 0;
+ int i;
+ int cycle_iters = 1000;
+ int err;
+
+ err = alloc_cycle_mem(cycle_iters, &post_cycles_start, &post_cycles_stop,
+ &poll_cycles_start, &poll_cycles_stop, &last_poll_cycles_start);
+
+ if (err) {
+ printf(PFX "%s kmalloc failed\n", __FUNCTION__);
+ return;
+ }
+
+ cb->rdma_sq_wr.opcode = IB_WR_RDMA_WRITE;
+ cb->rdma_sq_wr.wr.rdma.rkey = cb->remote_rkey;
+ cb->rdma_sq_wr.wr.rdma.remote_addr = cb->remote_addr;
+ cb->rdma_sq_wr.sg_list->length = cb->size;
+
+ if (cycle_iters > iters)
+ cycle_iters = iters;
+ microtime(&start_tv);
+ while (scnt < iters || ccnt < iters) {
+
+ while (scnt < iters && scnt - ccnt < cb->txdepth) {
+ struct ib_send_wr *bad_wr;
+
+ if (scnt < cycle_iters)
+ post_cycles_start[scnt] = get_cycles();
+ if (ib_post_send(cb->qp, &cb->rdma_sq_wr, &bad_wr)) {
+ printf(PFX "Couldn't post send: scnt=%d\n",
+ scnt);
+ return;
+ }
+ if (scnt < cycle_iters)
+ post_cycles_stop[scnt] = get_cycles();
+ ++scnt;
+ }
+
+ if (ccnt < iters) {
+ int ne;
+ struct ib_wc wc;
+
+ if (ccnt < cycle_iters)
+ poll_cycles_start[ccnt] = get_cycles();
+ do {
+ if (ccnt < cycle_iters)
+ last_poll_cycles_start[ccnt] = get_cycles();
+ ne = ib_poll_cq(cb->cq, 1, &wc);
+ } while (ne == 0);
+ if (ccnt < cycle_iters)
+ poll_cycles_stop[ccnt] = get_cycles();
+ ccnt += 1;
+
+ if (ne < 0) {
+ printf(PFX "poll CQ failed %d\n", ne);
+ return;
+ }
+ if (wc.status != IB_WC_SUCCESS) {
+ printf(PFX "Completion wth error at %s:\n",
+ cb->server ? "server" : "client");
+ printf(PFX "Failed status %d: wr_id %d\n",
+ wc.status, (int) wc.wr_id);
+ return;
+ }
+ }
+ }
+ microtime(&stop_tv);
+
+ if (stop_tv.tv_usec < start_tv.tv_usec) {
+ stop_tv.tv_usec += 1000000;
+ stop_tv.tv_sec -= 1;
+ }
+
+ for (i=0; i < cycle_iters; i++) {
+ sum_post += post_cycles_stop[i] - post_cycles_start[i];
+ sum_poll += poll_cycles_stop[i] - poll_cycles_start[i];
+ sum_last_poll += poll_cycles_stop[i] - last_poll_cycles_start[i];
+ }
+
+ printf(PFX "delta sec %lu delta usec %lu iter %d size %d cycle_iters %d sum_post %llu sum_poll %llu sum_last_poll %llu\n",
+ stop_tv.tv_sec - start_tv.tv_sec,
+ stop_tv.tv_usec - start_tv.tv_usec,
+ scnt, cb->size, cycle_iters,
+ (unsigned long long)sum_post, (unsigned long long)sum_poll,
+ (unsigned long long)sum_last_poll);
+
+ free_cycle_mem(post_cycles_start, post_cycles_stop, poll_cycles_start,
+ poll_cycles_stop, last_poll_cycles_start);
+}
+
+static void krping_rlat_test_server(struct krping_cb *cb)
+{
+ struct ib_send_wr *bad_wr;
+ struct ib_wc wc;
+ int ret;
+
+ /* Spin waiting for client's Start STAG/TO/Len */
+ while (cb->state < RDMA_READ_ADV) {
+ krping_cq_event_handler(cb->cq, cb);
+ }
+
+ /* Send STAG/TO/Len to client */
+ if (cb->dma_mr)
+ krping_format_send(cb, cb->start_addr, cb->dma_mr);
+ else
+ krping_format_send(cb, cb->start_addr, cb->start_mr);
+ ret = ib_post_send(cb->qp, &cb->sq_wr, &bad_wr);
+ if (ret) {
+ printf(PFX "post send error %d\n", ret);
+ return;
+ }
+
+ /* Spin waiting for send completion */
+ while ((ret = ib_poll_cq(cb->cq, 1, &wc) == 0));
+ if (ret < 0) {
+ printf(PFX "poll error %d\n", ret);
+ return;
+ }
+ if (wc.status) {
+ printf(PFX "send completiong error %d\n", wc.status);
+ return;
+ }
+
+ krping_wait(cb, ERROR);
+}
+
+static void krping_wlat_test_server(struct krping_cb *cb)
+{
+ struct ib_send_wr *bad_wr;
+ struct ib_wc wc;
+ int ret;
+
+ /* Spin waiting for client's Start STAG/TO/Len */
+ while (cb->state < RDMA_READ_ADV) {
+ krping_cq_event_handler(cb->cq, cb);
+ }
+
+ /* Send STAG/TO/Len to client */
+ if (cb->dma_mr)
+ krping_format_send(cb, cb->start_addr, cb->dma_mr);
+ else
+ krping_format_send(cb, cb->start_addr, cb->start_mr);
+ ret = ib_post_send(cb->qp, &cb->sq_wr, &bad_wr);
+ if (ret) {
+ printf(PFX "post send error %d\n", ret);
+ return;
+ }
+
+ /* Spin waiting for send completion */
+ while ((ret = ib_poll_cq(cb->cq, 1, &wc) == 0));
+ if (ret < 0) {
+ printf(PFX "poll error %d\n", ret);
+ return;
+ }
+ if (wc.status) {
+ printf(PFX "send completiong error %d\n", wc.status);
+ return;
+ }
+
+ wlat_test(cb);
+
+}
+
+static void krping_bw_test_server(struct krping_cb *cb)
+{
+ struct ib_send_wr *bad_wr;
+ struct ib_wc wc;
+ int ret;
+
+ /* Spin waiting for client's Start STAG/TO/Len */
+ while (cb->state < RDMA_READ_ADV) {
+ krping_cq_event_handler(cb->cq, cb);
+ }
+
+ /* Send STAG/TO/Len to client */
+ if (cb->dma_mr)
+ krping_format_send(cb, cb->start_addr, cb->dma_mr);
+ else
+ krping_format_send(cb, cb->start_addr, cb->start_mr);
+ ret = ib_post_send(cb->qp, &cb->sq_wr, &bad_wr);
+ if (ret) {
+ printf(PFX "post send error %d\n", ret);
+ return;
+ }
+
+ /* Spin waiting for send completion */
+ while ((ret = ib_poll_cq(cb->cq, 1, &wc) == 0));
+ if (ret < 0) {
+ printf(PFX "poll error %d\n", ret);
+ return;
+ }
+ if (wc.status) {
+ printf(PFX "send completiong error %d\n", wc.status);
+ return;
+ }
+
+ if (cb->duplex)
+ bw_test(cb);
+ krping_wait(cb, ERROR);
+}
+
static int krping_bind_server(struct krping_cb *cb)
{
struct sockaddr_in sin;
@@ -877,7 +1337,15 @@
goto err2;
}
- krping_test_server(cb);
+ if (cb->wlat)
+ krping_wlat_test_server(cb);
+ else if (cb->rlat)
+ krping_rlat_test_server(cb);
+ else if (cb->bw)
+ krping_bw_test_server(cb);
+ else
+ krping_test_server(cb);
+
rdma_disconnect(cb->child_cm_id);
rdma_destroy_id(cb->child_cm_id);
err2:
@@ -960,6 +1428,171 @@
}
}
+static void krping_rlat_test_client(struct krping_cb *cb)
+{
+ struct ib_send_wr *bad_wr;
+ struct ib_wc wc;
+ int ret;
+
+ cb->state = RDMA_READ_ADV;
+
+ /* Send STAG/TO/Len to client */
+ if (cb->dma_mr)
+ krping_format_send(cb, cb->start_addr, cb->dma_mr);
+ else
+ krping_format_send(cb, cb->start_addr, cb->rdma_mr);
+ ret = ib_post_send(cb->qp, &cb->sq_wr, &bad_wr);
+ if (ret) {
+ printf(PFX "post send error %d\n", ret);
+ return;
+ }
+
+ /* Spin waiting for send completion */
+ while ((ret = ib_poll_cq(cb->cq, 1, &wc) == 0));
+ if (ret < 0) {
+ printf(PFX "poll error %d\n", ret);
+ return;
+ }
+ if (wc.status) {
+ printf(PFX "send completion error %d\n", wc.status);
+ return;
+ }
+
+ /* Spin waiting for server's Start STAG/TO/Len */
+ while (cb->state < RDMA_WRITE_ADV) {
+ krping_cq_event_handler(cb->cq, cb);
+ }
+
+#if 0
+{
+ int i;
+ struct timeval start, stop;
+ time_t sec;
+ suseconds_t usec;
+ unsigned long long elapsed;
+ struct ib_wc wc;
+ struct ib_send_wr *bad_wr;
+ int ne;
+
+ cb->rdma_sq_wr.opcode = IB_WR_RDMA_WRITE;
+ cb->rdma_sq_wr.wr.rdma.rkey = cb->remote_rkey;
+ cb->rdma_sq_wr.wr.rdma.remote_addr = cb->remote_addr;
+ cb->rdma_sq_wr.sg_list->length = 0;
+ cb->rdma_sq_wr.num_sge = 0;
+
+ microtime(&start);
+ for (i=0; i < 100000; i++) {
+ if (ib_post_send(cb->qp, &cb->rdma_sq_wr, &bad_wr)) {
+ printf(PFX "Couldn't post send\n");
+ return;
+ }
+ do {
+ ne = ib_poll_cq(cb->cq, 1, &wc);
+ } while (ne == 0);
+ if (ne < 0) {
+ printf(PFX "poll CQ failed %d\n", ne);
+ return;
+ }
+ if (wc.status != IB_WC_SUCCESS) {
+ printf(PFX "Completion wth error at %s:\n",
+ cb->server ? "server" : "client");
+ printf(PFX "Failed status %d: wr_id %d\n",
+ wc.status, (int) wc.wr_id);
+ return;
+ }
+ }
+ microtime(&stop);
+
+ if (stop.tv_usec < start.tv_usec) {
+ stop.tv_usec += 1000000;
+ stop.tv_sec -= 1;
+ }
+ sec = stop.tv_sec - start.tv_sec;
+ usec = stop.tv_usec - start.tv_usec;
+ elapsed = sec * 1000000 + usec;
+ printf(PFX "0B-write-lat iters 100000 usec %llu\n", elapsed);
+}
+#endif
+
+ rlat_test(cb);
+}
+
+static void krping_wlat_test_client(struct krping_cb *cb)
+{
+ struct ib_send_wr *bad_wr;
+ struct ib_wc wc;
+ int ret;
+
+ cb->state = RDMA_READ_ADV;
+
+ /* Send STAG/TO/Len to client */
+ if (cb->dma_mr)
+ krping_format_send(cb, cb->start_addr, cb->dma_mr);
+ else
+ krping_format_send(cb, cb->start_addr, cb->start_mr);
+ ret = ib_post_send(cb->qp, &cb->sq_wr, &bad_wr);
+ if (ret) {
+ printf(PFX "post send error %d\n", ret);
+ return;
+ }
+
+ /* Spin waiting for send completion */
+ while ((ret = ib_poll_cq(cb->cq, 1, &wc) == 0));
+ if (ret < 0) {
+ printf(PFX "poll error %d\n", ret);
+ return;
+ }
+ if (wc.status) {
+ printf(PFX "send completion error %d\n", wc.status);
+ return;
+ }
+
+ /* Spin waiting for server's Start STAG/TO/Len */
+ while (cb->state < RDMA_WRITE_ADV) {
+ krping_cq_event_handler(cb->cq, cb);
+ }
+
+ wlat_test(cb);
+}
+
+static void krping_bw_test_client(struct krping_cb *cb)
+{
+ struct ib_send_wr *bad_wr;
+ struct ib_wc wc;
+ int ret;
+
+ cb->state = RDMA_READ_ADV;
+
+ /* Send STAG/TO/Len to client */
+ if (cb->dma_mr)
+ krping_format_send(cb, cb->start_addr, cb->dma_mr);
+ else
+ krping_format_send(cb, cb->start_addr, cb->start_mr);
+ ret = ib_post_send(cb->qp, &cb->sq_wr, &bad_wr);
+ if (ret) {
+ printf(PFX "post send error %d\n", ret);
+ return;
+ }
+
+ /* Spin waiting for send completion */
+ while ((ret = ib_poll_cq(cb->cq, 1, &wc) == 0));
+ if (ret < 0) {
+ printf(PFX "poll error %d\n", ret);
+ return;
+ }
+ if (wc.status) {
+ printf(PFX "send completion error %d\n", wc.status);
+ return;
+ }
+
+ /* Spin waiting for server's Start STAG/TO/Len */
+ while (cb->state < RDMA_WRITE_ADV) {
+ krping_cq_event_handler(cb->cq, cb);
+ }
+
+ bw_test(cb);
+}
+
static int krping_connect_client(struct krping_cb *cb)
{
struct rdma_conn_param conn_param;
@@ -1049,7 +1682,14 @@
goto err2;
}
- krping_test_client(cb);
+ if (cb->wlat)
+ krping_wlat_test_client(cb);
+ else if (cb->rlat)
+ krping_rlat_test_client(cb);
+ else if (cb->bw)
+ krping_bw_test_client(cb);
+ else
+ krping_test_client(cb);
rdma_disconnect(cb->cm_id);
err2:
krping_free_buffers(cb);
@@ -1064,6 +1704,7 @@
int ret = 0;
char *optarg;
unsigned long optint;
+ debug = 0;
cb = malloc(sizeof(*cb), M_DEVBUF, M_WAITOK);
if (!cb)
@@ -1077,6 +1718,7 @@
cb->server = -1;
cb->state = IDLE;
cb->size = 64;
+ cb->txdepth = RPING_SQ_DEPTH;
mtx_init(&cb->lock, "krping mtx", NULL, MTX_DUPOK|MTX_SPIN);
while ((op = krping_getopt("krping", &cmd, krping_opts, NULL, &optarg,
@@ -1134,6 +1776,19 @@
cb->validate++;
DEBUG_LOG(PFX "validate data\n");
break;
+ case 'L':
+ cb->rlat++;
+ break;
+ case 'l':
+ cb->wlat++;
+ break;
+ case 'B':
+ cb->bw++;
+ break;
+ case 't':
+ cb->txdepth = optint;
+ DEBUG_LOG(PFX "txdepth %d\n", cb->txdepth);
+ break;
case 'd':
debug++;
break;
@@ -1151,6 +1806,12 @@
ret = EINVAL;
goto out;
}
+ if ((cb->bw + cb->rlat + cb->wlat) > 1) {
+ printf(PFX "Pick only one test: bw, rlat, wlat\n");
+ ret = -EINVAL;
+ goto out;
+ }
+
cb->cm_id = rdma_create_id(krping_cma_event_handler, cb, RDMA_PS_TCP);
if (IS_ERR(cb->cm_id)) {
==== //depot/projects/iwarp/sys/contrib/rdma/krping/krping.h#5 (text+ko) ====
@@ -101,8 +101,24 @@
/* listener on service side. */
struct rdma_cm_id *child_cm_id; /* connection on server side */
TAILQ_ENTRY(krping_cb) list;
+
+ int rlat; /* run read latency test */
+ int wlat; /* run write latency test */
+ int bw; /* run write bw test */
+ int duplex; /* run write bw full duplex test */
+ int txdepth;
};
+static __inline uint64_t
+get_cycles(void)
+{
+ u_int32_t low, high;
+ __asm __volatile("rdtsc" : "=a" (low), "=d" (high));
+ return (low | ((u_int64_t)high << 32));
+}
+
+typedef uint64_t cycles_t;
+
extern struct mtx krping_mutex;
TAILQ_HEAD(krping_cb_list, krping_cb);
extern struct krping_cb_list krping_cbs;
More information about the p4-projects
mailing list