projects / fio.git / blame
? search:
summary | shortlog | log | commit | commitdiff | tree
blob | blame (incremental) | history | HEAD
implement 'unit_base' option to select between KB and Kbit et. al.
[fio.git] / engines / rdma.c
CommitLineData
21b8aee8 1/*
85286c5c 2 * RDMA I/O engine
21b8aee8 3 *
85286c5c
BVA		 4 * RDMA I/O engine based on the IB verbs and RDMA/CM user space libraries.
5 * Supports both RDMA memory semantics and channel semantics
6 * for the InfiniBand, RoCE and iWARP protocols.
21b8aee8 7 *
0ac5d398 8 * You will need the Linux RDMA software installed, either
85286c5c 9 * from your Linux distributor or directly from openfabrics.org:
21b8aee8 10 *
11 * http://www.openfabrics.org/downloads/OFED/
12 *
7d7803fa
YR		 13 * Exchanging steps of RDMA ioengine control messages:
14 * 1. client side sends test mode (RDMA_WRITE/RDMA_READ/SEND)
15 * to server side.
222757cc 16 * 2. server side parses test mode, and sends back confirmation
7d7803fa 17 * to client side. In RDMA WRITE/READ test, this confirmation
222757cc 18 * includes memory information, such as rkey, address.
7d7803fa 19 * 3. client side initiates test loop.
222757cc 20 * 4. In RDMA WRITE/READ test, client side sends a completion
7d7803fa 21 * notification to server side. Server side updates its
222757cc 22 * td->done as true.
7d7803fa 23 *
21b8aee8 24 */
25#include <stdio.h>
26#include <stdlib.h>
27#include <unistd.h>
28#include <errno.h>
29#include <assert.h>
30#include <netinet/in.h>
31#include <arpa/inet.h>
32#include <netdb.h>
33#include <sys/poll.h>
34#include <sys/types.h>
35#include <sys/socket.h>
36#include <sys/time.h>
37#include <sys/resource.h>
38
21b8aee8 39#include <pthread.h>
40#include <inttypes.h>
41
42#include "../fio.h"
ee88d056 43#include "../hash.h"
21b8aee8 44
21b8aee8 45#include <rdma/rdma_cma.h>
46#include <infiniband/arch.h>
47
7d7803fa 48#define FIO_RDMA_MAX_IO_DEPTH 512
21b8aee8 49
50enum rdma_io_mode {
51 FIO_RDMA_UNKNOWN = 0,
52 FIO_RDMA_MEM_WRITE,
53 FIO_RDMA_MEM_READ,
54 FIO_RDMA_CHA_SEND,
55 FIO_RDMA_CHA_RECV
56};
57
58struct remote_u {
59 uint64_t buf;
60 uint32_t rkey;
61 uint32_t size;
62};
63
64struct rdma_info_blk {
65 uint32_t mode; /* channel semantic or memory semantic */
66 uint32_t nr; /* client: io depth
67 server: number of records for memory semantic
68 */
76cc5224 69 struct remote_u rmt_us[FIO_RDMA_MAX_IO_DEPTH];
21b8aee8 70};
71
72struct rdma_io_u_data {
73 uint64_t wr_id;
74 struct ibv_send_wr sq_wr;
75 struct ibv_recv_wr rq_wr;
76 struct ibv_sge rdma_sgl;
77};
78
79struct rdmaio_data {
80 int is_client;
81 enum rdma_io_mode rdma_protocol;
82 char host[64];
83 struct sockaddr_in addr;
84
85 struct ibv_recv_wr rq_wr;
86 struct ibv_sge recv_sgl;
87 struct rdma_info_blk recv_buf;
88 struct ibv_mr *recv_mr;
89
90 struct ibv_send_wr sq_wr;
91 struct ibv_sge send_sgl;
92 struct rdma_info_blk send_buf;
93 struct ibv_mr *send_mr;
94
95 struct ibv_comp_channel *channel;
96 struct ibv_cq *cq;
97 struct ibv_pd *pd;
98 struct ibv_qp *qp;
99
100 pthread_t cmthread;
101 struct rdma_event_channel *cm_channel;
102 struct rdma_cm_id *cm_id;
103 struct rdma_cm_id *child_cm_id;
104
105 int cq_event_num;
106
107 struct remote_u *rmt_us;
108 int rmt_nr;
109 struct io_u **io_us_queued;
110 int io_u_queued_nr;
111 struct io_u **io_us_flight;
112 int io_u_flight_nr;
113 struct io_u **io_us_completed;
114 int io_u_completed_nr;
ea6209ad
JA		 115
116 struct frand_state rand_state;
21b8aee8 117};
118
119static int client_recv(struct thread_data *td, struct ibv_wc *wc)
120{
121 struct rdmaio_data *rd = td->io_ops->data;
122
123 if (wc->byte_len != sizeof(rd->recv_buf)) {
b6cf38f0 124 log_err("Received bogus data, size %d\n", wc->byte_len);
21b8aee8 125 return 1;
126 }
127
128 /* store mr info for MEMORY semantic */
129 if ((rd->rdma_protocol == FIO_RDMA_MEM_WRITE) ||
130 (rd->rdma_protocol == FIO_RDMA_MEM_READ)) {
131 /* struct flist_head *entry; */
132 int i = 0;
133
134 rd->rmt_nr = ntohl(rd->recv_buf.nr);
135
136 for (i = 0; i < rd->rmt_nr; i++) {
137 rd->rmt_us[i].buf = ntohll(rd->recv_buf.rmt_us[i].buf);
138 rd->rmt_us[i].rkey = ntohl(rd->recv_buf.rmt_us[i].rkey);
139 rd->rmt_us[i].size = ntohl(rd->recv_buf.rmt_us[i].size);
140
141 dprint(FD_IO,
142 "fio: Received rkey %x addr %" PRIx64
143 " len %d from peer\n", rd->rmt_us[i].rkey,
144 rd->rmt_us[i].buf, rd->rmt_us[i].size);
145 }
146 }
147
148 return 0;
149}
150
151static int server_recv(struct thread_data *td, struct ibv_wc *wc)
152{
153 struct rdmaio_data *rd = td->io_ops->data;
154
76cc5224 155 if (wc->wr_id == FIO_RDMA_MAX_IO_DEPTH) {
21b8aee8 156 rd->rdma_protocol = ntohl(rd->recv_buf.mode);
157
158 /* CHANNEL semantic, do nothing */
159 if (rd->rdma_protocol == FIO_RDMA_CHA_SEND)
160 rd->rdma_protocol = FIO_RDMA_CHA_RECV;
161 }
162
163 return 0;
164}
165
166static int cq_event_handler(struct thread_data *td, enum ibv_wc_opcode opcode)
167{
168 struct rdmaio_data *rd = td->io_ops->data;
169 struct ibv_wc wc;
170 struct rdma_io_u_data *r_io_u_d;
171 int ret;
172 int compevnum = 0;
173 int i;
174
175 while ((ret = ibv_poll_cq(rd->cq, 1, &wc)) == 1) {
176 ret = 0;
177 compevnum++;
178
179 if (wc.status) {
180 log_err("fio: cq completion status %d(%s)\n",
181 wc.status, ibv_wc_status_str(wc.status));
182 return -1;
183 }
184
185 switch (wc.opcode) {
186
187 case IBV_WC_RECV:
188 if (rd->is_client == 1)
189 client_recv(td, &wc);
190 else
191 server_recv(td, &wc);
192
76cc5224 193 if (wc.wr_id == FIO_RDMA_MAX_IO_DEPTH)
21b8aee8 194 break;
195
196 for (i = 0; i < rd->io_u_flight_nr; i++) {
197 r_io_u_d = rd->io_us_flight[i]->engine_data;
198
199 if (wc.wr_id == r_io_u_d->rq_wr.wr_id) {
200 rd->io_us_flight[i]->resid =
201 rd->io_us_flight[i]->buflen
202 - wc.byte_len;
203
204 rd->io_us_flight[i]->error = 0;
205
206 rd->io_us_completed[rd->
207 io_u_completed_nr]
208 = rd->io_us_flight[i];
209 rd->io_u_completed_nr++;
210 break;
211 }
212 }
213 if (i == rd->io_u_flight_nr)
e07f72d3 214 log_err("fio: recv wr %" PRId64 " not found\n",
21b8aee8 215 wc.wr_id);
216 else {
217 /* put the last one into middle of the list */
218 rd->io_us_flight[i] =
219 rd->io_us_flight[rd->io_u_flight_nr - 1];
220 rd->io_u_flight_nr--;
221 }
222
223 break;
224
225 case IBV_WC_SEND:
226 case IBV_WC_RDMA_WRITE:
227 case IBV_WC_RDMA_READ:
76cc5224 228 if (wc.wr_id == FIO_RDMA_MAX_IO_DEPTH)
21b8aee8 229 break;
230
231 for (i = 0; i < rd->io_u_flight_nr; i++) {
232 r_io_u_d = rd->io_us_flight[i]->engine_data;
233
234 if (wc.wr_id == r_io_u_d->sq_wr.wr_id) {
235 rd->io_us_completed[rd->
236 io_u_completed_nr]
237 = rd->io_us_flight[i];
238 rd->io_u_completed_nr++;
239 break;
240 }
241 }
242 if (i == rd->io_u_flight_nr)
e07f72d3 243 log_err("fio: send wr %" PRId64 " not found\n",
21b8aee8 244 wc.wr_id);
245 else {
246 /* put the last one into middle of the list */
247 rd->io_us_flight[i] =
248 rd->io_us_flight[rd->io_u_flight_nr - 1];
249 rd->io_u_flight_nr--;
250 }
251
252 break;
253
254 default:
255 log_info("fio: unknown completion event %d\n",
256 wc.opcode);
257 return -1;
258 }
259 rd->cq_event_num++;
260 }
261 if (ret) {
262 log_err("fio: poll error %d\n", ret);
263 return 1;
264 }
265
266 return compevnum;
267}
268
269/*
270 * Return -1 for error and 'nr events' for a positive number
271 * of events
272 */
273static int rdma_poll_wait(struct thread_data *td, enum ibv_wc_opcode opcode)
274{
275 struct rdmaio_data *rd = td->io_ops->data;
276 struct ibv_cq *ev_cq;
277 void *ev_ctx;
278 int ret;
279
280 if (rd->cq_event_num > 0) { /* previous left */
281 rd->cq_event_num--;
282 return 0;
283 }
284
285again:
286 if (ibv_get_cq_event(rd->channel, &ev_cq, &ev_ctx) != 0) {
287 log_err("fio: Failed to get cq event!\n");
288 return -1;
289 }
290 if (ev_cq != rd->cq) {
291 log_err("fio: Unknown CQ!\n");
292 return -1;
293 }
294 if (ibv_req_notify_cq(rd->cq, 0) != 0) {
295 log_err("fio: Failed to set notify!\n");
296 return -1;
297 }
298
299 ret = cq_event_handler(td, opcode);
300 if (ret < 1)
301 goto again;
302
303 ibv_ack_cq_events(rd->cq, ret);
304
305 rd->cq_event_num--;
306
307 return ret;
308}
309
310static int fio_rdmaio_setup_qp(struct thread_data *td)
311{
312 struct rdmaio_data *rd = td->io_ops->data;
313 struct ibv_qp_init_attr init_attr;
314 int qp_depth = td->o.iodepth * 2; /* 2 times of io depth */
315
316 if (rd->is_client == 0)
317 rd->pd = ibv_alloc_pd(rd->child_cm_id->verbs);
318 else
319 rd->pd = ibv_alloc_pd(rd->cm_id->verbs);
222757cc 320
21b8aee8 321 if (rd->pd == NULL) {
322 log_err("fio: ibv_alloc_pd fail\n");
323 return 1;
324 }
325
326 if (rd->is_client == 0)
327 rd->channel = ibv_create_comp_channel(rd->child_cm_id->verbs);
328 else
329 rd->channel = ibv_create_comp_channel(rd->cm_id->verbs);
330 if (rd->channel == NULL) {
331 log_err("fio: ibv_create_comp_channel fail\n");
332 goto err1;
333 }
334
335 if (qp_depth < 16)
336 qp_depth = 16;
337
338 if (rd->is_client == 0)
339 rd->cq = ibv_create_cq(rd->child_cm_id->verbs,
340 qp_depth, rd, rd->channel, 0);
341 else
342 rd->cq = ibv_create_cq(rd->cm_id->verbs,
343 qp_depth, rd, rd->channel, 0);
344 if (rd->cq == NULL) {
345 log_err("fio: ibv_create_cq failed\n");
346 goto err2;
347 }
348
349 if (ibv_req_notify_cq(rd->cq, 0) != 0) {
350 log_err("fio: ibv_create_cq failed\n");
351 goto err3;
352 }
353
354 /* create queue pair */
355 memset(&init_attr, 0, sizeof(init_attr));
356 init_attr.cap.max_send_wr = qp_depth;
357 init_attr.cap.max_recv_wr = qp_depth;
358 init_attr.cap.max_recv_sge = 1;
359 init_attr.cap.max_send_sge = 1;
360 init_attr.qp_type = IBV_QPT_RC;
361 init_attr.send_cq = rd->cq;
362 init_attr.recv_cq = rd->cq;
363
364 if (rd->is_client == 0) {
365 if (rdma_create_qp(rd->child_cm_id, rd->pd, &init_attr) != 0) {
366 log_err("fio: rdma_create_qp failed\n");
367 goto err3;
368 }
369 rd->qp = rd->child_cm_id->qp;
370 } else {
371 if (rdma_create_qp(rd->cm_id, rd->pd, &init_attr) != 0) {
372 log_err("fio: rdma_create_qp failed\n");
373 goto err3;
374 }
375 rd->qp = rd->cm_id->qp;
376 }
377
378 return 0;
379
380err3:
381 ibv_destroy_cq(rd->cq);
382err2:
383 ibv_destroy_comp_channel(rd->channel);
384err1:
385 ibv_dealloc_pd(rd->pd);
386
387 return 1;
388}
389
390static int fio_rdmaio_setup_control_msg_buffers(struct thread_data *td)
391{
392 struct rdmaio_data *rd = td->io_ops->data;
393
394 rd->recv_mr = ibv_reg_mr(rd->pd, &rd->recv_buf, sizeof(rd->recv_buf),
395 IBV_ACCESS_LOCAL_WRITE);
396 if (rd->recv_mr == NULL) {
397 log_err("fio: recv_buf reg_mr failed\n");
398 return 1;
399 }
400
401 rd->send_mr = ibv_reg_mr(rd->pd, &rd->send_buf, sizeof(rd->send_buf),
402 0);
403 if (rd->send_mr == NULL) {
404 log_err("fio: send_buf reg_mr failed\n");
405 ibv_dereg_mr(rd->recv_mr);
406 return 1;
407 }
408
409 /* setup work request */
410 /* recv wq */
411 rd->recv_sgl.addr = (uint64_t) (unsigned long)&rd->recv_buf;
222757cc 412 rd->recv_sgl.length = sizeof(rd->recv_buf);
21b8aee8 413 rd->recv_sgl.lkey = rd->recv_mr->lkey;
414 rd->rq_wr.sg_list = &rd->recv_sgl;
415 rd->rq_wr.num_sge = 1;
76cc5224 416 rd->rq_wr.wr_id = FIO_RDMA_MAX_IO_DEPTH;
21b8aee8 417
418 /* send wq */
419 rd->send_sgl.addr = (uint64_t) (unsigned long)&rd->send_buf;
222757cc 420 rd->send_sgl.length = sizeof(rd->send_buf);
21b8aee8 421 rd->send_sgl.lkey = rd->send_mr->lkey;
422
423 rd->sq_wr.opcode = IBV_WR_SEND;
424 rd->sq_wr.send_flags = IBV_SEND_SIGNALED;
425 rd->sq_wr.sg_list = &rd->send_sgl;
426 rd->sq_wr.num_sge = 1;
76cc5224 427 rd->sq_wr.wr_id = FIO_RDMA_MAX_IO_DEPTH;
21b8aee8 428
429 return 0;
430}
431
432static int get_next_channel_event(struct thread_data *td,
433 struct rdma_event_channel *channel,
434 enum rdma_cm_event_type wait_event)
435{
436 struct rdmaio_data *rd = td->io_ops->data;
21b8aee8 437 struct rdma_cm_event *event;
222757cc 438 int ret;
21b8aee8 439
440 ret = rdma_get_cm_event(channel, &event);
441 if (ret) {
222757cc 442 log_err("fio: rdma_get_cm_event: %d\n", ret);
21b8aee8 443 return 1;
444 }
445
446 if (event->event != wait_event) {
447 log_err("fio: event is %s instead of %s\n",
448 rdma_event_str(event->event),
449 rdma_event_str(wait_event));
450 return 1;
451 }
452
453 switch (event->event) {
454 case RDMA_CM_EVENT_CONNECT_REQUEST:
455 rd->child_cm_id = event->id;
456 break;
457 default:
458 break;
459 }
460
461 rdma_ack_cm_event(event);
462
463 return 0;
464}
465
466static int fio_rdmaio_prep(struct thread_data *td, struct io_u *io_u)
467{
468 struct rdmaio_data *rd = td->io_ops->data;
469 struct rdma_io_u_data *r_io_u_d;
470
471 r_io_u_d = io_u->engine_data;
472
473 switch (rd->rdma_protocol) {
474 case FIO_RDMA_MEM_WRITE:
475 case FIO_RDMA_MEM_READ:
476 r_io_u_d->rdma_sgl.addr = (uint64_t) (unsigned long)io_u->buf;
477 r_io_u_d->rdma_sgl.lkey = io_u->mr->lkey;
478 r_io_u_d->sq_wr.wr_id = r_io_u_d->wr_id;
479 r_io_u_d->sq_wr.send_flags = IBV_SEND_SIGNALED;
480 r_io_u_d->sq_wr.sg_list = &r_io_u_d->rdma_sgl;
481 r_io_u_d->sq_wr.num_sge = 1;
482 break;
483 case FIO_RDMA_CHA_SEND:
484 r_io_u_d->rdma_sgl.addr = (uint64_t) (unsigned long)io_u->buf;
485 r_io_u_d->rdma_sgl.lkey = io_u->mr->lkey;
486 r_io_u_d->rdma_sgl.length = io_u->buflen;
487 r_io_u_d->sq_wr.wr_id = r_io_u_d->wr_id;
488 r_io_u_d->sq_wr.opcode = IBV_WR_SEND;
489 r_io_u_d->sq_wr.send_flags = IBV_SEND_SIGNALED;
490 r_io_u_d->sq_wr.sg_list = &r_io_u_d->rdma_sgl;
491 r_io_u_d->sq_wr.num_sge = 1;
492 break;
493 case FIO_RDMA_CHA_RECV:
494 r_io_u_d->rdma_sgl.addr = (uint64_t) (unsigned long)io_u->buf;
495 r_io_u_d->rdma_sgl.lkey = io_u->mr->lkey;
496 r_io_u_d->rdma_sgl.length = io_u->buflen;
497 r_io_u_d->rq_wr.wr_id = r_io_u_d->wr_id;
498 r_io_u_d->rq_wr.sg_list = &r_io_u_d->rdma_sgl;
499 r_io_u_d->rq_wr.num_sge = 1;
500 break;
501 default:
502 log_err("fio: unknown rdma protocol - %d\n", rd->rdma_protocol);
503 break;
504 }
505
506 return 0;
507}
508
509static struct io_u *fio_rdmaio_event(struct thread_data *td, int event)
510{
511 struct rdmaio_data *rd = td->io_ops->data;
512 struct io_u *io_u;
513 int i;
514
515 io_u = rd->io_us_completed[0];
222757cc 516 for (i = 0; i < rd->io_u_completed_nr - 1; i++)
21b8aee8 517 rd->io_us_completed[i] = rd->io_us_completed[i + 1];
222757cc 518
21b8aee8 519 rd->io_u_completed_nr--;
520
521 dprint_io_u(io_u, "fio_rdmaio_event");
522
523 return io_u;
524}
525
526static int fio_rdmaio_getevents(struct thread_data *td, unsigned int min,
527 unsigned int max, struct timespec *t)
528{
529 struct rdmaio_data *rd = td->io_ops->data;
21b8aee8 530 enum ibv_wc_opcode comp_opcode;
531 comp_opcode = IBV_WC_RDMA_WRITE;
532 struct ibv_cq *ev_cq;
533 void *ev_ctx;
222757cc 534 int ret, r = 0;
21b8aee8 535
536 switch (rd->rdma_protocol) {
537 case FIO_RDMA_MEM_WRITE:
538 comp_opcode = IBV_WC_RDMA_WRITE;
539 break;
540 case FIO_RDMA_MEM_READ:
541 comp_opcode = IBV_WC_RDMA_READ;
542 break;
543 case FIO_RDMA_CHA_SEND:
544 comp_opcode = IBV_WC_SEND;
545 break;
546 case FIO_RDMA_CHA_RECV:
547 comp_opcode = IBV_WC_RECV;
548 break;
549 default:
550 log_err("fio: unknown rdma protocol - %d\n", rd->rdma_protocol);
551 break;
552 }
553
554 if (rd->cq_event_num > 0) { /* previous left */
555 rd->cq_event_num--;
556 return 0;
557 }
558
559again:
560 if (ibv_get_cq_event(rd->channel, &ev_cq, &ev_ctx) != 0) {
561 log_err("fio: Failed to get cq event!\n");
562 return -1;
563 }
564 if (ev_cq != rd->cq) {
565 log_err("fio: Unknown CQ!\n");
566 return -1;
567 }
568 if (ibv_req_notify_cq(rd->cq, 0) != 0) {
569 log_err("fio: Failed to set notify!\n");
570 return -1;
571 }
572
573 ret = cq_event_handler(td, comp_opcode);
574 if (ret < 1)
575 goto again;
576
577 ibv_ack_cq_events(rd->cq, ret);
578
579 r += ret;
580 if (r < min)
581 goto again;
582
583 rd->cq_event_num -= r;
584
585 return r;
586}
587
588static int fio_rdmaio_send(struct thread_data *td, struct io_u **io_us,
589 unsigned int nr)
590{
591 struct rdmaio_data *rd = td->io_ops->data;
592 struct ibv_send_wr *bad_wr;
e07f72d3 593#if 0
21b8aee8 594 enum ibv_wc_opcode comp_opcode;
595 comp_opcode = IBV_WC_RDMA_WRITE;
e07f72d3 596#endif
7d7803fa
YR		 597 int i;
598 long index;
21b8aee8 599 struct rdma_io_u_data *r_io_u_d;
600
601 r_io_u_d = NULL;
602
603 for (i = 0; i < nr; i++) {
604 /* RDMA_WRITE or RDMA_READ */
605 switch (rd->rdma_protocol) {
606 case FIO_RDMA_MEM_WRITE:
607 /* compose work request */
608 r_io_u_d = io_us[i]->engine_data;
222757cc 609 index = __rand(&rd->rand_state) % rd->rmt_nr;
21b8aee8 610 r_io_u_d->sq_wr.opcode = IBV_WR_RDMA_WRITE;
611 r_io_u_d->sq_wr.wr.rdma.rkey = rd->rmt_us[index].rkey;
b6cf38f0
YR		 612 r_io_u_d->sq_wr.wr.rdma.remote_addr = \
613 rd->rmt_us[index].buf;
21b8aee8 614 r_io_u_d->sq_wr.sg_list->length = io_us[i]->buflen;
615 break;
616 case FIO_RDMA_MEM_READ:
617 /* compose work request */
618 r_io_u_d = io_us[i]->engine_data;
222757cc 619 index = __rand(&rd->rand_state) % rd->rmt_nr;
21b8aee8 620 r_io_u_d->sq_wr.opcode = IBV_WR_RDMA_READ;
621 r_io_u_d->sq_wr.wr.rdma.rkey = rd->rmt_us[index].rkey;
b6cf38f0
YR		 622 r_io_u_d->sq_wr.wr.rdma.remote_addr = \
623 rd->rmt_us[index].buf;
21b8aee8 624 r_io_u_d->sq_wr.sg_list->length = io_us[i]->buflen;
625 break;
626 case FIO_RDMA_CHA_SEND:
627 r_io_u_d = io_us[i]->engine_data;
628 r_io_u_d->sq_wr.opcode = IBV_WR_SEND;
629 r_io_u_d->sq_wr.send_flags = IBV_SEND_SIGNALED;
630 break;
631 default:
632 log_err("fio: unknown rdma protocol - %d\n",
633 rd->rdma_protocol);
634 break;
635 }
636
637 if (ibv_post_send(rd->qp, &r_io_u_d->sq_wr, &bad_wr) != 0) {
638 log_err("fio: ibv_post_send fail\n");
639 return -1;
640 }
641
642 dprint_io_u(io_us[i], "fio_rdmaio_send");
643 }
644
645 /* wait for completion
646 rdma_poll_wait(td, comp_opcode); */
647
648 return i;
649}
650
651static int fio_rdmaio_recv(struct thread_data *td, struct io_u **io_us,
652 unsigned int nr)
653{
654 struct rdmaio_data *rd = td->io_ops->data;
655 struct ibv_recv_wr *bad_wr;
656 struct rdma_io_u_data *r_io_u_d;
657 int i;
658
659 i = 0;
660 if (rd->rdma_protocol == FIO_RDMA_CHA_RECV) {
661 /* post io_u into recv queue */
662 for (i = 0; i < nr; i++) {
663 r_io_u_d = io_us[i]->engine_data;
664 if (ibv_post_recv(rd->qp, &r_io_u_d->rq_wr, &bad_wr) !=
665 0) {
666 log_err("fio: ibv_post_recv fail\n");
667 return 1;
668 }
669 }
670 } else if ((rd->rdma_protocol == FIO_RDMA_MEM_READ)
671 || (rd->rdma_protocol == FIO_RDMA_MEM_WRITE)) {
672 /* re-post the rq_wr */
673 if (ibv_post_recv(rd->qp, &rd->rq_wr, &bad_wr) != 0) {
674 log_err("fio: ibv_post_recv fail\n");
675 return 1;
676 }
677
678 rdma_poll_wait(td, IBV_WC_RECV);
679
680 dprint(FD_IO, "fio: recv FINISH message\n");
a05d62b2
YR		 681 td->done = 1;
682 return 0;
21b8aee8 683 }
684
685 return i;
686}
687
688static int fio_rdmaio_queue(struct thread_data *td, struct io_u *io_u)
689{
690 struct rdmaio_data *rd = td->io_ops->data;
691
692 fio_ro_check(td, io_u);
693
694 if (rd->io_u_queued_nr == (int)td->o.iodepth)
695 return FIO_Q_BUSY;
696
697 rd->io_us_queued[rd->io_u_queued_nr] = io_u;
698 rd->io_u_queued_nr++;
699
700 dprint_io_u(io_u, "fio_rdmaio_queue");
701
702 return FIO_Q_QUEUED;
703}
704
705static void fio_rdmaio_queued(struct thread_data *td, struct io_u **io_us,
706 unsigned int nr)
707{
708 struct rdmaio_data *rd = td->io_ops->data;
709 struct timeval now;
710 unsigned int i;
711
712 if (!fio_fill_issue_time(td))
713 return;
714
715 fio_gettime(&now, NULL);
716
717 for (i = 0; i < nr; i++) {
718 struct io_u *io_u = io_us[i];
719
720 /* queued -> flight */
721 rd->io_us_flight[rd->io_u_flight_nr] = io_u;
722 rd->io_u_flight_nr++;
723
724 memcpy(&io_u->issue_time, &now, sizeof(now));
725 io_u_queued(td, io_u);
726 }
727}
728
729static int fio_rdmaio_commit(struct thread_data *td)
730{
731 struct rdmaio_data *rd = td->io_ops->data;
732 struct io_u **io_us;
733 int ret;
734
735 if (!rd->io_us_queued)
736 return 0;
737
738 io_us = rd->io_us_queued;
739 do {
740 /* RDMA_WRITE or RDMA_READ */
222757cc 741 if (rd->is_client)
21b8aee8 742 ret = fio_rdmaio_send(td, io_us, rd->io_u_queued_nr);
222757cc 743 else if (!rd->is_client)
21b8aee8 744 ret = fio_rdmaio_recv(td, io_us, rd->io_u_queued_nr);
222757cc 745 else
21b8aee8 746 ret = 0; /* must be a SYNC */
747
748 if (ret > 0) {
749 fio_rdmaio_queued(td, io_us, ret);
750 io_u_mark_submit(td, ret);
751 rd->io_u_queued_nr -= ret;
752 io_us += ret;
753 ret = 0;
754 } else
755 break;
756 } while (rd->io_u_queued_nr);
757
758 return ret;
759}
760
761static int fio_rdmaio_connect(struct thread_data *td, struct fio_file *f)
762{
763 struct rdmaio_data *rd = td->io_ops->data;
764 struct rdma_conn_param conn_param;
765 struct ibv_send_wr *bad_wr;
766
222757cc 767 memset(&conn_param, 0, sizeof(conn_param));
21b8aee8 768 conn_param.responder_resources = 1;
769 conn_param.initiator_depth = 1;
770 conn_param.retry_count = 10;
771
772 if (rdma_connect(rd->cm_id, &conn_param) != 0) {
773 log_err("fio: rdma_connect fail\n");
774 return 1;
775 }
776
777 if (get_next_channel_event
778 (td, rd->cm_channel, RDMA_CM_EVENT_ESTABLISHED) != 0) {
779 log_err("fio: wait for RDMA_CM_EVENT_ESTABLISHED\n");
780 return 1;
781 }
782
783 /* send task request */
784 rd->send_buf.mode = htonl(rd->rdma_protocol);
785 rd->send_buf.nr = htonl(td->o.iodepth);
786
787 if (ibv_post_send(rd->qp, &rd->sq_wr, &bad_wr) != 0) {
788 log_err("fio: ibv_post_send fail");
789 return 1;
790 }
791
792 rdma_poll_wait(td, IBV_WC_SEND);
793
794 /* wait for remote MR info from server side */
795 rdma_poll_wait(td, IBV_WC_RECV);
796
7d7803fa
YR		 797 /* In SEND/RECV test, it's a good practice to setup the iodepth of
798 * of the RECV side deeper than that of the SEND side to
799 * avoid RNR (receiver not ready) error. The
222757cc 800 * SEND side may send so many unsolicited message before
7d7803fa
YR		 801 * RECV side commits sufficient recv buffers into recv queue.
802 * This may lead to RNR error. Here, SEND side pauses for a while
803 * during which RECV side commits sufficient recv buffers.
804 */
805 usleep(500000);
806
21b8aee8 807 return 0;
808}
809
810static int fio_rdmaio_accept(struct thread_data *td, struct fio_file *f)
811{
812 struct rdmaio_data *rd = td->io_ops->data;
813 struct rdma_conn_param conn_param;
814 struct ibv_send_wr *bad_wr;
815
816 /* rdma_accept() - then wait for accept success */
222757cc 817 memset(&conn_param, 0, sizeof(conn_param));
21b8aee8 818 conn_param.responder_resources = 1;
819 conn_param.initiator_depth = 1;
820
821 if (rdma_accept(rd->child_cm_id, &conn_param) != 0) {
822 log_err("fio: rdma_accept\n");
823 return 1;
824 }
825
826 if (get_next_channel_event
827 (td, rd->cm_channel, RDMA_CM_EVENT_ESTABLISHED) != 0) {
828 log_err("fio: wait for RDMA_CM_EVENT_ESTABLISHED\n");
829 return 1;
830 }
831
832 /* wait for request */
833 rdma_poll_wait(td, IBV_WC_RECV);
834
835 if (ibv_post_send(rd->qp, &rd->sq_wr, &bad_wr) != 0) {
836 log_err("fio: ibv_post_send fail");
837 return 1;
838 }
839
840 rdma_poll_wait(td, IBV_WC_SEND);
841
842 return 0;
843}
844
845static int fio_rdmaio_open_file(struct thread_data *td, struct fio_file *f)
846{
847 if (td_read(td))
848 return fio_rdmaio_accept(td, f);
849 else
850 return fio_rdmaio_connect(td, f);
851}
852
853static int fio_rdmaio_close_file(struct thread_data *td, struct fio_file *f)
854{
855 struct rdmaio_data *rd = td->io_ops->data;
856 struct ibv_send_wr *bad_wr;
857
858 /* unregister rdma buffer */
859
860 /*
861 * Client sends notification to the server side
862 */
863 /* refer to: http://linux.die.net/man/7/rdma_cm */
864 if ((rd->is_client == 1) && ((rd->rdma_protocol == FIO_RDMA_MEM_WRITE)
865 || (rd->rdma_protocol ==
866 FIO_RDMA_MEM_READ))) {
867 if (ibv_post_send(rd->qp, &rd->sq_wr, &bad_wr) != 0) {
868 log_err("fio: ibv_post_send fail");
869 return 1;
870 }
871
872 dprint(FD_IO, "fio: close infomation sent success\n");
873 rdma_poll_wait(td, IBV_WC_SEND);
874 }
875
876 if (rd->is_client == 1)
877 rdma_disconnect(rd->cm_id);
878 else {
879 rdma_disconnect(rd->child_cm_id);
222757cc
JA		 880#if 0
881 rdma_disconnect(rd->cm_id);
882#endif
21b8aee8 883 }
884
222757cc
JA		 885#if 0
886 if (get_next_channel_event(td, rd->cm_channel, RDMA_CM_EVENT_DISCONNECTED) != 0) {
887 log_err("fio: wait for RDMA_CM_EVENT_DISCONNECTED\n");
888 return 1;
889 }
890#endif
21b8aee8 891
21b8aee8 892 ibv_destroy_cq(rd->cq);
7d7803fa 893 ibv_destroy_qp(rd->qp);
21b8aee8 894
895 if (rd->is_client == 1)
896 rdma_destroy_id(rd->cm_id);
897 else {
898 rdma_destroy_id(rd->child_cm_id);
899 rdma_destroy_id(rd->cm_id);
900 }
901
902 ibv_destroy_comp_channel(rd->channel);
903 ibv_dealloc_pd(rd->pd);
904
905 return 0;
906}
907
908static int fio_rdmaio_setup_connect(struct thread_data *td, const char *host,
909 unsigned short port)
910{
911 struct rdmaio_data *rd = td->io_ops->data;
912 struct ibv_recv_wr *bad_wr;
222757cc 913 int err;
21b8aee8 914
915 rd->addr.sin_family = AF_INET;
916 rd->addr.sin_port = htons(port);
917
918 if (inet_aton(host, &rd->addr.sin_addr) != 1) {
919 struct hostent *hent;
920
921 hent = gethostbyname(host);
922 if (!hent) {
923 td_verror(td, errno, "gethostbyname");
924 return 1;
925 }
926
927 memcpy(&rd->addr.sin_addr, hent->h_addr, 4);
928 }
929
930 /* resolve route */
222757cc
JA		 931 err = rdma_resolve_addr(rd->cm_id, NULL, (struct sockaddr *)&rd->addr, 2000);
932 if (err != 0) {
933 log_err("fio: rdma_resolve_addr: %d\n", err);
21b8aee8 934 return 1;
935 }
936
222757cc
JA		 937 err = get_next_channel_event(td, rd->cm_channel, RDMA_CM_EVENT_ADDR_RESOLVED);
938 if (err != 0) {
939 log_err("fio: get_next_channel_event: %d\n", err);
21b8aee8 940 return 1;
941 }
942
943 /* resolve route */
222757cc
JA		 944 err = rdma_resolve_route(rd->cm_id, 2000);
945 if (err != 0) {
946 log_err("fio: rdma_resolve_route: %d\n", err);
21b8aee8 947 return 1;
948 }
949
222757cc
JA		 950 err = get_next_channel_event(td, rd->cm_channel, RDMA_CM_EVENT_ROUTE_RESOLVED);
951 if (err != 0) {
952 log_err("fio: get_next_channel_event: %d\n", err);
21b8aee8 953 return 1;
954 }
955
956 /* create qp and buffer */
957 if (fio_rdmaio_setup_qp(td) != 0)
958 return 1;
959
960 if (fio_rdmaio_setup_control_msg_buffers(td) != 0)
961 return 1;
962
963 /* post recv buf */
222757cc
JA		 964 err = ibv_post_recv(rd->qp, &rd->rq_wr, &bad_wr);
965 if (err != 0) {
966 log_err("fio: ibv_post_recv fail: %d\n", err);
21b8aee8 967 return 1;
968 }
969
970 return 0;
971}
972
973static int fio_rdmaio_setup_listen(struct thread_data *td, short port)
974{
975 struct rdmaio_data *rd = td->io_ops->data;
976 struct ibv_recv_wr *bad_wr;
977
978 rd->addr.sin_family = AF_INET;
979 rd->addr.sin_addr.s_addr = htonl(INADDR_ANY);
980 rd->addr.sin_port = htons(port);
981
982 /* rdma_listen */
983 if (rdma_bind_addr(rd->cm_id, (struct sockaddr *)&rd->addr) != 0) {
984 log_err("fio: rdma_bind_addr fail\n");
985 return 1;
986 }
987
988 if (rdma_listen(rd->cm_id, 3) != 0) {
989 log_err("fio: rdma_listen fail\n");
990 return 1;
991 }
992
993 /* wait for CONNECT_REQUEST */
994 if (get_next_channel_event
995 (td, rd->cm_channel, RDMA_CM_EVENT_CONNECT_REQUEST) != 0) {
996 log_err("fio: wait for RDMA_CM_EVENT_CONNECT_REQUEST\n");
997 return 1;
998 }
999
1000 if (fio_rdmaio_setup_qp(td) != 0)
1001 return 1;
1002
1003 if (fio_rdmaio_setup_control_msg_buffers(td) != 0)
1004 return 1;
1005
1006 /* post recv buf */
1007 if (ibv_post_recv(rd->qp, &rd->rq_wr, &bad_wr) != 0) {
1008 log_err("fio: ibv_post_recv fail\n");
1009 return 1;
1010 }
1011
1012 return 0;
1013}
1014
38b9354a 1015static int check_set_rlimits(struct thread_data *td)
21b8aee8 1016{
38b9354a 1017#ifdef CONFIG_RLIMIT_MEMLOCK
21b8aee8 1018 struct rlimit rl;
1019
21b8aee8 1020 /* check RLIMIT_MEMLOCK */
1021 if (getrlimit(RLIMIT_MEMLOCK, &rl) != 0) {
1022 log_err("fio: getrlimit fail: %d(%s)\n",
1023 errno, strerror(errno));
1024 return 1;
1025 }
1026
1027 /* soft limit */
1028 if ((rl.rlim_cur != RLIM_INFINITY)
1029 && (rl.rlim_cur < td->orig_buffer_size)) {
e07f72d3
BVA		 1030 log_err("fio: soft RLIMIT_MEMLOCK is: %" PRId64 "\n",
1031 rl.rlim_cur);
1032 log_err("fio: total block size is: %zd\n",
21b8aee8 1033 td->orig_buffer_size);
1034 /* try to set larger RLIMIT_MEMLOCK */
1035 rl.rlim_cur = rl.rlim_max;
1036 if (setrlimit(RLIMIT_MEMLOCK, &rl) != 0) {
1037 log_err("fio: setrlimit fail: %d(%s)\n",
1038 errno, strerror(errno));
1039 log_err("fio: you may try enlarge MEMLOCK by root\n");
1040 log_err("# ulimit -l unlimited\n");
1041 return 1;
1042 }
1043 }
38b9354a
JA		 1044#endif
1045
1046 return 0;
1047}
1048
1049static int fio_rdmaio_init(struct thread_data *td)
1050{
1051 struct rdmaio_data *rd = td->io_ops->data;
1052 struct flist_head *entry;
1053 unsigned int max_bs;
1054 unsigned int port;
1055 char host[64], buf[128];
1056 char *sep, *portp, *modep;
1057 int ret, i = 0;
1058
1059 if (td_rw(td)) {
1060 log_err("fio: rdma connections must be read OR write\n");
1061 return 1;
1062 }
1063 if (td_random(td)) {
1064 log_err("fio: RDMA network IO can't be random\n");
1065 return 1;
1066 }
1067
1068 if (check_set_rlimits(td))
1069 return 1;
21b8aee8 1070
1071 strcpy(buf, td->o.filename);
1072
1073 sep = strchr(buf, '/');
1074 if (!sep)
1075 goto bad_host;
1076
1077 *sep = '\0';
1078 sep++;
1079 strcpy(host, buf);
1080 if (!strlen(host))
1081 goto bad_host;
1082
1083 modep = NULL;
1084 portp = sep;
1085 sep = strchr(portp, '/');
1086 if (sep) {
1087 *sep = '\0';
1088 modep = sep + 1;
1089 }
1090
1091 port = strtol(portp, NULL, 10);
1092 if (!port || port > 65535)
1093 goto bad_host;
1094
1095 if (modep) {
1096 if (!strncmp("rdma_write", modep, strlen(modep)) ||
1097 !strncmp("RDMA_WRITE", modep, strlen(modep)))
1098 rd->rdma_protocol = FIO_RDMA_MEM_WRITE;
1099 else if (!strncmp("rdma_read", modep, strlen(modep)) ||
1100 !strncmp("RDMA_READ", modep, strlen(modep)))
1101 rd->rdma_protocol = FIO_RDMA_MEM_READ;
1102 else if (!strncmp("send", modep, strlen(modep)) ||
1103 !strncmp("SEND", modep, strlen(modep)))
1104 rd->rdma_protocol = FIO_RDMA_CHA_SEND;
1105 else
1106 goto bad_host;
1107 } else
1108 rd->rdma_protocol = FIO_RDMA_MEM_WRITE;
1109
1110 rd->cq_event_num = 0;
1111
1112 rd->cm_channel = rdma_create_event_channel();
1113 if (!rd->cm_channel) {
1114 log_err("fio: rdma_create_event_channel fail\n");
1115 return 1;
1116 }
1117
1118 ret = rdma_create_id(rd->cm_channel, &rd->cm_id, rd, RDMA_PS_TCP);
1119 if (ret) {
1120 log_err("fio: rdma_create_id fail\n");
1121 return 1;
1122 }
1123
1124 if ((rd->rdma_protocol == FIO_RDMA_MEM_WRITE) ||
1125 (rd->rdma_protocol == FIO_RDMA_MEM_READ)) {
1126 rd->rmt_us =
76cc5224 1127 malloc(FIO_RDMA_MAX_IO_DEPTH * sizeof(struct remote_u));
21b8aee8 1128 memset(rd->rmt_us, 0,
76cc5224 1129 FIO_RDMA_MAX_IO_DEPTH * sizeof(struct remote_u));
21b8aee8 1130 rd->rmt_nr = 0;
1131 }
1132
1133 rd->io_us_queued = malloc(td->o.iodepth * sizeof(struct io_u *));
1134 memset(rd->io_us_queued, 0, td->o.iodepth * sizeof(struct io_u *));
1135 rd->io_u_queued_nr = 0;
1136
1137 rd->io_us_flight = malloc(td->o.iodepth * sizeof(struct io_u *));
1138 memset(rd->io_us_flight, 0, td->o.iodepth * sizeof(struct io_u *));
1139 rd->io_u_flight_nr = 0;
1140
1141 rd->io_us_completed = malloc(td->o.iodepth * sizeof(struct io_u *));
1142 memset(rd->io_us_completed, 0, td->o.iodepth * sizeof(struct io_u *));
1143 rd->io_u_completed_nr = 0;
1144
1145 if (td_read(td)) { /* READ as the server */
1146 rd->is_client = 0;
1147 /* server rd->rdma_buf_len will be setup after got request */
1148 ret = fio_rdmaio_setup_listen(td, port);
1149 } else { /* WRITE as the client */
1150 rd->is_client = 1;
1151 ret = fio_rdmaio_setup_connect(td, host, port);
1152 }
1153
21b8aee8 1154 max_bs = max(td->o.max_bs[DDIR_READ], td->o.max_bs[DDIR_WRITE]);
1155 /* register each io_u in the free list */
21b8aee8 1156 flist_for_each(entry, &td->io_u_freelist) {
1157 struct io_u *io_u = flist_entry(entry, struct io_u, list);
1158
1159 io_u->engine_data = malloc(sizeof(struct rdma_io_u_data));
1160 memset(io_u->engine_data, 0, sizeof(struct rdma_io_u_data));
1161 ((struct rdma_io_u_data *)io_u->engine_data)->wr_id = i;
1162
1163 io_u->mr = ibv_reg_mr(rd->pd, io_u->buf, max_bs,
1164 IBV_ACCESS_LOCAL_WRITE |
1165 IBV_ACCESS_REMOTE_READ |
1166 IBV_ACCESS_REMOTE_WRITE);
1167 if (io_u->mr == NULL) {
1168 log_err("fio: ibv_reg_mr io_u failed\n");
1169 return 1;
1170 }
1171
1172 rd->send_buf.rmt_us[i].buf =
1173 htonll((uint64_t) (unsigned long)io_u->buf);
1174 rd->send_buf.rmt_us[i].rkey = htonl(io_u->mr->rkey);
1175 rd->send_buf.rmt_us[i].size = htonl(max_bs);
1176
222757cc
JA		 1177#if 0
1178 log_info("fio: Send rkey %x addr %" PRIx64 " len %d to client\n", io_u->mr->rkey, io_u->buf, max_bs); */
1179#endif
21b8aee8 1180 i++;
1181 }
1182
1183 rd->send_buf.nr = htonl(i);
1184
1185 return ret;
1186bad_host:
1187 log_err("fio: bad rdma host/port/protocol: %s\n", td->o.filename);
1188 return 1;
1189}
1190
1191static void fio_rdmaio_cleanup(struct thread_data *td)
1192{
1193 struct rdmaio_data *rd = td->io_ops->data;
1194
222757cc 1195 if (rd)
21b8aee8 1196 free(rd);
21b8aee8 1197}
1198
1199static int fio_rdmaio_setup(struct thread_data *td)
1200{
1201 struct rdmaio_data *rd;
1202
1203 if (!td->io_ops->data) {
222757cc 1204 rd = malloc(sizeof(*rd));
21b8aee8 1205
1206 memset(rd, 0, sizeof(*rd));
ee88d056 1207 init_rand_seed(&rd->rand_state, (unsigned int) GOLDEN_RATIO_PRIME);
21b8aee8 1208 td->io_ops->data = rd;
1209 }
1210
1211 return 0;
1212}
1213
1214static struct ioengine_ops ioengine_rw = {
b6cf38f0
YR		 1215 .name = "rdma",
1216 .version = FIO_IOOPS_VERSION,
1217 .setup = fio_rdmaio_setup,
1218 .init = fio_rdmaio_init,
1219 .prep = fio_rdmaio_prep,
1220 .queue = fio_rdmaio_queue,
1221 .commit = fio_rdmaio_commit,
1222 .getevents = fio_rdmaio_getevents,
1223 .event = fio_rdmaio_event,
1224 .cleanup = fio_rdmaio_cleanup,
1225 .open_file = fio_rdmaio_open_file,
1226 .close_file = fio_rdmaio_close_file,
1227 .flags = FIO_DISKLESSIO | FIO_UNIDIR | FIO_PIPEIO,
21b8aee8 1228};
1229
21b8aee8 1230static void fio_init fio_rdmaio_register(void)
1231{
1232 register_ioengine(&ioengine_rw);
1233}
1234
1235static void fio_exit fio_rdmaio_unregister(void)
1236{
1237 unregister_ioengine(&ioengine_rw);
1238}
fio - Flexible IO Tester