projects / fio.git / blame_incremental
? search:
summary | shortlog | log | commit | commitdiff | tree
blob | blame (non-incremental) | history | HEAD
Add suppor for logging fsync (and friends) latencies
[fio.git] / engines / rdma.c
... / ...
CommitLineData
1/*
2 * RDMA I/O engine
3 *
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.
7 *
8 * You will need the Linux RDMA software installed, either
9 * from your Linux distributor or directly from openfabrics.org:
10 *
11 * http://www.openfabrics.org/downloads/OFED/
12 *
13 * Exchanging steps of RDMA ioengine control messages:
14 * 1. client side sends test mode (RDMA_WRITE/RDMA_READ/SEND)
15 * to server side.
16 * 2. server side parses test mode, and sends back confirmation
17 * to client side. In RDMA WRITE/READ test, this confirmation
18 * includes memory information, such as rkey, address.
19 * 3. client side initiates test loop.
20 * 4. In RDMA WRITE/READ test, client side sends a completion
21 * notification to server side. Server side updates its
22 * td->done as true.
23 *
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
39#include <pthread.h>
40#include <inttypes.h>
41
42#include "../fio.h"
43#include "../hash.h"
44#include "../optgroup.h"
45
46#include <rdma/rdma_cma.h>
47
48#define FIO_RDMA_MAX_IO_DEPTH 512
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 rdmaio_options {
59 struct thread_data *td;
60 unsigned int port;
61 enum rdma_io_mode verb;
62 char *bindname;
63};
64
65static int str_hostname_cb(void *data, const char *input)
66{
67 struct rdmaio_options *o = data;
68
69 if (o->td->o.filename)
70 free(o->td->o.filename);
71 o->td->o.filename = strdup(input);
72 return 0;
73}
74
75static struct fio_option options[] = {
76 {
77 .name = "hostname",
78 .lname = "rdma engine hostname",
79 .type = FIO_OPT_STR_STORE,
80 .cb = str_hostname_cb,
81 .help = "Hostname for RDMA IO engine",
82 .category = FIO_OPT_C_ENGINE,
83 .group = FIO_OPT_G_RDMA,
84 },
85 {
86 .name = "bindname",
87 .lname = "rdma engine bindname",
88 .type = FIO_OPT_STR_STORE,
89 .off1 = offsetof(struct rdmaio_options, bindname),
90 .help = "Bind for RDMA IO engine",
91 .def = "",
92 .category = FIO_OPT_C_ENGINE,
93 .group = FIO_OPT_G_RDMA,
94 },
95 {
96 .name = "port",
97 .lname = "rdma engine port",
98 .type = FIO_OPT_INT,
99 .off1 = offsetof(struct rdmaio_options, port),
100 .minval = 1,
101 .maxval = 65535,
102 .help = "Port to use for RDMA connections",
103 .category = FIO_OPT_C_ENGINE,
104 .group = FIO_OPT_G_RDMA,
105 },
106 {
107 .name = "verb",
108 .lname = "RDMA engine verb",
109 .alias = "proto",
110 .type = FIO_OPT_STR,
111 .off1 = offsetof(struct rdmaio_options, verb),
112 .help = "RDMA engine verb",
113 .def = "write",
114 .posval = {
115 { .ival = "write",
116 .oval = FIO_RDMA_MEM_WRITE,
117 .help = "Memory Write",
118 },
119 { .ival = "read",
120 .oval = FIO_RDMA_MEM_READ,
121 .help = "Memory Read",
122 },
123 { .ival = "send",
124 .oval = FIO_RDMA_CHA_SEND,
125 .help = "Posted Send",
126 },
127 { .ival = "recv",
128 .oval = FIO_RDMA_CHA_RECV,
129 .help = "Posted Receive",
130 },
131 },
132 .category = FIO_OPT_C_ENGINE,
133 .group = FIO_OPT_G_RDMA,
134 },
135 {
136 .name = NULL,
137 },
138};
139
140struct remote_u {
141 uint64_t buf;
142 uint32_t rkey;
143 uint32_t size;
144};
145
146struct rdma_info_blk {
147 uint32_t mode; /* channel semantic or memory semantic */
148 uint32_t nr; /* client: io depth
149 server: number of records for memory semantic
150 */
151 uint32_t max_bs; /* maximum block size */
152 struct remote_u rmt_us[FIO_RDMA_MAX_IO_DEPTH];
153};
154
155struct rdma_io_u_data {
156 uint64_t wr_id;
157 struct ibv_send_wr sq_wr;
158 struct ibv_recv_wr rq_wr;
159 struct ibv_sge rdma_sgl;
160};
161
162struct rdmaio_data {
163 int is_client;
164 enum rdma_io_mode rdma_protocol;
165 char host[64];
166 struct sockaddr_in addr;
167
168 struct ibv_recv_wr rq_wr;
169 struct ibv_sge recv_sgl;
170 struct rdma_info_blk recv_buf;
171 struct ibv_mr *recv_mr;
172
173 struct ibv_send_wr sq_wr;
174 struct ibv_sge send_sgl;
175 struct rdma_info_blk send_buf;
176 struct ibv_mr *send_mr;
177
178 struct ibv_comp_channel *channel;
179 struct ibv_cq *cq;
180 struct ibv_pd *pd;
181 struct ibv_qp *qp;
182
183 pthread_t cmthread;
184 struct rdma_event_channel *cm_channel;
185 struct rdma_cm_id *cm_id;
186 struct rdma_cm_id *child_cm_id;
187
188 int cq_event_num;
189
190 struct remote_u *rmt_us;
191 int rmt_nr;
192 struct io_u **io_us_queued;
193 int io_u_queued_nr;
194 struct io_u **io_us_flight;
195 int io_u_flight_nr;
196 struct io_u **io_us_completed;
197 int io_u_completed_nr;
198
199 struct frand_state rand_state;
200};
201
202static int client_recv(struct thread_data *td, struct ibv_wc *wc)
203{
204 struct rdmaio_data *rd = td->io_ops_data;
205 unsigned int max_bs;
206
207 if (wc->byte_len != sizeof(rd->recv_buf)) {
208 log_err("Received bogus data, size %d\n", wc->byte_len);
209 return 1;
210 }
211
212 max_bs = max(td->o.max_bs[DDIR_READ], td->o.max_bs[DDIR_WRITE]);
213 if (max_bs > ntohl(rd->recv_buf.max_bs)) {
214 log_err("fio: Server's block size (%d) must be greater than or "
215 "equal to the client's block size (%d)!\n",
216 ntohl(rd->recv_buf.max_bs), max_bs);
217 return 1;
218 }
219
220 /* store mr info for MEMORY semantic */
221 if ((rd->rdma_protocol == FIO_RDMA_MEM_WRITE) ||
222 (rd->rdma_protocol == FIO_RDMA_MEM_READ)) {
223 /* struct flist_head *entry; */
224 int i = 0;
225
226 rd->rmt_nr = ntohl(rd->recv_buf.nr);
227
228 for (i = 0; i < rd->rmt_nr; i++) {
229 rd->rmt_us[i].buf = be64_to_cpu(rd->recv_buf.rmt_us[i].buf);
230 rd->rmt_us[i].rkey = ntohl(rd->recv_buf.rmt_us[i].rkey);
231 rd->rmt_us[i].size = ntohl(rd->recv_buf.rmt_us[i].size);
232
233 dprint(FD_IO,
234 "fio: Received rkey %x addr %" PRIx64
235 " len %d from peer\n", rd->rmt_us[i].rkey,
236 rd->rmt_us[i].buf, rd->rmt_us[i].size);
237 }
238 }
239
240 return 0;
241}
242
243static int server_recv(struct thread_data *td, struct ibv_wc *wc)
244{
245 struct rdmaio_data *rd = td->io_ops_data;
246 unsigned int max_bs;
247
248 if (wc->wr_id == FIO_RDMA_MAX_IO_DEPTH) {
249 rd->rdma_protocol = ntohl(rd->recv_buf.mode);
250
251 /* CHANNEL semantic, do nothing */
252 if (rd->rdma_protocol == FIO_RDMA_CHA_SEND)
253 rd->rdma_protocol = FIO_RDMA_CHA_RECV;
254
255 max_bs = max(td->o.max_bs[DDIR_READ], td->o.max_bs[DDIR_WRITE]);
256 if (max_bs < ntohl(rd->recv_buf.max_bs)) {
257 log_err("fio: Server's block size (%d) must be greater than or "
258 "equal to the client's block size (%d)!\n",
259 ntohl(rd->recv_buf.max_bs), max_bs);
260 return 1;
261 }
262
263 }
264
265 return 0;
266}
267
268static int cq_event_handler(struct thread_data *td, enum ibv_wc_opcode opcode)
269{
270 struct rdmaio_data *rd = td->io_ops_data;
271 struct ibv_wc wc;
272 struct rdma_io_u_data *r_io_u_d;
273 int ret;
274 int compevnum = 0;
275 int i;
276
277 while ((ret = ibv_poll_cq(rd->cq, 1, &wc)) == 1) {
278 ret = 0;
279 compevnum++;
280
281 if (wc.status) {
282 log_err("fio: cq completion status %d(%s)\n",
283 wc.status, ibv_wc_status_str(wc.status));
284 return -1;
285 }
286
287 switch (wc.opcode) {
288
289 case IBV_WC_RECV:
290 if (rd->is_client == 1)
291 ret = client_recv(td, &wc);
292 else
293 ret = server_recv(td, &wc);
294
295 if (ret)
296 return -1;
297
298 if (wc.wr_id == FIO_RDMA_MAX_IO_DEPTH)
299 break;
300
301 for (i = 0; i < rd->io_u_flight_nr; i++) {
302 r_io_u_d = rd->io_us_flight[i]->engine_data;
303
304 if (wc.wr_id == r_io_u_d->rq_wr.wr_id) {
305 rd->io_us_flight[i]->resid =
306 rd->io_us_flight[i]->buflen
307 - wc.byte_len;
308
309 rd->io_us_flight[i]->error = 0;
310
311 rd->io_us_completed[rd->
312 io_u_completed_nr]
313 = rd->io_us_flight[i];
314 rd->io_u_completed_nr++;
315 break;
316 }
317 }
318 if (i == rd->io_u_flight_nr)
319 log_err("fio: recv wr %" PRId64 " not found\n",
320 wc.wr_id);
321 else {
322 /* put the last one into middle of the list */
323 rd->io_us_flight[i] =
324 rd->io_us_flight[rd->io_u_flight_nr - 1];
325 rd->io_u_flight_nr--;
326 }
327
328 break;
329
330 case IBV_WC_SEND:
331 case IBV_WC_RDMA_WRITE:
332 case IBV_WC_RDMA_READ:
333 if (wc.wr_id == FIO_RDMA_MAX_IO_DEPTH)
334 break;
335
336 for (i = 0; i < rd->io_u_flight_nr; i++) {
337 r_io_u_d = rd->io_us_flight[i]->engine_data;
338
339 if (wc.wr_id == r_io_u_d->sq_wr.wr_id) {
340 rd->io_us_completed[rd->
341 io_u_completed_nr]
342 = rd->io_us_flight[i];
343 rd->io_u_completed_nr++;
344 break;
345 }
346 }
347 if (i == rd->io_u_flight_nr)
348 log_err("fio: send wr %" PRId64 " not found\n",
349 wc.wr_id);
350 else {
351 /* put the last one into middle of the list */
352 rd->io_us_flight[i] =
353 rd->io_us_flight[rd->io_u_flight_nr - 1];
354 rd->io_u_flight_nr--;
355 }
356
357 break;
358
359 default:
360 log_info("fio: unknown completion event %d\n",
361 wc.opcode);
362 return -1;
363 }
364 rd->cq_event_num++;
365 }
366
367 if (ret) {
368 log_err("fio: poll error %d\n", ret);
369 return 1;
370 }
371
372 return compevnum;
373}
374
375/*
376 * Return -1 for error and 'nr events' for a positive number
377 * of events
378 */
379static int rdma_poll_wait(struct thread_data *td, enum ibv_wc_opcode opcode)
380{
381 struct rdmaio_data *rd = td->io_ops_data;
382 struct ibv_cq *ev_cq;
383 void *ev_ctx;
384 int ret;
385
386 if (rd->cq_event_num > 0) { /* previous left */
387 rd->cq_event_num--;
388 return 0;
389 }
390
391again:
392 if (ibv_get_cq_event(rd->channel, &ev_cq, &ev_ctx) != 0) {
393 log_err("fio: Failed to get cq event!\n");
394 return -1;
395 }
396 if (ev_cq != rd->cq) {
397 log_err("fio: Unknown CQ!\n");
398 return -1;
399 }
400 if (ibv_req_notify_cq(rd->cq, 0) != 0) {
401 log_err("fio: Failed to set notify!\n");
402 return -1;
403 }
404
405 ret = cq_event_handler(td, opcode);
406 if (ret == 0)
407 goto again;
408
409 ibv_ack_cq_events(rd->cq, ret);
410
411 rd->cq_event_num--;
412
413 return ret;
414}
415
416static int fio_rdmaio_setup_qp(struct thread_data *td)
417{
418 struct rdmaio_data *rd = td->io_ops_data;
419 struct ibv_qp_init_attr init_attr;
420 int qp_depth = td->o.iodepth * 2; /* 2 times of io depth */
421
422 if (rd->is_client == 0)
423 rd->pd = ibv_alloc_pd(rd->child_cm_id->verbs);
424 else
425 rd->pd = ibv_alloc_pd(rd->cm_id->verbs);
426
427 if (rd->pd == NULL) {
428 log_err("fio: ibv_alloc_pd fail: %m\n");
429 return 1;
430 }
431
432 if (rd->is_client == 0)
433 rd->channel = ibv_create_comp_channel(rd->child_cm_id->verbs);
434 else
435 rd->channel = ibv_create_comp_channel(rd->cm_id->verbs);
436 if (rd->channel == NULL) {
437 log_err("fio: ibv_create_comp_channel fail: %m\n");
438 goto err1;
439 }
440
441 if (qp_depth < 16)
442 qp_depth = 16;
443
444 if (rd->is_client == 0)
445 rd->cq = ibv_create_cq(rd->child_cm_id->verbs,
446 qp_depth, rd, rd->channel, 0);
447 else
448 rd->cq = ibv_create_cq(rd->cm_id->verbs,
449 qp_depth, rd, rd->channel, 0);
450 if (rd->cq == NULL) {
451 log_err("fio: ibv_create_cq failed: %m\n");
452 goto err2;
453 }
454
455 if (ibv_req_notify_cq(rd->cq, 0) != 0) {
456 log_err("fio: ibv_req_notify_cq failed: %m\n");
457 goto err3;
458 }
459
460 /* create queue pair */
461 memset(&init_attr, 0, sizeof(init_attr));
462 init_attr.cap.max_send_wr = qp_depth;
463 init_attr.cap.max_recv_wr = qp_depth;
464 init_attr.cap.max_recv_sge = 1;
465 init_attr.cap.max_send_sge = 1;
466 init_attr.qp_type = IBV_QPT_RC;
467 init_attr.send_cq = rd->cq;
468 init_attr.recv_cq = rd->cq;
469
470 if (rd->is_client == 0) {
471 if (rdma_create_qp(rd->child_cm_id, rd->pd, &init_attr) != 0) {
472 log_err("fio: rdma_create_qp failed: %m\n");
473 goto err3;
474 }
475 rd->qp = rd->child_cm_id->qp;
476 } else {
477 if (rdma_create_qp(rd->cm_id, rd->pd, &init_attr) != 0) {
478 log_err("fio: rdma_create_qp failed: %m\n");
479 goto err3;
480 }
481 rd->qp = rd->cm_id->qp;
482 }
483
484 return 0;
485
486err3:
487 ibv_destroy_cq(rd->cq);
488err2:
489 ibv_destroy_comp_channel(rd->channel);
490err1:
491 ibv_dealloc_pd(rd->pd);
492
493 return 1;
494}
495
496static int fio_rdmaio_setup_control_msg_buffers(struct thread_data *td)
497{
498 struct rdmaio_data *rd = td->io_ops_data;
499
500 rd->recv_mr = ibv_reg_mr(rd->pd, &rd->recv_buf, sizeof(rd->recv_buf),
501 IBV_ACCESS_LOCAL_WRITE);
502 if (rd->recv_mr == NULL) {
503 log_err("fio: recv_buf reg_mr failed: %m\n");
504 return 1;
505 }
506
507 rd->send_mr = ibv_reg_mr(rd->pd, &rd->send_buf, sizeof(rd->send_buf),
508 0);
509 if (rd->send_mr == NULL) {
510 log_err("fio: send_buf reg_mr failed: %m\n");
511 ibv_dereg_mr(rd->recv_mr);
512 return 1;
513 }
514
515 /* setup work request */
516 /* recv wq */
517 rd->recv_sgl.addr = (uint64_t) (unsigned long)&rd->recv_buf;
518 rd->recv_sgl.length = sizeof(rd->recv_buf);
519 rd->recv_sgl.lkey = rd->recv_mr->lkey;
520 rd->rq_wr.sg_list = &rd->recv_sgl;
521 rd->rq_wr.num_sge = 1;
522 rd->rq_wr.wr_id = FIO_RDMA_MAX_IO_DEPTH;
523
524 /* send wq */
525 rd->send_sgl.addr = (uint64_t) (unsigned long)&rd->send_buf;
526 rd->send_sgl.length = sizeof(rd->send_buf);
527 rd->send_sgl.lkey = rd->send_mr->lkey;
528
529 rd->sq_wr.opcode = IBV_WR_SEND;
530 rd->sq_wr.send_flags = IBV_SEND_SIGNALED;
531 rd->sq_wr.sg_list = &rd->send_sgl;
532 rd->sq_wr.num_sge = 1;
533 rd->sq_wr.wr_id = FIO_RDMA_MAX_IO_DEPTH;
534
535 return 0;
536}
537
538static int get_next_channel_event(struct thread_data *td,
539 struct rdma_event_channel *channel,
540 enum rdma_cm_event_type wait_event)
541{
542 struct rdmaio_data *rd = td->io_ops_data;
543 struct rdma_cm_event *event;
544 int ret;
545
546 ret = rdma_get_cm_event(channel, &event);
547 if (ret) {
548 log_err("fio: rdma_get_cm_event: %d\n", ret);
549 return 1;
550 }
551
552 if (event->event != wait_event) {
553 log_err("fio: event is %s instead of %s\n",
554 rdma_event_str(event->event),
555 rdma_event_str(wait_event));
556 return 1;
557 }
558
559 switch (event->event) {
560 case RDMA_CM_EVENT_CONNECT_REQUEST:
561 rd->child_cm_id = event->id;
562 break;
563 default:
564 break;
565 }
566
567 rdma_ack_cm_event(event);
568
569 return 0;
570}
571
572static int fio_rdmaio_prep(struct thread_data *td, struct io_u *io_u)
573{
574 struct rdmaio_data *rd = td->io_ops_data;
575 struct rdma_io_u_data *r_io_u_d;
576
577 r_io_u_d = io_u->engine_data;
578
579 switch (rd->rdma_protocol) {
580 case FIO_RDMA_MEM_WRITE:
581 case FIO_RDMA_MEM_READ:
582 r_io_u_d->rdma_sgl.addr = (uint64_t) (unsigned long)io_u->buf;
583 r_io_u_d->rdma_sgl.lkey = io_u->mr->lkey;
584 r_io_u_d->sq_wr.wr_id = r_io_u_d->wr_id;
585 r_io_u_d->sq_wr.send_flags = IBV_SEND_SIGNALED;
586 r_io_u_d->sq_wr.sg_list = &r_io_u_d->rdma_sgl;
587 r_io_u_d->sq_wr.num_sge = 1;
588 break;
589 case FIO_RDMA_CHA_SEND:
590 r_io_u_d->rdma_sgl.addr = (uint64_t) (unsigned long)io_u->buf;
591 r_io_u_d->rdma_sgl.lkey = io_u->mr->lkey;
592 r_io_u_d->rdma_sgl.length = io_u->buflen;
593 r_io_u_d->sq_wr.wr_id = r_io_u_d->wr_id;
594 r_io_u_d->sq_wr.opcode = IBV_WR_SEND;
595 r_io_u_d->sq_wr.send_flags = IBV_SEND_SIGNALED;
596 r_io_u_d->sq_wr.sg_list = &r_io_u_d->rdma_sgl;
597 r_io_u_d->sq_wr.num_sge = 1;
598 break;
599 case FIO_RDMA_CHA_RECV:
600 r_io_u_d->rdma_sgl.addr = (uint64_t) (unsigned long)io_u->buf;
601 r_io_u_d->rdma_sgl.lkey = io_u->mr->lkey;
602 r_io_u_d->rdma_sgl.length = io_u->buflen;
603 r_io_u_d->rq_wr.wr_id = r_io_u_d->wr_id;
604 r_io_u_d->rq_wr.sg_list = &r_io_u_d->rdma_sgl;
605 r_io_u_d->rq_wr.num_sge = 1;
606 break;
607 default:
608 log_err("fio: unknown rdma protocol - %d\n", rd->rdma_protocol);
609 break;
610 }
611
612 return 0;
613}
614
615static struct io_u *fio_rdmaio_event(struct thread_data *td, int event)
616{
617 struct rdmaio_data *rd = td->io_ops_data;
618 struct io_u *io_u;
619 int i;
620
621 io_u = rd->io_us_completed[0];
622 for (i = 0; i < rd->io_u_completed_nr - 1; i++)
623 rd->io_us_completed[i] = rd->io_us_completed[i + 1];
624
625 rd->io_u_completed_nr--;
626
627 dprint_io_u(io_u, "fio_rdmaio_event");
628
629 return io_u;
630}
631
632static int fio_rdmaio_getevents(struct thread_data *td, unsigned int min,
633 unsigned int max, const struct timespec *t)
634{
635 struct rdmaio_data *rd = td->io_ops_data;
636 enum ibv_wc_opcode comp_opcode;
637 struct ibv_cq *ev_cq;
638 void *ev_ctx;
639 int ret, r = 0;
640 comp_opcode = IBV_WC_RDMA_WRITE;
641
642 switch (rd->rdma_protocol) {
643 case FIO_RDMA_MEM_WRITE:
644 comp_opcode = IBV_WC_RDMA_WRITE;
645 break;
646 case FIO_RDMA_MEM_READ:
647 comp_opcode = IBV_WC_RDMA_READ;
648 break;
649 case FIO_RDMA_CHA_SEND:
650 comp_opcode = IBV_WC_SEND;
651 break;
652 case FIO_RDMA_CHA_RECV:
653 comp_opcode = IBV_WC_RECV;
654 break;
655 default:
656 log_err("fio: unknown rdma protocol - %d\n", rd->rdma_protocol);
657 break;
658 }
659
660 if (rd->cq_event_num > 0) { /* previous left */
661 rd->cq_event_num--;
662 return 0;
663 }
664
665again:
666 if (ibv_get_cq_event(rd->channel, &ev_cq, &ev_ctx) != 0) {
667 log_err("fio: Failed to get cq event!\n");
668 return -1;
669 }
670 if (ev_cq != rd->cq) {
671 log_err("fio: Unknown CQ!\n");
672 return -1;
673 }
674 if (ibv_req_notify_cq(rd->cq, 0) != 0) {
675 log_err("fio: Failed to set notify!\n");
676 return -1;
677 }
678
679 ret = cq_event_handler(td, comp_opcode);
680 if (ret < 1)
681 goto again;
682
683 ibv_ack_cq_events(rd->cq, ret);
684
685 r += ret;
686 if (r < min)
687 goto again;
688
689 rd->cq_event_num -= r;
690
691 return r;
692}
693
694static int fio_rdmaio_send(struct thread_data *td, struct io_u **io_us,
695 unsigned int nr)
696{
697 struct rdmaio_data *rd = td->io_ops_data;
698 struct ibv_send_wr *bad_wr;
699#if 0
700 enum ibv_wc_opcode comp_opcode;
701 comp_opcode = IBV_WC_RDMA_WRITE;
702#endif
703 int i;
704 long index;
705 struct rdma_io_u_data *r_io_u_d;
706
707 r_io_u_d = NULL;
708
709 for (i = 0; i < nr; i++) {
710 /* RDMA_WRITE or RDMA_READ */
711 switch (rd->rdma_protocol) {
712 case FIO_RDMA_MEM_WRITE:
713 /* compose work request */
714 r_io_u_d = io_us[i]->engine_data;
715 index = __rand(&rd->rand_state) % rd->rmt_nr;
716 r_io_u_d->sq_wr.opcode = IBV_WR_RDMA_WRITE;
717 r_io_u_d->sq_wr.wr.rdma.rkey = rd->rmt_us[index].rkey;
718 r_io_u_d->sq_wr.wr.rdma.remote_addr = \
719 rd->rmt_us[index].buf;
720 r_io_u_d->sq_wr.sg_list->length = io_us[i]->buflen;
721 break;
722 case FIO_RDMA_MEM_READ:
723 /* compose work request */
724 r_io_u_d = io_us[i]->engine_data;
725 index = __rand(&rd->rand_state) % rd->rmt_nr;
726 r_io_u_d->sq_wr.opcode = IBV_WR_RDMA_READ;
727 r_io_u_d->sq_wr.wr.rdma.rkey = rd->rmt_us[index].rkey;
728 r_io_u_d->sq_wr.wr.rdma.remote_addr = \
729 rd->rmt_us[index].buf;
730 r_io_u_d->sq_wr.sg_list->length = io_us[i]->buflen;
731 break;
732 case FIO_RDMA_CHA_SEND:
733 r_io_u_d = io_us[i]->engine_data;
734 r_io_u_d->sq_wr.opcode = IBV_WR_SEND;
735 r_io_u_d->sq_wr.send_flags = IBV_SEND_SIGNALED;
736 break;
737 default:
738 log_err("fio: unknown rdma protocol - %d\n",
739 rd->rdma_protocol);
740 break;
741 }
742
743 if (ibv_post_send(rd->qp, &r_io_u_d->sq_wr, &bad_wr) != 0) {
744 log_err("fio: ibv_post_send fail: %m\n");
745 return -1;
746 }
747
748 dprint_io_u(io_us[i], "fio_rdmaio_send");
749 }
750
751 /* wait for completion
752 rdma_poll_wait(td, comp_opcode); */
753
754 return i;
755}
756
757static int fio_rdmaio_recv(struct thread_data *td, struct io_u **io_us,
758 unsigned int nr)
759{
760 struct rdmaio_data *rd = td->io_ops_data;
761 struct ibv_recv_wr *bad_wr;
762 struct rdma_io_u_data *r_io_u_d;
763 int i;
764
765 i = 0;
766 if (rd->rdma_protocol == FIO_RDMA_CHA_RECV) {
767 /* post io_u into recv queue */
768 for (i = 0; i < nr; i++) {
769 r_io_u_d = io_us[i]->engine_data;
770 if (ibv_post_recv(rd->qp, &r_io_u_d->rq_wr, &bad_wr) !=
771 0) {
772 log_err("fio: ibv_post_recv fail: %m\n");
773 return 1;
774 }
775 }
776 } else if ((rd->rdma_protocol == FIO_RDMA_MEM_READ)
777 || (rd->rdma_protocol == FIO_RDMA_MEM_WRITE)) {
778 /* re-post the rq_wr */
779 if (ibv_post_recv(rd->qp, &rd->rq_wr, &bad_wr) != 0) {
780 log_err("fio: ibv_post_recv fail: %m\n");
781 return 1;
782 }
783
784 rdma_poll_wait(td, IBV_WC_RECV);
785
786 dprint(FD_IO, "fio: recv FINISH message\n");
787 td->done = 1;
788 return 0;
789 }
790
791 return i;
792}
793
794static int fio_rdmaio_queue(struct thread_data *td, struct io_u *io_u)
795{
796 struct rdmaio_data *rd = td->io_ops_data;
797
798 fio_ro_check(td, io_u);
799
800 if (rd->io_u_queued_nr == (int)td->o.iodepth)
801 return FIO_Q_BUSY;
802
803 rd->io_us_queued[rd->io_u_queued_nr] = io_u;
804 rd->io_u_queued_nr++;
805
806 dprint_io_u(io_u, "fio_rdmaio_queue");
807
808 return FIO_Q_QUEUED;
809}
810
811static void fio_rdmaio_queued(struct thread_data *td, struct io_u **io_us,
812 unsigned int nr)
813{
814 struct rdmaio_data *rd = td->io_ops_data;
815 struct timespec now;
816 unsigned int i;
817
818 if (!fio_fill_issue_time(td))
819 return;
820
821 fio_gettime(&now, NULL);
822
823 for (i = 0; i < nr; i++) {
824 struct io_u *io_u = io_us[i];
825
826 /* queued -> flight */
827 rd->io_us_flight[rd->io_u_flight_nr] = io_u;
828 rd->io_u_flight_nr++;
829
830 memcpy(&io_u->issue_time, &now, sizeof(now));
831 io_u_queued(td, io_u);
832 }
833}
834
835static int fio_rdmaio_commit(struct thread_data *td)
836{
837 struct rdmaio_data *rd = td->io_ops_data;
838 struct io_u **io_us;
839 int ret;
840
841 if (!rd->io_us_queued)
842 return 0;
843
844 io_us = rd->io_us_queued;
845 do {
846 /* RDMA_WRITE or RDMA_READ */
847 if (rd->is_client)
848 ret = fio_rdmaio_send(td, io_us, rd->io_u_queued_nr);
849 else if (!rd->is_client)
850 ret = fio_rdmaio_recv(td, io_us, rd->io_u_queued_nr);
851 else
852 ret = 0; /* must be a SYNC */
853
854 if (ret > 0) {
855 fio_rdmaio_queued(td, io_us, ret);
856 io_u_mark_submit(td, ret);
857 rd->io_u_queued_nr -= ret;
858 io_us += ret;
859 ret = 0;
860 } else
861 break;
862 } while (rd->io_u_queued_nr);
863
864 return ret;
865}
866
867static int fio_rdmaio_connect(struct thread_data *td, struct fio_file *f)
868{
869 struct rdmaio_data *rd = td->io_ops_data;
870 struct rdma_conn_param conn_param;
871 struct ibv_send_wr *bad_wr;
872
873 memset(&conn_param, 0, sizeof(conn_param));
874 conn_param.responder_resources = 1;
875 conn_param.initiator_depth = 1;
876 conn_param.retry_count = 10;
877
878 if (rdma_connect(rd->cm_id, &conn_param) != 0) {
879 log_err("fio: rdma_connect fail: %m\n");
880 return 1;
881 }
882
883 if (get_next_channel_event
884 (td, rd->cm_channel, RDMA_CM_EVENT_ESTABLISHED) != 0) {
885 log_err("fio: wait for RDMA_CM_EVENT_ESTABLISHED\n");
886 return 1;
887 }
888
889 /* send task request */
890 rd->send_buf.mode = htonl(rd->rdma_protocol);
891 rd->send_buf.nr = htonl(td->o.iodepth);
892
893 if (ibv_post_send(rd->qp, &rd->sq_wr, &bad_wr) != 0) {
894 log_err("fio: ibv_post_send fail: %m\n");
895 return 1;
896 }
897
898 if (rdma_poll_wait(td, IBV_WC_SEND) < 0)
899 return 1;
900
901 /* wait for remote MR info from server side */
902 if (rdma_poll_wait(td, IBV_WC_RECV) < 0)
903 return 1;
904
905 /* In SEND/RECV test, it's a good practice to setup the iodepth of
906 * of the RECV side deeper than that of the SEND side to
907 * avoid RNR (receiver not ready) error. The
908 * SEND side may send so many unsolicited message before
909 * RECV side commits sufficient recv buffers into recv queue.
910 * This may lead to RNR error. Here, SEND side pauses for a while
911 * during which RECV side commits sufficient recv buffers.
912 */
913 usleep(500000);
914
915 return 0;
916}
917
918static int fio_rdmaio_accept(struct thread_data *td, struct fio_file *f)
919{
920 struct rdmaio_data *rd = td->io_ops_data;
921 struct rdma_conn_param conn_param;
922 struct ibv_send_wr *bad_wr;
923 int ret = 0;
924
925 /* rdma_accept() - then wait for accept success */
926 memset(&conn_param, 0, sizeof(conn_param));
927 conn_param.responder_resources = 1;
928 conn_param.initiator_depth = 1;
929
930 if (rdma_accept(rd->child_cm_id, &conn_param) != 0) {
931 log_err("fio: rdma_accept: %m\n");
932 return 1;
933 }
934
935 if (get_next_channel_event
936 (td, rd->cm_channel, RDMA_CM_EVENT_ESTABLISHED) != 0) {
937 log_err("fio: wait for RDMA_CM_EVENT_ESTABLISHED\n");
938 return 1;
939 }
940
941 /* wait for request */
942 ret = rdma_poll_wait(td, IBV_WC_RECV) < 0;
943
944 if (ibv_post_send(rd->qp, &rd->sq_wr, &bad_wr) != 0) {
945 log_err("fio: ibv_post_send fail: %m\n");
946 return 1;
947 }
948
949 if (rdma_poll_wait(td, IBV_WC_SEND) < 0)
950 return 1;
951
952 return ret;
953}
954
955static int fio_rdmaio_open_file(struct thread_data *td, struct fio_file *f)
956{
957 if (td_read(td))
958 return fio_rdmaio_accept(td, f);
959 else
960 return fio_rdmaio_connect(td, f);
961}
962
963static int fio_rdmaio_close_file(struct thread_data *td, struct fio_file *f)
964{
965 struct rdmaio_data *rd = td->io_ops_data;
966 struct ibv_send_wr *bad_wr;
967
968 /* unregister rdma buffer */
969
970 /*
971 * Client sends notification to the server side
972 */
973 /* refer to: http://linux.die.net/man/7/rdma_cm */
974 if ((rd->is_client == 1) && ((rd->rdma_protocol == FIO_RDMA_MEM_WRITE)
975 || (rd->rdma_protocol ==
976 FIO_RDMA_MEM_READ))) {
977 if (ibv_post_send(rd->qp, &rd->sq_wr, &bad_wr) != 0) {
978 log_err("fio: ibv_post_send fail: %m\n");
979 return 1;
980 }
981
982 dprint(FD_IO, "fio: close information sent success\n");
983 rdma_poll_wait(td, IBV_WC_SEND);
984 }
985
986 if (rd->is_client == 1)
987 rdma_disconnect(rd->cm_id);
988 else {
989 rdma_disconnect(rd->child_cm_id);
990#if 0
991 rdma_disconnect(rd->cm_id);
992#endif
993 }
994
995#if 0
996 if (get_next_channel_event(td, rd->cm_channel, RDMA_CM_EVENT_DISCONNECTED) != 0) {
997 log_err("fio: wait for RDMA_CM_EVENT_DISCONNECTED\n");
998 return 1;
999 }
1000#endif
1001
1002 ibv_destroy_cq(rd->cq);
1003 ibv_destroy_qp(rd->qp);
1004
1005 if (rd->is_client == 1)
1006 rdma_destroy_id(rd->cm_id);
1007 else {
1008 rdma_destroy_id(rd->child_cm_id);
1009 rdma_destroy_id(rd->cm_id);
1010 }
1011
1012 ibv_destroy_comp_channel(rd->channel);
1013 ibv_dealloc_pd(rd->pd);
1014
1015 return 0;
1016}
1017
1018static int aton(struct thread_data *td, const char *host,
1019 struct sockaddr_in *addr)
1020{
1021 if (inet_aton(host, &addr->sin_addr) != 1) {
1022 struct hostent *hent;
1023
1024 hent = gethostbyname(host);
1025 if (!hent) {
1026 td_verror(td, errno, "gethostbyname");
1027 return 1;
1028 }
1029
1030 memcpy(&addr->sin_addr, hent->h_addr, 4);
1031 }
1032 return 0;
1033}
1034
1035static int fio_rdmaio_setup_connect(struct thread_data *td, const char *host,
1036 unsigned short port)
1037{
1038 struct rdmaio_data *rd = td->io_ops_data;
1039 struct rdmaio_options *o = td->eo;
1040 struct sockaddr_storage addrb;
1041 struct ibv_recv_wr *bad_wr;
1042 int err;
1043
1044 rd->addr.sin_family = AF_INET;
1045 rd->addr.sin_port = htons(port);
1046
1047 err = aton(td, host, &rd->addr);
1048 if (err)
1049 return err;
1050
1051 /* resolve route */
1052 if (strcmp(o->bindname, "") != 0) {
1053 addrb.ss_family = AF_INET;
1054 err = aton(td, o->bindname, (struct sockaddr_in *)&addrb);
1055 if (err)
1056 return err;
1057 err = rdma_resolve_addr(rd->cm_id, (struct sockaddr *)&addrb,
1058 (struct sockaddr *)&rd->addr, 2000);
1059
1060 } else {
1061 err = rdma_resolve_addr(rd->cm_id, NULL,
1062 (struct sockaddr *)&rd->addr, 2000);
1063 }
1064
1065 if (err != 0) {
1066 log_err("fio: rdma_resolve_addr: %d\n", err);
1067 return 1;
1068 }
1069
1070 err = get_next_channel_event(td, rd->cm_channel, RDMA_CM_EVENT_ADDR_RESOLVED);
1071 if (err != 0) {
1072 log_err("fio: get_next_channel_event: %d\n", err);
1073 return 1;
1074 }
1075
1076 /* resolve route */
1077 err = rdma_resolve_route(rd->cm_id, 2000);
1078 if (err != 0) {
1079 log_err("fio: rdma_resolve_route: %d\n", err);
1080 return 1;
1081 }
1082
1083 err = get_next_channel_event(td, rd->cm_channel, RDMA_CM_EVENT_ROUTE_RESOLVED);
1084 if (err != 0) {
1085 log_err("fio: get_next_channel_event: %d\n", err);
1086 return 1;
1087 }
1088
1089 /* create qp and buffer */
1090 if (fio_rdmaio_setup_qp(td) != 0)
1091 return 1;
1092
1093 if (fio_rdmaio_setup_control_msg_buffers(td) != 0)
1094 return 1;
1095
1096 /* post recv buf */
1097 err = ibv_post_recv(rd->qp, &rd->rq_wr, &bad_wr);
1098 if (err != 0) {
1099 log_err("fio: ibv_post_recv fail: %d\n", err);
1100 return 1;
1101 }
1102
1103 return 0;
1104}
1105
1106static int fio_rdmaio_setup_listen(struct thread_data *td, short port)
1107{
1108 struct rdmaio_data *rd = td->io_ops_data;
1109 struct rdmaio_options *o = td->eo;
1110 struct ibv_recv_wr *bad_wr;
1111 int state = td->runstate;
1112
1113 td_set_runstate(td, TD_SETTING_UP);
1114
1115 rd->addr.sin_family = AF_INET;
1116 rd->addr.sin_port = htons(port);
1117
1118 if (strcmp(o->bindname, "") == 0)
1119 rd->addr.sin_addr.s_addr = htonl(INADDR_ANY);
1120 else
1121 rd->addr.sin_addr.s_addr = htonl(*o->bindname);
1122
1123 /* rdma_listen */
1124 if (rdma_bind_addr(rd->cm_id, (struct sockaddr *)&rd->addr) != 0) {
1125 log_err("fio: rdma_bind_addr fail: %m\n");
1126 return 1;
1127 }
1128
1129 if (rdma_listen(rd->cm_id, 3) != 0) {
1130 log_err("fio: rdma_listen fail: %m\n");
1131 return 1;
1132 }
1133
1134 log_info("fio: waiting for connection\n");
1135
1136 /* wait for CONNECT_REQUEST */
1137 if (get_next_channel_event
1138 (td, rd->cm_channel, RDMA_CM_EVENT_CONNECT_REQUEST) != 0) {
1139 log_err("fio: wait for RDMA_CM_EVENT_CONNECT_REQUEST\n");
1140 return 1;
1141 }
1142
1143 if (fio_rdmaio_setup_qp(td) != 0)
1144 return 1;
1145
1146 if (fio_rdmaio_setup_control_msg_buffers(td) != 0)
1147 return 1;
1148
1149 /* post recv buf */
1150 if (ibv_post_recv(rd->qp, &rd->rq_wr, &bad_wr) != 0) {
1151 log_err("fio: ibv_post_recv fail: %m\n");
1152 return 1;
1153 }
1154
1155 td_set_runstate(td, state);
1156 return 0;
1157}
1158
1159static int check_set_rlimits(struct thread_data *td)
1160{
1161#ifdef CONFIG_RLIMIT_MEMLOCK
1162 struct rlimit rl;
1163
1164 /* check RLIMIT_MEMLOCK */
1165 if (getrlimit(RLIMIT_MEMLOCK, &rl) != 0) {
1166 log_err("fio: getrlimit fail: %d(%s)\n",
1167 errno, strerror(errno));
1168 return 1;
1169 }
1170
1171 /* soft limit */
1172 if ((rl.rlim_cur != RLIM_INFINITY)
1173 && (rl.rlim_cur < td->orig_buffer_size)) {
1174 log_err("fio: soft RLIMIT_MEMLOCK is: %" PRId64 "\n",
1175 rl.rlim_cur);
1176 log_err("fio: total block size is: %zd\n",
1177 td->orig_buffer_size);
1178 /* try to set larger RLIMIT_MEMLOCK */
1179 rl.rlim_cur = rl.rlim_max;
1180 if (setrlimit(RLIMIT_MEMLOCK, &rl) != 0) {
1181 log_err("fio: setrlimit fail: %d(%s)\n",
1182 errno, strerror(errno));
1183 log_err("fio: you may try enlarge MEMLOCK by root\n");
1184 log_err("# ulimit -l unlimited\n");
1185 return 1;
1186 }
1187 }
1188#endif
1189
1190 return 0;
1191}
1192
1193static int compat_options(struct thread_data *td)
1194{
1195 // The original RDMA engine had an ugly / seperator
1196 // on the filename for it's options. This function
1197 // retains backwards compatibility with it. Note we do not
1198 // support setting the bindname option is this legacy mode.
1199
1200 struct rdmaio_options *o = td->eo;
1201 char *modep, *portp;
1202 char *filename = td->o.filename;
1203
1204 if (!filename)
1205 return 0;
1206
1207 portp = strchr(filename, '/');
1208 if (portp == NULL)
1209 return 0;
1210
1211 *portp = '\0';
1212 portp++;
1213
1214 o->port = strtol(portp, NULL, 10);
1215 if (!o->port || o->port > 65535)
1216 goto bad_host;
1217
1218 modep = strchr(portp, '/');
1219 if (modep != NULL) {
1220 *modep = '\0';
1221 modep++;
1222 }
1223
1224 if (modep) {
1225 if (!strncmp("rdma_write", modep, strlen(modep)) ||
1226 !strncmp("RDMA_WRITE", modep, strlen(modep)))
1227 o->verb = FIO_RDMA_MEM_WRITE;
1228 else if (!strncmp("rdma_read", modep, strlen(modep)) ||
1229 !strncmp("RDMA_READ", modep, strlen(modep)))
1230 o->verb = FIO_RDMA_MEM_READ;
1231 else if (!strncmp("send", modep, strlen(modep)) ||
1232 !strncmp("SEND", modep, strlen(modep)))
1233 o->verb = FIO_RDMA_CHA_SEND;
1234 else
1235 goto bad_host;
1236 } else
1237 o->verb = FIO_RDMA_MEM_WRITE;
1238
1239
1240 return 0;
1241
1242bad_host:
1243 log_err("fio: bad rdma host/port/protocol: %s\n", td->o.filename);
1244 return 1;
1245}
1246
1247static int fio_rdmaio_init(struct thread_data *td)
1248{
1249 struct rdmaio_data *rd = td->io_ops_data;
1250 struct rdmaio_options *o = td->eo;
1251 unsigned int max_bs;
1252 int ret, i;
1253
1254 if (td_rw(td)) {
1255 log_err("fio: rdma connections must be read OR write\n");
1256 return 1;
1257 }
1258 if (td_random(td)) {
1259 log_err("fio: RDMA network IO can't be random\n");
1260 return 1;
1261 }
1262
1263 if (compat_options(td))
1264 return 1;
1265
1266 if (!o->port) {
1267 log_err("fio: no port has been specified which is required "
1268 "for the rdma engine\n");
1269 return 1;
1270 }
1271
1272 if (check_set_rlimits(td))
1273 return 1;
1274
1275 rd->rdma_protocol = o->verb;
1276 rd->cq_event_num = 0;
1277
1278 rd->cm_channel = rdma_create_event_channel();
1279 if (!rd->cm_channel) {
1280 log_err("fio: rdma_create_event_channel fail: %m\n");
1281 return 1;
1282 }
1283
1284 ret = rdma_create_id(rd->cm_channel, &rd->cm_id, rd, RDMA_PS_TCP);
1285 if (ret) {
1286 log_err("fio: rdma_create_id fail: %m\n");
1287 return 1;
1288 }
1289
1290 if ((rd->rdma_protocol == FIO_RDMA_MEM_WRITE) ||
1291 (rd->rdma_protocol == FIO_RDMA_MEM_READ)) {
1292 rd->rmt_us =
1293 malloc(FIO_RDMA_MAX_IO_DEPTH * sizeof(struct remote_u));
1294 memset(rd->rmt_us, 0,
1295 FIO_RDMA_MAX_IO_DEPTH * sizeof(struct remote_u));
1296 rd->rmt_nr = 0;
1297 }
1298
1299 rd->io_us_queued = malloc(td->o.iodepth * sizeof(struct io_u *));
1300 memset(rd->io_us_queued, 0, td->o.iodepth * sizeof(struct io_u *));
1301 rd->io_u_queued_nr = 0;
1302
1303 rd->io_us_flight = malloc(td->o.iodepth * sizeof(struct io_u *));
1304 memset(rd->io_us_flight, 0, td->o.iodepth * sizeof(struct io_u *));
1305 rd->io_u_flight_nr = 0;
1306
1307 rd->io_us_completed = malloc(td->o.iodepth * sizeof(struct io_u *));
1308 memset(rd->io_us_completed, 0, td->o.iodepth * sizeof(struct io_u *));
1309 rd->io_u_completed_nr = 0;
1310
1311 if (td_read(td)) { /* READ as the server */
1312 rd->is_client = 0;
1313 td->flags |= TD_F_NO_PROGRESS;
1314 /* server rd->rdma_buf_len will be setup after got request */
1315 ret = fio_rdmaio_setup_listen(td, o->port);
1316 } else { /* WRITE as the client */
1317 rd->is_client = 1;
1318 ret = fio_rdmaio_setup_connect(td, td->o.filename, o->port);
1319 }
1320
1321 max_bs = max(td->o.max_bs[DDIR_READ], td->o.max_bs[DDIR_WRITE]);
1322 rd->send_buf.max_bs = htonl(max_bs);
1323
1324 /* register each io_u in the free list */
1325 for (i = 0; i < td->io_u_freelist.nr; i++) {
1326 struct io_u *io_u = td->io_u_freelist.io_us[i];
1327
1328 io_u->engine_data = malloc(sizeof(struct rdma_io_u_data));
1329 memset(io_u->engine_data, 0, sizeof(struct rdma_io_u_data));
1330 ((struct rdma_io_u_data *)io_u->engine_data)->wr_id = i;
1331
1332 io_u->mr = ibv_reg_mr(rd->pd, io_u->buf, max_bs,
1333 IBV_ACCESS_LOCAL_WRITE |
1334 IBV_ACCESS_REMOTE_READ |
1335 IBV_ACCESS_REMOTE_WRITE);
1336 if (io_u->mr == NULL) {
1337 log_err("fio: ibv_reg_mr io_u failed: %m\n");
1338 return 1;
1339 }
1340
1341 rd->send_buf.rmt_us[i].buf =
1342 cpu_to_be64((uint64_t) (unsigned long)io_u->buf);
1343 rd->send_buf.rmt_us[i].rkey = htonl(io_u->mr->rkey);
1344 rd->send_buf.rmt_us[i].size = htonl(max_bs);
1345
1346#if 0
1347 log_info("fio: Send rkey %x addr %" PRIx64 " len %d to client\n", io_u->mr->rkey, io_u->buf, max_bs); */
1348#endif
1349 }
1350
1351 rd->send_buf.nr = htonl(i);
1352
1353 return ret;
1354}
1355
1356static void fio_rdmaio_cleanup(struct thread_data *td)
1357{
1358 struct rdmaio_data *rd = td->io_ops_data;
1359
1360 if (rd)
1361 free(rd);
1362}
1363
1364static int fio_rdmaio_setup(struct thread_data *td)
1365{
1366 struct rdmaio_data *rd;
1367
1368 if (!td->files_index) {
1369 add_file(td, td->o.filename ?: "rdma", 0, 0);
1370 td->o.nr_files = td->o.nr_files ?: 1;
1371 td->o.open_files++;
1372 }
1373
1374 if (!td->io_ops_data) {
1375 rd = malloc(sizeof(*rd));
1376
1377 memset(rd, 0, sizeof(*rd));
1378 init_rand_seed(&rd->rand_state, (unsigned int) GOLDEN_RATIO_PRIME, 0);
1379 td->io_ops_data = rd;
1380 }
1381
1382 return 0;
1383}
1384
1385static struct ioengine_ops ioengine_rw = {
1386 .name = "rdma",
1387 .version = FIO_IOOPS_VERSION,
1388 .setup = fio_rdmaio_setup,
1389 .init = fio_rdmaio_init,
1390 .prep = fio_rdmaio_prep,
1391 .queue = fio_rdmaio_queue,
1392 .commit = fio_rdmaio_commit,
1393 .getevents = fio_rdmaio_getevents,
1394 .event = fio_rdmaio_event,
1395 .cleanup = fio_rdmaio_cleanup,
1396 .open_file = fio_rdmaio_open_file,
1397 .close_file = fio_rdmaio_close_file,
1398 .flags = FIO_DISKLESSIO | FIO_UNIDIR | FIO_PIPEIO,
1399 .options = options,
1400 .option_struct_size = sizeof(struct rdmaio_options),
1401};
1402
1403static void fio_init fio_rdmaio_register(void)
1404{
1405 register_ioengine(&ioengine_rw);
1406}
1407
1408static void fio_exit fio_rdmaio_unregister(void)
1409{
1410 unregister_ioengine(&ioengine_rw);
1411}
fio - Flexible IO Tester