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#include <infiniband/arch.h>
48
49#define FIO_RDMA_MAX_IO_DEPTH    512
50
51enum rdma_io_mode {
52	FIO_RDMA_UNKNOWN = 0,
53	FIO_RDMA_MEM_WRITE,
54	FIO_RDMA_MEM_READ,
55	FIO_RDMA_CHA_SEND,
56	FIO_RDMA_CHA_RECV
57};
58
59struct rdmaio_options {
60	struct thread_data *td;
61	unsigned int port;
62	enum rdma_io_mode verb;
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	= "port",
87		.lname	= "rdma engine port",
88		.type	= FIO_OPT_INT,
89		.off1	= offsetof(struct rdmaio_options, port),
90		.minval	= 1,
91		.maxval	= 65535,
92		.help	= "Port to use for RDMA connections",
93		.category = FIO_OPT_C_ENGINE,
94		.group	= FIO_OPT_G_RDMA,
95	},
96	{
97		.name	= "verb",
98		.lname	= "RDMA engine verb",
99		.alias	= "proto",
100		.type	= FIO_OPT_STR,
101		.off1	= offsetof(struct rdmaio_options, verb),
102		.help	= "RDMA engine verb",
103		.def	= "write",
104		.posval = {
105			  { .ival = "write",
106			    .oval = FIO_RDMA_MEM_WRITE,
107			    .help = "Memory Write",
108			  },
109			  { .ival = "read",
110			    .oval = FIO_RDMA_MEM_READ,
111			    .help = "Memory Read",
112			  },
113			  { .ival = "send",
114			    .oval = FIO_RDMA_CHA_SEND,
115			    .help = "Posted Send",
116			  },
117			  { .ival = "recv",
118			    .oval = FIO_RDMA_CHA_RECV,
119			    .help = "Posted Receive",
120			  },
121		},
122		.category = FIO_OPT_C_ENGINE,
123		.group	= FIO_OPT_G_RDMA,
124	},
125	{
126		.name	= NULL,
127	},
128};
129
130struct remote_u {
131	uint64_t buf;
132	uint32_t rkey;
133	uint32_t size;
134};
135
136struct rdma_info_blk {
137	uint32_t mode;		/* channel semantic or memory semantic */
138	uint32_t nr;		/* client: io depth
139				   server: number of records for memory semantic
140				 */
141	uint32_t max_bs;        /* maximum block size */
142	struct remote_u rmt_us[FIO_RDMA_MAX_IO_DEPTH];
143};
144
145struct rdma_io_u_data {
146	uint64_t wr_id;
147	struct ibv_send_wr sq_wr;
148	struct ibv_recv_wr rq_wr;
149	struct ibv_sge rdma_sgl;
150};
151
152struct rdmaio_data {
153	int is_client;
154	enum rdma_io_mode rdma_protocol;
155	char host[64];
156	struct sockaddr_in addr;
157
158	struct ibv_recv_wr rq_wr;
159	struct ibv_sge recv_sgl;
160	struct rdma_info_blk recv_buf;
161	struct ibv_mr *recv_mr;
162
163	struct ibv_send_wr sq_wr;
164	struct ibv_sge send_sgl;
165	struct rdma_info_blk send_buf;
166	struct ibv_mr *send_mr;
167
168	struct ibv_comp_channel *channel;
169	struct ibv_cq *cq;
170	struct ibv_pd *pd;
171	struct ibv_qp *qp;
172
173	pthread_t cmthread;
174	struct rdma_event_channel *cm_channel;
175	struct rdma_cm_id *cm_id;
176	struct rdma_cm_id *child_cm_id;
177
178	int cq_event_num;
179
180	struct remote_u *rmt_us;
181	int rmt_nr;
182	struct io_u **io_us_queued;
183	int io_u_queued_nr;
184	struct io_u **io_us_flight;
185	int io_u_flight_nr;
186	struct io_u **io_us_completed;
187	int io_u_completed_nr;
188
189	struct frand_state rand_state;
190};
191
192static int client_recv(struct thread_data *td, struct ibv_wc *wc)
193{
194	struct rdmaio_data *rd = td->io_ops_data;
195	unsigned int max_bs;
196
197	if (wc->byte_len != sizeof(rd->recv_buf)) {
198		log_err("Received bogus data, size %d\n", wc->byte_len);
199		return 1;
200	}
201
202	max_bs = max(td->o.max_bs[DDIR_READ], td->o.max_bs[DDIR_WRITE]);
203	if (max_bs > ntohl(rd->recv_buf.max_bs)) {
204		log_err("fio: Server's block size (%d) must be greater than or "
205			"equal to the client's block size (%d)!\n",
206			ntohl(rd->recv_buf.max_bs), max_bs);
207		return 1;
208	}
209
210	/* store mr info for MEMORY semantic */
211	if ((rd->rdma_protocol == FIO_RDMA_MEM_WRITE) ||
212	    (rd->rdma_protocol == FIO_RDMA_MEM_READ)) {
213		/* struct flist_head *entry; */
214		int i = 0;
215
216		rd->rmt_nr = ntohl(rd->recv_buf.nr);
217
218		for (i = 0; i < rd->rmt_nr; i++) {
219			rd->rmt_us[i].buf = ntohll(rd->recv_buf.rmt_us[i].buf);
220			rd->rmt_us[i].rkey = ntohl(rd->recv_buf.rmt_us[i].rkey);
221			rd->rmt_us[i].size = ntohl(rd->recv_buf.rmt_us[i].size);
222
223			dprint(FD_IO,
224			       "fio: Received rkey %x addr %" PRIx64
225			       " len %d from peer\n", rd->rmt_us[i].rkey,
226			       rd->rmt_us[i].buf, rd->rmt_us[i].size);
227		}
228	}
229
230	return 0;
231}
232
233static int server_recv(struct thread_data *td, struct ibv_wc *wc)
234{
235	struct rdmaio_data *rd = td->io_ops_data;
236	unsigned int max_bs;
237
238	if (wc->wr_id == FIO_RDMA_MAX_IO_DEPTH) {
239		rd->rdma_protocol = ntohl(rd->recv_buf.mode);
240
241		/* CHANNEL semantic, do nothing */
242		if (rd->rdma_protocol == FIO_RDMA_CHA_SEND)
243			rd->rdma_protocol = FIO_RDMA_CHA_RECV;
244
245		max_bs = max(td->o.max_bs[DDIR_READ], td->o.max_bs[DDIR_WRITE]);
246		if (max_bs < ntohl(rd->recv_buf.max_bs)) {
247			log_err("fio: Server's block size (%d) must be greater than or "
248				"equal to the client's block size (%d)!\n",
249				ntohl(rd->recv_buf.max_bs), max_bs);
250			return 1;
251		}
252
253	}
254
255	return 0;
256}
257
258static int cq_event_handler(struct thread_data *td, enum ibv_wc_opcode opcode)
259{
260	struct rdmaio_data *rd = td->io_ops_data;
261	struct ibv_wc wc;
262	struct rdma_io_u_data *r_io_u_d;
263	int ret;
264	int compevnum = 0;
265	int i;
266
267	while ((ret = ibv_poll_cq(rd->cq, 1, &wc)) == 1) {
268		ret = 0;
269		compevnum++;
270
271		if (wc.status) {
272			log_err("fio: cq completion status %d(%s)\n",
273				wc.status, ibv_wc_status_str(wc.status));
274			return -1;
275		}
276
277		switch (wc.opcode) {
278
279		case IBV_WC_RECV:
280			if (rd->is_client == 1)
281				ret = client_recv(td, &wc);
282			else
283				ret = server_recv(td, &wc);
284
285			if (ret)
286				return -1;
287
288			if (wc.wr_id == FIO_RDMA_MAX_IO_DEPTH)
289				break;
290
291			for (i = 0; i < rd->io_u_flight_nr; i++) {
292				r_io_u_d = rd->io_us_flight[i]->engine_data;
293
294				if (wc.wr_id == r_io_u_d->rq_wr.wr_id) {
295					rd->io_us_flight[i]->resid =
296					    rd->io_us_flight[i]->buflen
297					    - wc.byte_len;
298
299					rd->io_us_flight[i]->error = 0;
300
301					rd->io_us_completed[rd->
302							    io_u_completed_nr]
303					    = rd->io_us_flight[i];
304					rd->io_u_completed_nr++;
305					break;
306				}
307			}
308			if (i == rd->io_u_flight_nr)
309				log_err("fio: recv wr %" PRId64 " not found\n",
310					wc.wr_id);
311			else {
312				/* put the last one into middle of the list */
313				rd->io_us_flight[i] =
314				    rd->io_us_flight[rd->io_u_flight_nr - 1];
315				rd->io_u_flight_nr--;
316			}
317
318			break;
319
320		case IBV_WC_SEND:
321		case IBV_WC_RDMA_WRITE:
322		case IBV_WC_RDMA_READ:
323			if (wc.wr_id == FIO_RDMA_MAX_IO_DEPTH)
324				break;
325
326			for (i = 0; i < rd->io_u_flight_nr; i++) {
327				r_io_u_d = rd->io_us_flight[i]->engine_data;
328
329				if (wc.wr_id == r_io_u_d->sq_wr.wr_id) {
330					rd->io_us_completed[rd->
331							    io_u_completed_nr]
332					    = rd->io_us_flight[i];
333					rd->io_u_completed_nr++;
334					break;
335				}
336			}
337			if (i == rd->io_u_flight_nr)
338				log_err("fio: send wr %" PRId64 " not found\n",
339					wc.wr_id);
340			else {
341				/* put the last one into middle of the list */
342				rd->io_us_flight[i] =
343				    rd->io_us_flight[rd->io_u_flight_nr - 1];
344				rd->io_u_flight_nr--;
345			}
346
347			break;
348
349		default:
350			log_info("fio: unknown completion event %d\n",
351				 wc.opcode);
352			return -1;
353		}
354		rd->cq_event_num++;
355	}
356
357	if (ret) {
358		log_err("fio: poll error %d\n", ret);
359		return 1;
360	}
361
362	return compevnum;
363}
364
365/*
366 * Return -1 for error and 'nr events' for a positive number
367 * of events
368 */
369static int rdma_poll_wait(struct thread_data *td, enum ibv_wc_opcode opcode)
370{
371	struct rdmaio_data *rd = td->io_ops_data;
372	struct ibv_cq *ev_cq;
373	void *ev_ctx;
374	int ret;
375
376	if (rd->cq_event_num > 0) {	/* previous left */
377		rd->cq_event_num--;
378		return 0;
379	}
380
381again:
382	if (ibv_get_cq_event(rd->channel, &ev_cq, &ev_ctx) != 0) {
383		log_err("fio: Failed to get cq event!\n");
384		return -1;
385	}
386	if (ev_cq != rd->cq) {
387		log_err("fio: Unknown CQ!\n");
388		return -1;
389	}
390	if (ibv_req_notify_cq(rd->cq, 0) != 0) {
391		log_err("fio: Failed to set notify!\n");
392		return -1;
393	}
394
395	ret = cq_event_handler(td, opcode);
396	if (ret == 0)
397		goto again;
398
399	ibv_ack_cq_events(rd->cq, ret);
400
401	rd->cq_event_num--;
402
403	return ret;
404}
405
406static int fio_rdmaio_setup_qp(struct thread_data *td)
407{
408	struct rdmaio_data *rd = td->io_ops_data;
409	struct ibv_qp_init_attr init_attr;
410	int qp_depth = td->o.iodepth * 2;	/* 2 times of io depth */
411
412	if (rd->is_client == 0)
413		rd->pd = ibv_alloc_pd(rd->child_cm_id->verbs);
414	else
415		rd->pd = ibv_alloc_pd(rd->cm_id->verbs);
416
417	if (rd->pd == NULL) {
418		log_err("fio: ibv_alloc_pd fail: %m\n");
419		return 1;
420	}
421
422	if (rd->is_client == 0)
423		rd->channel = ibv_create_comp_channel(rd->child_cm_id->verbs);
424	else
425		rd->channel = ibv_create_comp_channel(rd->cm_id->verbs);
426	if (rd->channel == NULL) {
427		log_err("fio: ibv_create_comp_channel fail: %m\n");
428		goto err1;
429	}
430
431	if (qp_depth < 16)
432		qp_depth = 16;
433
434	if (rd->is_client == 0)
435		rd->cq = ibv_create_cq(rd->child_cm_id->verbs,
436				       qp_depth, rd, rd->channel, 0);
437	else
438		rd->cq = ibv_create_cq(rd->cm_id->verbs,
439				       qp_depth, rd, rd->channel, 0);
440	if (rd->cq == NULL) {
441		log_err("fio: ibv_create_cq failed: %m\n");
442		goto err2;
443	}
444
445	if (ibv_req_notify_cq(rd->cq, 0) != 0) {
446		log_err("fio: ibv_req_notify_cq failed: %m\n");
447		goto err3;
448	}
449
450	/* create queue pair */
451	memset(&init_attr, 0, sizeof(init_attr));
452	init_attr.cap.max_send_wr = qp_depth;
453	init_attr.cap.max_recv_wr = qp_depth;
454	init_attr.cap.max_recv_sge = 1;
455	init_attr.cap.max_send_sge = 1;
456	init_attr.qp_type = IBV_QPT_RC;
457	init_attr.send_cq = rd->cq;
458	init_attr.recv_cq = rd->cq;
459
460	if (rd->is_client == 0) {
461		if (rdma_create_qp(rd->child_cm_id, rd->pd, &init_attr) != 0) {
462			log_err("fio: rdma_create_qp failed: %m\n");
463			goto err3;
464		}
465		rd->qp = rd->child_cm_id->qp;
466	} else {
467		if (rdma_create_qp(rd->cm_id, rd->pd, &init_attr) != 0) {
468			log_err("fio: rdma_create_qp failed: %m\n");
469			goto err3;
470		}
471		rd->qp = rd->cm_id->qp;
472	}
473
474	return 0;
475
476err3:
477	ibv_destroy_cq(rd->cq);
478err2:
479	ibv_destroy_comp_channel(rd->channel);
480err1:
481	ibv_dealloc_pd(rd->pd);
482
483	return 1;
484}
485
486static int fio_rdmaio_setup_control_msg_buffers(struct thread_data *td)
487{
488	struct rdmaio_data *rd = td->io_ops_data;
489
490	rd->recv_mr = ibv_reg_mr(rd->pd, &rd->recv_buf, sizeof(rd->recv_buf),
491				 IBV_ACCESS_LOCAL_WRITE);
492	if (rd->recv_mr == NULL) {
493		log_err("fio: recv_buf reg_mr failed: %m\n");
494		return 1;
495	}
496
497	rd->send_mr = ibv_reg_mr(rd->pd, &rd->send_buf, sizeof(rd->send_buf),
498				 0);
499	if (rd->send_mr == NULL) {
500		log_err("fio: send_buf reg_mr failed: %m\n");
501		ibv_dereg_mr(rd->recv_mr);
502		return 1;
503	}
504
505	/* setup work request */
506	/* recv wq */
507	rd->recv_sgl.addr = (uint64_t) (unsigned long)&rd->recv_buf;
508	rd->recv_sgl.length = sizeof(rd->recv_buf);
509	rd->recv_sgl.lkey = rd->recv_mr->lkey;
510	rd->rq_wr.sg_list = &rd->recv_sgl;
511	rd->rq_wr.num_sge = 1;
512	rd->rq_wr.wr_id = FIO_RDMA_MAX_IO_DEPTH;
513
514	/* send wq */
515	rd->send_sgl.addr = (uint64_t) (unsigned long)&rd->send_buf;
516	rd->send_sgl.length = sizeof(rd->send_buf);
517	rd->send_sgl.lkey = rd->send_mr->lkey;
518
519	rd->sq_wr.opcode = IBV_WR_SEND;
520	rd->sq_wr.send_flags = IBV_SEND_SIGNALED;
521	rd->sq_wr.sg_list = &rd->send_sgl;
522	rd->sq_wr.num_sge = 1;
523	rd->sq_wr.wr_id = FIO_RDMA_MAX_IO_DEPTH;
524
525	return 0;
526}
527
528static int get_next_channel_event(struct thread_data *td,
529				  struct rdma_event_channel *channel,
530				  enum rdma_cm_event_type wait_event)
531{
532	struct rdmaio_data *rd = td->io_ops_data;
533	struct rdma_cm_event *event;
534	int ret;
535
536	ret = rdma_get_cm_event(channel, &event);
537	if (ret) {
538		log_err("fio: rdma_get_cm_event: %d\n", ret);
539		return 1;
540	}
541
542	if (event->event != wait_event) {
543		log_err("fio: event is %s instead of %s\n",
544			rdma_event_str(event->event),
545			rdma_event_str(wait_event));
546		return 1;
547	}
548
549	switch (event->event) {
550	case RDMA_CM_EVENT_CONNECT_REQUEST:
551		rd->child_cm_id = event->id;
552		break;
553	default:
554		break;
555	}
556
557	rdma_ack_cm_event(event);
558
559	return 0;
560}
561
562static int fio_rdmaio_prep(struct thread_data *td, struct io_u *io_u)
563{
564	struct rdmaio_data *rd = td->io_ops_data;
565	struct rdma_io_u_data *r_io_u_d;
566
567	r_io_u_d = io_u->engine_data;
568
569	switch (rd->rdma_protocol) {
570	case FIO_RDMA_MEM_WRITE:
571	case FIO_RDMA_MEM_READ:
572		r_io_u_d->rdma_sgl.addr = (uint64_t) (unsigned long)io_u->buf;
573		r_io_u_d->rdma_sgl.lkey = io_u->mr->lkey;
574		r_io_u_d->sq_wr.wr_id = r_io_u_d->wr_id;
575		r_io_u_d->sq_wr.send_flags = IBV_SEND_SIGNALED;
576		r_io_u_d->sq_wr.sg_list = &r_io_u_d->rdma_sgl;
577		r_io_u_d->sq_wr.num_sge = 1;
578		break;
579	case FIO_RDMA_CHA_SEND:
580		r_io_u_d->rdma_sgl.addr = (uint64_t) (unsigned long)io_u->buf;
581		r_io_u_d->rdma_sgl.lkey = io_u->mr->lkey;
582		r_io_u_d->rdma_sgl.length = io_u->buflen;
583		r_io_u_d->sq_wr.wr_id = r_io_u_d->wr_id;
584		r_io_u_d->sq_wr.opcode = IBV_WR_SEND;
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_RECV:
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->rq_wr.wr_id = r_io_u_d->wr_id;
594		r_io_u_d->rq_wr.sg_list = &r_io_u_d->rdma_sgl;
595		r_io_u_d->rq_wr.num_sge = 1;
596		break;
597	default:
598		log_err("fio: unknown rdma protocol - %d\n", rd->rdma_protocol);
599		break;
600	}
601
602	return 0;
603}
604
605static struct io_u *fio_rdmaio_event(struct thread_data *td, int event)
606{
607	struct rdmaio_data *rd = td->io_ops_data;
608	struct io_u *io_u;
609	int i;
610
611	io_u = rd->io_us_completed[0];
612	for (i = 0; i < rd->io_u_completed_nr - 1; i++)
613		rd->io_us_completed[i] = rd->io_us_completed[i + 1];
614
615	rd->io_u_completed_nr--;
616
617	dprint_io_u(io_u, "fio_rdmaio_event");
618
619	return io_u;
620}
621
622static int fio_rdmaio_getevents(struct thread_data *td, unsigned int min,
623				unsigned int max, const struct timespec *t)
624{
625	struct rdmaio_data *rd = td->io_ops_data;
626	enum ibv_wc_opcode comp_opcode;
627	struct ibv_cq *ev_cq;
628	void *ev_ctx;
629	int ret, r = 0;
630	comp_opcode = IBV_WC_RDMA_WRITE;
631
632	switch (rd->rdma_protocol) {
633	case FIO_RDMA_MEM_WRITE:
634		comp_opcode = IBV_WC_RDMA_WRITE;
635		break;
636	case FIO_RDMA_MEM_READ:
637		comp_opcode = IBV_WC_RDMA_READ;
638		break;
639	case FIO_RDMA_CHA_SEND:
640		comp_opcode = IBV_WC_SEND;
641		break;
642	case FIO_RDMA_CHA_RECV:
643		comp_opcode = IBV_WC_RECV;
644		break;
645	default:
646		log_err("fio: unknown rdma protocol - %d\n", rd->rdma_protocol);
647		break;
648	}
649
650	if (rd->cq_event_num > 0) {	/* previous left */
651		rd->cq_event_num--;
652		return 0;
653	}
654
655again:
656	if (ibv_get_cq_event(rd->channel, &ev_cq, &ev_ctx) != 0) {
657		log_err("fio: Failed to get cq event!\n");
658		return -1;
659	}
660	if (ev_cq != rd->cq) {
661		log_err("fio: Unknown CQ!\n");
662		return -1;
663	}
664	if (ibv_req_notify_cq(rd->cq, 0) != 0) {
665		log_err("fio: Failed to set notify!\n");
666		return -1;
667	}
668
669	ret = cq_event_handler(td, comp_opcode);
670	if (ret < 1)
671		goto again;
672
673	ibv_ack_cq_events(rd->cq, ret);
674
675	r += ret;
676	if (r < min)
677		goto again;
678
679	rd->cq_event_num -= r;
680
681	return r;
682}
683
684static int fio_rdmaio_send(struct thread_data *td, struct io_u **io_us,
685			   unsigned int nr)
686{
687	struct rdmaio_data *rd = td->io_ops_data;
688	struct ibv_send_wr *bad_wr;
689#if 0
690	enum ibv_wc_opcode comp_opcode;
691	comp_opcode = IBV_WC_RDMA_WRITE;
692#endif
693	int i;
694	long index;
695	struct rdma_io_u_data *r_io_u_d;
696
697	r_io_u_d = NULL;
698
699	for (i = 0; i < nr; i++) {
700		/* RDMA_WRITE or RDMA_READ */
701		switch (rd->rdma_protocol) {
702		case FIO_RDMA_MEM_WRITE:
703			/* compose work request */
704			r_io_u_d = io_us[i]->engine_data;
705			index = __rand(&rd->rand_state) % rd->rmt_nr;
706			r_io_u_d->sq_wr.opcode = IBV_WR_RDMA_WRITE;
707			r_io_u_d->sq_wr.wr.rdma.rkey = rd->rmt_us[index].rkey;
708			r_io_u_d->sq_wr.wr.rdma.remote_addr = \
709				rd->rmt_us[index].buf;
710			r_io_u_d->sq_wr.sg_list->length = io_us[i]->buflen;
711			break;
712		case FIO_RDMA_MEM_READ:
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_READ;
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_CHA_SEND:
723			r_io_u_d = io_us[i]->engine_data;
724			r_io_u_d->sq_wr.opcode = IBV_WR_SEND;
725			r_io_u_d->sq_wr.send_flags = IBV_SEND_SIGNALED;
726			break;
727		default:
728			log_err("fio: unknown rdma protocol - %d\n",
729				rd->rdma_protocol);
730			break;
731		}
732
733		if (ibv_post_send(rd->qp, &r_io_u_d->sq_wr, &bad_wr) != 0) {
734			log_err("fio: ibv_post_send fail: %m\n");
735			return -1;
736		}
737
738		dprint_io_u(io_us[i], "fio_rdmaio_send");
739	}
740
741	/* wait for completion
742	   rdma_poll_wait(td, comp_opcode); */
743
744	return i;
745}
746
747static int fio_rdmaio_recv(struct thread_data *td, struct io_u **io_us,
748			   unsigned int nr)
749{
750	struct rdmaio_data *rd = td->io_ops_data;
751	struct ibv_recv_wr *bad_wr;
752	struct rdma_io_u_data *r_io_u_d;
753	int i;
754
755	i = 0;
756	if (rd->rdma_protocol == FIO_RDMA_CHA_RECV) {
757		/* post io_u into recv queue */
758		for (i = 0; i < nr; i++) {
759			r_io_u_d = io_us[i]->engine_data;
760			if (ibv_post_recv(rd->qp, &r_io_u_d->rq_wr, &bad_wr) !=
761			    0) {
762				log_err("fio: ibv_post_recv fail: %m\n");
763				return 1;
764			}
765		}
766	} else if ((rd->rdma_protocol == FIO_RDMA_MEM_READ)
767		   || (rd->rdma_protocol == FIO_RDMA_MEM_WRITE)) {
768		/* re-post the rq_wr */
769		if (ibv_post_recv(rd->qp, &rd->rq_wr, &bad_wr) != 0) {
770			log_err("fio: ibv_post_recv fail: %m\n");
771			return 1;
772		}
773
774		rdma_poll_wait(td, IBV_WC_RECV);
775
776		dprint(FD_IO, "fio: recv FINISH message\n");
777		td->done = 1;
778		return 0;
779	}
780
781	return i;
782}
783
784static int fio_rdmaio_queue(struct thread_data *td, struct io_u *io_u)
785{
786	struct rdmaio_data *rd = td->io_ops_data;
787
788	fio_ro_check(td, io_u);
789
790	if (rd->io_u_queued_nr == (int)td->o.iodepth)
791		return FIO_Q_BUSY;
792
793	rd->io_us_queued[rd->io_u_queued_nr] = io_u;
794	rd->io_u_queued_nr++;
795
796	dprint_io_u(io_u, "fio_rdmaio_queue");
797
798	return FIO_Q_QUEUED;
799}
800
801static void fio_rdmaio_queued(struct thread_data *td, struct io_u **io_us,
802			      unsigned int nr)
803{
804	struct rdmaio_data *rd = td->io_ops_data;
805	struct timeval now;
806	unsigned int i;
807
808	if (!fio_fill_issue_time(td))
809		return;
810
811	fio_gettime(&now, NULL);
812
813	for (i = 0; i < nr; i++) {
814		struct io_u *io_u = io_us[i];
815
816		/* queued -> flight */
817		rd->io_us_flight[rd->io_u_flight_nr] = io_u;
818		rd->io_u_flight_nr++;
819
820		memcpy(&io_u->issue_time, &now, sizeof(now));
821		io_u_queued(td, io_u);
822	}
823}
824
825static int fio_rdmaio_commit(struct thread_data *td)
826{
827	struct rdmaio_data *rd = td->io_ops_data;
828	struct io_u **io_us;
829	int ret;
830
831	if (!rd->io_us_queued)
832		return 0;
833
834	io_us = rd->io_us_queued;
835	do {
836		/* RDMA_WRITE or RDMA_READ */
837		if (rd->is_client)
838			ret = fio_rdmaio_send(td, io_us, rd->io_u_queued_nr);
839		else if (!rd->is_client)
840			ret = fio_rdmaio_recv(td, io_us, rd->io_u_queued_nr);
841		else
842			ret = 0;	/* must be a SYNC */
843
844		if (ret > 0) {
845			fio_rdmaio_queued(td, io_us, ret);
846			io_u_mark_submit(td, ret);
847			rd->io_u_queued_nr -= ret;
848			io_us += ret;
849			ret = 0;
850		} else
851			break;
852	} while (rd->io_u_queued_nr);
853
854	return ret;
855}
856
857static int fio_rdmaio_connect(struct thread_data *td, struct fio_file *f)
858{
859	struct rdmaio_data *rd = td->io_ops_data;
860	struct rdma_conn_param conn_param;
861	struct ibv_send_wr *bad_wr;
862
863	memset(&conn_param, 0, sizeof(conn_param));
864	conn_param.responder_resources = 1;
865	conn_param.initiator_depth = 1;
866	conn_param.retry_count = 10;
867
868	if (rdma_connect(rd->cm_id, &conn_param) != 0) {
869		log_err("fio: rdma_connect fail: %m\n");
870		return 1;
871	}
872
873	if (get_next_channel_event
874	    (td, rd->cm_channel, RDMA_CM_EVENT_ESTABLISHED) != 0) {
875		log_err("fio: wait for RDMA_CM_EVENT_ESTABLISHED\n");
876		return 1;
877	}
878
879	/* send task request */
880	rd->send_buf.mode = htonl(rd->rdma_protocol);
881	rd->send_buf.nr = htonl(td->o.iodepth);
882
883	if (ibv_post_send(rd->qp, &rd->sq_wr, &bad_wr) != 0) {
884		log_err("fio: ibv_post_send fail: %m\n");
885		return 1;
886	}
887
888	if (rdma_poll_wait(td, IBV_WC_SEND) < 0)
889		return 1;
890
891	/* wait for remote MR info from server side */
892	if (rdma_poll_wait(td, IBV_WC_RECV) < 0)
893		return 1;
894
895	/* In SEND/RECV test, it's a good practice to setup the iodepth of
896	 * of the RECV side deeper than that of the SEND side to
897	 * avoid RNR (receiver not ready) error. The
898	 * SEND side may send so many unsolicited message before
899	 * RECV side commits sufficient recv buffers into recv queue.
900	 * This may lead to RNR error. Here, SEND side pauses for a while
901	 * during which RECV side commits sufficient recv buffers.
902	 */
903	usleep(500000);
904
905	return 0;
906}
907
908static int fio_rdmaio_accept(struct thread_data *td, struct fio_file *f)
909{
910	struct rdmaio_data *rd = td->io_ops_data;
911	struct rdma_conn_param conn_param;
912	struct ibv_send_wr *bad_wr;
913	int ret = 0;
914
915	/* rdma_accept() - then wait for accept success */
916	memset(&conn_param, 0, sizeof(conn_param));
917	conn_param.responder_resources = 1;
918	conn_param.initiator_depth = 1;
919
920	if (rdma_accept(rd->child_cm_id, &conn_param) != 0) {
921		log_err("fio: rdma_accept: %m\n");
922		return 1;
923	}
924
925	if (get_next_channel_event
926	    (td, rd->cm_channel, RDMA_CM_EVENT_ESTABLISHED) != 0) {
927		log_err("fio: wait for RDMA_CM_EVENT_ESTABLISHED\n");
928		return 1;
929	}
930
931	/* wait for request */
932	ret = rdma_poll_wait(td, IBV_WC_RECV) < 0;
933
934	if (ibv_post_send(rd->qp, &rd->sq_wr, &bad_wr) != 0) {
935		log_err("fio: ibv_post_send fail: %m\n");
936		return 1;
937	}
938
939	if (rdma_poll_wait(td, IBV_WC_SEND) < 0)
940		return 1;
941
942	return ret;
943}
944
945static int fio_rdmaio_open_file(struct thread_data *td, struct fio_file *f)
946{
947	if (td_read(td))
948		return fio_rdmaio_accept(td, f);
949	else
950		return fio_rdmaio_connect(td, f);
951}
952
953static int fio_rdmaio_close_file(struct thread_data *td, struct fio_file *f)
954{
955	struct rdmaio_data *rd = td->io_ops_data;
956	struct ibv_send_wr *bad_wr;
957
958	/* unregister rdma buffer */
959
960	/*
961	 * Client sends notification to the server side
962	 */
963	/* refer to: http://linux.die.net/man/7/rdma_cm */
964	if ((rd->is_client == 1) && ((rd->rdma_protocol == FIO_RDMA_MEM_WRITE)
965				     || (rd->rdma_protocol ==
966					 FIO_RDMA_MEM_READ))) {
967		if (ibv_post_send(rd->qp, &rd->sq_wr, &bad_wr) != 0) {
968			log_err("fio: ibv_post_send fail: %m\n");
969			return 1;
970		}
971
972		dprint(FD_IO, "fio: close information sent success\n");
973		rdma_poll_wait(td, IBV_WC_SEND);
974	}
975
976	if (rd->is_client == 1)
977		rdma_disconnect(rd->cm_id);
978	else {
979		rdma_disconnect(rd->child_cm_id);
980#if 0
981		rdma_disconnect(rd->cm_id);
982#endif
983	}
984
985#if 0
986	if (get_next_channel_event(td, rd->cm_channel, RDMA_CM_EVENT_DISCONNECTED) != 0) {
987		log_err("fio: wait for RDMA_CM_EVENT_DISCONNECTED\n");
988		return 1;
989	}
990#endif
991
992	ibv_destroy_cq(rd->cq);
993	ibv_destroy_qp(rd->qp);
994
995	if (rd->is_client == 1)
996		rdma_destroy_id(rd->cm_id);
997	else {
998		rdma_destroy_id(rd->child_cm_id);
999		rdma_destroy_id(rd->cm_id);
1000	}
1001
1002	ibv_destroy_comp_channel(rd->channel);
1003	ibv_dealloc_pd(rd->pd);
1004
1005	return 0;
1006}
1007
1008static int fio_rdmaio_setup_connect(struct thread_data *td, const char *host,
1009				    unsigned short port)
1010{
1011	struct rdmaio_data *rd = td->io_ops_data;
1012	struct ibv_recv_wr *bad_wr;
1013	int err;
1014
1015	rd->addr.sin_family = AF_INET;
1016	rd->addr.sin_port = htons(port);
1017
1018	if (inet_aton(host, &rd->addr.sin_addr) != 1) {
1019		struct hostent *hent;
1020
1021		hent = gethostbyname(host);
1022		if (!hent) {
1023			td_verror(td, errno, "gethostbyname");
1024			return 1;
1025		}
1026
1027		memcpy(&rd->addr.sin_addr, hent->h_addr, 4);
1028	}
1029
1030	/* resolve route */
1031	err = rdma_resolve_addr(rd->cm_id, NULL, (struct sockaddr *)&rd->addr, 2000);
1032	if (err != 0) {
1033		log_err("fio: rdma_resolve_addr: %d\n", err);
1034		return 1;
1035	}
1036
1037	err = get_next_channel_event(td, rd->cm_channel, RDMA_CM_EVENT_ADDR_RESOLVED);
1038	if (err != 0) {
1039		log_err("fio: get_next_channel_event: %d\n", err);
1040		return 1;
1041	}
1042
1043	/* resolve route */
1044	err = rdma_resolve_route(rd->cm_id, 2000);
1045	if (err != 0) {
1046		log_err("fio: rdma_resolve_route: %d\n", err);
1047		return 1;
1048	}
1049
1050	err = get_next_channel_event(td, rd->cm_channel, RDMA_CM_EVENT_ROUTE_RESOLVED);
1051	if (err != 0) {
1052		log_err("fio: get_next_channel_event: %d\n", err);
1053		return 1;
1054	}
1055
1056	/* create qp and buffer */
1057	if (fio_rdmaio_setup_qp(td) != 0)
1058		return 1;
1059
1060	if (fio_rdmaio_setup_control_msg_buffers(td) != 0)
1061		return 1;
1062
1063	/* post recv buf */
1064	err = ibv_post_recv(rd->qp, &rd->rq_wr, &bad_wr);
1065	if (err != 0) {
1066		log_err("fio: ibv_post_recv fail: %d\n", err);
1067		return 1;
1068	}
1069
1070	return 0;
1071}
1072
1073static int fio_rdmaio_setup_listen(struct thread_data *td, short port)
1074{
1075	struct rdmaio_data *rd = td->io_ops_data;
1076	struct ibv_recv_wr *bad_wr;
1077	int state = td->runstate;
1078
1079	td_set_runstate(td, TD_SETTING_UP);
1080
1081	rd->addr.sin_family = AF_INET;
1082	rd->addr.sin_addr.s_addr = htonl(INADDR_ANY);
1083	rd->addr.sin_port = htons(port);
1084
1085	/* rdma_listen */
1086	if (rdma_bind_addr(rd->cm_id, (struct sockaddr *)&rd->addr) != 0) {
1087		log_err("fio: rdma_bind_addr fail: %m\n");
1088		return 1;
1089	}
1090
1091	if (rdma_listen(rd->cm_id, 3) != 0) {
1092		log_err("fio: rdma_listen fail: %m\n");
1093		return 1;
1094	}
1095
1096	log_info("fio: waiting for connection\n");
1097
1098	/* wait for CONNECT_REQUEST */
1099	if (get_next_channel_event
1100	    (td, rd->cm_channel, RDMA_CM_EVENT_CONNECT_REQUEST) != 0) {
1101		log_err("fio: wait for RDMA_CM_EVENT_CONNECT_REQUEST\n");
1102		return 1;
1103	}
1104
1105	if (fio_rdmaio_setup_qp(td) != 0)
1106		return 1;
1107
1108	if (fio_rdmaio_setup_control_msg_buffers(td) != 0)
1109		return 1;
1110
1111	/* post recv buf */
1112	if (ibv_post_recv(rd->qp, &rd->rq_wr, &bad_wr) != 0) {
1113		log_err("fio: ibv_post_recv fail: %m\n");
1114		return 1;
1115	}
1116
1117	td_set_runstate(td, state);
1118	return 0;
1119}
1120
1121static int check_set_rlimits(struct thread_data *td)
1122{
1123#ifdef CONFIG_RLIMIT_MEMLOCK
1124	struct rlimit rl;
1125
1126	/* check RLIMIT_MEMLOCK */
1127	if (getrlimit(RLIMIT_MEMLOCK, &rl) != 0) {
1128		log_err("fio: getrlimit fail: %d(%s)\n",
1129			errno, strerror(errno));
1130		return 1;
1131	}
1132
1133	/* soft limit */
1134	if ((rl.rlim_cur != RLIM_INFINITY)
1135	    && (rl.rlim_cur < td->orig_buffer_size)) {
1136		log_err("fio: soft RLIMIT_MEMLOCK is: %" PRId64 "\n",
1137			rl.rlim_cur);
1138		log_err("fio: total block size is:    %zd\n",
1139			td->orig_buffer_size);
1140		/* try to set larger RLIMIT_MEMLOCK */
1141		rl.rlim_cur = rl.rlim_max;
1142		if (setrlimit(RLIMIT_MEMLOCK, &rl) != 0) {
1143			log_err("fio: setrlimit fail: %d(%s)\n",
1144				errno, strerror(errno));
1145			log_err("fio: you may try enlarge MEMLOCK by root\n");
1146			log_err("# ulimit -l unlimited\n");
1147			return 1;
1148		}
1149	}
1150#endif
1151
1152	return 0;
1153}
1154
1155static int compat_options(struct thread_data *td)
1156{
1157	// The original RDMA engine had an ugly / seperator
1158	// on the filename for it's options. This function
1159	// retains backwards compatibility with it.100
1160
1161	struct rdmaio_options *o = td->eo;
1162	char *modep, *portp;
1163	char *filename = td->o.filename;
1164
1165	if (!filename)
1166		return 0;
1167
1168	portp = strchr(filename, '/');
1169	if (portp == NULL)
1170		return 0;
1171
1172	*portp = '\0';
1173	portp++;
1174
1175	o->port = strtol(portp, NULL, 10);
1176	if (!o->port || o->port > 65535)
1177		goto bad_host;
1178
1179	modep = strchr(portp, '/');
1180	if (modep != NULL) {
1181		*modep = '\0';
1182		modep++;
1183	}
1184
1185	if (modep) {
1186		if (!strncmp("rdma_write", modep, strlen(modep)) ||
1187		    !strncmp("RDMA_WRITE", modep, strlen(modep)))
1188			o->verb = FIO_RDMA_MEM_WRITE;
1189		else if (!strncmp("rdma_read", modep, strlen(modep)) ||
1190			 !strncmp("RDMA_READ", modep, strlen(modep)))
1191			o->verb = FIO_RDMA_MEM_READ;
1192		else if (!strncmp("send", modep, strlen(modep)) ||
1193			 !strncmp("SEND", modep, strlen(modep)))
1194			o->verb = FIO_RDMA_CHA_SEND;
1195		else
1196			goto bad_host;
1197	} else
1198		o->verb = FIO_RDMA_MEM_WRITE;
1199
1200
1201	return 0;
1202
1203bad_host:
1204	log_err("fio: bad rdma host/port/protocol: %s\n", td->o.filename);
1205	return 1;
1206}
1207
1208static int fio_rdmaio_init(struct thread_data *td)
1209{
1210	struct rdmaio_data *rd = td->io_ops_data;
1211	struct rdmaio_options *o = td->eo;
1212	unsigned int max_bs;
1213	int ret, i;
1214
1215	if (td_rw(td)) {
1216		log_err("fio: rdma connections must be read OR write\n");
1217		return 1;
1218	}
1219	if (td_random(td)) {
1220		log_err("fio: RDMA network IO can't be random\n");
1221		return 1;
1222	}
1223
1224	if (compat_options(td))
1225		return 1;
1226
1227	if (!o->port) {
1228		log_err("fio: no port has been specified which is required "
1229			"for the rdma engine\n");
1230		return 1;
1231	}
1232
1233	if (check_set_rlimits(td))
1234		return 1;
1235
1236	rd->rdma_protocol = o->verb;
1237	rd->cq_event_num = 0;
1238
1239	rd->cm_channel = rdma_create_event_channel();
1240	if (!rd->cm_channel) {
1241		log_err("fio: rdma_create_event_channel fail: %m\n");
1242		return 1;
1243	}
1244
1245	ret = rdma_create_id(rd->cm_channel, &rd->cm_id, rd, RDMA_PS_TCP);
1246	if (ret) {
1247		log_err("fio: rdma_create_id fail: %m\n");
1248		return 1;
1249	}
1250
1251	if ((rd->rdma_protocol == FIO_RDMA_MEM_WRITE) ||
1252	    (rd->rdma_protocol == FIO_RDMA_MEM_READ)) {
1253		rd->rmt_us =
1254			malloc(FIO_RDMA_MAX_IO_DEPTH * sizeof(struct remote_u));
1255		memset(rd->rmt_us, 0,
1256			FIO_RDMA_MAX_IO_DEPTH * sizeof(struct remote_u));
1257		rd->rmt_nr = 0;
1258	}
1259
1260	rd->io_us_queued = malloc(td->o.iodepth * sizeof(struct io_u *));
1261	memset(rd->io_us_queued, 0, td->o.iodepth * sizeof(struct io_u *));
1262	rd->io_u_queued_nr = 0;
1263
1264	rd->io_us_flight = malloc(td->o.iodepth * sizeof(struct io_u *));
1265	memset(rd->io_us_flight, 0, td->o.iodepth * sizeof(struct io_u *));
1266	rd->io_u_flight_nr = 0;
1267
1268	rd->io_us_completed = malloc(td->o.iodepth * sizeof(struct io_u *));
1269	memset(rd->io_us_completed, 0, td->o.iodepth * sizeof(struct io_u *));
1270	rd->io_u_completed_nr = 0;
1271
1272	if (td_read(td)) {	/* READ as the server */
1273		rd->is_client = 0;
1274		td->flags |= TD_F_NO_PROGRESS;
1275		/* server rd->rdma_buf_len will be setup after got request */
1276		ret = fio_rdmaio_setup_listen(td, o->port);
1277	} else {		/* WRITE as the client */
1278		rd->is_client = 1;
1279		ret = fio_rdmaio_setup_connect(td, td->o.filename, o->port);
1280	}
1281
1282	max_bs = max(td->o.max_bs[DDIR_READ], td->o.max_bs[DDIR_WRITE]);
1283	rd->send_buf.max_bs = htonl(max_bs);
1284
1285	/* register each io_u in the free list */
1286	for (i = 0; i < td->io_u_freelist.nr; i++) {
1287		struct io_u *io_u = td->io_u_freelist.io_us[i];
1288
1289		io_u->engine_data = malloc(sizeof(struct rdma_io_u_data));
1290		memset(io_u->engine_data, 0, sizeof(struct rdma_io_u_data));
1291		((struct rdma_io_u_data *)io_u->engine_data)->wr_id = i;
1292
1293		io_u->mr = ibv_reg_mr(rd->pd, io_u->buf, max_bs,
1294				      IBV_ACCESS_LOCAL_WRITE |
1295				      IBV_ACCESS_REMOTE_READ |
1296				      IBV_ACCESS_REMOTE_WRITE);
1297		if (io_u->mr == NULL) {
1298			log_err("fio: ibv_reg_mr io_u failed: %m\n");
1299			return 1;
1300		}
1301
1302		rd->send_buf.rmt_us[i].buf =
1303		    htonll((uint64_t) (unsigned long)io_u->buf);
1304		rd->send_buf.rmt_us[i].rkey = htonl(io_u->mr->rkey);
1305		rd->send_buf.rmt_us[i].size = htonl(max_bs);
1306
1307#if 0
1308		log_info("fio: Send rkey %x addr %" PRIx64 " len %d to client\n", io_u->mr->rkey, io_u->buf, max_bs); */
1309#endif
1310	}
1311
1312	rd->send_buf.nr = htonl(i);
1313
1314	return ret;
1315}
1316
1317static void fio_rdmaio_cleanup(struct thread_data *td)
1318{
1319	struct rdmaio_data *rd = td->io_ops_data;
1320
1321	if (rd)
1322		free(rd);
1323}
1324
1325static int fio_rdmaio_setup(struct thread_data *td)
1326{
1327	struct rdmaio_data *rd;
1328
1329	if (!td->files_index) {
1330		add_file(td, td->o.filename ?: "rdma", 0, 0);
1331		td->o.nr_files = td->o.nr_files ?: 1;
1332		td->o.open_files++;
1333	}
1334
1335	if (!td->io_ops_data) {
1336		rd = malloc(sizeof(*rd));
1337
1338		memset(rd, 0, sizeof(*rd));
1339		init_rand_seed(&rd->rand_state, (unsigned int) GOLDEN_RATIO_PRIME, 0);
1340		td->io_ops_data = rd;
1341	}
1342
1343	return 0;
1344}
1345
1346static struct ioengine_ops ioengine_rw = {
1347	.name			= "rdma",
1348	.version		= FIO_IOOPS_VERSION,
1349	.setup			= fio_rdmaio_setup,
1350	.init			= fio_rdmaio_init,
1351	.prep			= fio_rdmaio_prep,
1352	.queue			= fio_rdmaio_queue,
1353	.commit			= fio_rdmaio_commit,
1354	.getevents		= fio_rdmaio_getevents,
1355	.event			= fio_rdmaio_event,
1356	.cleanup		= fio_rdmaio_cleanup,
1357	.open_file		= fio_rdmaio_open_file,
1358	.close_file		= fio_rdmaio_close_file,
1359	.flags			= FIO_DISKLESSIO | FIO_UNIDIR | FIO_PIPEIO,
1360	.options		= options,
1361	.option_struct_size	= sizeof(struct rdmaio_options),
1362};
1363
1364static void fio_init fio_rdmaio_register(void)
1365{
1366	register_ioengine(&ioengine_rw);
1367}
1368
1369static void fio_exit fio_rdmaio_unregister(void)
1370{
1371	unregister_ioengine(&ioengine_rw);
1372}
1373