projects / fio.git / blame_incremental
? search:
summary | shortlog | log | commit | commitdiff | tree
blob | blame (non-incremental) | history | HEAD
RDMA engine: rewrite header comment block and man page section
[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 * This I/O engine is disabled by default. To enable it, execute:
9 *
10 * $ export EXTFLAGS="-DFIO_HAVE_RDMA"
11 * $ export EXTLIBS="-libverbs -lrdmacm"
12 *
13 * before running make. You will need the Linux RDMA software as well, either
14 * from your Linux distributor or directly from openfabrics.org:
15 *
16 * http://www.openfabrics.org/downloads/OFED/
17 *
18 */
19#include <stdio.h>
20#include <stdlib.h>
21#include <unistd.h>
22#include <errno.h>
23#include <assert.h>
24#include <netinet/in.h>
25#include <arpa/inet.h>
26#include <netdb.h>
27#include <sys/poll.h>
28#include <sys/types.h>
29#include <sys/socket.h>
30#include <sys/time.h>
31#include <sys/resource.h>
32
33#include <byteswap.h>
34#include <pthread.h>
35#include <inttypes.h>
36
37#include "../fio.h"
38
39#ifdef FIO_HAVE_RDMA
40
41#include <rdma/rdma_cma.h>
42#include <infiniband/arch.h>
43
44#define FIO_RDMA_MAX_IO_DEPTH 128
45
46enum rdma_io_mode {
47 FIO_RDMA_UNKNOWN = 0,
48 FIO_RDMA_MEM_WRITE,
49 FIO_RDMA_MEM_READ,
50 FIO_RDMA_CHA_SEND,
51 FIO_RDMA_CHA_RECV
52};
53
54struct remote_u {
55 uint64_t buf;
56 uint32_t rkey;
57 uint32_t size;
58};
59
60struct rdma_info_blk {
61 uint32_t mode; /* channel semantic or memory semantic */
62 uint32_t nr; /* client: io depth
63 server: number of records for memory semantic
64 */
65 struct remote_u rmt_us[FIO_RDMA_MAX_IO_DEPTH];
66};
67
68struct rdma_io_u_data {
69 uint64_t wr_id;
70 struct ibv_send_wr sq_wr;
71 struct ibv_recv_wr rq_wr;
72 struct ibv_sge rdma_sgl;
73};
74
75struct rdmaio_data {
76 int is_client;
77 enum rdma_io_mode rdma_protocol;
78 char host[64];
79 struct sockaddr_in addr;
80
81 struct ibv_recv_wr rq_wr;
82 struct ibv_sge recv_sgl;
83 struct rdma_info_blk recv_buf;
84 struct ibv_mr *recv_mr;
85
86 struct ibv_send_wr sq_wr;
87 struct ibv_sge send_sgl;
88 struct rdma_info_blk send_buf;
89 struct ibv_mr *send_mr;
90
91 struct ibv_comp_channel *channel;
92 struct ibv_cq *cq;
93 struct ibv_pd *pd;
94 struct ibv_qp *qp;
95
96 pthread_t cmthread;
97 struct rdma_event_channel *cm_channel;
98 struct rdma_cm_id *cm_id;
99 struct rdma_cm_id *child_cm_id;
100
101 int cq_event_num;
102
103 struct remote_u *rmt_us;
104 int rmt_nr;
105 struct io_u **io_us_queued;
106 int io_u_queued_nr;
107 struct io_u **io_us_flight;
108 int io_u_flight_nr;
109 struct io_u **io_us_completed;
110 int io_u_completed_nr;
111};
112
113static int client_recv(struct thread_data *td, struct ibv_wc *wc)
114{
115 struct rdmaio_data *rd = td->io_ops->data;
116
117 if (wc->byte_len != sizeof(rd->recv_buf)) {
118 log_err("Received bogus data, size %d\n", wc->byte_len);
119 return 1;
120 }
121
122 /* store mr info for MEMORY semantic */
123 if ((rd->rdma_protocol == FIO_RDMA_MEM_WRITE) ||
124 (rd->rdma_protocol == FIO_RDMA_MEM_READ)) {
125 /* struct flist_head *entry; */
126 int i = 0;
127
128 rd->rmt_nr = ntohl(rd->recv_buf.nr);
129
130 for (i = 0; i < rd->rmt_nr; i++) {
131 rd->rmt_us[i].buf = ntohll(rd->recv_buf.rmt_us[i].buf);
132 rd->rmt_us[i].rkey = ntohl(rd->recv_buf.rmt_us[i].rkey);
133 rd->rmt_us[i].size = ntohl(rd->recv_buf.rmt_us[i].size);
134
135 dprint(FD_IO,
136 "fio: Received rkey %x addr %" PRIx64
137 " len %d from peer\n", rd->rmt_us[i].rkey,
138 rd->rmt_us[i].buf, rd->rmt_us[i].size);
139 }
140 }
141
142 return 0;
143}
144
145static int server_recv(struct thread_data *td, struct ibv_wc *wc)
146{
147 struct rdmaio_data *rd = td->io_ops->data;
148
149 if (wc->wr_id == FIO_RDMA_MAX_IO_DEPTH) {
150 rd->rdma_protocol = ntohl(rd->recv_buf.mode);
151
152 /* CHANNEL semantic, do nothing */
153 if (rd->rdma_protocol == FIO_RDMA_CHA_SEND)
154 rd->rdma_protocol = FIO_RDMA_CHA_RECV;
155 }
156
157 return 0;
158}
159
160static int cq_event_handler(struct thread_data *td, enum ibv_wc_opcode opcode)
161{
162 struct rdmaio_data *rd = td->io_ops->data;
163 struct ibv_wc wc;
164 struct rdma_io_u_data *r_io_u_d;
165 int ret;
166 int compevnum = 0;
167 int i;
168
169 while ((ret = ibv_poll_cq(rd->cq, 1, &wc)) == 1) {
170 ret = 0;
171 compevnum++;
172
173 if (wc.status) {
174 log_err("fio: cq completion status %d(%s)\n",
175 wc.status, ibv_wc_status_str(wc.status));
176 return -1;
177 }
178
179 switch (wc.opcode) {
180
181 case IBV_WC_RECV:
182 if (rd->is_client == 1)
183 client_recv(td, &wc);
184 else
185 server_recv(td, &wc);
186
187 if (wc.wr_id == FIO_RDMA_MAX_IO_DEPTH)
188 break;
189
190 for (i = 0; i < rd->io_u_flight_nr; i++) {
191 r_io_u_d = rd->io_us_flight[i]->engine_data;
192
193 if (wc.wr_id == r_io_u_d->rq_wr.wr_id) {
194 rd->io_us_flight[i]->resid =
195 rd->io_us_flight[i]->buflen
196 - wc.byte_len;
197
198 rd->io_us_flight[i]->error = 0;
199
200 rd->io_us_completed[rd->
201 io_u_completed_nr]
202 = rd->io_us_flight[i];
203 rd->io_u_completed_nr++;
204 break;
205 }
206 }
207 if (i == rd->io_u_flight_nr)
208 log_err("fio: recv wr %ld not found\n",
209 wc.wr_id);
210 else {
211 /* put the last one into middle of the list */
212 rd->io_us_flight[i] =
213 rd->io_us_flight[rd->io_u_flight_nr - 1];
214 rd->io_u_flight_nr--;
215 }
216
217 break;
218
219 case IBV_WC_SEND:
220 case IBV_WC_RDMA_WRITE:
221 case IBV_WC_RDMA_READ:
222 if (wc.wr_id == FIO_RDMA_MAX_IO_DEPTH)
223 break;
224
225 for (i = 0; i < rd->io_u_flight_nr; i++) {
226 r_io_u_d = rd->io_us_flight[i]->engine_data;
227
228 if (wc.wr_id == r_io_u_d->sq_wr.wr_id) {
229 rd->io_us_completed[rd->
230 io_u_completed_nr]
231 = rd->io_us_flight[i];
232 rd->io_u_completed_nr++;
233 break;
234 }
235 }
236 if (i == rd->io_u_flight_nr)
237 log_err("fio: send wr %ld not found\n",
238 wc.wr_id);
239 else {
240 /* put the last one into middle of the list */
241 rd->io_us_flight[i] =
242 rd->io_us_flight[rd->io_u_flight_nr - 1];
243 rd->io_u_flight_nr--;
244 }
245
246 break;
247
248 default:
249 log_info("fio: unknown completion event %d\n",
250 wc.opcode);
251 return -1;
252 }
253 rd->cq_event_num++;
254 }
255 if (ret) {
256 log_err("fio: poll error %d\n", ret);
257 return 1;
258 }
259
260 return compevnum;
261}
262
263/*
264 * Return -1 for error and 'nr events' for a positive number
265 * of events
266 */
267static int rdma_poll_wait(struct thread_data *td, enum ibv_wc_opcode opcode)
268{
269 struct rdmaio_data *rd = td->io_ops->data;
270 struct ibv_cq *ev_cq;
271 void *ev_ctx;
272 int ret;
273
274 if (rd->cq_event_num > 0) { /* previous left */
275 rd->cq_event_num--;
276 return 0;
277 }
278
279again:
280 if (ibv_get_cq_event(rd->channel, &ev_cq, &ev_ctx) != 0) {
281 log_err("fio: Failed to get cq event!\n");
282 return -1;
283 }
284 if (ev_cq != rd->cq) {
285 log_err("fio: Unknown CQ!\n");
286 return -1;
287 }
288 if (ibv_req_notify_cq(rd->cq, 0) != 0) {
289 log_err("fio: Failed to set notify!\n");
290 return -1;
291 }
292
293 ret = cq_event_handler(td, opcode);
294 if (ret < 1)
295 goto again;
296
297 ibv_ack_cq_events(rd->cq, ret);
298
299 rd->cq_event_num--;
300
301 return ret;
302}
303
304static int fio_rdmaio_setup_qp(struct thread_data *td)
305{
306 struct rdmaio_data *rd = td->io_ops->data;
307 struct ibv_qp_init_attr init_attr;
308 int qp_depth = td->o.iodepth * 2; /* 2 times of io depth */
309
310 if (rd->is_client == 0)
311 rd->pd = ibv_alloc_pd(rd->child_cm_id->verbs);
312 else
313 rd->pd = ibv_alloc_pd(rd->cm_id->verbs);
314 if (rd->pd == NULL) {
315 log_err("fio: ibv_alloc_pd fail\n");
316 return 1;
317 }
318
319 if (rd->is_client == 0)
320 rd->channel = ibv_create_comp_channel(rd->child_cm_id->verbs);
321 else
322 rd->channel = ibv_create_comp_channel(rd->cm_id->verbs);
323 if (rd->channel == NULL) {
324 log_err("fio: ibv_create_comp_channel fail\n");
325 goto err1;
326 }
327
328 if (qp_depth < 16)
329 qp_depth = 16;
330
331 if (rd->is_client == 0)
332 rd->cq = ibv_create_cq(rd->child_cm_id->verbs,
333 qp_depth, rd, rd->channel, 0);
334 else
335 rd->cq = ibv_create_cq(rd->cm_id->verbs,
336 qp_depth, rd, rd->channel, 0);
337 if (rd->cq == NULL) {
338 log_err("fio: ibv_create_cq failed\n");
339 goto err2;
340 }
341
342 if (ibv_req_notify_cq(rd->cq, 0) != 0) {
343 log_err("fio: ibv_create_cq failed\n");
344 goto err3;
345 }
346
347 /* create queue pair */
348 memset(&init_attr, 0, sizeof(init_attr));
349 init_attr.cap.max_send_wr = qp_depth;
350 init_attr.cap.max_recv_wr = qp_depth;
351 init_attr.cap.max_recv_sge = 1;
352 init_attr.cap.max_send_sge = 1;
353 init_attr.qp_type = IBV_QPT_RC;
354 init_attr.send_cq = rd->cq;
355 init_attr.recv_cq = rd->cq;
356
357 if (rd->is_client == 0) {
358 if (rdma_create_qp(rd->child_cm_id, rd->pd, &init_attr) != 0) {
359 log_err("fio: rdma_create_qp failed\n");
360 goto err3;
361 }
362 rd->qp = rd->child_cm_id->qp;
363 } else {
364 if (rdma_create_qp(rd->cm_id, rd->pd, &init_attr) != 0) {
365 log_err("fio: rdma_create_qp failed\n");
366 goto err3;
367 }
368 rd->qp = rd->cm_id->qp;
369 }
370
371 return 0;
372
373err3:
374 ibv_destroy_cq(rd->cq);
375err2:
376 ibv_destroy_comp_channel(rd->channel);
377err1:
378 ibv_dealloc_pd(rd->pd);
379
380 return 1;
381}
382
383static int fio_rdmaio_setup_control_msg_buffers(struct thread_data *td)
384{
385 struct rdmaio_data *rd = td->io_ops->data;
386
387 rd->recv_mr = ibv_reg_mr(rd->pd, &rd->recv_buf, sizeof(rd->recv_buf),
388 IBV_ACCESS_LOCAL_WRITE);
389 if (rd->recv_mr == NULL) {
390 log_err("fio: recv_buf reg_mr failed\n");
391 return 1;
392 }
393
394 rd->send_mr = ibv_reg_mr(rd->pd, &rd->send_buf, sizeof(rd->send_buf),
395 0);
396 if (rd->send_mr == NULL) {
397 log_err("fio: send_buf reg_mr failed\n");
398 ibv_dereg_mr(rd->recv_mr);
399 return 1;
400 }
401
402 /* setup work request */
403 /* recv wq */
404 rd->recv_sgl.addr = (uint64_t) (unsigned long)&rd->recv_buf;
405 rd->recv_sgl.length = sizeof rd->recv_buf;
406 rd->recv_sgl.lkey = rd->recv_mr->lkey;
407 rd->rq_wr.sg_list = &rd->recv_sgl;
408 rd->rq_wr.num_sge = 1;
409 rd->rq_wr.wr_id = FIO_RDMA_MAX_IO_DEPTH;
410
411 /* send wq */
412 rd->send_sgl.addr = (uint64_t) (unsigned long)&rd->send_buf;
413 rd->send_sgl.length = sizeof rd->send_buf;
414 rd->send_sgl.lkey = rd->send_mr->lkey;
415
416 rd->sq_wr.opcode = IBV_WR_SEND;
417 rd->sq_wr.send_flags = IBV_SEND_SIGNALED;
418 rd->sq_wr.sg_list = &rd->send_sgl;
419 rd->sq_wr.num_sge = 1;
420 rd->sq_wr.wr_id = FIO_RDMA_MAX_IO_DEPTH;
421
422 return 0;
423}
424
425static int get_next_channel_event(struct thread_data *td,
426 struct rdma_event_channel *channel,
427 enum rdma_cm_event_type wait_event)
428{
429 struct rdmaio_data *rd = td->io_ops->data;
430
431 int ret;
432 struct rdma_cm_event *event;
433
434 ret = rdma_get_cm_event(channel, &event);
435 if (ret) {
436 log_err("fio: rdma_get_cm_event");
437 return 1;
438 }
439
440 if (event->event != wait_event) {
441 log_err("fio: event is %s instead of %s\n",
442 rdma_event_str(event->event),
443 rdma_event_str(wait_event));
444 return 1;
445 }
446
447 switch (event->event) {
448 case RDMA_CM_EVENT_CONNECT_REQUEST:
449 rd->child_cm_id = event->id;
450 break;
451 default:
452 break;
453 }
454
455 rdma_ack_cm_event(event);
456
457 return 0;
458}
459
460static int fio_rdmaio_prep(struct thread_data *td, struct io_u *io_u)
461{
462 struct rdmaio_data *rd = td->io_ops->data;
463 struct rdma_io_u_data *r_io_u_d;
464
465 r_io_u_d = io_u->engine_data;
466
467 switch (rd->rdma_protocol) {
468 case FIO_RDMA_MEM_WRITE:
469 case FIO_RDMA_MEM_READ:
470 r_io_u_d->rdma_sgl.addr = (uint64_t) (unsigned long)io_u->buf;
471 r_io_u_d->rdma_sgl.lkey = io_u->mr->lkey;
472 r_io_u_d->sq_wr.wr_id = r_io_u_d->wr_id;
473 r_io_u_d->sq_wr.send_flags = IBV_SEND_SIGNALED;
474 r_io_u_d->sq_wr.sg_list = &r_io_u_d->rdma_sgl;
475 r_io_u_d->sq_wr.num_sge = 1;
476 break;
477 case FIO_RDMA_CHA_SEND:
478 r_io_u_d->rdma_sgl.addr = (uint64_t) (unsigned long)io_u->buf;
479 r_io_u_d->rdma_sgl.lkey = io_u->mr->lkey;
480 r_io_u_d->rdma_sgl.length = io_u->buflen;
481 r_io_u_d->sq_wr.wr_id = r_io_u_d->wr_id;
482 r_io_u_d->sq_wr.opcode = IBV_WR_SEND;
483 r_io_u_d->sq_wr.send_flags = IBV_SEND_SIGNALED;
484 r_io_u_d->sq_wr.sg_list = &r_io_u_d->rdma_sgl;
485 r_io_u_d->sq_wr.num_sge = 1;
486 break;
487 case FIO_RDMA_CHA_RECV:
488 r_io_u_d->rdma_sgl.addr = (uint64_t) (unsigned long)io_u->buf;
489 r_io_u_d->rdma_sgl.lkey = io_u->mr->lkey;
490 r_io_u_d->rdma_sgl.length = io_u->buflen;
491 r_io_u_d->rq_wr.wr_id = r_io_u_d->wr_id;
492 r_io_u_d->rq_wr.sg_list = &r_io_u_d->rdma_sgl;
493 r_io_u_d->rq_wr.num_sge = 1;
494 break;
495 default:
496 log_err("fio: unknown rdma protocol - %d\n", rd->rdma_protocol);
497 break;
498 }
499
500 return 0;
501}
502
503static struct io_u *fio_rdmaio_event(struct thread_data *td, int event)
504{
505 struct rdmaio_data *rd = td->io_ops->data;
506 struct io_u *io_u;
507 int i;
508
509 io_u = rd->io_us_completed[0];
510 for (i = 0; i < rd->io_u_completed_nr - 1; i++) {
511 rd->io_us_completed[i] = rd->io_us_completed[i + 1];
512 }
513 rd->io_u_completed_nr--;
514
515 dprint_io_u(io_u, "fio_rdmaio_event");
516
517 return io_u;
518}
519
520static int fio_rdmaio_getevents(struct thread_data *td, unsigned int min,
521 unsigned int max, struct timespec *t)
522{
523 struct rdmaio_data *rd = td->io_ops->data;
524 int r;
525 enum ibv_wc_opcode comp_opcode;
526 comp_opcode = IBV_WC_RDMA_WRITE;
527 struct ibv_cq *ev_cq;
528 void *ev_ctx;
529 int ret;
530
531 r = 0;
532
533 switch (rd->rdma_protocol) {
534 case FIO_RDMA_MEM_WRITE:
535 comp_opcode = IBV_WC_RDMA_WRITE;
536 break;
537 case FIO_RDMA_MEM_READ:
538 comp_opcode = IBV_WC_RDMA_READ;
539 break;
540 case FIO_RDMA_CHA_SEND:
541 comp_opcode = IBV_WC_SEND;
542 break;
543 case FIO_RDMA_CHA_RECV:
544 comp_opcode = IBV_WC_RECV;
545 break;
546 default:
547 log_err("fio: unknown rdma protocol - %d\n", rd->rdma_protocol);
548 break;
549 }
550
551 if (rd->cq_event_num > 0) { /* previous left */
552 rd->cq_event_num--;
553 return 0;
554 }
555
556again:
557 if (ibv_get_cq_event(rd->channel, &ev_cq, &ev_ctx) != 0) {
558 log_err("fio: Failed to get cq event!\n");
559 return -1;
560 }
561 if (ev_cq != rd->cq) {
562 log_err("fio: Unknown CQ!\n");
563 return -1;
564 }
565 if (ibv_req_notify_cq(rd->cq, 0) != 0) {
566 log_err("fio: Failed to set notify!\n");
567 return -1;
568 }
569
570 ret = cq_event_handler(td, comp_opcode);
571 if (ret < 1)
572 goto again;
573
574 ibv_ack_cq_events(rd->cq, ret);
575
576 r += ret;
577 if (r < min)
578 goto again;
579
580 rd->cq_event_num -= r;
581
582 return r;
583}
584
585static int fio_rdmaio_send(struct thread_data *td, struct io_u **io_us,
586 unsigned int nr)
587{
588 struct rdmaio_data *rd = td->io_ops->data;
589 struct ibv_send_wr *bad_wr;
590 enum ibv_wc_opcode comp_opcode;
591 comp_opcode = IBV_WC_RDMA_WRITE;
592 int i, index;
593 struct rdma_io_u_data *r_io_u_d;
594
595 r_io_u_d = NULL;
596
597 for (i = 0; i < nr; i++) {
598 /* RDMA_WRITE or RDMA_READ */
599 switch (rd->rdma_protocol) {
600 case FIO_RDMA_MEM_WRITE:
601 /* compose work request */
602 r_io_u_d = io_us[i]->engine_data;
603 index = rand() % rd->rmt_nr;
604 r_io_u_d->sq_wr.opcode = IBV_WR_RDMA_WRITE;
605 r_io_u_d->sq_wr.wr.rdma.rkey = rd->rmt_us[index].rkey;
606 r_io_u_d->sq_wr.wr.rdma.remote_addr = \
607 rd->rmt_us[index].buf;
608 r_io_u_d->sq_wr.sg_list->length = io_us[i]->buflen;
609 break;
610 case FIO_RDMA_MEM_READ:
611 /* compose work request */
612 r_io_u_d = io_us[i]->engine_data;
613 index = rand() % rd->rmt_nr;
614 r_io_u_d->sq_wr.opcode = IBV_WR_RDMA_READ;
615 r_io_u_d->sq_wr.wr.rdma.rkey = rd->rmt_us[index].rkey;
616 r_io_u_d->sq_wr.wr.rdma.remote_addr = \
617 rd->rmt_us[index].buf;
618 r_io_u_d->sq_wr.sg_list->length = io_us[i]->buflen;
619 break;
620 case FIO_RDMA_CHA_SEND:
621 r_io_u_d = io_us[i]->engine_data;
622 r_io_u_d->sq_wr.opcode = IBV_WR_SEND;
623 r_io_u_d->sq_wr.send_flags = IBV_SEND_SIGNALED;
624 break;
625 default:
626 log_err("fio: unknown rdma protocol - %d\n",
627 rd->rdma_protocol);
628 break;
629 }
630
631 if (ibv_post_send(rd->qp, &r_io_u_d->sq_wr, &bad_wr) != 0) {
632 log_err("fio: ibv_post_send fail\n");
633 return -1;
634 }
635
636 dprint_io_u(io_us[i], "fio_rdmaio_send");
637 }
638
639 /* wait for completion
640 rdma_poll_wait(td, comp_opcode); */
641
642 return i;
643}
644
645static int fio_rdmaio_recv(struct thread_data *td, struct io_u **io_us,
646 unsigned int nr)
647{
648 struct rdmaio_data *rd = td->io_ops->data;
649 struct ibv_recv_wr *bad_wr;
650 struct rdma_io_u_data *r_io_u_d;
651 int i;
652
653 i = 0;
654 if (rd->rdma_protocol == FIO_RDMA_CHA_RECV) {
655 /* post io_u into recv queue */
656 for (i = 0; i < nr; i++) {
657 r_io_u_d = io_us[i]->engine_data;
658 if (ibv_post_recv(rd->qp, &r_io_u_d->rq_wr, &bad_wr) !=
659 0) {
660 log_err("fio: ibv_post_recv fail\n");
661 return 1;
662 }
663 }
664 } else if ((rd->rdma_protocol == FIO_RDMA_MEM_READ)
665 || (rd->rdma_protocol == FIO_RDMA_MEM_WRITE)) {
666 /* re-post the rq_wr */
667 if (ibv_post_recv(rd->qp, &rd->rq_wr, &bad_wr) != 0) {
668 log_err("fio: ibv_post_recv fail\n");
669 return 1;
670 }
671
672 rdma_poll_wait(td, IBV_WC_RECV);
673
674 dprint(FD_IO, "fio: recv FINISH message\n");
675 exit(0);
676 }
677
678 return i;
679}
680
681static int fio_rdmaio_queue(struct thread_data *td, struct io_u *io_u)
682{
683 struct rdmaio_data *rd = td->io_ops->data;
684
685 fio_ro_check(td, io_u);
686
687 if (rd->io_u_queued_nr == (int)td->o.iodepth)
688 return FIO_Q_BUSY;
689
690 rd->io_us_queued[rd->io_u_queued_nr] = io_u;
691 rd->io_u_queued_nr++;
692
693 dprint_io_u(io_u, "fio_rdmaio_queue");
694
695 return FIO_Q_QUEUED;
696}
697
698static void fio_rdmaio_queued(struct thread_data *td, struct io_u **io_us,
699 unsigned int nr)
700{
701 struct rdmaio_data *rd = td->io_ops->data;
702 struct timeval now;
703 unsigned int i;
704
705 if (!fio_fill_issue_time(td))
706 return;
707
708 fio_gettime(&now, NULL);
709
710 for (i = 0; i < nr; i++) {
711 struct io_u *io_u = io_us[i];
712
713 /* queued -> flight */
714 rd->io_us_flight[rd->io_u_flight_nr] = io_u;
715 rd->io_u_flight_nr++;
716
717 memcpy(&io_u->issue_time, &now, sizeof(now));
718 io_u_queued(td, io_u);
719 }
720}
721
722static int fio_rdmaio_commit(struct thread_data *td)
723{
724 struct rdmaio_data *rd = td->io_ops->data;
725 struct io_u **io_us;
726 int ret;
727
728 if (!rd->io_us_queued)
729 return 0;
730
731 io_us = rd->io_us_queued;
732 do {
733 /* RDMA_WRITE or RDMA_READ */
734 if (rd->is_client) {
735 ret = fio_rdmaio_send(td, io_us, rd->io_u_queued_nr);
736 } else if (!rd->is_client) {
737 ret = fio_rdmaio_recv(td, io_us, rd->io_u_queued_nr);
738 } else
739 ret = 0; /* must be a SYNC */
740
741 if (ret > 0) {
742 fio_rdmaio_queued(td, io_us, ret);
743 io_u_mark_submit(td, ret);
744 rd->io_u_queued_nr -= ret;
745 io_us += ret;
746 ret = 0;
747 } else
748 break;
749 } while (rd->io_u_queued_nr);
750
751 return ret;
752}
753
754static int fio_rdmaio_connect(struct thread_data *td, struct fio_file *f)
755{
756 struct rdmaio_data *rd = td->io_ops->data;
757 struct rdma_conn_param conn_param;
758 struct ibv_send_wr *bad_wr;
759
760 memset(&conn_param, 0, sizeof conn_param);
761 conn_param.responder_resources = 1;
762 conn_param.initiator_depth = 1;
763 conn_param.retry_count = 10;
764
765 if (rdma_connect(rd->cm_id, &conn_param) != 0) {
766 log_err("fio: rdma_connect fail\n");
767 return 1;
768 }
769
770 if (get_next_channel_event
771 (td, rd->cm_channel, RDMA_CM_EVENT_ESTABLISHED) != 0) {
772 log_err("fio: wait for RDMA_CM_EVENT_ESTABLISHED\n");
773 return 1;
774 }
775
776 /* send task request */
777 rd->send_buf.mode = htonl(rd->rdma_protocol);
778 rd->send_buf.nr = htonl(td->o.iodepth);
779
780 if (ibv_post_send(rd->qp, &rd->sq_wr, &bad_wr) != 0) {
781 log_err("fio: ibv_post_send fail");
782 return 1;
783 }
784
785 rdma_poll_wait(td, IBV_WC_SEND);
786
787 /* wait for remote MR info from server side */
788 rdma_poll_wait(td, IBV_WC_RECV);
789
790 return 0;
791}
792
793static int fio_rdmaio_accept(struct thread_data *td, struct fio_file *f)
794{
795 struct rdmaio_data *rd = td->io_ops->data;
796 struct rdma_conn_param conn_param;
797 struct ibv_send_wr *bad_wr;
798
799 /* rdma_accept() - then wait for accept success */
800 memset(&conn_param, 0, sizeof conn_param);
801 conn_param.responder_resources = 1;
802 conn_param.initiator_depth = 1;
803
804 if (rdma_accept(rd->child_cm_id, &conn_param) != 0) {
805 log_err("fio: rdma_accept\n");
806 return 1;
807 }
808
809 if (get_next_channel_event
810 (td, rd->cm_channel, RDMA_CM_EVENT_ESTABLISHED) != 0) {
811 log_err("fio: wait for RDMA_CM_EVENT_ESTABLISHED\n");
812 return 1;
813 }
814
815 /* wait for request */
816 rdma_poll_wait(td, IBV_WC_RECV);
817
818 if (ibv_post_send(rd->qp, &rd->sq_wr, &bad_wr) != 0) {
819 log_err("fio: ibv_post_send fail");
820 return 1;
821 }
822
823 rdma_poll_wait(td, IBV_WC_SEND);
824
825 return 0;
826}
827
828static int fio_rdmaio_open_file(struct thread_data *td, struct fio_file *f)
829{
830 if (td_read(td))
831 return fio_rdmaio_accept(td, f);
832 else
833 return fio_rdmaio_connect(td, f);
834}
835
836static int fio_rdmaio_close_file(struct thread_data *td, struct fio_file *f)
837{
838 struct rdmaio_data *rd = td->io_ops->data;
839 struct ibv_send_wr *bad_wr;
840
841 /* unregister rdma buffer */
842
843 /*
844 * Client sends notification to the server side
845 */
846 /* refer to: http://linux.die.net/man/7/rdma_cm */
847 if ((rd->is_client == 1) && ((rd->rdma_protocol == FIO_RDMA_MEM_WRITE)
848 || (rd->rdma_protocol ==
849 FIO_RDMA_MEM_READ))) {
850 if (ibv_post_send(rd->qp, &rd->sq_wr, &bad_wr) != 0) {
851 log_err("fio: ibv_post_send fail");
852 return 1;
853 }
854
855 dprint(FD_IO, "fio: close infomation sent success\n");
856 rdma_poll_wait(td, IBV_WC_SEND);
857 }
858
859 if (rd->is_client == 1)
860 rdma_disconnect(rd->cm_id);
861 else {
862 rdma_disconnect(rd->child_cm_id);
863/* rdma_disconnect(rd->cm_id); */
864 }
865
866/* if (get_next_channel_event(td, rd->cm_channel, RDMA_CM_EVENT_DISCONNECTED) != 0)
867 {
868 log_err("fio: wait for RDMA_CM_EVENT_DISCONNECTED\n");
869 return 1;
870 }*/
871
872 ibv_destroy_qp(rd->qp);
873 ibv_destroy_cq(rd->cq);
874
875 if (rd->is_client == 1)
876 rdma_destroy_id(rd->cm_id);
877 else {
878 rdma_destroy_id(rd->child_cm_id);
879 rdma_destroy_id(rd->cm_id);
880 }
881
882 ibv_destroy_comp_channel(rd->channel);
883 ibv_dealloc_pd(rd->pd);
884
885 return 0;
886}
887
888static int fio_rdmaio_setup_connect(struct thread_data *td, const char *host,
889 unsigned short port)
890{
891 struct rdmaio_data *rd = td->io_ops->data;
892 struct ibv_recv_wr *bad_wr;
893
894 rd->addr.sin_family = AF_INET;
895 rd->addr.sin_port = htons(port);
896
897 if (inet_aton(host, &rd->addr.sin_addr) != 1) {
898 struct hostent *hent;
899
900 hent = gethostbyname(host);
901 if (!hent) {
902 td_verror(td, errno, "gethostbyname");
903 return 1;
904 }
905
906 memcpy(&rd->addr.sin_addr, hent->h_addr, 4);
907 }
908
909 /* resolve route */
910 if (rdma_resolve_addr(rd->cm_id, NULL,
911 (struct sockaddr *)&rd->addr, 2000) != 0) {
912 log_err("fio: rdma_resolve_addr");
913 return 1;
914 }
915
916 if (get_next_channel_event
917 (td, rd->cm_channel, RDMA_CM_EVENT_ADDR_RESOLVED)
918 != 0) {
919 log_err("fio: get_next_channel_event");
920 return 1;
921 }
922
923 /* resolve route */
924 if (rdma_resolve_route(rd->cm_id, 2000) != 0) {
925 log_err("fio: rdma_resolve_route");
926 return 1;
927 }
928
929 if (get_next_channel_event
930 (td, rd->cm_channel, RDMA_CM_EVENT_ROUTE_RESOLVED) != 0) {
931 log_err("fio: get_next_channel_event");
932 return 1;
933 }
934
935 /* create qp and buffer */
936 if (fio_rdmaio_setup_qp(td) != 0)
937 return 1;
938
939 if (fio_rdmaio_setup_control_msg_buffers(td) != 0)
940 return 1;
941
942 /* post recv buf */
943 if (ibv_post_recv(rd->qp, &rd->rq_wr, &bad_wr) != 0) {
944 log_err("fio: ibv_post_recv fail\n");
945 return 1;
946 }
947
948 return 0;
949}
950
951static int fio_rdmaio_setup_listen(struct thread_data *td, short port)
952{
953 struct rdmaio_data *rd = td->io_ops->data;
954 struct ibv_recv_wr *bad_wr;
955
956 rd->addr.sin_family = AF_INET;
957 rd->addr.sin_addr.s_addr = htonl(INADDR_ANY);
958 rd->addr.sin_port = htons(port);
959
960 /* rdma_listen */
961 if (rdma_bind_addr(rd->cm_id, (struct sockaddr *)&rd->addr) != 0) {
962 log_err("fio: rdma_bind_addr fail\n");
963 return 1;
964 }
965
966 if (rdma_listen(rd->cm_id, 3) != 0) {
967 log_err("fio: rdma_listen fail\n");
968 return 1;
969 }
970
971 /* wait for CONNECT_REQUEST */
972 if (get_next_channel_event
973 (td, rd->cm_channel, RDMA_CM_EVENT_CONNECT_REQUEST) != 0) {
974 log_err("fio: wait for RDMA_CM_EVENT_CONNECT_REQUEST\n");
975 return 1;
976 }
977
978 if (fio_rdmaio_setup_qp(td) != 0)
979 return 1;
980
981 if (fio_rdmaio_setup_control_msg_buffers(td) != 0)
982 return 1;
983
984 /* post recv buf */
985 if (ibv_post_recv(rd->qp, &rd->rq_wr, &bad_wr) != 0) {
986 log_err("fio: ibv_post_recv fail\n");
987 return 1;
988 }
989
990 return 0;
991}
992
993static int fio_rdmaio_init(struct thread_data *td)
994{
995 struct rdmaio_data *rd = td->io_ops->data;
996 unsigned int port;
997 char host[64], buf[128];
998 char *sep, *portp, *modep;
999 int ret;
1000 struct rlimit rl;
1001
1002 if (td_rw(td)) {
1003 log_err("fio: rdma connections must be read OR write\n");
1004 return 1;
1005 }
1006 if (td_random(td)) {
1007 log_err("fio: RDMA network IO can't be random\n");
1008 return 1;
1009 }
1010
1011 /* check RLIMIT_MEMLOCK */
1012 if (getrlimit(RLIMIT_MEMLOCK, &rl) != 0) {
1013 log_err("fio: getrlimit fail: %d(%s)\n",
1014 errno, strerror(errno));
1015 return 1;
1016 }
1017
1018 /* soft limit */
1019 if ((rl.rlim_cur != RLIM_INFINITY)
1020 && (rl.rlim_cur < td->orig_buffer_size)) {
1021 log_err("fio: soft RLIMIT_MEMLOCK is: %ld\n", rl.rlim_cur);
1022 log_err("fio: total block size is: %ld\n",
1023 td->orig_buffer_size);
1024 /* try to set larger RLIMIT_MEMLOCK */
1025 rl.rlim_cur = rl.rlim_max;
1026 if (setrlimit(RLIMIT_MEMLOCK, &rl) != 0) {
1027 log_err("fio: setrlimit fail: %d(%s)\n",
1028 errno, strerror(errno));
1029 log_err("fio: you may try enlarge MEMLOCK by root\n");
1030 log_err("# ulimit -l unlimited\n");
1031 return 1;
1032 }
1033 }
1034
1035 strcpy(buf, td->o.filename);
1036
1037 sep = strchr(buf, '/');
1038 if (!sep)
1039 goto bad_host;
1040
1041 *sep = '\0';
1042 sep++;
1043 strcpy(host, buf);
1044 if (!strlen(host))
1045 goto bad_host;
1046
1047 modep = NULL;
1048 portp = sep;
1049 sep = strchr(portp, '/');
1050 if (sep) {
1051 *sep = '\0';
1052 modep = sep + 1;
1053 }
1054
1055 port = strtol(portp, NULL, 10);
1056 if (!port || port > 65535)
1057 goto bad_host;
1058
1059 if (modep) {
1060 if (!strncmp("rdma_write", modep, strlen(modep)) ||
1061 !strncmp("RDMA_WRITE", modep, strlen(modep)))
1062 rd->rdma_protocol = FIO_RDMA_MEM_WRITE;
1063 else if (!strncmp("rdma_read", modep, strlen(modep)) ||
1064 !strncmp("RDMA_READ", modep, strlen(modep)))
1065 rd->rdma_protocol = FIO_RDMA_MEM_READ;
1066 else if (!strncmp("send", modep, strlen(modep)) ||
1067 !strncmp("SEND", modep, strlen(modep)))
1068 rd->rdma_protocol = FIO_RDMA_CHA_SEND;
1069 else
1070 goto bad_host;
1071 } else
1072 rd->rdma_protocol = FIO_RDMA_MEM_WRITE;
1073
1074 rd->cq_event_num = 0;
1075
1076 rd->cm_channel = rdma_create_event_channel();
1077 if (!rd->cm_channel) {
1078 log_err("fio: rdma_create_event_channel fail\n");
1079 return 1;
1080 }
1081
1082 ret = rdma_create_id(rd->cm_channel, &rd->cm_id, rd, RDMA_PS_TCP);
1083 if (ret) {
1084 log_err("fio: rdma_create_id fail\n");
1085 return 1;
1086 }
1087
1088 if ((rd->rdma_protocol == FIO_RDMA_MEM_WRITE) ||
1089 (rd->rdma_protocol == FIO_RDMA_MEM_READ)) {
1090 rd->rmt_us =
1091 malloc(FIO_RDMA_MAX_IO_DEPTH * sizeof(struct remote_u));
1092 memset(rd->rmt_us, 0,
1093 FIO_RDMA_MAX_IO_DEPTH * sizeof(struct remote_u));
1094 rd->rmt_nr = 0;
1095 }
1096
1097 rd->io_us_queued = malloc(td->o.iodepth * sizeof(struct io_u *));
1098 memset(rd->io_us_queued, 0, td->o.iodepth * sizeof(struct io_u *));
1099 rd->io_u_queued_nr = 0;
1100
1101 rd->io_us_flight = malloc(td->o.iodepth * sizeof(struct io_u *));
1102 memset(rd->io_us_flight, 0, td->o.iodepth * sizeof(struct io_u *));
1103 rd->io_u_flight_nr = 0;
1104
1105 rd->io_us_completed = malloc(td->o.iodepth * sizeof(struct io_u *));
1106 memset(rd->io_us_completed, 0, td->o.iodepth * sizeof(struct io_u *));
1107 rd->io_u_completed_nr = 0;
1108
1109 if (td_read(td)) { /* READ as the server */
1110 rd->is_client = 0;
1111 /* server rd->rdma_buf_len will be setup after got request */
1112 ret = fio_rdmaio_setup_listen(td, port);
1113 } else { /* WRITE as the client */
1114 rd->is_client = 1;
1115 ret = fio_rdmaio_setup_connect(td, host, port);
1116 }
1117
1118 struct flist_head *entry;
1119 unsigned int max_bs;
1120 max_bs = max(td->o.max_bs[DDIR_READ], td->o.max_bs[DDIR_WRITE]);
1121 /* register each io_u in the free list */
1122 int i = 0;
1123 flist_for_each(entry, &td->io_u_freelist) {
1124 struct io_u *io_u = flist_entry(entry, struct io_u, list);
1125
1126 io_u->engine_data = malloc(sizeof(struct rdma_io_u_data));
1127 memset(io_u->engine_data, 0, sizeof(struct rdma_io_u_data));
1128 ((struct rdma_io_u_data *)io_u->engine_data)->wr_id = i;
1129
1130 io_u->mr = ibv_reg_mr(rd->pd, io_u->buf, max_bs,
1131 IBV_ACCESS_LOCAL_WRITE |
1132 IBV_ACCESS_REMOTE_READ |
1133 IBV_ACCESS_REMOTE_WRITE);
1134 if (io_u->mr == NULL) {
1135 log_err("fio: ibv_reg_mr io_u failed\n");
1136 return 1;
1137 }
1138
1139 rd->send_buf.rmt_us[i].buf =
1140 htonll((uint64_t) (unsigned long)io_u->buf);
1141 rd->send_buf.rmt_us[i].rkey = htonl(io_u->mr->rkey);
1142 rd->send_buf.rmt_us[i].size = htonl(max_bs);
1143
1144/* log_info("fio: Send rkey %x addr %" PRIx64 " len %d to client\n",
1145 io_u->mr->rkey, io_u->buf, max_bs); */
1146 i++;
1147 }
1148
1149 rd->send_buf.nr = htonl(i);
1150
1151 return ret;
1152bad_host:
1153 log_err("fio: bad rdma host/port/protocol: %s\n", td->o.filename);
1154 return 1;
1155}
1156
1157static void fio_rdmaio_cleanup(struct thread_data *td)
1158{
1159 struct rdmaio_data *rd = td->io_ops->data;
1160
1161 if (rd) {
1162/* if (nd->listenfd != -1)
1163 close(nd->listenfd);
1164 if (nd->pipes[0] != -1)
1165 close(nd->pipes[0]);
1166 if (nd->pipes[1] != -1)
1167 close(nd->pipes[1]);
1168*/
1169 free(rd);
1170 }
1171}
1172
1173static int fio_rdmaio_setup(struct thread_data *td)
1174{
1175 struct rdmaio_data *rd;
1176
1177 if (!td->io_ops->data) {
1178 rd = malloc(sizeof(*rd));;
1179
1180 memset(rd, 0, sizeof(*rd));
1181 td->io_ops->data = rd;
1182 }
1183
1184 return 0;
1185}
1186
1187static struct ioengine_ops ioengine_rw = {
1188 .name = "rdma",
1189 .version = FIO_IOOPS_VERSION,
1190 .setup = fio_rdmaio_setup,
1191 .init = fio_rdmaio_init,
1192 .prep = fio_rdmaio_prep,
1193 .queue = fio_rdmaio_queue,
1194 .commit = fio_rdmaio_commit,
1195 .getevents = fio_rdmaio_getevents,
1196 .event = fio_rdmaio_event,
1197 .cleanup = fio_rdmaio_cleanup,
1198 .open_file = fio_rdmaio_open_file,
1199 .close_file = fio_rdmaio_close_file,
1200 .flags = FIO_DISKLESSIO | FIO_UNIDIR | FIO_PIPEIO,
1201};
1202
1203#else /* FIO_HAVE_RDMA */
1204
1205static int fio_rdmaio_open_file(struct thread_data *td, struct fio_file *f)
1206{
1207 return 0;
1208}
1209
1210static int fio_rdmaio_close_file(struct thread_data *td, struct fio_file *f)
1211{
1212 return 0;
1213}
1214
1215static int fio_rdmaio_queue(struct thread_data *td, struct io_u *io_u)
1216{
1217 return FIO_Q_COMPLETED;
1218}
1219
1220static int fio_rdmaio_init(struct thread_data fio_unused * td)
1221{
1222 log_err("fio: rdma(librdmacm libibverbs) not available\n");
1223 log_err(" You haven't compiled rdma ioengine into fio.\n");
1224 log_err(" If you want to try rdma ioengine,\n");
1225 log_err(" make sure OFED is installed,\n");
1226 log_err(" $ ofed_info\n");
1227 log_err(" then try to make fio as follows:\n");
1228 log_err(" $ export EXTFLAGS=\"-DFIO_HAVE_RDMA\"\n");
1229 log_err(" $ export EXTLIBS=\"-libverbs -lrdmacm\"\n");
1230 log_err(" $ make clean && make\n");
1231 return 1;
1232}
1233
1234static struct ioengine_ops ioengine_rw = {
1235 .name = "rdma",
1236 .version = FIO_IOOPS_VERSION,
1237 .init = fio_rdmaio_init,
1238 .queue = fio_rdmaio_queue,
1239 .open_file = fio_rdmaio_open_file,
1240 .close_file = fio_rdmaio_close_file,
1241 .flags = FIO_SYNCIO | FIO_DISKLESSIO | FIO_UNIDIR | FIO_PIPEIO,
1242};
1243
1244#endif
1245
1246static void fio_init fio_rdmaio_register(void)
1247{
1248 register_ioengine(&ioengine_rw);
1249}
1250
1251static void fio_exit fio_rdmaio_unregister(void)
1252{
1253 unregister_ioengine(&ioengine_rw);
1254}
fio - Flexible IO Tester