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