Commit | Line | Data |
---|---|---|
21b8aee8 | 1 | /* |
85286c5c | 2 | * RDMA I/O engine |
21b8aee8 | 3 | * |
85286c5c BVA |
4 | * RDMA I/O engine based on the IB verbs and RDMA/CM user space libraries. |
5 | * Supports both RDMA memory semantics and channel semantics | |
6 | * for the InfiniBand, RoCE and iWARP protocols. | |
21b8aee8 | 7 | * |
0ac5d398 | 8 | * You will need the Linux RDMA software installed, either |
85286c5c | 9 | * from your Linux distributor or directly from openfabrics.org: |
21b8aee8 | 10 | * |
11 | * http://www.openfabrics.org/downloads/OFED/ | |
12 | * | |
7d7803fa YR |
13 | * Exchanging steps of RDMA ioengine control messages: |
14 | * 1. client side sends test mode (RDMA_WRITE/RDMA_READ/SEND) | |
15 | * to server side. | |
222757cc | 16 | * 2. server side parses test mode, and sends back confirmation |
7d7803fa | 17 | * to client side. In RDMA WRITE/READ test, this confirmation |
222757cc | 18 | * includes memory information, such as rkey, address. |
7d7803fa | 19 | * 3. client side initiates test loop. |
222757cc | 20 | * 4. In RDMA WRITE/READ test, client side sends a completion |
7d7803fa | 21 | * notification to server side. Server side updates its |
222757cc | 22 | * td->done as true. |
7d7803fa | 23 | * |
21b8aee8 | 24 | */ |
25 | #include <stdio.h> | |
26 | #include <stdlib.h> | |
27 | #include <unistd.h> | |
28 | #include <errno.h> | |
29 | #include <assert.h> | |
30 | #include <netinet/in.h> | |
31 | #include <arpa/inet.h> | |
32 | #include <netdb.h> | |
8393ca93 | 33 | #include <poll.h> |
21b8aee8 | 34 | #include <sys/types.h> |
35 | #include <sys/socket.h> | |
36 | #include <sys/time.h> | |
37 | #include <sys/resource.h> | |
38 | ||
21b8aee8 | 39 | #include <pthread.h> |
40 | #include <inttypes.h> | |
41 | ||
42 | #include "../fio.h" | |
ee88d056 | 43 | #include "../hash.h" |
d220c761 | 44 | #include "../optgroup.h" |
21b8aee8 | 45 | |
21b8aee8 | 46 | #include <rdma/rdma_cma.h> |
21b8aee8 | 47 | |
7d7803fa | 48 | #define FIO_RDMA_MAX_IO_DEPTH 512 |
21b8aee8 | 49 | |
50 | enum rdma_io_mode { | |
51 | FIO_RDMA_UNKNOWN = 0, | |
52 | FIO_RDMA_MEM_WRITE, | |
53 | FIO_RDMA_MEM_READ, | |
54 | FIO_RDMA_CHA_SEND, | |
55 | FIO_RDMA_CHA_RECV | |
56 | }; | |
57 | ||
cdf91594 LG |
58 | struct rdmaio_options { |
59 | struct thread_data *td; | |
60 | unsigned int port; | |
61 | enum rdma_io_mode verb; | |
2f28bb35 | 62 | char *bindname; |
cdf91594 LG |
63 | }; |
64 | ||
65 | static 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 | ||
75 | static 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 | }, | |
2f28bb35 SB |
85 | { |
86 | .name = "bindname", | |
87 | .lname = "rdma engine bindname", | |
88 | .type = FIO_OPT_STR_STORE, | |
89 | .off1 = offsetof(struct rdmaio_options, bindname), | |
90 | .help = "Bind for RDMA IO engine", | |
91 | .def = "", | |
92 | .category = FIO_OPT_C_ENGINE, | |
93 | .group = FIO_OPT_G_RDMA, | |
94 | }, | |
cdf91594 LG |
95 | { |
96 | .name = "port", | |
97 | .lname = "rdma engine port", | |
98 | .type = FIO_OPT_INT, | |
99 | .off1 = offsetof(struct rdmaio_options, port), | |
100 | .minval = 1, | |
101 | .maxval = 65535, | |
102 | .help = "Port to use for RDMA connections", | |
103 | .category = FIO_OPT_C_ENGINE, | |
104 | .group = FIO_OPT_G_RDMA, | |
105 | }, | |
106 | { | |
107 | .name = "verb", | |
108 | .lname = "RDMA engine verb", | |
109 | .alias = "proto", | |
110 | .type = FIO_OPT_STR, | |
111 | .off1 = offsetof(struct rdmaio_options, verb), | |
112 | .help = "RDMA engine verb", | |
113 | .def = "write", | |
114 | .posval = { | |
115 | { .ival = "write", | |
116 | .oval = FIO_RDMA_MEM_WRITE, | |
117 | .help = "Memory Write", | |
118 | }, | |
119 | { .ival = "read", | |
120 | .oval = FIO_RDMA_MEM_READ, | |
121 | .help = "Memory Read", | |
122 | }, | |
123 | { .ival = "send", | |
124 | .oval = FIO_RDMA_CHA_SEND, | |
125 | .help = "Posted Send", | |
126 | }, | |
127 | { .ival = "recv", | |
128 | .oval = FIO_RDMA_CHA_RECV, | |
ee386b78 | 129 | .help = "Posted Receive", |
cdf91594 LG |
130 | }, |
131 | }, | |
132 | .category = FIO_OPT_C_ENGINE, | |
133 | .group = FIO_OPT_G_RDMA, | |
134 | }, | |
135 | { | |
136 | .name = NULL, | |
137 | }, | |
138 | }; | |
139 | ||
21b8aee8 | 140 | struct remote_u { |
141 | uint64_t buf; | |
142 | uint32_t rkey; | |
143 | uint32_t size; | |
144 | }; | |
145 | ||
146 | struct rdma_info_blk { | |
147 | uint32_t mode; /* channel semantic or memory semantic */ | |
148 | uint32_t nr; /* client: io depth | |
149 | server: number of records for memory semantic | |
150 | */ | |
c4421263 | 151 | uint32_t max_bs; /* maximum block size */ |
76cc5224 | 152 | struct remote_u rmt_us[FIO_RDMA_MAX_IO_DEPTH]; |
21b8aee8 | 153 | }; |
154 | ||
155 | struct rdma_io_u_data { | |
156 | uint64_t wr_id; | |
157 | struct ibv_send_wr sq_wr; | |
158 | struct ibv_recv_wr rq_wr; | |
159 | struct ibv_sge rdma_sgl; | |
160 | }; | |
161 | ||
162 | struct rdmaio_data { | |
163 | int is_client; | |
164 | enum rdma_io_mode rdma_protocol; | |
165 | char host[64]; | |
166 | struct sockaddr_in addr; | |
167 | ||
168 | struct ibv_recv_wr rq_wr; | |
169 | struct ibv_sge recv_sgl; | |
170 | struct rdma_info_blk recv_buf; | |
171 | struct ibv_mr *recv_mr; | |
172 | ||
173 | struct ibv_send_wr sq_wr; | |
174 | struct ibv_sge send_sgl; | |
175 | struct rdma_info_blk send_buf; | |
176 | struct ibv_mr *send_mr; | |
177 | ||
178 | struct ibv_comp_channel *channel; | |
179 | struct ibv_cq *cq; | |
180 | struct ibv_pd *pd; | |
181 | struct ibv_qp *qp; | |
182 | ||
183 | pthread_t cmthread; | |
184 | struct rdma_event_channel *cm_channel; | |
185 | struct rdma_cm_id *cm_id; | |
186 | struct rdma_cm_id *child_cm_id; | |
187 | ||
188 | int cq_event_num; | |
189 | ||
190 | struct remote_u *rmt_us; | |
191 | int rmt_nr; | |
192 | struct io_u **io_us_queued; | |
193 | int io_u_queued_nr; | |
194 | struct io_u **io_us_flight; | |
195 | int io_u_flight_nr; | |
196 | struct io_u **io_us_completed; | |
197 | int io_u_completed_nr; | |
ea6209ad JA |
198 | |
199 | struct frand_state rand_state; | |
21b8aee8 | 200 | }; |
201 | ||
202 | static int client_recv(struct thread_data *td, struct ibv_wc *wc) | |
203 | { | |
565e784d | 204 | struct rdmaio_data *rd = td->io_ops_data; |
c4421263 | 205 | unsigned int max_bs; |
21b8aee8 | 206 | |
207 | if (wc->byte_len != sizeof(rd->recv_buf)) { | |
b6cf38f0 | 208 | log_err("Received bogus data, size %d\n", wc->byte_len); |
21b8aee8 | 209 | return 1; |
210 | } | |
211 | ||
c4421263 LG |
212 | max_bs = max(td->o.max_bs[DDIR_READ], td->o.max_bs[DDIR_WRITE]); |
213 | if (max_bs > ntohl(rd->recv_buf.max_bs)) { | |
214 | log_err("fio: Server's block size (%d) must be greater than or " | |
215 | "equal to the client's block size (%d)!\n", | |
216 | ntohl(rd->recv_buf.max_bs), max_bs); | |
217 | return 1; | |
218 | } | |
219 | ||
21b8aee8 | 220 | /* store mr info for MEMORY semantic */ |
221 | if ((rd->rdma_protocol == FIO_RDMA_MEM_WRITE) || | |
222 | (rd->rdma_protocol == FIO_RDMA_MEM_READ)) { | |
223 | /* struct flist_head *entry; */ | |
224 | int i = 0; | |
225 | ||
226 | rd->rmt_nr = ntohl(rd->recv_buf.nr); | |
227 | ||
228 | for (i = 0; i < rd->rmt_nr; i++) { | |
5a8a6a03 YK |
229 | rd->rmt_us[i].buf = __be64_to_cpu( |
230 | rd->recv_buf.rmt_us[i].buf); | |
21b8aee8 | 231 | rd->rmt_us[i].rkey = ntohl(rd->recv_buf.rmt_us[i].rkey); |
232 | rd->rmt_us[i].size = ntohl(rd->recv_buf.rmt_us[i].size); | |
233 | ||
234 | dprint(FD_IO, | |
235 | "fio: Received rkey %x addr %" PRIx64 | |
236 | " len %d from peer\n", rd->rmt_us[i].rkey, | |
237 | rd->rmt_us[i].buf, rd->rmt_us[i].size); | |
238 | } | |
239 | } | |
240 | ||
241 | return 0; | |
242 | } | |
243 | ||
244 | static int server_recv(struct thread_data *td, struct ibv_wc *wc) | |
245 | { | |
565e784d | 246 | struct rdmaio_data *rd = td->io_ops_data; |
c4421263 | 247 | unsigned int max_bs; |
21b8aee8 | 248 | |
76cc5224 | 249 | if (wc->wr_id == FIO_RDMA_MAX_IO_DEPTH) { |
21b8aee8 | 250 | rd->rdma_protocol = ntohl(rd->recv_buf.mode); |
251 | ||
252 | /* CHANNEL semantic, do nothing */ | |
253 | if (rd->rdma_protocol == FIO_RDMA_CHA_SEND) | |
254 | rd->rdma_protocol = FIO_RDMA_CHA_RECV; | |
c4421263 LG |
255 | |
256 | max_bs = max(td->o.max_bs[DDIR_READ], td->o.max_bs[DDIR_WRITE]); | |
257 | if (max_bs < ntohl(rd->recv_buf.max_bs)) { | |
258 | log_err("fio: Server's block size (%d) must be greater than or " | |
259 | "equal to the client's block size (%d)!\n", | |
260 | ntohl(rd->recv_buf.max_bs), max_bs); | |
261 | return 1; | |
262 | } | |
263 | ||
21b8aee8 | 264 | } |
265 | ||
266 | return 0; | |
267 | } | |
268 | ||
269 | static int cq_event_handler(struct thread_data *td, enum ibv_wc_opcode opcode) | |
270 | { | |
565e784d | 271 | struct rdmaio_data *rd = td->io_ops_data; |
21b8aee8 | 272 | struct ibv_wc wc; |
273 | struct rdma_io_u_data *r_io_u_d; | |
274 | int ret; | |
275 | int compevnum = 0; | |
276 | int i; | |
277 | ||
278 | while ((ret = ibv_poll_cq(rd->cq, 1, &wc)) == 1) { | |
279 | ret = 0; | |
280 | compevnum++; | |
281 | ||
282 | if (wc.status) { | |
283 | log_err("fio: cq completion status %d(%s)\n", | |
284 | wc.status, ibv_wc_status_str(wc.status)); | |
285 | return -1; | |
286 | } | |
287 | ||
288 | switch (wc.opcode) { | |
289 | ||
290 | case IBV_WC_RECV: | |
291 | if (rd->is_client == 1) | |
c4421263 | 292 | ret = client_recv(td, &wc); |
21b8aee8 | 293 | else |
c4421263 LG |
294 | ret = server_recv(td, &wc); |
295 | ||
296 | if (ret) | |
297 | return -1; | |
21b8aee8 | 298 | |
76cc5224 | 299 | if (wc.wr_id == FIO_RDMA_MAX_IO_DEPTH) |
21b8aee8 | 300 | break; |
301 | ||
302 | for (i = 0; i < rd->io_u_flight_nr; i++) { | |
303 | r_io_u_d = rd->io_us_flight[i]->engine_data; | |
304 | ||
305 | if (wc.wr_id == r_io_u_d->rq_wr.wr_id) { | |
306 | rd->io_us_flight[i]->resid = | |
307 | rd->io_us_flight[i]->buflen | |
308 | - wc.byte_len; | |
309 | ||
310 | rd->io_us_flight[i]->error = 0; | |
311 | ||
312 | rd->io_us_completed[rd-> | |
313 | io_u_completed_nr] | |
314 | = rd->io_us_flight[i]; | |
315 | rd->io_u_completed_nr++; | |
316 | break; | |
317 | } | |
318 | } | |
319 | if (i == rd->io_u_flight_nr) | |
e07f72d3 | 320 | log_err("fio: recv wr %" PRId64 " not found\n", |
21b8aee8 | 321 | wc.wr_id); |
322 | else { | |
323 | /* put the last one into middle of the list */ | |
324 | rd->io_us_flight[i] = | |
325 | rd->io_us_flight[rd->io_u_flight_nr - 1]; | |
326 | rd->io_u_flight_nr--; | |
327 | } | |
328 | ||
329 | break; | |
330 | ||
331 | case IBV_WC_SEND: | |
332 | case IBV_WC_RDMA_WRITE: | |
333 | case IBV_WC_RDMA_READ: | |
76cc5224 | 334 | if (wc.wr_id == FIO_RDMA_MAX_IO_DEPTH) |
21b8aee8 | 335 | break; |
336 | ||
337 | for (i = 0; i < rd->io_u_flight_nr; i++) { | |
338 | r_io_u_d = rd->io_us_flight[i]->engine_data; | |
339 | ||
340 | if (wc.wr_id == r_io_u_d->sq_wr.wr_id) { | |
341 | rd->io_us_completed[rd-> | |
342 | io_u_completed_nr] | |
343 | = rd->io_us_flight[i]; | |
344 | rd->io_u_completed_nr++; | |
345 | break; | |
346 | } | |
347 | } | |
348 | if (i == rd->io_u_flight_nr) | |
e07f72d3 | 349 | log_err("fio: send wr %" PRId64 " not found\n", |
21b8aee8 | 350 | wc.wr_id); |
351 | else { | |
352 | /* put the last one into middle of the list */ | |
353 | rd->io_us_flight[i] = | |
354 | rd->io_us_flight[rd->io_u_flight_nr - 1]; | |
355 | rd->io_u_flight_nr--; | |
356 | } | |
357 | ||
358 | break; | |
359 | ||
360 | default: | |
361 | log_info("fio: unknown completion event %d\n", | |
362 | wc.opcode); | |
363 | return -1; | |
364 | } | |
365 | rd->cq_event_num++; | |
366 | } | |
c4421263 | 367 | |
21b8aee8 | 368 | if (ret) { |
369 | log_err("fio: poll error %d\n", ret); | |
370 | return 1; | |
371 | } | |
372 | ||
373 | return compevnum; | |
374 | } | |
375 | ||
376 | /* | |
377 | * Return -1 for error and 'nr events' for a positive number | |
378 | * of events | |
379 | */ | |
380 | static int rdma_poll_wait(struct thread_data *td, enum ibv_wc_opcode opcode) | |
381 | { | |
565e784d | 382 | struct rdmaio_data *rd = td->io_ops_data; |
21b8aee8 | 383 | struct ibv_cq *ev_cq; |
384 | void *ev_ctx; | |
385 | int ret; | |
386 | ||
387 | if (rd->cq_event_num > 0) { /* previous left */ | |
388 | rd->cq_event_num--; | |
389 | return 0; | |
390 | } | |
391 | ||
392 | again: | |
393 | if (ibv_get_cq_event(rd->channel, &ev_cq, &ev_ctx) != 0) { | |
394 | log_err("fio: Failed to get cq event!\n"); | |
395 | return -1; | |
396 | } | |
397 | if (ev_cq != rd->cq) { | |
398 | log_err("fio: Unknown CQ!\n"); | |
399 | return -1; | |
400 | } | |
401 | if (ibv_req_notify_cq(rd->cq, 0) != 0) { | |
402 | log_err("fio: Failed to set notify!\n"); | |
403 | return -1; | |
404 | } | |
405 | ||
406 | ret = cq_event_handler(td, opcode); | |
c4421263 | 407 | if (ret == 0) |
21b8aee8 | 408 | goto again; |
409 | ||
410 | ibv_ack_cq_events(rd->cq, ret); | |
411 | ||
412 | rd->cq_event_num--; | |
413 | ||
414 | return ret; | |
415 | } | |
416 | ||
417 | static int fio_rdmaio_setup_qp(struct thread_data *td) | |
418 | { | |
565e784d | 419 | struct rdmaio_data *rd = td->io_ops_data; |
21b8aee8 | 420 | struct ibv_qp_init_attr init_attr; |
421 | int qp_depth = td->o.iodepth * 2; /* 2 times of io depth */ | |
422 | ||
423 | if (rd->is_client == 0) | |
424 | rd->pd = ibv_alloc_pd(rd->child_cm_id->verbs); | |
425 | else | |
426 | rd->pd = ibv_alloc_pd(rd->cm_id->verbs); | |
222757cc | 427 | |
21b8aee8 | 428 | if (rd->pd == NULL) { |
2a442a30 | 429 | log_err("fio: ibv_alloc_pd fail: %m\n"); |
21b8aee8 | 430 | return 1; |
431 | } | |
432 | ||
433 | if (rd->is_client == 0) | |
434 | rd->channel = ibv_create_comp_channel(rd->child_cm_id->verbs); | |
435 | else | |
436 | rd->channel = ibv_create_comp_channel(rd->cm_id->verbs); | |
437 | if (rd->channel == NULL) { | |
2a442a30 | 438 | log_err("fio: ibv_create_comp_channel fail: %m\n"); |
21b8aee8 | 439 | goto err1; |
440 | } | |
441 | ||
442 | if (qp_depth < 16) | |
443 | qp_depth = 16; | |
444 | ||
445 | if (rd->is_client == 0) | |
446 | rd->cq = ibv_create_cq(rd->child_cm_id->verbs, | |
447 | qp_depth, rd, rd->channel, 0); | |
448 | else | |
449 | rd->cq = ibv_create_cq(rd->cm_id->verbs, | |
450 | qp_depth, rd, rd->channel, 0); | |
451 | if (rd->cq == NULL) { | |
2a442a30 | 452 | log_err("fio: ibv_create_cq failed: %m\n"); |
21b8aee8 | 453 | goto err2; |
454 | } | |
455 | ||
456 | if (ibv_req_notify_cq(rd->cq, 0) != 0) { | |
2a442a30 | 457 | log_err("fio: ibv_req_notify_cq failed: %m\n"); |
21b8aee8 | 458 | goto err3; |
459 | } | |
460 | ||
461 | /* create queue pair */ | |
462 | memset(&init_attr, 0, sizeof(init_attr)); | |
463 | init_attr.cap.max_send_wr = qp_depth; | |
464 | init_attr.cap.max_recv_wr = qp_depth; | |
465 | init_attr.cap.max_recv_sge = 1; | |
466 | init_attr.cap.max_send_sge = 1; | |
467 | init_attr.qp_type = IBV_QPT_RC; | |
468 | init_attr.send_cq = rd->cq; | |
469 | init_attr.recv_cq = rd->cq; | |
470 | ||
471 | if (rd->is_client == 0) { | |
472 | if (rdma_create_qp(rd->child_cm_id, rd->pd, &init_attr) != 0) { | |
2a442a30 | 473 | log_err("fio: rdma_create_qp failed: %m\n"); |
21b8aee8 | 474 | goto err3; |
475 | } | |
476 | rd->qp = rd->child_cm_id->qp; | |
477 | } else { | |
478 | if (rdma_create_qp(rd->cm_id, rd->pd, &init_attr) != 0) { | |
2a442a30 | 479 | log_err("fio: rdma_create_qp failed: %m\n"); |
21b8aee8 | 480 | goto err3; |
481 | } | |
482 | rd->qp = rd->cm_id->qp; | |
483 | } | |
484 | ||
485 | return 0; | |
486 | ||
487 | err3: | |
488 | ibv_destroy_cq(rd->cq); | |
489 | err2: | |
490 | ibv_destroy_comp_channel(rd->channel); | |
491 | err1: | |
492 | ibv_dealloc_pd(rd->pd); | |
493 | ||
494 | return 1; | |
495 | } | |
496 | ||
497 | static int fio_rdmaio_setup_control_msg_buffers(struct thread_data *td) | |
498 | { | |
565e784d | 499 | struct rdmaio_data *rd = td->io_ops_data; |
21b8aee8 | 500 | |
501 | rd->recv_mr = ibv_reg_mr(rd->pd, &rd->recv_buf, sizeof(rd->recv_buf), | |
502 | IBV_ACCESS_LOCAL_WRITE); | |
503 | if (rd->recv_mr == NULL) { | |
2a442a30 | 504 | log_err("fio: recv_buf reg_mr failed: %m\n"); |
21b8aee8 | 505 | return 1; |
506 | } | |
507 | ||
508 | rd->send_mr = ibv_reg_mr(rd->pd, &rd->send_buf, sizeof(rd->send_buf), | |
509 | 0); | |
510 | if (rd->send_mr == NULL) { | |
2a442a30 | 511 | log_err("fio: send_buf reg_mr failed: %m\n"); |
21b8aee8 | 512 | ibv_dereg_mr(rd->recv_mr); |
513 | return 1; | |
514 | } | |
515 | ||
516 | /* setup work request */ | |
517 | /* recv wq */ | |
518 | rd->recv_sgl.addr = (uint64_t) (unsigned long)&rd->recv_buf; | |
222757cc | 519 | rd->recv_sgl.length = sizeof(rd->recv_buf); |
21b8aee8 | 520 | rd->recv_sgl.lkey = rd->recv_mr->lkey; |
521 | rd->rq_wr.sg_list = &rd->recv_sgl; | |
522 | rd->rq_wr.num_sge = 1; | |
76cc5224 | 523 | rd->rq_wr.wr_id = FIO_RDMA_MAX_IO_DEPTH; |
21b8aee8 | 524 | |
525 | /* send wq */ | |
526 | rd->send_sgl.addr = (uint64_t) (unsigned long)&rd->send_buf; | |
222757cc | 527 | rd->send_sgl.length = sizeof(rd->send_buf); |
21b8aee8 | 528 | rd->send_sgl.lkey = rd->send_mr->lkey; |
529 | ||
530 | rd->sq_wr.opcode = IBV_WR_SEND; | |
531 | rd->sq_wr.send_flags = IBV_SEND_SIGNALED; | |
532 | rd->sq_wr.sg_list = &rd->send_sgl; | |
533 | rd->sq_wr.num_sge = 1; | |
76cc5224 | 534 | rd->sq_wr.wr_id = FIO_RDMA_MAX_IO_DEPTH; |
21b8aee8 | 535 | |
536 | return 0; | |
537 | } | |
538 | ||
539 | static int get_next_channel_event(struct thread_data *td, | |
540 | struct rdma_event_channel *channel, | |
541 | enum rdma_cm_event_type wait_event) | |
542 | { | |
565e784d | 543 | struct rdmaio_data *rd = td->io_ops_data; |
21b8aee8 | 544 | struct rdma_cm_event *event; |
222757cc | 545 | int ret; |
21b8aee8 | 546 | |
547 | ret = rdma_get_cm_event(channel, &event); | |
548 | if (ret) { | |
222757cc | 549 | log_err("fio: rdma_get_cm_event: %d\n", ret); |
21b8aee8 | 550 | return 1; |
551 | } | |
552 | ||
553 | if (event->event != wait_event) { | |
554 | log_err("fio: event is %s instead of %s\n", | |
555 | rdma_event_str(event->event), | |
556 | rdma_event_str(wait_event)); | |
557 | return 1; | |
558 | } | |
559 | ||
560 | switch (event->event) { | |
561 | case RDMA_CM_EVENT_CONNECT_REQUEST: | |
562 | rd->child_cm_id = event->id; | |
563 | break; | |
564 | default: | |
565 | break; | |
566 | } | |
567 | ||
568 | rdma_ack_cm_event(event); | |
569 | ||
570 | return 0; | |
571 | } | |
572 | ||
573 | static int fio_rdmaio_prep(struct thread_data *td, struct io_u *io_u) | |
574 | { | |
565e784d | 575 | struct rdmaio_data *rd = td->io_ops_data; |
21b8aee8 | 576 | struct rdma_io_u_data *r_io_u_d; |
577 | ||
578 | r_io_u_d = io_u->engine_data; | |
579 | ||
580 | switch (rd->rdma_protocol) { | |
581 | case FIO_RDMA_MEM_WRITE: | |
582 | case FIO_RDMA_MEM_READ: | |
583 | r_io_u_d->rdma_sgl.addr = (uint64_t) (unsigned long)io_u->buf; | |
584 | r_io_u_d->rdma_sgl.lkey = io_u->mr->lkey; | |
585 | r_io_u_d->sq_wr.wr_id = r_io_u_d->wr_id; | |
586 | r_io_u_d->sq_wr.send_flags = IBV_SEND_SIGNALED; | |
587 | r_io_u_d->sq_wr.sg_list = &r_io_u_d->rdma_sgl; | |
588 | r_io_u_d->sq_wr.num_sge = 1; | |
589 | break; | |
590 | case FIO_RDMA_CHA_SEND: | |
591 | r_io_u_d->rdma_sgl.addr = (uint64_t) (unsigned long)io_u->buf; | |
592 | r_io_u_d->rdma_sgl.lkey = io_u->mr->lkey; | |
593 | r_io_u_d->rdma_sgl.length = io_u->buflen; | |
594 | r_io_u_d->sq_wr.wr_id = r_io_u_d->wr_id; | |
595 | r_io_u_d->sq_wr.opcode = IBV_WR_SEND; | |
596 | r_io_u_d->sq_wr.send_flags = IBV_SEND_SIGNALED; | |
597 | r_io_u_d->sq_wr.sg_list = &r_io_u_d->rdma_sgl; | |
598 | r_io_u_d->sq_wr.num_sge = 1; | |
599 | break; | |
600 | case FIO_RDMA_CHA_RECV: | |
601 | r_io_u_d->rdma_sgl.addr = (uint64_t) (unsigned long)io_u->buf; | |
602 | r_io_u_d->rdma_sgl.lkey = io_u->mr->lkey; | |
603 | r_io_u_d->rdma_sgl.length = io_u->buflen; | |
604 | r_io_u_d->rq_wr.wr_id = r_io_u_d->wr_id; | |
605 | r_io_u_d->rq_wr.sg_list = &r_io_u_d->rdma_sgl; | |
606 | r_io_u_d->rq_wr.num_sge = 1; | |
607 | break; | |
608 | default: | |
609 | log_err("fio: unknown rdma protocol - %d\n", rd->rdma_protocol); | |
610 | break; | |
611 | } | |
612 | ||
613 | return 0; | |
614 | } | |
615 | ||
616 | static struct io_u *fio_rdmaio_event(struct thread_data *td, int event) | |
617 | { | |
565e784d | 618 | struct rdmaio_data *rd = td->io_ops_data; |
21b8aee8 | 619 | struct io_u *io_u; |
620 | int i; | |
621 | ||
622 | io_u = rd->io_us_completed[0]; | |
222757cc | 623 | for (i = 0; i < rd->io_u_completed_nr - 1; i++) |
21b8aee8 | 624 | rd->io_us_completed[i] = rd->io_us_completed[i + 1]; |
222757cc | 625 | |
21b8aee8 | 626 | rd->io_u_completed_nr--; |
627 | ||
628 | dprint_io_u(io_u, "fio_rdmaio_event"); | |
629 | ||
630 | return io_u; | |
631 | } | |
632 | ||
633 | static int fio_rdmaio_getevents(struct thread_data *td, unsigned int min, | |
1f440ece | 634 | unsigned int max, const struct timespec *t) |
21b8aee8 | 635 | { |
565e784d | 636 | struct rdmaio_data *rd = td->io_ops_data; |
21b8aee8 | 637 | enum ibv_wc_opcode comp_opcode; |
21b8aee8 | 638 | struct ibv_cq *ev_cq; |
639 | void *ev_ctx; | |
222757cc | 640 | int ret, r = 0; |
954cd73a | 641 | comp_opcode = IBV_WC_RDMA_WRITE; |
21b8aee8 | 642 | |
643 | switch (rd->rdma_protocol) { | |
644 | case FIO_RDMA_MEM_WRITE: | |
645 | comp_opcode = IBV_WC_RDMA_WRITE; | |
646 | break; | |
647 | case FIO_RDMA_MEM_READ: | |
648 | comp_opcode = IBV_WC_RDMA_READ; | |
649 | break; | |
650 | case FIO_RDMA_CHA_SEND: | |
651 | comp_opcode = IBV_WC_SEND; | |
652 | break; | |
653 | case FIO_RDMA_CHA_RECV: | |
654 | comp_opcode = IBV_WC_RECV; | |
655 | break; | |
656 | default: | |
657 | log_err("fio: unknown rdma protocol - %d\n", rd->rdma_protocol); | |
658 | break; | |
659 | } | |
660 | ||
661 | if (rd->cq_event_num > 0) { /* previous left */ | |
662 | rd->cq_event_num--; | |
663 | return 0; | |
664 | } | |
665 | ||
666 | again: | |
667 | if (ibv_get_cq_event(rd->channel, &ev_cq, &ev_ctx) != 0) { | |
668 | log_err("fio: Failed to get cq event!\n"); | |
669 | return -1; | |
670 | } | |
671 | if (ev_cq != rd->cq) { | |
672 | log_err("fio: Unknown CQ!\n"); | |
673 | return -1; | |
674 | } | |
675 | if (ibv_req_notify_cq(rd->cq, 0) != 0) { | |
676 | log_err("fio: Failed to set notify!\n"); | |
677 | return -1; | |
678 | } | |
679 | ||
680 | ret = cq_event_handler(td, comp_opcode); | |
681 | if (ret < 1) | |
682 | goto again; | |
683 | ||
684 | ibv_ack_cq_events(rd->cq, ret); | |
685 | ||
686 | r += ret; | |
687 | if (r < min) | |
688 | goto again; | |
689 | ||
690 | rd->cq_event_num -= r; | |
691 | ||
692 | return r; | |
693 | } | |
694 | ||
695 | static int fio_rdmaio_send(struct thread_data *td, struct io_u **io_us, | |
696 | unsigned int nr) | |
697 | { | |
565e784d | 698 | struct rdmaio_data *rd = td->io_ops_data; |
21b8aee8 | 699 | struct ibv_send_wr *bad_wr; |
e07f72d3 | 700 | #if 0 |
21b8aee8 | 701 | enum ibv_wc_opcode comp_opcode; |
702 | comp_opcode = IBV_WC_RDMA_WRITE; | |
e07f72d3 | 703 | #endif |
7d7803fa YR |
704 | int i; |
705 | long index; | |
21b8aee8 | 706 | struct rdma_io_u_data *r_io_u_d; |
707 | ||
708 | r_io_u_d = NULL; | |
709 | ||
710 | for (i = 0; i < nr; i++) { | |
711 | /* RDMA_WRITE or RDMA_READ */ | |
712 | switch (rd->rdma_protocol) { | |
713 | case FIO_RDMA_MEM_WRITE: | |
714 | /* compose work request */ | |
715 | r_io_u_d = io_us[i]->engine_data; | |
222757cc | 716 | index = __rand(&rd->rand_state) % rd->rmt_nr; |
21b8aee8 | 717 | r_io_u_d->sq_wr.opcode = IBV_WR_RDMA_WRITE; |
718 | r_io_u_d->sq_wr.wr.rdma.rkey = rd->rmt_us[index].rkey; | |
b6cf38f0 YR |
719 | r_io_u_d->sq_wr.wr.rdma.remote_addr = \ |
720 | rd->rmt_us[index].buf; | |
21b8aee8 | 721 | r_io_u_d->sq_wr.sg_list->length = io_us[i]->buflen; |
722 | break; | |
723 | case FIO_RDMA_MEM_READ: | |
724 | /* compose work request */ | |
725 | r_io_u_d = io_us[i]->engine_data; | |
222757cc | 726 | index = __rand(&rd->rand_state) % rd->rmt_nr; |
21b8aee8 | 727 | r_io_u_d->sq_wr.opcode = IBV_WR_RDMA_READ; |
728 | r_io_u_d->sq_wr.wr.rdma.rkey = rd->rmt_us[index].rkey; | |
b6cf38f0 YR |
729 | r_io_u_d->sq_wr.wr.rdma.remote_addr = \ |
730 | rd->rmt_us[index].buf; | |
21b8aee8 | 731 | r_io_u_d->sq_wr.sg_list->length = io_us[i]->buflen; |
732 | break; | |
733 | case FIO_RDMA_CHA_SEND: | |
734 | r_io_u_d = io_us[i]->engine_data; | |
735 | r_io_u_d->sq_wr.opcode = IBV_WR_SEND; | |
736 | r_io_u_d->sq_wr.send_flags = IBV_SEND_SIGNALED; | |
737 | break; | |
738 | default: | |
739 | log_err("fio: unknown rdma protocol - %d\n", | |
740 | rd->rdma_protocol); | |
741 | break; | |
742 | } | |
743 | ||
744 | if (ibv_post_send(rd->qp, &r_io_u_d->sq_wr, &bad_wr) != 0) { | |
2a442a30 | 745 | log_err("fio: ibv_post_send fail: %m\n"); |
21b8aee8 | 746 | return -1; |
747 | } | |
748 | ||
749 | dprint_io_u(io_us[i], "fio_rdmaio_send"); | |
750 | } | |
751 | ||
752 | /* wait for completion | |
753 | rdma_poll_wait(td, comp_opcode); */ | |
754 | ||
755 | return i; | |
756 | } | |
757 | ||
758 | static int fio_rdmaio_recv(struct thread_data *td, struct io_u **io_us, | |
759 | unsigned int nr) | |
760 | { | |
565e784d | 761 | struct rdmaio_data *rd = td->io_ops_data; |
21b8aee8 | 762 | struct ibv_recv_wr *bad_wr; |
763 | struct rdma_io_u_data *r_io_u_d; | |
764 | int i; | |
765 | ||
766 | i = 0; | |
767 | if (rd->rdma_protocol == FIO_RDMA_CHA_RECV) { | |
768 | /* post io_u into recv queue */ | |
769 | for (i = 0; i < nr; i++) { | |
770 | r_io_u_d = io_us[i]->engine_data; | |
771 | if (ibv_post_recv(rd->qp, &r_io_u_d->rq_wr, &bad_wr) != | |
772 | 0) { | |
2a442a30 | 773 | log_err("fio: ibv_post_recv fail: %m\n"); |
21b8aee8 | 774 | return 1; |
775 | } | |
776 | } | |
777 | } else if ((rd->rdma_protocol == FIO_RDMA_MEM_READ) | |
778 | || (rd->rdma_protocol == FIO_RDMA_MEM_WRITE)) { | |
779 | /* re-post the rq_wr */ | |
780 | if (ibv_post_recv(rd->qp, &rd->rq_wr, &bad_wr) != 0) { | |
2a442a30 | 781 | log_err("fio: ibv_post_recv fail: %m\n"); |
21b8aee8 | 782 | return 1; |
783 | } | |
784 | ||
785 | rdma_poll_wait(td, IBV_WC_RECV); | |
786 | ||
787 | dprint(FD_IO, "fio: recv FINISH message\n"); | |
a05d62b2 YR |
788 | td->done = 1; |
789 | return 0; | |
21b8aee8 | 790 | } |
791 | ||
792 | return i; | |
793 | } | |
794 | ||
2e4ef4fb JA |
795 | static enum fio_q_status fio_rdmaio_queue(struct thread_data *td, |
796 | struct io_u *io_u) | |
21b8aee8 | 797 | { |
565e784d | 798 | struct rdmaio_data *rd = td->io_ops_data; |
21b8aee8 | 799 | |
800 | fio_ro_check(td, io_u); | |
801 | ||
802 | if (rd->io_u_queued_nr == (int)td->o.iodepth) | |
803 | return FIO_Q_BUSY; | |
804 | ||
805 | rd->io_us_queued[rd->io_u_queued_nr] = io_u; | |
806 | rd->io_u_queued_nr++; | |
807 | ||
808 | dprint_io_u(io_u, "fio_rdmaio_queue"); | |
809 | ||
810 | return FIO_Q_QUEUED; | |
811 | } | |
812 | ||
813 | static void fio_rdmaio_queued(struct thread_data *td, struct io_u **io_us, | |
814 | unsigned int nr) | |
815 | { | |
565e784d | 816 | struct rdmaio_data *rd = td->io_ops_data; |
8b6a404c | 817 | struct timespec now; |
21b8aee8 | 818 | unsigned int i; |
819 | ||
820 | if (!fio_fill_issue_time(td)) | |
821 | return; | |
822 | ||
823 | fio_gettime(&now, NULL); | |
824 | ||
825 | for (i = 0; i < nr; i++) { | |
826 | struct io_u *io_u = io_us[i]; | |
827 | ||
828 | /* queued -> flight */ | |
829 | rd->io_us_flight[rd->io_u_flight_nr] = io_u; | |
830 | rd->io_u_flight_nr++; | |
831 | ||
832 | memcpy(&io_u->issue_time, &now, sizeof(now)); | |
833 | io_u_queued(td, io_u); | |
834 | } | |
835 | } | |
836 | ||
837 | static int fio_rdmaio_commit(struct thread_data *td) | |
838 | { | |
565e784d | 839 | struct rdmaio_data *rd = td->io_ops_data; |
21b8aee8 | 840 | struct io_u **io_us; |
841 | int ret; | |
842 | ||
843 | if (!rd->io_us_queued) | |
844 | return 0; | |
845 | ||
846 | io_us = rd->io_us_queued; | |
847 | do { | |
848 | /* RDMA_WRITE or RDMA_READ */ | |
222757cc | 849 | if (rd->is_client) |
21b8aee8 | 850 | ret = fio_rdmaio_send(td, io_us, rd->io_u_queued_nr); |
222757cc | 851 | else if (!rd->is_client) |
21b8aee8 | 852 | ret = fio_rdmaio_recv(td, io_us, rd->io_u_queued_nr); |
222757cc | 853 | else |
21b8aee8 | 854 | ret = 0; /* must be a SYNC */ |
855 | ||
856 | if (ret > 0) { | |
857 | fio_rdmaio_queued(td, io_us, ret); | |
858 | io_u_mark_submit(td, ret); | |
859 | rd->io_u_queued_nr -= ret; | |
860 | io_us += ret; | |
861 | ret = 0; | |
862 | } else | |
863 | break; | |
864 | } while (rd->io_u_queued_nr); | |
865 | ||
866 | return ret; | |
867 | } | |
868 | ||
869 | static int fio_rdmaio_connect(struct thread_data *td, struct fio_file *f) | |
870 | { | |
565e784d | 871 | struct rdmaio_data *rd = td->io_ops_data; |
21b8aee8 | 872 | struct rdma_conn_param conn_param; |
873 | struct ibv_send_wr *bad_wr; | |
874 | ||
222757cc | 875 | memset(&conn_param, 0, sizeof(conn_param)); |
21b8aee8 | 876 | conn_param.responder_resources = 1; |
877 | conn_param.initiator_depth = 1; | |
878 | conn_param.retry_count = 10; | |
879 | ||
880 | if (rdma_connect(rd->cm_id, &conn_param) != 0) { | |
2a442a30 | 881 | log_err("fio: rdma_connect fail: %m\n"); |
21b8aee8 | 882 | return 1; |
883 | } | |
884 | ||
885 | if (get_next_channel_event | |
886 | (td, rd->cm_channel, RDMA_CM_EVENT_ESTABLISHED) != 0) { | |
887 | log_err("fio: wait for RDMA_CM_EVENT_ESTABLISHED\n"); | |
888 | return 1; | |
889 | } | |
890 | ||
891 | /* send task request */ | |
892 | rd->send_buf.mode = htonl(rd->rdma_protocol); | |
893 | rd->send_buf.nr = htonl(td->o.iodepth); | |
894 | ||
895 | if (ibv_post_send(rd->qp, &rd->sq_wr, &bad_wr) != 0) { | |
c5839011 | 896 | log_err("fio: ibv_post_send fail: %m\n"); |
21b8aee8 | 897 | return 1; |
898 | } | |
899 | ||
c4421263 LG |
900 | if (rdma_poll_wait(td, IBV_WC_SEND) < 0) |
901 | return 1; | |
21b8aee8 | 902 | |
903 | /* wait for remote MR info from server side */ | |
c4421263 LG |
904 | if (rdma_poll_wait(td, IBV_WC_RECV) < 0) |
905 | return 1; | |
21b8aee8 | 906 | |
7d7803fa YR |
907 | /* In SEND/RECV test, it's a good practice to setup the iodepth of |
908 | * of the RECV side deeper than that of the SEND side to | |
909 | * avoid RNR (receiver not ready) error. The | |
222757cc | 910 | * SEND side may send so many unsolicited message before |
7d7803fa YR |
911 | * RECV side commits sufficient recv buffers into recv queue. |
912 | * This may lead to RNR error. Here, SEND side pauses for a while | |
913 | * during which RECV side commits sufficient recv buffers. | |
914 | */ | |
915 | usleep(500000); | |
916 | ||
21b8aee8 | 917 | return 0; |
918 | } | |
919 | ||
920 | static int fio_rdmaio_accept(struct thread_data *td, struct fio_file *f) | |
921 | { | |
565e784d | 922 | struct rdmaio_data *rd = td->io_ops_data; |
21b8aee8 | 923 | struct rdma_conn_param conn_param; |
924 | struct ibv_send_wr *bad_wr; | |
c4421263 | 925 | int ret = 0; |
21b8aee8 | 926 | |
927 | /* rdma_accept() - then wait for accept success */ | |
222757cc | 928 | memset(&conn_param, 0, sizeof(conn_param)); |
21b8aee8 | 929 | conn_param.responder_resources = 1; |
930 | conn_param.initiator_depth = 1; | |
931 | ||
932 | if (rdma_accept(rd->child_cm_id, &conn_param) != 0) { | |
2a442a30 | 933 | log_err("fio: rdma_accept: %m\n"); |
21b8aee8 | 934 | return 1; |
935 | } | |
936 | ||
937 | if (get_next_channel_event | |
938 | (td, rd->cm_channel, RDMA_CM_EVENT_ESTABLISHED) != 0) { | |
939 | log_err("fio: wait for RDMA_CM_EVENT_ESTABLISHED\n"); | |
940 | return 1; | |
941 | } | |
942 | ||
943 | /* wait for request */ | |
c4421263 | 944 | ret = rdma_poll_wait(td, IBV_WC_RECV) < 0; |
21b8aee8 | 945 | |
946 | if (ibv_post_send(rd->qp, &rd->sq_wr, &bad_wr) != 0) { | |
c5839011 | 947 | log_err("fio: ibv_post_send fail: %m\n"); |
21b8aee8 | 948 | return 1; |
949 | } | |
950 | ||
c4421263 LG |
951 | if (rdma_poll_wait(td, IBV_WC_SEND) < 0) |
952 | return 1; | |
21b8aee8 | 953 | |
c4421263 | 954 | return ret; |
21b8aee8 | 955 | } |
956 | ||
957 | static int fio_rdmaio_open_file(struct thread_data *td, struct fio_file *f) | |
958 | { | |
959 | if (td_read(td)) | |
960 | return fio_rdmaio_accept(td, f); | |
961 | else | |
962 | return fio_rdmaio_connect(td, f); | |
963 | } | |
964 | ||
965 | static int fio_rdmaio_close_file(struct thread_data *td, struct fio_file *f) | |
966 | { | |
565e784d | 967 | struct rdmaio_data *rd = td->io_ops_data; |
21b8aee8 | 968 | struct ibv_send_wr *bad_wr; |
969 | ||
970 | /* unregister rdma buffer */ | |
971 | ||
972 | /* | |
973 | * Client sends notification to the server side | |
974 | */ | |
975 | /* refer to: http://linux.die.net/man/7/rdma_cm */ | |
976 | if ((rd->is_client == 1) && ((rd->rdma_protocol == FIO_RDMA_MEM_WRITE) | |
977 | || (rd->rdma_protocol == | |
978 | FIO_RDMA_MEM_READ))) { | |
979 | if (ibv_post_send(rd->qp, &rd->sq_wr, &bad_wr) != 0) { | |
c5839011 | 980 | log_err("fio: ibv_post_send fail: %m\n"); |
21b8aee8 | 981 | return 1; |
982 | } | |
983 | ||
de8f6de9 | 984 | dprint(FD_IO, "fio: close information sent success\n"); |
21b8aee8 | 985 | rdma_poll_wait(td, IBV_WC_SEND); |
986 | } | |
987 | ||
988 | if (rd->is_client == 1) | |
989 | rdma_disconnect(rd->cm_id); | |
990 | else { | |
991 | rdma_disconnect(rd->child_cm_id); | |
222757cc JA |
992 | #if 0 |
993 | rdma_disconnect(rd->cm_id); | |
994 | #endif | |
21b8aee8 | 995 | } |
996 | ||
222757cc JA |
997 | #if 0 |
998 | if (get_next_channel_event(td, rd->cm_channel, RDMA_CM_EVENT_DISCONNECTED) != 0) { | |
999 | log_err("fio: wait for RDMA_CM_EVENT_DISCONNECTED\n"); | |
1000 | return 1; | |
1001 | } | |
1002 | #endif | |
21b8aee8 | 1003 | |
21b8aee8 | 1004 | ibv_destroy_cq(rd->cq); |
7d7803fa | 1005 | ibv_destroy_qp(rd->qp); |
21b8aee8 | 1006 | |
1007 | if (rd->is_client == 1) | |
1008 | rdma_destroy_id(rd->cm_id); | |
1009 | else { | |
1010 | rdma_destroy_id(rd->child_cm_id); | |
1011 | rdma_destroy_id(rd->cm_id); | |
1012 | } | |
1013 | ||
1014 | ibv_destroy_comp_channel(rd->channel); | |
1015 | ibv_dealloc_pd(rd->pd); | |
1016 | ||
1017 | return 0; | |
1018 | } | |
1019 | ||
2f28bb35 SB |
1020 | static int aton(struct thread_data *td, const char *host, |
1021 | struct sockaddr_in *addr) | |
1022 | { | |
1023 | if (inet_aton(host, &addr->sin_addr) != 1) { | |
1024 | struct hostent *hent; | |
1025 | ||
1026 | hent = gethostbyname(host); | |
1027 | if (!hent) { | |
1028 | td_verror(td, errno, "gethostbyname"); | |
1029 | return 1; | |
1030 | } | |
1031 | ||
1032 | memcpy(&addr->sin_addr, hent->h_addr, 4); | |
1033 | } | |
1034 | return 0; | |
1035 | } | |
1036 | ||
21b8aee8 | 1037 | static int fio_rdmaio_setup_connect(struct thread_data *td, const char *host, |
1038 | unsigned short port) | |
1039 | { | |
565e784d | 1040 | struct rdmaio_data *rd = td->io_ops_data; |
2f28bb35 SB |
1041 | struct rdmaio_options *o = td->eo; |
1042 | struct sockaddr_storage addrb; | |
21b8aee8 | 1043 | struct ibv_recv_wr *bad_wr; |
222757cc | 1044 | int err; |
21b8aee8 | 1045 | |
1046 | rd->addr.sin_family = AF_INET; | |
1047 | rd->addr.sin_port = htons(port); | |
1048 | ||
2f28bb35 SB |
1049 | err = aton(td, host, &rd->addr); |
1050 | if (err) | |
1051 | return err; | |
21b8aee8 | 1052 | |
2f28bb35 | 1053 | /* resolve route */ |
de8f8675 | 1054 | if (o->bindname && strlen(o->bindname)) { |
2f28bb35 SB |
1055 | addrb.ss_family = AF_INET; |
1056 | err = aton(td, o->bindname, (struct sockaddr_in *)&addrb); | |
1057 | if (err) | |
1058 | return err; | |
1059 | err = rdma_resolve_addr(rd->cm_id, (struct sockaddr *)&addrb, | |
1060 | (struct sockaddr *)&rd->addr, 2000); | |
21b8aee8 | 1061 | |
2f28bb35 SB |
1062 | } else { |
1063 | err = rdma_resolve_addr(rd->cm_id, NULL, | |
1064 | (struct sockaddr *)&rd->addr, 2000); | |
21b8aee8 | 1065 | } |
1066 | ||
222757cc JA |
1067 | if (err != 0) { |
1068 | log_err("fio: rdma_resolve_addr: %d\n", err); | |
21b8aee8 | 1069 | return 1; |
1070 | } | |
1071 | ||
222757cc JA |
1072 | err = get_next_channel_event(td, rd->cm_channel, RDMA_CM_EVENT_ADDR_RESOLVED); |
1073 | if (err != 0) { | |
1074 | log_err("fio: get_next_channel_event: %d\n", err); | |
21b8aee8 | 1075 | return 1; |
1076 | } | |
1077 | ||
1078 | /* resolve route */ | |
222757cc JA |
1079 | err = rdma_resolve_route(rd->cm_id, 2000); |
1080 | if (err != 0) { | |
1081 | log_err("fio: rdma_resolve_route: %d\n", err); | |
21b8aee8 | 1082 | return 1; |
1083 | } | |
1084 | ||
222757cc JA |
1085 | err = get_next_channel_event(td, rd->cm_channel, RDMA_CM_EVENT_ROUTE_RESOLVED); |
1086 | if (err != 0) { | |
1087 | log_err("fio: get_next_channel_event: %d\n", err); | |
21b8aee8 | 1088 | return 1; |
1089 | } | |
1090 | ||
1091 | /* create qp and buffer */ | |
1092 | if (fio_rdmaio_setup_qp(td) != 0) | |
1093 | return 1; | |
1094 | ||
1095 | if (fio_rdmaio_setup_control_msg_buffers(td) != 0) | |
1096 | return 1; | |
1097 | ||
1098 | /* post recv buf */ | |
222757cc JA |
1099 | err = ibv_post_recv(rd->qp, &rd->rq_wr, &bad_wr); |
1100 | if (err != 0) { | |
1101 | log_err("fio: ibv_post_recv fail: %d\n", err); | |
21b8aee8 | 1102 | return 1; |
1103 | } | |
1104 | ||
1105 | return 0; | |
1106 | } | |
1107 | ||
1108 | static int fio_rdmaio_setup_listen(struct thread_data *td, short port) | |
1109 | { | |
565e784d | 1110 | struct rdmaio_data *rd = td->io_ops_data; |
2f28bb35 | 1111 | struct rdmaio_options *o = td->eo; |
21b8aee8 | 1112 | struct ibv_recv_wr *bad_wr; |
263775ad LG |
1113 | int state = td->runstate; |
1114 | ||
1115 | td_set_runstate(td, TD_SETTING_UP); | |
21b8aee8 | 1116 | |
1117 | rd->addr.sin_family = AF_INET; | |
21b8aee8 | 1118 | rd->addr.sin_port = htons(port); |
1119 | ||
de8f8675 | 1120 | if (!o->bindname || !strlen(o->bindname)) |
2f28bb35 SB |
1121 | rd->addr.sin_addr.s_addr = htonl(INADDR_ANY); |
1122 | else | |
1123 | rd->addr.sin_addr.s_addr = htonl(*o->bindname); | |
1124 | ||
21b8aee8 | 1125 | /* rdma_listen */ |
1126 | if (rdma_bind_addr(rd->cm_id, (struct sockaddr *)&rd->addr) != 0) { | |
2a442a30 | 1127 | log_err("fio: rdma_bind_addr fail: %m\n"); |
21b8aee8 | 1128 | return 1; |
1129 | } | |
1130 | ||
1131 | if (rdma_listen(rd->cm_id, 3) != 0) { | |
2a442a30 | 1132 | log_err("fio: rdma_listen fail: %m\n"); |
21b8aee8 | 1133 | return 1; |
1134 | } | |
1135 | ||
263775ad LG |
1136 | log_info("fio: waiting for connection\n"); |
1137 | ||
21b8aee8 | 1138 | /* wait for CONNECT_REQUEST */ |
1139 | if (get_next_channel_event | |
1140 | (td, rd->cm_channel, RDMA_CM_EVENT_CONNECT_REQUEST) != 0) { | |
1141 | log_err("fio: wait for RDMA_CM_EVENT_CONNECT_REQUEST\n"); | |
1142 | return 1; | |
1143 | } | |
1144 | ||
1145 | if (fio_rdmaio_setup_qp(td) != 0) | |
1146 | return 1; | |
1147 | ||
1148 | if (fio_rdmaio_setup_control_msg_buffers(td) != 0) | |
1149 | return 1; | |
1150 | ||
1151 | /* post recv buf */ | |
1152 | if (ibv_post_recv(rd->qp, &rd->rq_wr, &bad_wr) != 0) { | |
2a442a30 | 1153 | log_err("fio: ibv_post_recv fail: %m\n"); |
21b8aee8 | 1154 | return 1; |
1155 | } | |
1156 | ||
263775ad | 1157 | td_set_runstate(td, state); |
21b8aee8 | 1158 | return 0; |
1159 | } | |
1160 | ||
38b9354a | 1161 | static int check_set_rlimits(struct thread_data *td) |
21b8aee8 | 1162 | { |
38b9354a | 1163 | #ifdef CONFIG_RLIMIT_MEMLOCK |
21b8aee8 | 1164 | struct rlimit rl; |
1165 | ||
21b8aee8 | 1166 | /* check RLIMIT_MEMLOCK */ |
1167 | if (getrlimit(RLIMIT_MEMLOCK, &rl) != 0) { | |
1168 | log_err("fio: getrlimit fail: %d(%s)\n", | |
1169 | errno, strerror(errno)); | |
1170 | return 1; | |
1171 | } | |
1172 | ||
1173 | /* soft limit */ | |
1174 | if ((rl.rlim_cur != RLIM_INFINITY) | |
1175 | && (rl.rlim_cur < td->orig_buffer_size)) { | |
e07f72d3 BVA |
1176 | log_err("fio: soft RLIMIT_MEMLOCK is: %" PRId64 "\n", |
1177 | rl.rlim_cur); | |
1178 | log_err("fio: total block size is: %zd\n", | |
21b8aee8 | 1179 | td->orig_buffer_size); |
1180 | /* try to set larger RLIMIT_MEMLOCK */ | |
1181 | rl.rlim_cur = rl.rlim_max; | |
1182 | if (setrlimit(RLIMIT_MEMLOCK, &rl) != 0) { | |
1183 | log_err("fio: setrlimit fail: %d(%s)\n", | |
1184 | errno, strerror(errno)); | |
1185 | log_err("fio: you may try enlarge MEMLOCK by root\n"); | |
1186 | log_err("# ulimit -l unlimited\n"); | |
1187 | return 1; | |
1188 | } | |
1189 | } | |
38b9354a JA |
1190 | #endif |
1191 | ||
1192 | return 0; | |
1193 | } | |
1194 | ||
cdf91594 | 1195 | static int compat_options(struct thread_data *td) |
38b9354a | 1196 | { |
cdf91594 LG |
1197 | // The original RDMA engine had an ugly / seperator |
1198 | // on the filename for it's options. This function | |
2f28bb35 SB |
1199 | // retains backwards compatibility with it. Note we do not |
1200 | // support setting the bindname option is this legacy mode. | |
38b9354a | 1201 | |
cdf91594 LG |
1202 | struct rdmaio_options *o = td->eo; |
1203 | char *modep, *portp; | |
1204 | char *filename = td->o.filename; | |
38b9354a | 1205 | |
cdf91594 LG |
1206 | if (!filename) |
1207 | return 0; | |
21b8aee8 | 1208 | |
cdf91594 LG |
1209 | portp = strchr(filename, '/'); |
1210 | if (portp == NULL) | |
1211 | return 0; | |
21b8aee8 | 1212 | |
cdf91594 LG |
1213 | *portp = '\0'; |
1214 | portp++; | |
21b8aee8 | 1215 | |
cdf91594 LG |
1216 | o->port = strtol(portp, NULL, 10); |
1217 | if (!o->port || o->port > 65535) | |
21b8aee8 | 1218 | goto bad_host; |
1219 | ||
cdf91594 LG |
1220 | modep = strchr(portp, '/'); |
1221 | if (modep != NULL) { | |
1222 | *modep = '\0'; | |
1223 | modep++; | |
21b8aee8 | 1224 | } |
1225 | ||
21b8aee8 | 1226 | if (modep) { |
1227 | if (!strncmp("rdma_write", modep, strlen(modep)) || | |
1228 | !strncmp("RDMA_WRITE", modep, strlen(modep))) | |
cdf91594 | 1229 | o->verb = FIO_RDMA_MEM_WRITE; |
21b8aee8 | 1230 | else if (!strncmp("rdma_read", modep, strlen(modep)) || |
1231 | !strncmp("RDMA_READ", modep, strlen(modep))) | |
cdf91594 | 1232 | o->verb = FIO_RDMA_MEM_READ; |
21b8aee8 | 1233 | else if (!strncmp("send", modep, strlen(modep)) || |
1234 | !strncmp("SEND", modep, strlen(modep))) | |
cdf91594 | 1235 | o->verb = FIO_RDMA_CHA_SEND; |
21b8aee8 | 1236 | else |
1237 | goto bad_host; | |
1238 | } else | |
cdf91594 LG |
1239 | o->verb = FIO_RDMA_MEM_WRITE; |
1240 | ||
1241 | ||
1242 | return 0; | |
21b8aee8 | 1243 | |
cdf91594 LG |
1244 | bad_host: |
1245 | log_err("fio: bad rdma host/port/protocol: %s\n", td->o.filename); | |
1246 | return 1; | |
1247 | } | |
1248 | ||
1249 | static int fio_rdmaio_init(struct thread_data *td) | |
1250 | { | |
565e784d | 1251 | struct rdmaio_data *rd = td->io_ops_data; |
cdf91594 | 1252 | struct rdmaio_options *o = td->eo; |
1480ce7d | 1253 | int ret; |
cdf91594 LG |
1254 | |
1255 | if (td_rw(td)) { | |
1256 | log_err("fio: rdma connections must be read OR write\n"); | |
1257 | return 1; | |
1258 | } | |
1259 | if (td_random(td)) { | |
1260 | log_err("fio: RDMA network IO can't be random\n"); | |
1261 | return 1; | |
1262 | } | |
1263 | ||
1264 | if (compat_options(td)) | |
1265 | return 1; | |
1266 | ||
1267 | if (!o->port) { | |
1268 | log_err("fio: no port has been specified which is required " | |
1269 | "for the rdma engine\n"); | |
1270 | return 1; | |
1271 | } | |
1272 | ||
1273 | if (check_set_rlimits(td)) | |
1274 | return 1; | |
1275 | ||
1276 | rd->rdma_protocol = o->verb; | |
21b8aee8 | 1277 | rd->cq_event_num = 0; |
1278 | ||
1279 | rd->cm_channel = rdma_create_event_channel(); | |
1280 | if (!rd->cm_channel) { | |
2a442a30 | 1281 | log_err("fio: rdma_create_event_channel fail: %m\n"); |
21b8aee8 | 1282 | return 1; |
1283 | } | |
1284 | ||
1285 | ret = rdma_create_id(rd->cm_channel, &rd->cm_id, rd, RDMA_PS_TCP); | |
1286 | if (ret) { | |
2a442a30 | 1287 | log_err("fio: rdma_create_id fail: %m\n"); |
21b8aee8 | 1288 | return 1; |
1289 | } | |
1290 | ||
1291 | if ((rd->rdma_protocol == FIO_RDMA_MEM_WRITE) || | |
1292 | (rd->rdma_protocol == FIO_RDMA_MEM_READ)) { | |
1293 | rd->rmt_us = | |
76cc5224 | 1294 | malloc(FIO_RDMA_MAX_IO_DEPTH * sizeof(struct remote_u)); |
21b8aee8 | 1295 | memset(rd->rmt_us, 0, |
76cc5224 | 1296 | FIO_RDMA_MAX_IO_DEPTH * sizeof(struct remote_u)); |
21b8aee8 | 1297 | rd->rmt_nr = 0; |
1298 | } | |
1299 | ||
1300 | rd->io_us_queued = malloc(td->o.iodepth * sizeof(struct io_u *)); | |
1301 | memset(rd->io_us_queued, 0, td->o.iodepth * sizeof(struct io_u *)); | |
1302 | rd->io_u_queued_nr = 0; | |
1303 | ||
1304 | rd->io_us_flight = malloc(td->o.iodepth * sizeof(struct io_u *)); | |
1305 | memset(rd->io_us_flight, 0, td->o.iodepth * sizeof(struct io_u *)); | |
1306 | rd->io_u_flight_nr = 0; | |
1307 | ||
1308 | rd->io_us_completed = malloc(td->o.iodepth * sizeof(struct io_u *)); | |
1309 | memset(rd->io_us_completed, 0, td->o.iodepth * sizeof(struct io_u *)); | |
1310 | rd->io_u_completed_nr = 0; | |
1311 | ||
1312 | if (td_read(td)) { /* READ as the server */ | |
1313 | rd->is_client = 0; | |
263775ad | 1314 | td->flags |= TD_F_NO_PROGRESS; |
21b8aee8 | 1315 | /* server rd->rdma_buf_len will be setup after got request */ |
cdf91594 | 1316 | ret = fio_rdmaio_setup_listen(td, o->port); |
21b8aee8 | 1317 | } else { /* WRITE as the client */ |
1318 | rd->is_client = 1; | |
cdf91594 | 1319 | ret = fio_rdmaio_setup_connect(td, td->o.filename, o->port); |
21b8aee8 | 1320 | } |
1480ce7d DM |
1321 | return ret; |
1322 | } | |
1323 | static int fio_rdmaio_post_init(struct thread_data *td) | |
1324 | { | |
1325 | unsigned int max_bs; | |
1326 | int i; | |
1327 | struct rdmaio_data *rd = td->io_ops_data; | |
21b8aee8 | 1328 | |
21b8aee8 | 1329 | max_bs = max(td->o.max_bs[DDIR_READ], td->o.max_bs[DDIR_WRITE]); |
c4421263 LG |
1330 | rd->send_buf.max_bs = htonl(max_bs); |
1331 | ||
21b8aee8 | 1332 | /* register each io_u in the free list */ |
954cd73a YR |
1333 | for (i = 0; i < td->io_u_freelist.nr; i++) { |
1334 | struct io_u *io_u = td->io_u_freelist.io_us[i]; | |
21b8aee8 | 1335 | |
1336 | io_u->engine_data = malloc(sizeof(struct rdma_io_u_data)); | |
1337 | memset(io_u->engine_data, 0, sizeof(struct rdma_io_u_data)); | |
1338 | ((struct rdma_io_u_data *)io_u->engine_data)->wr_id = i; | |
1339 | ||
1340 | io_u->mr = ibv_reg_mr(rd->pd, io_u->buf, max_bs, | |
1341 | IBV_ACCESS_LOCAL_WRITE | | |
1342 | IBV_ACCESS_REMOTE_READ | | |
1343 | IBV_ACCESS_REMOTE_WRITE); | |
1344 | if (io_u->mr == NULL) { | |
2a442a30 | 1345 | log_err("fio: ibv_reg_mr io_u failed: %m\n"); |
21b8aee8 | 1346 | return 1; |
1347 | } | |
1348 | ||
1349 | rd->send_buf.rmt_us[i].buf = | |
b747af64 | 1350 | cpu_to_be64((uint64_t) (unsigned long)io_u->buf); |
21b8aee8 | 1351 | rd->send_buf.rmt_us[i].rkey = htonl(io_u->mr->rkey); |
1352 | rd->send_buf.rmt_us[i].size = htonl(max_bs); | |
1353 | ||
222757cc JA |
1354 | #if 0 |
1355 | log_info("fio: Send rkey %x addr %" PRIx64 " len %d to client\n", io_u->mr->rkey, io_u->buf, max_bs); */ | |
1356 | #endif | |
21b8aee8 | 1357 | } |
1358 | ||
1359 | rd->send_buf.nr = htonl(i); | |
1360 | ||
1480ce7d | 1361 | return 0; |
21b8aee8 | 1362 | } |
1363 | ||
1364 | static void fio_rdmaio_cleanup(struct thread_data *td) | |
1365 | { | |
565e784d | 1366 | struct rdmaio_data *rd = td->io_ops_data; |
21b8aee8 | 1367 | |
222757cc | 1368 | if (rd) |
21b8aee8 | 1369 | free(rd); |
21b8aee8 | 1370 | } |
1371 | ||
1372 | static int fio_rdmaio_setup(struct thread_data *td) | |
1373 | { | |
1374 | struct rdmaio_data *rd; | |
1375 | ||
cdf91594 LG |
1376 | if (!td->files_index) { |
1377 | add_file(td, td->o.filename ?: "rdma", 0, 0); | |
1378 | td->o.nr_files = td->o.nr_files ?: 1; | |
1379 | td->o.open_files++; | |
1380 | } | |
1381 | ||
565e784d | 1382 | if (!td->io_ops_data) { |
222757cc | 1383 | rd = malloc(sizeof(*rd)); |
21b8aee8 | 1384 | |
1385 | memset(rd, 0, sizeof(*rd)); | |
ec4dde59 | 1386 | init_rand_seed(&rd->rand_state, (unsigned int) GOLDEN_RATIO_PRIME, 0); |
565e784d | 1387 | td->io_ops_data = rd; |
21b8aee8 | 1388 | } |
1389 | ||
1390 | return 0; | |
1391 | } | |
1392 | ||
5a8a6a03 | 1393 | FIO_STATIC struct ioengine_ops ioengine = { |
cdf91594 LG |
1394 | .name = "rdma", |
1395 | .version = FIO_IOOPS_VERSION, | |
1396 | .setup = fio_rdmaio_setup, | |
1397 | .init = fio_rdmaio_init, | |
1480ce7d | 1398 | .post_init = fio_rdmaio_post_init, |
cdf91594 LG |
1399 | .prep = fio_rdmaio_prep, |
1400 | .queue = fio_rdmaio_queue, | |
1401 | .commit = fio_rdmaio_commit, | |
1402 | .getevents = fio_rdmaio_getevents, | |
1403 | .event = fio_rdmaio_event, | |
1404 | .cleanup = fio_rdmaio_cleanup, | |
1405 | .open_file = fio_rdmaio_open_file, | |
1406 | .close_file = fio_rdmaio_close_file, | |
1407 | .flags = FIO_DISKLESSIO | FIO_UNIDIR | FIO_PIPEIO, | |
1408 | .options = options, | |
1409 | .option_struct_size = sizeof(struct rdmaio_options), | |
21b8aee8 | 1410 | }; |
1411 | ||
21b8aee8 | 1412 | static void fio_init fio_rdmaio_register(void) |
1413 | { | |
5a8a6a03 | 1414 | register_ioengine(&ioengine); |
21b8aee8 | 1415 | } |
1416 | ||
1417 | static void fio_exit fio_rdmaio_unregister(void) | |
1418 | { | |
5a8a6a03 | 1419 | unregister_ioengine(&ioengine); |
21b8aee8 | 1420 | } |