Commit | Line | Data |
---|---|---|
e4c4625f JM |
1 | /* |
2 | * librpma_fio: librpma_apm and librpma_gpspm engines' common part. | |
3 | * | |
4 | * Copyright 2021, Intel Corporation | |
5 | * | |
6 | * This program is free software; you can redistribute it and/or | |
7 | * modify it under the terms of the GNU General Public License, | |
8 | * version 2 as published by the Free Software Foundation.. | |
9 | * | |
10 | * This program is distributed in the hope that it will be useful, | |
11 | * but WITHOUT ANY WARRANTY; without even the implied warranty of | |
12 | * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the | |
13 | * GNU General Public License for more details. | |
14 | */ | |
15 | ||
16 | #include "librpma_fio.h" | |
17 | ||
18 | #include <libpmem.h> | |
19 | ||
20 | struct fio_option librpma_fio_options[] = { | |
21 | { | |
22 | .name = "serverip", | |
23 | .lname = "rpma_server_ip", | |
24 | .type = FIO_OPT_STR_STORE, | |
25 | .off1 = offsetof(struct librpma_fio_options_values, server_ip), | |
26 | .help = "IP address the server is listening on", | |
27 | .def = "", | |
28 | .category = FIO_OPT_C_ENGINE, | |
29 | .group = FIO_OPT_G_LIBRPMA, | |
30 | }, | |
31 | { | |
32 | .name = "port", | |
33 | .lname = "rpma_server port", | |
34 | .type = FIO_OPT_STR_STORE, | |
35 | .off1 = offsetof(struct librpma_fio_options_values, port), | |
36 | .help = "port the server is listening on", | |
37 | .def = "7204", | |
38 | .category = FIO_OPT_C_ENGINE, | |
39 | .group = FIO_OPT_G_LIBRPMA, | |
40 | }, | |
41 | { | |
42 | .name = "direct_write_to_pmem", | |
43 | .lname = "Direct Write to PMem (via RDMA) from the remote host is possible", | |
44 | .type = FIO_OPT_BOOL, | |
45 | .off1 = offsetof(struct librpma_fio_options_values, | |
46 | direct_write_to_pmem), | |
47 | .help = "Set to true ONLY when Direct Write to PMem from the remote host is possible (https://pmem.io/rpma/documentation/basic-direct-write-to-pmem.html)", | |
48 | .def = "", | |
49 | .category = FIO_OPT_C_ENGINE, | |
50 | .group = FIO_OPT_G_LIBRPMA, | |
51 | }, | |
6a229978 OS |
52 | { |
53 | .name = "busy_wait_polling", | |
54 | .lname = "Set to 0 to wait for completion instead of busy-wait polling completion.", | |
55 | .type = FIO_OPT_BOOL, | |
56 | .off1 = offsetof(struct librpma_fio_options_values, | |
57 | busy_wait_polling), | |
58 | .help = "Set to false if you want to reduce CPU usage", | |
59 | .def = "1", | |
60 | .category = FIO_OPT_C_ENGINE, | |
61 | .group = FIO_OPT_G_LIBRPMA, | |
62 | }, | |
e4c4625f JM |
63 | { |
64 | .name = NULL, | |
65 | }, | |
66 | }; | |
67 | ||
68 | int librpma_fio_td_port(const char *port_base_str, struct thread_data *td, | |
69 | char *port_out) | |
70 | { | |
71 | unsigned long int port_ul = strtoul(port_base_str, NULL, 10); | |
72 | unsigned int port_new; | |
73 | ||
74 | port_out[0] = '\0'; | |
75 | ||
76 | if (port_ul == ULONG_MAX) { | |
77 | td_verror(td, errno, "strtoul"); | |
78 | return -1; | |
79 | } | |
80 | port_ul += td->thread_number - 1; | |
81 | if (port_ul >= UINT_MAX) { | |
82 | log_err("[%u] port number (%lu) bigger than UINT_MAX\n", | |
83 | td->thread_number, port_ul); | |
84 | return -1; | |
85 | } | |
86 | ||
87 | port_new = port_ul; | |
88 | snprintf(port_out, LIBRPMA_FIO_PORT_STR_LEN_MAX - 1, "%u", port_new); | |
89 | ||
90 | return 0; | |
91 | } | |
92 | ||
93 | char *librpma_fio_allocate_dram(struct thread_data *td, size_t size, | |
94 | struct librpma_fio_mem *mem) | |
95 | { | |
96 | char *mem_ptr = NULL; | |
97 | int ret; | |
98 | ||
99 | if ((ret = posix_memalign((void **)&mem_ptr, page_size, size))) { | |
100 | log_err("fio: posix_memalign() failed\n"); | |
101 | td_verror(td, ret, "posix_memalign"); | |
102 | return NULL; | |
103 | } | |
104 | ||
105 | mem->mem_ptr = mem_ptr; | |
106 | mem->size_mmap = 0; | |
107 | ||
108 | return mem_ptr; | |
109 | } | |
110 | ||
111 | char *librpma_fio_allocate_pmem(struct thread_data *td, const char *filename, | |
112 | size_t size, struct librpma_fio_mem *mem) | |
113 | { | |
114 | size_t size_mmap = 0; | |
115 | char *mem_ptr = NULL; | |
116 | int is_pmem = 0; | |
117 | size_t ws_offset; | |
118 | ||
119 | if (size % page_size) { | |
120 | log_err("fio: size (%zu) is not aligned to page size (%zu)\n", | |
121 | size, page_size); | |
122 | return NULL; | |
123 | } | |
124 | ||
125 | ws_offset = (td->thread_number - 1) * size; | |
126 | ||
127 | if (!filename) { | |
128 | log_err("fio: filename is not set\n"); | |
129 | return NULL; | |
130 | } | |
131 | ||
132 | /* map the file */ | |
133 | mem_ptr = pmem_map_file(filename, 0 /* len */, 0 /* flags */, | |
134 | 0 /* mode */, &size_mmap, &is_pmem); | |
135 | if (mem_ptr == NULL) { | |
136 | log_err("fio: pmem_map_file(%s) failed\n", filename); | |
137 | /* pmem_map_file() sets errno on failure */ | |
138 | td_verror(td, errno, "pmem_map_file"); | |
139 | return NULL; | |
140 | } | |
141 | ||
142 | /* pmem is expected */ | |
143 | if (!is_pmem) { | |
144 | log_err("fio: %s is not located in persistent memory\n", | |
145 | filename); | |
146 | goto err_unmap; | |
147 | } | |
148 | ||
149 | /* check size of allocated persistent memory */ | |
150 | if (size_mmap < ws_offset + size) { | |
151 | log_err( | |
152 | "fio: %s is too small to handle so many threads (%zu < %zu)\n", | |
153 | filename, size_mmap, ws_offset + size); | |
154 | goto err_unmap; | |
155 | } | |
156 | ||
157 | log_info("fio: size of memory mapped from the file %s: %zu\n", | |
158 | filename, size_mmap); | |
159 | ||
160 | mem->mem_ptr = mem_ptr; | |
161 | mem->size_mmap = size_mmap; | |
162 | ||
163 | return mem_ptr + ws_offset; | |
164 | ||
165 | err_unmap: | |
166 | (void) pmem_unmap(mem_ptr, size_mmap); | |
167 | return NULL; | |
168 | } | |
169 | ||
170 | void librpma_fio_free(struct librpma_fio_mem *mem) | |
171 | { | |
172 | if (mem->size_mmap) | |
173 | (void) pmem_unmap(mem->mem_ptr, mem->size_mmap); | |
174 | else | |
175 | free(mem->mem_ptr); | |
176 | } | |
177 | ||
178 | #define LIBRPMA_FIO_RETRY_MAX_NO 10 | |
179 | #define LIBRPMA_FIO_RETRY_DELAY_S 5 | |
180 | ||
181 | int librpma_fio_client_init(struct thread_data *td, | |
182 | struct rpma_conn_cfg *cfg) | |
183 | { | |
184 | struct librpma_fio_client_data *ccd; | |
185 | struct librpma_fio_options_values *o = td->eo; | |
186 | struct ibv_context *dev = NULL; | |
187 | char port_td[LIBRPMA_FIO_PORT_STR_LEN_MAX]; | |
188 | struct rpma_conn_req *req = NULL; | |
189 | enum rpma_conn_event event; | |
190 | struct rpma_conn_private_data pdata; | |
191 | enum rpma_log_level log_level_aux = RPMA_LOG_LEVEL_WARNING; | |
192 | int remote_flush_type; | |
193 | int retry; | |
194 | int ret; | |
195 | ||
196 | /* --debug=net sets RPMA_LOG_THRESHOLD_AUX to RPMA_LOG_LEVEL_INFO */ | |
197 | #ifdef FIO_INC_DEBUG | |
198 | if ((1UL << FD_NET) & fio_debug) | |
199 | log_level_aux = RPMA_LOG_LEVEL_INFO; | |
200 | #endif | |
201 | ||
202 | /* configure logging thresholds to see more details */ | |
203 | rpma_log_set_threshold(RPMA_LOG_THRESHOLD, RPMA_LOG_LEVEL_INFO); | |
204 | rpma_log_set_threshold(RPMA_LOG_THRESHOLD_AUX, log_level_aux); | |
205 | ||
206 | /* obtain an IBV context for a remote IP address */ | |
207 | if ((ret = rpma_utils_get_ibv_context(o->server_ip, | |
208 | RPMA_UTIL_IBV_CONTEXT_REMOTE, &dev))) { | |
209 | librpma_td_verror(td, ret, "rpma_utils_get_ibv_context"); | |
210 | return -1; | |
211 | } | |
212 | ||
213 | /* allocate client's data */ | |
214 | ccd = calloc(1, sizeof(*ccd)); | |
215 | if (ccd == NULL) { | |
216 | td_verror(td, errno, "calloc"); | |
217 | return -1; | |
218 | } | |
219 | ||
220 | /* allocate all in-memory queues */ | |
221 | ccd->io_us_queued = calloc(td->o.iodepth, sizeof(*ccd->io_us_queued)); | |
222 | if (ccd->io_us_queued == NULL) { | |
223 | td_verror(td, errno, "calloc"); | |
224 | goto err_free_ccd; | |
225 | } | |
226 | ||
227 | ccd->io_us_flight = calloc(td->o.iodepth, sizeof(*ccd->io_us_flight)); | |
228 | if (ccd->io_us_flight == NULL) { | |
229 | td_verror(td, errno, "calloc"); | |
230 | goto err_free_io_u_queues; | |
231 | } | |
232 | ||
233 | ccd->io_us_completed = calloc(td->o.iodepth, | |
234 | sizeof(*ccd->io_us_completed)); | |
235 | if (ccd->io_us_completed == NULL) { | |
236 | td_verror(td, errno, "calloc"); | |
237 | goto err_free_io_u_queues; | |
238 | } | |
239 | ||
240 | /* create a new peer object */ | |
241 | if ((ret = rpma_peer_new(dev, &ccd->peer))) { | |
242 | librpma_td_verror(td, ret, "rpma_peer_new"); | |
243 | goto err_free_io_u_queues; | |
244 | } | |
245 | ||
246 | /* create a connection request */ | |
247 | if (librpma_fio_td_port(o->port, td, port_td)) | |
248 | goto err_peer_delete; | |
249 | ||
250 | for (retry = 0; retry < LIBRPMA_FIO_RETRY_MAX_NO; retry++) { | |
251 | if ((ret = rpma_conn_req_new(ccd->peer, o->server_ip, port_td, | |
252 | cfg, &req))) { | |
253 | librpma_td_verror(td, ret, "rpma_conn_req_new"); | |
254 | goto err_peer_delete; | |
255 | } | |
256 | ||
257 | /* | |
258 | * Connect the connection request | |
259 | * and obtain the connection object. | |
260 | */ | |
261 | if ((ret = rpma_conn_req_connect(&req, NULL, &ccd->conn))) { | |
262 | librpma_td_verror(td, ret, "rpma_conn_req_connect"); | |
263 | goto err_req_delete; | |
264 | } | |
265 | ||
266 | /* wait for the connection to establish */ | |
267 | if ((ret = rpma_conn_next_event(ccd->conn, &event))) { | |
268 | librpma_td_verror(td, ret, "rpma_conn_next_event"); | |
269 | goto err_conn_delete; | |
270 | } else if (event == RPMA_CONN_ESTABLISHED) { | |
271 | break; | |
272 | } else if (event == RPMA_CONN_REJECTED) { | |
273 | (void) rpma_conn_disconnect(ccd->conn); | |
274 | (void) rpma_conn_delete(&ccd->conn); | |
275 | if (retry < LIBRPMA_FIO_RETRY_MAX_NO - 1) { | |
276 | log_err("Thread [%d]: Retrying (#%i) ...\n", | |
277 | td->thread_number, retry + 1); | |
278 | sleep(LIBRPMA_FIO_RETRY_DELAY_S); | |
279 | } else { | |
280 | log_err( | |
281 | "Thread [%d]: The maximum number of retries exceeded. Closing.\n", | |
282 | td->thread_number); | |
283 | } | |
284 | } else { | |
285 | log_err( | |
286 | "rpma_conn_next_event returned an unexptected event: (%s != RPMA_CONN_ESTABLISHED)\n", | |
287 | rpma_utils_conn_event_2str(event)); | |
288 | goto err_conn_delete; | |
289 | } | |
290 | } | |
291 | ||
292 | if (retry > 0) | |
293 | log_err("Thread [%d]: Connected after retry #%i\n", | |
294 | td->thread_number, retry); | |
295 | ||
296 | if (ccd->conn == NULL) | |
297 | goto err_peer_delete; | |
298 | ||
299 | /* get the connection's private data sent from the server */ | |
300 | if ((ret = rpma_conn_get_private_data(ccd->conn, &pdata))) { | |
301 | librpma_td_verror(td, ret, "rpma_conn_get_private_data"); | |
302 | goto err_conn_delete; | |
303 | } | |
304 | ||
305 | /* get the server's workspace representation */ | |
306 | ccd->ws = pdata.ptr; | |
307 | ||
308 | /* create the server's memory representation */ | |
309 | if ((ret = rpma_mr_remote_from_descriptor(&ccd->ws->descriptor[0], | |
310 | ccd->ws->mr_desc_size, &ccd->server_mr))) { | |
311 | librpma_td_verror(td, ret, "rpma_mr_remote_from_descriptor"); | |
312 | goto err_conn_delete; | |
313 | } | |
314 | ||
315 | /* get the total size of the shared server memory */ | |
316 | if ((ret = rpma_mr_remote_get_size(ccd->server_mr, &ccd->ws_size))) { | |
317 | librpma_td_verror(td, ret, "rpma_mr_remote_get_size"); | |
318 | goto err_conn_delete; | |
319 | } | |
320 | ||
321 | /* get flush type of the remote node */ | |
322 | if ((ret = rpma_mr_remote_get_flush_type(ccd->server_mr, | |
323 | &remote_flush_type))) { | |
324 | librpma_td_verror(td, ret, "rpma_mr_remote_get_flush_type"); | |
325 | goto err_conn_delete; | |
326 | } | |
327 | ||
328 | ccd->server_mr_flush_type = | |
329 | (remote_flush_type & RPMA_MR_USAGE_FLUSH_TYPE_PERSISTENT) ? | |
330 | RPMA_FLUSH_TYPE_PERSISTENT : RPMA_FLUSH_TYPE_VISIBILITY; | |
331 | ||
332 | /* | |
333 | * Assure an io_us buffer allocation is page-size-aligned which is required | |
334 | * to register for RDMA. User-provided value is intentionally ignored. | |
335 | */ | |
336 | td->o.mem_align = page_size; | |
337 | ||
338 | td->io_ops_data = ccd; | |
339 | ||
340 | return 0; | |
341 | ||
342 | err_conn_delete: | |
343 | (void) rpma_conn_disconnect(ccd->conn); | |
344 | (void) rpma_conn_delete(&ccd->conn); | |
345 | ||
346 | err_req_delete: | |
347 | (void) rpma_conn_req_delete(&req); | |
348 | ||
349 | err_peer_delete: | |
350 | (void) rpma_peer_delete(&ccd->peer); | |
351 | ||
352 | err_free_io_u_queues: | |
353 | free(ccd->io_us_queued); | |
354 | free(ccd->io_us_flight); | |
355 | free(ccd->io_us_completed); | |
356 | ||
357 | err_free_ccd: | |
358 | free(ccd); | |
359 | ||
360 | return -1; | |
361 | } | |
362 | ||
363 | void librpma_fio_client_cleanup(struct thread_data *td) | |
364 | { | |
365 | struct librpma_fio_client_data *ccd = td->io_ops_data; | |
366 | enum rpma_conn_event ev; | |
367 | int ret; | |
368 | ||
369 | if (ccd == NULL) | |
370 | return; | |
371 | ||
372 | /* delete the iou's memory registration */ | |
373 | if ((ret = rpma_mr_dereg(&ccd->orig_mr))) | |
374 | librpma_td_verror(td, ret, "rpma_mr_dereg"); | |
375 | /* delete the iou's memory registration */ | |
376 | if ((ret = rpma_mr_remote_delete(&ccd->server_mr))) | |
377 | librpma_td_verror(td, ret, "rpma_mr_remote_delete"); | |
378 | /* initiate disconnection */ | |
379 | if ((ret = rpma_conn_disconnect(ccd->conn))) | |
380 | librpma_td_verror(td, ret, "rpma_conn_disconnect"); | |
381 | /* wait for disconnection to end up */ | |
382 | if ((ret = rpma_conn_next_event(ccd->conn, &ev))) { | |
383 | librpma_td_verror(td, ret, "rpma_conn_next_event"); | |
384 | } else if (ev != RPMA_CONN_CLOSED) { | |
385 | log_err( | |
386 | "client_cleanup received an unexpected event (%s != RPMA_CONN_CLOSED)\n", | |
387 | rpma_utils_conn_event_2str(ev)); | |
388 | } | |
389 | /* delete the connection */ | |
390 | if ((ret = rpma_conn_delete(&ccd->conn))) | |
391 | librpma_td_verror(td, ret, "rpma_conn_delete"); | |
392 | /* delete the peer */ | |
393 | if ((ret = rpma_peer_delete(&ccd->peer))) | |
394 | librpma_td_verror(td, ret, "rpma_peer_delete"); | |
395 | /* free the software queues */ | |
396 | free(ccd->io_us_queued); | |
397 | free(ccd->io_us_flight); | |
398 | free(ccd->io_us_completed); | |
399 | free(ccd); | |
400 | td->io_ops_data = NULL; /* zero ccd */ | |
401 | } | |
402 | ||
403 | int librpma_fio_file_nop(struct thread_data *td, struct fio_file *f) | |
404 | { | |
405 | /* NOP */ | |
406 | return 0; | |
407 | } | |
408 | ||
409 | int librpma_fio_client_post_init(struct thread_data *td) | |
410 | { | |
411 | struct librpma_fio_client_data *ccd = td->io_ops_data; | |
412 | size_t io_us_size; | |
413 | int ret; | |
414 | ||
415 | /* | |
416 | * td->orig_buffer is not aligned. The engine requires aligned io_us | |
417 | * so FIO alignes up the address using the formula below. | |
418 | */ | |
419 | ccd->orig_buffer_aligned = PTR_ALIGN(td->orig_buffer, page_mask) + | |
420 | td->o.mem_align; | |
421 | ||
422 | /* | |
423 | * td->orig_buffer_size beside the space really consumed by io_us | |
424 | * has paddings which can be omitted for the memory registration. | |
425 | */ | |
426 | io_us_size = (unsigned long long)td_max_bs(td) * | |
427 | (unsigned long long)td->o.iodepth; | |
428 | ||
429 | if ((ret = rpma_mr_reg(ccd->peer, ccd->orig_buffer_aligned, io_us_size, | |
430 | RPMA_MR_USAGE_READ_DST | RPMA_MR_USAGE_READ_SRC | | |
431 | RPMA_MR_USAGE_WRITE_DST | RPMA_MR_USAGE_WRITE_SRC | | |
432 | RPMA_MR_USAGE_FLUSH_TYPE_PERSISTENT, &ccd->orig_mr))) | |
433 | librpma_td_verror(td, ret, "rpma_mr_reg"); | |
434 | return ret; | |
435 | } | |
436 | ||
437 | int librpma_fio_client_get_file_size(struct thread_data *td, | |
438 | struct fio_file *f) | |
439 | { | |
440 | struct librpma_fio_client_data *ccd = td->io_ops_data; | |
441 | ||
442 | f->real_file_size = ccd->ws_size; | |
443 | fio_file_set_size_known(f); | |
444 | ||
445 | return 0; | |
446 | } | |
447 | ||
448 | static enum fio_q_status client_queue_sync(struct thread_data *td, | |
449 | struct io_u *io_u) | |
450 | { | |
451 | struct librpma_fio_client_data *ccd = td->io_ops_data; | |
452 | struct rpma_completion cmpl; | |
453 | unsigned io_u_index; | |
454 | int ret; | |
455 | ||
456 | /* execute io_u */ | |
457 | if (io_u->ddir == DDIR_READ) { | |
458 | /* post an RDMA read operation */ | |
459 | if (librpma_fio_client_io_read(td, io_u, | |
460 | RPMA_F_COMPLETION_ALWAYS)) | |
461 | goto err; | |
462 | } else if (io_u->ddir == DDIR_WRITE) { | |
463 | /* post an RDMA write operation */ | |
464 | if (librpma_fio_client_io_write(td, io_u)) | |
465 | goto err; | |
466 | if (ccd->flush(td, io_u, io_u, io_u->xfer_buflen)) | |
467 | goto err; | |
468 | } else { | |
469 | log_err("unsupported IO mode: %s\n", io_ddir_name(io_u->ddir)); | |
470 | goto err; | |
471 | } | |
472 | ||
473 | do { | |
474 | /* get a completion */ | |
475 | ret = rpma_conn_completion_get(ccd->conn, &cmpl); | |
476 | if (ret == RPMA_E_NO_COMPLETION) { | |
477 | /* lack of completion is not an error */ | |
478 | continue; | |
479 | } else if (ret != 0) { | |
480 | /* an error occurred */ | |
481 | librpma_td_verror(td, ret, "rpma_conn_completion_get"); | |
482 | goto err; | |
483 | } | |
484 | ||
485 | /* if io_us has completed with an error */ | |
486 | if (cmpl.op_status != IBV_WC_SUCCESS) | |
487 | goto err; | |
488 | ||
489 | if (cmpl.op == RPMA_OP_SEND) | |
490 | ++ccd->op_send_completed; | |
491 | else { | |
492 | if (cmpl.op == RPMA_OP_RECV) | |
493 | ++ccd->op_recv_completed; | |
494 | ||
495 | break; | |
496 | } | |
497 | } while (1); | |
498 | ||
499 | if (ccd->get_io_u_index(&cmpl, &io_u_index) != 1) | |
500 | goto err; | |
501 | ||
502 | if (io_u->index != io_u_index) { | |
503 | log_err( | |
504 | "no matching io_u for received completion found (io_u_index=%u)\n", | |
505 | io_u_index); | |
506 | goto err; | |
507 | } | |
508 | ||
509 | /* make sure all SENDs are completed before exit - clean up SQ */ | |
510 | if (librpma_fio_client_io_complete_all_sends(td)) | |
511 | goto err; | |
512 | ||
513 | return FIO_Q_COMPLETED; | |
514 | ||
515 | err: | |
516 | io_u->error = -1; | |
517 | return FIO_Q_COMPLETED; | |
518 | } | |
519 | ||
520 | enum fio_q_status librpma_fio_client_queue(struct thread_data *td, | |
521 | struct io_u *io_u) | |
522 | { | |
523 | struct librpma_fio_client_data *ccd = td->io_ops_data; | |
524 | ||
525 | if (ccd->io_u_queued_nr == (int)td->o.iodepth) | |
526 | return FIO_Q_BUSY; | |
527 | ||
528 | if (td->o.sync_io) | |
529 | return client_queue_sync(td, io_u); | |
530 | ||
531 | /* io_u -> queued[] */ | |
532 | ccd->io_us_queued[ccd->io_u_queued_nr] = io_u; | |
533 | ccd->io_u_queued_nr++; | |
534 | ||
535 | return FIO_Q_QUEUED; | |
536 | } | |
537 | ||
538 | int librpma_fio_client_commit(struct thread_data *td) | |
539 | { | |
540 | struct librpma_fio_client_data *ccd = td->io_ops_data; | |
541 | int flags = RPMA_F_COMPLETION_ON_ERROR; | |
542 | struct timespec now; | |
543 | bool fill_time; | |
544 | int i; | |
545 | struct io_u *flush_first_io_u = NULL; | |
546 | unsigned long long int flush_len = 0; | |
547 | ||
548 | if (!ccd->io_us_queued) | |
549 | return -1; | |
550 | ||
551 | /* execute all io_us from queued[] */ | |
552 | for (i = 0; i < ccd->io_u_queued_nr; i++) { | |
553 | struct io_u *io_u = ccd->io_us_queued[i]; | |
554 | ||
555 | if (io_u->ddir == DDIR_READ) { | |
556 | if (i + 1 == ccd->io_u_queued_nr || | |
557 | ccd->io_us_queued[i + 1]->ddir == DDIR_WRITE) | |
558 | flags = RPMA_F_COMPLETION_ALWAYS; | |
559 | /* post an RDMA read operation */ | |
560 | if (librpma_fio_client_io_read(td, io_u, flags)) | |
561 | return -1; | |
562 | } else if (io_u->ddir == DDIR_WRITE) { | |
563 | /* post an RDMA write operation */ | |
564 | if (librpma_fio_client_io_write(td, io_u)) | |
565 | return -1; | |
566 | ||
567 | /* cache the first io_u in the sequence */ | |
568 | if (flush_first_io_u == NULL) | |
569 | flush_first_io_u = io_u; | |
570 | ||
571 | /* | |
572 | * the flush length is the sum of all io_u's creating | |
573 | * the sequence | |
574 | */ | |
575 | flush_len += io_u->xfer_buflen; | |
576 | ||
577 | /* | |
578 | * if io_u's are random the rpma_flush is required | |
579 | * after each one of them | |
580 | */ | |
581 | if (!td_random(td)) { | |
582 | /* | |
583 | * When the io_u's are sequential and | |
584 | * the current io_u is not the last one and | |
585 | * the next one is also a write operation | |
586 | * the flush can be postponed by one io_u and | |
587 | * cover all of them which build a continuous | |
588 | * sequence. | |
589 | */ | |
590 | if ((i + 1 < ccd->io_u_queued_nr) && | |
591 | (ccd->io_us_queued[i + 1]->ddir == DDIR_WRITE)) | |
592 | continue; | |
593 | } | |
594 | ||
595 | /* flush all writes which build a continuous sequence */ | |
596 | if (ccd->flush(td, flush_first_io_u, io_u, flush_len)) | |
597 | return -1; | |
598 | ||
599 | /* | |
600 | * reset the flush parameters in preparation for | |
601 | * the next one | |
602 | */ | |
603 | flush_first_io_u = NULL; | |
604 | flush_len = 0; | |
605 | } else { | |
606 | log_err("unsupported IO mode: %s\n", | |
607 | io_ddir_name(io_u->ddir)); | |
608 | return -1; | |
609 | } | |
610 | } | |
611 | ||
612 | if ((fill_time = fio_fill_issue_time(td))) | |
613 | fio_gettime(&now, NULL); | |
614 | ||
615 | /* move executed io_us from queued[] to flight[] */ | |
616 | for (i = 0; i < ccd->io_u_queued_nr; i++) { | |
617 | struct io_u *io_u = ccd->io_us_queued[i]; | |
618 | ||
619 | /* FIO does not do this if the engine is asynchronous */ | |
620 | if (fill_time) | |
621 | memcpy(&io_u->issue_time, &now, sizeof(now)); | |
622 | ||
623 | /* move executed io_us from queued[] to flight[] */ | |
624 | ccd->io_us_flight[ccd->io_u_flight_nr] = io_u; | |
625 | ccd->io_u_flight_nr++; | |
626 | ||
627 | /* | |
628 | * FIO says: | |
629 | * If an engine has the commit hook | |
630 | * it has to call io_u_queued() itself. | |
631 | */ | |
632 | io_u_queued(td, io_u); | |
633 | } | |
634 | ||
635 | /* FIO does not do this if an engine has the commit hook. */ | |
636 | io_u_mark_submit(td, ccd->io_u_queued_nr); | |
637 | ccd->io_u_queued_nr = 0; | |
638 | ||
639 | return 0; | |
640 | } | |
641 | ||
642 | /* | |
643 | * RETURN VALUE | |
644 | * - > 0 - a number of completed io_us | |
645 | * - 0 - when no complicitions received | |
646 | * - (-1) - when an error occurred | |
647 | */ | |
648 | static int client_getevent_process(struct thread_data *td) | |
649 | { | |
650 | struct librpma_fio_client_data *ccd = td->io_ops_data; | |
651 | struct rpma_completion cmpl; | |
652 | /* io_u->index of completed io_u (cmpl.op_context) */ | |
653 | unsigned int io_u_index; | |
654 | /* # of completed io_us */ | |
655 | int cmpl_num = 0; | |
656 | /* helpers */ | |
657 | struct io_u *io_u; | |
658 | int i; | |
659 | int ret; | |
660 | ||
661 | /* get a completion */ | |
662 | if ((ret = rpma_conn_completion_get(ccd->conn, &cmpl))) { | |
663 | /* lack of completion is not an error */ | |
664 | if (ret == RPMA_E_NO_COMPLETION) { | |
665 | /* lack of completion is not an error */ | |
666 | return 0; | |
667 | } | |
668 | ||
669 | /* an error occurred */ | |
670 | librpma_td_verror(td, ret, "rpma_conn_completion_get"); | |
671 | return -1; | |
672 | } | |
673 | ||
674 | /* if io_us has completed with an error */ | |
675 | if (cmpl.op_status != IBV_WC_SUCCESS) { | |
676 | td->error = cmpl.op_status; | |
677 | return -1; | |
678 | } | |
679 | ||
680 | if (cmpl.op == RPMA_OP_SEND) | |
681 | ++ccd->op_send_completed; | |
682 | else if (cmpl.op == RPMA_OP_RECV) | |
683 | ++ccd->op_recv_completed; | |
684 | ||
685 | if ((ret = ccd->get_io_u_index(&cmpl, &io_u_index)) != 1) | |
686 | return ret; | |
687 | ||
688 | /* look for an io_u being completed */ | |
689 | for (i = 0; i < ccd->io_u_flight_nr; ++i) { | |
690 | if (ccd->io_us_flight[i]->index == io_u_index) { | |
691 | cmpl_num = i + 1; | |
692 | break; | |
693 | } | |
694 | } | |
695 | ||
696 | /* if no matching io_u has been found */ | |
697 | if (cmpl_num == 0) { | |
698 | log_err( | |
699 | "no matching io_u for received completion found (io_u_index=%u)\n", | |
700 | io_u_index); | |
701 | return -1; | |
702 | } | |
703 | ||
704 | /* move completed io_us to the completed in-memory queue */ | |
705 | for (i = 0; i < cmpl_num; ++i) { | |
706 | /* get and prepare io_u */ | |
707 | io_u = ccd->io_us_flight[i]; | |
708 | ||
709 | /* append to the queue */ | |
710 | ccd->io_us_completed[ccd->io_u_completed_nr] = io_u; | |
711 | ccd->io_u_completed_nr++; | |
712 | } | |
713 | ||
714 | /* remove completed io_us from the flight queue */ | |
715 | for (i = cmpl_num; i < ccd->io_u_flight_nr; ++i) | |
716 | ccd->io_us_flight[i - cmpl_num] = ccd->io_us_flight[i]; | |
717 | ccd->io_u_flight_nr -= cmpl_num; | |
718 | ||
719 | return cmpl_num; | |
720 | } | |
721 | ||
722 | int librpma_fio_client_getevents(struct thread_data *td, unsigned int min, | |
723 | unsigned int max, const struct timespec *t) | |
724 | { | |
725 | struct librpma_fio_client_data *ccd = td->io_ops_data; | |
726 | /* total # of completed io_us */ | |
727 | int cmpl_num_total = 0; | |
728 | /* # of completed io_us from a single event */ | |
729 | int cmpl_num; | |
730 | ||
731 | do { | |
732 | cmpl_num = client_getevent_process(td); | |
733 | if (cmpl_num > 0) { | |
734 | /* new completions collected */ | |
735 | cmpl_num_total += cmpl_num; | |
736 | } else if (cmpl_num == 0) { | |
737 | /* | |
738 | * It is required to make sure that CQEs for SENDs | |
739 | * will flow at least at the same pace as CQEs for RECVs. | |
740 | */ | |
741 | if (cmpl_num_total >= min && | |
742 | ccd->op_send_completed >= ccd->op_recv_completed) | |
743 | break; | |
744 | ||
745 | /* | |
746 | * To reduce CPU consumption one can use | |
747 | * the rpma_conn_completion_wait() function. | |
748 | * Note this greatly increase the latency | |
749 | * and make the results less stable. | |
750 | * The bandwidth stays more or less the same. | |
751 | */ | |
752 | } else { | |
753 | /* an error occurred */ | |
754 | return -1; | |
755 | } | |
756 | ||
757 | /* | |
758 | * The expected max can be exceeded if CQEs for RECVs will come up | |
759 | * faster than CQEs for SENDs. But it is required to make sure CQEs for | |
760 | * SENDs will flow at least at the same pace as CQEs for RECVs. | |
761 | */ | |
762 | } while (cmpl_num_total < max || | |
763 | ccd->op_send_completed < ccd->op_recv_completed); | |
764 | ||
765 | /* | |
766 | * All posted SENDs are completed and RECVs for them (responses) are | |
767 | * completed. This is the initial situation so the counters are reset. | |
768 | */ | |
769 | if (ccd->op_send_posted == ccd->op_send_completed && | |
770 | ccd->op_send_completed == ccd->op_recv_completed) { | |
771 | ccd->op_send_posted = 0; | |
772 | ccd->op_send_completed = 0; | |
773 | ccd->op_recv_completed = 0; | |
774 | } | |
775 | ||
776 | return cmpl_num_total; | |
777 | } | |
778 | ||
779 | struct io_u *librpma_fio_client_event(struct thread_data *td, int event) | |
780 | { | |
781 | struct librpma_fio_client_data *ccd = td->io_ops_data; | |
782 | struct io_u *io_u; | |
783 | int i; | |
784 | ||
785 | /* get the first io_u from the queue */ | |
786 | io_u = ccd->io_us_completed[0]; | |
787 | ||
788 | /* remove the first io_u from the queue */ | |
789 | for (i = 1; i < ccd->io_u_completed_nr; ++i) | |
790 | ccd->io_us_completed[i - 1] = ccd->io_us_completed[i]; | |
791 | ccd->io_u_completed_nr--; | |
792 | ||
793 | dprint_io_u(io_u, "client_event"); | |
794 | ||
795 | return io_u; | |
796 | } | |
797 | ||
798 | char *librpma_fio_client_errdetails(struct io_u *io_u) | |
799 | { | |
800 | /* get the string representation of an error */ | |
801 | enum ibv_wc_status status = io_u->error; | |
802 | const char *status_str = ibv_wc_status_str(status); | |
803 | ||
804 | char *details = strdup(status_str); | |
805 | if (details == NULL) { | |
806 | fprintf(stderr, "Error: %s\n", status_str); | |
807 | fprintf(stderr, "Fatal error: out of memory. Aborting.\n"); | |
808 | abort(); | |
809 | } | |
810 | ||
811 | /* FIO frees the returned string when it becomes obsolete */ | |
812 | return details; | |
813 | } | |
814 | ||
815 | int librpma_fio_server_init(struct thread_data *td) | |
816 | { | |
817 | struct librpma_fio_options_values *o = td->eo; | |
818 | struct librpma_fio_server_data *csd; | |
819 | struct ibv_context *dev = NULL; | |
820 | enum rpma_log_level log_level_aux = RPMA_LOG_LEVEL_WARNING; | |
821 | int ret = -1; | |
822 | ||
823 | /* --debug=net sets RPMA_LOG_THRESHOLD_AUX to RPMA_LOG_LEVEL_INFO */ | |
824 | #ifdef FIO_INC_DEBUG | |
825 | if ((1UL << FD_NET) & fio_debug) | |
826 | log_level_aux = RPMA_LOG_LEVEL_INFO; | |
827 | #endif | |
828 | ||
829 | /* configure logging thresholds to see more details */ | |
830 | rpma_log_set_threshold(RPMA_LOG_THRESHOLD, RPMA_LOG_LEVEL_INFO); | |
831 | rpma_log_set_threshold(RPMA_LOG_THRESHOLD_AUX, log_level_aux); | |
832 | ||
833 | ||
834 | /* obtain an IBV context for a remote IP address */ | |
835 | if ((ret = rpma_utils_get_ibv_context(o->server_ip, | |
836 | RPMA_UTIL_IBV_CONTEXT_LOCAL, &dev))) { | |
837 | librpma_td_verror(td, ret, "rpma_utils_get_ibv_context"); | |
838 | return -1; | |
839 | } | |
840 | ||
841 | /* allocate server's data */ | |
842 | csd = calloc(1, sizeof(*csd)); | |
843 | if (csd == NULL) { | |
844 | td_verror(td, errno, "calloc"); | |
845 | return -1; | |
846 | } | |
847 | ||
848 | /* create a new peer object */ | |
849 | if ((ret = rpma_peer_new(dev, &csd->peer))) { | |
850 | librpma_td_verror(td, ret, "rpma_peer_new"); | |
851 | goto err_free_csd; | |
852 | } | |
853 | ||
854 | td->io_ops_data = csd; | |
855 | ||
856 | return 0; | |
857 | ||
858 | err_free_csd: | |
859 | free(csd); | |
860 | ||
861 | return -1; | |
862 | } | |
863 | ||
864 | void librpma_fio_server_cleanup(struct thread_data *td) | |
865 | { | |
866 | struct librpma_fio_server_data *csd = td->io_ops_data; | |
867 | int ret; | |
868 | ||
869 | if (csd == NULL) | |
870 | return; | |
871 | ||
872 | /* free the peer */ | |
873 | if ((ret = rpma_peer_delete(&csd->peer))) | |
874 | librpma_td_verror(td, ret, "rpma_peer_delete"); | |
875 | ||
876 | free(csd); | |
877 | } | |
878 | ||
879 | int librpma_fio_server_open_file(struct thread_data *td, struct fio_file *f, | |
880 | struct rpma_conn_cfg *cfg) | |
881 | { | |
882 | struct librpma_fio_server_data *csd = td->io_ops_data; | |
883 | struct librpma_fio_options_values *o = td->eo; | |
884 | enum rpma_conn_event conn_event = RPMA_CONN_UNDEFINED; | |
885 | struct librpma_fio_workspace ws = {0}; | |
886 | struct rpma_conn_private_data pdata; | |
887 | uint32_t max_msg_num; | |
888 | struct rpma_conn_req *conn_req; | |
889 | struct rpma_conn *conn; | |
890 | struct rpma_mr_local *mr; | |
891 | char port_td[LIBRPMA_FIO_PORT_STR_LEN_MAX]; | |
892 | struct rpma_ep *ep; | |
893 | size_t mem_size = td->o.size; | |
894 | size_t mr_desc_size; | |
895 | void *ws_ptr; | |
896 | int usage_mem_type; | |
897 | int ret; | |
898 | ||
899 | if (!f->file_name) { | |
900 | log_err("fio: filename is not set\n"); | |
901 | return -1; | |
902 | } | |
903 | ||
904 | /* start a listening endpoint at addr:port */ | |
905 | if (librpma_fio_td_port(o->port, td, port_td)) | |
906 | return -1; | |
907 | ||
908 | if ((ret = rpma_ep_listen(csd->peer, o->server_ip, port_td, &ep))) { | |
909 | librpma_td_verror(td, ret, "rpma_ep_listen"); | |
910 | return -1; | |
911 | } | |
912 | ||
913 | if (strcmp(f->file_name, "malloc") == 0) { | |
914 | /* allocation from DRAM using posix_memalign() */ | |
915 | ws_ptr = librpma_fio_allocate_dram(td, mem_size, &csd->mem); | |
916 | usage_mem_type = RPMA_MR_USAGE_FLUSH_TYPE_VISIBILITY; | |
917 | } else { | |
918 | /* allocation from PMEM using pmem_map_file() */ | |
919 | ws_ptr = librpma_fio_allocate_pmem(td, f->file_name, | |
920 | mem_size, &csd->mem); | |
921 | usage_mem_type = RPMA_MR_USAGE_FLUSH_TYPE_PERSISTENT; | |
922 | } | |
923 | ||
924 | if (ws_ptr == NULL) | |
925 | goto err_ep_shutdown; | |
926 | ||
927 | f->real_file_size = mem_size; | |
928 | ||
929 | if ((ret = rpma_mr_reg(csd->peer, ws_ptr, mem_size, | |
930 | RPMA_MR_USAGE_READ_DST | RPMA_MR_USAGE_READ_SRC | | |
931 | RPMA_MR_USAGE_WRITE_DST | RPMA_MR_USAGE_WRITE_SRC | | |
932 | usage_mem_type, &mr))) { | |
933 | librpma_td_verror(td, ret, "rpma_mr_reg"); | |
934 | goto err_free; | |
935 | } | |
936 | ||
937 | /* get size of the memory region's descriptor */ | |
938 | if ((ret = rpma_mr_get_descriptor_size(mr, &mr_desc_size))) { | |
939 | librpma_td_verror(td, ret, "rpma_mr_get_descriptor_size"); | |
940 | goto err_mr_dereg; | |
941 | } | |
942 | ||
943 | /* verify size of the memory region's descriptor */ | |
944 | if (mr_desc_size > LIBRPMA_FIO_DESCRIPTOR_MAX_SIZE) { | |
945 | log_err( | |
946 | "size of the memory region's descriptor is too big (max=%i)\n", | |
947 | LIBRPMA_FIO_DESCRIPTOR_MAX_SIZE); | |
948 | goto err_mr_dereg; | |
949 | } | |
950 | ||
951 | /* get the memory region's descriptor */ | |
952 | if ((ret = rpma_mr_get_descriptor(mr, &ws.descriptor[0]))) { | |
953 | librpma_td_verror(td, ret, "rpma_mr_get_descriptor"); | |
954 | goto err_mr_dereg; | |
955 | } | |
956 | ||
957 | if (cfg != NULL) { | |
958 | if ((ret = rpma_conn_cfg_get_rq_size(cfg, &max_msg_num))) { | |
959 | librpma_td_verror(td, ret, "rpma_conn_cfg_get_rq_size"); | |
960 | goto err_mr_dereg; | |
961 | } | |
962 | ||
963 | /* verify whether iodepth fits into uint16_t */ | |
964 | if (max_msg_num > UINT16_MAX) { | |
965 | log_err("fio: iodepth too big (%u > %u)\n", | |
966 | max_msg_num, UINT16_MAX); | |
967 | return -1; | |
968 | } | |
969 | ||
970 | ws.max_msg_num = max_msg_num; | |
971 | } | |
972 | ||
973 | /* prepare a workspace description */ | |
974 | ws.direct_write_to_pmem = o->direct_write_to_pmem; | |
975 | ws.mr_desc_size = mr_desc_size; | |
976 | pdata.ptr = &ws; | |
977 | pdata.len = sizeof(ws); | |
978 | ||
979 | /* receive an incoming connection request */ | |
980 | if ((ret = rpma_ep_next_conn_req(ep, cfg, &conn_req))) { | |
981 | librpma_td_verror(td, ret, "rpma_ep_next_conn_req"); | |
982 | goto err_mr_dereg; | |
983 | } | |
984 | ||
985 | if (csd->prepare_connection && csd->prepare_connection(td, conn_req)) | |
986 | goto err_req_delete; | |
987 | ||
988 | /* accept the connection request and obtain the connection object */ | |
989 | if ((ret = rpma_conn_req_connect(&conn_req, &pdata, &conn))) { | |
990 | librpma_td_verror(td, ret, "rpma_conn_req_connect"); | |
991 | goto err_req_delete; | |
992 | } | |
993 | ||
994 | /* wait for the connection to be established */ | |
995 | if ((ret = rpma_conn_next_event(conn, &conn_event))) { | |
996 | librpma_td_verror(td, ret, "rpma_conn_next_event"); | |
997 | goto err_conn_delete; | |
998 | } else if (conn_event != RPMA_CONN_ESTABLISHED) { | |
999 | log_err("rpma_conn_next_event returned an unexptected event\n"); | |
1000 | goto err_conn_delete; | |
1001 | } | |
1002 | ||
1003 | /* end-point is no longer needed */ | |
1004 | (void) rpma_ep_shutdown(&ep); | |
1005 | ||
1006 | csd->ws_mr = mr; | |
1007 | csd->ws_ptr = ws_ptr; | |
1008 | csd->conn = conn; | |
1009 | ||
1010 | return 0; | |
1011 | ||
1012 | err_conn_delete: | |
1013 | (void) rpma_conn_delete(&conn); | |
1014 | ||
1015 | err_req_delete: | |
1016 | (void) rpma_conn_req_delete(&conn_req); | |
1017 | ||
1018 | err_mr_dereg: | |
1019 | (void) rpma_mr_dereg(&mr); | |
1020 | ||
1021 | err_free: | |
1022 | librpma_fio_free(&csd->mem); | |
1023 | ||
1024 | err_ep_shutdown: | |
1025 | (void) rpma_ep_shutdown(&ep); | |
1026 | ||
1027 | return -1; | |
1028 | } | |
1029 | ||
1030 | int librpma_fio_server_close_file(struct thread_data *td, struct fio_file *f) | |
1031 | { | |
1032 | struct librpma_fio_server_data *csd = td->io_ops_data; | |
1033 | enum rpma_conn_event conn_event = RPMA_CONN_UNDEFINED; | |
1034 | int rv = 0; | |
1035 | int ret; | |
1036 | ||
1037 | /* wait for the connection to be closed */ | |
1038 | ret = rpma_conn_next_event(csd->conn, &conn_event); | |
1039 | if (!ret && conn_event != RPMA_CONN_CLOSED) { | |
1040 | log_err("rpma_conn_next_event returned an unexptected event\n"); | |
1041 | rv = -1; | |
1042 | } | |
1043 | ||
1044 | if ((ret = rpma_conn_disconnect(csd->conn))) { | |
1045 | librpma_td_verror(td, ret, "rpma_conn_disconnect"); | |
1046 | rv = -1; | |
1047 | } | |
1048 | ||
1049 | if ((ret = rpma_conn_delete(&csd->conn))) { | |
1050 | librpma_td_verror(td, ret, "rpma_conn_delete"); | |
1051 | rv = -1; | |
1052 | } | |
1053 | ||
1054 | if ((ret = rpma_mr_dereg(&csd->ws_mr))) { | |
1055 | librpma_td_verror(td, ret, "rpma_mr_dereg"); | |
1056 | rv = -1; | |
1057 | } | |
1058 | ||
1059 | librpma_fio_free(&csd->mem); | |
1060 | ||
1061 | return rv; | |
1062 | } |