Commit | Line | Data |
---|---|---|
fc5c0345 DG |
1 | /* |
2 | * rbd engine | |
3 | * | |
4 | * IO engine using Ceph's librbd to test RADOS Block Devices. | |
5 | * | |
6 | */ | |
7 | ||
8 | #include <rbd/librbd.h> | |
9 | ||
10 | #include "../fio.h" | |
d220c761 | 11 | #include "../optgroup.h" |
fc5c0345 | 12 | |
53b6c979 PL |
13 | #ifdef CONFIG_RBD_POLL |
14 | /* add for poll */ | |
15 | #include <poll.h> | |
16 | #include <sys/eventfd.h> | |
17 | #endif | |
18 | ||
fc5c0345 DG |
19 | struct fio_rbd_iou { |
20 | struct io_u *io_u; | |
d8b64af2 | 21 | rbd_completion_t completion; |
d8b64af2 | 22 | int io_seen; |
b8ecbef6 | 23 | int io_complete; |
fc5c0345 DG |
24 | }; |
25 | ||
26 | struct rbd_data { | |
27 | rados_t cluster; | |
28 | rados_ioctx_t io_ctx; | |
29 | rbd_image_t image; | |
30 | struct io_u **aio_events; | |
6f9961ac | 31 | struct io_u **sort_events; |
53b6c979 | 32 | int fd; /* add for poll */ |
e0f35b53 | 33 | bool connected; |
fc5c0345 DG |
34 | }; |
35 | ||
36 | struct rbd_options { | |
a1f871c7 | 37 | void *pad; |
6e20c6e7 | 38 | char *cluster_name; |
fc5c0345 DG |
39 | char *rbd_name; |
40 | char *pool_name; | |
41 | char *client_name; | |
d7d702c7 | 42 | int busy_poll; |
fc5c0345 DG |
43 | }; |
44 | ||
45 | static struct fio_option options[] = { | |
6e20c6e7 T |
46 | { |
47 | .name = "clustername", | |
48 | .lname = "ceph cluster name", | |
49 | .type = FIO_OPT_STR_STORE, | |
50 | .help = "Cluster name for ceph", | |
51 | .off1 = offsetof(struct rbd_options, cluster_name), | |
52 | .category = FIO_OPT_C_ENGINE, | |
53 | .group = FIO_OPT_G_RBD, | |
54 | }, | |
fc5c0345 | 55 | { |
d8b64af2 JA |
56 | .name = "rbdname", |
57 | .lname = "rbd engine rbdname", | |
58 | .type = FIO_OPT_STR_STORE, | |
59 | .help = "RBD name for RBD engine", | |
60 | .off1 = offsetof(struct rbd_options, rbd_name), | |
61 | .category = FIO_OPT_C_ENGINE, | |
62 | .group = FIO_OPT_G_RBD, | |
63 | }, | |
fc5c0345 | 64 | { |
d7d702c7 JA |
65 | .name = "pool", |
66 | .lname = "rbd engine pool", | |
67 | .type = FIO_OPT_STR_STORE, | |
68 | .help = "Name of the pool hosting the RBD for the RBD engine", | |
69 | .off1 = offsetof(struct rbd_options, pool_name), | |
70 | .category = FIO_OPT_C_ENGINE, | |
71 | .group = FIO_OPT_G_RBD, | |
d8b64af2 | 72 | }, |
fc5c0345 | 73 | { |
d7d702c7 JA |
74 | .name = "clientname", |
75 | .lname = "rbd engine clientname", | |
76 | .type = FIO_OPT_STR_STORE, | |
77 | .help = "Name of the ceph client to access the RBD for the RBD engine", | |
78 | .off1 = offsetof(struct rbd_options, client_name), | |
79 | .category = FIO_OPT_C_ENGINE, | |
80 | .group = FIO_OPT_G_RBD, | |
81 | }, | |
82 | { | |
83 | .name = "busy_poll", | |
84 | .lname = "Busy poll", | |
85 | .type = FIO_OPT_BOOL, | |
86 | .help = "Busy poll for completions instead of sleeping", | |
fea585d4 | 87 | .off1 = offsetof(struct rbd_options, busy_poll), |
d7d702c7 JA |
88 | .def = "0", |
89 | .category = FIO_OPT_C_ENGINE, | |
90 | .group = FIO_OPT_G_RBD, | |
d8b64af2 | 91 | }, |
fc5c0345 | 92 | { |
d8b64af2 JA |
93 | .name = NULL, |
94 | }, | |
fc5c0345 DG |
95 | }; |
96 | ||
97 | static int _fio_setup_rbd_data(struct thread_data *td, | |
98 | struct rbd_data **rbd_data_ptr) | |
99 | { | |
6f9961ac | 100 | struct rbd_data *rbd; |
fc5c0345 | 101 | |
565e784d | 102 | if (td->io_ops_data) |
fc5c0345 DG |
103 | return 0; |
104 | ||
6f9961ac JA |
105 | rbd = calloc(1, sizeof(struct rbd_data)); |
106 | if (!rbd) | |
fc5c0345 DG |
107 | goto failed; |
108 | ||
e0f35b53 PL |
109 | rbd->connected = false; |
110 | ||
53b6c979 PL |
111 | /* add for poll, init fd: -1 */ |
112 | rbd->fd = -1; | |
53b6c979 | 113 | |
6f9961ac JA |
114 | rbd->aio_events = calloc(td->o.iodepth, sizeof(struct io_u *)); |
115 | if (!rbd->aio_events) | |
fc5c0345 DG |
116 | goto failed; |
117 | ||
6f9961ac JA |
118 | rbd->sort_events = calloc(td->o.iodepth, sizeof(struct io_u *)); |
119 | if (!rbd->sort_events) | |
120 | goto failed; | |
fc5c0345 | 121 | |
6f9961ac | 122 | *rbd_data_ptr = rbd; |
fc5c0345 DG |
123 | return 0; |
124 | ||
125 | failed: | |
5a4adfd2 JQ |
126 | if (rbd) { |
127 | if (rbd->aio_events) | |
128 | free(rbd->aio_events); | |
129 | if (rbd->sort_events) | |
130 | free(rbd->sort_events); | |
6f9961ac | 131 | free(rbd); |
5a4adfd2 | 132 | } |
fc5c0345 DG |
133 | return 1; |
134 | ||
135 | } | |
136 | ||
73ae55c3 JA |
137 | #ifdef CONFIG_RBD_POLL |
138 | static bool _fio_rbd_setup_poll(struct rbd_data *rbd) | |
139 | { | |
140 | int r; | |
141 | ||
142 | /* add for rbd poll */ | |
4001b1ba | 143 | rbd->fd = eventfd(0, EFD_SEMAPHORE); |
73ae55c3 JA |
144 | if (rbd->fd < 0) { |
145 | log_err("eventfd failed.\n"); | |
146 | return false; | |
147 | } | |
148 | ||
149 | r = rbd_set_image_notification(rbd->image, rbd->fd, EVENT_TYPE_EVENTFD); | |
150 | if (r < 0) { | |
151 | log_err("rbd_set_image_notification failed.\n"); | |
152 | close(rbd->fd); | |
153 | rbd->fd = -1; | |
154 | return false; | |
155 | } | |
156 | ||
157 | return true; | |
158 | } | |
159 | #else | |
160 | static bool _fio_rbd_setup_poll(struct rbd_data *rbd) | |
161 | { | |
162 | return true; | |
163 | } | |
164 | #endif | |
165 | ||
fc5c0345 DG |
166 | static int _fio_rbd_connect(struct thread_data *td) |
167 | { | |
565e784d | 168 | struct rbd_data *rbd = td->io_ops_data; |
fc5c0345 DG |
169 | struct rbd_options *o = td->eo; |
170 | int r; | |
171 | ||
6e20c6e7 T |
172 | if (o->cluster_name) { |
173 | char *client_name = NULL; | |
174 | ||
175 | /* | |
fc002f14 | 176 | * If we specify cluster name, the rados_create2 |
6e20c6e7 T |
177 | * will not assume 'client.'. name is considered |
178 | * as a full type.id namestr | |
179 | */ | |
89e5ec96 JQ |
180 | if (o->client_name) { |
181 | if (!index(o->client_name, '.')) { | |
182 | client_name = calloc(1, strlen("client.") + | |
183 | strlen(o->client_name) + 1); | |
184 | strcat(client_name, "client."); | |
6ec0dcd1 JQ |
185 | strcat(client_name, o->client_name); |
186 | } else { | |
187 | client_name = o->client_name; | |
89e5ec96 | 188 | } |
6e20c6e7 | 189 | } |
6ec0dcd1 | 190 | |
6e20c6e7 | 191 | r = rados_create2(&rbd->cluster, o->cluster_name, |
6ec0dcd1 JQ |
192 | client_name, 0); |
193 | ||
194 | if (client_name && !index(o->client_name, '.')) | |
195 | free(client_name); | |
6e20c6e7 T |
196 | } else |
197 | r = rados_create(&rbd->cluster, o->client_name); | |
198 | ||
fc5c0345 DG |
199 | if (r < 0) { |
200 | log_err("rados_create failed.\n"); | |
201 | goto failed_early; | |
202 | } | |
de8f8675 DM |
203 | if (o->pool_name == NULL) { |
204 | log_err("rbd pool name must be provided.\n"); | |
205 | goto failed_early; | |
206 | } | |
207 | if (!o->rbd_name) { | |
208 | log_err("rbdname must be provided.\n"); | |
209 | goto failed_early; | |
210 | } | |
fc5c0345 | 211 | |
6f9961ac | 212 | r = rados_conf_read_file(rbd->cluster, NULL); |
fc5c0345 DG |
213 | if (r < 0) { |
214 | log_err("rados_conf_read_file failed.\n"); | |
215 | goto failed_early; | |
216 | } | |
217 | ||
6f9961ac | 218 | r = rados_connect(rbd->cluster); |
fc5c0345 DG |
219 | if (r < 0) { |
220 | log_err("rados_connect failed.\n"); | |
221 | goto failed_shutdown; | |
222 | } | |
223 | ||
6f9961ac | 224 | r = rados_ioctx_create(rbd->cluster, o->pool_name, &rbd->io_ctx); |
fc5c0345 DG |
225 | if (r < 0) { |
226 | log_err("rados_ioctx_create failed.\n"); | |
227 | goto failed_shutdown; | |
228 | } | |
229 | ||
ee584d96 JD |
230 | if (td->o.odirect) { |
231 | r = rados_conf_set(rbd->cluster, "rbd_cache", "false"); | |
232 | if (r < 0) { | |
233 | log_info("failed to disable RBD in-memory cache\n"); | |
234 | } | |
235 | } | |
236 | ||
6f9961ac | 237 | r = rbd_open(rbd->io_ctx, o->rbd_name, &rbd->image, NULL /*snap */ ); |
fc5c0345 DG |
238 | if (r < 0) { |
239 | log_err("rbd_open failed.\n"); | |
240 | goto failed_open; | |
241 | } | |
53b6c979 | 242 | |
8e4e151f JD |
243 | if (!td->o.odirect) { |
244 | /* | |
245 | * ensure cache enables writeback/around mode unless explicitly | |
246 | * configured for writethrough mode | |
247 | */ | |
248 | r = rbd_flush(rbd->image); | |
249 | if (r < 0) { | |
250 | log_info("rbd: failed to issue initial flush\n"); | |
251 | } | |
252 | } | |
253 | ||
73ae55c3 JA |
254 | if (!_fio_rbd_setup_poll(rbd)) |
255 | goto failed_poll; | |
53b6c979 | 256 | |
fc5c0345 DG |
257 | return 0; |
258 | ||
73ae55c3 JA |
259 | failed_poll: |
260 | rbd_close(rbd->image); | |
261 | rbd->image = NULL; | |
fc5c0345 | 262 | failed_open: |
6f9961ac JA |
263 | rados_ioctx_destroy(rbd->io_ctx); |
264 | rbd->io_ctx = NULL; | |
fc5c0345 | 265 | failed_shutdown: |
6f9961ac JA |
266 | rados_shutdown(rbd->cluster); |
267 | rbd->cluster = NULL; | |
fc5c0345 DG |
268 | failed_early: |
269 | return 1; | |
270 | } | |
271 | ||
6f9961ac | 272 | static void _fio_rbd_disconnect(struct rbd_data *rbd) |
fc5c0345 | 273 | { |
6f9961ac | 274 | if (!rbd) |
fc5c0345 DG |
275 | return; |
276 | ||
53b6c979 | 277 | /* close eventfd */ |
73ae55c3 | 278 | if (rbd->fd != -1) { |
53b6c979 PL |
279 | close(rbd->fd); |
280 | rbd->fd = -1; | |
281 | } | |
53b6c979 | 282 | |
fc5c0345 | 283 | /* shutdown everything */ |
6f9961ac JA |
284 | if (rbd->image) { |
285 | rbd_close(rbd->image); | |
286 | rbd->image = NULL; | |
fc5c0345 DG |
287 | } |
288 | ||
6f9961ac JA |
289 | if (rbd->io_ctx) { |
290 | rados_ioctx_destroy(rbd->io_ctx); | |
291 | rbd->io_ctx = NULL; | |
fc5c0345 DG |
292 | } |
293 | ||
6f9961ac JA |
294 | if (rbd->cluster) { |
295 | rados_shutdown(rbd->cluster); | |
296 | rbd->cluster = NULL; | |
fc5c0345 DG |
297 | } |
298 | } | |
299 | ||
d8b64af2 | 300 | static void _fio_rbd_finish_aiocb(rbd_completion_t comp, void *data) |
fc5c0345 | 301 | { |
dbf388d2 JA |
302 | struct fio_rbd_iou *fri = data; |
303 | struct io_u *io_u = fri->io_u; | |
d8b64af2 | 304 | ssize_t ret; |
fc5c0345 | 305 | |
d8b64af2 JA |
306 | /* |
307 | * Looks like return value is 0 for success, or < 0 for | |
308 | * a specific error. So we have to assume that it can't do | |
309 | * partial completions. | |
310 | */ | |
311 | ret = rbd_aio_get_return_value(fri->completion); | |
312 | if (ret < 0) { | |
d7658ced | 313 | io_u->error = -ret; |
d8b64af2 JA |
314 | io_u->resid = io_u->xfer_buflen; |
315 | } else | |
316 | io_u->error = 0; | |
20cf5aab JD |
317 | |
318 | fri->io_complete = 1; | |
d8b64af2 | 319 | } |
fc5c0345 | 320 | |
d8b64af2 JA |
321 | static struct io_u *fio_rbd_event(struct thread_data *td, int event) |
322 | { | |
565e784d | 323 | struct rbd_data *rbd = td->io_ops_data; |
fc5c0345 | 324 | |
6f9961ac | 325 | return rbd->aio_events[event]; |
fc5c0345 DG |
326 | } |
327 | ||
6f9961ac | 328 | static inline int fri_check_complete(struct rbd_data *rbd, struct io_u *io_u, |
d8b64af2 | 329 | unsigned int *events) |
fc5c0345 | 330 | { |
d8b64af2 | 331 | struct fio_rbd_iou *fri = io_u->engine_data; |
fc5c0345 | 332 | |
b8ecbef6 | 333 | if (fri->io_complete) { |
d8b64af2 | 334 | fri->io_seen = 1; |
6f9961ac | 335 | rbd->aio_events[*events] = io_u; |
d8b64af2 | 336 | (*events)++; |
fc5c0345 | 337 | |
d8b64af2 JA |
338 | rbd_aio_release(fri->completion); |
339 | return 1; | |
340 | } | |
fc5c0345 | 341 | |
d8b64af2 | 342 | return 0; |
fc5c0345 DG |
343 | } |
344 | ||
c7cc1c75 | 345 | #ifndef CONFIG_RBD_POLL |
6f9961ac JA |
346 | static inline int rbd_io_u_seen(struct io_u *io_u) |
347 | { | |
348 | struct fio_rbd_iou *fri = io_u->engine_data; | |
349 | ||
350 | return fri->io_seen; | |
351 | } | |
c7cc1c75 | 352 | #endif |
6f9961ac JA |
353 | |
354 | static void rbd_io_u_wait_complete(struct io_u *io_u) | |
355 | { | |
356 | struct fio_rbd_iou *fri = io_u->engine_data; | |
357 | ||
358 | rbd_aio_wait_for_complete(fri->completion); | |
359 | } | |
360 | ||
361 | static int rbd_io_u_cmp(const void *p1, const void *p2) | |
362 | { | |
363 | const struct io_u **a = (const struct io_u **) p1; | |
364 | const struct io_u **b = (const struct io_u **) p2; | |
365 | uint64_t at, bt; | |
366 | ||
367 | at = utime_since_now(&(*a)->start_time); | |
368 | bt = utime_since_now(&(*b)->start_time); | |
369 | ||
370 | if (at < bt) | |
371 | return -1; | |
372 | else if (at == bt) | |
373 | return 0; | |
374 | else | |
375 | return 1; | |
376 | } | |
377 | ||
d8b64af2 JA |
378 | static int rbd_iter_events(struct thread_data *td, unsigned int *events, |
379 | unsigned int min_evts, int wait) | |
82340a9f | 380 | { |
565e784d | 381 | struct rbd_data *rbd = td->io_ops_data; |
d8b64af2 JA |
382 | unsigned int this_events = 0; |
383 | struct io_u *io_u; | |
53b6c979 PL |
384 | int i, sidx = 0; |
385 | ||
386 | #ifdef CONFIG_RBD_POLL | |
387 | int ret = 0; | |
388 | int event_num = 0; | |
389 | struct fio_rbd_iou *fri = NULL; | |
390 | rbd_completion_t comps[min_evts]; | |
4001b1ba JD |
391 | uint64_t counter; |
392 | bool completed; | |
82340a9f | 393 | |
53b6c979 PL |
394 | struct pollfd pfd; |
395 | pfd.fd = rbd->fd; | |
396 | pfd.events = POLLIN; | |
397 | ||
4001b1ba | 398 | ret = poll(&pfd, 1, wait ? -1 : 0); |
73ae55c3 | 399 | if (ret <= 0) |
53b6c979 | 400 | return 0; |
4001b1ba JD |
401 | if (!(pfd.revents & POLLIN)) |
402 | return 0; | |
53b6c979 PL |
403 | |
404 | event_num = rbd_poll_io_events(rbd->image, comps, min_evts); | |
405 | ||
406 | for (i = 0; i < event_num; i++) { | |
407 | fri = rbd_aio_get_arg(comps[i]); | |
408 | io_u = fri->io_u; | |
4001b1ba JD |
409 | |
410 | /* best effort to decrement the semaphore */ | |
411 | ret = read(rbd->fd, &counter, sizeof(counter)); | |
412 | if (ret <= 0) | |
413 | log_err("rbd_iter_events failed to decrement semaphore.\n"); | |
414 | ||
415 | completed = fri_check_complete(rbd, io_u, events); | |
416 | assert(completed); | |
417 | ||
418 | this_events++; | |
419 | } | |
53b6c979 | 420 | #else |
d8b64af2 | 421 | io_u_qiter(&td->io_u_all, io_u, i) { |
d8b64af2 JA |
422 | if (!(io_u->flags & IO_U_F_FLIGHT)) |
423 | continue; | |
6f9961ac | 424 | if (rbd_io_u_seen(io_u)) |
d8b64af2 | 425 | continue; |
82340a9f | 426 | |
6f9961ac | 427 | if (fri_check_complete(rbd, io_u, events)) |
d8b64af2 | 428 | this_events++; |
6f9961ac JA |
429 | else if (wait) |
430 | rbd->sort_events[sidx++] = io_u; | |
431 | } | |
4001b1ba | 432 | #endif |
82340a9f | 433 | |
6f9961ac JA |
434 | if (!wait || !sidx) |
435 | return this_events; | |
436 | ||
437 | /* | |
438 | * Sort events, oldest issue first, then wait on as many as we | |
439 | * need in order of age. If we have enough events, stop waiting, | |
440 | * and just check if any of the older ones are done. | |
441 | */ | |
442 | if (sidx > 1) | |
443 | qsort(rbd->sort_events, sidx, sizeof(struct io_u *), rbd_io_u_cmp); | |
444 | ||
445 | for (i = 0; i < sidx; i++) { | |
446 | io_u = rbd->sort_events[i]; | |
447 | ||
448 | if (fri_check_complete(rbd, io_u, events)) { | |
449 | this_events++; | |
450 | continue; | |
d8b64af2 | 451 | } |
6f9961ac JA |
452 | |
453 | /* | |
454 | * Stop waiting when we have enough, but continue checking | |
455 | * all pending IOs if they are complete. | |
456 | */ | |
d8b64af2 | 457 | if (*events >= min_evts) |
6f9961ac JA |
458 | continue; |
459 | ||
460 | rbd_io_u_wait_complete(io_u); | |
461 | ||
462 | if (fri_check_complete(rbd, io_u, events)) | |
463 | this_events++; | |
d8b64af2 | 464 | } |
fc5c0345 | 465 | |
d8b64af2 | 466 | return this_events; |
fc5c0345 DG |
467 | } |
468 | ||
469 | static int fio_rbd_getevents(struct thread_data *td, unsigned int min, | |
1f440ece | 470 | unsigned int max, const struct timespec *t) |
fc5c0345 | 471 | { |
d8b64af2 | 472 | unsigned int this_events, events = 0; |
d7d702c7 | 473 | struct rbd_options *o = td->eo; |
d8b64af2 | 474 | int wait = 0; |
fc5c0345 DG |
475 | |
476 | do { | |
d8b64af2 | 477 | this_events = rbd_iter_events(td, &events, min, wait); |
fc5c0345 | 478 | |
d8b64af2 | 479 | if (events >= min) |
fc5c0345 | 480 | break; |
d8b64af2 JA |
481 | if (this_events) |
482 | continue; | |
fc5c0345 | 483 | |
d7d702c7 JA |
484 | if (!o->busy_poll) |
485 | wait = 1; | |
486 | else | |
487 | nop; | |
fc5c0345 DG |
488 | } while (1); |
489 | ||
490 | return events; | |
491 | } | |
492 | ||
2e4ef4fb JA |
493 | static enum fio_q_status fio_rbd_queue(struct thread_data *td, |
494 | struct io_u *io_u) | |
fc5c0345 | 495 | { |
565e784d | 496 | struct rbd_data *rbd = td->io_ops_data; |
d8b64af2 JA |
497 | struct fio_rbd_iou *fri = io_u->engine_data; |
498 | int r = -1; | |
fc5c0345 DG |
499 | |
500 | fio_ro_check(td, io_u); | |
501 | ||
d8b64af2 | 502 | fri->io_seen = 0; |
b8ecbef6 | 503 | fri->io_complete = 0; |
d8b64af2 | 504 | |
dbf388d2 | 505 | r = rbd_aio_create_completion(fri, _fio_rbd_finish_aiocb, |
d8b64af2 | 506 | &fri->completion); |
dbf388d2 JA |
507 | if (r < 0) { |
508 | log_err("rbd_aio_create_completion failed.\n"); | |
509 | goto failed; | |
510 | } | |
fc5c0345 | 511 | |
dbf388d2 | 512 | if (io_u->ddir == DDIR_WRITE) { |
6f9961ac JA |
513 | r = rbd_aio_write(rbd->image, io_u->offset, io_u->xfer_buflen, |
514 | io_u->xfer_buf, fri->completion); | |
fc5c0345 DG |
515 | if (r < 0) { |
516 | log_err("rbd_aio_write failed.\n"); | |
dbf388d2 | 517 | goto failed_comp; |
fc5c0345 DG |
518 | } |
519 | ||
520 | } else if (io_u->ddir == DDIR_READ) { | |
6f9961ac JA |
521 | r = rbd_aio_read(rbd->image, io_u->offset, io_u->xfer_buflen, |
522 | io_u->xfer_buf, fri->completion); | |
fc5c0345 DG |
523 | |
524 | if (r < 0) { | |
525 | log_err("rbd_aio_read failed.\n"); | |
dbf388d2 | 526 | goto failed_comp; |
fc5c0345 | 527 | } |
dbf388d2 | 528 | } else if (io_u->ddir == DDIR_TRIM) { |
6f9961ac JA |
529 | r = rbd_aio_discard(rbd->image, io_u->offset, |
530 | io_u->xfer_buflen, fri->completion); | |
82340a9f | 531 | if (r < 0) { |
dbf388d2 JA |
532 | log_err("rbd_aio_discard failed.\n"); |
533 | goto failed_comp; | |
82340a9f | 534 | } |
dbf388d2 | 535 | } else if (io_u->ddir == DDIR_SYNC) { |
6f9961ac | 536 | r = rbd_aio_flush(rbd->image, fri->completion); |
fc5c0345 DG |
537 | if (r < 0) { |
538 | log_err("rbd_flush failed.\n"); | |
dbf388d2 | 539 | goto failed_comp; |
fc5c0345 | 540 | } |
fc5c0345 DG |
541 | } else { |
542 | dprint(FD_IO, "%s: Warning: unhandled ddir: %d\n", __func__, | |
543 | io_u->ddir); | |
ef2f4a50 | 544 | r = -EINVAL; |
dbf388d2 | 545 | goto failed_comp; |
fc5c0345 DG |
546 | } |
547 | ||
548 | return FIO_Q_QUEUED; | |
dbf388d2 JA |
549 | failed_comp: |
550 | rbd_aio_release(fri->completion); | |
fc5c0345 | 551 | failed: |
d7658ced | 552 | io_u->error = -r; |
fc5c0345 DG |
553 | td_verror(td, io_u->error, "xfer"); |
554 | return FIO_Q_COMPLETED; | |
555 | } | |
556 | ||
557 | static int fio_rbd_init(struct thread_data *td) | |
558 | { | |
559 | int r; | |
e0f35b53 | 560 | struct rbd_data *rbd = td->io_ops_data; |
35448275 JA |
561 | |
562 | if (rbd->connected) | |
e0f35b53 | 563 | return 0; |
fc5c0345 DG |
564 | |
565 | r = _fio_rbd_connect(td); | |
566 | if (r) { | |
567 | log_err("fio_rbd_connect failed, return code: %d .\n", r); | |
568 | goto failed; | |
569 | } | |
570 | ||
571 | return 0; | |
572 | ||
573 | failed: | |
574 | return 1; | |
fc5c0345 DG |
575 | } |
576 | ||
577 | static void fio_rbd_cleanup(struct thread_data *td) | |
578 | { | |
565e784d | 579 | struct rbd_data *rbd = td->io_ops_data; |
fc5c0345 | 580 | |
6f9961ac JA |
581 | if (rbd) { |
582 | _fio_rbd_disconnect(rbd); | |
583 | free(rbd->aio_events); | |
584 | free(rbd->sort_events); | |
585 | free(rbd); | |
fc5c0345 | 586 | } |
fc5c0345 DG |
587 | } |
588 | ||
589 | static int fio_rbd_setup(struct thread_data *td) | |
590 | { | |
fc5c0345 DG |
591 | rbd_image_info_t info; |
592 | struct fio_file *f; | |
6f9961ac | 593 | struct rbd_data *rbd = NULL; |
6f9961ac | 594 | int r; |
fc5c0345 | 595 | |
fc5c0345 | 596 | /* allocate engine specific structure to deal with librbd. */ |
6f9961ac | 597 | r = _fio_setup_rbd_data(td, &rbd); |
fc5c0345 DG |
598 | if (r) { |
599 | log_err("fio_setup_rbd_data failed.\n"); | |
600 | goto cleanup; | |
601 | } | |
565e784d | 602 | td->io_ops_data = rbd; |
fc5c0345 | 603 | |
d8b64af2 JA |
604 | /* librbd does not allow us to run first in the main thread and later |
605 | * in a fork child. It needs to be the same process context all the | |
606 | * time. | |
fc5c0345 DG |
607 | */ |
608 | td->o.use_thread = 1; | |
609 | ||
610 | /* connect in the main thread to determine to determine | |
611 | * the size of the given RADOS block device. And disconnect | |
612 | * later on. | |
613 | */ | |
614 | r = _fio_rbd_connect(td); | |
615 | if (r) { | |
616 | log_err("fio_rbd_connect failed.\n"); | |
617 | goto cleanup; | |
618 | } | |
e0f35b53 | 619 | rbd->connected = true; |
fc5c0345 DG |
620 | |
621 | /* get size of the RADOS block device */ | |
6f9961ac | 622 | r = rbd_stat(rbd->image, &info, sizeof(info)); |
fc5c0345 DG |
623 | if (r < 0) { |
624 | log_err("rbd_status failed.\n"); | |
6747f469 | 625 | goto cleanup; |
206c546d PL |
626 | } else if (info.size == 0) { |
627 | log_err("image size should be larger than zero.\n"); | |
628 | r = -EINVAL; | |
6747f469 | 629 | goto cleanup; |
fc5c0345 | 630 | } |
206c546d | 631 | |
50ea012c | 632 | dprint(FD_IO, "rbd-engine: image size: %" PRIu64 "\n", info.size); |
fc5c0345 DG |
633 | |
634 | /* taken from "net" engine. Pretend we deal with files, | |
635 | * even if we do not have any ideas about files. | |
fc002f14 | 636 | * The size of the RBD is set instead of an artificial file. |
fc5c0345 DG |
637 | */ |
638 | if (!td->files_index) { | |
5903e7b7 | 639 | add_file(td, td->o.filename ? : "rbd", 0, 0); |
fc5c0345 | 640 | td->o.nr_files = td->o.nr_files ? : 1; |
b53f2c54 | 641 | td->o.open_files++; |
fc5c0345 DG |
642 | } |
643 | f = td->files[0]; | |
644 | f->real_file_size = info.size; | |
645 | ||
fc5c0345 DG |
646 | return 0; |
647 | ||
fc5c0345 DG |
648 | cleanup: |
649 | fio_rbd_cleanup(td); | |
650 | return r; | |
651 | } | |
652 | ||
653 | static int fio_rbd_open(struct thread_data *td, struct fio_file *f) | |
654 | { | |
655 | return 0; | |
656 | } | |
657 | ||
d9b100fc JA |
658 | static int fio_rbd_invalidate(struct thread_data *td, struct fio_file *f) |
659 | { | |
903b2812 | 660 | #if defined(CONFIG_RBD_INVAL) |
565e784d | 661 | struct rbd_data *rbd = td->io_ops_data; |
903b2812 | 662 | |
6f9961ac | 663 | return rbd_invalidate_cache(rbd->image); |
903b2812 | 664 | #else |
d9b100fc | 665 | return 0; |
903b2812 | 666 | #endif |
d9b100fc JA |
667 | } |
668 | ||
fc5c0345 DG |
669 | static void fio_rbd_io_u_free(struct thread_data *td, struct io_u *io_u) |
670 | { | |
d8b64af2 | 671 | struct fio_rbd_iou *fri = io_u->engine_data; |
fc5c0345 | 672 | |
d8b64af2 | 673 | if (fri) { |
fc5c0345 | 674 | io_u->engine_data = NULL; |
d8b64af2 | 675 | free(fri); |
fc5c0345 DG |
676 | } |
677 | } | |
678 | ||
679 | static int fio_rbd_io_u_init(struct thread_data *td, struct io_u *io_u) | |
680 | { | |
d8b64af2 | 681 | struct fio_rbd_iou *fri; |
fc5c0345 | 682 | |
d8b64af2 JA |
683 | fri = calloc(1, sizeof(*fri)); |
684 | fri->io_u = io_u; | |
685 | io_u->engine_data = fri; | |
fc5c0345 DG |
686 | return 0; |
687 | } | |
688 | ||
5a8a6a03 | 689 | FIO_STATIC struct ioengine_ops ioengine = { |
d9b100fc JA |
690 | .name = "rbd", |
691 | .version = FIO_IOOPS_VERSION, | |
692 | .setup = fio_rbd_setup, | |
693 | .init = fio_rbd_init, | |
694 | .queue = fio_rbd_queue, | |
695 | .getevents = fio_rbd_getevents, | |
696 | .event = fio_rbd_event, | |
697 | .cleanup = fio_rbd_cleanup, | |
698 | .open_file = fio_rbd_open, | |
699 | .invalidate = fio_rbd_invalidate, | |
700 | .options = options, | |
701 | .io_u_init = fio_rbd_io_u_init, | |
702 | .io_u_free = fio_rbd_io_u_free, | |
703 | .option_struct_size = sizeof(struct rbd_options), | |
fc5c0345 DG |
704 | }; |
705 | ||
706 | static void fio_init fio_rbd_register(void) | |
707 | { | |
708 | register_ioengine(&ioengine); | |
709 | } | |
710 | ||
711 | static void fio_exit fio_rbd_unregister(void) | |
712 | { | |
713 | unregister_ioengine(&ioengine); | |
714 | } |