| 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" |
| 11 | |
| 12 | struct fio_rbd_iou { |
| 13 | struct io_u *io_u; |
| 14 | rbd_completion_t completion; |
| 15 | int io_seen; |
| 16 | }; |
| 17 | |
| 18 | struct rbd_data { |
| 19 | rados_t cluster; |
| 20 | rados_ioctx_t io_ctx; |
| 21 | rbd_image_t image; |
| 22 | struct io_u **aio_events; |
| 23 | }; |
| 24 | |
| 25 | struct rbd_options { |
| 26 | char *rbd_name; |
| 27 | char *pool_name; |
| 28 | char *client_name; |
| 29 | }; |
| 30 | |
| 31 | static struct fio_option options[] = { |
| 32 | { |
| 33 | .name = "rbdname", |
| 34 | .lname = "rbd engine rbdname", |
| 35 | .type = FIO_OPT_STR_STORE, |
| 36 | .help = "RBD name for RBD engine", |
| 37 | .off1 = offsetof(struct rbd_options, rbd_name), |
| 38 | .category = FIO_OPT_C_ENGINE, |
| 39 | .group = FIO_OPT_G_RBD, |
| 40 | }, |
| 41 | { |
| 42 | .name = "pool", |
| 43 | .lname = "rbd engine pool", |
| 44 | .type = FIO_OPT_STR_STORE, |
| 45 | .help = "Name of the pool hosting the RBD for the RBD engine", |
| 46 | .off1 = offsetof(struct rbd_options, pool_name), |
| 47 | .category = FIO_OPT_C_ENGINE, |
| 48 | .group = FIO_OPT_G_RBD, |
| 49 | }, |
| 50 | { |
| 51 | .name = "clientname", |
| 52 | .lname = "rbd engine clientname", |
| 53 | .type = FIO_OPT_STR_STORE, |
| 54 | .help = "Name of the ceph client to access the RBD for the RBD engine", |
| 55 | .off1 = offsetof(struct rbd_options, client_name), |
| 56 | .category = FIO_OPT_C_ENGINE, |
| 57 | .group = FIO_OPT_G_RBD, |
| 58 | }, |
| 59 | { |
| 60 | .name = NULL, |
| 61 | }, |
| 62 | }; |
| 63 | |
| 64 | static int _fio_setup_rbd_data(struct thread_data *td, |
| 65 | struct rbd_data **rbd_data_ptr) |
| 66 | { |
| 67 | struct rbd_data *rbd_data; |
| 68 | |
| 69 | if (td->io_ops->data) |
| 70 | return 0; |
| 71 | |
| 72 | rbd_data = malloc(sizeof(struct rbd_data)); |
| 73 | if (!rbd_data) |
| 74 | goto failed; |
| 75 | |
| 76 | memset(rbd_data, 0, sizeof(struct rbd_data)); |
| 77 | |
| 78 | rbd_data->aio_events = malloc(td->o.iodepth * sizeof(struct io_u *)); |
| 79 | if (!rbd_data->aio_events) |
| 80 | goto failed; |
| 81 | |
| 82 | memset(rbd_data->aio_events, 0, td->o.iodepth * sizeof(struct io_u *)); |
| 83 | |
| 84 | *rbd_data_ptr = rbd_data; |
| 85 | |
| 86 | return 0; |
| 87 | |
| 88 | failed: |
| 89 | if (rbd_data) |
| 90 | free(rbd_data); |
| 91 | return 1; |
| 92 | |
| 93 | } |
| 94 | |
| 95 | static int _fio_rbd_connect(struct thread_data *td) |
| 96 | { |
| 97 | struct rbd_data *rbd_data = td->io_ops->data; |
| 98 | struct rbd_options *o = td->eo; |
| 99 | int r; |
| 100 | |
| 101 | r = rados_create(&rbd_data->cluster, o->client_name); |
| 102 | if (r < 0) { |
| 103 | log_err("rados_create failed.\n"); |
| 104 | goto failed_early; |
| 105 | } |
| 106 | |
| 107 | r = rados_conf_read_file(rbd_data->cluster, NULL); |
| 108 | if (r < 0) { |
| 109 | log_err("rados_conf_read_file failed.\n"); |
| 110 | goto failed_early; |
| 111 | } |
| 112 | |
| 113 | r = rados_connect(rbd_data->cluster); |
| 114 | if (r < 0) { |
| 115 | log_err("rados_connect failed.\n"); |
| 116 | goto failed_shutdown; |
| 117 | } |
| 118 | |
| 119 | r = rados_ioctx_create(rbd_data->cluster, o->pool_name, |
| 120 | &rbd_data->io_ctx); |
| 121 | if (r < 0) { |
| 122 | log_err("rados_ioctx_create failed.\n"); |
| 123 | goto failed_shutdown; |
| 124 | } |
| 125 | |
| 126 | r = rbd_open(rbd_data->io_ctx, o->rbd_name, &rbd_data->image, |
| 127 | NULL /*snap */ ); |
| 128 | if (r < 0) { |
| 129 | log_err("rbd_open failed.\n"); |
| 130 | goto failed_open; |
| 131 | } |
| 132 | return 0; |
| 133 | |
| 134 | failed_open: |
| 135 | rados_ioctx_destroy(rbd_data->io_ctx); |
| 136 | rbd_data->io_ctx = NULL; |
| 137 | failed_shutdown: |
| 138 | rados_shutdown(rbd_data->cluster); |
| 139 | rbd_data->cluster = NULL; |
| 140 | failed_early: |
| 141 | return 1; |
| 142 | } |
| 143 | |
| 144 | static void _fio_rbd_disconnect(struct rbd_data *rbd_data) |
| 145 | { |
| 146 | if (!rbd_data) |
| 147 | return; |
| 148 | |
| 149 | /* shutdown everything */ |
| 150 | if (rbd_data->image) { |
| 151 | rbd_close(rbd_data->image); |
| 152 | rbd_data->image = NULL; |
| 153 | } |
| 154 | |
| 155 | if (rbd_data->io_ctx) { |
| 156 | rados_ioctx_destroy(rbd_data->io_ctx); |
| 157 | rbd_data->io_ctx = NULL; |
| 158 | } |
| 159 | |
| 160 | if (rbd_data->cluster) { |
| 161 | rados_shutdown(rbd_data->cluster); |
| 162 | rbd_data->cluster = NULL; |
| 163 | } |
| 164 | } |
| 165 | |
| 166 | static void _fio_rbd_finish_aiocb(rbd_completion_t comp, void *data) |
| 167 | { |
| 168 | struct fio_rbd_iou *fri = data; |
| 169 | struct io_u *io_u = fri->io_u; |
| 170 | ssize_t ret; |
| 171 | |
| 172 | /* |
| 173 | * Looks like return value is 0 for success, or < 0 for |
| 174 | * a specific error. So we have to assume that it can't do |
| 175 | * partial completions. |
| 176 | */ |
| 177 | ret = rbd_aio_get_return_value(fri->completion); |
| 178 | if (ret < 0) { |
| 179 | io_u->error = ret; |
| 180 | io_u->resid = io_u->xfer_buflen; |
| 181 | } else |
| 182 | io_u->error = 0; |
| 183 | } |
| 184 | |
| 185 | static struct io_u *fio_rbd_event(struct thread_data *td, int event) |
| 186 | { |
| 187 | struct rbd_data *rbd_data = td->io_ops->data; |
| 188 | |
| 189 | return rbd_data->aio_events[event]; |
| 190 | } |
| 191 | |
| 192 | static inline int fri_check_complete(struct rbd_data *rbd_data, |
| 193 | struct io_u *io_u, |
| 194 | unsigned int *events) |
| 195 | { |
| 196 | struct fio_rbd_iou *fri = io_u->engine_data; |
| 197 | |
| 198 | if (rbd_aio_is_complete(fri->completion)) { |
| 199 | fri->io_seen = 1; |
| 200 | rbd_data->aio_events[*events] = io_u; |
| 201 | (*events)++; |
| 202 | |
| 203 | rbd_aio_release(fri->completion); |
| 204 | return 1; |
| 205 | } |
| 206 | |
| 207 | return 0; |
| 208 | } |
| 209 | |
| 210 | static int rbd_iter_events(struct thread_data *td, unsigned int *events, |
| 211 | unsigned int min_evts, int wait) |
| 212 | { |
| 213 | struct rbd_data *rbd_data = td->io_ops->data; |
| 214 | unsigned int this_events = 0; |
| 215 | struct io_u *io_u; |
| 216 | int i; |
| 217 | |
| 218 | io_u_qiter(&td->io_u_all, io_u, i) { |
| 219 | struct fio_rbd_iou *fri = io_u->engine_data; |
| 220 | |
| 221 | if (!(io_u->flags & IO_U_F_FLIGHT)) |
| 222 | continue; |
| 223 | if (fri->io_seen) |
| 224 | continue; |
| 225 | |
| 226 | if (fri_check_complete(rbd_data, io_u, events)) |
| 227 | this_events++; |
| 228 | else if (wait) { |
| 229 | rbd_aio_wait_for_complete(fri->completion); |
| 230 | |
| 231 | if (fri_check_complete(rbd_data, io_u, events)) |
| 232 | this_events++; |
| 233 | } |
| 234 | if (*events >= min_evts) |
| 235 | break; |
| 236 | } |
| 237 | |
| 238 | return this_events; |
| 239 | } |
| 240 | |
| 241 | static int fio_rbd_getevents(struct thread_data *td, unsigned int min, |
| 242 | unsigned int max, const struct timespec *t) |
| 243 | { |
| 244 | unsigned int this_events, events = 0; |
| 245 | int wait = 0; |
| 246 | |
| 247 | do { |
| 248 | this_events = rbd_iter_events(td, &events, min, wait); |
| 249 | |
| 250 | if (events >= min) |
| 251 | break; |
| 252 | if (this_events) |
| 253 | continue; |
| 254 | |
| 255 | wait = 1; |
| 256 | } while (1); |
| 257 | |
| 258 | return events; |
| 259 | } |
| 260 | |
| 261 | static int fio_rbd_queue(struct thread_data *td, struct io_u *io_u) |
| 262 | { |
| 263 | struct rbd_data *rbd_data = td->io_ops->data; |
| 264 | struct fio_rbd_iou *fri = io_u->engine_data; |
| 265 | int r = -1; |
| 266 | |
| 267 | fio_ro_check(td, io_u); |
| 268 | |
| 269 | fri->io_seen = 0; |
| 270 | |
| 271 | r = rbd_aio_create_completion(fri, _fio_rbd_finish_aiocb, |
| 272 | &fri->completion); |
| 273 | if (r < 0) { |
| 274 | log_err("rbd_aio_create_completion failed.\n"); |
| 275 | goto failed; |
| 276 | } |
| 277 | |
| 278 | if (io_u->ddir == DDIR_WRITE) { |
| 279 | r = rbd_aio_write(rbd_data->image, io_u->offset, |
| 280 | io_u->xfer_buflen, io_u->xfer_buf, |
| 281 | fri->completion); |
| 282 | if (r < 0) { |
| 283 | log_err("rbd_aio_write failed.\n"); |
| 284 | goto failed_comp; |
| 285 | } |
| 286 | |
| 287 | } else if (io_u->ddir == DDIR_READ) { |
| 288 | r = rbd_aio_read(rbd_data->image, io_u->offset, |
| 289 | io_u->xfer_buflen, io_u->xfer_buf, |
| 290 | fri->completion); |
| 291 | |
| 292 | if (r < 0) { |
| 293 | log_err("rbd_aio_read failed.\n"); |
| 294 | goto failed_comp; |
| 295 | } |
| 296 | } else if (io_u->ddir == DDIR_TRIM) { |
| 297 | r = rbd_aio_discard(rbd_data->image, io_u->offset, |
| 298 | io_u->xfer_buflen, fri->completion); |
| 299 | if (r < 0) { |
| 300 | log_err("rbd_aio_discard failed.\n"); |
| 301 | goto failed_comp; |
| 302 | } |
| 303 | } else if (io_u->ddir == DDIR_SYNC) { |
| 304 | r = rbd_aio_flush(rbd_data->image, fri->completion); |
| 305 | if (r < 0) { |
| 306 | log_err("rbd_flush failed.\n"); |
| 307 | goto failed_comp; |
| 308 | } |
| 309 | } else { |
| 310 | dprint(FD_IO, "%s: Warning: unhandled ddir: %d\n", __func__, |
| 311 | io_u->ddir); |
| 312 | goto failed_comp; |
| 313 | } |
| 314 | |
| 315 | return FIO_Q_QUEUED; |
| 316 | failed_comp: |
| 317 | rbd_aio_release(fri->completion); |
| 318 | failed: |
| 319 | io_u->error = r; |
| 320 | td_verror(td, io_u->error, "xfer"); |
| 321 | return FIO_Q_COMPLETED; |
| 322 | } |
| 323 | |
| 324 | static int fio_rbd_init(struct thread_data *td) |
| 325 | { |
| 326 | int r; |
| 327 | |
| 328 | r = _fio_rbd_connect(td); |
| 329 | if (r) { |
| 330 | log_err("fio_rbd_connect failed, return code: %d .\n", r); |
| 331 | goto failed; |
| 332 | } |
| 333 | |
| 334 | return 0; |
| 335 | |
| 336 | failed: |
| 337 | return 1; |
| 338 | } |
| 339 | |
| 340 | static void fio_rbd_cleanup(struct thread_data *td) |
| 341 | { |
| 342 | struct rbd_data *rbd_data = td->io_ops->data; |
| 343 | |
| 344 | if (rbd_data) { |
| 345 | _fio_rbd_disconnect(rbd_data); |
| 346 | free(rbd_data->aio_events); |
| 347 | free(rbd_data); |
| 348 | } |
| 349 | |
| 350 | } |
| 351 | |
| 352 | static int fio_rbd_setup(struct thread_data *td) |
| 353 | { |
| 354 | int r = 0; |
| 355 | rbd_image_info_t info; |
| 356 | struct fio_file *f; |
| 357 | struct rbd_data *rbd_data = NULL; |
| 358 | int major, minor, extra; |
| 359 | |
| 360 | /* log version of librbd. No cluster connection required. */ |
| 361 | rbd_version(&major, &minor, &extra); |
| 362 | log_info("rbd engine: RBD version: %d.%d.%d\n", major, minor, extra); |
| 363 | |
| 364 | /* allocate engine specific structure to deal with librbd. */ |
| 365 | r = _fio_setup_rbd_data(td, &rbd_data); |
| 366 | if (r) { |
| 367 | log_err("fio_setup_rbd_data failed.\n"); |
| 368 | goto cleanup; |
| 369 | } |
| 370 | td->io_ops->data = rbd_data; |
| 371 | |
| 372 | /* librbd does not allow us to run first in the main thread and later |
| 373 | * in a fork child. It needs to be the same process context all the |
| 374 | * time. |
| 375 | */ |
| 376 | td->o.use_thread = 1; |
| 377 | |
| 378 | /* connect in the main thread to determine to determine |
| 379 | * the size of the given RADOS block device. And disconnect |
| 380 | * later on. |
| 381 | */ |
| 382 | r = _fio_rbd_connect(td); |
| 383 | if (r) { |
| 384 | log_err("fio_rbd_connect failed.\n"); |
| 385 | goto cleanup; |
| 386 | } |
| 387 | |
| 388 | /* get size of the RADOS block device */ |
| 389 | r = rbd_stat(rbd_data->image, &info, sizeof(info)); |
| 390 | if (r < 0) { |
| 391 | log_err("rbd_status failed.\n"); |
| 392 | goto disconnect; |
| 393 | } |
| 394 | dprint(FD_IO, "rbd-engine: image size: %lu\n", info.size); |
| 395 | |
| 396 | /* taken from "net" engine. Pretend we deal with files, |
| 397 | * even if we do not have any ideas about files. |
| 398 | * The size of the RBD is set instead of a artificial file. |
| 399 | */ |
| 400 | if (!td->files_index) { |
| 401 | add_file(td, td->o.filename ? : "rbd", 0, 0); |
| 402 | td->o.nr_files = td->o.nr_files ? : 1; |
| 403 | td->o.open_files++; |
| 404 | } |
| 405 | f = td->files[0]; |
| 406 | f->real_file_size = info.size; |
| 407 | |
| 408 | /* disconnect, then we were only connected to determine |
| 409 | * the size of the RBD. |
| 410 | */ |
| 411 | _fio_rbd_disconnect(rbd_data); |
| 412 | return 0; |
| 413 | |
| 414 | disconnect: |
| 415 | _fio_rbd_disconnect(rbd_data); |
| 416 | cleanup: |
| 417 | fio_rbd_cleanup(td); |
| 418 | return r; |
| 419 | } |
| 420 | |
| 421 | static int fio_rbd_open(struct thread_data *td, struct fio_file *f) |
| 422 | { |
| 423 | return 0; |
| 424 | } |
| 425 | |
| 426 | static int fio_rbd_invalidate(struct thread_data *td, struct fio_file *f) |
| 427 | { |
| 428 | #if defined(CONFIG_RBD_INVAL) |
| 429 | struct rbd_data *rbd_data = td->io_ops->data; |
| 430 | |
| 431 | return rbd_invalidate_cache(rbd_data->image); |
| 432 | #else |
| 433 | return 0; |
| 434 | #endif |
| 435 | } |
| 436 | |
| 437 | static void fio_rbd_io_u_free(struct thread_data *td, struct io_u *io_u) |
| 438 | { |
| 439 | struct fio_rbd_iou *fri = io_u->engine_data; |
| 440 | |
| 441 | if (fri) { |
| 442 | io_u->engine_data = NULL; |
| 443 | free(fri); |
| 444 | } |
| 445 | } |
| 446 | |
| 447 | static int fio_rbd_io_u_init(struct thread_data *td, struct io_u *io_u) |
| 448 | { |
| 449 | struct fio_rbd_iou *fri; |
| 450 | |
| 451 | fri = calloc(1, sizeof(*fri)); |
| 452 | fri->io_u = io_u; |
| 453 | io_u->engine_data = fri; |
| 454 | return 0; |
| 455 | } |
| 456 | |
| 457 | static struct ioengine_ops ioengine = { |
| 458 | .name = "rbd", |
| 459 | .version = FIO_IOOPS_VERSION, |
| 460 | .setup = fio_rbd_setup, |
| 461 | .init = fio_rbd_init, |
| 462 | .queue = fio_rbd_queue, |
| 463 | .getevents = fio_rbd_getevents, |
| 464 | .event = fio_rbd_event, |
| 465 | .cleanup = fio_rbd_cleanup, |
| 466 | .open_file = fio_rbd_open, |
| 467 | .invalidate = fio_rbd_invalidate, |
| 468 | .options = options, |
| 469 | .io_u_init = fio_rbd_io_u_init, |
| 470 | .io_u_free = fio_rbd_io_u_free, |
| 471 | .option_struct_size = sizeof(struct rbd_options), |
| 472 | }; |
| 473 | |
| 474 | static void fio_init fio_rbd_register(void) |
| 475 | { |
| 476 | register_ioengine(&ioengine); |
| 477 | } |
| 478 | |
| 479 | static void fio_exit fio_rbd_unregister(void) |
| 480 | { |
| 481 | unregister_ioengine(&ioengine); |
| 482 | } |