| 1 | /* |
| 2 | * libblkio engine |
| 3 | * |
| 4 | * IO engine using libblkio to access various block I/O interfaces: |
| 5 | * https://gitlab.com/libblkio/libblkio |
| 6 | */ |
| 7 | |
| 8 | #include <assert.h> |
| 9 | #include <errno.h> |
| 10 | #include <stdbool.h> |
| 11 | #include <stddef.h> |
| 12 | #include <stdint.h> |
| 13 | #include <stdlib.h> |
| 14 | #include <string.h> |
| 15 | |
| 16 | #include <blkio.h> |
| 17 | |
| 18 | #include "../fio.h" |
| 19 | #include "../optgroup.h" |
| 20 | #include "../options.h" |
| 21 | #include "../parse.h" |
| 22 | |
| 23 | /* per-process state */ |
| 24 | static struct { |
| 25 | pthread_mutex_t mutex; |
| 26 | int initted_threads; |
| 27 | int initted_hipri_threads; |
| 28 | struct blkio *b; |
| 29 | } proc_state = { PTHREAD_MUTEX_INITIALIZER, 0, 0, NULL }; |
| 30 | |
| 31 | static void fio_blkio_proc_lock(void) { |
| 32 | int ret; |
| 33 | ret = pthread_mutex_lock(&proc_state.mutex); |
| 34 | assert(ret == 0); |
| 35 | } |
| 36 | |
| 37 | static void fio_blkio_proc_unlock(void) { |
| 38 | int ret; |
| 39 | ret = pthread_mutex_unlock(&proc_state.mutex); |
| 40 | assert(ret == 0); |
| 41 | } |
| 42 | |
| 43 | /* per-thread state */ |
| 44 | struct fio_blkio_data { |
| 45 | struct blkioq *q; |
| 46 | int completion_fd; /* may be -1 if not FIO_BLKIO_WAIT_MODE_EVENTFD */ |
| 47 | |
| 48 | bool has_mem_region; /* whether mem_region is valid */ |
| 49 | struct blkio_mem_region mem_region; /* only if allocated by libblkio */ |
| 50 | |
| 51 | struct iovec *iovecs; /* for vectored requests */ |
| 52 | struct blkio_completion *completions; |
| 53 | }; |
| 54 | |
| 55 | enum fio_blkio_wait_mode { |
| 56 | FIO_BLKIO_WAIT_MODE_BLOCK, |
| 57 | FIO_BLKIO_WAIT_MODE_EVENTFD, |
| 58 | FIO_BLKIO_WAIT_MODE_LOOP, |
| 59 | }; |
| 60 | |
| 61 | struct fio_blkio_options { |
| 62 | void *pad; /* option fields must not have offset 0 */ |
| 63 | |
| 64 | char *driver; |
| 65 | |
| 66 | char *path; |
| 67 | char *pre_connect_props; |
| 68 | |
| 69 | int num_entries; |
| 70 | int queue_size; |
| 71 | char *pre_start_props; |
| 72 | |
| 73 | unsigned int hipri; |
| 74 | unsigned int vectored; |
| 75 | unsigned int write_zeroes_on_trim; |
| 76 | enum fio_blkio_wait_mode wait_mode; |
| 77 | unsigned int force_enable_completion_eventfd; |
| 78 | }; |
| 79 | |
| 80 | static struct fio_option options[] = { |
| 81 | { |
| 82 | .name = "libblkio_driver", |
| 83 | .lname = "libblkio driver name", |
| 84 | .type = FIO_OPT_STR_STORE, |
| 85 | .off1 = offsetof(struct fio_blkio_options, driver), |
| 86 | .help = "Name of the driver to be used by libblkio", |
| 87 | .category = FIO_OPT_C_ENGINE, |
| 88 | .group = FIO_OPT_G_LIBBLKIO, |
| 89 | }, |
| 90 | { |
| 91 | .name = "libblkio_path", |
| 92 | .lname = "libblkio \"path\" property", |
| 93 | .type = FIO_OPT_STR_STORE, |
| 94 | .off1 = offsetof(struct fio_blkio_options, path), |
| 95 | .help = "Value to set the \"path\" property to", |
| 96 | .category = FIO_OPT_C_ENGINE, |
| 97 | .group = FIO_OPT_G_LIBBLKIO, |
| 98 | }, |
| 99 | { |
| 100 | .name = "libblkio_pre_connect_props", |
| 101 | .lname = "Additional properties to be set before blkio_connect()", |
| 102 | .type = FIO_OPT_STR_STORE, |
| 103 | .off1 = offsetof(struct fio_blkio_options, pre_connect_props), |
| 104 | .help = "", |
| 105 | .category = FIO_OPT_C_ENGINE, |
| 106 | .group = FIO_OPT_G_LIBBLKIO, |
| 107 | }, |
| 108 | { |
| 109 | .name = "libblkio_num_entries", |
| 110 | .lname = "libblkio \"num-entries\" property", |
| 111 | .type = FIO_OPT_INT, |
| 112 | .off1 = offsetof(struct fio_blkio_options, num_entries), |
| 113 | .help = "Value to set the \"num-entries\" property to", |
| 114 | .minval = 1, |
| 115 | .interval = 1, |
| 116 | .category = FIO_OPT_C_ENGINE, |
| 117 | .group = FIO_OPT_G_LIBBLKIO, |
| 118 | }, |
| 119 | { |
| 120 | .name = "libblkio_queue_size", |
| 121 | .lname = "libblkio \"queue-size\" property", |
| 122 | .type = FIO_OPT_INT, |
| 123 | .off1 = offsetof(struct fio_blkio_options, queue_size), |
| 124 | .help = "Value to set the \"queue-size\" property to", |
| 125 | .minval = 1, |
| 126 | .interval = 1, |
| 127 | .category = FIO_OPT_C_ENGINE, |
| 128 | .group = FIO_OPT_G_LIBBLKIO, |
| 129 | }, |
| 130 | { |
| 131 | .name = "libblkio_pre_start_props", |
| 132 | .lname = "Additional properties to be set before blkio_start()", |
| 133 | .type = FIO_OPT_STR_STORE, |
| 134 | .off1 = offsetof(struct fio_blkio_options, pre_start_props), |
| 135 | .help = "", |
| 136 | .category = FIO_OPT_C_ENGINE, |
| 137 | .group = FIO_OPT_G_LIBBLKIO, |
| 138 | }, |
| 139 | { |
| 140 | .name = "hipri", |
| 141 | .lname = "Use poll queues", |
| 142 | .type = FIO_OPT_STR_SET, |
| 143 | .off1 = offsetof(struct fio_blkio_options, hipri), |
| 144 | .help = "Use poll queues", |
| 145 | .category = FIO_OPT_C_ENGINE, |
| 146 | .group = FIO_OPT_G_LIBBLKIO, |
| 147 | }, |
| 148 | { |
| 149 | .name = "libblkio_vectored", |
| 150 | .lname = "Use blkioq_{readv,writev}()", |
| 151 | .type = FIO_OPT_STR_SET, |
| 152 | .off1 = offsetof(struct fio_blkio_options, vectored), |
| 153 | .help = "Use blkioq_{readv,writev}() instead of blkioq_{read,write}()", |
| 154 | .category = FIO_OPT_C_ENGINE, |
| 155 | .group = FIO_OPT_G_LIBBLKIO, |
| 156 | }, |
| 157 | { |
| 158 | .name = "libblkio_write_zeroes_on_trim", |
| 159 | .lname = "Use blkioq_write_zeroes() for TRIM", |
| 160 | .type = FIO_OPT_STR_SET, |
| 161 | .off1 = offsetof(struct fio_blkio_options, |
| 162 | write_zeroes_on_trim), |
| 163 | .help = "Use blkioq_write_zeroes() for TRIM instead of blkioq_discard()", |
| 164 | .category = FIO_OPT_C_ENGINE, |
| 165 | .group = FIO_OPT_G_LIBBLKIO, |
| 166 | }, |
| 167 | { |
| 168 | .name = "libblkio_wait_mode", |
| 169 | .lname = "How to wait for completions", |
| 170 | .type = FIO_OPT_STR, |
| 171 | .off1 = offsetof(struct fio_blkio_options, wait_mode), |
| 172 | .help = "How to wait for completions", |
| 173 | .def = "block", |
| 174 | .posval = { |
| 175 | { .ival = "block", |
| 176 | .oval = FIO_BLKIO_WAIT_MODE_BLOCK, |
| 177 | .help = "Blocking blkioq_do_io()", |
| 178 | }, |
| 179 | { .ival = "eventfd", |
| 180 | .oval = FIO_BLKIO_WAIT_MODE_EVENTFD, |
| 181 | .help = "Blocking read() on the completion eventfd", |
| 182 | }, |
| 183 | { .ival = "loop", |
| 184 | .oval = FIO_BLKIO_WAIT_MODE_LOOP, |
| 185 | .help = "Busy loop with non-blocking blkioq_do_io()", |
| 186 | }, |
| 187 | }, |
| 188 | .category = FIO_OPT_C_ENGINE, |
| 189 | .group = FIO_OPT_G_LIBBLKIO, |
| 190 | }, |
| 191 | { |
| 192 | .name = "libblkio_force_enable_completion_eventfd", |
| 193 | .lname = "Force enable the completion eventfd, even if unused", |
| 194 | .type = FIO_OPT_STR_SET, |
| 195 | .off1 = offsetof(struct fio_blkio_options, |
| 196 | force_enable_completion_eventfd), |
| 197 | .help = "This can impact performance", |
| 198 | .category = FIO_OPT_C_ENGINE, |
| 199 | .group = FIO_OPT_G_LIBBLKIO, |
| 200 | }, |
| 201 | { |
| 202 | .name = NULL, |
| 203 | }, |
| 204 | }; |
| 205 | |
| 206 | static int fio_blkio_set_props_from_str(struct blkio *b, const char *opt_name, |
| 207 | const char *str) { |
| 208 | int ret = 0; |
| 209 | char *new_str, *name, *value; |
| 210 | |
| 211 | if (!str) |
| 212 | return 0; |
| 213 | |
| 214 | /* iteration can mutate string, so copy it */ |
| 215 | new_str = strdup(str); |
| 216 | if (!new_str) { |
| 217 | log_err("fio: strdup() failed\n"); |
| 218 | return 1; |
| 219 | } |
| 220 | |
| 221 | /* iterate over property name-value pairs */ |
| 222 | while ((name = get_next_str(&new_str))) { |
| 223 | /* split into property name and value */ |
| 224 | value = strchr(name, '='); |
| 225 | if (!value) { |
| 226 | log_err("fio: missing '=' in option %s\n", opt_name); |
| 227 | ret = 1; |
| 228 | break; |
| 229 | } |
| 230 | |
| 231 | *value = '\0'; |
| 232 | ++value; |
| 233 | |
| 234 | /* strip whitespace from property name */ |
| 235 | strip_blank_front(&name); |
| 236 | strip_blank_end(name); |
| 237 | |
| 238 | if (name[0] == '\0') { |
| 239 | log_err("fio: empty property name in option %s\n", |
| 240 | opt_name); |
| 241 | ret = 1; |
| 242 | break; |
| 243 | } |
| 244 | |
| 245 | /* strip whitespace from property value */ |
| 246 | strip_blank_front(&value); |
| 247 | strip_blank_end(value); |
| 248 | |
| 249 | /* set property */ |
| 250 | if (blkio_set_str(b, name, value) != 0) { |
| 251 | log_err("fio: error setting property '%s' to '%s': %s\n", |
| 252 | name, value, blkio_get_error_msg()); |
| 253 | ret = 1; |
| 254 | break; |
| 255 | } |
| 256 | } |
| 257 | |
| 258 | free(new_str); |
| 259 | return ret; |
| 260 | } |
| 261 | |
| 262 | /* |
| 263 | * Log the failure of a libblkio function. |
| 264 | * |
| 265 | * `(void)func` is to ensure `func` exists and prevent typos |
| 266 | */ |
| 267 | #define fio_blkio_log_err(func) \ |
| 268 | ({ \ |
| 269 | (void)func; \ |
| 270 | log_err("fio: %s() failed: %s\n", #func, \ |
| 271 | blkio_get_error_msg()); \ |
| 272 | }) |
| 273 | |
| 274 | static bool possibly_null_strs_equal(const char *a, const char *b) |
| 275 | { |
| 276 | return (!a && !b) || (a && b && strcmp(a, b) == 0); |
| 277 | } |
| 278 | |
| 279 | /* |
| 280 | * Returns the total number of subjobs using the 'libblkio' ioengine and setting |
| 281 | * the 'thread' option in the entire workload that have the given value for the |
| 282 | * 'hipri' option. |
| 283 | */ |
| 284 | static int total_threaded_subjobs(bool hipri) |
| 285 | { |
| 286 | int count = 0; |
| 287 | |
| 288 | for_each_td(td) { |
| 289 | const struct fio_blkio_options *options = td->eo; |
| 290 | if (strcmp(td->o.ioengine, "libblkio") == 0 && |
| 291 | td->o.use_thread && (bool)options->hipri == hipri) |
| 292 | ++count; |
| 293 | } end_for_each(); |
| 294 | |
| 295 | return count; |
| 296 | } |
| 297 | |
| 298 | static struct { |
| 299 | bool set_up; |
| 300 | bool direct; |
| 301 | struct fio_blkio_options opts; |
| 302 | } first_threaded_subjob = { 0 }; |
| 303 | |
| 304 | static void fio_blkio_log_opt_compat_err(const char *option_name) |
| 305 | { |
| 306 | log_err("fio: jobs using engine libblkio and sharing a process must agree on the %s option\n", |
| 307 | option_name); |
| 308 | } |
| 309 | |
| 310 | /* |
| 311 | * If td represents a subjob with option 'thread', check if its options are |
| 312 | * compatible with those of other threaded subjobs that were already set up. |
| 313 | */ |
| 314 | static int fio_blkio_check_opt_compat(struct thread_data *td) |
| 315 | { |
| 316 | const struct fio_blkio_options *options = td->eo, *prev_options; |
| 317 | |
| 318 | if (!td->o.use_thread) |
| 319 | return 0; /* subjob doesn't use 'thread' */ |
| 320 | |
| 321 | if (!first_threaded_subjob.set_up) { |
| 322 | /* first subjob using 'thread', store options for later */ |
| 323 | first_threaded_subjob.set_up = true; |
| 324 | first_threaded_subjob.direct = td->o.odirect; |
| 325 | first_threaded_subjob.opts = *options; |
| 326 | return 0; |
| 327 | } |
| 328 | |
| 329 | /* not first subjob using 'thread', check option compatibility */ |
| 330 | prev_options = &first_threaded_subjob.opts; |
| 331 | |
| 332 | if (td->o.odirect != first_threaded_subjob.direct) { |
| 333 | fio_blkio_log_opt_compat_err("direct/buffered"); |
| 334 | return 1; |
| 335 | } |
| 336 | |
| 337 | if (strcmp(options->driver, prev_options->driver) != 0) { |
| 338 | fio_blkio_log_opt_compat_err("libblkio_driver"); |
| 339 | return 1; |
| 340 | } |
| 341 | |
| 342 | if (!possibly_null_strs_equal(options->path, prev_options->path)) { |
| 343 | fio_blkio_log_opt_compat_err("libblkio_path"); |
| 344 | return 1; |
| 345 | } |
| 346 | |
| 347 | if (!possibly_null_strs_equal(options->pre_connect_props, |
| 348 | prev_options->pre_connect_props)) { |
| 349 | fio_blkio_log_opt_compat_err("libblkio_pre_connect_props"); |
| 350 | return 1; |
| 351 | } |
| 352 | |
| 353 | if (options->num_entries != prev_options->num_entries) { |
| 354 | fio_blkio_log_opt_compat_err("libblkio_num_entries"); |
| 355 | return 1; |
| 356 | } |
| 357 | |
| 358 | if (options->queue_size != prev_options->queue_size) { |
| 359 | fio_blkio_log_opt_compat_err("libblkio_queue_size"); |
| 360 | return 1; |
| 361 | } |
| 362 | |
| 363 | if (!possibly_null_strs_equal(options->pre_start_props, |
| 364 | prev_options->pre_start_props)) { |
| 365 | fio_blkio_log_opt_compat_err("libblkio_pre_start_props"); |
| 366 | return 1; |
| 367 | } |
| 368 | |
| 369 | return 0; |
| 370 | } |
| 371 | |
| 372 | static int fio_blkio_create_and_connect(struct thread_data *td, |
| 373 | struct blkio **out_blkio) |
| 374 | { |
| 375 | const struct fio_blkio_options *options = td->eo; |
| 376 | struct blkio *b; |
| 377 | int ret; |
| 378 | |
| 379 | if (!options->driver) { |
| 380 | log_err("fio: engine libblkio requires option libblkio_driver to be set\n"); |
| 381 | return 1; |
| 382 | } |
| 383 | |
| 384 | if (blkio_create(options->driver, &b) != 0) { |
| 385 | fio_blkio_log_err(blkio_create); |
| 386 | return 1; |
| 387 | } |
| 388 | |
| 389 | /* don't fail if driver doesn't have a "direct" property */ |
| 390 | ret = blkio_set_bool(b, "direct", td->o.odirect); |
| 391 | if (ret != 0 && ret != -ENOENT) { |
| 392 | fio_blkio_log_err(blkio_set_bool); |
| 393 | goto err_blkio_destroy; |
| 394 | } |
| 395 | |
| 396 | if (blkio_set_bool(b, "read-only", read_only) != 0) { |
| 397 | fio_blkio_log_err(blkio_set_bool); |
| 398 | goto err_blkio_destroy; |
| 399 | } |
| 400 | |
| 401 | if (options->path) { |
| 402 | if (blkio_set_str(b, "path", options->path) != 0) { |
| 403 | fio_blkio_log_err(blkio_set_str); |
| 404 | goto err_blkio_destroy; |
| 405 | } |
| 406 | } |
| 407 | |
| 408 | if (fio_blkio_set_props_from_str(b, "libblkio_pre_connect_props", |
| 409 | options->pre_connect_props) != 0) |
| 410 | goto err_blkio_destroy; |
| 411 | |
| 412 | if (blkio_connect(b) != 0) { |
| 413 | fio_blkio_log_err(blkio_connect); |
| 414 | goto err_blkio_destroy; |
| 415 | } |
| 416 | |
| 417 | if (options->num_entries != 0) { |
| 418 | if (blkio_set_int(b, "num-entries", |
| 419 | options->num_entries) != 0) { |
| 420 | fio_blkio_log_err(blkio_set_int); |
| 421 | goto err_blkio_destroy; |
| 422 | } |
| 423 | } |
| 424 | |
| 425 | if (options->queue_size != 0) { |
| 426 | if (blkio_set_int(b, "queue-size", options->queue_size) != 0) { |
| 427 | fio_blkio_log_err(blkio_set_int); |
| 428 | goto err_blkio_destroy; |
| 429 | } |
| 430 | } |
| 431 | |
| 432 | if (fio_blkio_set_props_from_str(b, "libblkio_pre_start_props", |
| 433 | options->pre_start_props) != 0) |
| 434 | goto err_blkio_destroy; |
| 435 | |
| 436 | *out_blkio = b; |
| 437 | return 0; |
| 438 | |
| 439 | err_blkio_destroy: |
| 440 | blkio_destroy(&b); |
| 441 | return 1; |
| 442 | } |
| 443 | |
| 444 | static bool incompatible_threaded_subjob_options = false; |
| 445 | |
| 446 | /* |
| 447 | * This callback determines the device/file size, so it creates and connects a |
| 448 | * blkio instance. But it is invoked from the main thread in the original fio |
| 449 | * process, not from the processes in which jobs will actually run. It thus |
| 450 | * subsequently destroys the blkio, which is recreated in the init() callback. |
| 451 | */ |
| 452 | static int fio_blkio_setup(struct thread_data *td) |
| 453 | { |
| 454 | const struct fio_blkio_options *options = td->eo; |
| 455 | struct blkio *b; |
| 456 | int ret = 0; |
| 457 | uint64_t capacity; |
| 458 | |
| 459 | assert(td->files_index == 1); |
| 460 | |
| 461 | if (fio_blkio_check_opt_compat(td) != 0) { |
| 462 | incompatible_threaded_subjob_options = true; |
| 463 | return 1; |
| 464 | } |
| 465 | |
| 466 | if (options->hipri && |
| 467 | options->wait_mode == FIO_BLKIO_WAIT_MODE_EVENTFD) { |
| 468 | log_err("fio: option hipri is incompatible with option libblkio_wait_mode=eventfd\n"); |
| 469 | return 1; |
| 470 | } |
| 471 | |
| 472 | if (options->hipri && options->force_enable_completion_eventfd) { |
| 473 | log_err("fio: option hipri is incompatible with option libblkio_force_enable_completion_eventfd\n"); |
| 474 | return 1; |
| 475 | } |
| 476 | |
| 477 | if (fio_blkio_create_and_connect(td, &b) != 0) |
| 478 | return 1; |
| 479 | |
| 480 | if (blkio_get_uint64(b, "capacity", &capacity) != 0) { |
| 481 | fio_blkio_log_err(blkio_get_uint64); |
| 482 | ret = 1; |
| 483 | goto out_blkio_destroy; |
| 484 | } |
| 485 | |
| 486 | td->files[0]->real_file_size = capacity; |
| 487 | fio_file_set_size_known(td->files[0]); |
| 488 | |
| 489 | out_blkio_destroy: |
| 490 | blkio_destroy(&b); |
| 491 | return ret; |
| 492 | } |
| 493 | |
| 494 | static int fio_blkio_init(struct thread_data *td) |
| 495 | { |
| 496 | const struct fio_blkio_options *options = td->eo; |
| 497 | struct fio_blkio_data *data; |
| 498 | int flags; |
| 499 | |
| 500 | if (td->o.use_thread && incompatible_threaded_subjob_options) { |
| 501 | /* |
| 502 | * Different subjobs using option 'thread' specified |
| 503 | * incompatible options. We don't know which configuration |
| 504 | * should win, so we just fail all such subjobs. |
| 505 | */ |
| 506 | return 1; |
| 507 | } |
| 508 | |
| 509 | /* |
| 510 | * Request enqueueing is fast, and it's not possible to know exactly |
| 511 | * when a request is submitted, so never report submission latencies. |
| 512 | */ |
| 513 | td->o.disable_slat = 1; |
| 514 | |
| 515 | data = calloc(1, sizeof(*data)); |
| 516 | if (!data) { |
| 517 | log_err("fio: calloc() failed\n"); |
| 518 | return 1; |
| 519 | } |
| 520 | |
| 521 | data->iovecs = calloc(td->o.iodepth, sizeof(data->iovecs[0])); |
| 522 | data->completions = calloc(td->o.iodepth, sizeof(data->completions[0])); |
| 523 | if (!data->iovecs || !data->completions) { |
| 524 | log_err("fio: calloc() failed\n"); |
| 525 | goto err_free; |
| 526 | } |
| 527 | |
| 528 | fio_blkio_proc_lock(); |
| 529 | |
| 530 | if (proc_state.initted_threads == 0) { |
| 531 | /* initialize per-process blkio */ |
| 532 | int num_queues, num_poll_queues; |
| 533 | |
| 534 | if (td->o.use_thread) { |
| 535 | num_queues = total_threaded_subjobs(false); |
| 536 | num_poll_queues = total_threaded_subjobs(true); |
| 537 | } else { |
| 538 | num_queues = options->hipri ? 0 : 1; |
| 539 | num_poll_queues = options->hipri ? 1 : 0; |
| 540 | } |
| 541 | |
| 542 | if (fio_blkio_create_and_connect(td, &proc_state.b) != 0) |
| 543 | goto err_unlock; |
| 544 | |
| 545 | if (blkio_set_int(proc_state.b, "num-queues", |
| 546 | num_queues) != 0) { |
| 547 | fio_blkio_log_err(blkio_set_int); |
| 548 | goto err_blkio_destroy; |
| 549 | } |
| 550 | |
| 551 | if (blkio_set_int(proc_state.b, "num-poll-queues", |
| 552 | num_poll_queues) != 0) { |
| 553 | fio_blkio_log_err(blkio_set_int); |
| 554 | goto err_blkio_destroy; |
| 555 | } |
| 556 | |
| 557 | if (blkio_start(proc_state.b) != 0) { |
| 558 | fio_blkio_log_err(blkio_start); |
| 559 | goto err_blkio_destroy; |
| 560 | } |
| 561 | } |
| 562 | |
| 563 | if (options->hipri) { |
| 564 | int i = proc_state.initted_hipri_threads; |
| 565 | data->q = blkio_get_poll_queue(proc_state.b, i); |
| 566 | } else { |
| 567 | int i = proc_state.initted_threads - |
| 568 | proc_state.initted_hipri_threads; |
| 569 | data->q = blkio_get_queue(proc_state.b, i); |
| 570 | } |
| 571 | |
| 572 | if (options->wait_mode == FIO_BLKIO_WAIT_MODE_EVENTFD || |
| 573 | options->force_enable_completion_eventfd) { |
| 574 | /* enable completion fd and make it blocking */ |
| 575 | blkioq_set_completion_fd_enabled(data->q, true); |
| 576 | data->completion_fd = blkioq_get_completion_fd(data->q); |
| 577 | |
| 578 | flags = fcntl(data->completion_fd, F_GETFL); |
| 579 | if (flags < 0) { |
| 580 | log_err("fio: fcntl(F_GETFL) failed: %s\n", |
| 581 | strerror(errno)); |
| 582 | goto err_blkio_destroy; |
| 583 | } |
| 584 | |
| 585 | if (fcntl(data->completion_fd, F_SETFL, |
| 586 | flags & ~O_NONBLOCK) != 0) { |
| 587 | log_err("fio: fcntl(F_SETFL) failed: %s\n", |
| 588 | strerror(errno)); |
| 589 | goto err_blkio_destroy; |
| 590 | } |
| 591 | } else { |
| 592 | data->completion_fd = -1; |
| 593 | } |
| 594 | |
| 595 | ++proc_state.initted_threads; |
| 596 | if (options->hipri) |
| 597 | ++proc_state.initted_hipri_threads; |
| 598 | |
| 599 | /* Set data last so cleanup() does nothing if init() fails. */ |
| 600 | td->io_ops_data = data; |
| 601 | |
| 602 | fio_blkio_proc_unlock(); |
| 603 | |
| 604 | return 0; |
| 605 | |
| 606 | err_blkio_destroy: |
| 607 | if (proc_state.initted_threads == 0) |
| 608 | blkio_destroy(&proc_state.b); |
| 609 | err_unlock: |
| 610 | if (proc_state.initted_threads == 0) |
| 611 | proc_state.b = NULL; |
| 612 | fio_blkio_proc_unlock(); |
| 613 | err_free: |
| 614 | free(data->completions); |
| 615 | free(data->iovecs); |
| 616 | free(data); |
| 617 | return 1; |
| 618 | } |
| 619 | |
| 620 | static int fio_blkio_post_init(struct thread_data *td) |
| 621 | { |
| 622 | struct fio_blkio_data *data = td->io_ops_data; |
| 623 | |
| 624 | if (!data->has_mem_region) { |
| 625 | /* |
| 626 | * Memory was allocated by the fio core and not iomem_alloc(), |
| 627 | * so we need to register it as a memory region here. |
| 628 | * |
| 629 | * `td->orig_buffer_size` is computed like `len` below, but then |
| 630 | * fio can add some padding to it to make sure it is |
| 631 | * sufficiently aligned to the page size and the mem_align |
| 632 | * option. However, this can make it become unaligned to the |
| 633 | * "mem-region-alignment" property in ways that the user can't |
| 634 | * control, so we essentially recompute `td->orig_buffer_size` |
| 635 | * here but without adding that padding. |
| 636 | */ |
| 637 | |
| 638 | unsigned long long max_block_size; |
| 639 | struct blkio_mem_region region; |
| 640 | |
| 641 | max_block_size = max(td->o.max_bs[DDIR_READ], |
| 642 | max(td->o.max_bs[DDIR_WRITE], |
| 643 | td->o.max_bs[DDIR_TRIM])); |
| 644 | |
| 645 | region = (struct blkio_mem_region) { |
| 646 | .addr = td->orig_buffer, |
| 647 | .len = (size_t)max_block_size * |
| 648 | (size_t)td->o.iodepth, |
| 649 | .fd = -1, |
| 650 | }; |
| 651 | |
| 652 | if (blkio_map_mem_region(proc_state.b, ®ion) != 0) { |
| 653 | fio_blkio_log_err(blkio_map_mem_region); |
| 654 | return 1; |
| 655 | } |
| 656 | } |
| 657 | |
| 658 | return 0; |
| 659 | } |
| 660 | |
| 661 | static void fio_blkio_cleanup(struct thread_data *td) |
| 662 | { |
| 663 | struct fio_blkio_data *data = td->io_ops_data; |
| 664 | |
| 665 | /* |
| 666 | * Subjobs from different jobs can be terminated at different times, so |
| 667 | * this callback may be invoked for one subjob while another is still |
| 668 | * doing I/O. Those subjobs may share the process, so we must wait until |
| 669 | * the last subjob in the process wants to clean up to actually destroy |
| 670 | * the blkio. |
| 671 | */ |
| 672 | |
| 673 | if (data) { |
| 674 | free(data->completions); |
| 675 | free(data->iovecs); |
| 676 | free(data); |
| 677 | |
| 678 | fio_blkio_proc_lock(); |
| 679 | if (--proc_state.initted_threads == 0) { |
| 680 | blkio_destroy(&proc_state.b); |
| 681 | proc_state.b = NULL; |
| 682 | } |
| 683 | fio_blkio_proc_unlock(); |
| 684 | } |
| 685 | } |
| 686 | |
| 687 | #define align_up(x, y) ((((x) + (y) - 1) / (y)) * (y)) |
| 688 | |
| 689 | static int fio_blkio_iomem_alloc(struct thread_data *td, size_t size) |
| 690 | { |
| 691 | struct fio_blkio_data *data = td->io_ops_data; |
| 692 | int ret; |
| 693 | uint64_t mem_region_alignment; |
| 694 | |
| 695 | if (blkio_get_uint64(proc_state.b, "mem-region-alignment", |
| 696 | &mem_region_alignment) != 0) { |
| 697 | fio_blkio_log_err(blkio_get_uint64); |
| 698 | return 1; |
| 699 | } |
| 700 | |
| 701 | /* round up size to satisfy mem-region-alignment */ |
| 702 | size = align_up(size, (size_t)mem_region_alignment); |
| 703 | |
| 704 | fio_blkio_proc_lock(); |
| 705 | |
| 706 | if (blkio_alloc_mem_region(proc_state.b, &data->mem_region, |
| 707 | size) != 0) { |
| 708 | fio_blkio_log_err(blkio_alloc_mem_region); |
| 709 | ret = 1; |
| 710 | goto out; |
| 711 | } |
| 712 | |
| 713 | if (blkio_map_mem_region(proc_state.b, &data->mem_region) != 0) { |
| 714 | fio_blkio_log_err(blkio_map_mem_region); |
| 715 | ret = 1; |
| 716 | goto out_free; |
| 717 | } |
| 718 | |
| 719 | td->orig_buffer = data->mem_region.addr; |
| 720 | data->has_mem_region = true; |
| 721 | |
| 722 | ret = 0; |
| 723 | goto out; |
| 724 | |
| 725 | out_free: |
| 726 | blkio_free_mem_region(proc_state.b, &data->mem_region); |
| 727 | out: |
| 728 | fio_blkio_proc_unlock(); |
| 729 | return ret; |
| 730 | } |
| 731 | |
| 732 | static void fio_blkio_iomem_free(struct thread_data *td) |
| 733 | { |
| 734 | struct fio_blkio_data *data = td->io_ops_data; |
| 735 | |
| 736 | if (data && data->has_mem_region) { |
| 737 | fio_blkio_proc_lock(); |
| 738 | blkio_unmap_mem_region(proc_state.b, &data->mem_region); |
| 739 | blkio_free_mem_region(proc_state.b, &data->mem_region); |
| 740 | fio_blkio_proc_unlock(); |
| 741 | |
| 742 | data->has_mem_region = false; |
| 743 | } |
| 744 | } |
| 745 | |
| 746 | static int fio_blkio_open_file(struct thread_data *td, struct fio_file *f) |
| 747 | { |
| 748 | return 0; |
| 749 | } |
| 750 | |
| 751 | static enum fio_q_status fio_blkio_queue(struct thread_data *td, |
| 752 | struct io_u *io_u) |
| 753 | { |
| 754 | const struct fio_blkio_options *options = td->eo; |
| 755 | struct fio_blkio_data *data = td->io_ops_data; |
| 756 | |
| 757 | fio_ro_check(td, io_u); |
| 758 | |
| 759 | switch (io_u->ddir) { |
| 760 | case DDIR_READ: |
| 761 | if (options->vectored) { |
| 762 | struct iovec *iov = &data->iovecs[io_u->index]; |
| 763 | iov->iov_base = io_u->xfer_buf; |
| 764 | iov->iov_len = (size_t)io_u->xfer_buflen; |
| 765 | |
| 766 | blkioq_readv(data->q, io_u->offset, iov, 1, |
| 767 | io_u, 0); |
| 768 | } else { |
| 769 | blkioq_read(data->q, io_u->offset, |
| 770 | io_u->xfer_buf, |
| 771 | (size_t)io_u->xfer_buflen, io_u, 0); |
| 772 | } |
| 773 | break; |
| 774 | case DDIR_WRITE: |
| 775 | if (options->vectored) { |
| 776 | struct iovec *iov = &data->iovecs[io_u->index]; |
| 777 | iov->iov_base = io_u->xfer_buf; |
| 778 | iov->iov_len = (size_t)io_u->xfer_buflen; |
| 779 | |
| 780 | blkioq_writev(data->q, io_u->offset, iov, 1, |
| 781 | io_u, 0); |
| 782 | } else { |
| 783 | blkioq_write(data->q, io_u->offset, |
| 784 | io_u->xfer_buf, |
| 785 | (size_t)io_u->xfer_buflen, io_u, |
| 786 | 0); |
| 787 | } |
| 788 | break; |
| 789 | case DDIR_TRIM: |
| 790 | if (options->write_zeroes_on_trim) { |
| 791 | blkioq_write_zeroes(data->q, io_u->offset, |
| 792 | io_u->xfer_buflen, io_u, 0); |
| 793 | } else { |
| 794 | blkioq_discard(data->q, io_u->offset, |
| 795 | io_u->xfer_buflen, io_u, 0); |
| 796 | } |
| 797 | break; |
| 798 | case DDIR_SYNC: |
| 799 | case DDIR_DATASYNC: |
| 800 | blkioq_flush(data->q, io_u, 0); |
| 801 | break; |
| 802 | default: |
| 803 | io_u->error = ENOTSUP; |
| 804 | io_u_log_error(td, io_u); |
| 805 | return FIO_Q_COMPLETED; |
| 806 | } |
| 807 | |
| 808 | return FIO_Q_QUEUED; |
| 809 | } |
| 810 | |
| 811 | static int fio_blkio_getevents(struct thread_data *td, unsigned int min, |
| 812 | unsigned int max, const struct timespec *t) |
| 813 | { |
| 814 | const struct fio_blkio_options *options = td->eo; |
| 815 | struct fio_blkio_data *data = td->io_ops_data; |
| 816 | int ret, n; |
| 817 | uint64_t event; |
| 818 | |
| 819 | switch (options->wait_mode) { |
| 820 | case FIO_BLKIO_WAIT_MODE_BLOCK: |
| 821 | n = blkioq_do_io(data->q, data->completions, (int)min, (int)max, |
| 822 | NULL); |
| 823 | if (n < 0) { |
| 824 | fio_blkio_log_err(blkioq_do_io); |
| 825 | return -1; |
| 826 | } |
| 827 | return n; |
| 828 | case FIO_BLKIO_WAIT_MODE_EVENTFD: |
| 829 | n = blkioq_do_io(data->q, data->completions, 0, (int)max, NULL); |
| 830 | if (n < 0) { |
| 831 | fio_blkio_log_err(blkioq_do_io); |
| 832 | return -1; |
| 833 | } |
| 834 | while (n < (int)min) { |
| 835 | ret = read(data->completion_fd, &event, sizeof(event)); |
| 836 | if (ret != sizeof(event)) { |
| 837 | log_err("fio: read() on the completion fd returned %d\n", |
| 838 | ret); |
| 839 | return -1; |
| 840 | } |
| 841 | |
| 842 | ret = blkioq_do_io(data->q, data->completions + n, 0, |
| 843 | (int)max - n, NULL); |
| 844 | if (ret < 0) { |
| 845 | fio_blkio_log_err(blkioq_do_io); |
| 846 | return -1; |
| 847 | } |
| 848 | |
| 849 | n += ret; |
| 850 | } |
| 851 | return n; |
| 852 | case FIO_BLKIO_WAIT_MODE_LOOP: |
| 853 | for (n = 0; n < (int)min; ) { |
| 854 | ret = blkioq_do_io(data->q, data->completions + n, 0, |
| 855 | (int)max - n, NULL); |
| 856 | if (ret < 0) { |
| 857 | fio_blkio_log_err(blkioq_do_io); |
| 858 | return -1; |
| 859 | } |
| 860 | |
| 861 | n += ret; |
| 862 | } |
| 863 | return n; |
| 864 | default: |
| 865 | return -1; |
| 866 | } |
| 867 | } |
| 868 | |
| 869 | static struct io_u *fio_blkio_event(struct thread_data *td, int event) |
| 870 | { |
| 871 | struct fio_blkio_data *data = td->io_ops_data; |
| 872 | struct blkio_completion *completion = &data->completions[event]; |
| 873 | struct io_u *io_u = completion->user_data; |
| 874 | |
| 875 | io_u->error = -completion->ret; |
| 876 | |
| 877 | return io_u; |
| 878 | } |
| 879 | |
| 880 | FIO_STATIC struct ioengine_ops ioengine = { |
| 881 | .name = "libblkio", |
| 882 | .version = FIO_IOOPS_VERSION, |
| 883 | .flags = FIO_DISKLESSIO | FIO_NOEXTEND | |
| 884 | FIO_NO_OFFLOAD | FIO_SKIPPABLE_IOMEM_ALLOC, |
| 885 | |
| 886 | .setup = fio_blkio_setup, |
| 887 | .init = fio_blkio_init, |
| 888 | .post_init = fio_blkio_post_init, |
| 889 | .cleanup = fio_blkio_cleanup, |
| 890 | |
| 891 | .iomem_alloc = fio_blkio_iomem_alloc, |
| 892 | .iomem_free = fio_blkio_iomem_free, |
| 893 | |
| 894 | .open_file = fio_blkio_open_file, |
| 895 | |
| 896 | .queue = fio_blkio_queue, |
| 897 | .getevents = fio_blkio_getevents, |
| 898 | .event = fio_blkio_event, |
| 899 | |
| 900 | .options = options, |
| 901 | .option_struct_size = sizeof(struct fio_blkio_options), |
| 902 | }; |
| 903 | |
| 904 | static void fio_init fio_blkio_register(void) |
| 905 | { |
| 906 | register_ioengine(&ioengine); |
| 907 | } |
| 908 | |
| 909 | static void fio_exit fio_blkio_unregister(void) |
| 910 | { |
| 911 | unregister_ioengine(&ioengine); |
| 912 | } |