1 /* SPDX-License-Identifier: BSD-3-Clause 2 * Copyright (C) 2016 Intel Corporation. 3 * Copyright (c) 2022-2023 NVIDIA CORPORATION & AFFILIATES. 4 * All rights reserved. 5 */ 6 7 #include "spdk/stdinc.h" 8 9 #include "spdk/bdev.h" 10 #include "spdk/accel.h" 11 #include "spdk/endian.h" 12 #include "spdk/env.h" 13 #include "spdk/event.h" 14 #include "spdk/log.h" 15 #include "spdk/util.h" 16 #include "spdk/thread.h" 17 #include "spdk/string.h" 18 #include "spdk/rpc.h" 19 #include "spdk/bit_array.h" 20 #include "spdk/conf.h" 21 #include "spdk/zipf.h" 22 #include "spdk/histogram_data.h" 23 24 #define BDEVPERF_CONFIG_MAX_FILENAME 1024 25 #define BDEVPERF_CONFIG_UNDEFINED -1 26 #define BDEVPERF_CONFIG_ERROR -2 27 #define PATTERN_TYPES_STR "(read, write, randread, randwrite, rw, randrw, verify, reset, unmap, flush, write_zeroes)" 28 29 struct bdevperf_task { 30 struct iovec iov; 31 struct bdevperf_job *job; 32 struct spdk_bdev_io *bdev_io; 33 void *buf; 34 void *verify_buf; 35 void *md_buf; 36 uint64_t offset_blocks; 37 struct bdevperf_task *task_to_abort; 38 enum spdk_bdev_io_type io_type; 39 TAILQ_ENTRY(bdevperf_task) link; 40 struct spdk_bdev_io_wait_entry bdev_io_wait; 41 }; 42 43 static char *g_workload_type = NULL; 44 static int g_io_size = 0; 45 /* initialize to invalid value so we can detect if user overrides it. */ 46 static int g_rw_percentage = -1; 47 static bool g_verify = false; 48 static bool g_reset = false; 49 static bool g_continue_on_failure = false; 50 static bool g_abort = false; 51 static bool g_error_to_exit = false; 52 static int g_queue_depth = 0; 53 static uint64_t g_time_in_usec; 54 static int g_show_performance_real_time = 0; 55 static uint64_t g_show_performance_period_in_usec = SPDK_SEC_TO_USEC; 56 static uint64_t g_show_performance_period_num = 0; 57 static uint64_t g_show_performance_ema_period = 0; 58 static int g_run_rc = 0; 59 static bool g_shutdown = false; 60 static uint64_t g_start_tsc; 61 static uint64_t g_shutdown_tsc; 62 static bool g_zcopy = false; 63 static struct spdk_thread *g_main_thread; 64 static int g_time_in_sec = 0; 65 static bool g_mix_specified = false; 66 static const char *g_job_bdev_name; 67 static bool g_wait_for_tests = false; 68 static struct spdk_jsonrpc_request *g_request = NULL; 69 static bool g_multithread_mode = false; 70 static int g_timeout_in_sec; 71 static struct spdk_conf *g_bdevperf_conf = NULL; 72 static const char *g_bdevperf_conf_file = NULL; 73 static double g_zipf_theta; 74 static bool g_random_map = false; 75 static bool g_unique_writes = false; 76 77 static struct spdk_cpuset g_all_cpuset; 78 static struct spdk_poller *g_perf_timer = NULL; 79 80 static void bdevperf_submit_single(struct bdevperf_job *job, struct bdevperf_task *task); 81 static void rpc_perform_tests_cb(void); 82 static int bdevperf_parse_arg(int ch, char *arg); 83 static int verify_test_params(void); 84 static void bdevperf_usage(void); 85 86 static uint32_t g_bdev_count = 0; 87 static uint32_t g_latency_display_level; 88 89 static bool g_one_thread_per_lcore = false; 90 91 static const double g_latency_cutoffs[] = { 92 0.01, 93 0.10, 94 0.25, 95 0.50, 96 0.75, 97 0.90, 98 0.95, 99 0.98, 100 0.99, 101 0.995, 102 0.999, 103 0.9999, 104 0.99999, 105 0.999999, 106 0.9999999, 107 -1, 108 }; 109 110 static const char *g_rpc_log_file_name = NULL; 111 static FILE *g_rpc_log_file = NULL; 112 113 struct latency_info { 114 uint64_t min; 115 uint64_t max; 116 uint64_t total; 117 }; 118 119 120 enum job_config_rw { 121 JOB_CONFIG_RW_READ = 0, 122 JOB_CONFIG_RW_WRITE, 123 JOB_CONFIG_RW_RANDREAD, 124 JOB_CONFIG_RW_RANDWRITE, 125 JOB_CONFIG_RW_RW, 126 JOB_CONFIG_RW_RANDRW, 127 JOB_CONFIG_RW_VERIFY, 128 JOB_CONFIG_RW_RESET, 129 JOB_CONFIG_RW_UNMAP, 130 JOB_CONFIG_RW_FLUSH, 131 JOB_CONFIG_RW_WRITE_ZEROES, 132 }; 133 134 struct bdevperf_job { 135 char *name; 136 struct spdk_bdev *bdev; 137 struct spdk_bdev_desc *bdev_desc; 138 struct spdk_io_channel *ch; 139 TAILQ_ENTRY(bdevperf_job) link; 140 struct spdk_thread *thread; 141 142 enum job_config_rw workload_type; 143 int io_size; 144 int rw_percentage; 145 bool is_random; 146 bool verify; 147 bool reset; 148 bool continue_on_failure; 149 bool unmap; 150 bool write_zeroes; 151 bool flush; 152 bool abort; 153 int queue_depth; 154 unsigned int seed; 155 156 uint64_t io_completed; 157 uint64_t io_failed; 158 uint64_t io_timeout; 159 uint64_t prev_io_completed; 160 double ema_io_per_second; 161 int current_queue_depth; 162 uint64_t size_in_ios; 163 uint64_t ios_base; 164 uint64_t offset_in_ios; 165 uint64_t io_size_blocks; 166 uint64_t buf_size; 167 uint32_t dif_check_flags; 168 bool is_draining; 169 struct spdk_poller *run_timer; 170 struct spdk_poller *reset_timer; 171 struct spdk_bit_array *outstanding; 172 struct spdk_zipf *zipf; 173 TAILQ_HEAD(, bdevperf_task) task_list; 174 uint64_t run_time_in_usec; 175 176 /* keep channel's histogram data before being destroyed */ 177 struct spdk_histogram_data *histogram; 178 struct spdk_bit_array *random_map; 179 180 /* counter used for generating unique write data (-U option) */ 181 uint32_t write_io_count; 182 }; 183 184 struct spdk_bdevperf { 185 TAILQ_HEAD(, bdevperf_job) jobs; 186 uint32_t running_jobs; 187 }; 188 189 static struct spdk_bdevperf g_bdevperf = { 190 .jobs = TAILQ_HEAD_INITIALIZER(g_bdevperf.jobs), 191 .running_jobs = 0, 192 }; 193 194 /* Storing values from a section of job config file */ 195 struct job_config { 196 const char *name; 197 const char *filename; 198 struct spdk_cpuset cpumask; 199 int bs; 200 int iodepth; 201 int rwmixread; 202 uint32_t lcore; 203 int64_t offset; 204 uint64_t length; 205 enum job_config_rw rw; 206 TAILQ_ENTRY(job_config) link; 207 }; 208 209 TAILQ_HEAD(, job_config) job_config_list 210 = TAILQ_HEAD_INITIALIZER(job_config_list); 211 212 static bool g_performance_dump_active = false; 213 214 struct bdevperf_aggregate_stats { 215 struct bdevperf_job *current_job; 216 uint64_t io_time_in_usec; 217 uint64_t ema_period; 218 double total_io_per_second; 219 double total_mb_per_second; 220 double total_failed_per_second; 221 double total_timeout_per_second; 222 double min_latency; 223 double max_latency; 224 uint64_t total_io_completed; 225 uint64_t total_tsc; 226 }; 227 228 static struct bdevperf_aggregate_stats g_stats = {.min_latency = (double)UINT64_MAX}; 229 230 struct lcore_thread { 231 struct spdk_thread *thread; 232 uint32_t lcore; 233 TAILQ_ENTRY(lcore_thread) link; 234 }; 235 236 TAILQ_HEAD(, lcore_thread) g_lcore_thread_list 237 = TAILQ_HEAD_INITIALIZER(g_lcore_thread_list); 238 239 240 static char * 241 parse_workload_type(enum job_config_rw ret) 242 { 243 switch (ret) { 244 case JOB_CONFIG_RW_READ: 245 return "read"; 246 case JOB_CONFIG_RW_RANDREAD: 247 return "randread"; 248 case JOB_CONFIG_RW_WRITE: 249 return "write"; 250 case JOB_CONFIG_RW_RANDWRITE: 251 return "randwrite"; 252 case JOB_CONFIG_RW_VERIFY: 253 return "verify"; 254 case JOB_CONFIG_RW_RESET: 255 return "reset"; 256 case JOB_CONFIG_RW_UNMAP: 257 return "unmap"; 258 case JOB_CONFIG_RW_WRITE_ZEROES: 259 return "write_zeroes"; 260 case JOB_CONFIG_RW_FLUSH: 261 return "flush"; 262 case JOB_CONFIG_RW_RW: 263 return "rw"; 264 case JOB_CONFIG_RW_RANDRW: 265 return "randrw"; 266 default: 267 fprintf(stderr, "wrong workload_type code\n"); 268 } 269 270 return NULL; 271 } 272 273 /* 274 * Cumulative Moving Average (CMA): average of all data up to current 275 * Exponential Moving Average (EMA): weighted mean of the previous n data and more weight is given to recent 276 * Simple Moving Average (SMA): unweighted mean of the previous n data 277 * 278 * Bdevperf supports CMA and EMA. 279 */ 280 static double 281 get_cma_io_per_second(struct bdevperf_job *job, uint64_t io_time_in_usec) 282 { 283 return (double)job->io_completed * SPDK_SEC_TO_USEC / io_time_in_usec; 284 } 285 286 static double 287 get_ema_io_per_second(struct bdevperf_job *job, uint64_t ema_period) 288 { 289 double io_completed, io_per_second; 290 291 io_completed = job->io_completed; 292 io_per_second = (double)(io_completed - job->prev_io_completed) * SPDK_SEC_TO_USEC 293 / g_show_performance_period_in_usec; 294 job->prev_io_completed = io_completed; 295 296 job->ema_io_per_second += (io_per_second - job->ema_io_per_second) * 2 297 / (ema_period + 1); 298 return job->ema_io_per_second; 299 } 300 301 static void 302 get_avg_latency(void *ctx, uint64_t start, uint64_t end, uint64_t count, 303 uint64_t total, uint64_t so_far) 304 { 305 struct latency_info *latency_info = ctx; 306 307 if (count == 0) { 308 return; 309 } 310 311 latency_info->total += (start + end) / 2 * count; 312 313 if (so_far == count) { 314 latency_info->min = start; 315 } 316 317 if (so_far == total) { 318 latency_info->max = end; 319 } 320 } 321 322 static void 323 performance_dump_job(struct bdevperf_aggregate_stats *stats, struct bdevperf_job *job) 324 { 325 double io_per_second, mb_per_second, failed_per_second, timeout_per_second; 326 double average_latency = 0.0, min_latency, max_latency; 327 uint64_t time_in_usec; 328 uint64_t tsc_rate; 329 uint64_t total_io; 330 struct latency_info latency_info = {}; 331 332 if (job->workload_type == JOB_CONFIG_RW_RW || job->workload_type == JOB_CONFIG_RW_RANDRW) { 333 printf("\r Job: %s (Core Mask 0x%s, workload: %s, percentage: %d, depth: %d, IO size: %d)\n", 334 job->name, spdk_cpuset_fmt(spdk_thread_get_cpumask(job->thread)), 335 parse_workload_type(job->workload_type), job->rw_percentage, 336 job->queue_depth, job->io_size); 337 } else { 338 printf("\r Job: %s (Core Mask 0x%s, workload: %s, depth: %d, IO size: %d)\n", 339 job->name, spdk_cpuset_fmt(spdk_thread_get_cpumask(job->thread)), 340 parse_workload_type(job->workload_type), job->queue_depth, job->io_size); 341 } 342 343 344 if (job->io_failed > 0 && !job->reset && !job->continue_on_failure) { 345 printf("\r Job: %s ended in about %.2f seconds with error\n", 346 job->name, (double)job->run_time_in_usec / SPDK_SEC_TO_USEC); 347 } 348 if (job->verify) { 349 printf("\t Verification LBA range: start 0x%" PRIx64 " length 0x%" PRIx64 "\n", 350 job->ios_base, job->size_in_ios); 351 } 352 353 if (g_performance_dump_active == true) { 354 /* Use job's actual run time as Job has ended */ 355 if (job->io_failed > 0 && !job->continue_on_failure) { 356 time_in_usec = job->run_time_in_usec; 357 } else { 358 time_in_usec = stats->io_time_in_usec; 359 } 360 } else { 361 time_in_usec = job->run_time_in_usec; 362 } 363 364 if (stats->ema_period == 0) { 365 io_per_second = get_cma_io_per_second(job, time_in_usec); 366 } else { 367 io_per_second = get_ema_io_per_second(job, stats->ema_period); 368 } 369 370 tsc_rate = spdk_get_ticks_hz(); 371 mb_per_second = io_per_second * job->io_size / (1024 * 1024); 372 373 spdk_histogram_data_iterate(job->histogram, get_avg_latency, &latency_info); 374 375 total_io = job->io_completed + job->io_failed; 376 if (total_io != 0) { 377 average_latency = (double)latency_info.total / total_io * SPDK_SEC_TO_USEC / tsc_rate; 378 } 379 min_latency = (double)latency_info.min * SPDK_SEC_TO_USEC / tsc_rate; 380 max_latency = (double)latency_info.max * SPDK_SEC_TO_USEC / tsc_rate; 381 382 failed_per_second = (double)job->io_failed * SPDK_SEC_TO_USEC / time_in_usec; 383 timeout_per_second = (double)job->io_timeout * SPDK_SEC_TO_USEC / time_in_usec; 384 385 printf("\t %-20s: %10.2f %10.2f %10.2f", 386 job->name, (float)time_in_usec / SPDK_SEC_TO_USEC, io_per_second, mb_per_second); 387 printf(" %10.2f %8.2f", 388 failed_per_second, timeout_per_second); 389 printf(" %10.2f %10.2f %10.2f\n", 390 average_latency, min_latency, max_latency); 391 392 stats->total_io_per_second += io_per_second; 393 stats->total_mb_per_second += mb_per_second; 394 stats->total_failed_per_second += failed_per_second; 395 stats->total_timeout_per_second += timeout_per_second; 396 stats->total_io_completed += job->io_completed + job->io_failed; 397 stats->total_tsc += latency_info.total; 398 if (min_latency < stats->min_latency) { 399 stats->min_latency = min_latency; 400 } 401 if (max_latency > stats->max_latency) { 402 stats->max_latency = max_latency; 403 } 404 } 405 406 static void 407 generate_data(struct bdevperf_job *job, void *buf, void *md_buf, bool unique) 408 { 409 int offset_blocks = 0, md_offset, data_block_size, inner_offset; 410 int buf_len = job->buf_size; 411 int block_size = spdk_bdev_get_block_size(job->bdev); 412 int md_size = spdk_bdev_get_md_size(job->bdev); 413 int num_blocks = job->io_size_blocks; 414 415 if (buf_len < num_blocks * block_size) { 416 return; 417 } 418 419 if (md_buf == NULL) { 420 data_block_size = block_size - md_size; 421 md_buf = (char *)buf + data_block_size; 422 md_offset = block_size; 423 } else { 424 data_block_size = block_size; 425 md_offset = md_size; 426 } 427 428 if (unique) { 429 uint64_t io_count = job->write_io_count++; 430 unsigned int i; 431 432 assert(md_size == 0 || md_size >= (int)sizeof(uint64_t)); 433 434 while (offset_blocks < num_blocks) { 435 inner_offset = 0; 436 while (inner_offset < data_block_size) { 437 *(uint64_t *)buf = (io_count << 32) | (offset_blocks + inner_offset); 438 inner_offset += sizeof(uint64_t); 439 buf += sizeof(uint64_t); 440 } 441 for (i = 0; i < md_size / sizeof(uint64_t); i++) { 442 ((uint64_t *)md_buf)[i] = (io_count << 32) | offset_blocks; 443 } 444 md_buf += md_offset; 445 offset_blocks++; 446 } 447 return; 448 } 449 450 while (offset_blocks < num_blocks) { 451 inner_offset = 0; 452 while (inner_offset < data_block_size) { 453 *(uint32_t *)buf = offset_blocks + inner_offset; 454 inner_offset += sizeof(uint32_t); 455 buf += sizeof(uint32_t); 456 } 457 memset(md_buf, offset_blocks, md_size); 458 md_buf += md_offset; 459 offset_blocks++; 460 } 461 } 462 463 static bool 464 copy_data(void *wr_buf, int wr_buf_len, void *rd_buf, int rd_buf_len, int block_size, 465 void *wr_md_buf, void *rd_md_buf, int md_size, int num_blocks) 466 { 467 if (wr_buf_len < num_blocks * block_size || rd_buf_len < num_blocks * block_size) { 468 return false; 469 } 470 471 assert((wr_md_buf != NULL) == (rd_md_buf != NULL)); 472 473 memcpy(wr_buf, rd_buf, block_size * num_blocks); 474 475 if (wr_md_buf != NULL) { 476 memcpy(wr_md_buf, rd_md_buf, md_size * num_blocks); 477 } 478 479 return true; 480 } 481 482 static bool 483 verify_data(void *wr_buf, int wr_buf_len, void *rd_buf, int rd_buf_len, int block_size, 484 void *wr_md_buf, void *rd_md_buf, int md_size, int num_blocks, bool md_check) 485 { 486 int offset_blocks = 0, md_offset, data_block_size; 487 488 if (wr_buf_len < num_blocks * block_size || rd_buf_len < num_blocks * block_size) { 489 return false; 490 } 491 492 assert((wr_md_buf != NULL) == (rd_md_buf != NULL)); 493 494 if (wr_md_buf == NULL) { 495 data_block_size = block_size - md_size; 496 wr_md_buf = (char *)wr_buf + data_block_size; 497 rd_md_buf = (char *)rd_buf + data_block_size; 498 md_offset = block_size; 499 } else { 500 data_block_size = block_size; 501 md_offset = md_size; 502 } 503 504 while (offset_blocks < num_blocks) { 505 if (memcmp(wr_buf, rd_buf, data_block_size) != 0) { 506 printf("data_block_size %d, num_blocks %d, offset %d\n", data_block_size, num_blocks, 507 offset_blocks); 508 spdk_log_dump(stdout, "rd_buf", rd_buf, data_block_size); 509 spdk_log_dump(stdout, "wr_buf", wr_buf, data_block_size); 510 return false; 511 } 512 513 wr_buf += block_size; 514 rd_buf += block_size; 515 516 if (md_check) { 517 if (memcmp(wr_md_buf, rd_md_buf, md_size) != 0) { 518 printf("md_size %d, num_blocks %d, offset %d\n", md_size, num_blocks, offset_blocks); 519 spdk_log_dump(stdout, "rd_md_buf", rd_md_buf, md_size); 520 spdk_log_dump(stdout, "wr_md_buf", wr_md_buf, md_size); 521 return false; 522 } 523 524 wr_md_buf += md_offset; 525 rd_md_buf += md_offset; 526 } 527 528 offset_blocks++; 529 } 530 531 return true; 532 } 533 534 static void 535 free_job_config(void) 536 { 537 struct job_config *config, *tmp; 538 539 spdk_conf_free(g_bdevperf_conf); 540 g_bdevperf_conf = NULL; 541 542 TAILQ_FOREACH_SAFE(config, &job_config_list, link, tmp) { 543 TAILQ_REMOVE(&job_config_list, config, link); 544 free(config); 545 } 546 } 547 548 static void 549 bdevperf_job_free(struct bdevperf_job *job) 550 { 551 spdk_histogram_data_free(job->histogram); 552 spdk_bit_array_free(&job->outstanding); 553 spdk_bit_array_free(&job->random_map); 554 spdk_zipf_free(&job->zipf); 555 free(job->name); 556 free(job); 557 } 558 559 static void 560 job_thread_exit(void *ctx) 561 { 562 spdk_thread_exit(spdk_get_thread()); 563 } 564 565 static void 566 check_cutoff(void *ctx, uint64_t start, uint64_t end, uint64_t count, 567 uint64_t total, uint64_t so_far) 568 { 569 double so_far_pct; 570 double **cutoff = ctx; 571 uint64_t tsc_rate; 572 573 if (count == 0) { 574 return; 575 } 576 577 tsc_rate = spdk_get_ticks_hz(); 578 so_far_pct = (double)so_far / total; 579 while (so_far_pct >= **cutoff && **cutoff > 0) { 580 printf("%9.5f%% : %9.3fus\n", **cutoff * 100, (double)end * SPDK_SEC_TO_USEC / tsc_rate); 581 (*cutoff)++; 582 } 583 } 584 585 static void 586 print_bucket(void *ctx, uint64_t start, uint64_t end, uint64_t count, 587 uint64_t total, uint64_t so_far) 588 { 589 double so_far_pct; 590 uint64_t tsc_rate; 591 592 if (count == 0) { 593 return; 594 } 595 596 tsc_rate = spdk_get_ticks_hz(); 597 so_far_pct = (double)so_far * 100 / total; 598 printf("%9.3f - %9.3f: %9.4f%% (%9ju)\n", 599 (double)start * SPDK_SEC_TO_USEC / tsc_rate, 600 (double)end * SPDK_SEC_TO_USEC / tsc_rate, 601 so_far_pct, count); 602 } 603 604 static void 605 bdevperf_test_done(void *ctx) 606 { 607 struct bdevperf_job *job, *jtmp; 608 struct bdevperf_task *task, *ttmp; 609 struct lcore_thread *lthread, *lttmp; 610 double average_latency = 0.0; 611 uint64_t time_in_usec; 612 int rc; 613 614 if (g_time_in_usec) { 615 g_stats.io_time_in_usec = g_time_in_usec; 616 617 if (!g_run_rc && g_performance_dump_active) { 618 spdk_thread_send_msg(spdk_get_thread(), bdevperf_test_done, NULL); 619 return; 620 } 621 } 622 623 if (g_show_performance_real_time) { 624 spdk_poller_unregister(&g_perf_timer); 625 } 626 627 if (g_shutdown) { 628 g_shutdown_tsc = spdk_get_ticks() - g_start_tsc; 629 time_in_usec = g_shutdown_tsc * SPDK_SEC_TO_USEC / spdk_get_ticks_hz(); 630 g_time_in_usec = (g_time_in_usec > time_in_usec) ? time_in_usec : g_time_in_usec; 631 printf("Received shutdown signal, test time was about %.6f seconds\n", 632 (double)g_time_in_usec / SPDK_SEC_TO_USEC); 633 } 634 635 printf("\n%*s\n", 107, "Latency(us)"); 636 printf("\r %-*s: %10s %10s %10s %10s %8s %10s %10s %10s\n", 637 28, "Device Information", "runtime(s)", "IOPS", "MiB/s", "Fail/s", "TO/s", "Average", "min", "max"); 638 639 TAILQ_FOREACH_SAFE(job, &g_bdevperf.jobs, link, jtmp) { 640 performance_dump_job(&g_stats, job); 641 } 642 643 printf("\r ==================================================================================" 644 "=================================\n"); 645 printf("\r %-28s: %10s %10.2f %10.2f", 646 "Total", "", g_stats.total_io_per_second, g_stats.total_mb_per_second); 647 printf(" %10.2f %8.2f", 648 g_stats.total_failed_per_second, g_stats.total_timeout_per_second); 649 650 if (g_stats.total_io_completed != 0) { 651 average_latency = ((double)g_stats.total_tsc / g_stats.total_io_completed) * SPDK_SEC_TO_USEC / 652 spdk_get_ticks_hz(); 653 } 654 printf(" %10.2f %10.2f %10.2f\n", average_latency, g_stats.min_latency, g_stats.max_latency); 655 656 if (g_latency_display_level == 0 || g_stats.total_io_completed == 0) { 657 goto clean; 658 } 659 660 printf("\n Latency summary\n"); 661 TAILQ_FOREACH_SAFE(job, &g_bdevperf.jobs, link, jtmp) { 662 printf("\r =============================================\n"); 663 printf("\r Job: %s (Core Mask 0x%s)\n", job->name, 664 spdk_cpuset_fmt(spdk_thread_get_cpumask(job->thread))); 665 666 const double *cutoff = g_latency_cutoffs; 667 668 spdk_histogram_data_iterate(job->histogram, check_cutoff, &cutoff); 669 670 printf("\n"); 671 } 672 673 if (g_latency_display_level == 1) { 674 goto clean; 675 } 676 677 printf("\r Latency histogram\n"); 678 TAILQ_FOREACH_SAFE(job, &g_bdevperf.jobs, link, jtmp) { 679 printf("\r =============================================\n"); 680 printf("\r Job: %s (Core Mask 0x%s)\n", job->name, 681 spdk_cpuset_fmt(spdk_thread_get_cpumask(job->thread))); 682 683 spdk_histogram_data_iterate(job->histogram, print_bucket, NULL); 684 printf("\n"); 685 } 686 687 clean: 688 fflush(stdout); 689 690 TAILQ_FOREACH_SAFE(job, &g_bdevperf.jobs, link, jtmp) { 691 TAILQ_REMOVE(&g_bdevperf.jobs, job, link); 692 693 if (!g_one_thread_per_lcore) { 694 spdk_thread_send_msg(job->thread, job_thread_exit, NULL); 695 } 696 697 TAILQ_FOREACH_SAFE(task, &job->task_list, link, ttmp) { 698 TAILQ_REMOVE(&job->task_list, task, link); 699 spdk_free(task->buf); 700 spdk_free(task->verify_buf); 701 spdk_free(task->md_buf); 702 free(task); 703 } 704 705 bdevperf_job_free(job); 706 } 707 708 if (g_one_thread_per_lcore) { 709 TAILQ_FOREACH_SAFE(lthread, &g_lcore_thread_list, link, lttmp) { 710 TAILQ_REMOVE(&g_lcore_thread_list, lthread, link); 711 spdk_thread_send_msg(lthread->thread, job_thread_exit, NULL); 712 free(lthread); 713 } 714 } 715 716 if (g_bdevperf_conf == NULL) { 717 free_job_config(); 718 } 719 720 rc = g_run_rc; 721 if (g_request && !g_shutdown) { 722 rpc_perform_tests_cb(); 723 if (rc != 0) { 724 spdk_app_stop(rc); 725 } 726 } else { 727 spdk_app_stop(rc); 728 } 729 } 730 731 static void 732 bdevperf_job_end(void *ctx) 733 { 734 assert(g_main_thread == spdk_get_thread()); 735 736 if (--g_bdevperf.running_jobs == 0) { 737 bdevperf_test_done(NULL); 738 } 739 } 740 741 static void 742 bdevperf_channel_get_histogram_cb(void *cb_arg, int status, struct spdk_histogram_data *histogram) 743 { 744 struct spdk_histogram_data *job_hist = cb_arg; 745 746 if (status == 0) { 747 spdk_histogram_data_merge(job_hist, histogram); 748 } 749 } 750 751 static void 752 bdevperf_job_empty(struct bdevperf_job *job) 753 { 754 uint64_t end_tsc = 0; 755 756 end_tsc = spdk_get_ticks() - g_start_tsc; 757 job->run_time_in_usec = end_tsc * SPDK_SEC_TO_USEC / spdk_get_ticks_hz(); 758 /* keep histogram info before channel is destroyed */ 759 spdk_bdev_channel_get_histogram(job->ch, bdevperf_channel_get_histogram_cb, 760 job->histogram); 761 spdk_put_io_channel(job->ch); 762 spdk_bdev_close(job->bdev_desc); 763 spdk_thread_send_msg(g_main_thread, bdevperf_job_end, NULL); 764 } 765 766 static void 767 bdevperf_end_task(struct bdevperf_task *task) 768 { 769 struct bdevperf_job *job = task->job; 770 771 TAILQ_INSERT_TAIL(&job->task_list, task, link); 772 if (job->is_draining) { 773 if (job->current_queue_depth == 0) { 774 bdevperf_job_empty(job); 775 } 776 } 777 } 778 779 static void 780 bdevperf_queue_io_wait_with_cb(struct bdevperf_task *task, spdk_bdev_io_wait_cb cb_fn) 781 { 782 struct bdevperf_job *job = task->job; 783 784 task->bdev_io_wait.bdev = job->bdev; 785 task->bdev_io_wait.cb_fn = cb_fn; 786 task->bdev_io_wait.cb_arg = task; 787 spdk_bdev_queue_io_wait(job->bdev, job->ch, &task->bdev_io_wait); 788 } 789 790 static int 791 bdevperf_job_drain(void *ctx) 792 { 793 struct bdevperf_job *job = ctx; 794 795 spdk_poller_unregister(&job->run_timer); 796 if (job->reset) { 797 spdk_poller_unregister(&job->reset_timer); 798 } 799 800 job->is_draining = true; 801 802 return -1; 803 } 804 805 static int 806 bdevperf_job_drain_timer(void *ctx) 807 { 808 struct bdevperf_job *job = ctx; 809 810 bdevperf_job_drain(ctx); 811 if (job->current_queue_depth == 0) { 812 bdevperf_job_empty(job); 813 } 814 815 return SPDK_POLLER_BUSY; 816 } 817 818 static void 819 bdevperf_abort_complete(struct spdk_bdev_io *bdev_io, bool success, void *cb_arg) 820 { 821 struct bdevperf_task *task = cb_arg; 822 struct bdevperf_job *job = task->job; 823 824 job->current_queue_depth--; 825 826 if (success) { 827 job->io_completed++; 828 } else { 829 job->io_failed++; 830 if (!job->continue_on_failure) { 831 bdevperf_job_drain(job); 832 g_run_rc = -1; 833 } 834 } 835 836 spdk_bdev_free_io(bdev_io); 837 bdevperf_end_task(task); 838 } 839 840 static int 841 bdevperf_verify_dif(struct bdevperf_task *task) 842 { 843 struct bdevperf_job *job = task->job; 844 struct spdk_bdev *bdev = job->bdev; 845 struct spdk_dif_ctx dif_ctx; 846 struct spdk_dif_error err_blk = {}; 847 int rc; 848 struct spdk_dif_ctx_init_ext_opts dif_opts; 849 850 dif_opts.size = SPDK_SIZEOF(&dif_opts, dif_pi_format); 851 dif_opts.dif_pi_format = spdk_bdev_get_dif_pi_format(bdev); 852 rc = spdk_dif_ctx_init(&dif_ctx, 853 spdk_bdev_get_block_size(bdev), 854 spdk_bdev_get_md_size(bdev), 855 spdk_bdev_is_md_interleaved(bdev), 856 spdk_bdev_is_dif_head_of_md(bdev), 857 spdk_bdev_get_dif_type(bdev), 858 job->dif_check_flags, 859 task->offset_blocks, 0, 0, 0, 0, &dif_opts); 860 if (rc != 0) { 861 fprintf(stderr, "Initialization of DIF context failed\n"); 862 return rc; 863 } 864 865 if (spdk_bdev_is_md_interleaved(bdev)) { 866 rc = spdk_dif_verify(&task->iov, 1, job->io_size_blocks, &dif_ctx, &err_blk); 867 } else { 868 struct iovec md_iov = { 869 .iov_base = task->md_buf, 870 .iov_len = spdk_bdev_get_md_size(bdev) * job->io_size_blocks, 871 }; 872 873 rc = spdk_dix_verify(&task->iov, 1, &md_iov, job->io_size_blocks, &dif_ctx, &err_blk); 874 } 875 876 if (rc != 0) { 877 fprintf(stderr, "DIF/DIX error detected. type=%d, offset=%" PRIu32 "\n", 878 err_blk.err_type, err_blk.err_offset); 879 } 880 881 return rc; 882 } 883 884 static void 885 bdevperf_complete(struct spdk_bdev_io *bdev_io, bool success, void *cb_arg) 886 { 887 struct bdevperf_job *job; 888 struct bdevperf_task *task = cb_arg; 889 bool md_check; 890 uint64_t offset_in_ios; 891 int rc; 892 893 job = task->job; 894 md_check = spdk_bdev_get_dif_type(job->bdev) == SPDK_DIF_DISABLE; 895 896 if (g_error_to_exit == true) { 897 bdevperf_job_drain(job); 898 } else if (!success) { 899 if (!job->reset && !job->continue_on_failure) { 900 bdevperf_job_drain(job); 901 g_run_rc = -1; 902 g_error_to_exit = true; 903 printf("task offset: %" PRIu64 " on job bdev=%s fails\n", 904 task->offset_blocks, job->name); 905 } 906 } else if (job->verify || job->reset) { 907 if (!verify_data(task->buf, job->buf_size, 908 task->iov.iov_base, job->buf_size, 909 spdk_bdev_get_block_size(job->bdev), 910 task->md_buf, spdk_bdev_io_get_md_buf(bdev_io), 911 spdk_bdev_get_md_size(job->bdev), 912 job->io_size_blocks, md_check)) { 913 printf("Buffer mismatch! Target: %s Disk Offset: %" PRIu64 "\n", job->name, task->offset_blocks); 914 bdevperf_job_drain(job); 915 g_run_rc = -1; 916 } 917 } else if (job->dif_check_flags != 0) { 918 if (task->io_type == SPDK_BDEV_IO_TYPE_READ && spdk_bdev_get_md_size(job->bdev) != 0) { 919 rc = bdevperf_verify_dif(task); 920 if (rc != 0) { 921 printf("DIF error detected. task offset: %" PRIu64 " on job bdev=%s\n", 922 task->offset_blocks, job->name); 923 924 success = false; 925 if (!job->reset && !job->continue_on_failure) { 926 bdevperf_job_drain(job); 927 g_run_rc = -1; 928 g_error_to_exit = true; 929 } 930 } 931 } 932 } 933 934 job->current_queue_depth--; 935 936 if (success) { 937 job->io_completed++; 938 } else { 939 job->io_failed++; 940 } 941 942 if (job->verify) { 943 assert(task->offset_blocks / job->io_size_blocks >= job->ios_base); 944 offset_in_ios = task->offset_blocks / job->io_size_blocks - job->ios_base; 945 946 assert(spdk_bit_array_get(job->outstanding, offset_in_ios) == true); 947 spdk_bit_array_clear(job->outstanding, offset_in_ios); 948 } 949 950 spdk_bdev_free_io(bdev_io); 951 952 /* 953 * is_draining indicates when time has expired for the test run 954 * and we are just waiting for the previously submitted I/O 955 * to complete. In this case, do not submit a new I/O to replace 956 * the one just completed. 957 */ 958 if (!job->is_draining) { 959 bdevperf_submit_single(job, task); 960 } else { 961 bdevperf_end_task(task); 962 } 963 } 964 965 static void 966 bdevperf_verify_submit_read(void *cb_arg) 967 { 968 struct bdevperf_job *job; 969 struct bdevperf_task *task = cb_arg; 970 int rc; 971 972 job = task->job; 973 974 task->iov.iov_base = task->verify_buf; 975 task->iov.iov_len = job->buf_size; 976 977 /* Read the data back in */ 978 rc = spdk_bdev_readv_blocks_with_md(job->bdev_desc, job->ch, &task->iov, 1, NULL, 979 task->offset_blocks, job->io_size_blocks, 980 bdevperf_complete, task); 981 982 if (rc == -ENOMEM) { 983 bdevperf_queue_io_wait_with_cb(task, bdevperf_verify_submit_read); 984 } else if (rc != 0) { 985 printf("Failed to submit read: %d\n", rc); 986 bdevperf_job_drain(job); 987 g_run_rc = rc; 988 } 989 } 990 991 static void 992 bdevperf_verify_write_complete(struct spdk_bdev_io *bdev_io, bool success, 993 void *cb_arg) 994 { 995 if (success) { 996 spdk_bdev_free_io(bdev_io); 997 bdevperf_verify_submit_read(cb_arg); 998 } else { 999 bdevperf_complete(bdev_io, success, cb_arg); 1000 } 1001 } 1002 1003 static void 1004 bdevperf_zcopy_populate_complete(struct spdk_bdev_io *bdev_io, bool success, void *cb_arg) 1005 { 1006 if (!success) { 1007 bdevperf_complete(bdev_io, success, cb_arg); 1008 return; 1009 } 1010 1011 spdk_bdev_zcopy_end(bdev_io, false, bdevperf_complete, cb_arg); 1012 } 1013 1014 static int 1015 bdevperf_generate_dif(struct bdevperf_task *task) 1016 { 1017 struct bdevperf_job *job = task->job; 1018 struct spdk_bdev *bdev = job->bdev; 1019 struct spdk_dif_ctx dif_ctx; 1020 int rc; 1021 struct spdk_dif_ctx_init_ext_opts dif_opts; 1022 1023 dif_opts.size = SPDK_SIZEOF(&dif_opts, dif_pi_format); 1024 dif_opts.dif_pi_format = spdk_bdev_get_dif_pi_format(bdev); 1025 rc = spdk_dif_ctx_init(&dif_ctx, 1026 spdk_bdev_get_block_size(bdev), 1027 spdk_bdev_get_md_size(bdev), 1028 spdk_bdev_is_md_interleaved(bdev), 1029 spdk_bdev_is_dif_head_of_md(bdev), 1030 spdk_bdev_get_dif_type(bdev), 1031 job->dif_check_flags, 1032 task->offset_blocks, 0, 0, 0, 0, &dif_opts); 1033 if (rc != 0) { 1034 fprintf(stderr, "Initialization of DIF context failed\n"); 1035 return rc; 1036 } 1037 1038 if (spdk_bdev_is_md_interleaved(bdev)) { 1039 rc = spdk_dif_generate(&task->iov, 1, job->io_size_blocks, &dif_ctx); 1040 } else { 1041 struct iovec md_iov = { 1042 .iov_base = task->md_buf, 1043 .iov_len = spdk_bdev_get_md_size(bdev) * job->io_size_blocks, 1044 }; 1045 1046 rc = spdk_dix_generate(&task->iov, 1, &md_iov, job->io_size_blocks, &dif_ctx); 1047 } 1048 1049 if (rc != 0) { 1050 fprintf(stderr, "Generation of DIF/DIX failed\n"); 1051 } 1052 1053 return rc; 1054 } 1055 1056 static void 1057 bdevperf_submit_task(void *arg) 1058 { 1059 struct bdevperf_task *task = arg; 1060 struct bdevperf_job *job = task->job; 1061 struct spdk_bdev_desc *desc; 1062 struct spdk_io_channel *ch; 1063 spdk_bdev_io_completion_cb cb_fn; 1064 uint64_t offset_in_ios; 1065 int rc = 0; 1066 1067 desc = job->bdev_desc; 1068 ch = job->ch; 1069 1070 switch (task->io_type) { 1071 case SPDK_BDEV_IO_TYPE_WRITE: 1072 if (spdk_bdev_get_md_size(job->bdev) != 0 && job->dif_check_flags != 0) { 1073 rc = bdevperf_generate_dif(task); 1074 } 1075 if (rc == 0) { 1076 cb_fn = (job->verify || job->reset) ? bdevperf_verify_write_complete : bdevperf_complete; 1077 1078 if (g_zcopy) { 1079 spdk_bdev_zcopy_end(task->bdev_io, true, cb_fn, task); 1080 return; 1081 } else { 1082 rc = spdk_bdev_writev_blocks_with_md(desc, ch, &task->iov, 1, 1083 task->md_buf, 1084 task->offset_blocks, 1085 job->io_size_blocks, 1086 cb_fn, task); 1087 } 1088 } 1089 break; 1090 case SPDK_BDEV_IO_TYPE_FLUSH: 1091 rc = spdk_bdev_flush_blocks(desc, ch, task->offset_blocks, 1092 job->io_size_blocks, bdevperf_complete, task); 1093 break; 1094 case SPDK_BDEV_IO_TYPE_UNMAP: 1095 rc = spdk_bdev_unmap_blocks(desc, ch, task->offset_blocks, 1096 job->io_size_blocks, bdevperf_complete, task); 1097 break; 1098 case SPDK_BDEV_IO_TYPE_WRITE_ZEROES: 1099 rc = spdk_bdev_write_zeroes_blocks(desc, ch, task->offset_blocks, 1100 job->io_size_blocks, bdevperf_complete, task); 1101 break; 1102 case SPDK_BDEV_IO_TYPE_READ: 1103 if (g_zcopy) { 1104 rc = spdk_bdev_zcopy_start(desc, ch, NULL, 0, task->offset_blocks, job->io_size_blocks, 1105 true, bdevperf_zcopy_populate_complete, task); 1106 } else { 1107 rc = spdk_bdev_readv_blocks_with_md(desc, ch, &task->iov, 1, 1108 task->md_buf, 1109 task->offset_blocks, 1110 job->io_size_blocks, 1111 bdevperf_complete, task); 1112 } 1113 break; 1114 case SPDK_BDEV_IO_TYPE_ABORT: 1115 rc = spdk_bdev_abort(desc, ch, task->task_to_abort, bdevperf_abort_complete, task); 1116 break; 1117 default: 1118 assert(false); 1119 rc = -EINVAL; 1120 break; 1121 } 1122 1123 if (rc == -ENOMEM) { 1124 bdevperf_queue_io_wait_with_cb(task, bdevperf_submit_task); 1125 return; 1126 } else if (rc != 0) { 1127 printf("Failed to submit bdev_io: %d\n", rc); 1128 if (job->verify) { 1129 assert(task->offset_blocks / job->io_size_blocks >= job->ios_base); 1130 offset_in_ios = task->offset_blocks / job->io_size_blocks - job->ios_base; 1131 1132 assert(spdk_bit_array_get(job->outstanding, offset_in_ios) == true); 1133 spdk_bit_array_clear(job->outstanding, offset_in_ios); 1134 } 1135 bdevperf_job_drain(job); 1136 g_run_rc = rc; 1137 return; 1138 } 1139 1140 job->current_queue_depth++; 1141 } 1142 1143 static void 1144 bdevperf_zcopy_get_buf_complete(struct spdk_bdev_io *bdev_io, bool success, void *cb_arg) 1145 { 1146 struct bdevperf_task *task = cb_arg; 1147 struct bdevperf_job *job = task->job; 1148 struct iovec *iovs; 1149 int iovcnt; 1150 1151 if (!success) { 1152 bdevperf_job_drain(job); 1153 g_run_rc = -1; 1154 return; 1155 } 1156 1157 task->bdev_io = bdev_io; 1158 task->io_type = SPDK_BDEV_IO_TYPE_WRITE; 1159 1160 if (job->verify || job->reset) { 1161 /* When job->verify or job->reset is enabled, task->buf is used for 1162 * verification of read after write. For write I/O, when zcopy APIs 1163 * are used, task->buf cannot be used, and data must be written to 1164 * the data buffer allocated underneath bdev layer instead. 1165 * Hence we copy task->buf to the allocated data buffer here. 1166 */ 1167 spdk_bdev_io_get_iovec(bdev_io, &iovs, &iovcnt); 1168 assert(iovcnt == 1); 1169 assert(iovs != NULL); 1170 1171 copy_data(iovs[0].iov_base, iovs[0].iov_len, task->buf, job->buf_size, 1172 spdk_bdev_get_block_size(job->bdev), 1173 spdk_bdev_io_get_md_buf(bdev_io), task->md_buf, 1174 spdk_bdev_get_md_size(job->bdev), job->io_size_blocks); 1175 } 1176 1177 bdevperf_submit_task(task); 1178 } 1179 1180 static void 1181 bdevperf_prep_zcopy_write_task(void *arg) 1182 { 1183 struct bdevperf_task *task = arg; 1184 struct bdevperf_job *job = task->job; 1185 int rc; 1186 1187 rc = spdk_bdev_zcopy_start(job->bdev_desc, job->ch, NULL, 0, 1188 task->offset_blocks, job->io_size_blocks, 1189 false, bdevperf_zcopy_get_buf_complete, task); 1190 if (rc != 0) { 1191 assert(rc == -ENOMEM); 1192 bdevperf_queue_io_wait_with_cb(task, bdevperf_prep_zcopy_write_task); 1193 return; 1194 } 1195 1196 job->current_queue_depth++; 1197 } 1198 1199 static struct bdevperf_task * 1200 bdevperf_job_get_task(struct bdevperf_job *job) 1201 { 1202 struct bdevperf_task *task; 1203 1204 task = TAILQ_FIRST(&job->task_list); 1205 if (!task) { 1206 printf("Task allocation failed\n"); 1207 abort(); 1208 } 1209 1210 TAILQ_REMOVE(&job->task_list, task, link); 1211 return task; 1212 } 1213 1214 static void 1215 bdevperf_submit_single(struct bdevperf_job *job, struct bdevperf_task *task) 1216 { 1217 uint64_t offset_in_ios; 1218 uint64_t rand_value; 1219 uint32_t first_clear; 1220 1221 if (job->zipf) { 1222 offset_in_ios = spdk_zipf_generate(job->zipf); 1223 } else if (job->is_random) { 1224 /* RAND_MAX is only INT32_MAX, so use 2 calls to rand_r to 1225 * get a large enough value to ensure we are issuing I/O 1226 * uniformly across the whole bdev. 1227 */ 1228 rand_value = (uint64_t)rand_r(&job->seed) * RAND_MAX + rand_r(&job->seed); 1229 offset_in_ios = rand_value % job->size_in_ios; 1230 1231 if (g_random_map) { 1232 /* Make sure, that the offset does not exceed the maximum size 1233 * of the bit array (verified during job creation) 1234 */ 1235 assert(offset_in_ios < UINT32_MAX); 1236 1237 first_clear = spdk_bit_array_find_first_clear(job->random_map, (uint32_t)offset_in_ios); 1238 1239 if (first_clear == UINT32_MAX) { 1240 first_clear = spdk_bit_array_find_first_clear(job->random_map, 0); 1241 1242 if (first_clear == UINT32_MAX) { 1243 /* If there are no more clear bits in the array, we start over 1244 * and select the previously selected random value. 1245 */ 1246 spdk_bit_array_clear_mask(job->random_map); 1247 first_clear = (uint32_t)offset_in_ios; 1248 } 1249 } 1250 1251 spdk_bit_array_set(job->random_map, first_clear); 1252 1253 offset_in_ios = first_clear; 1254 } 1255 } else { 1256 offset_in_ios = job->offset_in_ios++; 1257 if (job->offset_in_ios == job->size_in_ios) { 1258 job->offset_in_ios = 0; 1259 } 1260 1261 /* Increment of offset_in_ios if there's already an outstanding IO 1262 * to that location. We only need this with job->verify as random 1263 * offsets are not supported with job->verify at this time. 1264 */ 1265 if (job->verify) { 1266 assert(spdk_bit_array_find_first_clear(job->outstanding, 0) != UINT32_MAX); 1267 1268 while (spdk_bit_array_get(job->outstanding, offset_in_ios)) { 1269 offset_in_ios = job->offset_in_ios++; 1270 if (job->offset_in_ios == job->size_in_ios) { 1271 job->offset_in_ios = 0; 1272 } 1273 } 1274 spdk_bit_array_set(job->outstanding, offset_in_ios); 1275 } 1276 } 1277 1278 /* For multi-thread to same job, offset_in_ios is relative 1279 * to the LBA range assigned for that job. job->offset_blocks 1280 * is absolute (entire bdev LBA range). 1281 */ 1282 task->offset_blocks = (offset_in_ios + job->ios_base) * job->io_size_blocks; 1283 1284 if (job->flush) { 1285 task->io_type = SPDK_BDEV_IO_TYPE_FLUSH; 1286 } else if (job->unmap) { 1287 task->io_type = SPDK_BDEV_IO_TYPE_UNMAP; 1288 } else if (job->write_zeroes) { 1289 task->io_type = SPDK_BDEV_IO_TYPE_WRITE_ZEROES; 1290 } else if ((job->rw_percentage == 100) || 1291 (job->rw_percentage != 0 && ((rand_r(&job->seed) % 100) < job->rw_percentage))) { 1292 assert(!job->verify); 1293 task->io_type = SPDK_BDEV_IO_TYPE_READ; 1294 if (!g_zcopy) { 1295 task->iov.iov_base = task->buf; 1296 task->iov.iov_len = job->buf_size; 1297 } 1298 } else { 1299 if (job->verify || job->reset || g_unique_writes) { 1300 generate_data(job, task->buf, task->md_buf, g_unique_writes); 1301 } 1302 if (g_zcopy) { 1303 bdevperf_prep_zcopy_write_task(task); 1304 return; 1305 } else { 1306 task->iov.iov_base = task->buf; 1307 task->iov.iov_len = job->buf_size; 1308 task->io_type = SPDK_BDEV_IO_TYPE_WRITE; 1309 } 1310 } 1311 1312 bdevperf_submit_task(task); 1313 } 1314 1315 static int reset_job(void *arg); 1316 1317 static void 1318 reset_cb(struct spdk_bdev_io *bdev_io, bool success, void *cb_arg) 1319 { 1320 struct bdevperf_task *task = cb_arg; 1321 struct bdevperf_job *job = task->job; 1322 1323 if (!success) { 1324 printf("Reset blockdev=%s failed\n", spdk_bdev_get_name(job->bdev)); 1325 bdevperf_job_drain(job); 1326 g_run_rc = -1; 1327 } 1328 1329 TAILQ_INSERT_TAIL(&job->task_list, task, link); 1330 spdk_bdev_free_io(bdev_io); 1331 1332 job->reset_timer = SPDK_POLLER_REGISTER(reset_job, job, 1333 10 * SPDK_SEC_TO_USEC); 1334 } 1335 1336 static int 1337 reset_job(void *arg) 1338 { 1339 struct bdevperf_job *job = arg; 1340 struct bdevperf_task *task; 1341 int rc; 1342 1343 spdk_poller_unregister(&job->reset_timer); 1344 1345 /* Do reset. */ 1346 task = bdevperf_job_get_task(job); 1347 rc = spdk_bdev_reset(job->bdev_desc, job->ch, 1348 reset_cb, task); 1349 if (rc) { 1350 printf("Reset failed: %d\n", rc); 1351 bdevperf_job_drain(job); 1352 g_run_rc = -1; 1353 } 1354 1355 return -1; 1356 } 1357 1358 static void 1359 bdevperf_timeout_cb(void *cb_arg, struct spdk_bdev_io *bdev_io) 1360 { 1361 struct bdevperf_job *job = cb_arg; 1362 struct bdevperf_task *task; 1363 1364 job->io_timeout++; 1365 1366 if (job->is_draining || !job->abort || 1367 !spdk_bdev_io_type_supported(job->bdev, SPDK_BDEV_IO_TYPE_ABORT)) { 1368 return; 1369 } 1370 1371 task = bdevperf_job_get_task(job); 1372 if (task == NULL) { 1373 return; 1374 } 1375 1376 task->task_to_abort = spdk_bdev_io_get_cb_arg(bdev_io); 1377 task->io_type = SPDK_BDEV_IO_TYPE_ABORT; 1378 1379 bdevperf_submit_task(task); 1380 } 1381 1382 static void 1383 bdevperf_job_run(void *ctx) 1384 { 1385 struct bdevperf_job *job = ctx; 1386 struct bdevperf_task *task; 1387 int i; 1388 1389 /* Submit initial I/O for this job. Each time one 1390 * completes, another will be submitted. */ 1391 1392 /* Start a timer to stop this I/O chain when the run is over */ 1393 job->run_timer = SPDK_POLLER_REGISTER(bdevperf_job_drain_timer, job, g_time_in_usec); 1394 if (job->reset) { 1395 job->reset_timer = SPDK_POLLER_REGISTER(reset_job, job, 1396 10 * SPDK_SEC_TO_USEC); 1397 } 1398 1399 spdk_bdev_set_timeout(job->bdev_desc, g_timeout_in_sec, bdevperf_timeout_cb, job); 1400 1401 for (i = 0; i < job->queue_depth; i++) { 1402 task = bdevperf_job_get_task(job); 1403 bdevperf_submit_single(job, task); 1404 } 1405 } 1406 1407 static void 1408 _performance_dump_done(void *ctx) 1409 { 1410 struct bdevperf_aggregate_stats *stats = ctx; 1411 double average_latency; 1412 1413 printf("\r ==================================================================================" 1414 "=================================\n"); 1415 printf("\r %-28s: %10s %10.2f %10.2f", 1416 "Total", "", stats->total_io_per_second, stats->total_mb_per_second); 1417 printf(" %10.2f %8.2f", 1418 stats->total_failed_per_second, stats->total_timeout_per_second); 1419 1420 average_latency = ((double)stats->total_tsc / stats->total_io_completed) * SPDK_SEC_TO_USEC / 1421 spdk_get_ticks_hz(); 1422 printf(" %10.2f %10.2f %10.2f\n", average_latency, stats->min_latency, stats->max_latency); 1423 printf("\n"); 1424 1425 fflush(stdout); 1426 1427 g_performance_dump_active = false; 1428 1429 free(stats); 1430 } 1431 1432 static void 1433 _performance_dump(void *ctx) 1434 { 1435 struct bdevperf_aggregate_stats *stats = ctx; 1436 1437 performance_dump_job(stats, stats->current_job); 1438 1439 /* This assumes the jobs list is static after start up time. 1440 * That's true right now, but if that ever changed this would need a lock. */ 1441 stats->current_job = TAILQ_NEXT(stats->current_job, link); 1442 if (stats->current_job == NULL) { 1443 spdk_thread_send_msg(g_main_thread, _performance_dump_done, stats); 1444 } else { 1445 spdk_thread_send_msg(stats->current_job->thread, _performance_dump, stats); 1446 } 1447 } 1448 1449 static int 1450 performance_statistics_thread(void *arg) 1451 { 1452 struct bdevperf_aggregate_stats *stats; 1453 1454 if (g_performance_dump_active) { 1455 return -1; 1456 } 1457 1458 g_performance_dump_active = true; 1459 1460 stats = calloc(1, sizeof(*stats)); 1461 if (stats == NULL) { 1462 return -1; 1463 } 1464 1465 stats->min_latency = (double)UINT64_MAX; 1466 1467 g_show_performance_period_num++; 1468 1469 stats->io_time_in_usec = g_show_performance_period_num * g_show_performance_period_in_usec; 1470 stats->ema_period = g_show_performance_ema_period; 1471 1472 /* Iterate all of the jobs to gather stats 1473 * These jobs will not get removed here until a final performance dump is run, 1474 * so this should be safe without locking. 1475 */ 1476 stats->current_job = TAILQ_FIRST(&g_bdevperf.jobs); 1477 if (stats->current_job == NULL) { 1478 spdk_thread_send_msg(g_main_thread, _performance_dump_done, stats); 1479 } else { 1480 spdk_thread_send_msg(stats->current_job->thread, _performance_dump, stats); 1481 } 1482 1483 return -1; 1484 } 1485 1486 static void 1487 bdevperf_test(void) 1488 { 1489 struct bdevperf_job *job; 1490 1491 printf("Running I/O for %" PRIu64 " seconds...\n", g_time_in_usec / (uint64_t)SPDK_SEC_TO_USEC); 1492 fflush(stdout); 1493 1494 /* Start a timer to dump performance numbers */ 1495 g_start_tsc = spdk_get_ticks(); 1496 if (g_show_performance_real_time && !g_perf_timer) { 1497 printf("%*s\n", 107, "Latency(us)"); 1498 printf("\r %-*s: %10s %10s %10s %10s %8s %10s %10s %10s\n", 1499 28, "Device Information", "runtime(s)", "IOPS", "MiB/s", "Fail/s", "TO/s", "Average", "min", "max"); 1500 1501 g_perf_timer = SPDK_POLLER_REGISTER(performance_statistics_thread, NULL, 1502 g_show_performance_period_in_usec); 1503 } 1504 1505 /* Iterate jobs to start all I/O */ 1506 TAILQ_FOREACH(job, &g_bdevperf.jobs, link) { 1507 g_bdevperf.running_jobs++; 1508 spdk_thread_send_msg(job->thread, bdevperf_job_run, job); 1509 } 1510 } 1511 1512 static void 1513 bdevperf_bdev_removed(enum spdk_bdev_event_type type, struct spdk_bdev *bdev, void *event_ctx) 1514 { 1515 struct bdevperf_job *job = event_ctx; 1516 1517 if (SPDK_BDEV_EVENT_REMOVE == type) { 1518 bdevperf_job_drain(job); 1519 } 1520 } 1521 1522 static void 1523 bdevperf_histogram_status_cb(void *cb_arg, int status) 1524 { 1525 if (status != 0) { 1526 g_run_rc = status; 1527 if (g_continue_on_failure == false) { 1528 g_error_to_exit = true; 1529 } 1530 } 1531 1532 if (--g_bdev_count == 0) { 1533 if (g_run_rc == 0) { 1534 /* Ready to run the test */ 1535 bdevperf_test(); 1536 } else { 1537 bdevperf_test_done(NULL); 1538 } 1539 } 1540 } 1541 1542 static uint32_t g_construct_job_count = 0; 1543 1544 static int 1545 _bdevperf_enable_histogram(void *ctx, struct spdk_bdev *bdev) 1546 { 1547 bool *enable = ctx; 1548 1549 g_bdev_count++; 1550 1551 spdk_bdev_histogram_enable(bdev, bdevperf_histogram_status_cb, NULL, *enable); 1552 1553 return 0; 1554 } 1555 1556 static void 1557 bdevperf_enable_histogram(bool enable) 1558 { 1559 struct spdk_bdev *bdev; 1560 int rc; 1561 1562 /* increment initial g_bdev_count so that it will never reach 0 in the middle of iteration */ 1563 g_bdev_count = 1; 1564 1565 if (g_job_bdev_name != NULL) { 1566 bdev = spdk_bdev_get_by_name(g_job_bdev_name); 1567 if (bdev) { 1568 rc = _bdevperf_enable_histogram(&enable, bdev); 1569 } else { 1570 fprintf(stderr, "Unable to find bdev '%s'\n", g_job_bdev_name); 1571 rc = -1; 1572 } 1573 } else { 1574 rc = spdk_for_each_bdev_leaf(&enable, _bdevperf_enable_histogram); 1575 } 1576 1577 bdevperf_histogram_status_cb(NULL, rc); 1578 } 1579 1580 static void 1581 _bdevperf_construct_job_done(void *ctx) 1582 { 1583 if (--g_construct_job_count == 0) { 1584 if (g_run_rc != 0) { 1585 /* Something failed. */ 1586 bdevperf_test_done(NULL); 1587 return; 1588 } 1589 1590 /* always enable histogram. */ 1591 bdevperf_enable_histogram(true); 1592 } else if (g_run_rc != 0) { 1593 /* Reset error as some jobs constructed right */ 1594 g_run_rc = 0; 1595 if (g_continue_on_failure == false) { 1596 g_error_to_exit = true; 1597 } 1598 } 1599 } 1600 1601 /* Checkformat will not allow to use inlined type, 1602 this is a workaround */ 1603 typedef struct spdk_thread *spdk_thread_t; 1604 1605 static spdk_thread_t 1606 construct_job_thread(struct spdk_cpuset *cpumask, const char *tag) 1607 { 1608 struct spdk_cpuset tmp; 1609 1610 /* This function runs on the main thread. */ 1611 assert(g_main_thread == spdk_get_thread()); 1612 1613 /* Handle default mask */ 1614 if (spdk_cpuset_count(cpumask) == 0) { 1615 cpumask = &g_all_cpuset; 1616 } 1617 1618 /* Warn user that mask might need to be changed */ 1619 spdk_cpuset_copy(&tmp, cpumask); 1620 spdk_cpuset_or(&tmp, &g_all_cpuset); 1621 if (!spdk_cpuset_equal(&tmp, &g_all_cpuset)) { 1622 fprintf(stderr, "cpumask for '%s' is too big\n", tag); 1623 } 1624 1625 return spdk_thread_create(tag, cpumask); 1626 } 1627 1628 static uint32_t 1629 _get_next_core(void) 1630 { 1631 static uint32_t current_core = SPDK_ENV_LCORE_ID_ANY; 1632 1633 if (current_core == SPDK_ENV_LCORE_ID_ANY) { 1634 current_core = spdk_env_get_first_core(); 1635 return current_core; 1636 } 1637 1638 current_core = spdk_env_get_next_core(current_core); 1639 if (current_core == SPDK_ENV_LCORE_ID_ANY) { 1640 current_core = spdk_env_get_first_core(); 1641 } 1642 1643 return current_core; 1644 } 1645 1646 static void 1647 _bdevperf_construct_job(void *ctx) 1648 { 1649 struct bdevperf_job *job = ctx; 1650 int rc; 1651 1652 rc = spdk_bdev_open_ext(spdk_bdev_get_name(job->bdev), true, bdevperf_bdev_removed, job, 1653 &job->bdev_desc); 1654 if (rc != 0) { 1655 SPDK_ERRLOG("Could not open leaf bdev %s, error=%d\n", spdk_bdev_get_name(job->bdev), rc); 1656 g_run_rc = -EINVAL; 1657 goto end; 1658 } 1659 1660 if (g_zcopy) { 1661 if (!spdk_bdev_io_type_supported(job->bdev, SPDK_BDEV_IO_TYPE_ZCOPY)) { 1662 printf("Test requires ZCOPY but bdev module does not support ZCOPY\n"); 1663 g_run_rc = -ENOTSUP; 1664 goto end; 1665 } 1666 } 1667 1668 job->ch = spdk_bdev_get_io_channel(job->bdev_desc); 1669 if (!job->ch) { 1670 SPDK_ERRLOG("Could not get io_channel for device %s, error=%d\n", spdk_bdev_get_name(job->bdev), 1671 rc); 1672 spdk_bdev_close(job->bdev_desc); 1673 TAILQ_REMOVE(&g_bdevperf.jobs, job, link); 1674 g_run_rc = -ENOMEM; 1675 goto end; 1676 } 1677 1678 end: 1679 spdk_thread_send_msg(g_main_thread, _bdevperf_construct_job_done, NULL); 1680 } 1681 1682 static void 1683 job_init_rw(struct bdevperf_job *job, enum job_config_rw rw) 1684 { 1685 switch (rw) { 1686 case JOB_CONFIG_RW_READ: 1687 job->rw_percentage = 100; 1688 break; 1689 case JOB_CONFIG_RW_WRITE: 1690 job->rw_percentage = 0; 1691 break; 1692 case JOB_CONFIG_RW_RANDREAD: 1693 job->is_random = true; 1694 job->rw_percentage = 100; 1695 job->seed = rand(); 1696 break; 1697 case JOB_CONFIG_RW_RANDWRITE: 1698 job->is_random = true; 1699 job->rw_percentage = 0; 1700 job->seed = rand(); 1701 break; 1702 case JOB_CONFIG_RW_RW: 1703 job->is_random = false; 1704 break; 1705 case JOB_CONFIG_RW_RANDRW: 1706 job->is_random = true; 1707 job->seed = rand(); 1708 break; 1709 case JOB_CONFIG_RW_RESET: 1710 /* Reset shares the flow with verify. */ 1711 job->reset = true; 1712 /* fallthrough */ 1713 case JOB_CONFIG_RW_VERIFY: 1714 job->verify = true; 1715 /* For verify flow read is done on write completion 1716 * callback only, rw_percentage shall not be used. */ 1717 job->rw_percentage = 0; 1718 break; 1719 case JOB_CONFIG_RW_UNMAP: 1720 job->unmap = true; 1721 break; 1722 case JOB_CONFIG_RW_FLUSH: 1723 job->flush = true; 1724 break; 1725 case JOB_CONFIG_RW_WRITE_ZEROES: 1726 job->write_zeroes = true; 1727 break; 1728 } 1729 } 1730 1731 static int 1732 bdevperf_construct_job(struct spdk_bdev *bdev, struct job_config *config, 1733 struct spdk_thread *thread) 1734 { 1735 struct bdevperf_job *job; 1736 struct bdevperf_task *task; 1737 int block_size, data_block_size; 1738 int rc; 1739 int task_num, n; 1740 1741 block_size = spdk_bdev_get_block_size(bdev); 1742 data_block_size = spdk_bdev_get_data_block_size(bdev); 1743 1744 job = calloc(1, sizeof(struct bdevperf_job)); 1745 if (!job) { 1746 fprintf(stderr, "Unable to allocate memory for new job.\n"); 1747 return -ENOMEM; 1748 } 1749 1750 job->name = strdup(spdk_bdev_get_name(bdev)); 1751 if (!job->name) { 1752 fprintf(stderr, "Unable to allocate memory for job name.\n"); 1753 bdevperf_job_free(job); 1754 return -ENOMEM; 1755 } 1756 1757 job->workload_type = config->rw; 1758 job->io_size = config->bs; 1759 job->rw_percentage = config->rwmixread; 1760 job->continue_on_failure = g_continue_on_failure; 1761 job->queue_depth = config->iodepth; 1762 job->bdev = bdev; 1763 job->io_size_blocks = job->io_size / data_block_size; 1764 job->buf_size = job->io_size_blocks * block_size; 1765 job->abort = g_abort; 1766 job_init_rw(job, config->rw); 1767 1768 if ((job->io_size % data_block_size) != 0) { 1769 SPDK_ERRLOG("IO size (%d) is not multiples of data block size of bdev %s (%"PRIu32")\n", 1770 job->io_size, spdk_bdev_get_name(bdev), data_block_size); 1771 bdevperf_job_free(job); 1772 return -ENOTSUP; 1773 } 1774 1775 if (job->unmap && !spdk_bdev_io_type_supported(bdev, SPDK_BDEV_IO_TYPE_UNMAP)) { 1776 printf("Skipping %s because it does not support unmap\n", spdk_bdev_get_name(bdev)); 1777 bdevperf_job_free(job); 1778 return -ENOTSUP; 1779 } 1780 1781 if (spdk_bdev_is_dif_check_enabled(bdev, SPDK_DIF_CHECK_TYPE_REFTAG)) { 1782 job->dif_check_flags |= SPDK_DIF_FLAGS_REFTAG_CHECK; 1783 } 1784 if (spdk_bdev_is_dif_check_enabled(bdev, SPDK_DIF_CHECK_TYPE_GUARD)) { 1785 job->dif_check_flags |= SPDK_DIF_FLAGS_GUARD_CHECK; 1786 } 1787 1788 job->offset_in_ios = 0; 1789 1790 if (config->length != 0) { 1791 /* Use subset of disk */ 1792 job->size_in_ios = config->length / job->io_size_blocks; 1793 job->ios_base = config->offset / job->io_size_blocks; 1794 } else { 1795 /* Use whole disk */ 1796 job->size_in_ios = spdk_bdev_get_num_blocks(bdev) / job->io_size_blocks; 1797 job->ios_base = 0; 1798 } 1799 1800 if (job->is_random && g_zipf_theta > 0) { 1801 job->zipf = spdk_zipf_create(job->size_in_ios, g_zipf_theta, 0); 1802 } 1803 1804 if (job->verify) { 1805 if (job->size_in_ios >= UINT32_MAX) { 1806 SPDK_ERRLOG("Due to constraints of verify operation, the job storage capacity is too large\n"); 1807 bdevperf_job_free(job); 1808 return -ENOMEM; 1809 } 1810 job->outstanding = spdk_bit_array_create(job->size_in_ios); 1811 if (job->outstanding == NULL) { 1812 SPDK_ERRLOG("Could not create outstanding array bitmap for bdev %s\n", 1813 spdk_bdev_get_name(bdev)); 1814 bdevperf_job_free(job); 1815 return -ENOMEM; 1816 } 1817 if (job->queue_depth > (int)job->size_in_ios) { 1818 SPDK_WARNLOG("Due to constraints of verify job, queue depth (-q, %d) can't exceed the number of IO " 1819 "requests which can be submitted to the bdev %s simultaneously (%"PRIu64"). " 1820 "Queue depth is limited to %"PRIu64"\n", 1821 job->queue_depth, job->name, job->size_in_ios, job->size_in_ios); 1822 job->queue_depth = (int)job->size_in_ios; 1823 } 1824 } 1825 1826 job->histogram = spdk_histogram_data_alloc(); 1827 if (job->histogram == NULL) { 1828 fprintf(stderr, "Failed to allocate histogram\n"); 1829 bdevperf_job_free(job); 1830 return -ENOMEM; 1831 } 1832 1833 TAILQ_INIT(&job->task_list); 1834 1835 if (g_random_map) { 1836 if (job->size_in_ios >= UINT32_MAX) { 1837 SPDK_ERRLOG("Due to constraints of the random map, the job storage capacity is too large\n"); 1838 bdevperf_job_free(job); 1839 return -ENOMEM; 1840 } 1841 job->random_map = spdk_bit_array_create(job->size_in_ios); 1842 if (job->random_map == NULL) { 1843 SPDK_ERRLOG("Could not create random_map array bitmap for bdev %s\n", 1844 spdk_bdev_get_name(bdev)); 1845 bdevperf_job_free(job); 1846 return -ENOMEM; 1847 } 1848 } 1849 1850 task_num = job->queue_depth; 1851 if (job->reset) { 1852 task_num += 1; 1853 } 1854 if (job->abort) { 1855 task_num += job->queue_depth; 1856 } 1857 1858 TAILQ_INSERT_TAIL(&g_bdevperf.jobs, job, link); 1859 1860 for (n = 0; n < task_num; n++) { 1861 task = calloc(1, sizeof(struct bdevperf_task)); 1862 if (!task) { 1863 fprintf(stderr, "Failed to allocate task from memory\n"); 1864 spdk_zipf_free(&job->zipf); 1865 return -ENOMEM; 1866 } 1867 1868 task->buf = spdk_zmalloc(job->buf_size, spdk_bdev_get_buf_align(job->bdev), NULL, 1869 SPDK_ENV_LCORE_ID_ANY, SPDK_MALLOC_DMA); 1870 if (!task->buf) { 1871 fprintf(stderr, "Cannot allocate buf for task=%p\n", task); 1872 spdk_zipf_free(&job->zipf); 1873 free(task); 1874 return -ENOMEM; 1875 } 1876 1877 if (job->verify && job->buf_size > SPDK_BDEV_LARGE_BUF_MAX_SIZE) { 1878 task->verify_buf = spdk_zmalloc(job->buf_size, spdk_bdev_get_buf_align(job->bdev), NULL, 1879 SPDK_ENV_LCORE_ID_ANY, SPDK_MALLOC_DMA); 1880 if (!task->verify_buf) { 1881 fprintf(stderr, "Cannot allocate buf_verify for task=%p\n", task); 1882 spdk_free(task->buf); 1883 spdk_zipf_free(&job->zipf); 1884 free(task); 1885 return -ENOMEM; 1886 } 1887 1888 } 1889 1890 if (spdk_bdev_is_md_separate(job->bdev)) { 1891 task->md_buf = spdk_zmalloc(job->io_size_blocks * 1892 spdk_bdev_get_md_size(job->bdev), 0, NULL, 1893 SPDK_ENV_LCORE_ID_ANY, SPDK_MALLOC_DMA); 1894 if (!task->md_buf) { 1895 fprintf(stderr, "Cannot allocate md buf for task=%p\n", task); 1896 spdk_zipf_free(&job->zipf); 1897 spdk_free(task->verify_buf); 1898 spdk_free(task->buf); 1899 free(task); 1900 return -ENOMEM; 1901 } 1902 } 1903 1904 task->job = job; 1905 TAILQ_INSERT_TAIL(&job->task_list, task, link); 1906 } 1907 1908 job->thread = thread; 1909 1910 g_construct_job_count++; 1911 1912 rc = spdk_thread_send_msg(thread, _bdevperf_construct_job, job); 1913 assert(rc == 0); 1914 1915 return rc; 1916 } 1917 1918 static int 1919 parse_rw(const char *str, enum job_config_rw ret) 1920 { 1921 if (str == NULL) { 1922 return ret; 1923 } 1924 1925 if (!strcmp(str, "read")) { 1926 ret = JOB_CONFIG_RW_READ; 1927 } else if (!strcmp(str, "randread")) { 1928 ret = JOB_CONFIG_RW_RANDREAD; 1929 } else if (!strcmp(str, "write")) { 1930 ret = JOB_CONFIG_RW_WRITE; 1931 } else if (!strcmp(str, "randwrite")) { 1932 ret = JOB_CONFIG_RW_RANDWRITE; 1933 } else if (!strcmp(str, "verify")) { 1934 ret = JOB_CONFIG_RW_VERIFY; 1935 } else if (!strcmp(str, "reset")) { 1936 ret = JOB_CONFIG_RW_RESET; 1937 } else if (!strcmp(str, "unmap")) { 1938 ret = JOB_CONFIG_RW_UNMAP; 1939 } else if (!strcmp(str, "write_zeroes")) { 1940 ret = JOB_CONFIG_RW_WRITE_ZEROES; 1941 } else if (!strcmp(str, "flush")) { 1942 ret = JOB_CONFIG_RW_FLUSH; 1943 } else if (!strcmp(str, "rw")) { 1944 ret = JOB_CONFIG_RW_RW; 1945 } else if (!strcmp(str, "randrw")) { 1946 ret = JOB_CONFIG_RW_RANDRW; 1947 } else { 1948 fprintf(stderr, "rw must be one of\n" 1949 PATTERN_TYPES_STR "\n"); 1950 ret = BDEVPERF_CONFIG_ERROR; 1951 } 1952 1953 return ret; 1954 } 1955 1956 static const char * 1957 config_filename_next(const char *filename, char *out) 1958 { 1959 int i, k; 1960 1961 if (filename == NULL) { 1962 out[0] = '\0'; 1963 return NULL; 1964 } 1965 1966 if (filename[0] == ':') { 1967 filename++; 1968 } 1969 1970 for (i = 0, k = 0; 1971 filename[i] != '\0' && 1972 filename[i] != ':' && 1973 i < BDEVPERF_CONFIG_MAX_FILENAME && 1974 k < (BDEVPERF_CONFIG_MAX_FILENAME - 1); 1975 i++) { 1976 if (filename[i] == ' ' || filename[i] == '\t') { 1977 continue; 1978 } 1979 1980 out[k++] = filename[i]; 1981 } 1982 out[k] = 0; 1983 1984 return filename + i; 1985 } 1986 1987 static struct spdk_thread * 1988 get_lcore_thread(uint32_t lcore) 1989 { 1990 struct lcore_thread *lthread; 1991 1992 TAILQ_FOREACH(lthread, &g_lcore_thread_list, link) { 1993 if (lthread->lcore == lcore) { 1994 return lthread->thread; 1995 } 1996 } 1997 1998 return NULL; 1999 } 2000 2001 static void 2002 create_lcore_thread(uint32_t lcore) 2003 { 2004 struct lcore_thread *lthread; 2005 struct spdk_cpuset cpumask = {}; 2006 char name[32]; 2007 2008 lthread = calloc(1, sizeof(*lthread)); 2009 assert(lthread != NULL); 2010 2011 lthread->lcore = lcore; 2012 2013 snprintf(name, sizeof(name), "lcore_%u", lcore); 2014 spdk_cpuset_set_cpu(&cpumask, lcore, true); 2015 2016 lthread->thread = spdk_thread_create(name, &cpumask); 2017 assert(lthread->thread != NULL); 2018 2019 TAILQ_INSERT_TAIL(&g_lcore_thread_list, lthread, link); 2020 } 2021 2022 static void 2023 bdevperf_construct_jobs(void) 2024 { 2025 char filename[BDEVPERF_CONFIG_MAX_FILENAME]; 2026 struct spdk_thread *thread; 2027 struct job_config *config; 2028 struct spdk_bdev *bdev; 2029 const char *filenames; 2030 uint32_t i; 2031 int rc; 2032 2033 if (g_one_thread_per_lcore) { 2034 SPDK_ENV_FOREACH_CORE(i) { 2035 create_lcore_thread(i); 2036 } 2037 } 2038 2039 TAILQ_FOREACH(config, &job_config_list, link) { 2040 filenames = config->filename; 2041 2042 if (!g_one_thread_per_lcore) { 2043 thread = construct_job_thread(&config->cpumask, config->name); 2044 } else { 2045 thread = get_lcore_thread(config->lcore); 2046 } 2047 assert(thread); 2048 2049 while (filenames) { 2050 filenames = config_filename_next(filenames, filename); 2051 if (strlen(filename) == 0) { 2052 break; 2053 } 2054 2055 bdev = spdk_bdev_get_by_name(filename); 2056 if (!bdev) { 2057 fprintf(stderr, "Unable to find bdev '%s'\n", filename); 2058 g_run_rc = -EINVAL; 2059 return; 2060 } 2061 2062 rc = bdevperf_construct_job(bdev, config, thread); 2063 if (rc < 0) { 2064 g_run_rc = rc; 2065 return; 2066 } 2067 } 2068 } 2069 } 2070 2071 static int 2072 make_cli_job_config(const char *filename, int64_t offset, uint64_t range) 2073 { 2074 struct job_config *config = calloc(1, sizeof(*config)); 2075 2076 if (config == NULL) { 2077 fprintf(stderr, "Unable to allocate memory for job config\n"); 2078 return -ENOMEM; 2079 } 2080 2081 config->name = filename; 2082 config->filename = filename; 2083 config->lcore = _get_next_core(); 2084 spdk_cpuset_zero(&config->cpumask); 2085 spdk_cpuset_set_cpu(&config->cpumask, config->lcore, true); 2086 config->bs = g_io_size; 2087 config->iodepth = g_queue_depth; 2088 config->rwmixread = g_rw_percentage; 2089 config->offset = offset; 2090 config->length = range; 2091 config->rw = parse_rw(g_workload_type, BDEVPERF_CONFIG_ERROR); 2092 if ((int)config->rw == BDEVPERF_CONFIG_ERROR) { 2093 free(config); 2094 return -EINVAL; 2095 } 2096 2097 TAILQ_INSERT_TAIL(&job_config_list, config, link); 2098 return 0; 2099 } 2100 2101 static int 2102 bdevperf_construct_multithread_job_config(void *ctx, struct spdk_bdev *bdev) 2103 { 2104 uint32_t *num_cores = ctx; 2105 uint32_t i; 2106 uint64_t blocks_per_job; 2107 int64_t offset; 2108 int rc; 2109 2110 blocks_per_job = spdk_bdev_get_num_blocks(bdev) / *num_cores; 2111 offset = 0; 2112 2113 SPDK_ENV_FOREACH_CORE(i) { 2114 rc = make_cli_job_config(spdk_bdev_get_name(bdev), offset, blocks_per_job); 2115 if (rc) { 2116 return rc; 2117 } 2118 2119 offset += blocks_per_job; 2120 } 2121 2122 return 0; 2123 } 2124 2125 static void 2126 bdevperf_construct_multithread_job_configs(void) 2127 { 2128 struct spdk_bdev *bdev; 2129 uint32_t i; 2130 uint32_t num_cores; 2131 2132 num_cores = 0; 2133 SPDK_ENV_FOREACH_CORE(i) { 2134 num_cores++; 2135 } 2136 2137 if (num_cores == 0) { 2138 g_run_rc = -EINVAL; 2139 return; 2140 } 2141 2142 if (g_job_bdev_name != NULL) { 2143 bdev = spdk_bdev_get_by_name(g_job_bdev_name); 2144 if (!bdev) { 2145 fprintf(stderr, "Unable to find bdev '%s'\n", g_job_bdev_name); 2146 return; 2147 } 2148 g_run_rc = bdevperf_construct_multithread_job_config(&num_cores, bdev); 2149 } else { 2150 g_run_rc = spdk_for_each_bdev_leaf(&num_cores, bdevperf_construct_multithread_job_config); 2151 } 2152 2153 } 2154 2155 static int 2156 bdevperf_construct_job_config(void *ctx, struct spdk_bdev *bdev) 2157 { 2158 /* Construct the job */ 2159 return make_cli_job_config(spdk_bdev_get_name(bdev), 0, 0); 2160 } 2161 2162 static void 2163 bdevperf_construct_job_configs(void) 2164 { 2165 struct spdk_bdev *bdev; 2166 2167 /* There are three different modes for allocating jobs. Standard mode 2168 * (the default) creates one spdk_thread per bdev and runs the I/O job there. 2169 * 2170 * The -C flag places bdevperf into "multithread" mode, meaning it creates 2171 * one spdk_thread per bdev PER CORE, and runs a copy of the job on each. 2172 * This runs multiple threads per bdev, effectively. 2173 * 2174 * The -j flag implies "FIO" mode which tries to mimic semantic of FIO jobs. 2175 * In "FIO" mode, threads are spawned per-job instead of per-bdev. 2176 * Each FIO job can be individually parameterized by filename, cpu mask, etc, 2177 * which is different from other modes in that they only support global options. 2178 * 2179 * Both for standard mode and "multithread" mode, if the -E flag is specified, 2180 * it creates one spdk_thread PER CORE. On each core, one spdk_thread is shared by 2181 * multiple jobs. 2182 */ 2183 2184 if (g_bdevperf_conf) { 2185 goto end; 2186 } 2187 2188 if (g_multithread_mode) { 2189 bdevperf_construct_multithread_job_configs(); 2190 } else if (g_job_bdev_name != NULL) { 2191 bdev = spdk_bdev_get_by_name(g_job_bdev_name); 2192 if (bdev) { 2193 /* Construct the job */ 2194 g_run_rc = make_cli_job_config(g_job_bdev_name, 0, 0); 2195 } else { 2196 fprintf(stderr, "Unable to find bdev '%s'\n", g_job_bdev_name); 2197 } 2198 } else { 2199 g_run_rc = spdk_for_each_bdev_leaf(NULL, bdevperf_construct_job_config); 2200 } 2201 2202 end: 2203 /* Increment initial construct_jobs count so that it will never reach 0 in the middle 2204 * of iteration. 2205 */ 2206 g_construct_job_count = 1; 2207 2208 if (g_run_rc == 0) { 2209 bdevperf_construct_jobs(); 2210 } 2211 2212 _bdevperf_construct_job_done(NULL); 2213 } 2214 2215 static int 2216 parse_uint_option(struct spdk_conf_section *s, const char *name, int def) 2217 { 2218 const char *job_name; 2219 int tmp; 2220 2221 tmp = spdk_conf_section_get_intval(s, name); 2222 if (tmp == -1) { 2223 /* Field was not found. Check default value 2224 * In [global] section it is ok to have undefined values 2225 * but for other sections it is not ok */ 2226 if (def == BDEVPERF_CONFIG_UNDEFINED) { 2227 job_name = spdk_conf_section_get_name(s); 2228 if (strcmp(job_name, "global") == 0) { 2229 return def; 2230 } 2231 2232 fprintf(stderr, 2233 "Job '%s' has no '%s' assigned\n", 2234 job_name, name); 2235 return BDEVPERF_CONFIG_ERROR; 2236 } 2237 return def; 2238 } 2239 2240 /* NOTE: get_intval returns nonnegative on success */ 2241 if (tmp < 0) { 2242 fprintf(stderr, "Job '%s' has bad '%s' value.\n", 2243 spdk_conf_section_get_name(s), name); 2244 return BDEVPERF_CONFIG_ERROR; 2245 } 2246 2247 return tmp; 2248 } 2249 2250 /* CLI arguments override parameters for global sections */ 2251 static void 2252 config_set_cli_args(struct job_config *config) 2253 { 2254 if (g_job_bdev_name) { 2255 config->filename = g_job_bdev_name; 2256 } 2257 if (g_io_size > 0) { 2258 config->bs = g_io_size; 2259 } 2260 if (g_queue_depth > 0) { 2261 config->iodepth = g_queue_depth; 2262 } 2263 if (g_rw_percentage > 0) { 2264 config->rwmixread = g_rw_percentage; 2265 } 2266 if (g_workload_type) { 2267 config->rw = parse_rw(g_workload_type, config->rw); 2268 } 2269 } 2270 2271 static int 2272 read_job_config(void) 2273 { 2274 struct job_config global_default_config; 2275 struct job_config global_config; 2276 struct spdk_conf_section *s; 2277 struct job_config *config = NULL; 2278 const char *cpumask; 2279 const char *rw; 2280 bool is_global; 2281 int n = 0; 2282 int val; 2283 2284 if (g_bdevperf_conf_file == NULL) { 2285 return 0; 2286 } 2287 2288 g_bdevperf_conf = spdk_conf_allocate(); 2289 if (g_bdevperf_conf == NULL) { 2290 fprintf(stderr, "Could not allocate job config structure\n"); 2291 return 1; 2292 } 2293 2294 spdk_conf_disable_sections_merge(g_bdevperf_conf); 2295 if (spdk_conf_read(g_bdevperf_conf, g_bdevperf_conf_file)) { 2296 fprintf(stderr, "Invalid job config"); 2297 return 1; 2298 } 2299 2300 /* Initialize global defaults */ 2301 global_default_config.filename = NULL; 2302 /* Zero mask is the same as g_all_cpuset 2303 * The g_all_cpuset is not initialized yet, 2304 * so use zero mask as the default instead */ 2305 spdk_cpuset_zero(&global_default_config.cpumask); 2306 global_default_config.bs = BDEVPERF_CONFIG_UNDEFINED; 2307 global_default_config.iodepth = BDEVPERF_CONFIG_UNDEFINED; 2308 /* bdevperf has no default for -M option but in FIO the default is 50 */ 2309 global_default_config.rwmixread = 50; 2310 global_default_config.offset = 0; 2311 /* length 0 means 100% */ 2312 global_default_config.length = 0; 2313 global_default_config.rw = BDEVPERF_CONFIG_UNDEFINED; 2314 config_set_cli_args(&global_default_config); 2315 2316 if ((int)global_default_config.rw == BDEVPERF_CONFIG_ERROR) { 2317 return 1; 2318 } 2319 2320 /* There is only a single instance of global job_config 2321 * We just reset its value when we encounter new [global] section */ 2322 global_config = global_default_config; 2323 2324 for (s = spdk_conf_first_section(g_bdevperf_conf); 2325 s != NULL; 2326 s = spdk_conf_next_section(s)) { 2327 config = calloc(1, sizeof(*config)); 2328 if (config == NULL) { 2329 fprintf(stderr, "Unable to allocate memory for job config\n"); 2330 return 1; 2331 } 2332 2333 config->name = spdk_conf_section_get_name(s); 2334 is_global = strcmp(config->name, "global") == 0; 2335 2336 if (is_global) { 2337 global_config = global_default_config; 2338 } 2339 2340 config->filename = spdk_conf_section_get_val(s, "filename"); 2341 if (config->filename == NULL) { 2342 config->filename = global_config.filename; 2343 } 2344 if (!is_global) { 2345 if (config->filename == NULL) { 2346 fprintf(stderr, "Job '%s' expects 'filename' parameter\n", config->name); 2347 goto error; 2348 } else if (strnlen(config->filename, BDEVPERF_CONFIG_MAX_FILENAME) 2349 >= BDEVPERF_CONFIG_MAX_FILENAME) { 2350 fprintf(stderr, 2351 "filename for '%s' job is too long. Max length is %d\n", 2352 config->name, BDEVPERF_CONFIG_MAX_FILENAME); 2353 goto error; 2354 } 2355 } 2356 2357 cpumask = spdk_conf_section_get_val(s, "cpumask"); 2358 if (cpumask == NULL) { 2359 config->cpumask = global_config.cpumask; 2360 } else if (spdk_cpuset_parse(&config->cpumask, cpumask)) { 2361 fprintf(stderr, "Job '%s' has bad 'cpumask' value\n", config->name); 2362 goto error; 2363 } 2364 2365 config->bs = parse_uint_option(s, "bs", global_config.bs); 2366 if (config->bs == BDEVPERF_CONFIG_ERROR) { 2367 goto error; 2368 } else if (config->bs == 0) { 2369 fprintf(stderr, "'bs' of job '%s' must be greater than 0\n", config->name); 2370 goto error; 2371 } 2372 2373 config->iodepth = parse_uint_option(s, "iodepth", global_config.iodepth); 2374 if (config->iodepth == BDEVPERF_CONFIG_ERROR) { 2375 goto error; 2376 } else if (config->iodepth == 0) { 2377 fprintf(stderr, 2378 "'iodepth' of job '%s' must be greater than 0\n", 2379 config->name); 2380 goto error; 2381 } 2382 2383 config->rwmixread = parse_uint_option(s, "rwmixread", global_config.rwmixread); 2384 if (config->rwmixread == BDEVPERF_CONFIG_ERROR) { 2385 goto error; 2386 } else if (config->rwmixread > 100) { 2387 fprintf(stderr, 2388 "'rwmixread' value of '%s' job is not in 0-100 range\n", 2389 config->name); 2390 goto error; 2391 } 2392 2393 config->offset = parse_uint_option(s, "offset", global_config.offset); 2394 if (config->offset == BDEVPERF_CONFIG_ERROR) { 2395 goto error; 2396 } 2397 2398 val = parse_uint_option(s, "length", global_config.length); 2399 if (val == BDEVPERF_CONFIG_ERROR) { 2400 goto error; 2401 } 2402 config->length = val; 2403 2404 rw = spdk_conf_section_get_val(s, "rw"); 2405 config->rw = parse_rw(rw, global_config.rw); 2406 if ((int)config->rw == BDEVPERF_CONFIG_ERROR) { 2407 fprintf(stderr, "Job '%s' has bad 'rw' value\n", config->name); 2408 goto error; 2409 } else if (!is_global && (int)config->rw == BDEVPERF_CONFIG_UNDEFINED) { 2410 fprintf(stderr, "Job '%s' has no 'rw' assigned\n", config->name); 2411 goto error; 2412 } 2413 2414 if (is_global) { 2415 config_set_cli_args(config); 2416 global_config = *config; 2417 free(config); 2418 config = NULL; 2419 } else { 2420 TAILQ_INSERT_TAIL(&job_config_list, config, link); 2421 n++; 2422 } 2423 } 2424 2425 if (g_rpc_log_file_name != NULL) { 2426 g_rpc_log_file = fopen(g_rpc_log_file_name, "a"); 2427 if (g_rpc_log_file == NULL) { 2428 fprintf(stderr, "Failed to open %s\n", g_rpc_log_file_name); 2429 goto error; 2430 } 2431 } 2432 2433 printf("Using job config with %d jobs\n", n); 2434 return 0; 2435 error: 2436 free(config); 2437 return 1; 2438 } 2439 2440 static void 2441 bdevperf_run(void *arg1) 2442 { 2443 uint32_t i; 2444 2445 g_main_thread = spdk_get_thread(); 2446 2447 spdk_cpuset_zero(&g_all_cpuset); 2448 SPDK_ENV_FOREACH_CORE(i) { 2449 spdk_cpuset_set_cpu(&g_all_cpuset, i, true); 2450 } 2451 2452 if (g_wait_for_tests) { 2453 /* Do not perform any tests until RPC is received */ 2454 return; 2455 } 2456 2457 bdevperf_construct_job_configs(); 2458 } 2459 2460 static void 2461 rpc_perform_tests_reset(void) 2462 { 2463 /* Reset g_run_rc to 0 for the next test run. */ 2464 g_run_rc = 0; 2465 2466 /* Reset g_stats to 0 for the next test run. */ 2467 memset(&g_stats, 0, sizeof(g_stats)); 2468 2469 /* Reset g_show_performance_period_num to 0 for the next test run. */ 2470 g_show_performance_period_num = 0; 2471 } 2472 2473 static void 2474 rpc_perform_tests_cb(void) 2475 { 2476 struct spdk_json_write_ctx *w; 2477 struct spdk_jsonrpc_request *request = g_request; 2478 2479 g_request = NULL; 2480 2481 if (g_run_rc == 0) { 2482 w = spdk_jsonrpc_begin_result(request); 2483 spdk_json_write_uint32(w, g_run_rc); 2484 spdk_jsonrpc_end_result(request, w); 2485 } else { 2486 spdk_jsonrpc_send_error_response_fmt(request, SPDK_JSONRPC_ERROR_INTERNAL_ERROR, 2487 "bdevperf failed with error %s", spdk_strerror(-g_run_rc)); 2488 } 2489 2490 rpc_perform_tests_reset(); 2491 } 2492 2493 struct rpc_bdevperf_params { 2494 int time_in_sec; 2495 char *workload_type; 2496 int queue_depth; 2497 char *io_size; 2498 int rw_percentage; 2499 }; 2500 2501 static const struct spdk_json_object_decoder rpc_bdevperf_params_decoders[] = { 2502 {"time_in_sec", offsetof(struct rpc_bdevperf_params, time_in_sec), spdk_json_decode_int32, true}, 2503 {"workload_type", offsetof(struct rpc_bdevperf_params, workload_type), spdk_json_decode_string, true}, 2504 {"queue_depth", offsetof(struct rpc_bdevperf_params, queue_depth), spdk_json_decode_int32, true}, 2505 {"io_size", offsetof(struct rpc_bdevperf_params, io_size), spdk_json_decode_string, true}, 2506 {"rw_percentage", offsetof(struct rpc_bdevperf_params, rw_percentage), spdk_json_decode_int32, true}, 2507 }; 2508 2509 static void 2510 rpc_apply_bdevperf_params(struct rpc_bdevperf_params *params) 2511 { 2512 if (params->workload_type) { 2513 /* we need to clear previously settled parameter to avoid memory leak */ 2514 free(g_workload_type); 2515 g_workload_type = strdup(params->workload_type); 2516 } 2517 if (params->queue_depth) { 2518 g_queue_depth = params->queue_depth; 2519 } 2520 if (params->io_size) { 2521 bdevperf_parse_arg('o', params->io_size); 2522 } 2523 if (params->time_in_sec) { 2524 g_time_in_sec = params->time_in_sec; 2525 } 2526 if (params->rw_percentage) { 2527 g_rw_percentage = params->rw_percentage; 2528 g_mix_specified = true; 2529 } else { 2530 g_mix_specified = false; 2531 } 2532 } 2533 2534 static void 2535 rpc_perform_tests(struct spdk_jsonrpc_request *request, const struct spdk_json_val *params) 2536 { 2537 struct rpc_bdevperf_params req = {}, backup = {}; 2538 int rc; 2539 2540 if (g_request != NULL) { 2541 fprintf(stderr, "Another test is already in progress.\n"); 2542 spdk_jsonrpc_send_error_response(request, SPDK_JSONRPC_ERROR_INTERNAL_ERROR, 2543 spdk_strerror(-EINPROGRESS)); 2544 return; 2545 } 2546 2547 if (params) { 2548 if (spdk_json_decode_object_relaxed(params, rpc_bdevperf_params_decoders, 2549 SPDK_COUNTOF(rpc_bdevperf_params_decoders), 2550 &req)) { 2551 spdk_jsonrpc_send_error_response(request, SPDK_JSONRPC_ERROR_PARSE_ERROR, 2552 "spdk_json_decode_object failed"); 2553 return; 2554 } 2555 2556 if (g_workload_type) { 2557 backup.workload_type = strdup(g_workload_type); 2558 } 2559 backup.queue_depth = g_queue_depth; 2560 if (asprintf(&backup.io_size, "%d", g_io_size) < 0) { 2561 fprintf(stderr, "Couldn't allocate memory for queue depth"); 2562 goto rpc_error; 2563 } 2564 backup.time_in_sec = g_time_in_sec; 2565 backup.rw_percentage = g_rw_percentage; 2566 2567 rpc_apply_bdevperf_params(&req); 2568 2569 free(req.workload_type); 2570 free(req.io_size); 2571 } 2572 2573 rc = verify_test_params(); 2574 2575 if (rc) { 2576 spdk_jsonrpc_send_error_response(request, SPDK_JSONRPC_ERROR_PARSE_ERROR, 2577 "Invalid parameters provided"); 2578 /* restore old params on error */ 2579 rpc_apply_bdevperf_params(&backup); 2580 goto rpc_error; 2581 } 2582 2583 g_request = request; 2584 2585 /* Only construct job configs at the first test run. */ 2586 if (TAILQ_EMPTY(&job_config_list)) { 2587 bdevperf_construct_job_configs(); 2588 } else { 2589 bdevperf_construct_jobs(); 2590 } 2591 2592 rpc_error: 2593 free(backup.io_size); 2594 free(backup.workload_type); 2595 } 2596 SPDK_RPC_REGISTER("perform_tests", rpc_perform_tests, SPDK_RPC_RUNTIME) 2597 2598 static void 2599 _bdevperf_job_drain(void *ctx) 2600 { 2601 bdevperf_job_drain(ctx); 2602 } 2603 2604 static void 2605 spdk_bdevperf_shutdown_cb(void) 2606 { 2607 g_shutdown = true; 2608 struct bdevperf_job *job, *tmp; 2609 2610 if (g_bdevperf.running_jobs == 0) { 2611 bdevperf_test_done(NULL); 2612 return; 2613 } 2614 2615 /* Iterate jobs to stop all I/O */ 2616 TAILQ_FOREACH_SAFE(job, &g_bdevperf.jobs, link, tmp) { 2617 spdk_thread_send_msg(job->thread, _bdevperf_job_drain, job); 2618 } 2619 } 2620 2621 static int 2622 bdevperf_parse_arg(int ch, char *arg) 2623 { 2624 long long tmp; 2625 2626 if (ch == 'w') { 2627 g_workload_type = strdup(arg); 2628 } else if (ch == 'T') { 2629 g_job_bdev_name = arg; 2630 } else if (ch == 'z') { 2631 g_wait_for_tests = true; 2632 } else if (ch == 'Z') { 2633 g_zcopy = true; 2634 } else if (ch == 'X') { 2635 g_abort = true; 2636 } else if (ch == 'C') { 2637 g_multithread_mode = true; 2638 } else if (ch == 'f') { 2639 g_continue_on_failure = true; 2640 } else if (ch == 'j') { 2641 g_bdevperf_conf_file = arg; 2642 } else if (ch == 'F') { 2643 char *endptr; 2644 2645 errno = 0; 2646 g_zipf_theta = strtod(arg, &endptr); 2647 if (errno || arg == endptr || g_zipf_theta < 0) { 2648 fprintf(stderr, "Illegal zipf theta value %s\n", arg); 2649 return -EINVAL; 2650 } 2651 } else if (ch == 'l') { 2652 g_latency_display_level++; 2653 } else if (ch == 'D') { 2654 g_random_map = true; 2655 } else if (ch == 'E') { 2656 g_one_thread_per_lcore = true; 2657 } else if (ch == 'J') { 2658 g_rpc_log_file_name = arg; 2659 } else if (ch == 'o') { 2660 uint64_t size; 2661 2662 if (spdk_parse_capacity(arg, &size, NULL) != 0) { 2663 fprintf(stderr, "Invalid IO size: %s\n", arg); 2664 return -EINVAL; 2665 } 2666 g_io_size = (int)size; 2667 } else if (ch == 'U') { 2668 g_unique_writes = true; 2669 } else { 2670 tmp = spdk_strtoll(arg, 10); 2671 if (tmp < 0) { 2672 fprintf(stderr, "Parse failed for the option %c.\n", ch); 2673 return tmp; 2674 } else if (tmp >= INT_MAX) { 2675 fprintf(stderr, "Parsed option was too large %c.\n", ch); 2676 return -ERANGE; 2677 } 2678 2679 switch (ch) { 2680 case 'q': 2681 g_queue_depth = tmp; 2682 break; 2683 case 't': 2684 g_time_in_sec = tmp; 2685 break; 2686 case 'k': 2687 g_timeout_in_sec = tmp; 2688 break; 2689 case 'M': 2690 g_rw_percentage = tmp; 2691 g_mix_specified = true; 2692 break; 2693 case 'P': 2694 g_show_performance_ema_period = tmp; 2695 break; 2696 case 'S': 2697 g_show_performance_real_time = 1; 2698 g_show_performance_period_in_usec = tmp * SPDK_SEC_TO_USEC; 2699 break; 2700 default: 2701 return -EINVAL; 2702 } 2703 } 2704 return 0; 2705 } 2706 2707 static void 2708 bdevperf_usage(void) 2709 { 2710 printf(" -q <depth> io depth\n"); 2711 printf(" -o <size> io size in bytes\n"); 2712 printf(" -w <type> io pattern type, must be one of " PATTERN_TYPES_STR "\n"); 2713 printf(" -t <time> time in seconds\n"); 2714 printf(" -k <timeout> timeout in seconds to detect starved I/O (default is 0 and disabled)\n"); 2715 printf(" -M <percent> rwmixread (100 for reads, 0 for writes)\n"); 2716 printf(" -P <num> number of moving average period\n"); 2717 printf("\t\t(If set to n, show weighted mean of the previous n IO/s in real time)\n"); 2718 printf("\t\t(Formula: M = 2 / (n + 1), EMA[i+1] = IO/s * M + (1 - M) * EMA[i])\n"); 2719 printf("\t\t(only valid with -S)\n"); 2720 printf(" -S <period> show performance result in real time every <period> seconds\n"); 2721 printf(" -T <bdev> bdev to run against. Default: all available bdevs.\n"); 2722 printf(" -f continue processing I/O even after failures\n"); 2723 printf(" -F <zipf theta> use zipf distribution for random I/O\n"); 2724 printf(" -Z enable using zcopy bdev API for read or write I/O\n"); 2725 printf(" -z start bdevperf, but wait for perform_tests RPC to start tests\n"); 2726 printf(" (See examples/bdev/bdevperf/bdevperf.py)\n"); 2727 printf(" -X abort timed out I/O\n"); 2728 printf(" -C enable every core to send I/Os to each bdev\n"); 2729 printf(" -j <filename> use job config file\n"); 2730 printf(" -l display latency histogram, default: disable. -l display summary, -ll display details\n"); 2731 printf(" -D use a random map for picking offsets not previously read or written (for all jobs)\n"); 2732 printf(" -E share per lcore thread among jobs. Available only if -j is not used.\n"); 2733 printf(" -J File name to open with append mode and log JSON RPC calls.\n"); 2734 printf(" -U generate unique data for each write I/O, has no effect on non-write I/O\n"); 2735 } 2736 2737 static void 2738 bdevperf_fini(void) 2739 { 2740 free_job_config(); 2741 free(g_workload_type); 2742 2743 if (g_rpc_log_file != NULL) { 2744 fclose(g_rpc_log_file); 2745 g_rpc_log_file = NULL; 2746 } 2747 } 2748 2749 static int 2750 verify_test_params(void) 2751 { 2752 if (!g_bdevperf_conf_file && g_queue_depth <= 0) { 2753 goto out; 2754 } 2755 if (!g_bdevperf_conf_file && g_io_size <= 0) { 2756 goto out; 2757 } 2758 if (!g_bdevperf_conf_file && !g_workload_type) { 2759 goto out; 2760 } 2761 if (g_bdevperf_conf_file && g_one_thread_per_lcore) { 2762 printf("If bdevperf's config file is used, per lcore thread cannot be used\n"); 2763 goto out; 2764 } 2765 if (g_time_in_sec <= 0) { 2766 goto out; 2767 } 2768 g_time_in_usec = g_time_in_sec * SPDK_SEC_TO_USEC; 2769 2770 if (g_timeout_in_sec < 0) { 2771 goto out; 2772 } 2773 2774 if (g_abort && !g_timeout_in_sec) { 2775 printf("Timeout must be set for abort option, Ignoring g_abort\n"); 2776 } 2777 2778 if (g_show_performance_ema_period > 0 && 2779 g_show_performance_real_time == 0) { 2780 fprintf(stderr, "-P option must be specified with -S option\n"); 2781 return 1; 2782 } 2783 2784 if (g_io_size > SPDK_BDEV_LARGE_BUF_MAX_SIZE) { 2785 printf("I/O size of %d is greater than zero copy threshold (%d).\n", 2786 g_io_size, SPDK_BDEV_LARGE_BUF_MAX_SIZE); 2787 printf("Zero copy mechanism will not be used.\n"); 2788 g_zcopy = false; 2789 } 2790 2791 if (g_bdevperf_conf_file) { 2792 /* workload_type verification happens during config file parsing */ 2793 return 0; 2794 } 2795 2796 if (!strcmp(g_workload_type, "verify") || 2797 !strcmp(g_workload_type, "reset")) { 2798 g_rw_percentage = 50; 2799 g_verify = true; 2800 if (!strcmp(g_workload_type, "reset")) { 2801 g_reset = true; 2802 } 2803 } 2804 2805 if (!strcmp(g_workload_type, "read") || 2806 !strcmp(g_workload_type, "randread") || 2807 !strcmp(g_workload_type, "write") || 2808 !strcmp(g_workload_type, "randwrite") || 2809 !strcmp(g_workload_type, "verify") || 2810 !strcmp(g_workload_type, "reset") || 2811 !strcmp(g_workload_type, "unmap") || 2812 !strcmp(g_workload_type, "write_zeroes") || 2813 !strcmp(g_workload_type, "flush")) { 2814 if (g_mix_specified) { 2815 fprintf(stderr, "Ignoring -M option... Please use -M option" 2816 " only when using rw or randrw.\n"); 2817 } 2818 } 2819 2820 if (!strcmp(g_workload_type, "rw") || 2821 !strcmp(g_workload_type, "randrw")) { 2822 if (g_rw_percentage < 0 || g_rw_percentage > 100) { 2823 fprintf(stderr, 2824 "-M must be specified to value from 0 to 100 " 2825 "for rw or randrw.\n"); 2826 return 1; 2827 } 2828 } 2829 2830 if (strcmp(g_workload_type, "randread") && 2831 strcmp(g_workload_type, "randwrite") && 2832 strcmp(g_workload_type, "randrw")) { 2833 if (g_random_map) { 2834 fprintf(stderr, "Ignoring -D option... Please use -D option" 2835 " only when using randread, randwrite or randrw.\n"); 2836 return 1; 2837 } 2838 } 2839 2840 return 0; 2841 out: 2842 return 1; 2843 } 2844 2845 int 2846 main(int argc, char **argv) 2847 { 2848 struct spdk_app_opts opts = {}; 2849 int rc; 2850 2851 /* Use the runtime PID to set the random seed */ 2852 srand(getpid()); 2853 2854 spdk_app_opts_init(&opts, sizeof(opts)); 2855 opts.name = "bdevperf"; 2856 opts.rpc_addr = NULL; 2857 opts.shutdown_cb = spdk_bdevperf_shutdown_cb; 2858 2859 if ((rc = spdk_app_parse_args(argc, argv, &opts, "Zzfq:o:t:w:k:CEF:J:M:P:S:T:Xlj:DU", NULL, 2860 bdevperf_parse_arg, bdevperf_usage)) != 2861 SPDK_APP_PARSE_ARGS_SUCCESS) { 2862 return rc; 2863 } 2864 2865 /* Set the default address if no rpc_addr was provided in args 2866 * and RPC is used for starting tests */ 2867 if (g_wait_for_tests && opts.rpc_addr == NULL) { 2868 opts.rpc_addr = SPDK_DEFAULT_RPC_ADDR; 2869 } 2870 2871 if (read_job_config()) { 2872 bdevperf_fini(); 2873 return 1; 2874 } 2875 2876 if (g_rpc_log_file != NULL) { 2877 opts.rpc_log_file = g_rpc_log_file; 2878 } 2879 2880 if (verify_test_params() != 0 && !g_wait_for_tests) { 2881 spdk_app_usage(); 2882 bdevperf_usage(); 2883 bdevperf_fini(); 2884 exit(1); 2885 } 2886 2887 rc = spdk_app_start(&opts, bdevperf_run, NULL); 2888 2889 spdk_app_fini(); 2890 bdevperf_fini(); 2891 return rc; 2892 } 2893