| 1 | /* |
| 2 | * fio - the flexible io tester |
| 3 | * |
| 4 | * Copyright (C) 2005 Jens Axboe <axboe@suse.de> |
| 5 | * Copyright (C) 2006-2012 Jens Axboe <axboe@kernel.dk> |
| 6 | * |
| 7 | * The license below covers all files distributed with fio unless otherwise |
| 8 | * noted in the file itself. |
| 9 | * |
| 10 | * This program is free software; you can redistribute it and/or modify |
| 11 | * it under the terms of the GNU General Public License version 2 as |
| 12 | * published by the Free Software Foundation. |
| 13 | * |
| 14 | * This program is distributed in the hope that it will be useful, |
| 15 | * but WITHOUT ANY WARRANTY; without even the implied warranty of |
| 16 | * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the |
| 17 | * GNU General Public License for more details. |
| 18 | * |
| 19 | * You should have received a copy of the GNU General Public License |
| 20 | * along with this program; if not, write to the Free Software |
| 21 | * Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA 02111-1307 USA |
| 22 | * |
| 23 | */ |
| 24 | #include <unistd.h> |
| 25 | #include <fcntl.h> |
| 26 | #include <string.h> |
| 27 | #include <limits.h> |
| 28 | #include <signal.h> |
| 29 | #include <time.h> |
| 30 | #include <locale.h> |
| 31 | #include <assert.h> |
| 32 | #include <time.h> |
| 33 | #include <inttypes.h> |
| 34 | #include <sys/stat.h> |
| 35 | #include <sys/wait.h> |
| 36 | #include <sys/ipc.h> |
| 37 | #include <sys/mman.h> |
| 38 | #include <math.h> |
| 39 | |
| 40 | #include "fio.h" |
| 41 | #ifndef FIO_NO_HAVE_SHM_H |
| 42 | #include <sys/shm.h> |
| 43 | #endif |
| 44 | #include "hash.h" |
| 45 | #include "smalloc.h" |
| 46 | #include "verify.h" |
| 47 | #include "trim.h" |
| 48 | #include "diskutil.h" |
| 49 | #include "cgroup.h" |
| 50 | #include "profile.h" |
| 51 | #include "lib/rand.h" |
| 52 | #include "memalign.h" |
| 53 | #include "server.h" |
| 54 | #include "lib/getrusage.h" |
| 55 | #include "idletime.h" |
| 56 | #include "err.h" |
| 57 | #include "lib/tp.h" |
| 58 | #include "workqueue.h" |
| 59 | #include "lib/mountcheck.h" |
| 60 | |
| 61 | static pthread_t helper_thread; |
| 62 | static pthread_mutex_t helper_lock; |
| 63 | pthread_cond_t helper_cond; |
| 64 | int helper_do_stat = 0; |
| 65 | |
| 66 | static struct fio_mutex *startup_mutex; |
| 67 | static struct flist_head *cgroup_list; |
| 68 | static char *cgroup_mnt; |
| 69 | static int exit_value; |
| 70 | static volatile int fio_abort; |
| 71 | static unsigned int nr_process = 0; |
| 72 | static unsigned int nr_thread = 0; |
| 73 | |
| 74 | struct io_log *agg_io_log[DDIR_RWDIR_CNT]; |
| 75 | |
| 76 | int groupid = 0; |
| 77 | unsigned int thread_number = 0; |
| 78 | unsigned int stat_number = 0; |
| 79 | int shm_id = 0; |
| 80 | int temp_stall_ts; |
| 81 | unsigned long done_secs = 0; |
| 82 | volatile int helper_exit = 0; |
| 83 | |
| 84 | #define PAGE_ALIGN(buf) \ |
| 85 | (char *) (((uintptr_t) (buf) + page_mask) & ~page_mask) |
| 86 | |
| 87 | #define JOB_START_TIMEOUT (5 * 1000) |
| 88 | |
| 89 | static void sig_int(int sig) |
| 90 | { |
| 91 | if (threads) { |
| 92 | if (is_backend) |
| 93 | fio_server_got_signal(sig); |
| 94 | else { |
| 95 | log_info("\nfio: terminating on signal %d\n", sig); |
| 96 | log_info_flush(); |
| 97 | exit_value = 128; |
| 98 | } |
| 99 | |
| 100 | fio_terminate_threads(TERMINATE_ALL); |
| 101 | } |
| 102 | } |
| 103 | |
| 104 | void sig_show_status(int sig) |
| 105 | { |
| 106 | show_running_run_stats(); |
| 107 | } |
| 108 | |
| 109 | static void set_sig_handlers(void) |
| 110 | { |
| 111 | struct sigaction act; |
| 112 | |
| 113 | memset(&act, 0, sizeof(act)); |
| 114 | act.sa_handler = sig_int; |
| 115 | act.sa_flags = SA_RESTART; |
| 116 | sigaction(SIGINT, &act, NULL); |
| 117 | |
| 118 | memset(&act, 0, sizeof(act)); |
| 119 | act.sa_handler = sig_int; |
| 120 | act.sa_flags = SA_RESTART; |
| 121 | sigaction(SIGTERM, &act, NULL); |
| 122 | |
| 123 | /* Windows uses SIGBREAK as a quit signal from other applications */ |
| 124 | #ifdef WIN32 |
| 125 | memset(&act, 0, sizeof(act)); |
| 126 | act.sa_handler = sig_int; |
| 127 | act.sa_flags = SA_RESTART; |
| 128 | sigaction(SIGBREAK, &act, NULL); |
| 129 | #endif |
| 130 | |
| 131 | memset(&act, 0, sizeof(act)); |
| 132 | act.sa_handler = sig_show_status; |
| 133 | act.sa_flags = SA_RESTART; |
| 134 | sigaction(SIGUSR1, &act, NULL); |
| 135 | |
| 136 | if (is_backend) { |
| 137 | memset(&act, 0, sizeof(act)); |
| 138 | act.sa_handler = sig_int; |
| 139 | act.sa_flags = SA_RESTART; |
| 140 | sigaction(SIGPIPE, &act, NULL); |
| 141 | } |
| 142 | } |
| 143 | |
| 144 | /* |
| 145 | * Check if we are above the minimum rate given. |
| 146 | */ |
| 147 | static int __check_min_rate(struct thread_data *td, struct timeval *now, |
| 148 | enum fio_ddir ddir) |
| 149 | { |
| 150 | unsigned long long bytes = 0; |
| 151 | unsigned long iops = 0; |
| 152 | unsigned long spent; |
| 153 | unsigned long rate; |
| 154 | unsigned int ratemin = 0; |
| 155 | unsigned int rate_iops = 0; |
| 156 | unsigned int rate_iops_min = 0; |
| 157 | |
| 158 | assert(ddir_rw(ddir)); |
| 159 | |
| 160 | if (!td->o.ratemin[ddir] && !td->o.rate_iops_min[ddir]) |
| 161 | return 0; |
| 162 | |
| 163 | /* |
| 164 | * allow a 2 second settle period in the beginning |
| 165 | */ |
| 166 | if (mtime_since(&td->start, now) < 2000) |
| 167 | return 0; |
| 168 | |
| 169 | iops += td->this_io_blocks[ddir]; |
| 170 | bytes += td->this_io_bytes[ddir]; |
| 171 | ratemin += td->o.ratemin[ddir]; |
| 172 | rate_iops += td->o.rate_iops[ddir]; |
| 173 | rate_iops_min += td->o.rate_iops_min[ddir]; |
| 174 | |
| 175 | /* |
| 176 | * if rate blocks is set, sample is running |
| 177 | */ |
| 178 | if (td->rate_bytes[ddir] || td->rate_blocks[ddir]) { |
| 179 | spent = mtime_since(&td->lastrate[ddir], now); |
| 180 | if (spent < td->o.ratecycle) |
| 181 | return 0; |
| 182 | |
| 183 | if (td->o.rate[ddir] || td->o.ratemin[ddir]) { |
| 184 | /* |
| 185 | * check bandwidth specified rate |
| 186 | */ |
| 187 | if (bytes < td->rate_bytes[ddir]) { |
| 188 | log_err("%s: min rate %u not met\n", td->o.name, |
| 189 | ratemin); |
| 190 | return 1; |
| 191 | } else { |
| 192 | if (spent) |
| 193 | rate = ((bytes - td->rate_bytes[ddir]) * 1000) / spent; |
| 194 | else |
| 195 | rate = 0; |
| 196 | |
| 197 | if (rate < ratemin || |
| 198 | bytes < td->rate_bytes[ddir]) { |
| 199 | log_err("%s: min rate %u not met, got" |
| 200 | " %luKB/sec\n", td->o.name, |
| 201 | ratemin, rate); |
| 202 | return 1; |
| 203 | } |
| 204 | } |
| 205 | } else { |
| 206 | /* |
| 207 | * checks iops specified rate |
| 208 | */ |
| 209 | if (iops < rate_iops) { |
| 210 | log_err("%s: min iops rate %u not met\n", |
| 211 | td->o.name, rate_iops); |
| 212 | return 1; |
| 213 | } else { |
| 214 | if (spent) |
| 215 | rate = ((iops - td->rate_blocks[ddir]) * 1000) / spent; |
| 216 | else |
| 217 | rate = 0; |
| 218 | |
| 219 | if (rate < rate_iops_min || |
| 220 | iops < td->rate_blocks[ddir]) { |
| 221 | log_err("%s: min iops rate %u not met," |
| 222 | " got %lu\n", td->o.name, |
| 223 | rate_iops_min, rate); |
| 224 | return 1; |
| 225 | } |
| 226 | } |
| 227 | } |
| 228 | } |
| 229 | |
| 230 | td->rate_bytes[ddir] = bytes; |
| 231 | td->rate_blocks[ddir] = iops; |
| 232 | memcpy(&td->lastrate[ddir], now, sizeof(*now)); |
| 233 | return 0; |
| 234 | } |
| 235 | |
| 236 | static int check_min_rate(struct thread_data *td, struct timeval *now) |
| 237 | { |
| 238 | int ret = 0; |
| 239 | |
| 240 | if (td->bytes_done[DDIR_READ]) |
| 241 | ret |= __check_min_rate(td, now, DDIR_READ); |
| 242 | if (td->bytes_done[DDIR_WRITE]) |
| 243 | ret |= __check_min_rate(td, now, DDIR_WRITE); |
| 244 | if (td->bytes_done[DDIR_TRIM]) |
| 245 | ret |= __check_min_rate(td, now, DDIR_TRIM); |
| 246 | |
| 247 | return ret; |
| 248 | } |
| 249 | |
| 250 | /* |
| 251 | * When job exits, we can cancel the in-flight IO if we are using async |
| 252 | * io. Attempt to do so. |
| 253 | */ |
| 254 | static void cleanup_pending_aio(struct thread_data *td) |
| 255 | { |
| 256 | int r; |
| 257 | |
| 258 | /* |
| 259 | * get immediately available events, if any |
| 260 | */ |
| 261 | r = io_u_queued_complete(td, 0); |
| 262 | if (r < 0) |
| 263 | return; |
| 264 | |
| 265 | /* |
| 266 | * now cancel remaining active events |
| 267 | */ |
| 268 | if (td->io_ops->cancel) { |
| 269 | struct io_u *io_u; |
| 270 | int i; |
| 271 | |
| 272 | io_u_qiter(&td->io_u_all, io_u, i) { |
| 273 | if (io_u->flags & IO_U_F_FLIGHT) { |
| 274 | r = td->io_ops->cancel(td, io_u); |
| 275 | if (!r) |
| 276 | put_io_u(td, io_u); |
| 277 | } |
| 278 | } |
| 279 | } |
| 280 | |
| 281 | if (td->cur_depth) |
| 282 | r = io_u_queued_complete(td, td->cur_depth); |
| 283 | } |
| 284 | |
| 285 | /* |
| 286 | * Helper to handle the final sync of a file. Works just like the normal |
| 287 | * io path, just does everything sync. |
| 288 | */ |
| 289 | static int fio_io_sync(struct thread_data *td, struct fio_file *f) |
| 290 | { |
| 291 | struct io_u *io_u = __get_io_u(td); |
| 292 | int ret; |
| 293 | |
| 294 | if (!io_u) |
| 295 | return 1; |
| 296 | |
| 297 | io_u->ddir = DDIR_SYNC; |
| 298 | io_u->file = f; |
| 299 | |
| 300 | if (td_io_prep(td, io_u)) { |
| 301 | put_io_u(td, io_u); |
| 302 | return 1; |
| 303 | } |
| 304 | |
| 305 | requeue: |
| 306 | ret = td_io_queue(td, io_u); |
| 307 | if (ret < 0) { |
| 308 | td_verror(td, io_u->error, "td_io_queue"); |
| 309 | put_io_u(td, io_u); |
| 310 | return 1; |
| 311 | } else if (ret == FIO_Q_QUEUED) { |
| 312 | if (io_u_queued_complete(td, 1) < 0) |
| 313 | return 1; |
| 314 | } else if (ret == FIO_Q_COMPLETED) { |
| 315 | if (io_u->error) { |
| 316 | td_verror(td, io_u->error, "td_io_queue"); |
| 317 | return 1; |
| 318 | } |
| 319 | |
| 320 | if (io_u_sync_complete(td, io_u) < 0) |
| 321 | return 1; |
| 322 | } else if (ret == FIO_Q_BUSY) { |
| 323 | if (td_io_commit(td)) |
| 324 | return 1; |
| 325 | goto requeue; |
| 326 | } |
| 327 | |
| 328 | return 0; |
| 329 | } |
| 330 | |
| 331 | static int fio_file_fsync(struct thread_data *td, struct fio_file *f) |
| 332 | { |
| 333 | int ret; |
| 334 | |
| 335 | if (fio_file_open(f)) |
| 336 | return fio_io_sync(td, f); |
| 337 | |
| 338 | if (td_io_open_file(td, f)) |
| 339 | return 1; |
| 340 | |
| 341 | ret = fio_io_sync(td, f); |
| 342 | td_io_close_file(td, f); |
| 343 | return ret; |
| 344 | } |
| 345 | |
| 346 | static inline void __update_tv_cache(struct thread_data *td) |
| 347 | { |
| 348 | fio_gettime(&td->tv_cache, NULL); |
| 349 | } |
| 350 | |
| 351 | static inline void update_tv_cache(struct thread_data *td) |
| 352 | { |
| 353 | if ((++td->tv_cache_nr & td->tv_cache_mask) == td->tv_cache_mask) |
| 354 | __update_tv_cache(td); |
| 355 | } |
| 356 | |
| 357 | static inline int runtime_exceeded(struct thread_data *td, struct timeval *t) |
| 358 | { |
| 359 | if (in_ramp_time(td)) |
| 360 | return 0; |
| 361 | if (!td->o.timeout) |
| 362 | return 0; |
| 363 | if (utime_since(&td->epoch, t) >= td->o.timeout) |
| 364 | return 1; |
| 365 | |
| 366 | return 0; |
| 367 | } |
| 368 | |
| 369 | /* |
| 370 | * We need to update the runtime consistently in ms, but keep a running |
| 371 | * tally of the current elapsed time in microseconds for sub millisecond |
| 372 | * updates. |
| 373 | */ |
| 374 | static inline void update_runtime(struct thread_data *td, |
| 375 | unsigned long long *elapsed_us, |
| 376 | const enum fio_ddir ddir) |
| 377 | { |
| 378 | if (ddir == DDIR_WRITE && td_write(td) && td->o.verify_only) |
| 379 | return; |
| 380 | |
| 381 | td->ts.runtime[ddir] -= (elapsed_us[ddir] + 999) / 1000; |
| 382 | elapsed_us[ddir] += utime_since_now(&td->start); |
| 383 | td->ts.runtime[ddir] += (elapsed_us[ddir] + 999) / 1000; |
| 384 | } |
| 385 | |
| 386 | static int break_on_this_error(struct thread_data *td, enum fio_ddir ddir, |
| 387 | int *retptr) |
| 388 | { |
| 389 | int ret = *retptr; |
| 390 | |
| 391 | if (ret < 0 || td->error) { |
| 392 | int err = td->error; |
| 393 | enum error_type_bit eb; |
| 394 | |
| 395 | if (ret < 0) |
| 396 | err = -ret; |
| 397 | |
| 398 | eb = td_error_type(ddir, err); |
| 399 | if (!(td->o.continue_on_error & (1 << eb))) |
| 400 | return 1; |
| 401 | |
| 402 | if (td_non_fatal_error(td, eb, err)) { |
| 403 | /* |
| 404 | * Continue with the I/Os in case of |
| 405 | * a non fatal error. |
| 406 | */ |
| 407 | update_error_count(td, err); |
| 408 | td_clear_error(td); |
| 409 | *retptr = 0; |
| 410 | return 0; |
| 411 | } else if (td->o.fill_device && err == ENOSPC) { |
| 412 | /* |
| 413 | * We expect to hit this error if |
| 414 | * fill_device option is set. |
| 415 | */ |
| 416 | td_clear_error(td); |
| 417 | fio_mark_td_terminate(td); |
| 418 | return 1; |
| 419 | } else { |
| 420 | /* |
| 421 | * Stop the I/O in case of a fatal |
| 422 | * error. |
| 423 | */ |
| 424 | update_error_count(td, err); |
| 425 | return 1; |
| 426 | } |
| 427 | } |
| 428 | |
| 429 | return 0; |
| 430 | } |
| 431 | |
| 432 | static void check_update_rusage(struct thread_data *td) |
| 433 | { |
| 434 | if (td->update_rusage) { |
| 435 | td->update_rusage = 0; |
| 436 | update_rusage_stat(td); |
| 437 | fio_mutex_up(td->rusage_sem); |
| 438 | } |
| 439 | } |
| 440 | |
| 441 | static int wait_for_completions(struct thread_data *td, struct timeval *time) |
| 442 | { |
| 443 | const int full = queue_full(td); |
| 444 | int min_evts = 0; |
| 445 | int ret; |
| 446 | |
| 447 | /* |
| 448 | * if the queue is full, we MUST reap at least 1 event |
| 449 | */ |
| 450 | min_evts = min(td->o.iodepth_batch_complete_min, td->cur_depth); |
| 451 | if ((full && !min_evts) || !td->o.iodepth_batch_complete_min) |
| 452 | min_evts = 1; |
| 453 | |
| 454 | if (time && (__should_check_rate(td, DDIR_READ) || |
| 455 | __should_check_rate(td, DDIR_WRITE) || |
| 456 | __should_check_rate(td, DDIR_TRIM))) |
| 457 | fio_gettime(time, NULL); |
| 458 | |
| 459 | do { |
| 460 | ret = io_u_queued_complete(td, min_evts); |
| 461 | if (ret < 0) |
| 462 | break; |
| 463 | } while (full && (td->cur_depth > td->o.iodepth_low)); |
| 464 | |
| 465 | return ret; |
| 466 | } |
| 467 | |
| 468 | int io_queue_event(struct thread_data *td, struct io_u *io_u, int *ret, |
| 469 | enum fio_ddir ddir, uint64_t *bytes_issued, int from_verify, |
| 470 | struct timeval *comp_time) |
| 471 | { |
| 472 | int ret2; |
| 473 | |
| 474 | switch (*ret) { |
| 475 | case FIO_Q_COMPLETED: |
| 476 | if (io_u->error) { |
| 477 | *ret = -io_u->error; |
| 478 | clear_io_u(td, io_u); |
| 479 | } else if (io_u->resid) { |
| 480 | int bytes = io_u->xfer_buflen - io_u->resid; |
| 481 | struct fio_file *f = io_u->file; |
| 482 | |
| 483 | if (bytes_issued) |
| 484 | *bytes_issued += bytes; |
| 485 | |
| 486 | if (!from_verify) |
| 487 | trim_io_piece(td, io_u); |
| 488 | |
| 489 | /* |
| 490 | * zero read, fail |
| 491 | */ |
| 492 | if (!bytes) { |
| 493 | if (!from_verify) |
| 494 | unlog_io_piece(td, io_u); |
| 495 | td_verror(td, EIO, "full resid"); |
| 496 | put_io_u(td, io_u); |
| 497 | break; |
| 498 | } |
| 499 | |
| 500 | io_u->xfer_buflen = io_u->resid; |
| 501 | io_u->xfer_buf += bytes; |
| 502 | io_u->offset += bytes; |
| 503 | |
| 504 | if (ddir_rw(io_u->ddir)) |
| 505 | td->ts.short_io_u[io_u->ddir]++; |
| 506 | |
| 507 | f = io_u->file; |
| 508 | if (io_u->offset == f->real_file_size) |
| 509 | goto sync_done; |
| 510 | |
| 511 | requeue_io_u(td, &io_u); |
| 512 | } else { |
| 513 | sync_done: |
| 514 | if (comp_time && (__should_check_rate(td, DDIR_READ) || |
| 515 | __should_check_rate(td, DDIR_WRITE) || |
| 516 | __should_check_rate(td, DDIR_TRIM))) |
| 517 | fio_gettime(comp_time, NULL); |
| 518 | |
| 519 | *ret = io_u_sync_complete(td, io_u); |
| 520 | if (*ret < 0) |
| 521 | break; |
| 522 | } |
| 523 | return 0; |
| 524 | case FIO_Q_QUEUED: |
| 525 | /* |
| 526 | * if the engine doesn't have a commit hook, |
| 527 | * the io_u is really queued. if it does have such |
| 528 | * a hook, it has to call io_u_queued() itself. |
| 529 | */ |
| 530 | if (td->io_ops->commit == NULL) |
| 531 | io_u_queued(td, io_u); |
| 532 | if (bytes_issued) |
| 533 | *bytes_issued += io_u->xfer_buflen; |
| 534 | break; |
| 535 | case FIO_Q_BUSY: |
| 536 | if (!from_verify) |
| 537 | unlog_io_piece(td, io_u); |
| 538 | requeue_io_u(td, &io_u); |
| 539 | ret2 = td_io_commit(td); |
| 540 | if (ret2 < 0) |
| 541 | *ret = ret2; |
| 542 | break; |
| 543 | default: |
| 544 | assert(*ret < 0); |
| 545 | td_verror(td, -(*ret), "td_io_queue"); |
| 546 | break; |
| 547 | } |
| 548 | |
| 549 | if (break_on_this_error(td, ddir, ret)) |
| 550 | return 1; |
| 551 | |
| 552 | return 0; |
| 553 | } |
| 554 | |
| 555 | static inline int io_in_polling(struct thread_data *td) |
| 556 | { |
| 557 | return !td->o.iodepth_batch_complete_min && |
| 558 | !td->o.iodepth_batch_complete_max; |
| 559 | } |
| 560 | |
| 561 | /* |
| 562 | * The main verify engine. Runs over the writes we previously submitted, |
| 563 | * reads the blocks back in, and checks the crc/md5 of the data. |
| 564 | */ |
| 565 | static void do_verify(struct thread_data *td, uint64_t verify_bytes) |
| 566 | { |
| 567 | struct fio_file *f; |
| 568 | struct io_u *io_u; |
| 569 | int ret, min_events; |
| 570 | unsigned int i; |
| 571 | |
| 572 | dprint(FD_VERIFY, "starting loop\n"); |
| 573 | |
| 574 | /* |
| 575 | * sync io first and invalidate cache, to make sure we really |
| 576 | * read from disk. |
| 577 | */ |
| 578 | for_each_file(td, f, i) { |
| 579 | if (!fio_file_open(f)) |
| 580 | continue; |
| 581 | if (fio_io_sync(td, f)) |
| 582 | break; |
| 583 | if (file_invalidate_cache(td, f)) |
| 584 | break; |
| 585 | } |
| 586 | |
| 587 | check_update_rusage(td); |
| 588 | |
| 589 | if (td->error) |
| 590 | return; |
| 591 | |
| 592 | td_set_runstate(td, TD_VERIFYING); |
| 593 | |
| 594 | io_u = NULL; |
| 595 | while (!td->terminate) { |
| 596 | enum fio_ddir ddir; |
| 597 | int full; |
| 598 | |
| 599 | update_tv_cache(td); |
| 600 | check_update_rusage(td); |
| 601 | |
| 602 | if (runtime_exceeded(td, &td->tv_cache)) { |
| 603 | __update_tv_cache(td); |
| 604 | if (runtime_exceeded(td, &td->tv_cache)) { |
| 605 | fio_mark_td_terminate(td); |
| 606 | break; |
| 607 | } |
| 608 | } |
| 609 | |
| 610 | if (flow_threshold_exceeded(td)) |
| 611 | continue; |
| 612 | |
| 613 | if (!td->o.experimental_verify) { |
| 614 | io_u = __get_io_u(td); |
| 615 | if (!io_u) |
| 616 | break; |
| 617 | |
| 618 | if (get_next_verify(td, io_u)) { |
| 619 | put_io_u(td, io_u); |
| 620 | break; |
| 621 | } |
| 622 | |
| 623 | if (td_io_prep(td, io_u)) { |
| 624 | put_io_u(td, io_u); |
| 625 | break; |
| 626 | } |
| 627 | } else { |
| 628 | if (ddir_rw_sum(td->bytes_done) + td->o.rw_min_bs > verify_bytes) |
| 629 | break; |
| 630 | |
| 631 | while ((io_u = get_io_u(td)) != NULL) { |
| 632 | if (IS_ERR(io_u)) { |
| 633 | io_u = NULL; |
| 634 | ret = FIO_Q_BUSY; |
| 635 | goto reap; |
| 636 | } |
| 637 | |
| 638 | /* |
| 639 | * We are only interested in the places where |
| 640 | * we wrote or trimmed IOs. Turn those into |
| 641 | * reads for verification purposes. |
| 642 | */ |
| 643 | if (io_u->ddir == DDIR_READ) { |
| 644 | /* |
| 645 | * Pretend we issued it for rwmix |
| 646 | * accounting |
| 647 | */ |
| 648 | td->io_issues[DDIR_READ]++; |
| 649 | put_io_u(td, io_u); |
| 650 | continue; |
| 651 | } else if (io_u->ddir == DDIR_TRIM) { |
| 652 | io_u->ddir = DDIR_READ; |
| 653 | io_u_set(io_u, IO_U_F_TRIMMED); |
| 654 | break; |
| 655 | } else if (io_u->ddir == DDIR_WRITE) { |
| 656 | io_u->ddir = DDIR_READ; |
| 657 | break; |
| 658 | } else { |
| 659 | put_io_u(td, io_u); |
| 660 | continue; |
| 661 | } |
| 662 | } |
| 663 | |
| 664 | if (!io_u) |
| 665 | break; |
| 666 | } |
| 667 | |
| 668 | if (verify_state_should_stop(td, io_u)) { |
| 669 | put_io_u(td, io_u); |
| 670 | break; |
| 671 | } |
| 672 | |
| 673 | if (td->o.verify_async) |
| 674 | io_u->end_io = verify_io_u_async; |
| 675 | else |
| 676 | io_u->end_io = verify_io_u; |
| 677 | |
| 678 | ddir = io_u->ddir; |
| 679 | if (!td->o.disable_slat) |
| 680 | fio_gettime(&io_u->start_time, NULL); |
| 681 | |
| 682 | ret = td_io_queue(td, io_u); |
| 683 | |
| 684 | if (io_queue_event(td, io_u, &ret, ddir, NULL, 1, NULL)) |
| 685 | break; |
| 686 | |
| 687 | /* |
| 688 | * if we can queue more, do so. but check if there are |
| 689 | * completed io_u's first. Note that we can get BUSY even |
| 690 | * without IO queued, if the system is resource starved. |
| 691 | */ |
| 692 | reap: |
| 693 | full = queue_full(td) || (ret == FIO_Q_BUSY && td->cur_depth); |
| 694 | if (full || io_in_polling(td)) |
| 695 | ret = wait_for_completions(td, NULL); |
| 696 | |
| 697 | if (ret < 0) |
| 698 | break; |
| 699 | } |
| 700 | |
| 701 | check_update_rusage(td); |
| 702 | |
| 703 | if (!td->error) { |
| 704 | min_events = td->cur_depth; |
| 705 | |
| 706 | if (min_events) |
| 707 | ret = io_u_queued_complete(td, min_events); |
| 708 | } else |
| 709 | cleanup_pending_aio(td); |
| 710 | |
| 711 | td_set_runstate(td, TD_RUNNING); |
| 712 | |
| 713 | dprint(FD_VERIFY, "exiting loop\n"); |
| 714 | } |
| 715 | |
| 716 | static unsigned int exceeds_number_ios(struct thread_data *td) |
| 717 | { |
| 718 | unsigned long long number_ios; |
| 719 | |
| 720 | if (!td->o.number_ios) |
| 721 | return 0; |
| 722 | |
| 723 | number_ios = ddir_rw_sum(td->io_blocks); |
| 724 | number_ios += td->io_u_queued + td->io_u_in_flight; |
| 725 | |
| 726 | return number_ios >= (td->o.number_ios * td->loops); |
| 727 | } |
| 728 | |
| 729 | static int io_issue_bytes_exceeded(struct thread_data *td) |
| 730 | { |
| 731 | unsigned long long bytes, limit; |
| 732 | |
| 733 | if (td_rw(td)) |
| 734 | bytes = td->io_issue_bytes[DDIR_READ] + td->io_issue_bytes[DDIR_WRITE]; |
| 735 | else if (td_write(td)) |
| 736 | bytes = td->io_issue_bytes[DDIR_WRITE]; |
| 737 | else if (td_read(td)) |
| 738 | bytes = td->io_issue_bytes[DDIR_READ]; |
| 739 | else |
| 740 | bytes = td->io_issue_bytes[DDIR_TRIM]; |
| 741 | |
| 742 | if (td->o.io_limit) |
| 743 | limit = td->o.io_limit; |
| 744 | else |
| 745 | limit = td->o.size; |
| 746 | |
| 747 | limit *= td->loops; |
| 748 | return bytes >= limit || exceeds_number_ios(td); |
| 749 | } |
| 750 | |
| 751 | static int io_complete_bytes_exceeded(struct thread_data *td) |
| 752 | { |
| 753 | unsigned long long bytes, limit; |
| 754 | |
| 755 | if (td_rw(td)) |
| 756 | bytes = td->this_io_bytes[DDIR_READ] + td->this_io_bytes[DDIR_WRITE]; |
| 757 | else if (td_write(td)) |
| 758 | bytes = td->this_io_bytes[DDIR_WRITE]; |
| 759 | else if (td_read(td)) |
| 760 | bytes = td->this_io_bytes[DDIR_READ]; |
| 761 | else |
| 762 | bytes = td->this_io_bytes[DDIR_TRIM]; |
| 763 | |
| 764 | if (td->o.io_limit) |
| 765 | limit = td->o.io_limit; |
| 766 | else |
| 767 | limit = td->o.size; |
| 768 | |
| 769 | limit *= td->loops; |
| 770 | return bytes >= limit || exceeds_number_ios(td); |
| 771 | } |
| 772 | |
| 773 | /* |
| 774 | * used to calculate the next io time for rate control |
| 775 | * |
| 776 | */ |
| 777 | static long long usec_for_io(struct thread_data *td, enum fio_ddir ddir) |
| 778 | { |
| 779 | uint64_t secs, remainder, bps, bytes, iops; |
| 780 | |
| 781 | assert(!(td->flags & TD_F_CHILD)); |
| 782 | bytes = td->rate_io_issue_bytes[ddir]; |
| 783 | bps = td->rate_bps[ddir]; |
| 784 | |
| 785 | if (td->o.poisson_request) { |
| 786 | iops = bps / td->o.bs[ddir]; |
| 787 | td->last_usec += (long long)(1000000 / iops) * |
| 788 | (-logf(((float)rand() + 1) / ((float)RAND_MAX + 1))); |
| 789 | return td->last_usec; |
| 790 | } else if (bps) { |
| 791 | secs = bytes / bps; |
| 792 | remainder = bytes % bps; |
| 793 | return remainder * 1000000 / bps + secs * 1000000; |
| 794 | } else |
| 795 | return 0; |
| 796 | } |
| 797 | |
| 798 | /* |
| 799 | * Main IO worker function. It retrieves io_u's to process and queues |
| 800 | * and reaps them, checking for rate and errors along the way. |
| 801 | * |
| 802 | * Returns number of bytes written and trimmed. |
| 803 | */ |
| 804 | static uint64_t do_io(struct thread_data *td) |
| 805 | { |
| 806 | unsigned int i; |
| 807 | int ret = 0; |
| 808 | uint64_t total_bytes, bytes_issued = 0; |
| 809 | |
| 810 | if (in_ramp_time(td)) |
| 811 | td_set_runstate(td, TD_RAMP); |
| 812 | else |
| 813 | td_set_runstate(td, TD_RUNNING); |
| 814 | |
| 815 | lat_target_init(td); |
| 816 | |
| 817 | total_bytes = td->o.size; |
| 818 | /* |
| 819 | * Allow random overwrite workloads to write up to io_limit |
| 820 | * before starting verification phase as 'size' doesn't apply. |
| 821 | */ |
| 822 | if (td_write(td) && td_random(td) && td->o.norandommap) |
| 823 | total_bytes = max(total_bytes, (uint64_t) td->o.io_limit); |
| 824 | /* |
| 825 | * If verify_backlog is enabled, we'll run the verify in this |
| 826 | * handler as well. For that case, we may need up to twice the |
| 827 | * amount of bytes. |
| 828 | */ |
| 829 | if (td->o.verify != VERIFY_NONE && |
| 830 | (td_write(td) && td->o.verify_backlog)) |
| 831 | total_bytes += td->o.size; |
| 832 | |
| 833 | /* In trimwrite mode, each byte is trimmed and then written, so |
| 834 | * allow total_bytes to be twice as big */ |
| 835 | if (td_trimwrite(td)) |
| 836 | total_bytes += td->total_io_size; |
| 837 | |
| 838 | while ((td->o.read_iolog_file && !flist_empty(&td->io_log_list)) || |
| 839 | (!flist_empty(&td->trim_list)) || !io_issue_bytes_exceeded(td) || |
| 840 | td->o.time_based) { |
| 841 | struct timeval comp_time; |
| 842 | struct io_u *io_u; |
| 843 | int full; |
| 844 | enum fio_ddir ddir; |
| 845 | |
| 846 | check_update_rusage(td); |
| 847 | |
| 848 | if (td->terminate || td->done) |
| 849 | break; |
| 850 | |
| 851 | update_tv_cache(td); |
| 852 | |
| 853 | if (runtime_exceeded(td, &td->tv_cache)) { |
| 854 | __update_tv_cache(td); |
| 855 | if (runtime_exceeded(td, &td->tv_cache)) { |
| 856 | fio_mark_td_terminate(td); |
| 857 | break; |
| 858 | } |
| 859 | } |
| 860 | |
| 861 | if (flow_threshold_exceeded(td)) |
| 862 | continue; |
| 863 | |
| 864 | if (bytes_issued >= total_bytes) |
| 865 | break; |
| 866 | |
| 867 | io_u = get_io_u(td); |
| 868 | if (IS_ERR_OR_NULL(io_u)) { |
| 869 | int err = PTR_ERR(io_u); |
| 870 | |
| 871 | io_u = NULL; |
| 872 | if (err == -EBUSY) { |
| 873 | ret = FIO_Q_BUSY; |
| 874 | goto reap; |
| 875 | } |
| 876 | if (td->o.latency_target) |
| 877 | goto reap; |
| 878 | break; |
| 879 | } |
| 880 | |
| 881 | ddir = io_u->ddir; |
| 882 | |
| 883 | /* |
| 884 | * Add verification end_io handler if: |
| 885 | * - Asked to verify (!td_rw(td)) |
| 886 | * - Or the io_u is from our verify list (mixed write/ver) |
| 887 | */ |
| 888 | if (td->o.verify != VERIFY_NONE && io_u->ddir == DDIR_READ && |
| 889 | ((io_u->flags & IO_U_F_VER_LIST) || !td_rw(td))) { |
| 890 | |
| 891 | if (!td->o.verify_pattern_bytes) { |
| 892 | io_u->rand_seed = __rand(&td->verify_state); |
| 893 | if (sizeof(int) != sizeof(long *)) |
| 894 | io_u->rand_seed *= __rand(&td->verify_state); |
| 895 | } |
| 896 | |
| 897 | if (verify_state_should_stop(td, io_u)) { |
| 898 | put_io_u(td, io_u); |
| 899 | break; |
| 900 | } |
| 901 | |
| 902 | if (td->o.verify_async) |
| 903 | io_u->end_io = verify_io_u_async; |
| 904 | else |
| 905 | io_u->end_io = verify_io_u; |
| 906 | td_set_runstate(td, TD_VERIFYING); |
| 907 | } else if (in_ramp_time(td)) |
| 908 | td_set_runstate(td, TD_RAMP); |
| 909 | else |
| 910 | td_set_runstate(td, TD_RUNNING); |
| 911 | |
| 912 | /* |
| 913 | * Always log IO before it's issued, so we know the specific |
| 914 | * order of it. The logged unit will track when the IO has |
| 915 | * completed. |
| 916 | */ |
| 917 | if (td_write(td) && io_u->ddir == DDIR_WRITE && |
| 918 | td->o.do_verify && |
| 919 | td->o.verify != VERIFY_NONE && |
| 920 | !td->o.experimental_verify) |
| 921 | log_io_piece(td, io_u); |
| 922 | |
| 923 | if (td->o.io_submit_mode == IO_MODE_OFFLOAD) { |
| 924 | if (td->error) |
| 925 | break; |
| 926 | ret = workqueue_enqueue(&td->io_wq, io_u); |
| 927 | |
| 928 | if (should_check_rate(td)) |
| 929 | td->rate_next_io_time[ddir] = usec_for_io(td, ddir); |
| 930 | |
| 931 | } else { |
| 932 | ret = td_io_queue(td, io_u); |
| 933 | |
| 934 | if (should_check_rate(td)) |
| 935 | td->rate_next_io_time[ddir] = usec_for_io(td, ddir); |
| 936 | |
| 937 | if (io_queue_event(td, io_u, &ret, ddir, &bytes_issued, 0, &comp_time)) |
| 938 | break; |
| 939 | |
| 940 | /* |
| 941 | * See if we need to complete some commands. Note that |
| 942 | * we can get BUSY even without IO queued, if the |
| 943 | * system is resource starved. |
| 944 | */ |
| 945 | reap: |
| 946 | full = queue_full(td) || |
| 947 | (ret == FIO_Q_BUSY && td->cur_depth); |
| 948 | if (full || io_in_polling(td)) |
| 949 | ret = wait_for_completions(td, &comp_time); |
| 950 | } |
| 951 | if (ret < 0) |
| 952 | break; |
| 953 | if (!ddir_rw_sum(td->bytes_done) && |
| 954 | !(td->io_ops->flags & FIO_NOIO)) |
| 955 | continue; |
| 956 | |
| 957 | if (!in_ramp_time(td) && should_check_rate(td)) { |
| 958 | if (check_min_rate(td, &comp_time)) { |
| 959 | if (exitall_on_terminate) |
| 960 | fio_terminate_threads(td->groupid); |
| 961 | td_verror(td, EIO, "check_min_rate"); |
| 962 | break; |
| 963 | } |
| 964 | } |
| 965 | if (!in_ramp_time(td) && td->o.latency_target) |
| 966 | lat_target_check(td); |
| 967 | |
| 968 | if (td->o.thinktime) { |
| 969 | unsigned long long b; |
| 970 | |
| 971 | b = ddir_rw_sum(td->io_blocks); |
| 972 | if (!(b % td->o.thinktime_blocks)) { |
| 973 | int left; |
| 974 | |
| 975 | io_u_quiesce(td); |
| 976 | |
| 977 | if (td->o.thinktime_spin) |
| 978 | usec_spin(td->o.thinktime_spin); |
| 979 | |
| 980 | left = td->o.thinktime - td->o.thinktime_spin; |
| 981 | if (left) |
| 982 | usec_sleep(td, left); |
| 983 | } |
| 984 | } |
| 985 | } |
| 986 | |
| 987 | check_update_rusage(td); |
| 988 | |
| 989 | if (td->trim_entries) |
| 990 | log_err("fio: %lu trim entries leaked?\n", td->trim_entries); |
| 991 | |
| 992 | if (td->o.fill_device && td->error == ENOSPC) { |
| 993 | td->error = 0; |
| 994 | fio_mark_td_terminate(td); |
| 995 | } |
| 996 | if (!td->error) { |
| 997 | struct fio_file *f; |
| 998 | |
| 999 | if (td->o.io_submit_mode == IO_MODE_OFFLOAD) { |
| 1000 | workqueue_flush(&td->io_wq); |
| 1001 | i = 0; |
| 1002 | } else |
| 1003 | i = td->cur_depth; |
| 1004 | |
| 1005 | if (i) { |
| 1006 | ret = io_u_queued_complete(td, i); |
| 1007 | if (td->o.fill_device && td->error == ENOSPC) |
| 1008 | td->error = 0; |
| 1009 | } |
| 1010 | |
| 1011 | if (should_fsync(td) && td->o.end_fsync) { |
| 1012 | td_set_runstate(td, TD_FSYNCING); |
| 1013 | |
| 1014 | for_each_file(td, f, i) { |
| 1015 | if (!fio_file_fsync(td, f)) |
| 1016 | continue; |
| 1017 | |
| 1018 | log_err("fio: end_fsync failed for file %s\n", |
| 1019 | f->file_name); |
| 1020 | } |
| 1021 | } |
| 1022 | } else |
| 1023 | cleanup_pending_aio(td); |
| 1024 | |
| 1025 | /* |
| 1026 | * stop job if we failed doing any IO |
| 1027 | */ |
| 1028 | if (!ddir_rw_sum(td->this_io_bytes)) |
| 1029 | td->done = 1; |
| 1030 | |
| 1031 | return td->bytes_done[DDIR_WRITE] + td->bytes_done[DDIR_TRIM]; |
| 1032 | } |
| 1033 | |
| 1034 | static void cleanup_io_u(struct thread_data *td) |
| 1035 | { |
| 1036 | struct io_u *io_u; |
| 1037 | |
| 1038 | while ((io_u = io_u_qpop(&td->io_u_freelist)) != NULL) { |
| 1039 | |
| 1040 | if (td->io_ops->io_u_free) |
| 1041 | td->io_ops->io_u_free(td, io_u); |
| 1042 | |
| 1043 | fio_memfree(io_u, sizeof(*io_u)); |
| 1044 | } |
| 1045 | |
| 1046 | free_io_mem(td); |
| 1047 | |
| 1048 | io_u_rexit(&td->io_u_requeues); |
| 1049 | io_u_qexit(&td->io_u_freelist); |
| 1050 | io_u_qexit(&td->io_u_all); |
| 1051 | |
| 1052 | if (td->last_write_comp) |
| 1053 | sfree(td->last_write_comp); |
| 1054 | } |
| 1055 | |
| 1056 | static int init_io_u(struct thread_data *td) |
| 1057 | { |
| 1058 | struct io_u *io_u; |
| 1059 | unsigned int max_bs, min_write; |
| 1060 | int cl_align, i, max_units; |
| 1061 | int data_xfer = 1, err; |
| 1062 | char *p; |
| 1063 | |
| 1064 | max_units = td->o.iodepth; |
| 1065 | max_bs = td_max_bs(td); |
| 1066 | min_write = td->o.min_bs[DDIR_WRITE]; |
| 1067 | td->orig_buffer_size = (unsigned long long) max_bs |
| 1068 | * (unsigned long long) max_units; |
| 1069 | |
| 1070 | if ((td->io_ops->flags & FIO_NOIO) || !(td_read(td) || td_write(td))) |
| 1071 | data_xfer = 0; |
| 1072 | |
| 1073 | err = 0; |
| 1074 | err += io_u_rinit(&td->io_u_requeues, td->o.iodepth); |
| 1075 | err += io_u_qinit(&td->io_u_freelist, td->o.iodepth); |
| 1076 | err += io_u_qinit(&td->io_u_all, td->o.iodepth); |
| 1077 | |
| 1078 | if (err) { |
| 1079 | log_err("fio: failed setting up IO queues\n"); |
| 1080 | return 1; |
| 1081 | } |
| 1082 | |
| 1083 | /* |
| 1084 | * if we may later need to do address alignment, then add any |
| 1085 | * possible adjustment here so that we don't cause a buffer |
| 1086 | * overflow later. this adjustment may be too much if we get |
| 1087 | * lucky and the allocator gives us an aligned address. |
| 1088 | */ |
| 1089 | if (td->o.odirect || td->o.mem_align || td->o.oatomic || |
| 1090 | (td->io_ops->flags & FIO_RAWIO)) |
| 1091 | td->orig_buffer_size += page_mask + td->o.mem_align; |
| 1092 | |
| 1093 | if (td->o.mem_type == MEM_SHMHUGE || td->o.mem_type == MEM_MMAPHUGE) { |
| 1094 | unsigned long bs; |
| 1095 | |
| 1096 | bs = td->orig_buffer_size + td->o.hugepage_size - 1; |
| 1097 | td->orig_buffer_size = bs & ~(td->o.hugepage_size - 1); |
| 1098 | } |
| 1099 | |
| 1100 | if (td->orig_buffer_size != (size_t) td->orig_buffer_size) { |
| 1101 | log_err("fio: IO memory too large. Reduce max_bs or iodepth\n"); |
| 1102 | return 1; |
| 1103 | } |
| 1104 | |
| 1105 | if (data_xfer && allocate_io_mem(td)) |
| 1106 | return 1; |
| 1107 | |
| 1108 | if (td->o.odirect || td->o.mem_align || td->o.oatomic || |
| 1109 | (td->io_ops->flags & FIO_RAWIO)) |
| 1110 | p = PAGE_ALIGN(td->orig_buffer) + td->o.mem_align; |
| 1111 | else |
| 1112 | p = td->orig_buffer; |
| 1113 | |
| 1114 | cl_align = os_cache_line_size(); |
| 1115 | |
| 1116 | for (i = 0; i < max_units; i++) { |
| 1117 | void *ptr; |
| 1118 | |
| 1119 | if (td->terminate) |
| 1120 | return 1; |
| 1121 | |
| 1122 | ptr = fio_memalign(cl_align, sizeof(*io_u)); |
| 1123 | if (!ptr) { |
| 1124 | log_err("fio: unable to allocate aligned memory\n"); |
| 1125 | break; |
| 1126 | } |
| 1127 | |
| 1128 | io_u = ptr; |
| 1129 | memset(io_u, 0, sizeof(*io_u)); |
| 1130 | INIT_FLIST_HEAD(&io_u->verify_list); |
| 1131 | dprint(FD_MEM, "io_u alloc %p, index %u\n", io_u, i); |
| 1132 | |
| 1133 | if (data_xfer) { |
| 1134 | io_u->buf = p; |
| 1135 | dprint(FD_MEM, "io_u %p, mem %p\n", io_u, io_u->buf); |
| 1136 | |
| 1137 | if (td_write(td)) |
| 1138 | io_u_fill_buffer(td, io_u, min_write, max_bs); |
| 1139 | if (td_write(td) && td->o.verify_pattern_bytes) { |
| 1140 | /* |
| 1141 | * Fill the buffer with the pattern if we are |
| 1142 | * going to be doing writes. |
| 1143 | */ |
| 1144 | fill_verify_pattern(td, io_u->buf, max_bs, io_u, 0, 0); |
| 1145 | } |
| 1146 | } |
| 1147 | |
| 1148 | io_u->index = i; |
| 1149 | io_u->flags = IO_U_F_FREE; |
| 1150 | io_u_qpush(&td->io_u_freelist, io_u); |
| 1151 | |
| 1152 | /* |
| 1153 | * io_u never leaves this stack, used for iteration of all |
| 1154 | * io_u buffers. |
| 1155 | */ |
| 1156 | io_u_qpush(&td->io_u_all, io_u); |
| 1157 | |
| 1158 | if (td->io_ops->io_u_init) { |
| 1159 | int ret = td->io_ops->io_u_init(td, io_u); |
| 1160 | |
| 1161 | if (ret) { |
| 1162 | log_err("fio: failed to init engine data: %d\n", ret); |
| 1163 | return 1; |
| 1164 | } |
| 1165 | } |
| 1166 | |
| 1167 | p += max_bs; |
| 1168 | } |
| 1169 | |
| 1170 | if (td->o.verify != VERIFY_NONE) { |
| 1171 | td->last_write_comp = scalloc(max_units, sizeof(uint64_t)); |
| 1172 | if (!td->last_write_comp) { |
| 1173 | log_err("fio: failed to alloc write comp data\n"); |
| 1174 | return 1; |
| 1175 | } |
| 1176 | } |
| 1177 | |
| 1178 | return 0; |
| 1179 | } |
| 1180 | |
| 1181 | static int switch_ioscheduler(struct thread_data *td) |
| 1182 | { |
| 1183 | char tmp[256], tmp2[128]; |
| 1184 | FILE *f; |
| 1185 | int ret; |
| 1186 | |
| 1187 | if (td->io_ops->flags & FIO_DISKLESSIO) |
| 1188 | return 0; |
| 1189 | |
| 1190 | sprintf(tmp, "%s/queue/scheduler", td->sysfs_root); |
| 1191 | |
| 1192 | f = fopen(tmp, "r+"); |
| 1193 | if (!f) { |
| 1194 | if (errno == ENOENT) { |
| 1195 | log_err("fio: os or kernel doesn't support IO scheduler" |
| 1196 | " switching\n"); |
| 1197 | return 0; |
| 1198 | } |
| 1199 | td_verror(td, errno, "fopen iosched"); |
| 1200 | return 1; |
| 1201 | } |
| 1202 | |
| 1203 | /* |
| 1204 | * Set io scheduler. |
| 1205 | */ |
| 1206 | ret = fwrite(td->o.ioscheduler, strlen(td->o.ioscheduler), 1, f); |
| 1207 | if (ferror(f) || ret != 1) { |
| 1208 | td_verror(td, errno, "fwrite"); |
| 1209 | fclose(f); |
| 1210 | return 1; |
| 1211 | } |
| 1212 | |
| 1213 | rewind(f); |
| 1214 | |
| 1215 | /* |
| 1216 | * Read back and check that the selected scheduler is now the default. |
| 1217 | */ |
| 1218 | memset(tmp, 0, sizeof(tmp)); |
| 1219 | ret = fread(tmp, sizeof(tmp), 1, f); |
| 1220 | if (ferror(f) || ret < 0) { |
| 1221 | td_verror(td, errno, "fread"); |
| 1222 | fclose(f); |
| 1223 | return 1; |
| 1224 | } |
| 1225 | /* |
| 1226 | * either a list of io schedulers or "none\n" is expected. |
| 1227 | */ |
| 1228 | tmp[strlen(tmp) - 1] = '\0'; |
| 1229 | |
| 1230 | |
| 1231 | sprintf(tmp2, "[%s]", td->o.ioscheduler); |
| 1232 | if (!strstr(tmp, tmp2)) { |
| 1233 | log_err("fio: io scheduler %s not found\n", td->o.ioscheduler); |
| 1234 | td_verror(td, EINVAL, "iosched_switch"); |
| 1235 | fclose(f); |
| 1236 | return 1; |
| 1237 | } |
| 1238 | |
| 1239 | fclose(f); |
| 1240 | return 0; |
| 1241 | } |
| 1242 | |
| 1243 | static int keep_running(struct thread_data *td) |
| 1244 | { |
| 1245 | unsigned long long limit; |
| 1246 | |
| 1247 | if (td->done) |
| 1248 | return 0; |
| 1249 | if (td->o.time_based) |
| 1250 | return 1; |
| 1251 | if (td->o.loops) { |
| 1252 | td->o.loops--; |
| 1253 | return 1; |
| 1254 | } |
| 1255 | if (exceeds_number_ios(td)) |
| 1256 | return 0; |
| 1257 | |
| 1258 | if (td->o.io_limit) |
| 1259 | limit = td->o.io_limit; |
| 1260 | else |
| 1261 | limit = td->o.size; |
| 1262 | |
| 1263 | if (limit != -1ULL && ddir_rw_sum(td->io_bytes) < limit) { |
| 1264 | uint64_t diff; |
| 1265 | |
| 1266 | /* |
| 1267 | * If the difference is less than the minimum IO size, we |
| 1268 | * are done. |
| 1269 | */ |
| 1270 | diff = limit - ddir_rw_sum(td->io_bytes); |
| 1271 | if (diff < td_max_bs(td)) |
| 1272 | return 0; |
| 1273 | |
| 1274 | if (fio_files_done(td)) |
| 1275 | return 0; |
| 1276 | |
| 1277 | return 1; |
| 1278 | } |
| 1279 | |
| 1280 | return 0; |
| 1281 | } |
| 1282 | |
| 1283 | static int exec_string(struct thread_options *o, const char *string, const char *mode) |
| 1284 | { |
| 1285 | size_t newlen = strlen(string) + strlen(o->name) + strlen(mode) + 9 + 1; |
| 1286 | int ret; |
| 1287 | char *str; |
| 1288 | |
| 1289 | str = malloc(newlen); |
| 1290 | sprintf(str, "%s &> %s.%s.txt", string, o->name, mode); |
| 1291 | |
| 1292 | log_info("%s : Saving output of %s in %s.%s.txt\n",o->name, mode, o->name, mode); |
| 1293 | ret = system(str); |
| 1294 | if (ret == -1) |
| 1295 | log_err("fio: exec of cmd <%s> failed\n", str); |
| 1296 | |
| 1297 | free(str); |
| 1298 | return ret; |
| 1299 | } |
| 1300 | |
| 1301 | /* |
| 1302 | * Dry run to compute correct state of numberio for verification. |
| 1303 | */ |
| 1304 | static uint64_t do_dry_run(struct thread_data *td) |
| 1305 | { |
| 1306 | td_set_runstate(td, TD_RUNNING); |
| 1307 | |
| 1308 | while ((td->o.read_iolog_file && !flist_empty(&td->io_log_list)) || |
| 1309 | (!flist_empty(&td->trim_list)) || !io_complete_bytes_exceeded(td)) { |
| 1310 | struct io_u *io_u; |
| 1311 | int ret; |
| 1312 | |
| 1313 | if (td->terminate || td->done) |
| 1314 | break; |
| 1315 | |
| 1316 | io_u = get_io_u(td); |
| 1317 | if (!io_u) |
| 1318 | break; |
| 1319 | |
| 1320 | io_u_set(io_u, IO_U_F_FLIGHT); |
| 1321 | io_u->error = 0; |
| 1322 | io_u->resid = 0; |
| 1323 | if (ddir_rw(acct_ddir(io_u))) |
| 1324 | td->io_issues[acct_ddir(io_u)]++; |
| 1325 | if (ddir_rw(io_u->ddir)) { |
| 1326 | io_u_mark_depth(td, 1); |
| 1327 | td->ts.total_io_u[io_u->ddir]++; |
| 1328 | } |
| 1329 | |
| 1330 | if (td_write(td) && io_u->ddir == DDIR_WRITE && |
| 1331 | td->o.do_verify && |
| 1332 | td->o.verify != VERIFY_NONE && |
| 1333 | !td->o.experimental_verify) |
| 1334 | log_io_piece(td, io_u); |
| 1335 | |
| 1336 | ret = io_u_sync_complete(td, io_u); |
| 1337 | (void) ret; |
| 1338 | } |
| 1339 | |
| 1340 | return td->bytes_done[DDIR_WRITE] + td->bytes_done[DDIR_TRIM]; |
| 1341 | } |
| 1342 | |
| 1343 | static void io_workqueue_fn(struct thread_data *td, struct io_u *io_u) |
| 1344 | { |
| 1345 | const enum fio_ddir ddir = io_u->ddir; |
| 1346 | int ret; |
| 1347 | |
| 1348 | dprint(FD_RATE, "io_u %p queued by %u\n", io_u, gettid()); |
| 1349 | |
| 1350 | io_u_set(io_u, IO_U_F_NO_FILE_PUT); |
| 1351 | |
| 1352 | td->cur_depth++; |
| 1353 | |
| 1354 | ret = td_io_queue(td, io_u); |
| 1355 | |
| 1356 | dprint(FD_RATE, "io_u %p ret %d by %u\n", io_u, ret, gettid()); |
| 1357 | |
| 1358 | io_queue_event(td, io_u, &ret, ddir, NULL, 0, NULL); |
| 1359 | |
| 1360 | if (ret == FIO_Q_QUEUED) |
| 1361 | ret = io_u_queued_complete(td, 1); |
| 1362 | |
| 1363 | td->cur_depth--; |
| 1364 | } |
| 1365 | |
| 1366 | /* |
| 1367 | * Entry point for the thread based jobs. The process based jobs end up |
| 1368 | * here as well, after a little setup. |
| 1369 | */ |
| 1370 | static void *thread_main(void *data) |
| 1371 | { |
| 1372 | unsigned long long elapsed_us[DDIR_RWDIR_CNT] = { 0, }; |
| 1373 | struct thread_data *td = data; |
| 1374 | struct thread_options *o = &td->o; |
| 1375 | pthread_condattr_t attr; |
| 1376 | int clear_state; |
| 1377 | int ret; |
| 1378 | |
| 1379 | if (!o->use_thread) { |
| 1380 | setsid(); |
| 1381 | td->pid = getpid(); |
| 1382 | } else |
| 1383 | td->pid = gettid(); |
| 1384 | |
| 1385 | fio_local_clock_init(o->use_thread); |
| 1386 | |
| 1387 | dprint(FD_PROCESS, "jobs pid=%d started\n", (int) td->pid); |
| 1388 | |
| 1389 | if (is_backend) |
| 1390 | fio_server_send_start(td); |
| 1391 | |
| 1392 | INIT_FLIST_HEAD(&td->io_log_list); |
| 1393 | INIT_FLIST_HEAD(&td->io_hist_list); |
| 1394 | INIT_FLIST_HEAD(&td->verify_list); |
| 1395 | INIT_FLIST_HEAD(&td->trim_list); |
| 1396 | INIT_FLIST_HEAD(&td->next_rand_list); |
| 1397 | pthread_mutex_init(&td->io_u_lock, NULL); |
| 1398 | td->io_hist_tree = RB_ROOT; |
| 1399 | |
| 1400 | pthread_condattr_init(&attr); |
| 1401 | pthread_cond_init(&td->verify_cond, &attr); |
| 1402 | pthread_cond_init(&td->free_cond, &attr); |
| 1403 | |
| 1404 | td_set_runstate(td, TD_INITIALIZED); |
| 1405 | dprint(FD_MUTEX, "up startup_mutex\n"); |
| 1406 | fio_mutex_up(startup_mutex); |
| 1407 | dprint(FD_MUTEX, "wait on td->mutex\n"); |
| 1408 | fio_mutex_down(td->mutex); |
| 1409 | dprint(FD_MUTEX, "done waiting on td->mutex\n"); |
| 1410 | |
| 1411 | /* |
| 1412 | * A new gid requires privilege, so we need to do this before setting |
| 1413 | * the uid. |
| 1414 | */ |
| 1415 | if (o->gid != -1U && setgid(o->gid)) { |
| 1416 | td_verror(td, errno, "setgid"); |
| 1417 | goto err; |
| 1418 | } |
| 1419 | if (o->uid != -1U && setuid(o->uid)) { |
| 1420 | td_verror(td, errno, "setuid"); |
| 1421 | goto err; |
| 1422 | } |
| 1423 | |
| 1424 | /* |
| 1425 | * If we have a gettimeofday() thread, make sure we exclude that |
| 1426 | * thread from this job |
| 1427 | */ |
| 1428 | if (o->gtod_cpu) |
| 1429 | fio_cpu_clear(&o->cpumask, o->gtod_cpu); |
| 1430 | |
| 1431 | /* |
| 1432 | * Set affinity first, in case it has an impact on the memory |
| 1433 | * allocations. |
| 1434 | */ |
| 1435 | if (fio_option_is_set(o, cpumask)) { |
| 1436 | if (o->cpus_allowed_policy == FIO_CPUS_SPLIT) { |
| 1437 | ret = fio_cpus_split(&o->cpumask, td->thread_number - 1); |
| 1438 | if (!ret) { |
| 1439 | log_err("fio: no CPUs set\n"); |
| 1440 | log_err("fio: Try increasing number of available CPUs\n"); |
| 1441 | td_verror(td, EINVAL, "cpus_split"); |
| 1442 | goto err; |
| 1443 | } |
| 1444 | } |
| 1445 | ret = fio_setaffinity(td->pid, o->cpumask); |
| 1446 | if (ret == -1) { |
| 1447 | td_verror(td, errno, "cpu_set_affinity"); |
| 1448 | goto err; |
| 1449 | } |
| 1450 | } |
| 1451 | |
| 1452 | #ifdef CONFIG_LIBNUMA |
| 1453 | /* numa node setup */ |
| 1454 | if (fio_option_is_set(o, numa_cpunodes) || |
| 1455 | fio_option_is_set(o, numa_memnodes)) { |
| 1456 | struct bitmask *mask; |
| 1457 | |
| 1458 | if (numa_available() < 0) { |
| 1459 | td_verror(td, errno, "Does not support NUMA API\n"); |
| 1460 | goto err; |
| 1461 | } |
| 1462 | |
| 1463 | if (fio_option_is_set(o, numa_cpunodes)) { |
| 1464 | mask = numa_parse_nodestring(o->numa_cpunodes); |
| 1465 | ret = numa_run_on_node_mask(mask); |
| 1466 | numa_free_nodemask(mask); |
| 1467 | if (ret == -1) { |
| 1468 | td_verror(td, errno, \ |
| 1469 | "numa_run_on_node_mask failed\n"); |
| 1470 | goto err; |
| 1471 | } |
| 1472 | } |
| 1473 | |
| 1474 | if (fio_option_is_set(o, numa_memnodes)) { |
| 1475 | mask = NULL; |
| 1476 | if (o->numa_memnodes) |
| 1477 | mask = numa_parse_nodestring(o->numa_memnodes); |
| 1478 | |
| 1479 | switch (o->numa_mem_mode) { |
| 1480 | case MPOL_INTERLEAVE: |
| 1481 | numa_set_interleave_mask(mask); |
| 1482 | break; |
| 1483 | case MPOL_BIND: |
| 1484 | numa_set_membind(mask); |
| 1485 | break; |
| 1486 | case MPOL_LOCAL: |
| 1487 | numa_set_localalloc(); |
| 1488 | break; |
| 1489 | case MPOL_PREFERRED: |
| 1490 | numa_set_preferred(o->numa_mem_prefer_node); |
| 1491 | break; |
| 1492 | case MPOL_DEFAULT: |
| 1493 | default: |
| 1494 | break; |
| 1495 | } |
| 1496 | |
| 1497 | if (mask) |
| 1498 | numa_free_nodemask(mask); |
| 1499 | |
| 1500 | } |
| 1501 | } |
| 1502 | #endif |
| 1503 | |
| 1504 | if (fio_pin_memory(td)) |
| 1505 | goto err; |
| 1506 | |
| 1507 | /* |
| 1508 | * May alter parameters that init_io_u() will use, so we need to |
| 1509 | * do this first. |
| 1510 | */ |
| 1511 | if (init_iolog(td)) |
| 1512 | goto err; |
| 1513 | |
| 1514 | if (init_io_u(td)) |
| 1515 | goto err; |
| 1516 | |
| 1517 | if (o->verify_async && verify_async_init(td)) |
| 1518 | goto err; |
| 1519 | |
| 1520 | if (fio_option_is_set(o, ioprio) || |
| 1521 | fio_option_is_set(o, ioprio_class)) { |
| 1522 | ret = ioprio_set(IOPRIO_WHO_PROCESS, 0, o->ioprio_class, o->ioprio); |
| 1523 | if (ret == -1) { |
| 1524 | td_verror(td, errno, "ioprio_set"); |
| 1525 | goto err; |
| 1526 | } |
| 1527 | } |
| 1528 | |
| 1529 | if (o->cgroup && cgroup_setup(td, cgroup_list, &cgroup_mnt)) |
| 1530 | goto err; |
| 1531 | |
| 1532 | errno = 0; |
| 1533 | if (nice(o->nice) == -1 && errno != 0) { |
| 1534 | td_verror(td, errno, "nice"); |
| 1535 | goto err; |
| 1536 | } |
| 1537 | |
| 1538 | if (o->ioscheduler && switch_ioscheduler(td)) |
| 1539 | goto err; |
| 1540 | |
| 1541 | if (!o->create_serialize && setup_files(td)) |
| 1542 | goto err; |
| 1543 | |
| 1544 | if (td_io_init(td)) |
| 1545 | goto err; |
| 1546 | |
| 1547 | if (init_random_map(td)) |
| 1548 | goto err; |
| 1549 | |
| 1550 | if (o->exec_prerun && exec_string(o, o->exec_prerun, (const char *)"prerun")) |
| 1551 | goto err; |
| 1552 | |
| 1553 | if (o->pre_read) { |
| 1554 | if (pre_read_files(td) < 0) |
| 1555 | goto err; |
| 1556 | } |
| 1557 | |
| 1558 | if (td->flags & TD_F_COMPRESS_LOG) |
| 1559 | tp_init(&td->tp_data); |
| 1560 | |
| 1561 | fio_verify_init(td); |
| 1562 | |
| 1563 | if ((o->io_submit_mode == IO_MODE_OFFLOAD) && |
| 1564 | workqueue_init(td, &td->io_wq, io_workqueue_fn, td->o.iodepth)) |
| 1565 | goto err; |
| 1566 | |
| 1567 | fio_gettime(&td->epoch, NULL); |
| 1568 | fio_getrusage(&td->ru_start); |
| 1569 | memcpy(&td->bw_sample_time, &td->epoch, sizeof(td->epoch)); |
| 1570 | memcpy(&td->iops_sample_time, &td->epoch, sizeof(td->epoch)); |
| 1571 | |
| 1572 | if (o->ratemin[DDIR_READ] || o->ratemin[DDIR_WRITE] || |
| 1573 | o->ratemin[DDIR_TRIM]) { |
| 1574 | memcpy(&td->lastrate[DDIR_READ], &td->bw_sample_time, |
| 1575 | sizeof(td->bw_sample_time)); |
| 1576 | memcpy(&td->lastrate[DDIR_WRITE], &td->bw_sample_time, |
| 1577 | sizeof(td->bw_sample_time)); |
| 1578 | memcpy(&td->lastrate[DDIR_TRIM], &td->bw_sample_time, |
| 1579 | sizeof(td->bw_sample_time)); |
| 1580 | } |
| 1581 | |
| 1582 | clear_state = 0; |
| 1583 | while (keep_running(td)) { |
| 1584 | uint64_t verify_bytes; |
| 1585 | |
| 1586 | fio_gettime(&td->start, NULL); |
| 1587 | memcpy(&td->tv_cache, &td->start, sizeof(td->start)); |
| 1588 | |
| 1589 | if (clear_state) |
| 1590 | clear_io_state(td, 0); |
| 1591 | |
| 1592 | prune_io_piece_log(td); |
| 1593 | |
| 1594 | if (td->o.verify_only && (td_write(td) || td_rw(td))) |
| 1595 | verify_bytes = do_dry_run(td); |
| 1596 | else |
| 1597 | verify_bytes = do_io(td); |
| 1598 | |
| 1599 | clear_state = 1; |
| 1600 | |
| 1601 | /* |
| 1602 | * Make sure we've successfully updated the rusage stats |
| 1603 | * before waiting on the stat mutex. Otherwise we could have |
| 1604 | * the stat thread holding stat mutex and waiting for |
| 1605 | * the rusage_sem, which would never get upped because |
| 1606 | * this thread is waiting for the stat mutex. |
| 1607 | */ |
| 1608 | check_update_rusage(td); |
| 1609 | |
| 1610 | fio_mutex_down(stat_mutex); |
| 1611 | if (td_read(td) && td->io_bytes[DDIR_READ]) |
| 1612 | update_runtime(td, elapsed_us, DDIR_READ); |
| 1613 | if (td_write(td) && td->io_bytes[DDIR_WRITE]) |
| 1614 | update_runtime(td, elapsed_us, DDIR_WRITE); |
| 1615 | if (td_trim(td) && td->io_bytes[DDIR_TRIM]) |
| 1616 | update_runtime(td, elapsed_us, DDIR_TRIM); |
| 1617 | fio_gettime(&td->start, NULL); |
| 1618 | fio_mutex_up(stat_mutex); |
| 1619 | |
| 1620 | if (td->error || td->terminate) |
| 1621 | break; |
| 1622 | |
| 1623 | if (!o->do_verify || |
| 1624 | o->verify == VERIFY_NONE || |
| 1625 | (td->io_ops->flags & FIO_UNIDIR)) |
| 1626 | continue; |
| 1627 | |
| 1628 | clear_io_state(td, 0); |
| 1629 | |
| 1630 | fio_gettime(&td->start, NULL); |
| 1631 | |
| 1632 | do_verify(td, verify_bytes); |
| 1633 | |
| 1634 | /* |
| 1635 | * See comment further up for why this is done here. |
| 1636 | */ |
| 1637 | check_update_rusage(td); |
| 1638 | |
| 1639 | fio_mutex_down(stat_mutex); |
| 1640 | update_runtime(td, elapsed_us, DDIR_READ); |
| 1641 | fio_gettime(&td->start, NULL); |
| 1642 | fio_mutex_up(stat_mutex); |
| 1643 | |
| 1644 | if (td->error || td->terminate) |
| 1645 | break; |
| 1646 | } |
| 1647 | |
| 1648 | update_rusage_stat(td); |
| 1649 | td->ts.total_run_time = mtime_since_now(&td->epoch); |
| 1650 | td->ts.io_bytes[DDIR_READ] = td->io_bytes[DDIR_READ]; |
| 1651 | td->ts.io_bytes[DDIR_WRITE] = td->io_bytes[DDIR_WRITE]; |
| 1652 | td->ts.io_bytes[DDIR_TRIM] = td->io_bytes[DDIR_TRIM]; |
| 1653 | |
| 1654 | if (td->o.verify_state_save && !(td->flags & TD_F_VSTATE_SAVED) && |
| 1655 | (td->o.verify != VERIFY_NONE && td_write(td))) |
| 1656 | verify_save_state(td->thread_number); |
| 1657 | |
| 1658 | fio_unpin_memory(td); |
| 1659 | |
| 1660 | fio_writeout_logs(td); |
| 1661 | |
| 1662 | if (o->io_submit_mode == IO_MODE_OFFLOAD) |
| 1663 | workqueue_exit(&td->io_wq); |
| 1664 | |
| 1665 | if (td->flags & TD_F_COMPRESS_LOG) |
| 1666 | tp_exit(&td->tp_data); |
| 1667 | |
| 1668 | if (o->exec_postrun) |
| 1669 | exec_string(o, o->exec_postrun, (const char *)"postrun"); |
| 1670 | |
| 1671 | if (exitall_on_terminate) |
| 1672 | fio_terminate_threads(td->groupid); |
| 1673 | |
| 1674 | err: |
| 1675 | if (td->error) |
| 1676 | log_info("fio: pid=%d, err=%d/%s\n", (int) td->pid, td->error, |
| 1677 | td->verror); |
| 1678 | |
| 1679 | if (o->verify_async) |
| 1680 | verify_async_exit(td); |
| 1681 | |
| 1682 | close_and_free_files(td); |
| 1683 | cleanup_io_u(td); |
| 1684 | close_ioengine(td); |
| 1685 | cgroup_shutdown(td, &cgroup_mnt); |
| 1686 | verify_free_state(td); |
| 1687 | |
| 1688 | if (fio_option_is_set(o, cpumask)) { |
| 1689 | ret = fio_cpuset_exit(&o->cpumask); |
| 1690 | if (ret) |
| 1691 | td_verror(td, ret, "fio_cpuset_exit"); |
| 1692 | } |
| 1693 | |
| 1694 | /* |
| 1695 | * do this very late, it will log file closing as well |
| 1696 | */ |
| 1697 | if (o->write_iolog_file) |
| 1698 | write_iolog_close(td); |
| 1699 | |
| 1700 | fio_mutex_remove(td->mutex); |
| 1701 | td->mutex = NULL; |
| 1702 | |
| 1703 | td_set_runstate(td, TD_EXITED); |
| 1704 | |
| 1705 | /* |
| 1706 | * Do this last after setting our runstate to exited, so we |
| 1707 | * know that the stat thread is signaled. |
| 1708 | */ |
| 1709 | check_update_rusage(td); |
| 1710 | |
| 1711 | return (void *) (uintptr_t) td->error; |
| 1712 | } |
| 1713 | |
| 1714 | |
| 1715 | /* |
| 1716 | * We cannot pass the td data into a forked process, so attach the td and |
| 1717 | * pass it to the thread worker. |
| 1718 | */ |
| 1719 | static int fork_main(int shmid, int offset) |
| 1720 | { |
| 1721 | struct thread_data *td; |
| 1722 | void *data, *ret; |
| 1723 | |
| 1724 | #if !defined(__hpux) && !defined(CONFIG_NO_SHM) |
| 1725 | data = shmat(shmid, NULL, 0); |
| 1726 | if (data == (void *) -1) { |
| 1727 | int __err = errno; |
| 1728 | |
| 1729 | perror("shmat"); |
| 1730 | return __err; |
| 1731 | } |
| 1732 | #else |
| 1733 | /* |
| 1734 | * HP-UX inherits shm mappings? |
| 1735 | */ |
| 1736 | data = threads; |
| 1737 | #endif |
| 1738 | |
| 1739 | td = data + offset * sizeof(struct thread_data); |
| 1740 | ret = thread_main(td); |
| 1741 | shmdt(data); |
| 1742 | return (int) (uintptr_t) ret; |
| 1743 | } |
| 1744 | |
| 1745 | static void dump_td_info(struct thread_data *td) |
| 1746 | { |
| 1747 | log_err("fio: job '%s' hasn't exited in %lu seconds, it appears to " |
| 1748 | "be stuck. Doing forceful exit of this job.\n", td->o.name, |
| 1749 | (unsigned long) time_since_now(&td->terminate_time)); |
| 1750 | } |
| 1751 | |
| 1752 | /* |
| 1753 | * Run over the job map and reap the threads that have exited, if any. |
| 1754 | */ |
| 1755 | static void reap_threads(unsigned int *nr_running, unsigned int *t_rate, |
| 1756 | unsigned int *m_rate) |
| 1757 | { |
| 1758 | struct thread_data *td; |
| 1759 | unsigned int cputhreads, realthreads, pending; |
| 1760 | int i, status, ret; |
| 1761 | |
| 1762 | /* |
| 1763 | * reap exited threads (TD_EXITED -> TD_REAPED) |
| 1764 | */ |
| 1765 | realthreads = pending = cputhreads = 0; |
| 1766 | for_each_td(td, i) { |
| 1767 | int flags = 0; |
| 1768 | |
| 1769 | /* |
| 1770 | * ->io_ops is NULL for a thread that has closed its |
| 1771 | * io engine |
| 1772 | */ |
| 1773 | if (td->io_ops && !strcmp(td->io_ops->name, "cpuio")) |
| 1774 | cputhreads++; |
| 1775 | else |
| 1776 | realthreads++; |
| 1777 | |
| 1778 | if (!td->pid) { |
| 1779 | pending++; |
| 1780 | continue; |
| 1781 | } |
| 1782 | if (td->runstate == TD_REAPED) |
| 1783 | continue; |
| 1784 | if (td->o.use_thread) { |
| 1785 | if (td->runstate == TD_EXITED) { |
| 1786 | td_set_runstate(td, TD_REAPED); |
| 1787 | goto reaped; |
| 1788 | } |
| 1789 | continue; |
| 1790 | } |
| 1791 | |
| 1792 | flags = WNOHANG; |
| 1793 | if (td->runstate == TD_EXITED) |
| 1794 | flags = 0; |
| 1795 | |
| 1796 | /* |
| 1797 | * check if someone quit or got killed in an unusual way |
| 1798 | */ |
| 1799 | ret = waitpid(td->pid, &status, flags); |
| 1800 | if (ret < 0) { |
| 1801 | if (errno == ECHILD) { |
| 1802 | log_err("fio: pid=%d disappeared %d\n", |
| 1803 | (int) td->pid, td->runstate); |
| 1804 | td->sig = ECHILD; |
| 1805 | td_set_runstate(td, TD_REAPED); |
| 1806 | goto reaped; |
| 1807 | } |
| 1808 | perror("waitpid"); |
| 1809 | } else if (ret == td->pid) { |
| 1810 | if (WIFSIGNALED(status)) { |
| 1811 | int sig = WTERMSIG(status); |
| 1812 | |
| 1813 | if (sig != SIGTERM && sig != SIGUSR2) |
| 1814 | log_err("fio: pid=%d, got signal=%d\n", |
| 1815 | (int) td->pid, sig); |
| 1816 | td->sig = sig; |
| 1817 | td_set_runstate(td, TD_REAPED); |
| 1818 | goto reaped; |
| 1819 | } |
| 1820 | if (WIFEXITED(status)) { |
| 1821 | if (WEXITSTATUS(status) && !td->error) |
| 1822 | td->error = WEXITSTATUS(status); |
| 1823 | |
| 1824 | td_set_runstate(td, TD_REAPED); |
| 1825 | goto reaped; |
| 1826 | } |
| 1827 | } |
| 1828 | |
| 1829 | /* |
| 1830 | * If the job is stuck, do a forceful timeout of it and |
| 1831 | * move on. |
| 1832 | */ |
| 1833 | if (td->terminate && |
| 1834 | time_since_now(&td->terminate_time) >= FIO_REAP_TIMEOUT) { |
| 1835 | dump_td_info(td); |
| 1836 | td_set_runstate(td, TD_REAPED); |
| 1837 | goto reaped; |
| 1838 | } |
| 1839 | |
| 1840 | /* |
| 1841 | * thread is not dead, continue |
| 1842 | */ |
| 1843 | pending++; |
| 1844 | continue; |
| 1845 | reaped: |
| 1846 | (*nr_running)--; |
| 1847 | (*m_rate) -= ddir_rw_sum(td->o.ratemin); |
| 1848 | (*t_rate) -= ddir_rw_sum(td->o.rate); |
| 1849 | if (!td->pid) |
| 1850 | pending--; |
| 1851 | |
| 1852 | if (td->error) |
| 1853 | exit_value++; |
| 1854 | |
| 1855 | done_secs += mtime_since_now(&td->epoch) / 1000; |
| 1856 | profile_td_exit(td); |
| 1857 | } |
| 1858 | |
| 1859 | if (*nr_running == cputhreads && !pending && realthreads) |
| 1860 | fio_terminate_threads(TERMINATE_ALL); |
| 1861 | } |
| 1862 | |
| 1863 | static int __check_trigger_file(void) |
| 1864 | { |
| 1865 | struct stat sb; |
| 1866 | |
| 1867 | if (!trigger_file) |
| 1868 | return 0; |
| 1869 | |
| 1870 | if (stat(trigger_file, &sb)) |
| 1871 | return 0; |
| 1872 | |
| 1873 | if (unlink(trigger_file) < 0) |
| 1874 | log_err("fio: failed to unlink %s: %s\n", trigger_file, |
| 1875 | strerror(errno)); |
| 1876 | |
| 1877 | return 1; |
| 1878 | } |
| 1879 | |
| 1880 | static int trigger_timedout(void) |
| 1881 | { |
| 1882 | if (trigger_timeout) |
| 1883 | return time_since_genesis() >= trigger_timeout; |
| 1884 | |
| 1885 | return 0; |
| 1886 | } |
| 1887 | |
| 1888 | void exec_trigger(const char *cmd) |
| 1889 | { |
| 1890 | int ret; |
| 1891 | |
| 1892 | if (!cmd) |
| 1893 | return; |
| 1894 | |
| 1895 | ret = system(cmd); |
| 1896 | if (ret == -1) |
| 1897 | log_err("fio: failed executing %s trigger\n", cmd); |
| 1898 | } |
| 1899 | |
| 1900 | void check_trigger_file(void) |
| 1901 | { |
| 1902 | if (__check_trigger_file() || trigger_timedout()) { |
| 1903 | if (nr_clients) |
| 1904 | fio_clients_send_trigger(trigger_remote_cmd); |
| 1905 | else { |
| 1906 | verify_save_state(IO_LIST_ALL); |
| 1907 | fio_terminate_threads(TERMINATE_ALL); |
| 1908 | exec_trigger(trigger_cmd); |
| 1909 | } |
| 1910 | } |
| 1911 | } |
| 1912 | |
| 1913 | static int fio_verify_load_state(struct thread_data *td) |
| 1914 | { |
| 1915 | int ret; |
| 1916 | |
| 1917 | if (!td->o.verify_state) |
| 1918 | return 0; |
| 1919 | |
| 1920 | if (is_backend) { |
| 1921 | void *data; |
| 1922 | int ver; |
| 1923 | |
| 1924 | ret = fio_server_get_verify_state(td->o.name, |
| 1925 | td->thread_number - 1, &data, &ver); |
| 1926 | if (!ret) |
| 1927 | verify_convert_assign_state(td, data, ver); |
| 1928 | } else |
| 1929 | ret = verify_load_state(td, "local"); |
| 1930 | |
| 1931 | return ret; |
| 1932 | } |
| 1933 | |
| 1934 | static void do_usleep(unsigned int usecs) |
| 1935 | { |
| 1936 | check_for_running_stats(); |
| 1937 | check_trigger_file(); |
| 1938 | usleep(usecs); |
| 1939 | } |
| 1940 | |
| 1941 | static int check_mount_writes(struct thread_data *td) |
| 1942 | { |
| 1943 | struct fio_file *f; |
| 1944 | unsigned int i; |
| 1945 | |
| 1946 | if (!td_write(td) || td->o.allow_mounted_write) |
| 1947 | return 0; |
| 1948 | |
| 1949 | for_each_file(td, f, i) { |
| 1950 | if (f->filetype != FIO_TYPE_BD) |
| 1951 | continue; |
| 1952 | if (device_is_mounted(f->file_name)) |
| 1953 | goto mounted; |
| 1954 | } |
| 1955 | |
| 1956 | return 0; |
| 1957 | mounted: |
| 1958 | log_err("fio: %s appears mounted, and 'allow_mounted_write' isn't set. Aborting.", f->file_name); |
| 1959 | return 1; |
| 1960 | } |
| 1961 | |
| 1962 | /* |
| 1963 | * Main function for kicking off and reaping jobs, as needed. |
| 1964 | */ |
| 1965 | static void run_threads(void) |
| 1966 | { |
| 1967 | struct thread_data *td; |
| 1968 | unsigned int i, todo, nr_running, m_rate, t_rate, nr_started; |
| 1969 | uint64_t spent; |
| 1970 | |
| 1971 | if (fio_gtod_offload && fio_start_gtod_thread()) |
| 1972 | return; |
| 1973 | |
| 1974 | fio_idle_prof_init(); |
| 1975 | |
| 1976 | set_sig_handlers(); |
| 1977 | |
| 1978 | nr_thread = nr_process = 0; |
| 1979 | for_each_td(td, i) { |
| 1980 | if (check_mount_writes(td)) |
| 1981 | return; |
| 1982 | if (td->o.use_thread) |
| 1983 | nr_thread++; |
| 1984 | else |
| 1985 | nr_process++; |
| 1986 | } |
| 1987 | |
| 1988 | if (output_format & FIO_OUTPUT_NORMAL) { |
| 1989 | log_info("Starting "); |
| 1990 | if (nr_thread) |
| 1991 | log_info("%d thread%s", nr_thread, |
| 1992 | nr_thread > 1 ? "s" : ""); |
| 1993 | if (nr_process) { |
| 1994 | if (nr_thread) |
| 1995 | log_info(" and "); |
| 1996 | log_info("%d process%s", nr_process, |
| 1997 | nr_process > 1 ? "es" : ""); |
| 1998 | } |
| 1999 | log_info("\n"); |
| 2000 | log_info_flush(); |
| 2001 | } |
| 2002 | |
| 2003 | todo = thread_number; |
| 2004 | nr_running = 0; |
| 2005 | nr_started = 0; |
| 2006 | m_rate = t_rate = 0; |
| 2007 | |
| 2008 | for_each_td(td, i) { |
| 2009 | print_status_init(td->thread_number - 1); |
| 2010 | |
| 2011 | if (!td->o.create_serialize) |
| 2012 | continue; |
| 2013 | |
| 2014 | if (fio_verify_load_state(td)) |
| 2015 | goto reap; |
| 2016 | |
| 2017 | /* |
| 2018 | * do file setup here so it happens sequentially, |
| 2019 | * we don't want X number of threads getting their |
| 2020 | * client data interspersed on disk |
| 2021 | */ |
| 2022 | if (setup_files(td)) { |
| 2023 | reap: |
| 2024 | exit_value++; |
| 2025 | if (td->error) |
| 2026 | log_err("fio: pid=%d, err=%d/%s\n", |
| 2027 | (int) td->pid, td->error, td->verror); |
| 2028 | td_set_runstate(td, TD_REAPED); |
| 2029 | todo--; |
| 2030 | } else { |
| 2031 | struct fio_file *f; |
| 2032 | unsigned int j; |
| 2033 | |
| 2034 | /* |
| 2035 | * for sharing to work, each job must always open |
| 2036 | * its own files. so close them, if we opened them |
| 2037 | * for creation |
| 2038 | */ |
| 2039 | for_each_file(td, f, j) { |
| 2040 | if (fio_file_open(f)) |
| 2041 | td_io_close_file(td, f); |
| 2042 | } |
| 2043 | } |
| 2044 | } |
| 2045 | |
| 2046 | /* start idle threads before io threads start to run */ |
| 2047 | fio_idle_prof_start(); |
| 2048 | |
| 2049 | set_genesis_time(); |
| 2050 | |
| 2051 | while (todo) { |
| 2052 | struct thread_data *map[REAL_MAX_JOBS]; |
| 2053 | struct timeval this_start; |
| 2054 | int this_jobs = 0, left; |
| 2055 | |
| 2056 | /* |
| 2057 | * create threads (TD_NOT_CREATED -> TD_CREATED) |
| 2058 | */ |
| 2059 | for_each_td(td, i) { |
| 2060 | if (td->runstate != TD_NOT_CREATED) |
| 2061 | continue; |
| 2062 | |
| 2063 | /* |
| 2064 | * never got a chance to start, killed by other |
| 2065 | * thread for some reason |
| 2066 | */ |
| 2067 | if (td->terminate) { |
| 2068 | todo--; |
| 2069 | continue; |
| 2070 | } |
| 2071 | |
| 2072 | if (td->o.start_delay) { |
| 2073 | spent = utime_since_genesis(); |
| 2074 | |
| 2075 | if (td->o.start_delay > spent) |
| 2076 | continue; |
| 2077 | } |
| 2078 | |
| 2079 | if (td->o.stonewall && (nr_started || nr_running)) { |
| 2080 | dprint(FD_PROCESS, "%s: stonewall wait\n", |
| 2081 | td->o.name); |
| 2082 | break; |
| 2083 | } |
| 2084 | |
| 2085 | init_disk_util(td); |
| 2086 | |
| 2087 | td->rusage_sem = fio_mutex_init(FIO_MUTEX_LOCKED); |
| 2088 | td->update_rusage = 0; |
| 2089 | |
| 2090 | /* |
| 2091 | * Set state to created. Thread will transition |
| 2092 | * to TD_INITIALIZED when it's done setting up. |
| 2093 | */ |
| 2094 | td_set_runstate(td, TD_CREATED); |
| 2095 | map[this_jobs++] = td; |
| 2096 | nr_started++; |
| 2097 | |
| 2098 | if (td->o.use_thread) { |
| 2099 | int ret; |
| 2100 | |
| 2101 | dprint(FD_PROCESS, "will pthread_create\n"); |
| 2102 | ret = pthread_create(&td->thread, NULL, |
| 2103 | thread_main, td); |
| 2104 | if (ret) { |
| 2105 | log_err("pthread_create: %s\n", |
| 2106 | strerror(ret)); |
| 2107 | nr_started--; |
| 2108 | break; |
| 2109 | } |
| 2110 | ret = pthread_detach(td->thread); |
| 2111 | if (ret) |
| 2112 | log_err("pthread_detach: %s", |
| 2113 | strerror(ret)); |
| 2114 | } else { |
| 2115 | pid_t pid; |
| 2116 | dprint(FD_PROCESS, "will fork\n"); |
| 2117 | pid = fork(); |
| 2118 | if (!pid) { |
| 2119 | int ret = fork_main(shm_id, i); |
| 2120 | |
| 2121 | _exit(ret); |
| 2122 | } else if (i == fio_debug_jobno) |
| 2123 | *fio_debug_jobp = pid; |
| 2124 | } |
| 2125 | dprint(FD_MUTEX, "wait on startup_mutex\n"); |
| 2126 | if (fio_mutex_down_timeout(startup_mutex, 10)) { |
| 2127 | log_err("fio: job startup hung? exiting.\n"); |
| 2128 | fio_terminate_threads(TERMINATE_ALL); |
| 2129 | fio_abort = 1; |
| 2130 | nr_started--; |
| 2131 | break; |
| 2132 | } |
| 2133 | dprint(FD_MUTEX, "done waiting on startup_mutex\n"); |
| 2134 | } |
| 2135 | |
| 2136 | /* |
| 2137 | * Wait for the started threads to transition to |
| 2138 | * TD_INITIALIZED. |
| 2139 | */ |
| 2140 | fio_gettime(&this_start, NULL); |
| 2141 | left = this_jobs; |
| 2142 | while (left && !fio_abort) { |
| 2143 | if (mtime_since_now(&this_start) > JOB_START_TIMEOUT) |
| 2144 | break; |
| 2145 | |
| 2146 | do_usleep(100000); |
| 2147 | |
| 2148 | for (i = 0; i < this_jobs; i++) { |
| 2149 | td = map[i]; |
| 2150 | if (!td) |
| 2151 | continue; |
| 2152 | if (td->runstate == TD_INITIALIZED) { |
| 2153 | map[i] = NULL; |
| 2154 | left--; |
| 2155 | } else if (td->runstate >= TD_EXITED) { |
| 2156 | map[i] = NULL; |
| 2157 | left--; |
| 2158 | todo--; |
| 2159 | nr_running++; /* work-around... */ |
| 2160 | } |
| 2161 | } |
| 2162 | } |
| 2163 | |
| 2164 | if (left) { |
| 2165 | log_err("fio: %d job%s failed to start\n", left, |
| 2166 | left > 1 ? "s" : ""); |
| 2167 | for (i = 0; i < this_jobs; i++) { |
| 2168 | td = map[i]; |
| 2169 | if (!td) |
| 2170 | continue; |
| 2171 | kill(td->pid, SIGTERM); |
| 2172 | } |
| 2173 | break; |
| 2174 | } |
| 2175 | |
| 2176 | /* |
| 2177 | * start created threads (TD_INITIALIZED -> TD_RUNNING). |
| 2178 | */ |
| 2179 | for_each_td(td, i) { |
| 2180 | if (td->runstate != TD_INITIALIZED) |
| 2181 | continue; |
| 2182 | |
| 2183 | if (in_ramp_time(td)) |
| 2184 | td_set_runstate(td, TD_RAMP); |
| 2185 | else |
| 2186 | td_set_runstate(td, TD_RUNNING); |
| 2187 | nr_running++; |
| 2188 | nr_started--; |
| 2189 | m_rate += ddir_rw_sum(td->o.ratemin); |
| 2190 | t_rate += ddir_rw_sum(td->o.rate); |
| 2191 | todo--; |
| 2192 | fio_mutex_up(td->mutex); |
| 2193 | } |
| 2194 | |
| 2195 | reap_threads(&nr_running, &t_rate, &m_rate); |
| 2196 | |
| 2197 | if (todo) |
| 2198 | do_usleep(100000); |
| 2199 | } |
| 2200 | |
| 2201 | while (nr_running) { |
| 2202 | reap_threads(&nr_running, &t_rate, &m_rate); |
| 2203 | do_usleep(10000); |
| 2204 | } |
| 2205 | |
| 2206 | fio_idle_prof_stop(); |
| 2207 | |
| 2208 | update_io_ticks(); |
| 2209 | } |
| 2210 | |
| 2211 | static void wait_for_helper_thread_exit(void) |
| 2212 | { |
| 2213 | void *ret; |
| 2214 | |
| 2215 | helper_exit = 1; |
| 2216 | pthread_cond_signal(&helper_cond); |
| 2217 | pthread_join(helper_thread, &ret); |
| 2218 | } |
| 2219 | |
| 2220 | static void free_disk_util(void) |
| 2221 | { |
| 2222 | disk_util_prune_entries(); |
| 2223 | |
| 2224 | pthread_cond_destroy(&helper_cond); |
| 2225 | } |
| 2226 | |
| 2227 | static void *helper_thread_main(void *data) |
| 2228 | { |
| 2229 | int ret = 0; |
| 2230 | |
| 2231 | fio_mutex_up(startup_mutex); |
| 2232 | |
| 2233 | while (!ret) { |
| 2234 | uint64_t sec = DISK_UTIL_MSEC / 1000; |
| 2235 | uint64_t nsec = (DISK_UTIL_MSEC % 1000) * 1000000; |
| 2236 | struct timespec ts; |
| 2237 | struct timeval tv; |
| 2238 | |
| 2239 | gettimeofday(&tv, NULL); |
| 2240 | ts.tv_sec = tv.tv_sec + sec; |
| 2241 | ts.tv_nsec = (tv.tv_usec * 1000) + nsec; |
| 2242 | |
| 2243 | if (ts.tv_nsec >= 1000000000ULL) { |
| 2244 | ts.tv_nsec -= 1000000000ULL; |
| 2245 | ts.tv_sec++; |
| 2246 | } |
| 2247 | |
| 2248 | pthread_cond_timedwait(&helper_cond, &helper_lock, &ts); |
| 2249 | |
| 2250 | ret = update_io_ticks(); |
| 2251 | |
| 2252 | if (helper_do_stat) { |
| 2253 | helper_do_stat = 0; |
| 2254 | __show_running_run_stats(); |
| 2255 | } |
| 2256 | |
| 2257 | if (!is_backend) |
| 2258 | print_thread_status(); |
| 2259 | } |
| 2260 | |
| 2261 | return NULL; |
| 2262 | } |
| 2263 | |
| 2264 | static int create_helper_thread(void) |
| 2265 | { |
| 2266 | int ret; |
| 2267 | |
| 2268 | setup_disk_util(); |
| 2269 | |
| 2270 | pthread_cond_init(&helper_cond, NULL); |
| 2271 | pthread_mutex_init(&helper_lock, NULL); |
| 2272 | |
| 2273 | ret = pthread_create(&helper_thread, NULL, helper_thread_main, NULL); |
| 2274 | if (ret) { |
| 2275 | log_err("Can't create helper thread: %s\n", strerror(ret)); |
| 2276 | return 1; |
| 2277 | } |
| 2278 | |
| 2279 | dprint(FD_MUTEX, "wait on startup_mutex\n"); |
| 2280 | fio_mutex_down(startup_mutex); |
| 2281 | dprint(FD_MUTEX, "done waiting on startup_mutex\n"); |
| 2282 | return 0; |
| 2283 | } |
| 2284 | |
| 2285 | int fio_backend(void) |
| 2286 | { |
| 2287 | struct thread_data *td; |
| 2288 | int i; |
| 2289 | |
| 2290 | if (exec_profile) { |
| 2291 | if (load_profile(exec_profile)) |
| 2292 | return 1; |
| 2293 | free(exec_profile); |
| 2294 | exec_profile = NULL; |
| 2295 | } |
| 2296 | if (!thread_number) |
| 2297 | return 0; |
| 2298 | |
| 2299 | if (write_bw_log) { |
| 2300 | struct log_params p = { |
| 2301 | .log_type = IO_LOG_TYPE_BW, |
| 2302 | }; |
| 2303 | |
| 2304 | setup_log(&agg_io_log[DDIR_READ], &p, "agg-read_bw.log"); |
| 2305 | setup_log(&agg_io_log[DDIR_WRITE], &p, "agg-write_bw.log"); |
| 2306 | setup_log(&agg_io_log[DDIR_TRIM], &p, "agg-trim_bw.log"); |
| 2307 | } |
| 2308 | |
| 2309 | startup_mutex = fio_mutex_init(FIO_MUTEX_LOCKED); |
| 2310 | if (startup_mutex == NULL) |
| 2311 | return 1; |
| 2312 | |
| 2313 | set_genesis_time(); |
| 2314 | stat_init(); |
| 2315 | create_helper_thread(); |
| 2316 | |
| 2317 | cgroup_list = smalloc(sizeof(*cgroup_list)); |
| 2318 | INIT_FLIST_HEAD(cgroup_list); |
| 2319 | |
| 2320 | run_threads(); |
| 2321 | |
| 2322 | wait_for_helper_thread_exit(); |
| 2323 | |
| 2324 | if (!fio_abort) { |
| 2325 | __show_run_stats(); |
| 2326 | if (write_bw_log) { |
| 2327 | for (i = 0; i < DDIR_RWDIR_CNT; i++) { |
| 2328 | struct io_log *log = agg_io_log[i]; |
| 2329 | |
| 2330 | flush_log(log, 0); |
| 2331 | free_log(log); |
| 2332 | } |
| 2333 | } |
| 2334 | } |
| 2335 | |
| 2336 | for_each_td(td, i) { |
| 2337 | fio_options_free(td); |
| 2338 | if (td->rusage_sem) { |
| 2339 | fio_mutex_remove(td->rusage_sem); |
| 2340 | td->rusage_sem = NULL; |
| 2341 | } |
| 2342 | } |
| 2343 | |
| 2344 | free_disk_util(); |
| 2345 | cgroup_kill(cgroup_list); |
| 2346 | sfree(cgroup_list); |
| 2347 | sfree(cgroup_mnt); |
| 2348 | |
| 2349 | fio_mutex_remove(startup_mutex); |
| 2350 | stat_exit(); |
| 2351 | return exit_value; |
| 2352 | } |