fio.c

   1 /*
   2  * fio - the flexible io tester
   3  *
   4  * Copyright (C) 2005 Jens Axboe <axboe@suse.de>
   5  * Copyright (C) 2006 Jens Axboe <axboe@kernel.dk>
   6  *
   7  *  This program is free software; you can redistribute it and/or modify
   8  *  it under the terms of the GNU General Public License as published by
   9  *  the Free Software Foundation; either version 2 of the License, or
  10  *  (at your option) any later version.
  11  *
  12  *  This program is distributed in the hope that it will be useful,
  13  *  but WITHOUT ANY WARRANTY; without even the implied warranty of
  14  *  MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
  15  *  GNU General Public License for more details.
  16  *
  17  *  You should have received a copy of the GNU General Public License
  18  *  along with this program; if not, write to the Free Software
  19  *  Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA  02111-1307  USA
  20  *
  21  */
  22 #include <unistd.h>
  23 #include <fcntl.h>
  24 #include <string.h>
  25 #include <signal.h>
  26 #include <time.h>
  27 #include <assert.h>
  28 #include <sys/stat.h>
  29 #include <sys/wait.h>
  30 #include <sys/ipc.h>
  31 #include <sys/shm.h>
  32 #include <sys/ioctl.h>
  33 #include <sys/mman.h>
  34
  35 #include "fio.h"
  36 #include "os.h"
  37
  38 #define MASK    (4095)
  39
  40 #define ALIGN(buf)      (char *) (((unsigned long) (buf) + MASK) & ~(MASK))
  41
  42 int groupid = 0;
  43 int thread_number = 0;
  44 int shm_id = 0;
  45 int temp_stall_ts;
  46 char *fio_inst_prefix = _INST_PREFIX;
  47
  48 #define should_fsync(td)        ((td_write(td) || td_rw(td)) && (!(td)->odirect || (td)->override_sync))
  49
  50 static volatile int startup_sem;
  51
  52 #define TERMINATE_ALL           (-1)
  53 #define JOB_START_TIMEOUT       (5 * 1000)
  54
  55 static void terminate_threads(int group_id)
  56 {
  57         struct thread_data *td;
  58         int i;
  59
  60         for_each_td(td, i) {
  61                 if (group_id == TERMINATE_ALL || groupid == td->groupid) {
  62                         td->terminate = 1;
  63                         td->start_delay = 0;
  64                 }
  65         }
  66 }
  67
  68 static void sig_handler(int sig)
  69 {
  70         switch (sig) {
  71                 case SIGALRM:
  72                         update_io_ticks();
  73                         disk_util_timer_arm();
  74                         print_thread_status();
  75                         break;
  76                 default:
  77                         printf("\nfio: terminating on signal\n");
  78                         fflush(stdout);
  79                         terminate_threads(TERMINATE_ALL);
  80                         break;
  81         }
  82 }
  83
  84 /*
  85  * Check if we are above the minimum rate given.
  86  */
  87 static int check_min_rate(struct thread_data *td, struct timeval *now)
  88 {
  89         unsigned long spent;
  90         unsigned long rate;
  91         int ddir = td->ddir;
  92
  93         /*
  94          * allow a 2 second settle period in the beginning
  95          */
  96         if (mtime_since(&td->start, now) < 2000)
  97                 return 0;
  98
  99         /*
 100          * if rate blocks is set, sample is running
 101          */
 102         if (td->rate_bytes) {
 103                 spent = mtime_since(&td->lastrate, now);
 104                 if (spent < td->ratecycle)
 105                         return 0;
 106
 107                 rate = (td->this_io_bytes[ddir] - td->rate_bytes) / spent;
 108                 if (rate < td->ratemin) {
 109                         fprintf(f_out, "%s: min rate %d not met, got %ldKiB/sec\n", td->name, td->ratemin, rate);
 110                         return 1;
 111                 }
 112         }
 113
 114         td->rate_bytes = td->this_io_bytes[ddir];
 115         memcpy(&td->lastrate, now, sizeof(*now));
 116         return 0;
 117 }
 118
 119 static inline int runtime_exceeded(struct thread_data *td, struct timeval *t)
 120 {
 121         if (!td->timeout)
 122                 return 0;
 123         if (mtime_since(&td->epoch, t) >= td->timeout * 1000)
 124                 return 1;
 125
 126         return 0;
 127 }
 128
 129 static inline void td_set_runstate(struct thread_data *td, int runstate)
 130 {
 131         td->runstate = runstate;
 132 }
 133
 134 static struct fio_file *get_next_file(struct thread_data *td)
 135 {
 136         unsigned int old_next_file = td->next_file;
 137         struct fio_file *f;
 138
 139         do {
 140                 f = &td->files[td->next_file];
 141
 142                 td->next_file++;
 143                 if (td->next_file >= td->nr_files)
 144                         td->next_file = 0;
 145
 146                 if (f->fd != -1)
 147                         break;
 148
 149                 f = NULL;
 150         } while (td->next_file != old_next_file);
 151
 152         return f;
 153 }
 154
 155 /*
 156  * When job exits, we can cancel the in-flight IO if we are using async
 157  * io. Attempt to do so.
 158  */
 159 static void cleanup_pending_aio(struct thread_data *td)
 160 {
 161         struct timespec ts = { .tv_sec = 0, .tv_nsec = 0};
 162         struct list_head *entry, *n;
 163         struct io_completion_data icd;
 164         struct io_u *io_u;
 165         int r;
 166
 167         /*
 168          * get immediately available events, if any
 169          */
 170         r = td_io_getevents(td, 0, td->cur_depth, &ts);
 171         if (r > 0) {
 172                 icd.nr = r;
 173                 ios_completed(td, &icd);
 174         }
 175
 176         /*
 177          * now cancel remaining active events
 178          */
 179         if (td->io_ops->cancel) {
 180                 list_for_each_safe(entry, n, &td->io_u_busylist) {
 181                         io_u = list_entry(entry, struct io_u, list);
 182
 183                         r = td->io_ops->cancel(td, io_u);
 184                         if (!r)
 185                                 put_io_u(td, io_u);
 186                 }
 187         }
 188
 189         if (td->cur_depth) {
 190                 r = td_io_getevents(td, td->cur_depth, td->cur_depth, NULL);
 191                 if (r > 0) {
 192                         icd.nr = r;
 193                         ios_completed(td, &icd);
 194                 }
 195         }
 196 }
 197
 198 /*
 199  * The main verify engine. Runs over the writes we previusly submitted,
 200  * reads the blocks back in, and checks the crc/md5 of the data.
 201  */
 202 void do_verify(struct thread_data *td)
 203 {
 204         struct timeval t;
 205         struct io_u *io_u, *v_io_u = NULL;
 206         struct io_completion_data icd;
 207         struct fio_file *f;
 208         int ret, i;
 209
 210         /*
 211          * sync io first and invalidate cache, to make sure we really
 212          * read from disk.
 213          */
 214         for_each_file(td, f, i) {
 215                 td_io_sync(td, f);
 216                 file_invalidate_cache(td, f);
 217         }
 218
 219         td_set_runstate(td, TD_VERIFYING);
 220
 221         do {
 222                 if (td->terminate)
 223                         break;
 224
 225                 gettimeofday(&t, NULL);
 226                 if (runtime_exceeded(td, &t))
 227                         break;
 228
 229                 io_u = __get_io_u(td);
 230                 if (!io_u)
 231                         break;
 232
 233                 if (get_next_verify(td, io_u)) {
 234                         put_io_u(td, io_u);
 235                         break;
 236                 }
 237
 238                 f = get_next_file(td);
 239                 if (!f)
 240                         break;
 241
 242                 io_u->file = f;
 243
 244                 if (td_io_prep(td, io_u)) {
 245                         put_io_u(td, io_u);
 246                         break;
 247                 }
 248
 249                 ret = td_io_queue(td, io_u);
 250                 if (ret) {
 251                         put_io_u(td, io_u);
 252                         td_verror(td, ret);
 253                         break;
 254                 }
 255
 256                 /*
 257                  * we have one pending to verify, do that while
 258                  * we are doing io on the next one
 259                  */
 260                 if (do_io_u_verify(td, &v_io_u))
 261                         break;
 262
 263                 ret = td_io_getevents(td, 1, 1, NULL);
 264                 if (ret != 1) {
 265                         if (ret < 0)
 266                                 td_verror(td, ret);
 267                         break;
 268                 }
 269
 270                 v_io_u = td->io_ops->event(td, 0);
 271                 icd.nr = 1;
 272                 icd.error = 0;
 273                 io_completed(td, v_io_u, &icd);
 274
 275                 if (icd.error) {
 276                         td_verror(td, icd.error);
 277                         put_io_u(td, v_io_u);
 278                         v_io_u = NULL;
 279                         break;
 280                 }
 281
 282                 /*
 283                  * if we can't submit more io, we need to verify now
 284                  */
 285                 if (queue_full(td) && do_io_u_verify(td, &v_io_u))
 286                         break;
 287
 288         } while (1);
 289
 290         do_io_u_verify(td, &v_io_u);
 291
 292         if (td->cur_depth)
 293                 cleanup_pending_aio(td);
 294
 295         td_set_runstate(td, TD_RUNNING);
 296 }
 297
 298 /*
 299  * Not really an io thread, all it does is burn CPU cycles in the specified
 300  * manner.
 301  */
 302 static void do_cpuio(struct thread_data *td)
 303 {
 304         struct timeval e;
 305         int split = 100 / td->cpuload;
 306         int i = 0;
 307
 308         while (!td->terminate) {
 309                 gettimeofday(&e, NULL);
 310
 311                 if (runtime_exceeded(td, &e))
 312                         break;
 313
 314                 if (!(i % split))
 315                         __usec_sleep(10000);
 316                 else
 317                         usec_sleep(td, 10000);
 318
 319                 i++;
 320         }
 321 }
 322
 323 /*
 324  * Main IO worker function. It retrieves io_u's to process and queues
 325  * and reaps them, checking for rate and errors along the way.
 326  */
 327 static void do_io(struct thread_data *td)
 328 {
 329         struct io_completion_data icd;
 330         struct timeval s, e;
 331         unsigned long usec;
 332         struct fio_file *f;
 333         int i, ret = 0;
 334
 335         td_set_runstate(td, TD_RUNNING);
 336
 337         while (td->this_io_bytes[td->ddir] < td->io_size) {
 338                 struct timespec ts = { .tv_sec = 0, .tv_nsec = 0};
 339                 struct timespec *timeout;
 340                 int min_evts = 0;
 341                 struct io_u *io_u;
 342
 343                 if (td->terminate)
 344                         break;
 345
 346                 f = get_next_file(td);
 347                 if (!f)
 348                         break;
 349
 350                 io_u = get_io_u(td, f);
 351                 if (!io_u)
 352                         break;
 353
 354                 memcpy(&s, &io_u->start_time, sizeof(s));
 355
 356                 ret = td_io_queue(td, io_u);
 357                 if (ret) {
 358                         put_io_u(td, io_u);
 359                         td_verror(td, ret);
 360                         break;
 361                 }
 362
 363                 add_slat_sample(td, io_u->ddir, mtime_since(&io_u->start_time, &io_u->issue_time));
 364
 365                 if (td->cur_depth < td->iodepth) {
 366                         timeout = &ts;
 367                         min_evts = 0;
 368                 } else {
 369                         timeout = NULL;
 370                         min_evts = 1;
 371                 }
 372
 373                 ret = td_io_getevents(td, min_evts, td->cur_depth, timeout);
 374                 if (ret < 0) {
 375                         td_verror(td, -ret);
 376                         break;
 377                 } else if (!ret)
 378                         continue;
 379
 380                 icd.nr = ret;
 381                 ios_completed(td, &icd);
 382                 if (icd.error) {
 383                         td_verror(td, icd.error);
 384                         break;
 385                 }
 386
 387                 /*
 388                  * the rate is batched for now, it should work for batches
 389                  * of completions except the very first one which may look
 390                  * a little bursty
 391                  */
 392                 gettimeofday(&e, NULL);
 393                 usec = utime_since(&s, &e);
 394
 395                 rate_throttle(td, usec, icd.bytes_done[td->ddir]);
 396
 397                 if (check_min_rate(td, &e)) {
 398                         if (rate_quit)
 399                                 terminate_threads(td->groupid);
 400                         td_verror(td, ENOMEM);
 401                         break;
 402                 }
 403
 404                 if (runtime_exceeded(td, &e))
 405                         break;
 406
 407                 if (td->thinktime)
 408                         usec_sleep(td, td->thinktime);
 409
 410                 if (should_fsync(td) && td->fsync_blocks &&
 411                     (td->io_blocks[DDIR_WRITE] % td->fsync_blocks) == 0)
 412                         td_io_sync(td, f);
 413         }
 414
 415         if (!ret) {
 416                 if (td->cur_depth)
 417                         cleanup_pending_aio(td);
 418
 419                 if (should_fsync(td) && td->end_fsync) {
 420                         td_set_runstate(td, TD_FSYNCING);
 421                         for_each_file(td, f, i)
 422                                 td_io_sync(td, f);
 423                 }
 424         }
 425 }
 426
 427 static void cleanup_io_u(struct thread_data *td)
 428 {
 429         struct list_head *entry, *n;
 430         struct io_u *io_u;
 431
 432         list_for_each_safe(entry, n, &td->io_u_freelist) {
 433                 io_u = list_entry(entry, struct io_u, list);
 434
 435                 list_del(&io_u->list);
 436                 free(io_u);
 437         }
 438
 439         free_io_mem(td);
 440 }
 441
 442 static int init_io_u(struct thread_data *td)
 443 {
 444         struct io_u *io_u;
 445         int i, max_units;
 446         char *p;
 447
 448         if (td->io_ops->flags & FIO_CPUIO)
 449                 return 0;
 450
 451         if (td->io_ops->flags & FIO_SYNCIO)
 452                 max_units = 1;
 453         else
 454                 max_units = td->iodepth;
 455
 456         td->orig_buffer_size = td->max_bs * max_units + MASK;
 457
 458         if (allocate_io_mem(td))
 459                 return 1;
 460
 461         p = ALIGN(td->orig_buffer);
 462         for (i = 0; i < max_units; i++) {
 463                 io_u = malloc(sizeof(*io_u));
 464                 memset(io_u, 0, sizeof(*io_u));
 465                 INIT_LIST_HEAD(&io_u->list);
 466
 467                 io_u->buf = p + td->max_bs * i;
 468                 io_u->index = i;
 469                 list_add(&io_u->list, &td->io_u_freelist);
 470         }
 471
 472         return 0;
 473 }
 474
 475 static int switch_ioscheduler(struct thread_data *td)
 476 {
 477         char tmp[256], tmp2[128];
 478         FILE *f;
 479         int ret;
 480
 481         sprintf(tmp, "%s/queue/scheduler", td->sysfs_root);
 482
 483         f = fopen(tmp, "r+");
 484         if (!f) {
 485                 td_verror(td, errno);
 486                 return 1;
 487         }
 488
 489         /*
 490          * Set io scheduler.
 491          */
 492         ret = fwrite(td->ioscheduler, strlen(td->ioscheduler), 1, f);
 493         if (ferror(f) || ret != 1) {
 494                 td_verror(td, errno);
 495                 fclose(f);
 496                 return 1;
 497         }
 498
 499         rewind(f);
 500
 501         /*
 502          * Read back and check that the selected scheduler is now the default.
 503          */
 504         ret = fread(tmp, 1, sizeof(tmp), f);
 505         if (ferror(f) || ret < 0) {
 506                 td_verror(td, errno);
 507                 fclose(f);
 508                 return 1;
 509         }
 510
 511         sprintf(tmp2, "[%s]", td->ioscheduler);
 512         if (!strstr(tmp, tmp2)) {
 513                 log_err("fio: io scheduler %s not found\n", td->ioscheduler);
 514                 td_verror(td, EINVAL);
 515                 fclose(f);
 516                 return 1;
 517         }
 518
 519         fclose(f);
 520         return 0;
 521 }
 522
 523 static void clear_io_state(struct thread_data *td)
 524 {
 525         struct fio_file *f;
 526         int i;
 527
 528         td->stat_io_bytes[0] = td->stat_io_bytes[1] = 0;
 529         td->this_io_bytes[0] = td->this_io_bytes[1] = 0;
 530         td->zone_bytes = 0;
 531
 532         for_each_file(td, f, i) {
 533                 f->last_pos = 0;
 534                 if (td->io_ops->flags & FIO_SYNCIO)
 535                         lseek(f->fd, SEEK_SET, 0);
 536
 537                 if (f->file_map)
 538                         memset(f->file_map, 0, f->num_maps * sizeof(long));
 539         }
 540 }
 541
 542 /*
 543  * Entry point for the thread based jobs. The process based jobs end up
 544  * here as well, after a little setup.
 545  */
 546 static void *thread_main(void *data)
 547 {
 548         struct thread_data *td = data;
 549
 550         if (!td->use_thread)
 551                 setsid();
 552
 553         td->pid = getpid();
 554
 555         INIT_LIST_HEAD(&td->io_u_freelist);
 556         INIT_LIST_HEAD(&td->io_u_busylist);
 557         INIT_LIST_HEAD(&td->io_hist_list);
 558         INIT_LIST_HEAD(&td->io_log_list);
 559
 560         if (init_io_u(td))
 561                 goto err;
 562
 563         if (fio_setaffinity(td) == -1) {
 564                 td_verror(td, errno);
 565                 goto err;
 566         }
 567
 568         if (td_io_init(td))
 569                 goto err;
 570
 571         if (init_iolog(td))
 572                 goto err;
 573
 574         if (td->ioprio) {
 575                 if (ioprio_set(IOPRIO_WHO_PROCESS, 0, td->ioprio) == -1) {
 576                         td_verror(td, errno);
 577                         goto err;
 578                 }
 579         }
 580
 581         if (nice(td->nice) == -1) {
 582                 td_verror(td, errno);
 583                 goto err;
 584         }
 585
 586         if (init_random_state(td))
 587                 goto err;
 588
 589         if (td->ioscheduler && switch_ioscheduler(td))
 590                 goto err;
 591
 592         td_set_runstate(td, TD_INITIALIZED);
 593         fio_sem_up(&startup_sem);
 594         fio_sem_down(&td->mutex);
 595
 596         if (!td->create_serialize && setup_files(td))
 597                 goto err;
 598
 599         gettimeofday(&td->epoch, NULL);
 600
 601         if (td->exec_prerun)
 602                 system(td->exec_prerun);
 603
 604         while (td->loops--) {
 605                 getrusage(RUSAGE_SELF, &td->ru_start);
 606                 gettimeofday(&td->start, NULL);
 607                 memcpy(&td->stat_sample_time, &td->start, sizeof(td->start));
 608
 609                 if (td->ratemin)
 610                         memcpy(&td->lastrate, &td->stat_sample_time, sizeof(td->lastrate));
 611
 612                 clear_io_state(td);
 613                 prune_io_piece_log(td);
 614
 615                 if (td->io_ops->flags & FIO_CPUIO)
 616                         do_cpuio(td);
 617                 else
 618                         do_io(td);
 619
 620                 td->runtime[td->ddir] += mtime_since_now(&td->start);
 621                 if (td_rw(td) && td->io_bytes[td->ddir ^ 1])
 622                         td->runtime[td->ddir ^ 1] = td->runtime[td->ddir];
 623
 624                 update_rusage_stat(td);
 625
 626                 if (td->error || td->terminate)
 627                         break;
 628
 629                 if (td->verify == VERIFY_NONE)
 630                         continue;
 631
 632                 clear_io_state(td);
 633                 gettimeofday(&td->start, NULL);
 634
 635                 do_verify(td);
 636
 637                 td->runtime[DDIR_READ] += mtime_since_now(&td->start);
 638
 639                 if (td->error || td->terminate)
 640                         break;
 641         }
 642
 643         if (td->bw_log)
 644                 finish_log(td, td->bw_log, "bw");
 645         if (td->slat_log)
 646                 finish_log(td, td->slat_log, "slat");
 647         if (td->clat_log)
 648                 finish_log(td, td->clat_log, "clat");
 649         if (td->write_iolog)
 650                 write_iolog_close(td);
 651         if (td->exec_postrun)
 652                 system(td->exec_postrun);
 653
 654         if (exitall_on_terminate)
 655                 terminate_threads(td->groupid);
 656
 657 err:
 658         close_files(td);
 659         close_ioengine(td);
 660         cleanup_io_u(td);
 661         td_set_runstate(td, TD_EXITED);
 662         return NULL;
 663
 664 }
 665
 666 /*
 667  * We cannot pass the td data into a forked process, so attach the td and
 668  * pass it to the thread worker.
 669  */
 670 static void *fork_main(int shmid, int offset)
 671 {
 672         struct thread_data *td;
 673         void *data;
 674
 675         data = shmat(shmid, NULL, 0);
 676         if (data == (void *) -1) {
 677                 perror("shmat");
 678                 return NULL;
 679         }
 680
 681         td = data + offset * sizeof(struct thread_data);
 682         thread_main(td);
 683         shmdt(data);
 684         return NULL;
 685 }
 686
 687 /*
 688  * Run over the job map and reap the threads that have exited, if any.
 689  */
 690 static void reap_threads(int *nr_running, int *t_rate, int *m_rate)
 691 {
 692         struct thread_data *td;
 693         int i, cputhreads;
 694
 695         /*
 696          * reap exited threads (TD_EXITED -> TD_REAPED)
 697          */
 698         cputhreads = 0;
 699         for_each_td(td, i) {
 700                 /*
 701                  * ->io_ops is NULL for a thread that has closed its
 702                  * io engine
 703                  */
 704                 if (td->io_ops && td->io_ops->flags & FIO_CPUIO)
 705                         cputhreads++;
 706
 707                 if (td->runstate != TD_EXITED)
 708                         continue;
 709
 710                 td_set_runstate(td, TD_REAPED);
 711
 712                 if (td->use_thread) {
 713                         long ret;
 714
 715                         if (pthread_join(td->thread, (void *) &ret))
 716                                 perror("thread_join");
 717                 } else
 718                         waitpid(td->pid, NULL, 0);
 719
 720                 (*nr_running)--;
 721                 (*m_rate) -= td->ratemin;
 722                 (*t_rate) -= td->rate;
 723         }
 724
 725         if (*nr_running == cputhreads)
 726                 terminate_threads(TERMINATE_ALL);
 727 }
 728
 729 /*
 730  * Main function for kicking off and reaping jobs, as needed.
 731  */
 732 static void run_threads(void)
 733 {
 734         struct thread_data *td;
 735         unsigned long spent;
 736         int i, todo, nr_running, m_rate, t_rate, nr_started;
 737
 738         if (fio_pin_memory())
 739                 return;
 740
 741         if (!terse_output) {
 742                 printf("Starting %d thread%s\n", thread_number, thread_number > 1 ? "s" : "");
 743                 fflush(stdout);
 744         }
 745
 746         signal(SIGINT, sig_handler);
 747         signal(SIGALRM, sig_handler);
 748
 749         todo = thread_number;
 750         nr_running = 0;
 751         nr_started = 0;
 752         m_rate = t_rate = 0;
 753
 754         for_each_td(td, i) {
 755                 print_status_init(td->thread_number - 1);
 756
 757                 init_disk_util(td);
 758
 759                 if (!td->create_serialize)
 760                         continue;
 761
 762                 /*
 763                  * do file setup here so it happens sequentially,
 764                  * we don't want X number of threads getting their
 765                  * client data interspersed on disk
 766                  */
 767                 if (setup_files(td)) {
 768                         td_set_runstate(td, TD_REAPED);
 769                         todo--;
 770                 }
 771         }
 772
 773         time_init();
 774
 775         while (todo) {
 776                 struct thread_data *map[MAX_JOBS];
 777                 struct timeval this_start;
 778                 int this_jobs = 0, left;
 779
 780                 /*
 781                  * create threads (TD_NOT_CREATED -> TD_CREATED)
 782                  */
 783                 for_each_td(td, i) {
 784                         if (td->runstate != TD_NOT_CREATED)
 785                                 continue;
 786
 787                         /*
 788                          * never got a chance to start, killed by other
 789                          * thread for some reason
 790                          */
 791                         if (td->terminate) {
 792                                 todo--;
 793                                 continue;
 794                         }
 795
 796                         if (td->start_delay) {
 797                                 spent = mtime_since_genesis();
 798
 799                                 if (td->start_delay * 1000 > spent)
 800                                         continue;
 801                         }
 802
 803                         if (td->stonewall && (nr_started || nr_running))
 804                                 break;
 805
 806                         /*
 807                          * Set state to created. Thread will transition
 808                          * to TD_INITIALIZED when it's done setting up.
 809                          */
 810                         td_set_runstate(td, TD_CREATED);
 811                         map[this_jobs++] = td;
 812                         fio_sem_init(&startup_sem, 1);
 813                         nr_started++;
 814
 815                         if (td->use_thread) {
 816                                 if (pthread_create(&td->thread, NULL, thread_main, td)) {
 817                                         perror("thread_create");
 818                                         nr_started--;
 819                                 }
 820                         } else {
 821                                 if (fork())
 822                                         fio_sem_down(&startup_sem);
 823                                 else {
 824                                         fork_main(shm_id, i);
 825                                         exit(0);
 826                                 }
 827                         }
 828                 }
 829
 830                 /*
 831                  * Wait for the started threads to transition to
 832                  * TD_INITIALIZED.
 833                  */
 834                 gettimeofday(&this_start, NULL);
 835                 left = this_jobs;
 836                 while (left) {
 837                         if (mtime_since_now(&this_start) > JOB_START_TIMEOUT)
 838                                 break;
 839
 840                         usleep(100000);
 841
 842                         for (i = 0; i < this_jobs; i++) {
 843                                 td = map[i];
 844                                 if (!td)
 845                                         continue;
 846                                 if (td->runstate == TD_INITIALIZED) {
 847                                         map[i] = NULL;
 848                                         left--;
 849                                 } else if (td->runstate >= TD_EXITED) {
 850                                         map[i] = NULL;
 851                                         left--;
 852                                         todo--;
 853                                         nr_running++; /* work-around... */
 854                                 }
 855                         }
 856                 }
 857
 858                 if (left) {
 859                         log_err("fio: %d jobs failed to start\n", left);
 860                         for (i = 0; i < this_jobs; i++) {
 861                                 td = map[i];
 862                                 if (!td)
 863                                         continue;
 864                                 kill(td->pid, SIGTERM);
 865                         }
 866                         break;
 867                 }
 868
 869                 /*
 870                  * start created threads (TD_INITIALIZED -> TD_RUNNING).
 871                  */
 872                 for_each_td(td, i) {
 873                         if (td->runstate != TD_INITIALIZED)
 874                                 continue;
 875
 876                         td_set_runstate(td, TD_RUNNING);
 877                         nr_running++;
 878                         nr_started--;
 879                         m_rate += td->ratemin;
 880                         t_rate += td->rate;
 881                         todo--;
 882                         fio_sem_up(&td->mutex);
 883                 }
 884
 885                 reap_threads(&nr_running, &t_rate, &m_rate);
 886
 887                 if (todo)
 888                         usleep(100000);
 889         }
 890
 891         while (nr_running) {
 892                 reap_threads(&nr_running, &t_rate, &m_rate);
 893                 usleep(10000);
 894         }
 895
 896         update_io_ticks();
 897         fio_unpin_memory();
 898 }
 899
 900 int main(int argc, char *argv[])
 901 {
 902         if (parse_options(argc, argv))
 903                 return 1;
 904
 905         if (!thread_number) {
 906                 log_err("Nothing to do\n");
 907                 return 1;
 908         }
 909
 910         disk_util_timer_arm();
 911
 912         run_threads();
 913         show_run_stats();
 914
 915         return 0;
 916 }