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  * 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 <signal.h>
  28 #include <time.h>
  29 #include <assert.h>
  30 #include <sys/stat.h>
  31 #include <sys/wait.h>
  32 #include <sys/ipc.h>
  33 #include <sys/shm.h>
  34 #include <sys/ioctl.h>
  35 #include <sys/mman.h>
  36
  37 #include "fio.h"
  38 #include "os.h"
  39
  40 #define MASK    (4095)
  41
  42 #define ALIGN(buf)      (char *) (((unsigned long) (buf) + MASK) & ~(MASK))
  43
  44 int groupid = 0;
  45 int thread_number = 0;
  46 int shm_id = 0;
  47 int temp_stall_ts;
  48
  49 static volatile int startup_sem;
  50
  51 #define TERMINATE_ALL           (-1)
  52 #define JOB_START_TIMEOUT       (5 * 1000)
  53
  54 static void terminate_threads(int group_id)
  55 {
  56         struct thread_data *td;
  57         int i;
  58
  59         for_each_td(td, i) {
  60                 if (group_id == TERMINATE_ALL || groupid == td->groupid) {
  61                         td->terminate = 1;
  62                         td->start_delay = 0;
  63                 }
  64         }
  65 }
  66
  67 static void sig_handler(int sig)
  68 {
  69         switch (sig) {
  70                 case SIGALRM:
  71                         update_io_ticks();
  72                         disk_util_timer_arm();
  73                         print_thread_status();
  74                         break;
  75                 default:
  76                         printf("\nfio: terminating on signal\n");
  77                         fflush(stdout);
  78                         terminate_threads(TERMINATE_ALL);
  79                         break;
  80         }
  81 }
  82
  83 /*
  84  * Check if we are above the minimum rate given.
  85  */
  86 static int check_min_rate(struct thread_data *td, struct timeval *now)
  87 {
  88         unsigned long spent;
  89         unsigned long rate;
  90         int ddir = td->ddir;
  91
  92         /*
  93          * allow a 2 second settle period in the beginning
  94          */
  95         if (mtime_since(&td->start, now) < 2000)
  96                 return 0;
  97
  98         /*
  99          * if rate blocks is set, sample is running
 100          */
 101         if (td->rate_bytes) {
 102                 spent = mtime_since(&td->lastrate, now);
 103                 if (spent < td->ratecycle)
 104                         return 0;
 105
 106                 rate = (td->this_io_bytes[ddir] - td->rate_bytes) / spent;
 107                 if (rate < td->ratemin) {
 108                         fprintf(f_out, "%s: min rate %u not met, got %luKiB/sec\n", td->name, td->ratemin, rate);
 109                         return 1;
 110                 }
 111         }
 112
 113         td->rate_bytes = td->this_io_bytes[ddir];
 114         memcpy(&td->lastrate, now, sizeof(*now));
 115         return 0;
 116 }
 117
 118 static inline int runtime_exceeded(struct thread_data *td, struct timeval *t)
 119 {
 120         if (!td->timeout)
 121                 return 0;
 122         if (mtime_since(&td->epoch, t) >= td->timeout * 1000)
 123                 return 1;
 124
 125         return 0;
 126 }
 127
 128 static inline void td_set_runstate(struct thread_data *td, int runstate)
 129 {
 130         td->runstate = runstate;
 131 }
 132
 133 static struct fio_file *get_next_file(struct thread_data *td)
 134 {
 135         unsigned int old_next_file = td->next_file;
 136         struct fio_file *f;
 137
 138         do {
 139                 f = &td->files[td->next_file];
 140
 141                 td->next_file++;
 142                 if (td->next_file >= td->nr_files)
 143                         td->next_file = 0;
 144
 145                 if (f->fd != -1)
 146                         break;
 147
 148                 f = NULL;
 149         } while (td->next_file != old_next_file);
 150
 151         return f;
 152 }
 153
 154 /*
 155  * When job exits, we can cancel the in-flight IO if we are using async
 156  * io. Attempt to do so.
 157  */
 158 static void cleanup_pending_aio(struct thread_data *td)
 159 {
 160         struct timespec ts = { .tv_sec = 0, .tv_nsec = 0};
 161         struct list_head *entry, *n;
 162         struct io_completion_data icd;
 163         struct io_u *io_u;
 164         int r;
 165
 166         /*
 167          * get immediately available events, if any
 168          */
 169         r = td_io_getevents(td, 0, td->cur_depth, &ts);
 170         if (r > 0) {
 171                 icd.nr = r;
 172                 ios_completed(td, &icd);
 173         }
 174
 175         /*
 176          * now cancel remaining active events
 177          */
 178         if (td->io_ops->cancel) {
 179                 list_for_each_safe(entry, n, &td->io_u_busylist) {
 180                         io_u = list_entry(entry, struct io_u, list);
 181
 182                         r = td->io_ops->cancel(td, io_u);
 183                         if (!r)
 184                                 put_io_u(td, io_u);
 185                 }
 186         }
 187
 188         if (td->cur_depth) {
 189                 r = td_io_getevents(td, td->cur_depth, td->cur_depth, NULL);
 190                 if (r > 0) {
 191                         icd.nr = r;
 192                         ios_completed(td, &icd);
 193                 }
 194         }
 195 }
 196
 197 /*
 198  * Helper to handle the final sync of a file. Works just like the normal
 199  * io path, just does everything sync.
 200  */
 201 static int fio_io_sync(struct thread_data *td, struct fio_file *f)
 202 {
 203         struct io_u *io_u = __get_io_u(td);
 204         struct io_completion_data icd;
 205         int ret;
 206
 207         if (!io_u)
 208                 return 1;
 209
 210         io_u->ddir = DDIR_SYNC;
 211         io_u->file = f;
 212
 213         if (td_io_prep(td, io_u)) {
 214                 put_io_u(td, io_u);
 215                 return 1;
 216         }
 217
 218         ret = td_io_queue(td, io_u);
 219         if (ret) {
 220                 td_verror(td, io_u->error);
 221                 put_io_u(td, io_u);
 222                 return 1;
 223         }
 224
 225         ret = td_io_getevents(td, 1, td->cur_depth, NULL);
 226         if (ret < 0) {
 227                 td_verror(td, ret);
 228                 return 1;
 229         }
 230
 231         icd.nr = ret;
 232         ios_completed(td, &icd);
 233         if (icd.error) {
 234                 td_verror(td, icd.error);
 235                 return 1;
 236         }
 237
 238         return 0;
 239 }
 240
 241 /*
 242  * The main verify engine. Runs over the writes we previusly submitted,
 243  * reads the blocks back in, and checks the crc/md5 of the data.
 244  */
 245 static void do_verify(struct thread_data *td)
 246 {
 247         struct io_u *io_u, *v_io_u = NULL;
 248         struct io_completion_data icd;
 249         struct fio_file *f;
 250         int ret, i;
 251
 252         /*
 253          * sync io first and invalidate cache, to make sure we really
 254          * read from disk.
 255          */
 256         for_each_file(td, f, i) {
 257                 fio_io_sync(td, f);
 258                 file_invalidate_cache(td, f);
 259         }
 260
 261         td_set_runstate(td, TD_VERIFYING);
 262
 263         do {
 264                 if (td->terminate)
 265                         break;
 266
 267                 io_u = __get_io_u(td);
 268                 if (!io_u)
 269                         break;
 270
 271                 if (runtime_exceeded(td, &io_u->start_time)) {
 272                         put_io_u(td, io_u);
 273                         break;
 274                 }
 275
 276                 if (get_next_verify(td, io_u)) {
 277                         put_io_u(td, io_u);
 278                         break;
 279                 }
 280
 281                 f = get_next_file(td);
 282                 if (!f)
 283                         break;
 284
 285                 io_u->file = f;
 286
 287                 if (td_io_prep(td, io_u)) {
 288                         put_io_u(td, io_u);
 289                         break;
 290                 }
 291
 292                 ret = td_io_queue(td, io_u);
 293                 if (ret) {
 294                         td_verror(td, io_u->error);
 295                         put_io_u(td, io_u);
 296                         break;
 297                 }
 298
 299                 /*
 300                  * we have one pending to verify, do that while
 301                  * we are doing io on the next one
 302                  */
 303                 if (do_io_u_verify(td, &v_io_u))
 304                         break;
 305
 306                 ret = td_io_getevents(td, 1, 1, NULL);
 307                 if (ret != 1) {
 308                         if (ret < 0)
 309                                 td_verror(td, ret);
 310                         break;
 311                 }
 312
 313                 v_io_u = td->io_ops->event(td, 0);
 314                 icd.nr = 1;
 315                 icd.error = 0;
 316                 fio_gettime(&icd.time, NULL);
 317                 io_completed(td, v_io_u, &icd);
 318
 319                 if (icd.error) {
 320                         td_verror(td, icd.error);
 321                         put_io_u(td, v_io_u);
 322                         v_io_u = NULL;
 323                         break;
 324                 }
 325
 326                 /*
 327                  * if we can't submit more io, we need to verify now
 328                  */
 329                 if (queue_full(td) && do_io_u_verify(td, &v_io_u))
 330                         break;
 331
 332         } while (1);
 333
 334         do_io_u_verify(td, &v_io_u);
 335
 336         if (td->cur_depth)
 337                 cleanup_pending_aio(td);
 338
 339         td_set_runstate(td, TD_RUNNING);
 340 }
 341
 342 /*
 343  * Not really an io thread, all it does is burn CPU cycles in the specified
 344  * manner.
 345  */
 346 static void do_cpuio(struct thread_data *td)
 347 {
 348         struct timeval e;
 349         int split = 100 / td->cpuload;
 350         int i = 0;
 351
 352         while (!td->terminate) {
 353                 fio_gettime(&e, NULL);
 354
 355                 if (runtime_exceeded(td, &e))
 356                         break;
 357
 358                 if (!(i % split))
 359                         __usec_sleep(10000);
 360                 else
 361                         usec_sleep(td, 10000);
 362
 363                 i++;
 364         }
 365 }
 366
 367 /*
 368  * Main IO worker function. It retrieves io_u's to process and queues
 369  * and reaps them, checking for rate and errors along the way.
 370  */
 371 static void do_io(struct thread_data *td)
 372 {
 373         struct io_completion_data icd;
 374         struct timeval s;
 375         unsigned long usec;
 376         struct fio_file *f;
 377         int i, ret = 0;
 378
 379         td_set_runstate(td, TD_RUNNING);
 380
 381         while (td->this_io_bytes[td->ddir] < td->io_size) {
 382                 struct timespec *timeout;
 383                 int min_evts = 0;
 384                 struct io_u *io_u;
 385
 386                 if (td->terminate)
 387                         break;
 388
 389                 f = get_next_file(td);
 390                 if (!f)
 391                         break;
 392
 393                 io_u = get_io_u(td, f);
 394                 if (!io_u)
 395                         break;
 396
 397                 memcpy(&s, &io_u->start_time, sizeof(s));
 398
 399                 ret = td_io_queue(td, io_u);
 400                 if (ret) {
 401                         td_verror(td, io_u->error);
 402                         put_io_u(td, io_u);
 403                         break;
 404                 }
 405
 406                 add_slat_sample(td, io_u->ddir, mtime_since(&io_u->start_time, &io_u->issue_time));
 407
 408                 if (td->cur_depth < td->iodepth) {
 409                         struct timespec ts = { .tv_sec = 0, .tv_nsec = 0};
 410
 411                         timeout = &ts;
 412                         min_evts = 0;
 413                 } else {
 414                         timeout = NULL;
 415                         min_evts = 1;
 416                 }
 417
 418                 ret = td_io_getevents(td, min_evts, td->cur_depth, timeout);
 419                 if (ret < 0) {
 420                         td_verror(td, ret);
 421                         break;
 422                 } else if (!ret)
 423                         continue;
 424
 425                 icd.nr = ret;
 426                 ios_completed(td, &icd);
 427                 if (icd.error) {
 428                         td_verror(td, icd.error);
 429                         break;
 430                 }
 431
 432                 /*
 433                  * the rate is batched for now, it should work for batches
 434                  * of completions except the very first one which may look
 435                  * a little bursty
 436                  */
 437                 usec = utime_since(&s, &icd.time);
 438
 439                 rate_throttle(td, usec, icd.bytes_done[td->ddir], td->ddir);
 440
 441                 if (check_min_rate(td, &icd.time)) {
 442                         if (exitall_on_terminate)
 443                                 terminate_threads(td->groupid);
 444                         td_verror(td, ENOMEM);
 445                         break;
 446                 }
 447
 448                 if (runtime_exceeded(td, &icd.time))
 449                         break;
 450
 451                 if (td->thinktime) {
 452                         unsigned long long b;
 453
 454                         b = td->io_blocks[0] + td->io_blocks[1];
 455                         if (!(b % td->thinktime_blocks))
 456                                 usec_sleep(td, td->thinktime);
 457                 }
 458         }
 459
 460         if (!td->error) {
 461                 if (td->cur_depth)
 462                         cleanup_pending_aio(td);
 463
 464                 if (should_fsync(td) && td->end_fsync) {
 465                         td_set_runstate(td, TD_FSYNCING);
 466                         for_each_file(td, f, i)
 467                                 fio_io_sync(td, f);
 468                 }
 469         }
 470 }
 471
 472 static void cleanup_io_u(struct thread_data *td)
 473 {
 474         struct list_head *entry, *n;
 475         struct io_u *io_u;
 476
 477         list_for_each_safe(entry, n, &td->io_u_freelist) {
 478                 io_u = list_entry(entry, struct io_u, list);
 479
 480                 list_del(&io_u->list);
 481                 free(io_u);
 482         }
 483
 484         free_io_mem(td);
 485 }
 486
 487 /*
 488  * "randomly" fill the buffer contents
 489  */
 490 static void fill_rand_buf(struct io_u *io_u, int max_bs)
 491 {
 492         int *ptr = io_u->buf;
 493
 494         while ((void *) ptr - io_u->buf < max_bs) {
 495                 *ptr = rand() * 0x9e370001;
 496                 ptr++;
 497         }
 498 }
 499
 500 static int init_io_u(struct thread_data *td)
 501 {
 502         struct io_u *io_u;
 503         unsigned int max_bs;
 504         int i, max_units;
 505         char *p;
 506
 507         if (td->io_ops->flags & FIO_CPUIO)
 508                 return 0;
 509
 510         if (td->io_ops->flags & FIO_SYNCIO)
 511                 max_units = 1;
 512         else
 513                 max_units = td->iodepth;
 514
 515         max_bs = max(td->max_bs[DDIR_READ], td->max_bs[DDIR_WRITE]);
 516         td->orig_buffer_size = max_bs * max_units;
 517
 518         if (td->mem_type == MEM_SHMHUGE || td->mem_type == MEM_MMAPHUGE)
 519                 td->orig_buffer_size = (td->orig_buffer_size + td->hugepage_size - 1) & ~(td->hugepage_size - 1);
 520         else
 521                 td->orig_buffer_size += MASK;
 522
 523         if (allocate_io_mem(td))
 524                 return 1;
 525
 526         p = ALIGN(td->orig_buffer);
 527         for (i = 0; i < max_units; i++) {
 528                 io_u = malloc(sizeof(*io_u));
 529                 memset(io_u, 0, sizeof(*io_u));
 530                 INIT_LIST_HEAD(&io_u->list);
 531
 532                 io_u->buf = p + max_bs * i;
 533                 if (td_write(td) || td_rw(td))
 534                         fill_rand_buf(io_u, max_bs);
 535
 536                 io_u->index = i;
 537                 list_add(&io_u->list, &td->io_u_freelist);
 538         }
 539
 540         return 0;
 541 }
 542
 543 static int switch_ioscheduler(struct thread_data *td)
 544 {
 545         char tmp[256], tmp2[128];
 546         FILE *f;
 547         int ret;
 548
 549         if (td->io_ops->flags & FIO_CPUIO)
 550                 return 0;
 551
 552         sprintf(tmp, "%s/queue/scheduler", td->sysfs_root);
 553
 554         f = fopen(tmp, "r+");
 555         if (!f) {
 556                 td_verror(td, errno);
 557                 return 1;
 558         }
 559
 560         /*
 561          * Set io scheduler.
 562          */
 563         ret = fwrite(td->ioscheduler, strlen(td->ioscheduler), 1, f);
 564         if (ferror(f) || ret != 1) {
 565                 td_verror(td, errno);
 566                 fclose(f);
 567                 return 1;
 568         }
 569
 570         rewind(f);
 571
 572         /*
 573          * Read back and check that the selected scheduler is now the default.
 574          */
 575         ret = fread(tmp, 1, sizeof(tmp), f);
 576         if (ferror(f) || ret < 0) {
 577                 td_verror(td, errno);
 578                 fclose(f);
 579                 return 1;
 580         }
 581
 582         sprintf(tmp2, "[%s]", td->ioscheduler);
 583         if (!strstr(tmp, tmp2)) {
 584                 log_err("fio: io scheduler %s not found\n", td->ioscheduler);
 585                 td_verror(td, EINVAL);
 586                 fclose(f);
 587                 return 1;
 588         }
 589
 590         fclose(f);
 591         return 0;
 592 }
 593
 594 static void clear_io_state(struct thread_data *td)
 595 {
 596         struct fio_file *f;
 597         int i;
 598
 599         td->stat_io_bytes[0] = td->stat_io_bytes[1] = 0;
 600         td->this_io_bytes[0] = td->this_io_bytes[1] = 0;
 601         td->zone_bytes = 0;
 602
 603         for_each_file(td, f, i) {
 604                 f->last_pos = 0;
 605                 if (td->io_ops->flags & FIO_SYNCIO)
 606                         lseek(f->fd, SEEK_SET, 0);
 607
 608                 if (f->file_map)
 609                         memset(f->file_map, 0, f->num_maps * sizeof(long));
 610         }
 611 }
 612
 613 /*
 614  * Entry point for the thread based jobs. The process based jobs end up
 615  * here as well, after a little setup.
 616  */
 617 static void *thread_main(void *data)
 618 {
 619         unsigned long long runtime[2];
 620         struct thread_data *td = data;
 621
 622         if (!td->use_thread)
 623                 setsid();
 624
 625         td->pid = getpid();
 626
 627         INIT_LIST_HEAD(&td->io_u_freelist);
 628         INIT_LIST_HEAD(&td->io_u_busylist);
 629         INIT_LIST_HEAD(&td->io_hist_list);
 630         INIT_LIST_HEAD(&td->io_log_list);
 631
 632         if (init_io_u(td))
 633                 goto err;
 634
 635         if (fio_setaffinity(td) == -1) {
 636                 td_verror(td, errno);
 637                 goto err;
 638         }
 639
 640         if (td_io_init(td))
 641                 goto err;
 642
 643         if (init_iolog(td))
 644                 goto err;
 645
 646         if (td->ioprio) {
 647                 if (ioprio_set(IOPRIO_WHO_PROCESS, 0, td->ioprio) == -1) {
 648                         td_verror(td, errno);
 649                         goto err;
 650                 }
 651         }
 652
 653         if (nice(td->nice) == -1) {
 654                 td_verror(td, errno);
 655                 goto err;
 656         }
 657
 658         if (init_random_state(td))
 659                 goto err;
 660
 661         if (td->ioscheduler && switch_ioscheduler(td))
 662                 goto err;
 663
 664         td_set_runstate(td, TD_INITIALIZED);
 665         fio_sem_up(&startup_sem);
 666         fio_sem_down(&td->mutex);
 667
 668         if (!td->create_serialize && setup_files(td))
 669                 goto err;
 670         if (open_files(td))
 671                 goto err;
 672
 673         if (td->exec_prerun)
 674                 system(td->exec_prerun);
 675
 676         fio_gettime(&td->epoch, NULL);
 677         getrusage(RUSAGE_SELF, &td->ru_start);
 678
 679         runtime[0] = runtime[1] = 0;
 680         while (td->loops--) {
 681                 fio_gettime(&td->start, NULL);
 682                 memcpy(&td->stat_sample_time, &td->start, sizeof(td->start));
 683
 684                 if (td->ratemin)
 685                         memcpy(&td->lastrate, &td->stat_sample_time, sizeof(td->lastrate));
 686
 687                 clear_io_state(td);
 688                 prune_io_piece_log(td);
 689
 690                 if (td->io_ops->flags & FIO_CPUIO)
 691                         do_cpuio(td);
 692                 else
 693                         do_io(td);
 694
 695                 runtime[td->ddir] += utime_since_now(&td->start);
 696                 if (td_rw(td) && td->io_bytes[td->ddir ^ 1])
 697                         runtime[td->ddir ^ 1] = runtime[td->ddir];
 698
 699                 if (td->error || td->terminate)
 700                         break;
 701
 702                 if (td->verify == VERIFY_NONE)
 703                         continue;
 704
 705                 clear_io_state(td);
 706                 fio_gettime(&td->start, NULL);
 707
 708                 do_verify(td);
 709
 710                 runtime[DDIR_READ] += utime_since_now(&td->start);
 711
 712                 if (td->error || td->terminate)
 713                         break;
 714         }
 715
 716         update_rusage_stat(td);
 717         fio_gettime(&td->end_time, NULL);
 718         td->runtime[0] = runtime[0] / 1000;
 719         td->runtime[1] = runtime[1] / 1000;
 720
 721         if (td->bw_log)
 722                 finish_log(td, td->bw_log, "bw");
 723         if (td->slat_log)
 724                 finish_log(td, td->slat_log, "slat");
 725         if (td->clat_log)
 726                 finish_log(td, td->clat_log, "clat");
 727         if (td->write_iolog_file)
 728                 write_iolog_close(td);
 729         if (td->exec_postrun)
 730                 system(td->exec_postrun);
 731
 732         if (exitall_on_terminate)
 733                 terminate_threads(td->groupid);
 734
 735 err:
 736         close_files(td);
 737         close_ioengine(td);
 738         cleanup_io_u(td);
 739         td_set_runstate(td, TD_EXITED);
 740         return NULL;
 741
 742 }
 743
 744 /*
 745  * We cannot pass the td data into a forked process, so attach the td and
 746  * pass it to the thread worker.
 747  */
 748 static void *fork_main(int shmid, int offset)
 749 {
 750         struct thread_data *td;
 751         void *data;
 752
 753         data = shmat(shmid, NULL, 0);
 754         if (data == (void *) -1) {
 755                 perror("shmat");
 756                 return NULL;
 757         }
 758
 759         td = data + offset * sizeof(struct thread_data);
 760         thread_main(td);
 761         shmdt(data);
 762         return NULL;
 763 }
 764
 765 /*
 766  * Run over the job map and reap the threads that have exited, if any.
 767  */
 768 static void reap_threads(int *nr_running, int *t_rate, int *m_rate)
 769 {
 770         struct thread_data *td;
 771         int i, cputhreads, pending;
 772
 773         /*
 774          * reap exited threads (TD_EXITED -> TD_REAPED)
 775          */
 776         pending = cputhreads = 0;
 777         for_each_td(td, i) {
 778                 /*
 779                  * ->io_ops is NULL for a thread that has closed its
 780                  * io engine
 781                  */
 782                 if (td->io_ops && td->io_ops->flags & FIO_CPUIO)
 783                         cputhreads++;
 784
 785                 if (td->runstate != TD_EXITED) {
 786                         if (td->runstate < TD_RUNNING)
 787                                 pending++;
 788
 789                         continue;
 790                 }
 791
 792                 td_set_runstate(td, TD_REAPED);
 793
 794                 if (td->use_thread) {
 795                         long ret;
 796
 797                         if (pthread_join(td->thread, (void *) &ret))
 798                                 perror("thread_join");
 799                 } else
 800                         waitpid(td->pid, NULL, 0);
 801
 802                 (*nr_running)--;
 803                 (*m_rate) -= td->ratemin;
 804                 (*t_rate) -= td->rate;
 805         }
 806
 807         if (*nr_running == cputhreads && !pending)
 808                 terminate_threads(TERMINATE_ALL);
 809 }
 810
 811 /*
 812  * Main function for kicking off and reaping jobs, as needed.
 813  */
 814 static void run_threads(void)
 815 {
 816         struct thread_data *td;
 817         unsigned long spent;
 818         int i, todo, nr_running, m_rate, t_rate, nr_started;
 819
 820         if (fio_pin_memory())
 821                 return;
 822
 823         if (!terse_output) {
 824                 printf("Starting %d thread%s\n", thread_number, thread_number > 1 ? "s" : "");
 825                 fflush(stdout);
 826         }
 827
 828         signal(SIGINT, sig_handler);
 829         signal(SIGALRM, sig_handler);
 830
 831         todo = thread_number;
 832         nr_running = 0;
 833         nr_started = 0;
 834         m_rate = t_rate = 0;
 835
 836         for_each_td(td, i) {
 837                 print_status_init(td->thread_number - 1);
 838
 839                 init_disk_util(td);
 840
 841                 if (!td->create_serialize)
 842                         continue;
 843
 844                 /*
 845                  * do file setup here so it happens sequentially,
 846                  * we don't want X number of threads getting their
 847                  * client data interspersed on disk
 848                  */
 849                 if (setup_files(td)) {
 850                         td_set_runstate(td, TD_REAPED);
 851                         todo--;
 852                 }
 853         }
 854
 855         while (todo) {
 856                 struct thread_data *map[MAX_JOBS];
 857                 struct timeval this_start;
 858                 int this_jobs = 0, left;
 859
 860                 /*
 861                  * create threads (TD_NOT_CREATED -> TD_CREATED)
 862                  */
 863                 for_each_td(td, i) {
 864                         if (td->runstate != TD_NOT_CREATED)
 865                                 continue;
 866
 867                         /*
 868                          * never got a chance to start, killed by other
 869                          * thread for some reason
 870                          */
 871                         if (td->terminate) {
 872                                 todo--;
 873                                 continue;
 874                         }
 875
 876                         if (td->start_delay) {
 877                                 spent = mtime_since_genesis();
 878
 879                                 if (td->start_delay * 1000 > spent)
 880                                         continue;
 881                         }
 882
 883                         if (td->stonewall && (nr_started || nr_running))
 884                                 break;
 885
 886                         /*
 887                          * Set state to created. Thread will transition
 888                          * to TD_INITIALIZED when it's done setting up.
 889                          */
 890                         td_set_runstate(td, TD_CREATED);
 891                         map[this_jobs++] = td;
 892                         fio_sem_init(&startup_sem, 1);
 893                         nr_started++;
 894
 895                         if (td->use_thread) {
 896                                 if (pthread_create(&td->thread, NULL, thread_main, td)) {
 897                                         perror("thread_create");
 898                                         nr_started--;
 899                                 }
 900                         } else {
 901                                 if (fork())
 902                                         fio_sem_down(&startup_sem);
 903                                 else {
 904                                         fork_main(shm_id, i);
 905                                         exit(0);
 906                                 }
 907                         }
 908                 }
 909
 910                 /*
 911                  * Wait for the started threads to transition to
 912                  * TD_INITIALIZED.
 913                  */
 914                 fio_gettime(&this_start, NULL);
 915                 left = this_jobs;
 916                 while (left) {
 917                         if (mtime_since_now(&this_start) > JOB_START_TIMEOUT)
 918                                 break;
 919
 920                         usleep(100000);
 921
 922                         for (i = 0; i < this_jobs; i++) {
 923                                 td = map[i];
 924                                 if (!td)
 925                                         continue;
 926                                 if (td->runstate == TD_INITIALIZED) {
 927                                         map[i] = NULL;
 928                                         left--;
 929                                 } else if (td->runstate >= TD_EXITED) {
 930                                         map[i] = NULL;
 931                                         left--;
 932                                         todo--;
 933                                         nr_running++; /* work-around... */
 934                                 }
 935                         }
 936                 }
 937
 938                 if (left) {
 939                         log_err("fio: %d jobs failed to start\n", left);
 940                         for (i = 0; i < this_jobs; i++) {
 941                                 td = map[i];
 942                                 if (!td)
 943                                         continue;
 944                                 kill(td->pid, SIGTERM);
 945                         }
 946                         break;
 947                 }
 948
 949                 /*
 950                  * start created threads (TD_INITIALIZED -> TD_RUNNING).
 951                  */
 952                 for_each_td(td, i) {
 953                         if (td->runstate != TD_INITIALIZED)
 954                                 continue;
 955
 956                         td_set_runstate(td, TD_RUNNING);
 957                         nr_running++;
 958                         nr_started--;
 959                         m_rate += td->ratemin;
 960                         t_rate += td->rate;
 961                         todo--;
 962                         fio_sem_up(&td->mutex);
 963                 }
 964
 965                 reap_threads(&nr_running, &t_rate, &m_rate);
 966
 967                 if (todo)
 968                         usleep(100000);
 969         }
 970
 971         while (nr_running) {
 972                 reap_threads(&nr_running, &t_rate, &m_rate);
 973                 usleep(10000);
 974         }
 975
 976         update_io_ticks();
 977         fio_unpin_memory();
 978 }
 979
 980 int main(int argc, char *argv[])
 981 {
 982         if (parse_options(argc, argv))
 983                 return 1;
 984
 985         if (!thread_number) {
 986                 log_err("Nothing to do\n");
 987                 return 1;
 988         }
 989
 990         disk_util_timer_arm();
 991
 992         run_threads();
 993         show_run_stats();
 994
 995         return 0;
 996 }