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 ts = { .tv_sec = 0, .tv_nsec = 0};
 383                 struct timespec *timeout;
 384                 int min_evts = 0;
 385                 struct io_u *io_u;
 386
 387                 if (td->terminate)
 388                         break;
 389
 390                 f = get_next_file(td);
 391                 if (!f)
 392                         break;
 393
 394                 io_u = get_io_u(td, f);
 395                 if (!io_u)
 396                         break;
 397
 398                 memcpy(&s, &io_u->start_time, sizeof(s));
 399
 400                 ret = td_io_queue(td, io_u);
 401                 if (ret) {
 402                         td_verror(td, io_u->error);
 403                         put_io_u(td, io_u);
 404                         break;
 405                 }
 406
 407                 add_slat_sample(td, io_u->ddir, mtime_since(&io_u->start_time, &io_u->issue_time));
 408
 409                 if (td->cur_depth < td->iodepth) {
 410                         timeout = &ts;
 411                         min_evts = 0;
 412                 } else {
 413                         timeout = NULL;
 414                         min_evts = 1;
 415                 }
 416
 417                 ret = td_io_getevents(td, min_evts, td->cur_depth, timeout);
 418                 if (ret < 0) {
 419                         td_verror(td, ret);
 420                         break;
 421                 } else if (!ret)
 422                         continue;
 423
 424                 icd.nr = ret;
 425                 ios_completed(td, &icd);
 426                 if (icd.error) {
 427                         td_verror(td, icd.error);
 428                         break;
 429                 }
 430
 431                 /*
 432                  * the rate is batched for now, it should work for batches
 433                  * of completions except the very first one which may look
 434                  * a little bursty
 435                  */
 436                 usec = utime_since(&s, &icd.time);
 437
 438                 rate_throttle(td, usec, icd.bytes_done[td->ddir], td->ddir);
 439
 440                 if (check_min_rate(td, &icd.time)) {
 441                         if (exitall_on_terminate)
 442                                 terminate_threads(td->groupid);
 443                         td_verror(td, ENOMEM);
 444                         break;
 445                 }
 446
 447                 if (runtime_exceeded(td, &icd.time))
 448                         break;
 449
 450                 if (td->thinktime) {
 451                         unsigned long long b;
 452
 453                         b = td->io_blocks[0] + td->io_blocks[1];
 454                         if (!(td->thinktime_blocks % b))
 455                                 usec_sleep(td, td->thinktime);
 456                 }
 457         }
 458
 459         if (!td->error) {
 460                 if (td->cur_depth)
 461                         cleanup_pending_aio(td);
 462
 463                 if (should_fsync(td) && td->end_fsync) {
 464                         td_set_runstate(td, TD_FSYNCING);
 465                         for_each_file(td, f, i)
 466                                 fio_io_sync(td, f);
 467                 }
 468         }
 469 }
 470
 471 static void cleanup_io_u(struct thread_data *td)
 472 {
 473         struct list_head *entry, *n;
 474         struct io_u *io_u;
 475
 476         list_for_each_safe(entry, n, &td->io_u_freelist) {
 477                 io_u = list_entry(entry, struct io_u, list);
 478
 479                 list_del(&io_u->list);
 480                 free(io_u);
 481         }
 482
 483         free_io_mem(td);
 484 }
 485
 486 /*
 487  * "randomly" fill the buffer contents
 488  */
 489 static void fill_rand_buf(struct io_u *io_u, int max_bs)
 490 {
 491         int *ptr = io_u->buf;
 492
 493         while ((void *) ptr - io_u->buf < max_bs) {
 494                 *ptr = rand() * 0x9e370001;
 495                 ptr++;
 496         }
 497 }
 498
 499 static int init_io_u(struct thread_data *td)
 500 {
 501         struct io_u *io_u;
 502         unsigned int max_bs;
 503         int i, max_units;
 504         char *p;
 505
 506         if (td->io_ops->flags & FIO_CPUIO)
 507                 return 0;
 508
 509         if (td->io_ops->flags & FIO_SYNCIO)
 510                 max_units = 1;
 511         else
 512                 max_units = td->iodepth;
 513
 514         max_bs = max(td->max_bs[DDIR_READ], td->max_bs[DDIR_WRITE]);
 515         td->orig_buffer_size = max_bs * max_units;
 516
 517         if (td->mem_type == MEM_SHMHUGE || td->mem_type == MEM_MMAPHUGE)
 518                 td->orig_buffer_size = (td->orig_buffer_size + td->hugepage_size - 1) & ~(td->hugepage_size - 1);
 519         else
 520                 td->orig_buffer_size += MASK;
 521
 522         if (allocate_io_mem(td))
 523                 return 1;
 524
 525         p = ALIGN(td->orig_buffer);
 526         for (i = 0; i < max_units; i++) {
 527                 io_u = malloc(sizeof(*io_u));
 528                 memset(io_u, 0, sizeof(*io_u));
 529                 INIT_LIST_HEAD(&io_u->list);
 530
 531                 io_u->buf = p + max_bs * i;
 532                 if (td_write(td) || td_rw(td))
 533                         fill_rand_buf(io_u, max_bs);
 534
 535                 io_u->index = i;
 536                 list_add(&io_u->list, &td->io_u_freelist);
 537         }
 538
 539         return 0;
 540 }
 541
 542 static int switch_ioscheduler(struct thread_data *td)
 543 {
 544         char tmp[256], tmp2[128];
 545         FILE *f;
 546         int ret;
 547
 548         if (td->io_ops->flags & FIO_CPUIO)
 549                 return 0;
 550
 551         sprintf(tmp, "%s/queue/scheduler", td->sysfs_root);
 552
 553         f = fopen(tmp, "r+");
 554         if (!f) {
 555                 td_verror(td, errno);
 556                 return 1;
 557         }
 558
 559         /*
 560          * Set io scheduler.
 561          */
 562         ret = fwrite(td->ioscheduler, strlen(td->ioscheduler), 1, f);
 563         if (ferror(f) || ret != 1) {
 564                 td_verror(td, errno);
 565                 fclose(f);
 566                 return 1;
 567         }
 568
 569         rewind(f);
 570
 571         /*
 572          * Read back and check that the selected scheduler is now the default.
 573          */
 574         ret = fread(tmp, 1, sizeof(tmp), f);
 575         if (ferror(f) || ret < 0) {
 576                 td_verror(td, errno);
 577                 fclose(f);
 578                 return 1;
 579         }
 580
 581         sprintf(tmp2, "[%s]", td->ioscheduler);
 582         if (!strstr(tmp, tmp2)) {
 583                 log_err("fio: io scheduler %s not found\n", td->ioscheduler);
 584                 td_verror(td, EINVAL);
 585                 fclose(f);
 586                 return 1;
 587         }
 588
 589         fclose(f);
 590         return 0;
 591 }
 592
 593 static void clear_io_state(struct thread_data *td)
 594 {
 595         struct fio_file *f;
 596         int i;
 597
 598         td->stat_io_bytes[0] = td->stat_io_bytes[1] = 0;
 599         td->this_io_bytes[0] = td->this_io_bytes[1] = 0;
 600         td->zone_bytes = 0;
 601
 602         for_each_file(td, f, i) {
 603                 f->last_pos = 0;
 604                 if (td->io_ops->flags & FIO_SYNCIO)
 605                         lseek(f->fd, SEEK_SET, 0);
 606
 607                 if (f->file_map)
 608                         memset(f->file_map, 0, f->num_maps * sizeof(long));
 609         }
 610 }
 611
 612 /*
 613  * Entry point for the thread based jobs. The process based jobs end up
 614  * here as well, after a little setup.
 615  */
 616 static void *thread_main(void *data)
 617 {
 618         unsigned long long runtime[2];
 619         struct thread_data *td = data;
 620
 621         if (!td->use_thread)
 622                 setsid();
 623
 624         td->pid = getpid();
 625
 626         INIT_LIST_HEAD(&td->io_u_freelist);
 627         INIT_LIST_HEAD(&td->io_u_busylist);
 628         INIT_LIST_HEAD(&td->io_hist_list);
 629         INIT_LIST_HEAD(&td->io_log_list);
 630
 631         if (init_io_u(td))
 632                 goto err;
 633
 634         if (fio_setaffinity(td) == -1) {
 635                 td_verror(td, errno);
 636                 goto err;
 637         }
 638
 639         if (td_io_init(td))
 640                 goto err;
 641
 642         if (init_iolog(td))
 643                 goto err;
 644
 645         if (td->ioprio) {
 646                 if (ioprio_set(IOPRIO_WHO_PROCESS, 0, td->ioprio) == -1) {
 647                         td_verror(td, errno);
 648                         goto err;
 649                 }
 650         }
 651
 652         if (nice(td->nice) == -1) {
 653                 td_verror(td, errno);
 654                 goto err;
 655         }
 656
 657         if (init_random_state(td))
 658                 goto err;
 659
 660         if (td->ioscheduler && switch_ioscheduler(td))
 661                 goto err;
 662
 663         td_set_runstate(td, TD_INITIALIZED);
 664         fio_sem_up(&startup_sem);
 665         fio_sem_down(&td->mutex);
 666
 667         if (!td->create_serialize && setup_files(td))
 668                 goto err;
 669         if (open_files(td))
 670                 goto err;
 671
 672         if (td->exec_prerun)
 673                 system(td->exec_prerun);
 674
 675         fio_gettime(&td->epoch, NULL);
 676         getrusage(RUSAGE_SELF, &td->ru_start);
 677
 678         runtime[0] = runtime[1] = 0;
 679         while (td->loops--) {
 680                 fio_gettime(&td->start, NULL);
 681                 memcpy(&td->stat_sample_time, &td->start, sizeof(td->start));
 682
 683                 if (td->ratemin)
 684                         memcpy(&td->lastrate, &td->stat_sample_time, sizeof(td->lastrate));
 685
 686                 clear_io_state(td);
 687                 prune_io_piece_log(td);
 688
 689                 if (td->io_ops->flags & FIO_CPUIO)
 690                         do_cpuio(td);
 691                 else
 692                         do_io(td);
 693
 694                 runtime[td->ddir] += utime_since_now(&td->start);
 695                 if (td_rw(td) && td->io_bytes[td->ddir ^ 1])
 696                         runtime[td->ddir ^ 1] = runtime[td->ddir];
 697
 698                 if (td->error || td->terminate)
 699                         break;
 700
 701                 if (td->verify == VERIFY_NONE)
 702                         continue;
 703
 704                 clear_io_state(td);
 705                 fio_gettime(&td->start, NULL);
 706
 707                 do_verify(td);
 708
 709                 runtime[DDIR_READ] += utime_since_now(&td->start);
 710
 711                 if (td->error || td->terminate)
 712                         break;
 713         }
 714
 715         update_rusage_stat(td);
 716         fio_gettime(&td->end_time, NULL);
 717         td->runtime[0] = runtime[0] / 1000;
 718         td->runtime[1] = runtime[1] / 1000;
 719
 720         if (td->bw_log)
 721                 finish_log(td, td->bw_log, "bw");
 722         if (td->slat_log)
 723                 finish_log(td, td->slat_log, "slat");
 724         if (td->clat_log)
 725                 finish_log(td, td->clat_log, "clat");
 726         if (td->write_iolog_file)
 727                 write_iolog_close(td);
 728         if (td->exec_postrun)
 729                 system(td->exec_postrun);
 730
 731         if (exitall_on_terminate)
 732                 terminate_threads(td->groupid);
 733
 734 err:
 735         close_files(td);
 736         close_ioengine(td);
 737         cleanup_io_u(td);
 738         td_set_runstate(td, TD_EXITED);
 739         return NULL;
 740
 741 }
 742
 743 /*
 744  * We cannot pass the td data into a forked process, so attach the td and
 745  * pass it to the thread worker.
 746  */
 747 static void *fork_main(int shmid, int offset)
 748 {
 749         struct thread_data *td;
 750         void *data;
 751
 752         data = shmat(shmid, NULL, 0);
 753         if (data == (void *) -1) {
 754                 perror("shmat");
 755                 return NULL;
 756         }
 757
 758         td = data + offset * sizeof(struct thread_data);
 759         thread_main(td);
 760         shmdt(data);
 761         return NULL;
 762 }
 763
 764 /*
 765  * Run over the job map and reap the threads that have exited, if any.
 766  */
 767 static void reap_threads(int *nr_running, int *t_rate, int *m_rate)
 768 {
 769         struct thread_data *td;
 770         int i, cputhreads, pending;
 771
 772         /*
 773          * reap exited threads (TD_EXITED -> TD_REAPED)
 774          */
 775         pending = cputhreads = 0;
 776         for_each_td(td, i) {
 777                 /*
 778                  * ->io_ops is NULL for a thread that has closed its
 779                  * io engine
 780                  */
 781                 if (td->io_ops && td->io_ops->flags & FIO_CPUIO)
 782                         cputhreads++;
 783
 784                 if (td->runstate != TD_EXITED) {
 785                         if (td->runstate < TD_RUNNING)
 786                                 pending++;
 787
 788                         continue;
 789                 }
 790
 791                 td_set_runstate(td, TD_REAPED);
 792
 793                 if (td->use_thread) {
 794                         long ret;
 795
 796                         if (pthread_join(td->thread, (void *) &ret))
 797                                 perror("thread_join");
 798                 } else
 799                         waitpid(td->pid, NULL, 0);
 800
 801                 (*nr_running)--;
 802                 (*m_rate) -= td->ratemin;
 803                 (*t_rate) -= td->rate;
 804         }
 805
 806         if (*nr_running == cputhreads && !pending)
 807                 terminate_threads(TERMINATE_ALL);
 808 }
 809
 810 /*
 811  * Main function for kicking off and reaping jobs, as needed.
 812  */
 813 static void run_threads(void)
 814 {
 815         struct thread_data *td;
 816         unsigned long spent;
 817         int i, todo, nr_running, m_rate, t_rate, nr_started;
 818
 819         if (fio_pin_memory())
 820                 return;
 821
 822         if (!terse_output) {
 823                 printf("Starting %d thread%s\n", thread_number, thread_number > 1 ? "s" : "");
 824                 fflush(stdout);
 825         }
 826
 827         signal(SIGINT, sig_handler);
 828         signal(SIGALRM, sig_handler);
 829
 830         todo = thread_number;
 831         nr_running = 0;
 832         nr_started = 0;
 833         m_rate = t_rate = 0;
 834
 835         for_each_td(td, i) {
 836                 print_status_init(td->thread_number - 1);
 837
 838                 init_disk_util(td);
 839
 840                 if (!td->create_serialize)
 841                         continue;
 842
 843                 /*
 844                  * do file setup here so it happens sequentially,
 845                  * we don't want X number of threads getting their
 846                  * client data interspersed on disk
 847                  */
 848                 if (setup_files(td)) {
 849                         td_set_runstate(td, TD_REAPED);
 850                         todo--;
 851                 }
 852         }
 853
 854         while (todo) {
 855                 struct thread_data *map[MAX_JOBS];
 856                 struct timeval this_start;
 857                 int this_jobs = 0, left;
 858
 859                 /*
 860                  * create threads (TD_NOT_CREATED -> TD_CREATED)
 861                  */
 862                 for_each_td(td, i) {
 863                         if (td->runstate != TD_NOT_CREATED)
 864                                 continue;
 865
 866                         /*
 867                          * never got a chance to start, killed by other
 868                          * thread for some reason
 869                          */
 870                         if (td->terminate) {
 871                                 todo--;
 872                                 continue;
 873                         }
 874
 875                         if (td->start_delay) {
 876                                 spent = mtime_since_genesis();
 877
 878                                 if (td->start_delay * 1000 > spent)
 879                                         continue;
 880                         }
 881
 882                         if (td->stonewall && (nr_started || nr_running))
 883                                 break;
 884
 885                         /*
 886                          * Set state to created. Thread will transition
 887                          * to TD_INITIALIZED when it's done setting up.
 888                          */
 889                         td_set_runstate(td, TD_CREATED);
 890                         map[this_jobs++] = td;
 891                         fio_sem_init(&startup_sem, 1);
 892                         nr_started++;
 893
 894                         if (td->use_thread) {
 895                                 if (pthread_create(&td->thread, NULL, thread_main, td)) {
 896                                         perror("thread_create");
 897                                         nr_started--;
 898                                 }
 899                         } else {
 900                                 if (fork())
 901                                         fio_sem_down(&startup_sem);
 902                                 else {
 903                                         fork_main(shm_id, i);
 904                                         exit(0);
 905                                 }
 906                         }
 907                 }
 908
 909                 /*
 910                  * Wait for the started threads to transition to
 911                  * TD_INITIALIZED.
 912                  */
 913                 fio_gettime(&this_start, NULL);
 914                 left = this_jobs;
 915                 while (left) {
 916                         if (mtime_since_now(&this_start) > JOB_START_TIMEOUT)
 917                                 break;
 918
 919                         usleep(100000);
 920
 921                         for (i = 0; i < this_jobs; i++) {
 922                                 td = map[i];
 923                                 if (!td)
 924                                         continue;
 925                                 if (td->runstate == TD_INITIALIZED) {
 926                                         map[i] = NULL;
 927                                         left--;
 928                                 } else if (td->runstate >= TD_EXITED) {
 929                                         map[i] = NULL;
 930                                         left--;
 931                                         todo--;
 932                                         nr_running++; /* work-around... */
 933                                 }
 934                         }
 935                 }
 936
 937                 if (left) {
 938                         log_err("fio: %d jobs failed to start\n", left);
 939                         for (i = 0; i < this_jobs; i++) {
 940                                 td = map[i];
 941                                 if (!td)
 942                                         continue;
 943                                 kill(td->pid, SIGTERM);
 944                         }
 945                         break;
 946                 }
 947
 948                 /*
 949                  * start created threads (TD_INITIALIZED -> TD_RUNNING).
 950                  */
 951                 for_each_td(td, i) {
 952                         if (td->runstate != TD_INITIALIZED)
 953                                 continue;
 954
 955                         td_set_runstate(td, TD_RUNNING);
 956                         nr_running++;
 957                         nr_started--;
 958                         m_rate += td->ratemin;
 959                         t_rate += td->rate;
 960                         todo--;
 961                         fio_sem_up(&td->mutex);
 962                 }
 963
 964                 reap_threads(&nr_running, &t_rate, &m_rate);
 965
 966                 if (todo)
 967                         usleep(100000);
 968         }
 969
 970         while (nr_running) {
 971                 reap_threads(&nr_running, &t_rate, &m_rate);
 972                 usleep(10000);
 973         }
 974
 975         update_io_ticks();
 976         fio_unpin_memory();
 977 }
 978
 979 int main(int argc, char *argv[])
 980 {
 981         if (parse_options(argc, argv))
 982                 return 1;
 983
 984         if (!thread_number) {
 985                 log_err("Nothing to do\n");
 986                 return 1;
 987         }
 988
 989         disk_util_timer_arm();
 990
 991         run_threads();
 992         show_run_stats();
 993
 994         return 0;
 995 }