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 static unsigned long page_mask;
  41 #define ALIGN(buf)      \
  42         (char *) (((unsigned long) (buf) + page_mask) & ~page_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 static volatile int fio_abort;
  51
  52 struct io_log *agg_io_log[2];
  53
  54 #define TERMINATE_ALL           (-1)
  55 #define JOB_START_TIMEOUT       (5 * 1000)
  56
  57 static inline void td_set_runstate(struct thread_data *td, int runstate)
  58 {
  59         td->runstate = runstate;
  60 }
  61
  62 static void terminate_threads(int group_id, int forced_kill)
  63 {
  64         struct thread_data *td;
  65         int i;
  66
  67         for_each_td(td, i) {
  68                 if (group_id == TERMINATE_ALL || groupid == td->groupid) {
  69                         td->terminate = 1;
  70                         td->start_delay = 0;
  71                         if (forced_kill)
  72                                 td_set_runstate(td, TD_EXITED);
  73                 }
  74         }
  75 }
  76
  77 static void sig_handler(int sig)
  78 {
  79         switch (sig) {
  80                 case SIGALRM:
  81                         update_io_ticks();
  82                         disk_util_timer_arm();
  83                         print_thread_status();
  84                         break;
  85                 case SIGSEGV:
  86                         fprintf(stderr, "fio: got segfault, aborting\n");
  87                         terminate_threads(TERMINATE_ALL, 1);
  88                         fio_abort = 1;
  89                         exit(0);
  90                 default:
  91                         printf("\nfio: terminating on signal %d\n", sig);
  92                         fflush(stdout);
  93                         terminate_threads(TERMINATE_ALL, 0);
  94                         break;
  95         }
  96 }
  97
  98 /*
  99  * Check if we are above the minimum rate given.
 100  */
 101 static int check_min_rate(struct thread_data *td, struct timeval *now)
 102 {
 103         unsigned long spent;
 104         unsigned long rate;
 105         int ddir = td->ddir;
 106
 107         /*
 108          * allow a 2 second settle period in the beginning
 109          */
 110         if (mtime_since(&td->start, now) < 2000)
 111                 return 0;
 112
 113         /*
 114          * if rate blocks is set, sample is running
 115          */
 116         if (td->rate_bytes) {
 117                 spent = mtime_since(&td->lastrate, now);
 118                 if (spent < td->ratecycle)
 119                         return 0;
 120
 121                 rate = (td->this_io_bytes[ddir] - td->rate_bytes) / spent;
 122                 if (rate < td->ratemin) {
 123                         fprintf(f_out, "%s: min rate %u not met, got %luKiB/sec\n", td->name, td->ratemin, rate);
 124                         return 1;
 125                 }
 126         }
 127
 128         td->rate_bytes = td->this_io_bytes[ddir];
 129         memcpy(&td->lastrate, now, sizeof(*now));
 130         return 0;
 131 }
 132
 133 static inline int runtime_exceeded(struct thread_data *td, struct timeval *t)
 134 {
 135         if (!td->timeout)
 136                 return 0;
 137         if (mtime_since(&td->epoch, t) >= td->timeout * 1000)
 138                 return 1;
 139
 140         return 0;
 141 }
 142
 143 static struct fio_file *get_next_file(struct thread_data *td)
 144 {
 145         unsigned int old_next_file = td->next_file;
 146         struct fio_file *f;
 147
 148         do {
 149                 f = &td->files[td->next_file];
 150
 151                 td->next_file++;
 152                 if (td->next_file >= td->nr_files)
 153                         td->next_file = 0;
 154
 155                 if (f->fd != -1)
 156                         break;
 157
 158                 f = NULL;
 159         } while (td->next_file != old_next_file);
 160
 161         return f;
 162 }
 163
 164 /*
 165  * When job exits, we can cancel the in-flight IO if we are using async
 166  * io. Attempt to do so.
 167  */
 168 static void cleanup_pending_aio(struct thread_data *td)
 169 {
 170         struct timespec ts = { .tv_sec = 0, .tv_nsec = 0};
 171         struct list_head *entry, *n;
 172         struct io_completion_data icd;
 173         struct io_u *io_u;
 174         int r;
 175
 176         /*
 177          * get immediately available events, if any
 178          */
 179         r = td_io_getevents(td, 0, td->cur_depth, &ts);
 180         if (r > 0) {
 181                 icd.nr = r;
 182                 ios_completed(td, &icd);
 183         }
 184
 185         /*
 186          * now cancel remaining active events
 187          */
 188         if (td->io_ops->cancel) {
 189                 list_for_each_safe(entry, n, &td->io_u_busylist) {
 190                         io_u = list_entry(entry, struct io_u, list);
 191
 192                         r = td->io_ops->cancel(td, io_u);
 193                         if (!r)
 194                                 put_io_u(td, io_u);
 195                 }
 196         }
 197
 198         if (td->cur_depth) {
 199                 r = td_io_getevents(td, td->cur_depth, td->cur_depth, NULL);
 200                 if (r > 0) {
 201                         icd.nr = r;
 202                         ios_completed(td, &icd);
 203                 }
 204         }
 205 }
 206
 207 /*
 208  * Helper to handle the final sync of a file. Works just like the normal
 209  * io path, just does everything sync.
 210  */
 211 static int fio_io_sync(struct thread_data *td, struct fio_file *f)
 212 {
 213         struct io_u *io_u = __get_io_u(td);
 214         struct io_completion_data icd;
 215         int ret;
 216
 217         if (!io_u)
 218                 return 1;
 219
 220         io_u->ddir = DDIR_SYNC;
 221         io_u->file = f;
 222
 223         if (td_io_prep(td, io_u)) {
 224                 put_io_u(td, io_u);
 225                 return 1;
 226         }
 227
 228         ret = td_io_queue(td, io_u);
 229         if (ret) {
 230                 td_verror(td, io_u->error);
 231                 put_io_u(td, io_u);
 232                 return 1;
 233         }
 234
 235         ret = td_io_getevents(td, 1, td->cur_depth, NULL);
 236         if (ret < 0) {
 237                 td_verror(td, ret);
 238                 return 1;
 239         }
 240
 241         icd.nr = ret;
 242         ios_completed(td, &icd);
 243         if (icd.error) {
 244                 td_verror(td, icd.error);
 245                 return 1;
 246         }
 247
 248         return 0;
 249 }
 250
 251 /*
 252  * The main verify engine. Runs over the writes we previusly submitted,
 253  * reads the blocks back in, and checks the crc/md5 of the data.
 254  */
 255 static void do_verify(struct thread_data *td)
 256 {
 257         struct io_u *io_u, *v_io_u = NULL;
 258         struct io_completion_data icd;
 259         struct fio_file *f;
 260         int ret, i;
 261
 262         /*
 263          * sync io first and invalidate cache, to make sure we really
 264          * read from disk.
 265          */
 266         for_each_file(td, f, i) {
 267                 fio_io_sync(td, f);
 268                 file_invalidate_cache(td, f);
 269         }
 270
 271         td_set_runstate(td, TD_VERIFYING);
 272
 273         do {
 274                 if (td->terminate)
 275                         break;
 276
 277                 io_u = __get_io_u(td);
 278                 if (!io_u)
 279                         break;
 280
 281                 if (runtime_exceeded(td, &io_u->start_time)) {
 282                         put_io_u(td, io_u);
 283                         break;
 284                 }
 285
 286                 if (get_next_verify(td, io_u)) {
 287                         put_io_u(td, io_u);
 288                         break;
 289                 }
 290
 291                 f = get_next_file(td);
 292                 if (!f)
 293                         break;
 294
 295                 io_u->file = f;
 296
 297                 if (td_io_prep(td, io_u)) {
 298                         put_io_u(td, io_u);
 299                         break;
 300                 }
 301
 302                 ret = td_io_queue(td, io_u);
 303                 if (ret) {
 304                         td_verror(td, io_u->error);
 305                         put_io_u(td, io_u);
 306                         break;
 307                 }
 308
 309                 /*
 310                  * we have one pending to verify, do that while
 311                  * we are doing io on the next one
 312                  */
 313                 if (do_io_u_verify(td, &v_io_u))
 314                         break;
 315
 316                 ret = td_io_getevents(td, 1, 1, NULL);
 317                 if (ret != 1) {
 318                         if (ret < 0)
 319                                 td_verror(td, ret);
 320                         break;
 321                 }
 322
 323                 v_io_u = td->io_ops->event(td, 0);
 324                 icd.nr = 1;
 325                 icd.error = 0;
 326                 fio_gettime(&icd.time, NULL);
 327                 io_completed(td, v_io_u, &icd);
 328
 329                 if (icd.error) {
 330                         td_verror(td, icd.error);
 331                         put_io_u(td, v_io_u);
 332                         v_io_u = NULL;
 333                         break;
 334                 }
 335
 336                 /*
 337                  * if we can't submit more io, we need to verify now
 338                  */
 339                 if (queue_full(td) && do_io_u_verify(td, &v_io_u))
 340                         break;
 341
 342         } while (1);
 343
 344         do_io_u_verify(td, &v_io_u);
 345
 346         if (td->cur_depth)
 347                 cleanup_pending_aio(td);
 348
 349         td_set_runstate(td, TD_RUNNING);
 350 }
 351
 352 /*
 353  * Not really an io thread, all it does is burn CPU cycles in the specified
 354  * manner.
 355  */
 356 static void do_cpuio(struct thread_data *td)
 357 {
 358         struct timeval e;
 359         int split = 100 / td->cpuload;
 360         int i = 0;
 361
 362         while (!td->terminate) {
 363                 fio_gettime(&e, NULL);
 364
 365                 if (runtime_exceeded(td, &e))
 366                         break;
 367
 368                 if (!(i % split))
 369                         __usec_sleep(10000);
 370                 else
 371                         usec_sleep(td, 10000);
 372
 373                 i++;
 374         }
 375 }
 376
 377 /*
 378  * Main IO worker function. It retrieves io_u's to process and queues
 379  * and reaps them, checking for rate and errors along the way.
 380  */
 381 static void do_io(struct thread_data *td)
 382 {
 383         struct io_completion_data icd;
 384         struct timeval s;
 385         unsigned long usec;
 386         struct fio_file *f;
 387         int i, ret = 0;
 388
 389         td_set_runstate(td, TD_RUNNING);
 390
 391         while (td->this_io_bytes[td->ddir] < td->io_size) {
 392                 struct timespec *timeout;
 393                 int min_evts = 0;
 394                 struct io_u *io_u;
 395
 396                 if (td->terminate)
 397                         break;
 398
 399                 f = get_next_file(td);
 400                 if (!f)
 401                         break;
 402
 403                 io_u = get_io_u(td, f);
 404                 if (!io_u)
 405                         break;
 406
 407                 memcpy(&s, &io_u->start_time, sizeof(s));
 408
 409                 ret = td_io_queue(td, io_u);
 410                 if (ret) {
 411                         td_verror(td, io_u->error);
 412                         put_io_u(td, io_u);
 413                         break;
 414                 }
 415
 416                 add_slat_sample(td, io_u->ddir, mtime_since(&io_u->start_time, &io_u->issue_time));
 417
 418                 if (td->cur_depth < td->iodepth) {
 419                         struct timespec ts = { .tv_sec = 0, .tv_nsec = 0};
 420
 421                         timeout = &ts;
 422                         min_evts = 0;
 423                 } else {
 424                         timeout = NULL;
 425                         min_evts = 1;
 426                 }
 427
 428                 ret = td_io_getevents(td, min_evts, td->cur_depth, timeout);
 429                 if (ret < 0) {
 430                         td_verror(td, ret);
 431                         break;
 432                 } else if (!ret)
 433                         continue;
 434
 435                 icd.nr = ret;
 436                 ios_completed(td, &icd);
 437                 if (icd.error) {
 438                         td_verror(td, icd.error);
 439                         break;
 440                 }
 441
 442                 /*
 443                  * the rate is batched for now, it should work for batches
 444                  * of completions except the very first one which may look
 445                  * a little bursty
 446                  */
 447                 usec = utime_since(&s, &icd.time);
 448
 449                 rate_throttle(td, usec, icd.bytes_done[td->ddir], td->ddir);
 450
 451                 if (check_min_rate(td, &icd.time)) {
 452                         if (exitall_on_terminate)
 453                                 terminate_threads(td->groupid, 0);
 454                         td_verror(td, ENODATA);
 455                         break;
 456                 }
 457
 458                 if (runtime_exceeded(td, &icd.time))
 459                         break;
 460
 461                 if (td->thinktime) {
 462                         unsigned long long b;
 463
 464                         b = td->io_blocks[0] + td->io_blocks[1];
 465                         if (!(b % td->thinktime_blocks))
 466                                 usec_sleep(td, td->thinktime);
 467                 }
 468         }
 469
 470         if (!td->error) {
 471                 if (td->cur_depth)
 472                         cleanup_pending_aio(td);
 473
 474                 if (should_fsync(td) && td->end_fsync) {
 475                         td_set_runstate(td, TD_FSYNCING);
 476                         for_each_file(td, f, i)
 477                                 fio_io_sync(td, f);
 478                 }
 479         }
 480 }
 481
 482 static void cleanup_io_u(struct thread_data *td)
 483 {
 484         struct list_head *entry, *n;
 485         struct io_u *io_u;
 486
 487         list_for_each_safe(entry, n, &td->io_u_freelist) {
 488                 io_u = list_entry(entry, struct io_u, list);
 489
 490                 list_del(&io_u->list);
 491                 free(io_u);
 492         }
 493
 494         free_io_mem(td);
 495 }
 496
 497 /*
 498  * "randomly" fill the buffer contents
 499  */
 500 static void fill_rand_buf(struct io_u *io_u, int max_bs)
 501 {
 502         int *ptr = io_u->buf;
 503
 504         while ((void *) ptr - io_u->buf < max_bs) {
 505                 *ptr = rand() * 0x9e370001;
 506                 ptr++;
 507         }
 508 }
 509
 510 static int init_io_u(struct thread_data *td)
 511 {
 512         struct io_u *io_u;
 513         unsigned int max_bs;
 514         int i, max_units;
 515         char *p;
 516
 517         if (td->io_ops->flags & FIO_CPUIO)
 518                 return 0;
 519
 520         if (td->io_ops->flags & FIO_SYNCIO)
 521                 max_units = 1;
 522         else
 523                 max_units = td->iodepth;
 524
 525         max_bs = max(td->max_bs[DDIR_READ], td->max_bs[DDIR_WRITE]);
 526         td->orig_buffer_size = max_bs * max_units;
 527
 528         if (td->mem_type == MEM_SHMHUGE || td->mem_type == MEM_MMAPHUGE)
 529                 td->orig_buffer_size = (td->orig_buffer_size + td->hugepage_size - 1) & ~(td->hugepage_size - 1);
 530         else
 531                 td->orig_buffer_size += page_mask;
 532
 533         if (allocate_io_mem(td))
 534                 return 1;
 535
 536         p = ALIGN(td->orig_buffer);
 537         for (i = 0; i < max_units; i++) {
 538                 io_u = malloc(sizeof(*io_u));
 539                 memset(io_u, 0, sizeof(*io_u));
 540                 INIT_LIST_HEAD(&io_u->list);
 541
 542                 io_u->buf = p + max_bs * i;
 543                 if (td_write(td) || td_rw(td))
 544                         fill_rand_buf(io_u, max_bs);
 545
 546                 io_u->index = i;
 547                 list_add(&io_u->list, &td->io_u_freelist);
 548         }
 549
 550         return 0;
 551 }
 552
 553 static int switch_ioscheduler(struct thread_data *td)
 554 {
 555         char tmp[256], tmp2[128];
 556         FILE *f;
 557         int ret;
 558
 559         if (td->io_ops->flags & FIO_CPUIO)
 560                 return 0;
 561
 562         sprintf(tmp, "%s/queue/scheduler", td->sysfs_root);
 563
 564         f = fopen(tmp, "r+");
 565         if (!f) {
 566                 td_verror(td, errno);
 567                 return 1;
 568         }
 569
 570         /*
 571          * Set io scheduler.
 572          */
 573         ret = fwrite(td->ioscheduler, strlen(td->ioscheduler), 1, f);
 574         if (ferror(f) || ret != 1) {
 575                 td_verror(td, errno);
 576                 fclose(f);
 577                 return 1;
 578         }
 579
 580         rewind(f);
 581
 582         /*
 583          * Read back and check that the selected scheduler is now the default.
 584          */
 585         ret = fread(tmp, 1, sizeof(tmp), f);
 586         if (ferror(f) || ret < 0) {
 587                 td_verror(td, errno);
 588                 fclose(f);
 589                 return 1;
 590         }
 591
 592         sprintf(tmp2, "[%s]", td->ioscheduler);
 593         if (!strstr(tmp, tmp2)) {
 594                 log_err("fio: io scheduler %s not found\n", td->ioscheduler);
 595                 td_verror(td, EINVAL);
 596                 fclose(f);
 597                 return 1;
 598         }
 599
 600         fclose(f);
 601         return 0;
 602 }
 603
 604 static void clear_io_state(struct thread_data *td)
 605 {
 606         struct fio_file *f;
 607         int i;
 608
 609         td->stat_io_bytes[0] = td->stat_io_bytes[1] = 0;
 610         td->this_io_bytes[0] = td->this_io_bytes[1] = 0;
 611         td->zone_bytes = 0;
 612
 613         for_each_file(td, f, i) {
 614                 f->last_pos = 0;
 615                 if (td->io_ops->flags & FIO_SYNCIO)
 616                         lseek(f->fd, SEEK_SET, 0);
 617
 618                 if (f->file_map)
 619                         memset(f->file_map, 0, f->num_maps * sizeof(long));
 620         }
 621 }
 622
 623 /*
 624  * Entry point for the thread based jobs. The process based jobs end up
 625  * here as well, after a little setup.
 626  */
 627 static void *thread_main(void *data)
 628 {
 629         unsigned long long runtime[2];
 630         struct thread_data *td = data;
 631
 632         if (!td->use_thread)
 633                 setsid();
 634
 635         td->pid = getpid();
 636
 637         INIT_LIST_HEAD(&td->io_u_freelist);
 638         INIT_LIST_HEAD(&td->io_u_busylist);
 639         INIT_LIST_HEAD(&td->io_hist_list);
 640         INIT_LIST_HEAD(&td->io_log_list);
 641
 642         if (init_io_u(td))
 643                 goto err;
 644
 645         if (fio_setaffinity(td) == -1) {
 646                 td_verror(td, errno);
 647                 goto err;
 648         }
 649
 650         if (init_iolog(td))
 651                 goto err;
 652
 653         if (td->ioprio) {
 654                 if (ioprio_set(IOPRIO_WHO_PROCESS, 0, td->ioprio) == -1) {
 655                         td_verror(td, errno);
 656                         goto err;
 657                 }
 658         }
 659
 660         if (nice(td->nice) == -1) {
 661                 td_verror(td, errno);
 662                 goto err;
 663         }
 664
 665         if (init_random_state(td))
 666                 goto err;
 667
 668         if (td->ioscheduler && switch_ioscheduler(td))
 669                 goto err;
 670
 671         td_set_runstate(td, TD_INITIALIZED);
 672         fio_sem_up(&startup_sem);
 673         fio_sem_down(&td->mutex);
 674
 675         if (!td->create_serialize && setup_files(td))
 676                 goto err;
 677         if (open_files(td))
 678                 goto err;
 679
 680         /*
 681          * Do this late, as some IO engines would like to have the
 682          * files setup prior to initializing structures.
 683          */
 684         if (td_io_init(td))
 685                 goto err;
 686
 687         if (td->exec_prerun)
 688                 system(td->exec_prerun);
 689
 690         fio_gettime(&td->epoch, NULL);
 691         getrusage(RUSAGE_SELF, &td->ru_start);
 692
 693         runtime[0] = runtime[1] = 0;
 694         while (td->loops--) {
 695                 fio_gettime(&td->start, NULL);
 696                 memcpy(&td->stat_sample_time, &td->start, sizeof(td->start));
 697
 698                 if (td->ratemin)
 699                         memcpy(&td->lastrate, &td->stat_sample_time, sizeof(td->lastrate));
 700
 701                 clear_io_state(td);
 702                 prune_io_piece_log(td);
 703
 704                 if (td->io_ops->flags & FIO_CPUIO)
 705                         do_cpuio(td);
 706                 else
 707                         do_io(td);
 708
 709                 runtime[td->ddir] += utime_since_now(&td->start);
 710                 if (td_rw(td) && td->io_bytes[td->ddir ^ 1])
 711                         runtime[td->ddir ^ 1] = runtime[td->ddir];
 712
 713                 if (td->error || td->terminate)
 714                         break;
 715
 716                 if (td->verify == VERIFY_NONE)
 717                         continue;
 718
 719                 clear_io_state(td);
 720                 fio_gettime(&td->start, NULL);
 721
 722                 do_verify(td);
 723
 724                 runtime[DDIR_READ] += utime_since_now(&td->start);
 725
 726                 if (td->error || td->terminate)
 727                         break;
 728         }
 729
 730         update_rusage_stat(td);
 731         fio_gettime(&td->end_time, NULL);
 732         td->runtime[0] = runtime[0] / 1000;
 733         td->runtime[1] = runtime[1] / 1000;
 734
 735         if (td->bw_log)
 736                 finish_log(td, td->bw_log, "bw");
 737         if (td->slat_log)
 738                 finish_log(td, td->slat_log, "slat");
 739         if (td->clat_log)
 740                 finish_log(td, td->clat_log, "clat");
 741         if (td->write_iolog_file)
 742                 write_iolog_close(td);
 743         if (td->exec_postrun)
 744                 system(td->exec_postrun);
 745
 746         if (exitall_on_terminate)
 747                 terminate_threads(td->groupid, 0);
 748
 749 err:
 750         close_files(td);
 751         close_ioengine(td);
 752         cleanup_io_u(td);
 753         td_set_runstate(td, TD_EXITED);
 754         return NULL;
 755
 756 }
 757
 758 /*
 759  * We cannot pass the td data into a forked process, so attach the td and
 760  * pass it to the thread worker.
 761  */
 762 static void *fork_main(int shmid, int offset)
 763 {
 764         struct thread_data *td;
 765         void *data;
 766
 767         data = shmat(shmid, NULL, 0);
 768         if (data == (void *) -1) {
 769                 perror("shmat");
 770                 return NULL;
 771         }
 772
 773         td = data + offset * sizeof(struct thread_data);
 774         thread_main(td);
 775         shmdt(data);
 776         return NULL;
 777 }
 778
 779 /*
 780  * Run over the job map and reap the threads that have exited, if any.
 781  */
 782 static void reap_threads(int *nr_running, int *t_rate, int *m_rate)
 783 {
 784         struct thread_data *td;
 785         int i, cputhreads, pending;
 786
 787         /*
 788          * reap exited threads (TD_EXITED -> TD_REAPED)
 789          */
 790         pending = cputhreads = 0;
 791         for_each_td(td, i) {
 792                 /*
 793                  * ->io_ops is NULL for a thread that has closed its
 794                  * io engine
 795                  */
 796                 if (td->io_ops && td->io_ops->flags & FIO_CPUIO)
 797                         cputhreads++;
 798
 799                 if (td->runstate != TD_EXITED) {
 800                         if (td->runstate < TD_RUNNING)
 801                                 pending++;
 802
 803                         continue;
 804                 }
 805
 806                 td_set_runstate(td, TD_REAPED);
 807
 808                 if (td->use_thread) {
 809                         long ret;
 810
 811                         if (pthread_join(td->thread, (void *) &ret))
 812                                 perror("thread_join");
 813                 } else
 814                         waitpid(td->pid, NULL, 0);
 815
 816                 (*nr_running)--;
 817                 (*m_rate) -= td->ratemin;
 818                 (*t_rate) -= td->rate;
 819         }
 820
 821         if (*nr_running == cputhreads && !pending)
 822                 terminate_threads(TERMINATE_ALL, 0);
 823 }
 824
 825 /*
 826  * Main function for kicking off and reaping jobs, as needed.
 827  */
 828 static void run_threads(void)
 829 {
 830         struct thread_data *td;
 831         unsigned long spent;
 832         int i, todo, nr_running, m_rate, t_rate, nr_started;
 833
 834         if (fio_pin_memory())
 835                 return;
 836
 837         if (!terse_output) {
 838                 printf("Starting %d thread%s\n", thread_number, thread_number > 1 ? "s" : "");
 839                 fflush(stdout);
 840         }
 841
 842         signal(SIGINT, sig_handler);
 843         signal(SIGALRM, sig_handler);
 844
 845         todo = thread_number;
 846         nr_running = 0;
 847         nr_started = 0;
 848         m_rate = t_rate = 0;
 849
 850         for_each_td(td, i) {
 851                 print_status_init(td->thread_number - 1);
 852
 853                 if (!td->create_serialize) {
 854                         init_disk_util(td);
 855                         continue;
 856                 }
 857
 858                 /*
 859                  * do file setup here so it happens sequentially,
 860                  * we don't want X number of threads getting their
 861                  * client data interspersed on disk
 862                  */
 863                 if (setup_files(td)) {
 864                         td_set_runstate(td, TD_REAPED);
 865                         todo--;
 866                 }
 867
 868                 init_disk_util(td);
 869         }
 870
 871         while (todo) {
 872                 struct thread_data *map[MAX_JOBS];
 873                 struct timeval this_start;
 874                 int this_jobs = 0, left;
 875
 876                 /*
 877                  * create threads (TD_NOT_CREATED -> TD_CREATED)
 878                  */
 879                 for_each_td(td, i) {
 880                         if (td->runstate != TD_NOT_CREATED)
 881                                 continue;
 882
 883                         /*
 884                          * never got a chance to start, killed by other
 885                          * thread for some reason
 886                          */
 887                         if (td->terminate) {
 888                                 todo--;
 889                                 continue;
 890                         }
 891
 892                         if (td->start_delay) {
 893                                 spent = mtime_since_genesis();
 894
 895                                 if (td->start_delay * 1000 > spent)
 896                                         continue;
 897                         }
 898
 899                         if (td->stonewall && (nr_started || nr_running))
 900                                 break;
 901
 902                         /*
 903                          * Set state to created. Thread will transition
 904                          * to TD_INITIALIZED when it's done setting up.
 905                          */
 906                         td_set_runstate(td, TD_CREATED);
 907                         map[this_jobs++] = td;
 908                         fio_sem_init(&startup_sem, 1);
 909                         nr_started++;
 910
 911                         if (td->use_thread) {
 912                                 if (pthread_create(&td->thread, NULL, thread_main, td)) {
 913                                         perror("thread_create");
 914                                         nr_started--;
 915                                 }
 916                         } else {
 917                                 if (fork())
 918                                         fio_sem_down(&startup_sem);
 919                                 else {
 920                                         fork_main(shm_id, i);
 921                                         exit(0);
 922                                 }
 923                         }
 924                 }
 925
 926                 /*
 927                  * Wait for the started threads to transition to
 928                  * TD_INITIALIZED.
 929                  */
 930                 fio_gettime(&this_start, NULL);
 931                 left = this_jobs;
 932                 while (left && !fio_abort) {
 933                         if (mtime_since_now(&this_start) > JOB_START_TIMEOUT)
 934                                 break;
 935
 936                         usleep(100000);
 937
 938                         for (i = 0; i < this_jobs; i++) {
 939                                 td = map[i];
 940                                 if (!td)
 941                                         continue;
 942                                 if (td->runstate == TD_INITIALIZED) {
 943                                         map[i] = NULL;
 944                                         left--;
 945                                 } else if (td->runstate >= TD_EXITED) {
 946                                         map[i] = NULL;
 947                                         left--;
 948                                         todo--;
 949                                         nr_running++; /* work-around... */
 950                                 }
 951                         }
 952                 }
 953
 954                 if (left) {
 955                         log_err("fio: %d jobs failed to start\n", left);
 956                         for (i = 0; i < this_jobs; i++) {
 957                                 td = map[i];
 958                                 if (!td)
 959                                         continue;
 960                                 kill(td->pid, SIGTERM);
 961                         }
 962                         break;
 963                 }
 964
 965                 /*
 966                  * start created threads (TD_INITIALIZED -> TD_RUNNING).
 967                  */
 968                 for_each_td(td, i) {
 969                         if (td->runstate != TD_INITIALIZED)
 970                                 continue;
 971
 972                         td_set_runstate(td, TD_RUNNING);
 973                         nr_running++;
 974                         nr_started--;
 975                         m_rate += td->ratemin;
 976                         t_rate += td->rate;
 977                         todo--;
 978                         fio_sem_up(&td->mutex);
 979                 }
 980
 981                 reap_threads(&nr_running, &t_rate, &m_rate);
 982
 983                 if (todo)
 984                         usleep(100000);
 985         }
 986
 987         while (nr_running) {
 988                 reap_threads(&nr_running, &t_rate, &m_rate);
 989                 usleep(10000);
 990         }
 991
 992         update_io_ticks();
 993         fio_unpin_memory();
 994 }
 995
 996 int main(int argc, char *argv[])
 997 {
 998         long ps;
 999
1000         if (parse_options(argc, argv))
1001                 return 1;
1002
1003         if (!thread_number) {
1004                 log_err("Nothing to do\n");
1005                 return 1;
1006         }
1007
1008         ps = sysconf(_SC_PAGESIZE);
1009         if (ps < 0) {
1010                 log_err("Failed to get page size\n");
1011                 return 1;
1012         }
1013
1014         page_mask = ps - 1;
1015
1016         if (write_bw_log) {
1017                 setup_log(&agg_io_log[DDIR_READ]);
1018                 setup_log(&agg_io_log[DDIR_WRITE]);
1019         }
1020
1021         disk_util_timer_arm();
1022
1023         run_threads();
1024
1025         if (!fio_abort) {
1026                 show_run_stats();
1027                 if (write_bw_log) {
1028                         __finish_log(agg_io_log[DDIR_READ],"agg-read_bw.log");
1029                         __finish_log(agg_io_log[DDIR_WRITE],"agg-write_bw.log");
1030                 }
1031         }
1032
1033         return 0;
1034 }