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 <locale.h>
  30 #include <assert.h>
  31 #include <sys/stat.h>
  32 #include <sys/wait.h>
  33 #include <sys/ipc.h>
  34 #include <sys/shm.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 static int exit_value;
  52
  53 struct io_log *agg_io_log[2];
  54
  55 #define TERMINATE_ALL           (-1)
  56 #define JOB_START_TIMEOUT       (5 * 1000)
  57
  58 static inline void td_set_runstate(struct thread_data *td, int runstate)
  59 {
  60         td->runstate = runstate;
  61 }
  62
  63 static void terminate_threads(int group_id, int forced_kill)
  64 {
  65         struct thread_data *td;
  66         int i;
  67
  68         for_each_td(td, i) {
  69                 if (group_id == TERMINATE_ALL || groupid == td->groupid) {
  70                         td->terminate = 1;
  71                         td->start_delay = 0;
  72                         if (forced_kill)
  73                                 td_set_runstate(td, TD_EXITED);
  74                 }
  75         }
  76 }
  77
  78 static void sig_handler(int sig)
  79 {
  80         switch (sig) {
  81                 case SIGALRM:
  82                         update_io_ticks();
  83                         disk_util_timer_arm();
  84                         print_thread_status();
  85                         break;
  86                 default:
  87                         printf("\nfio: terminating on signal %d\n", sig);
  88                         fflush(stdout);
  89                         terminate_threads(TERMINATE_ALL, 0);
  90                         break;
  91         }
  92 }
  93
  94 /*
  95  * Check if we are above the minimum rate given.
  96  */
  97 static int check_min_rate(struct thread_data *td, struct timeval *now)
  98 {
  99         unsigned long spent;
 100         unsigned long rate;
 101         int ddir = td->ddir;
 102
 103         /*
 104          * allow a 2 second settle period in the beginning
 105          */
 106         if (mtime_since(&td->start, now) < 2000)
 107                 return 0;
 108
 109         /*
 110          * if rate blocks is set, sample is running
 111          */
 112         if (td->rate_bytes) {
 113                 spent = mtime_since(&td->lastrate, now);
 114                 if (spent < td->ratecycle)
 115                         return 0;
 116
 117                 rate = (td->this_io_bytes[ddir] - td->rate_bytes) / spent;
 118                 if (rate < td->ratemin) {
 119                         fprintf(f_out, "%s: min rate %u not met, got %luKiB/sec\n", td->name, td->ratemin, rate);
 120                         return 1;
 121                 }
 122         }
 123
 124         td->rate_bytes = td->this_io_bytes[ddir];
 125         memcpy(&td->lastrate, now, sizeof(*now));
 126         return 0;
 127 }
 128
 129 static inline int runtime_exceeded(struct thread_data *td, struct timeval *t)
 130 {
 131         if (!td->timeout)
 132                 return 0;
 133         if (mtime_since(&td->epoch, t) >= td->timeout * 1000)
 134                 return 1;
 135
 136         return 0;
 137 }
 138
 139 static struct fio_file *get_next_file(struct thread_data *td)
 140 {
 141         unsigned int old_next_file = td->next_file;
 142         struct fio_file *f;
 143
 144         do {
 145                 f = &td->files[td->next_file];
 146
 147                 td->next_file++;
 148                 if (td->next_file >= td->nr_files)
 149                         td->next_file = 0;
 150
 151                 if (f->fd != -1)
 152                         break;
 153
 154                 f = NULL;
 155         } while (td->next_file != old_next_file);
 156
 157         return f;
 158 }
 159
 160 /*
 161  * When job exits, we can cancel the in-flight IO if we are using async
 162  * io. Attempt to do so.
 163  */
 164 static void cleanup_pending_aio(struct thread_data *td)
 165 {
 166         struct list_head *entry, *n;
 167         struct io_u *io_u;
 168         int r;
 169
 170         /*
 171          * get immediately available events, if any
 172          */
 173         io_u_queued_complete(td, 0, NULL);
 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                 io_u_queued_complete(td, td->cur_depth, NULL);
 190 }
 191
 192 /*
 193  * Helper to handle the final sync of a file. Works just like the normal
 194  * io path, just does everything sync.
 195  */
 196 static int fio_io_sync(struct thread_data *td, struct fio_file *f)
 197 {
 198         struct io_u *io_u = __get_io_u(td);
 199         int ret;
 200
 201         if (!io_u)
 202                 return 1;
 203
 204         io_u->ddir = DDIR_SYNC;
 205         io_u->file = f;
 206
 207         if (td_io_prep(td, io_u)) {
 208                 put_io_u(td, io_u);
 209                 return 1;
 210         }
 211
 212         ret = td_io_queue(td, io_u);
 213         if (ret < 0) {
 214                 td_verror(td, io_u->error);
 215                 put_io_u(td, io_u);
 216                 return 1;
 217         } else if (ret == FIO_Q_QUEUED) {
 218                 if (io_u_queued_complete(td, 1, NULL))
 219                         return 1;
 220         } else if (ret == FIO_Q_COMPLETED) {
 221                 if (io_u->error) {
 222                         td_verror(td, io_u->error);
 223                         return 1;
 224                 }
 225
 226                 io_u_sync_complete(td, io_u, NULL);
 227         }
 228
 229         return 0;
 230 }
 231
 232 /*
 233  * The main verify engine. Runs over the writes we previusly submitted,
 234  * reads the blocks back in, and checks the crc/md5 of the data.
 235  */
 236 static void do_verify(struct thread_data *td)
 237 {
 238         struct fio_file *f;
 239         struct io_u *io_u;
 240         int ret, i, min_events;
 241
 242         /*
 243          * sync io first and invalidate cache, to make sure we really
 244          * read from disk.
 245          */
 246         for_each_file(td, f, i) {
 247                 fio_io_sync(td, f);
 248                 file_invalidate_cache(td, f);
 249         }
 250
 251         td_set_runstate(td, TD_VERIFYING);
 252
 253         io_u = NULL;
 254         while (!td->terminate) {
 255                 io_u = __get_io_u(td);
 256                 if (!io_u)
 257                         break;
 258
 259                 if (runtime_exceeded(td, &io_u->start_time))
 260                         break;
 261
 262                 if (get_next_verify(td, io_u))
 263                         break;
 264
 265                 if (td_io_prep(td, io_u))
 266                         break;
 267
 268 requeue:
 269                 ret = td_io_queue(td, io_u);
 270
 271                 switch (ret) {
 272                 case FIO_Q_COMPLETED:
 273                         if (io_u->error)
 274                                 ret = -io_u->error;
 275                         if (io_u->xfer_buflen != io_u->resid && io_u->resid) {
 276                                 int bytes = io_u->xfer_buflen - io_u->resid;
 277
 278                                 io_u->xfer_buflen = io_u->resid;
 279                                 io_u->xfer_buf += bytes;
 280                                 goto requeue;
 281                         }
 282                         if (io_u_sync_complete(td, io_u, verify_io_u))
 283                                 break;
 284                         continue;
 285                 case FIO_Q_QUEUED:
 286                         break;
 287                 default:
 288                         assert(ret < 0);
 289                         td_verror(td, -ret);
 290                         break;
 291                 }
 292
 293                 if (ret < 0)
 294                         break;
 295
 296                 /*
 297                  * if we can queue more, do so. but check if there are
 298                  * completed io_u's first.
 299                  */
 300                 min_events = 0;
 301                 if (queue_full(td))
 302                         min_events = 1;
 303
 304                 /*
 305                  * Reap required number of io units, if any, and do the
 306                  * verification on them through the callback handler
 307                  */
 308                 if (io_u_queued_complete(td, min_events, verify_io_u))
 309                         break;
 310         }
 311
 312         if (io_u)
 313                 put_io_u(td, io_u);
 314
 315         if (td->cur_depth)
 316                 cleanup_pending_aio(td);
 317
 318         td_set_runstate(td, TD_RUNNING);
 319 }
 320
 321 /*
 322  * Not really an io thread, all it does is burn CPU cycles in the specified
 323  * manner.
 324  */
 325 static void do_cpuio(struct thread_data *td)
 326 {
 327         struct timeval e;
 328         int split = 100 / td->cpuload;
 329         int i = 0;
 330
 331         while (!td->terminate) {
 332                 fio_gettime(&e, NULL);
 333
 334                 if (runtime_exceeded(td, &e))
 335                         break;
 336
 337                 if (!(i % split))
 338                         __usec_sleep(10000);
 339                 else
 340                         usec_sleep(td, 10000);
 341
 342                 i++;
 343         }
 344 }
 345
 346 /*
 347  * Main IO worker function. It retrieves io_u's to process and queues
 348  * and reaps them, checking for rate and errors along the way.
 349  */
 350 static void do_io(struct thread_data *td)
 351 {
 352         struct timeval s;
 353         unsigned long usec;
 354         struct fio_file *f;
 355         int i, ret = 0;
 356
 357         td_set_runstate(td, TD_RUNNING);
 358
 359         while ((td->this_io_bytes[0] + td->this_io_bytes[1]) < td->io_size) {
 360                 struct timeval comp_time;
 361                 long bytes_done = 0;
 362                 int min_evts = 0;
 363                 struct io_u *io_u;
 364
 365                 if (td->terminate)
 366                         break;
 367
 368                 f = get_next_file(td);
 369                 if (!f)
 370                         break;
 371
 372                 io_u = get_io_u(td, f);
 373                 if (!io_u)
 374                         break;
 375
 376                 memcpy(&s, &io_u->start_time, sizeof(s));
 377
 378                 if (runtime_exceeded(td, &s)) {
 379                         put_io_u(td, io_u);
 380                         break;
 381                 }
 382 requeue:
 383                 ret = td_io_queue(td, io_u);
 384
 385                 switch (ret) {
 386                 case FIO_Q_COMPLETED:
 387                         if (io_u->error) {
 388                                 ret = io_u->error;
 389                                 break;
 390                         }
 391                         if (io_u->xfer_buflen != io_u->resid && io_u->resid) {
 392                                 int bytes = io_u->xfer_buflen - io_u->resid;
 393
 394                                 io_u->xfer_buflen = io_u->resid;
 395                                 io_u->xfer_buf += bytes;
 396                                 goto requeue;
 397                         }
 398                         fio_gettime(&comp_time, NULL);
 399                         bytes_done = io_u_sync_complete(td, io_u, NULL);
 400                         break;
 401                 case FIO_Q_QUEUED:
 402                         break;
 403                 default:
 404                         assert(ret < 0);
 405                         put_io_u(td, io_u);
 406                         break;
 407                 }
 408
 409                 if (ret < 0)
 410                         break;
 411
 412                 add_slat_sample(td, io_u->ddir, mtime_since(&io_u->start_time, &io_u->issue_time));
 413
 414                 /*
 415                  * See if we need to complete some commands
 416                  */
 417                 if (ret == FIO_Q_QUEUED) {
 418                         min_evts = 0;
 419                         if (queue_full(td))
 420                                 min_evts = 1;
 421
 422                         fio_gettime(&comp_time, NULL);
 423                         bytes_done = io_u_queued_complete(td, min_evts, NULL);
 424                         if (bytes_done < 0)
 425                                 break;
 426                 }
 427
 428                 if (!bytes_done)
 429                         continue;
 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, &comp_time);
 437
 438                 rate_throttle(td, usec, bytes_done, td->ddir);
 439
 440                 if (check_min_rate(td, &comp_time)) {
 441                         if (exitall_on_terminate)
 442                                 terminate_threads(td->groupid, 0);
 443                         td_verror(td, ENODATA);
 444                         break;
 445                 }
 446
 447                 if (td->thinktime) {
 448                         unsigned long long b;
 449
 450                         b = td->io_blocks[0] + td->io_blocks[1];
 451                         if (!(b % td->thinktime_blocks)) {
 452                                 int left;
 453
 454                                 if (td->thinktime_spin)
 455                                         __usec_sleep(td->thinktime_spin);
 456
 457                                 left = td->thinktime - td->thinktime_spin;
 458                                 if (left)
 459                                         usec_sleep(td, left);
 460                         }
 461                 }
 462         }
 463
 464         if (!td->error) {
 465                 if (td->cur_depth)
 466                         cleanup_pending_aio(td);
 467
 468                 if (should_fsync(td) && td->end_fsync) {
 469                         td_set_runstate(td, TD_FSYNCING);
 470                         for_each_file(td, f, i)
 471                                 fio_io_sync(td, f);
 472                 }
 473         }
 474 }
 475
 476 static void cleanup_io_u(struct thread_data *td)
 477 {
 478         struct list_head *entry, *n;
 479         struct io_u *io_u;
 480
 481         list_for_each_safe(entry, n, &td->io_u_freelist) {
 482                 io_u = list_entry(entry, struct io_u, list);
 483
 484                 list_del(&io_u->list);
 485                 free(io_u);
 486         }
 487
 488         free_io_mem(td);
 489 }
 490
 491 /*
 492  * "randomly" fill the buffer contents
 493  */
 494 static void fill_rand_buf(struct io_u *io_u, int max_bs)
 495 {
 496         int *ptr = io_u->buf;
 497
 498         while ((void *) ptr - io_u->buf < max_bs) {
 499                 *ptr = rand() * 0x9e370001;
 500                 ptr++;
 501         }
 502 }
 503
 504 static int init_io_u(struct thread_data *td)
 505 {
 506         struct io_u *io_u;
 507         unsigned int max_bs;
 508         int i, max_units;
 509         char *p;
 510
 511         if (td->io_ops->flags & FIO_CPUIO)
 512                 return 0;
 513
 514         if (td->io_ops->flags & FIO_SYNCIO)
 515                 max_units = 1;
 516         else
 517                 max_units = td->iodepth;
 518
 519         max_bs = max(td->max_bs[DDIR_READ], td->max_bs[DDIR_WRITE]);
 520         td->orig_buffer_size = max_bs * max_units;
 521
 522         if (td->mem_type == MEM_SHMHUGE || td->mem_type == MEM_MMAPHUGE)
 523                 td->orig_buffer_size = (td->orig_buffer_size + td->hugepage_size - 1) & ~(td->hugepage_size - 1);
 524         else
 525                 td->orig_buffer_size += page_mask;
 526
 527         if (allocate_io_mem(td))
 528                 return 1;
 529
 530         p = ALIGN(td->orig_buffer);
 531         for (i = 0; i < max_units; i++) {
 532                 io_u = malloc(sizeof(*io_u));
 533                 memset(io_u, 0, sizeof(*io_u));
 534                 INIT_LIST_HEAD(&io_u->list);
 535
 536                 io_u->buf = p + max_bs * i;
 537                 if (td_write(td) || td_rw(td))
 538                         fill_rand_buf(io_u, max_bs);
 539
 540                 io_u->index = i;
 541                 list_add(&io_u->list, &td->io_u_freelist);
 542         }
 543
 544         return 0;
 545 }
 546
 547 static int switch_ioscheduler(struct thread_data *td)
 548 {
 549         char tmp[256], tmp2[128];
 550         FILE *f;
 551         int ret;
 552
 553         if (td->io_ops->flags & FIO_CPUIO)
 554                 return 0;
 555
 556         sprintf(tmp, "%s/queue/scheduler", td->sysfs_root);
 557
 558         f = fopen(tmp, "r+");
 559         if (!f) {
 560                 td_verror(td, errno);
 561                 return 1;
 562         }
 563
 564         /*
 565          * Set io scheduler.
 566          */
 567         ret = fwrite(td->ioscheduler, strlen(td->ioscheduler), 1, f);
 568         if (ferror(f) || ret != 1) {
 569                 td_verror(td, errno);
 570                 fclose(f);
 571                 return 1;
 572         }
 573
 574         rewind(f);
 575
 576         /*
 577          * Read back and check that the selected scheduler is now the default.
 578          */
 579         ret = fread(tmp, 1, sizeof(tmp), f);
 580         if (ferror(f) || ret < 0) {
 581                 td_verror(td, errno);
 582                 fclose(f);
 583                 return 1;
 584         }
 585
 586         sprintf(tmp2, "[%s]", td->ioscheduler);
 587         if (!strstr(tmp, tmp2)) {
 588                 log_err("fio: io scheduler %s not found\n", td->ioscheduler);
 589                 td_verror(td, EINVAL);
 590                 fclose(f);
 591                 return 1;
 592         }
 593
 594         fclose(f);
 595         return 0;
 596 }
 597
 598 static void clear_io_state(struct thread_data *td)
 599 {
 600         struct fio_file *f;
 601         int i;
 602
 603         td->stat_io_bytes[0] = td->stat_io_bytes[1] = 0;
 604         td->this_io_bytes[0] = td->this_io_bytes[1] = 0;
 605         td->zone_bytes = 0;
 606
 607         for_each_file(td, f, i) {
 608                 f->last_pos = 0;
 609                 if (td->io_ops->flags & FIO_SYNCIO)
 610                         lseek(f->fd, SEEK_SET, 0);
 611
 612                 if (f->file_map)
 613                         memset(f->file_map, 0, f->num_maps * sizeof(long));
 614         }
 615 }
 616
 617 /*
 618  * Entry point for the thread based jobs. The process based jobs end up
 619  * here as well, after a little setup.
 620  */
 621 static void *thread_main(void *data)
 622 {
 623         unsigned long long runtime[2];
 624         struct thread_data *td = data;
 625
 626         if (!td->use_thread)
 627                 setsid();
 628
 629         td->pid = getpid();
 630
 631         INIT_LIST_HEAD(&td->io_u_freelist);
 632         INIT_LIST_HEAD(&td->io_u_busylist);
 633         INIT_LIST_HEAD(&td->io_hist_list);
 634         INIT_LIST_HEAD(&td->io_log_list);
 635
 636         if (init_io_u(td))
 637                 goto err;
 638
 639         if (fio_setaffinity(td) == -1) {
 640                 td_verror(td, errno);
 641                 goto err;
 642         }
 643
 644         if (init_iolog(td))
 645                 goto err;
 646
 647         if (td->ioprio) {
 648                 if (ioprio_set(IOPRIO_WHO_PROCESS, 0, td->ioprio) == -1) {
 649                         td_verror(td, errno);
 650                         goto err;
 651                 }
 652         }
 653
 654         if (nice(td->nice) == -1) {
 655                 td_verror(td, errno);
 656                 goto err;
 657         }
 658
 659         if (init_random_state(td))
 660                 goto err;
 661
 662         if (td->ioscheduler && switch_ioscheduler(td))
 663                 goto err;
 664
 665         td_set_runstate(td, TD_INITIALIZED);
 666         fio_sem_up(&startup_sem);
 667         fio_sem_down(&td->mutex);
 668
 669         if (!td->create_serialize && setup_files(td))
 670                 goto err;
 671         if (open_files(td))
 672                 goto err;
 673
 674         /*
 675          * Do this late, as some IO engines would like to have the
 676          * files setup prior to initializing structures.
 677          */
 678         if (td_io_init(td))
 679                 goto err;
 680
 681         if (td->exec_prerun) {
 682                 if (system(td->exec_prerun) < 0)
 683                         goto err;
 684         }
 685
 686         fio_gettime(&td->epoch, NULL);
 687         getrusage(RUSAGE_SELF, &td->ru_start);
 688
 689         runtime[0] = runtime[1] = 0;
 690         while (td->loops--) {
 691                 fio_gettime(&td->start, NULL);
 692                 memcpy(&td->stat_sample_time, &td->start, sizeof(td->start));
 693
 694                 if (td->ratemin)
 695                         memcpy(&td->lastrate, &td->stat_sample_time, sizeof(td->lastrate));
 696
 697                 clear_io_state(td);
 698                 prune_io_piece_log(td);
 699
 700                 if (td->io_ops->flags & FIO_CPUIO)
 701                         do_cpuio(td);
 702                 else
 703                         do_io(td);
 704
 705                 runtime[td->ddir] += utime_since_now(&td->start);
 706                 if (td_rw(td) && td->io_bytes[td->ddir ^ 1])
 707                         runtime[td->ddir ^ 1] = runtime[td->ddir];
 708
 709                 if (td->error || td->terminate)
 710                         break;
 711
 712                 if (td->verify == VERIFY_NONE)
 713                         continue;
 714
 715                 clear_io_state(td);
 716                 fio_gettime(&td->start, NULL);
 717
 718                 do_verify(td);
 719
 720                 runtime[DDIR_READ] += utime_since_now(&td->start);
 721
 722                 if (td->error || td->terminate)
 723                         break;
 724         }
 725
 726         update_rusage_stat(td);
 727         fio_gettime(&td->end_time, NULL);
 728         td->runtime[0] = runtime[0] / 1000;
 729         td->runtime[1] = runtime[1] / 1000;
 730
 731         if (td->bw_log)
 732                 finish_log(td, td->bw_log, "bw");
 733         if (td->slat_log)
 734                 finish_log(td, td->slat_log, "slat");
 735         if (td->clat_log)
 736                 finish_log(td, td->clat_log, "clat");
 737         if (td->write_iolog_file)
 738                 write_iolog_close(td);
 739         if (td->exec_postrun) {
 740                 if (system(td->exec_postrun) < 0)
 741                         log_err("fio: postrun %s failed\n", td->exec_postrun);
 742         }
 743
 744         if (exitall_on_terminate)
 745                 terminate_threads(td->groupid, 0);
 746
 747 err:
 748         if (td->error)
 749                 printf("fio: pid=%d, err=%d/%s\n", td->pid, td->error, td->verror);
 750         close_files(td);
 751         close_ioengine(td);
 752         cleanup_io_u(td);
 753         td_set_runstate(td, TD_EXITED);
 754         return (void *) td->error;
 755 }
 756
 757 /*
 758  * We cannot pass the td data into a forked process, so attach the td and
 759  * pass it to the thread worker.
 760  */
 761 static int fork_main(int shmid, int offset)
 762 {
 763         struct thread_data *td;
 764         void *data, *ret;
 765
 766         data = shmat(shmid, NULL, 0);
 767         if (data == (void *) -1) {
 768                 int __err = errno;
 769
 770                 perror("shmat");
 771                 return __err;
 772         }
 773
 774         td = data + offset * sizeof(struct thread_data);
 775         ret = thread_main(td);
 776         shmdt(data);
 777         return (int) ret;
 778 }
 779
 780 /*
 781  * Run over the job map and reap the threads that have exited, if any.
 782  */
 783 static void reap_threads(int *nr_running, int *t_rate, int *m_rate)
 784 {
 785         struct thread_data *td;
 786         int i, cputhreads, pending, status, ret;
 787
 788         /*
 789          * reap exited threads (TD_EXITED -> TD_REAPED)
 790          */
 791         pending = cputhreads = 0;
 792         for_each_td(td, i) {
 793                 /*
 794                  * ->io_ops is NULL for a thread that has closed its
 795                  * io engine
 796                  */
 797                 if (td->io_ops && td->io_ops->flags & FIO_CPUIO)
 798                         cputhreads++;
 799
 800                 if (td->runstate < TD_EXITED) {
 801                         /*
 802                          * check if someone quit or got killed in an unusual way
 803                          */
 804                         ret = waitpid(td->pid, &status, WNOHANG);
 805                         if (ret < 0)
 806                                 perror("waitpid");
 807                         else if ((ret == td->pid) && WIFSIGNALED(status)) {
 808                                 int sig = WTERMSIG(status);
 809
 810                                 log_err("fio: pid=%d, got signal=%d\n", td->pid, sig);
 811                                 td_set_runstate(td, TD_REAPED);
 812                                 goto reaped;
 813                         }
 814                 }
 815
 816                 if (td->runstate != TD_EXITED) {
 817                         if (td->runstate < TD_RUNNING)
 818                                 pending++;
 819
 820                         continue;
 821                 }
 822
 823                 if (td->error)
 824                         exit_value++;
 825
 826                 td_set_runstate(td, TD_REAPED);
 827
 828                 if (td->use_thread) {
 829                         long ret;
 830
 831                         if (pthread_join(td->thread, (void *) &ret))
 832                                 perror("thread_join");
 833                 } else {
 834                         int status;
 835
 836                         ret = waitpid(td->pid, &status, 0);
 837                         if (ret < 0)
 838                                 perror("waitpid");
 839                         else if (WIFEXITED(status) && WEXITSTATUS(status)) {
 840                                 if (!exit_value)
 841                                         exit_value++;
 842                         }
 843                 }
 844
 845 reaped:
 846                 (*nr_running)--;
 847                 (*m_rate) -= td->ratemin;
 848                 (*t_rate) -= td->rate;
 849         }
 850
 851         if (*nr_running == cputhreads && !pending)
 852                 terminate_threads(TERMINATE_ALL, 0);
 853 }
 854
 855 /*
 856  * Main function for kicking off and reaping jobs, as needed.
 857  */
 858 static void run_threads(void)
 859 {
 860         struct thread_data *td;
 861         unsigned long spent;
 862         int i, todo, nr_running, m_rate, t_rate, nr_started;
 863
 864         if (fio_pin_memory())
 865                 return;
 866
 867         if (!terse_output) {
 868                 printf("Starting %d thread%s\n", thread_number, thread_number > 1 ? "s" : "");
 869                 fflush(stdout);
 870         }
 871
 872         signal(SIGINT, sig_handler);
 873         signal(SIGALRM, sig_handler);
 874
 875         todo = thread_number;
 876         nr_running = 0;
 877         nr_started = 0;
 878         m_rate = t_rate = 0;
 879
 880         for_each_td(td, i) {
 881                 print_status_init(td->thread_number - 1);
 882
 883                 if (!td->create_serialize) {
 884                         init_disk_util(td);
 885                         continue;
 886                 }
 887
 888                 /*
 889                  * do file setup here so it happens sequentially,
 890                  * we don't want X number of threads getting their
 891                  * client data interspersed on disk
 892                  */
 893                 if (setup_files(td)) {
 894                         exit_value++;
 895                         if (td->error)
 896                                 log_err("fio: pid=%d, err=%d/%s\n", td->pid, td->error, td->verror);
 897                         td_set_runstate(td, TD_REAPED);
 898                         todo--;
 899                 }
 900
 901                 init_disk_util(td);
 902         }
 903
 904         while (todo) {
 905                 struct thread_data *map[MAX_JOBS];
 906                 struct timeval this_start;
 907                 int this_jobs = 0, left;
 908
 909                 /*
 910                  * create threads (TD_NOT_CREATED -> TD_CREATED)
 911                  */
 912                 for_each_td(td, i) {
 913                         if (td->runstate != TD_NOT_CREATED)
 914                                 continue;
 915
 916                         /*
 917                          * never got a chance to start, killed by other
 918                          * thread for some reason
 919                          */
 920                         if (td->terminate) {
 921                                 todo--;
 922                                 continue;
 923                         }
 924
 925                         if (td->start_delay) {
 926                                 spent = mtime_since_genesis();
 927
 928                                 if (td->start_delay * 1000 > spent)
 929                                         continue;
 930                         }
 931
 932                         if (td->stonewall && (nr_started || nr_running))
 933                                 break;
 934
 935                         /*
 936                          * Set state to created. Thread will transition
 937                          * to TD_INITIALIZED when it's done setting up.
 938                          */
 939                         td_set_runstate(td, TD_CREATED);
 940                         map[this_jobs++] = td;
 941                         fio_sem_init(&startup_sem, 1);
 942                         nr_started++;
 943
 944                         if (td->use_thread) {
 945                                 if (pthread_create(&td->thread, NULL, thread_main, td)) {
 946                                         perror("thread_create");
 947                                         nr_started--;
 948                                 }
 949                         } else {
 950                                 if (fork())
 951                                         fio_sem_down(&startup_sem);
 952                                 else {
 953                                         int ret = fork_main(shm_id, i);
 954
 955                                         exit(ret);
 956                                 }
 957                         }
 958                 }
 959
 960                 /*
 961                  * Wait for the started threads to transition to
 962                  * TD_INITIALIZED.
 963                  */
 964                 fio_gettime(&this_start, NULL);
 965                 left = this_jobs;
 966                 while (left && !fio_abort) {
 967                         if (mtime_since_now(&this_start) > JOB_START_TIMEOUT)
 968                                 break;
 969
 970                         usleep(100000);
 971
 972                         for (i = 0; i < this_jobs; i++) {
 973                                 td = map[i];
 974                                 if (!td)
 975                                         continue;
 976                                 if (td->runstate == TD_INITIALIZED) {
 977                                         map[i] = NULL;
 978                                         left--;
 979                                 } else if (td->runstate >= TD_EXITED) {
 980                                         map[i] = NULL;
 981                                         left--;
 982                                         todo--;
 983                                         nr_running++; /* work-around... */
 984                                 }
 985                         }
 986                 }
 987
 988                 if (left) {
 989                         log_err("fio: %d jobs failed to start\n", left);
 990                         for (i = 0; i < this_jobs; i++) {
 991                                 td = map[i];
 992                                 if (!td)
 993                                         continue;
 994                                 kill(td->pid, SIGTERM);
 995                         }
 996                         break;
 997                 }
 998
 999                 /*
1000                  * start created threads (TD_INITIALIZED -> TD_RUNNING).
1001                  */
1002                 for_each_td(td, i) {
1003                         if (td->runstate != TD_INITIALIZED)
1004                                 continue;
1005
1006                         td_set_runstate(td, TD_RUNNING);
1007                         nr_running++;
1008                         nr_started--;
1009                         m_rate += td->ratemin;
1010                         t_rate += td->rate;
1011                         todo--;
1012                         fio_sem_up(&td->mutex);
1013                 }
1014
1015                 reap_threads(&nr_running, &t_rate, &m_rate);
1016
1017                 if (todo)
1018                         usleep(100000);
1019         }
1020
1021         while (nr_running) {
1022                 reap_threads(&nr_running, &t_rate, &m_rate);
1023                 usleep(10000);
1024         }
1025
1026         update_io_ticks();
1027         fio_unpin_memory();
1028 }
1029
1030 int main(int argc, char *argv[])
1031 {
1032         long ps;
1033
1034         /*
1035          * We need locale for number printing, if it isn't set then just
1036          * go with the US format.
1037          */
1038         if (!getenv("LC_NUMERIC"))
1039                 setlocale(LC_NUMERIC, "en_US");
1040
1041         if (parse_options(argc, argv))
1042                 return 1;
1043
1044         if (!thread_number) {
1045                 log_err("Nothing to do\n");
1046                 return 1;
1047         }
1048
1049         ps = sysconf(_SC_PAGESIZE);
1050         if (ps < 0) {
1051                 log_err("Failed to get page size\n");
1052                 return 1;
1053         }
1054
1055         page_mask = ps - 1;
1056
1057         if (write_bw_log) {
1058                 setup_log(&agg_io_log[DDIR_READ]);
1059                 setup_log(&agg_io_log[DDIR_WRITE]);
1060         }
1061
1062         disk_util_timer_arm();
1063
1064         run_threads();
1065
1066         if (!fio_abort) {
1067                 show_run_stats();
1068                 if (write_bw_log) {
1069                         __finish_log(agg_io_log[DDIR_READ],"agg-read_bw.log");
1070                         __finish_log(agg_io_log[DDIR_WRITE],"agg-write_bw.log");
1071                 }
1072         }
1073
1074         return exit_value;
1075 }