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 requeue:
 410                 ret = td_io_queue(td, io_u);
 411                 if (ret) {
 412                         if (ret > 0 && (io_u->xfer_buflen != io_u->resid) &&
 413                             io_u->resid) {
 414                                 /*
 415                                  * short read/write. requeue.
 416                                  */
 417                                 io_u->xfer_buflen = io_u->resid;
 418                                 io_u->xfer_buf += ret;
 419                                 goto requeue;
 420                         } else {
 421                                 td_verror(td, io_u->error);
 422                                 put_io_u(td, io_u);
 423                                 break;
 424                         }
 425                 }
 426
 427                 add_slat_sample(td, io_u->ddir, mtime_since(&io_u->start_time, &io_u->issue_time));
 428
 429                 if (td->cur_depth < td->iodepth) {
 430                         struct timespec ts = { .tv_sec = 0, .tv_nsec = 0};
 431
 432                         timeout = &ts;
 433                         min_evts = 0;
 434                 } else {
 435                         timeout = NULL;
 436                         min_evts = 1;
 437                 }
 438
 439                 ret = td_io_getevents(td, min_evts, td->cur_depth, timeout);
 440                 if (ret < 0) {
 441                         td_verror(td, ret);
 442                         break;
 443                 } else if (!ret)
 444                         continue;
 445
 446                 icd.nr = ret;
 447                 ios_completed(td, &icd);
 448                 if (icd.error) {
 449                         td_verror(td, icd.error);
 450                         break;
 451                 }
 452
 453                 /*
 454                  * the rate is batched for now, it should work for batches
 455                  * of completions except the very first one which may look
 456                  * a little bursty
 457                  */
 458                 usec = utime_since(&s, &icd.time);
 459
 460                 rate_throttle(td, usec, icd.bytes_done[td->ddir], td->ddir);
 461
 462                 if (check_min_rate(td, &icd.time)) {
 463                         if (exitall_on_terminate)
 464                                 terminate_threads(td->groupid, 0);
 465                         td_verror(td, ENODATA);
 466                         break;
 467                 }
 468
 469                 if (runtime_exceeded(td, &icd.time))
 470                         break;
 471
 472                 if (td->thinktime) {
 473                         unsigned long long b;
 474
 475                         b = td->io_blocks[0] + td->io_blocks[1];
 476                         if (!(b % td->thinktime_blocks))
 477                                 usec_sleep(td, td->thinktime);
 478                 }
 479         }
 480
 481         if (!td->error) {
 482                 if (td->cur_depth)
 483                         cleanup_pending_aio(td);
 484
 485                 if (should_fsync(td) && td->end_fsync) {
 486                         td_set_runstate(td, TD_FSYNCING);
 487                         for_each_file(td, f, i)
 488                                 fio_io_sync(td, f);
 489                 }
 490         }
 491 }
 492
 493 static void cleanup_io_u(struct thread_data *td)
 494 {
 495         struct list_head *entry, *n;
 496         struct io_u *io_u;
 497
 498         list_for_each_safe(entry, n, &td->io_u_freelist) {
 499                 io_u = list_entry(entry, struct io_u, list);
 500
 501                 list_del(&io_u->list);
 502                 free(io_u);
 503         }
 504
 505         free_io_mem(td);
 506 }
 507
 508 /*
 509  * "randomly" fill the buffer contents
 510  */
 511 static void fill_rand_buf(struct io_u *io_u, int max_bs)
 512 {
 513         int *ptr = io_u->buf;
 514
 515         while ((void *) ptr - io_u->buf < max_bs) {
 516                 *ptr = rand() * 0x9e370001;
 517                 ptr++;
 518         }
 519 }
 520
 521 static int init_io_u(struct thread_data *td)
 522 {
 523         struct io_u *io_u;
 524         unsigned int max_bs;
 525         int i, max_units;
 526         char *p;
 527
 528         if (td->io_ops->flags & FIO_CPUIO)
 529                 return 0;
 530
 531         if (td->io_ops->flags & FIO_SYNCIO)
 532                 max_units = 1;
 533         else
 534                 max_units = td->iodepth;
 535
 536         max_bs = max(td->max_bs[DDIR_READ], td->max_bs[DDIR_WRITE]);
 537         td->orig_buffer_size = max_bs * max_units;
 538
 539         if (td->mem_type == MEM_SHMHUGE || td->mem_type == MEM_MMAPHUGE)
 540                 td->orig_buffer_size = (td->orig_buffer_size + td->hugepage_size - 1) & ~(td->hugepage_size - 1);
 541         else
 542                 td->orig_buffer_size += page_mask;
 543
 544         if (allocate_io_mem(td))
 545                 return 1;
 546
 547         p = ALIGN(td->orig_buffer);
 548         for (i = 0; i < max_units; i++) {
 549                 io_u = malloc(sizeof(*io_u));
 550                 memset(io_u, 0, sizeof(*io_u));
 551                 INIT_LIST_HEAD(&io_u->list);
 552
 553                 io_u->buf = p + max_bs * i;
 554                 if (td_write(td) || td_rw(td))
 555                         fill_rand_buf(io_u, max_bs);
 556
 557                 io_u->index = i;
 558                 list_add(&io_u->list, &td->io_u_freelist);
 559         }
 560
 561         return 0;
 562 }
 563
 564 static int switch_ioscheduler(struct thread_data *td)
 565 {
 566         char tmp[256], tmp2[128];
 567         FILE *f;
 568         int ret;
 569
 570         if (td->io_ops->flags & FIO_CPUIO)
 571                 return 0;
 572
 573         sprintf(tmp, "%s/queue/scheduler", td->sysfs_root);
 574
 575         f = fopen(tmp, "r+");
 576         if (!f) {
 577                 td_verror(td, errno);
 578                 return 1;
 579         }
 580
 581         /*
 582          * Set io scheduler.
 583          */
 584         ret = fwrite(td->ioscheduler, strlen(td->ioscheduler), 1, f);
 585         if (ferror(f) || ret != 1) {
 586                 td_verror(td, errno);
 587                 fclose(f);
 588                 return 1;
 589         }
 590
 591         rewind(f);
 592
 593         /*
 594          * Read back and check that the selected scheduler is now the default.
 595          */
 596         ret = fread(tmp, 1, sizeof(tmp), f);
 597         if (ferror(f) || ret < 0) {
 598                 td_verror(td, errno);
 599                 fclose(f);
 600                 return 1;
 601         }
 602
 603         sprintf(tmp2, "[%s]", td->ioscheduler);
 604         if (!strstr(tmp, tmp2)) {
 605                 log_err("fio: io scheduler %s not found\n", td->ioscheduler);
 606                 td_verror(td, EINVAL);
 607                 fclose(f);
 608                 return 1;
 609         }
 610
 611         fclose(f);
 612         return 0;
 613 }
 614
 615 static void clear_io_state(struct thread_data *td)
 616 {
 617         struct fio_file *f;
 618         int i;
 619
 620         td->stat_io_bytes[0] = td->stat_io_bytes[1] = 0;
 621         td->this_io_bytes[0] = td->this_io_bytes[1] = 0;
 622         td->zone_bytes = 0;
 623
 624         for_each_file(td, f, i) {
 625                 f->last_pos = 0;
 626                 if (td->io_ops->flags & FIO_SYNCIO)
 627                         lseek(f->fd, SEEK_SET, 0);
 628
 629                 if (f->file_map)
 630                         memset(f->file_map, 0, f->num_maps * sizeof(long));
 631         }
 632 }
 633
 634 /*
 635  * Entry point for the thread based jobs. The process based jobs end up
 636  * here as well, after a little setup.
 637  */
 638 static void *thread_main(void *data)
 639 {
 640         unsigned long long runtime[2];
 641         struct thread_data *td = data;
 642
 643         if (!td->use_thread)
 644                 setsid();
 645
 646         td->pid = getpid();
 647
 648         INIT_LIST_HEAD(&td->io_u_freelist);
 649         INIT_LIST_HEAD(&td->io_u_busylist);
 650         INIT_LIST_HEAD(&td->io_hist_list);
 651         INIT_LIST_HEAD(&td->io_log_list);
 652
 653         if (init_io_u(td))
 654                 goto err;
 655
 656         if (fio_setaffinity(td) == -1) {
 657                 td_verror(td, errno);
 658                 goto err;
 659         }
 660
 661         if (init_iolog(td))
 662                 goto err;
 663
 664         if (td->ioprio) {
 665                 if (ioprio_set(IOPRIO_WHO_PROCESS, 0, td->ioprio) == -1) {
 666                         td_verror(td, errno);
 667                         goto err;
 668                 }
 669         }
 670
 671         if (nice(td->nice) == -1) {
 672                 td_verror(td, errno);
 673                 goto err;
 674         }
 675
 676         if (init_random_state(td))
 677                 goto err;
 678
 679         if (td->ioscheduler && switch_ioscheduler(td))
 680                 goto err;
 681
 682         td_set_runstate(td, TD_INITIALIZED);
 683         fio_sem_up(&startup_sem);
 684         fio_sem_down(&td->mutex);
 685
 686         if (!td->create_serialize && setup_files(td))
 687                 goto err;
 688         if (open_files(td))
 689                 goto err;
 690
 691         /*
 692          * Do this late, as some IO engines would like to have the
 693          * files setup prior to initializing structures.
 694          */
 695         if (td_io_init(td))
 696                 goto err;
 697
 698         if (td->exec_prerun)
 699                 system(td->exec_prerun);
 700
 701         fio_gettime(&td->epoch, NULL);
 702         getrusage(RUSAGE_SELF, &td->ru_start);
 703
 704         runtime[0] = runtime[1] = 0;
 705         while (td->loops--) {
 706                 fio_gettime(&td->start, NULL);
 707                 memcpy(&td->stat_sample_time, &td->start, sizeof(td->start));
 708
 709                 if (td->ratemin)
 710                         memcpy(&td->lastrate, &td->stat_sample_time, sizeof(td->lastrate));
 711
 712                 clear_io_state(td);
 713                 prune_io_piece_log(td);
 714
 715                 if (td->io_ops->flags & FIO_CPUIO)
 716                         do_cpuio(td);
 717                 else
 718                         do_io(td);
 719
 720                 runtime[td->ddir] += utime_since_now(&td->start);
 721                 if (td_rw(td) && td->io_bytes[td->ddir ^ 1])
 722                         runtime[td->ddir ^ 1] = runtime[td->ddir];
 723
 724                 if (td->error || td->terminate)
 725                         break;
 726
 727                 if (td->verify == VERIFY_NONE)
 728                         continue;
 729
 730                 clear_io_state(td);
 731                 fio_gettime(&td->start, NULL);
 732
 733                 do_verify(td);
 734
 735                 runtime[DDIR_READ] += utime_since_now(&td->start);
 736
 737                 if (td->error || td->terminate)
 738                         break;
 739         }
 740
 741         update_rusage_stat(td);
 742         fio_gettime(&td->end_time, NULL);
 743         td->runtime[0] = runtime[0] / 1000;
 744         td->runtime[1] = runtime[1] / 1000;
 745
 746         if (td->bw_log)
 747                 finish_log(td, td->bw_log, "bw");
 748         if (td->slat_log)
 749                 finish_log(td, td->slat_log, "slat");
 750         if (td->clat_log)
 751                 finish_log(td, td->clat_log, "clat");
 752         if (td->write_iolog_file)
 753                 write_iolog_close(td);
 754         if (td->exec_postrun)
 755                 system(td->exec_postrun);
 756
 757         if (exitall_on_terminate)
 758                 terminate_threads(td->groupid, 0);
 759
 760 err:
 761         close_files(td);
 762         close_ioengine(td);
 763         cleanup_io_u(td);
 764         td_set_runstate(td, TD_EXITED);
 765         return NULL;
 766
 767 }
 768
 769 /*
 770  * We cannot pass the td data into a forked process, so attach the td and
 771  * pass it to the thread worker.
 772  */
 773 static void *fork_main(int shmid, int offset)
 774 {
 775         struct thread_data *td;
 776         void *data;
 777
 778         data = shmat(shmid, NULL, 0);
 779         if (data == (void *) -1) {
 780                 perror("shmat");
 781                 return NULL;
 782         }
 783
 784         td = data + offset * sizeof(struct thread_data);
 785         thread_main(td);
 786         shmdt(data);
 787         return NULL;
 788 }
 789
 790 /*
 791  * Run over the job map and reap the threads that have exited, if any.
 792  */
 793 static void reap_threads(int *nr_running, int *t_rate, int *m_rate)
 794 {
 795         struct thread_data *td;
 796         int i, cputhreads, pending;
 797
 798         /*
 799          * reap exited threads (TD_EXITED -> TD_REAPED)
 800          */
 801         pending = cputhreads = 0;
 802         for_each_td(td, i) {
 803                 /*
 804                  * ->io_ops is NULL for a thread that has closed its
 805                  * io engine
 806                  */
 807                 if (td->io_ops && td->io_ops->flags & FIO_CPUIO)
 808                         cputhreads++;
 809
 810                 if (td->runstate != TD_EXITED) {
 811                         if (td->runstate < TD_RUNNING)
 812                                 pending++;
 813
 814                         continue;
 815                 }
 816
 817                 td_set_runstate(td, TD_REAPED);
 818
 819                 if (td->use_thread) {
 820                         long ret;
 821
 822                         if (pthread_join(td->thread, (void *) &ret))
 823                                 perror("thread_join");
 824                 } else
 825                         waitpid(td->pid, NULL, 0);
 826
 827                 (*nr_running)--;
 828                 (*m_rate) -= td->ratemin;
 829                 (*t_rate) -= td->rate;
 830         }
 831
 832         if (*nr_running == cputhreads && !pending)
 833                 terminate_threads(TERMINATE_ALL, 0);
 834 }
 835
 836 /*
 837  * Main function for kicking off and reaping jobs, as needed.
 838  */
 839 static void run_threads(void)
 840 {
 841         struct thread_data *td;
 842         unsigned long spent;
 843         int i, todo, nr_running, m_rate, t_rate, nr_started;
 844
 845         if (fio_pin_memory())
 846                 return;
 847
 848         if (!terse_output) {
 849                 printf("Starting %d thread%s\n", thread_number, thread_number > 1 ? "s" : "");
 850                 fflush(stdout);
 851         }
 852
 853         signal(SIGINT, sig_handler);
 854         signal(SIGALRM, sig_handler);
 855
 856         todo = thread_number;
 857         nr_running = 0;
 858         nr_started = 0;
 859         m_rate = t_rate = 0;
 860
 861         for_each_td(td, i) {
 862                 print_status_init(td->thread_number - 1);
 863
 864                 if (!td->create_serialize) {
 865                         init_disk_util(td);
 866                         continue;
 867                 }
 868
 869                 /*
 870                  * do file setup here so it happens sequentially,
 871                  * we don't want X number of threads getting their
 872                  * client data interspersed on disk
 873                  */
 874                 if (setup_files(td)) {
 875                         td_set_runstate(td, TD_REAPED);
 876                         todo--;
 877                 }
 878
 879                 init_disk_util(td);
 880         }
 881
 882         while (todo) {
 883                 struct thread_data *map[MAX_JOBS];
 884                 struct timeval this_start;
 885                 int this_jobs = 0, left;
 886
 887                 /*
 888                  * create threads (TD_NOT_CREATED -> TD_CREATED)
 889                  */
 890                 for_each_td(td, i) {
 891                         if (td->runstate != TD_NOT_CREATED)
 892                                 continue;
 893
 894                         /*
 895                          * never got a chance to start, killed by other
 896                          * thread for some reason
 897                          */
 898                         if (td->terminate) {
 899                                 todo--;
 900                                 continue;
 901                         }
 902
 903                         if (td->start_delay) {
 904                                 spent = mtime_since_genesis();
 905
 906                                 if (td->start_delay * 1000 > spent)
 907                                         continue;
 908                         }
 909
 910                         if (td->stonewall && (nr_started || nr_running))
 911                                 break;
 912
 913                         /*
 914                          * Set state to created. Thread will transition
 915                          * to TD_INITIALIZED when it's done setting up.
 916                          */
 917                         td_set_runstate(td, TD_CREATED);
 918                         map[this_jobs++] = td;
 919                         fio_sem_init(&startup_sem, 1);
 920                         nr_started++;
 921
 922                         if (td->use_thread) {
 923                                 if (pthread_create(&td->thread, NULL, thread_main, td)) {
 924                                         perror("thread_create");
 925                                         nr_started--;
 926                                 }
 927                         } else {
 928                                 if (fork())
 929                                         fio_sem_down(&startup_sem);
 930                                 else {
 931                                         fork_main(shm_id, i);
 932                                         exit(0);
 933                                 }
 934                         }
 935                 }
 936
 937                 /*
 938                  * Wait for the started threads to transition to
 939                  * TD_INITIALIZED.
 940                  */
 941                 fio_gettime(&this_start, NULL);
 942                 left = this_jobs;
 943                 while (left && !fio_abort) {
 944                         if (mtime_since_now(&this_start) > JOB_START_TIMEOUT)
 945                                 break;
 946
 947                         usleep(100000);
 948
 949                         for (i = 0; i < this_jobs; i++) {
 950                                 td = map[i];
 951                                 if (!td)
 952                                         continue;
 953                                 if (td->runstate == TD_INITIALIZED) {
 954                                         map[i] = NULL;
 955                                         left--;
 956                                 } else if (td->runstate >= TD_EXITED) {
 957                                         map[i] = NULL;
 958                                         left--;
 959                                         todo--;
 960                                         nr_running++; /* work-around... */
 961                                 }
 962                         }
 963                 }
 964
 965                 if (left) {
 966                         log_err("fio: %d jobs failed to start\n", left);
 967                         for (i = 0; i < this_jobs; i++) {
 968                                 td = map[i];
 969                                 if (!td)
 970                                         continue;
 971                                 kill(td->pid, SIGTERM);
 972                         }
 973                         break;
 974                 }
 975
 976                 /*
 977                  * start created threads (TD_INITIALIZED -> TD_RUNNING).
 978                  */
 979                 for_each_td(td, i) {
 980                         if (td->runstate != TD_INITIALIZED)
 981                                 continue;
 982
 983                         td_set_runstate(td, TD_RUNNING);
 984                         nr_running++;
 985                         nr_started--;
 986                         m_rate += td->ratemin;
 987                         t_rate += td->rate;
 988                         todo--;
 989                         fio_sem_up(&td->mutex);
 990                 }
 991
 992                 reap_threads(&nr_running, &t_rate, &m_rate);
 993
 994                 if (todo)
 995                         usleep(100000);
 996         }
 997
 998         while (nr_running) {
 999                 reap_threads(&nr_running, &t_rate, &m_rate);
1000                 usleep(10000);
1001         }
1002
1003         update_io_ticks();
1004         fio_unpin_memory();
1005 }
1006
1007 int main(int argc, char *argv[])
1008 {
1009         long ps;
1010
1011         if (parse_options(argc, argv))
1012                 return 1;
1013
1014         if (!thread_number) {
1015                 log_err("Nothing to do\n");
1016                 return 1;
1017         }
1018
1019         ps = sysconf(_SC_PAGESIZE);
1020         if (ps < 0) {
1021                 log_err("Failed to get page size\n");
1022                 return 1;
1023         }
1024
1025         page_mask = ps - 1;
1026
1027         if (write_bw_log) {
1028                 setup_log(&agg_io_log[DDIR_READ]);
1029                 setup_log(&agg_io_log[DDIR_WRITE]);
1030         }
1031
1032         disk_util_timer_arm();
1033
1034         run_threads();
1035
1036         if (!fio_abort) {
1037                 show_run_stats();
1038                 if (write_bw_log) {
1039                         __finish_log(agg_io_log[DDIR_READ],"agg-read_bw.log");
1040                         __finish_log(agg_io_log[DDIR_WRITE],"agg-write_bw.log");
1041                 }
1042         }
1043
1044         return 0;
1045 }