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 <sys/stat.h>
  31 #include <sys/wait.h>
  32 #include <sys/ipc.h>
  33 #include <sys/shm.h>
  34 #include <sys/mman.h>
  35
  36 #include "fio.h"
  37 #include "os.h"
  38
  39 static unsigned long page_mask;
  40 #define ALIGN(buf)      \
  41         (char *) (((unsigned long) (buf) + page_mask) & ~page_mask)
  42
  43 int groupid = 0;
  44 int thread_number = 0;
  45 int shm_id = 0;
  46 int temp_stall_ts;
  47
  48 static volatile int startup_sem;
  49 static volatile int fio_abort;
  50 static int exit_value;
  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                 default:
  86                         printf("\nfio: terminating on signal %d\n", sig);
  87                         fflush(stdout);
  88                         terminate_threads(TERMINATE_ALL, 0);
  89                         break;
  90         }
  91 }
  92
  93 /*
  94  * Check if we are above the minimum rate given.
  95  */
  96 static int check_min_rate(struct thread_data *td, struct timeval *now)
  97 {
  98         unsigned long spent;
  99         unsigned long rate;
 100         int ddir = td->ddir;
 101
 102         /*
 103          * allow a 2 second settle period in the beginning
 104          */
 105         if (mtime_since(&td->start, now) < 2000)
 106                 return 0;
 107
 108         /*
 109          * if rate blocks is set, sample is running
 110          */
 111         if (td->rate_bytes) {
 112                 spent = mtime_since(&td->lastrate, now);
 113                 if (spent < td->ratecycle)
 114                         return 0;
 115
 116                 rate = (td->this_io_bytes[ddir] - td->rate_bytes) / spent;
 117                 if (rate < td->ratemin) {
 118                         fprintf(f_out, "%s: min rate %u not met, got %luKiB/sec\n", td->name, td->ratemin, rate);
 119                         return 1;
 120                 }
 121         }
 122
 123         td->rate_bytes = td->this_io_bytes[ddir];
 124         memcpy(&td->lastrate, now, sizeof(*now));
 125         return 0;
 126 }
 127
 128 static inline int runtime_exceeded(struct thread_data *td, struct timeval *t)
 129 {
 130         if (!td->timeout)
 131                 return 0;
 132         if (mtime_since(&td->epoch, t) >= td->timeout * 1000)
 133                 return 1;
 134
 135         return 0;
 136 }
 137
 138 static struct fio_file *get_next_file(struct thread_data *td)
 139 {
 140         unsigned int old_next_file = td->next_file;
 141         struct fio_file *f;
 142
 143         do {
 144                 f = &td->files[td->next_file];
 145
 146                 td->next_file++;
 147                 if (td->next_file >= td->nr_files)
 148                         td->next_file = 0;
 149
 150                 if (f->fd != -1)
 151                         break;
 152
 153                 f = NULL;
 154         } while (td->next_file != old_next_file);
 155
 156         return f;
 157 }
 158
 159 /*
 160  * When job exits, we can cancel the in-flight IO if we are using async
 161  * io. Attempt to do so.
 162  */
 163 static void cleanup_pending_aio(struct thread_data *td)
 164 {
 165         struct timespec ts = { .tv_sec = 0, .tv_nsec = 0};
 166         struct list_head *entry, *n;
 167         struct io_completion_data icd;
 168         struct io_u *io_u;
 169         int r;
 170
 171         /*
 172          * get immediately available events, if any
 173          */
 174         r = td_io_getevents(td, 0, td->cur_depth, &ts);
 175         if (r > 0) {
 176                 icd.nr = r;
 177                 ios_completed(td, &icd);
 178         }
 179
 180         /*
 181          * now cancel remaining active events
 182          */
 183         if (td->io_ops->cancel) {
 184                 list_for_each_safe(entry, n, &td->io_u_busylist) {
 185                         io_u = list_entry(entry, struct io_u, list);
 186
 187                         r = td->io_ops->cancel(td, io_u);
 188                         if (!r)
 189                                 put_io_u(td, io_u);
 190                 }
 191         }
 192
 193         if (td->cur_depth) {
 194                 r = td_io_getevents(td, td->cur_depth, td->cur_depth, NULL);
 195                 if (r > 0) {
 196                         icd.nr = r;
 197                         ios_completed(td, &icd);
 198                 }
 199         }
 200 }
 201
 202 /*
 203  * Helper to handle the final sync of a file. Works just like the normal
 204  * io path, just does everything sync.
 205  */
 206 static int fio_io_sync(struct thread_data *td, struct fio_file *f)
 207 {
 208         struct io_u *io_u = __get_io_u(td);
 209         struct io_completion_data icd;
 210         int ret;
 211
 212         if (!io_u)
 213                 return 1;
 214
 215         io_u->ddir = DDIR_SYNC;
 216         io_u->file = f;
 217
 218         if (td_io_prep(td, io_u)) {
 219                 put_io_u(td, io_u);
 220                 return 1;
 221         }
 222
 223         ret = td_io_queue(td, io_u);
 224         if (ret) {
 225                 td_verror(td, io_u->error);
 226                 put_io_u(td, io_u);
 227                 return 1;
 228         }
 229
 230         ret = td_io_getevents(td, 1, td->cur_depth, NULL);
 231         if (ret < 0) {
 232                 td_verror(td, ret);
 233                 return 1;
 234         }
 235
 236         icd.nr = ret;
 237         ios_completed(td, &icd);
 238         if (icd.error) {
 239                 td_verror(td, icd.error);
 240                 return 1;
 241         }
 242
 243         return 0;
 244 }
 245
 246 /*
 247  * The main verify engine. Runs over the writes we previusly submitted,
 248  * reads the blocks back in, and checks the crc/md5 of the data.
 249  */
 250 static void do_verify(struct thread_data *td)
 251 {
 252         struct io_u *io_u, *v_io_u = NULL;
 253         struct io_completion_data icd;
 254         struct fio_file *f;
 255         int ret, i;
 256
 257         /*
 258          * sync io first and invalidate cache, to make sure we really
 259          * read from disk.
 260          */
 261         for_each_file(td, f, i) {
 262                 fio_io_sync(td, f);
 263                 file_invalidate_cache(td, f);
 264         }
 265
 266         td_set_runstate(td, TD_VERIFYING);
 267
 268         do {
 269                 if (td->terminate)
 270                         break;
 271
 272                 io_u = __get_io_u(td);
 273                 if (!io_u)
 274                         break;
 275
 276                 if (runtime_exceeded(td, &io_u->start_time)) {
 277                         put_io_u(td, io_u);
 278                         break;
 279                 }
 280
 281                 if (get_next_verify(td, io_u)) {
 282                         put_io_u(td, io_u);
 283                         break;
 284                 }
 285
 286                 f = get_next_file(td);
 287                 if (!f)
 288                         break;
 289
 290                 io_u->file = f;
 291
 292                 if (td_io_prep(td, io_u)) {
 293                         put_io_u(td, io_u);
 294                         break;
 295                 }
 296
 297                 ret = td_io_queue(td, io_u);
 298                 if (ret) {
 299                         td_verror(td, io_u->error);
 300                         put_io_u(td, io_u);
 301                         break;
 302                 }
 303
 304                 /*
 305                  * we have one pending to verify, do that while
 306                  * we are doing io on the next one
 307                  */
 308                 if (do_io_u_verify(td, &v_io_u))
 309                         break;
 310
 311                 ret = td_io_getevents(td, 1, 1, NULL);
 312                 if (ret != 1) {
 313                         if (ret < 0)
 314                                 td_verror(td, ret);
 315                         break;
 316                 }
 317
 318                 v_io_u = td->io_ops->event(td, 0);
 319                 icd.nr = 1;
 320                 icd.error = 0;
 321                 fio_gettime(&icd.time, NULL);
 322                 io_completed(td, v_io_u, &icd);
 323
 324                 if (icd.error) {
 325                         td_verror(td, icd.error);
 326                         put_io_u(td, v_io_u);
 327                         v_io_u = NULL;
 328                         break;
 329                 }
 330
 331                 /*
 332                  * if we can't submit more io, we need to verify now
 333                  */
 334                 if (queue_full(td) && do_io_u_verify(td, &v_io_u))
 335                         break;
 336
 337         } while (1);
 338
 339         do_io_u_verify(td, &v_io_u);
 340
 341         if (td->cur_depth)
 342                 cleanup_pending_aio(td);
 343
 344         td_set_runstate(td, TD_RUNNING);
 345 }
 346
 347 /*
 348  * Not really an io thread, all it does is burn CPU cycles in the specified
 349  * manner.
 350  */
 351 static void do_cpuio(struct thread_data *td)
 352 {
 353         struct timeval e;
 354         int split = 100 / td->cpuload;
 355         int i = 0;
 356
 357         while (!td->terminate) {
 358                 fio_gettime(&e, NULL);
 359
 360                 if (runtime_exceeded(td, &e))
 361                         break;
 362
 363                 if (!(i % split))
 364                         __usec_sleep(10000);
 365                 else
 366                         usec_sleep(td, 10000);
 367
 368                 i++;
 369         }
 370 }
 371
 372 /*
 373  * Main IO worker function. It retrieves io_u's to process and queues
 374  * and reaps them, checking for rate and errors along the way.
 375  */
 376 static void do_io(struct thread_data *td)
 377 {
 378         struct io_completion_data icd;
 379         struct timeval s;
 380         unsigned long usec;
 381         struct fio_file *f;
 382         int i, ret = 0;
 383
 384         td_set_runstate(td, TD_RUNNING);
 385
 386         while ((td->this_io_bytes[0] + td->this_io_bytes[1]) < td->io_size) {
 387                 struct timespec *timeout;
 388                 int min_evts = 0;
 389                 struct io_u *io_u;
 390
 391                 if (td->terminate)
 392                         break;
 393
 394                 f = get_next_file(td);
 395                 if (!f)
 396                         break;
 397
 398                 io_u = get_io_u(td, f);
 399                 if (!io_u)
 400                         break;
 401
 402                 memcpy(&s, &io_u->start_time, sizeof(s));
 403 requeue:
 404                 ret = td_io_queue(td, io_u);
 405                 if (ret) {
 406                         if (ret > 0 && (io_u->xfer_buflen != io_u->resid) &&
 407                             io_u->resid) {
 408                                 /*
 409                                  * short read/write. requeue.
 410                                  */
 411                                 io_u->xfer_buflen = io_u->resid;
 412                                 io_u->xfer_buf += ret;
 413                                 goto requeue;
 414                         } else {
 415                                 put_io_u(td, io_u);
 416                                 break;
 417                         }
 418                 }
 419
 420                 add_slat_sample(td, io_u->ddir, mtime_since(&io_u->start_time, &io_u->issue_time));
 421
 422                 if (td->cur_depth < td->iodepth) {
 423                         struct timespec ts = { .tv_sec = 0, .tv_nsec = 0};
 424
 425                         timeout = &ts;
 426                         min_evts = 0;
 427                 } else {
 428                         timeout = NULL;
 429                         min_evts = 1;
 430                 }
 431
 432                 ret = td_io_getevents(td, min_evts, td->cur_depth, timeout);
 433                 if (ret < 0) {
 434                         td_verror(td, ret);
 435                         break;
 436                 } else if (!ret)
 437                         continue;
 438
 439                 icd.nr = ret;
 440                 ios_completed(td, &icd);
 441                 if (icd.error) {
 442                         td_verror(td, icd.error);
 443                         break;
 444                 }
 445
 446                 /*
 447                  * the rate is batched for now, it should work for batches
 448                  * of completions except the very first one which may look
 449                  * a little bursty
 450                  */
 451                 usec = utime_since(&s, &icd.time);
 452
 453                 rate_throttle(td, usec, icd.bytes_done[td->ddir], td->ddir);
 454
 455                 if (check_min_rate(td, &icd.time)) {
 456                         if (exitall_on_terminate)
 457                                 terminate_threads(td->groupid, 0);
 458                         td_verror(td, ENODATA);
 459                         break;
 460                 }
 461
 462                 if (runtime_exceeded(td, &icd.time))
 463                         break;
 464
 465                 if (td->thinktime) {
 466                         unsigned long long b;
 467
 468                         b = td->io_blocks[0] + td->io_blocks[1];
 469                         if (!(b % td->thinktime_blocks))
 470                                 usec_sleep(td, td->thinktime);
 471                 }
 472         }
 473
 474         if (!td->error) {
 475                 if (td->cur_depth)
 476                         cleanup_pending_aio(td);
 477
 478                 if (should_fsync(td) && td->end_fsync) {
 479                         td_set_runstate(td, TD_FSYNCING);
 480                         for_each_file(td, f, i)
 481                                 fio_io_sync(td, f);
 482                 }
 483         }
 484 }
 485
 486 static void cleanup_io_u(struct thread_data *td)
 487 {
 488         struct list_head *entry, *n;
 489         struct io_u *io_u;
 490
 491         list_for_each_safe(entry, n, &td->io_u_freelist) {
 492                 io_u = list_entry(entry, struct io_u, list);
 493
 494                 list_del(&io_u->list);
 495                 free(io_u);
 496         }
 497
 498         free_io_mem(td);
 499 }
 500
 501 /*
 502  * "randomly" fill the buffer contents
 503  */
 504 static void fill_rand_buf(struct io_u *io_u, int max_bs)
 505 {
 506         int *ptr = io_u->buf;
 507
 508         while ((void *) ptr - io_u->buf < max_bs) {
 509                 *ptr = rand() * 0x9e370001;
 510                 ptr++;
 511         }
 512 }
 513
 514 static int init_io_u(struct thread_data *td)
 515 {
 516         struct io_u *io_u;
 517         unsigned int max_bs;
 518         int i, max_units;
 519         char *p;
 520
 521         if (td->io_ops->flags & FIO_CPUIO)
 522                 return 0;
 523
 524         if (td->io_ops->flags & FIO_SYNCIO)
 525                 max_units = 1;
 526         else
 527                 max_units = td->iodepth;
 528
 529         max_bs = max(td->max_bs[DDIR_READ], td->max_bs[DDIR_WRITE]);
 530         td->orig_buffer_size = max_bs * max_units;
 531
 532         if (td->mem_type == MEM_SHMHUGE || td->mem_type == MEM_MMAPHUGE)
 533                 td->orig_buffer_size = (td->orig_buffer_size + td->hugepage_size - 1) & ~(td->hugepage_size - 1);
 534         else
 535                 td->orig_buffer_size += page_mask;
 536
 537         if (allocate_io_mem(td))
 538                 return 1;
 539
 540         p = ALIGN(td->orig_buffer);
 541         for (i = 0; i < max_units; i++) {
 542                 io_u = malloc(sizeof(*io_u));
 543                 memset(io_u, 0, sizeof(*io_u));
 544                 INIT_LIST_HEAD(&io_u->list);
 545
 546                 io_u->buf = p + max_bs * i;
 547                 if (td_write(td) || td_rw(td))
 548                         fill_rand_buf(io_u, max_bs);
 549
 550                 io_u->index = i;
 551                 list_add(&io_u->list, &td->io_u_freelist);
 552         }
 553
 554         return 0;
 555 }
 556
 557 static int switch_ioscheduler(struct thread_data *td)
 558 {
 559         char tmp[256], tmp2[128];
 560         FILE *f;
 561         int ret;
 562
 563         if (td->io_ops->flags & FIO_CPUIO)
 564                 return 0;
 565
 566         sprintf(tmp, "%s/queue/scheduler", td->sysfs_root);
 567
 568         f = fopen(tmp, "r+");
 569         if (!f) {
 570                 td_verror(td, errno);
 571                 return 1;
 572         }
 573
 574         /*
 575          * Set io scheduler.
 576          */
 577         ret = fwrite(td->ioscheduler, strlen(td->ioscheduler), 1, f);
 578         if (ferror(f) || ret != 1) {
 579                 td_verror(td, errno);
 580                 fclose(f);
 581                 return 1;
 582         }
 583
 584         rewind(f);
 585
 586         /*
 587          * Read back and check that the selected scheduler is now the default.
 588          */
 589         ret = fread(tmp, 1, sizeof(tmp), f);
 590         if (ferror(f) || ret < 0) {
 591                 td_verror(td, errno);
 592                 fclose(f);
 593                 return 1;
 594         }
 595
 596         sprintf(tmp2, "[%s]", td->ioscheduler);
 597         if (!strstr(tmp, tmp2)) {
 598                 log_err("fio: io scheduler %s not found\n", td->ioscheduler);
 599                 td_verror(td, EINVAL);
 600                 fclose(f);
 601                 return 1;
 602         }
 603
 604         fclose(f);
 605         return 0;
 606 }
 607
 608 static void clear_io_state(struct thread_data *td)
 609 {
 610         struct fio_file *f;
 611         int i;
 612
 613         td->stat_io_bytes[0] = td->stat_io_bytes[1] = 0;
 614         td->this_io_bytes[0] = td->this_io_bytes[1] = 0;
 615         td->zone_bytes = 0;
 616
 617         for_each_file(td, f, i) {
 618                 f->last_pos = 0;
 619                 if (td->io_ops->flags & FIO_SYNCIO)
 620                         lseek(f->fd, SEEK_SET, 0);
 621
 622                 if (f->file_map)
 623                         memset(f->file_map, 0, f->num_maps * sizeof(long));
 624         }
 625 }
 626
 627 /*
 628  * Entry point for the thread based jobs. The process based jobs end up
 629  * here as well, after a little setup.
 630  */
 631 static void *thread_main(void *data)
 632 {
 633         unsigned long long runtime[2];
 634         struct thread_data *td = data;
 635
 636         if (!td->use_thread)
 637                 setsid();
 638
 639         td->pid = getpid();
 640
 641         INIT_LIST_HEAD(&td->io_u_freelist);
 642         INIT_LIST_HEAD(&td->io_u_busylist);
 643         INIT_LIST_HEAD(&td->io_hist_list);
 644         INIT_LIST_HEAD(&td->io_log_list);
 645
 646         if (init_io_u(td))
 647                 goto err;
 648
 649         if (fio_setaffinity(td) == -1) {
 650                 td_verror(td, errno);
 651                 goto err;
 652         }
 653
 654         if (init_iolog(td))
 655                 goto err;
 656
 657         if (td->ioprio) {
 658                 if (ioprio_set(IOPRIO_WHO_PROCESS, 0, td->ioprio) == -1) {
 659                         td_verror(td, errno);
 660                         goto err;
 661                 }
 662         }
 663
 664         if (nice(td->nice) == -1) {
 665                 td_verror(td, errno);
 666                 goto err;
 667         }
 668
 669         if (init_random_state(td))
 670                 goto err;
 671
 672         if (td->ioscheduler && switch_ioscheduler(td))
 673                 goto err;
 674
 675         td_set_runstate(td, TD_INITIALIZED);
 676         fio_sem_up(&startup_sem);
 677         fio_sem_down(&td->mutex);
 678
 679         if (!td->create_serialize && setup_files(td))
 680                 goto err;
 681         if (open_files(td))
 682                 goto err;
 683
 684         /*
 685          * Do this late, as some IO engines would like to have the
 686          * files setup prior to initializing structures.
 687          */
 688         if (td_io_init(td))
 689                 goto err;
 690
 691         if (td->exec_prerun) {
 692                 if (system(td->exec_prerun) < 0)
 693                         goto err;
 694         }
 695
 696         fio_gettime(&td->epoch, NULL);
 697         getrusage(RUSAGE_SELF, &td->ru_start);
 698
 699         runtime[0] = runtime[1] = 0;
 700         while (td->loops--) {
 701                 fio_gettime(&td->start, NULL);
 702                 memcpy(&td->stat_sample_time, &td->start, sizeof(td->start));
 703
 704                 if (td->ratemin)
 705                         memcpy(&td->lastrate, &td->stat_sample_time, sizeof(td->lastrate));
 706
 707                 clear_io_state(td);
 708                 prune_io_piece_log(td);
 709
 710                 if (td->io_ops->flags & FIO_CPUIO)
 711                         do_cpuio(td);
 712                 else
 713                         do_io(td);
 714
 715                 runtime[td->ddir] += utime_since_now(&td->start);
 716                 if (td_rw(td) && td->io_bytes[td->ddir ^ 1])
 717                         runtime[td->ddir ^ 1] = runtime[td->ddir];
 718
 719                 if (td->error || td->terminate)
 720                         break;
 721
 722                 if (td->verify == VERIFY_NONE)
 723                         continue;
 724
 725                 clear_io_state(td);
 726                 fio_gettime(&td->start, NULL);
 727
 728                 do_verify(td);
 729
 730                 runtime[DDIR_READ] += utime_since_now(&td->start);
 731
 732                 if (td->error || td->terminate)
 733                         break;
 734         }
 735
 736         update_rusage_stat(td);
 737         fio_gettime(&td->end_time, NULL);
 738         td->runtime[0] = runtime[0] / 1000;
 739         td->runtime[1] = runtime[1] / 1000;
 740
 741         if (td->bw_log)
 742                 finish_log(td, td->bw_log, "bw");
 743         if (td->slat_log)
 744                 finish_log(td, td->slat_log, "slat");
 745         if (td->clat_log)
 746                 finish_log(td, td->clat_log, "clat");
 747         if (td->write_iolog_file)
 748                 write_iolog_close(td);
 749         if (td->exec_postrun) {
 750                 if (system(td->exec_postrun) < 0)
 751                         log_err("fio: postrun %s failed\n", td->exec_postrun);
 752         }
 753
 754         if (exitall_on_terminate)
 755                 terminate_threads(td->groupid, 0);
 756
 757 err:
 758         if (td->error)
 759                 printf("fio: pid=%d, err=%d/%s\n", td->pid, td->error, td->verror);
 760         close_files(td);
 761         close_ioengine(td);
 762         cleanup_io_u(td);
 763         td_set_runstate(td, TD_EXITED);
 764         return NULL;
 765 }
 766
 767 /*
 768  * We cannot pass the td data into a forked process, so attach the td and
 769  * pass it to the thread worker.
 770  */
 771 static void *fork_main(int shmid, int offset)
 772 {
 773         struct thread_data *td;
 774         void *data;
 775
 776         data = shmat(shmid, NULL, 0);
 777         if (data == (void *) -1) {
 778                 perror("shmat");
 779                 return NULL;
 780         }
 781
 782         td = data + offset * sizeof(struct thread_data);
 783         thread_main(td);
 784         shmdt(data);
 785         return NULL;
 786 }
 787
 788 /*
 789  * Run over the job map and reap the threads that have exited, if any.
 790  */
 791 static void reap_threads(int *nr_running, int *t_rate, int *m_rate)
 792 {
 793         struct thread_data *td;
 794         int i, cputhreads, pending;
 795
 796         /*
 797          * reap exited threads (TD_EXITED -> TD_REAPED)
 798          */
 799         pending = cputhreads = 0;
 800         for_each_td(td, i) {
 801                 /*
 802                  * ->io_ops is NULL for a thread that has closed its
 803                  * io engine
 804                  */
 805                 if (td->io_ops && td->io_ops->flags & FIO_CPUIO)
 806                         cputhreads++;
 807
 808                 if (td->runstate != TD_EXITED) {
 809                         if (td->runstate < TD_RUNNING)
 810                                 pending++;
 811
 812                         continue;
 813                 }
 814
 815                 if (td->error)
 816                         exit_value++;
 817
 818                 td_set_runstate(td, TD_REAPED);
 819
 820                 if (td->use_thread) {
 821                         long ret;
 822
 823                         if (pthread_join(td->thread, (void *) &ret))
 824                                 perror("thread_join");
 825                 } else
 826                         waitpid(td->pid, NULL, 0);
 827
 828                 (*nr_running)--;
 829                 (*m_rate) -= td->ratemin;
 830                 (*t_rate) -= td->rate;
 831         }
 832
 833         if (*nr_running == cputhreads && !pending)
 834                 terminate_threads(TERMINATE_ALL, 0);
 835 }
 836
 837 /*
 838  * Main function for kicking off and reaping jobs, as needed.
 839  */
 840 static void run_threads(void)
 841 {
 842         struct thread_data *td;
 843         unsigned long spent;
 844         int i, todo, nr_running, m_rate, t_rate, nr_started;
 845
 846         if (fio_pin_memory())
 847                 return;
 848
 849         if (!terse_output) {
 850                 printf("Starting %d thread%s\n", thread_number, thread_number > 1 ? "s" : "");
 851                 fflush(stdout);
 852         }
 853
 854         signal(SIGINT, sig_handler);
 855         signal(SIGALRM, sig_handler);
 856
 857         todo = thread_number;
 858         nr_running = 0;
 859         nr_started = 0;
 860         m_rate = t_rate = 0;
 861
 862         for_each_td(td, i) {
 863                 print_status_init(td->thread_number - 1);
 864
 865                 if (!td->create_serialize) {
 866                         init_disk_util(td);
 867                         continue;
 868                 }
 869
 870                 /*
 871                  * do file setup here so it happens sequentially,
 872                  * we don't want X number of threads getting their
 873                  * client data interspersed on disk
 874                  */
 875                 if (setup_files(td)) {
 876                         exit_value++;
 877                         if (td->error)
 878                                 log_err("fio: pid=%d, err=%d/%s\n", td->pid, td->error, td->verror);
 879                         td_set_runstate(td, TD_REAPED);
 880                         todo--;
 881                 }
 882
 883                 init_disk_util(td);
 884         }
 885
 886         while (todo) {
 887                 struct thread_data *map[MAX_JOBS];
 888                 struct timeval this_start;
 889                 int this_jobs = 0, left;
 890
 891                 /*
 892                  * create threads (TD_NOT_CREATED -> TD_CREATED)
 893                  */
 894                 for_each_td(td, i) {
 895                         if (td->runstate != TD_NOT_CREATED)
 896                                 continue;
 897
 898                         /*
 899                          * never got a chance to start, killed by other
 900                          * thread for some reason
 901                          */
 902                         if (td->terminate) {
 903                                 todo--;
 904                                 continue;
 905                         }
 906
 907                         if (td->start_delay) {
 908                                 spent = mtime_since_genesis();
 909
 910                                 if (td->start_delay * 1000 > spent)
 911                                         continue;
 912                         }
 913
 914                         if (td->stonewall && (nr_started || nr_running))
 915                                 break;
 916
 917                         /*
 918                          * Set state to created. Thread will transition
 919                          * to TD_INITIALIZED when it's done setting up.
 920                          */
 921                         td_set_runstate(td, TD_CREATED);
 922                         map[this_jobs++] = td;
 923                         fio_sem_init(&startup_sem, 1);
 924                         nr_started++;
 925
 926                         if (td->use_thread) {
 927                                 if (pthread_create(&td->thread, NULL, thread_main, td)) {
 928                                         perror("thread_create");
 929                                         nr_started--;
 930                                 }
 931                         } else {
 932                                 if (fork())
 933                                         fio_sem_down(&startup_sem);
 934                                 else {
 935                                         fork_main(shm_id, i);
 936                                         exit(0);
 937                                 }
 938                         }
 939                 }
 940
 941                 /*
 942                  * Wait for the started threads to transition to
 943                  * TD_INITIALIZED.
 944                  */
 945                 fio_gettime(&this_start, NULL);
 946                 left = this_jobs;
 947                 while (left && !fio_abort) {
 948                         if (mtime_since_now(&this_start) > JOB_START_TIMEOUT)
 949                                 break;
 950
 951                         usleep(100000);
 952
 953                         for (i = 0; i < this_jobs; i++) {
 954                                 td = map[i];
 955                                 if (!td)
 956                                         continue;
 957                                 if (td->runstate == TD_INITIALIZED) {
 958                                         map[i] = NULL;
 959                                         left--;
 960                                 } else if (td->runstate >= TD_EXITED) {
 961                                         map[i] = NULL;
 962                                         left--;
 963                                         todo--;
 964                                         nr_running++; /* work-around... */
 965                                 }
 966                         }
 967                 }
 968
 969                 if (left) {
 970                         log_err("fio: %d jobs failed to start\n", left);
 971                         for (i = 0; i < this_jobs; i++) {
 972                                 td = map[i];
 973                                 if (!td)
 974                                         continue;
 975                                 kill(td->pid, SIGTERM);
 976                         }
 977                         break;
 978                 }
 979
 980                 /*
 981                  * start created threads (TD_INITIALIZED -> TD_RUNNING).
 982                  */
 983                 for_each_td(td, i) {
 984                         if (td->runstate != TD_INITIALIZED)
 985                                 continue;
 986
 987                         td_set_runstate(td, TD_RUNNING);
 988                         nr_running++;
 989                         nr_started--;
 990                         m_rate += td->ratemin;
 991                         t_rate += td->rate;
 992                         todo--;
 993                         fio_sem_up(&td->mutex);
 994                 }
 995
 996                 reap_threads(&nr_running, &t_rate, &m_rate);
 997
 998                 if (todo)
 999                         usleep(100000);
1000         }
1001
1002         while (nr_running) {
1003                 reap_threads(&nr_running, &t_rate, &m_rate);
1004                 usleep(10000);
1005         }
1006
1007         update_io_ticks();
1008         fio_unpin_memory();
1009 }
1010
1011 int main(int argc, char *argv[])
1012 {
1013         long ps;
1014
1015         /*
1016          * We need locale for number printing, if it isn't set then just
1017          * go with the US format.
1018          */
1019         if (!getenv("LC_NUMERIC"))
1020                 setlocale(LC_NUMERIC, "en_US");
1021
1022         if (parse_options(argc, argv))
1023                 return 1;
1024
1025         if (!thread_number) {
1026                 log_err("Nothing to do\n");
1027                 return 1;
1028         }
1029
1030         ps = sysconf(_SC_PAGESIZE);
1031         if (ps < 0) {
1032                 log_err("Failed to get page size\n");
1033                 return 1;
1034         }
1035
1036         page_mask = ps - 1;
1037
1038         if (write_bw_log) {
1039                 setup_log(&agg_io_log[DDIR_READ]);
1040                 setup_log(&agg_io_log[DDIR_WRITE]);
1041         }
1042
1043         disk_util_timer_arm();
1044
1045         run_threads();
1046
1047         if (!fio_abort) {
1048                 show_run_stats();
1049                 if (write_bw_log) {
1050                         __finish_log(agg_io_log[DDIR_READ],"agg-read_bw.log");
1051                         __finish_log(agg_io_log[DDIR_WRITE],"agg-write_bw.log");
1052                 }
1053         }
1054
1055         return exit_value;
1056 }