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 /*
 140  * When job exits, we can cancel the in-flight IO if we are using async
 141  * io. Attempt to do so.
 142  */
 143 static void cleanup_pending_aio(struct thread_data *td)
 144 {
 145         struct list_head *entry, *n;
 146         struct io_u *io_u;
 147         int r;
 148
 149         /*
 150          * get immediately available events, if any
 151          */
 152         r = io_u_queued_complete(td, 0, NULL);
 153         if (r < 0)
 154                 return;
 155
 156         /*
 157          * now cancel remaining active events
 158          */
 159         if (td->io_ops->cancel) {
 160                 list_for_each_safe(entry, n, &td->io_u_busylist) {
 161                         io_u = list_entry(entry, struct io_u, list);
 162
 163                         /*
 164                          * if the io_u isn't in flight, then that generally
 165                          * means someone leaked an io_u. complain but fix
 166                          * it up, so we don't stall here.
 167                          */
 168                         if ((io_u->flags & IO_U_F_FLIGHT) == 0) {
 169                                 log_err("fio: non-busy IO on busy list\n");
 170                                 put_io_u(td, io_u);
 171                         } else {
 172                                 r = td->io_ops->cancel(td, io_u);
 173                                 if (!r)
 174                                         put_io_u(td, io_u);
 175                         }
 176                 }
 177         }
 178
 179         if (td->cur_depth)
 180                 r = io_u_queued_complete(td, td->cur_depth, NULL);
 181 }
 182
 183 /*
 184  * Helper to handle the final sync of a file. Works just like the normal
 185  * io path, just does everything sync.
 186  */
 187 static int fio_io_sync(struct thread_data *td, struct fio_file *f)
 188 {
 189         struct io_u *io_u = __get_io_u(td);
 190         int ret;
 191
 192         if (!io_u)
 193                 return 1;
 194
 195         io_u->ddir = DDIR_SYNC;
 196         io_u->file = f;
 197
 198         if (td_io_prep(td, io_u)) {
 199                 put_io_u(td, io_u);
 200                 return 1;
 201         }
 202
 203 requeue:
 204         ret = td_io_queue(td, io_u);
 205         if (ret < 0) {
 206                 td_verror(td, io_u->error);
 207                 put_io_u(td, io_u);
 208                 return 1;
 209         } else if (ret == FIO_Q_QUEUED) {
 210                 if (io_u_queued_complete(td, 1, NULL) < 0)
 211                         return 1;
 212         } else if (ret == FIO_Q_COMPLETED) {
 213                 if (io_u->error) {
 214                         td_verror(td, io_u->error);
 215                         return 1;
 216                 }
 217
 218                 if (io_u_sync_complete(td, io_u, NULL) < 0)
 219                         return 1;
 220         } else if (ret == FIO_Q_BUSY) {
 221                 if (td_io_commit(td))
 222                         return 1;
 223                 goto requeue;
 224         }
 225
 226         return 0;
 227 }
 228
 229 /*
 230  * The main verify engine. Runs over the writes we previusly submitted,
 231  * reads the blocks back in, and checks the crc/md5 of the data.
 232  */
 233 static void do_verify(struct thread_data *td)
 234 {
 235         struct fio_file *f;
 236         struct io_u *io_u;
 237         int ret, i, min_events;
 238
 239         /*
 240          * sync io first and invalidate cache, to make sure we really
 241          * read from disk.
 242          */
 243         for_each_file(td, f, i) {
 244                 if (fio_io_sync(td, f))
 245                         break;
 246                 if (file_invalidate_cache(td, f))
 247                         break;
 248         }
 249
 250         if (td->error)
 251                 return;
 252
 253         td_set_runstate(td, TD_VERIFYING);
 254
 255         io_u = NULL;
 256         while (!td->terminate) {
 257                 io_u = __get_io_u(td);
 258                 if (!io_u)
 259                         break;
 260
 261                 if (runtime_exceeded(td, &io_u->start_time)) {
 262                         put_io_u(td, io_u);
 263                         break;
 264                 }
 265
 266                 if (get_next_verify(td, io_u)) {
 267                         put_io_u(td, io_u);
 268                         break;
 269                 }
 270
 271                 if (td_io_prep(td, io_u)) {
 272                         put_io_u(td, io_u);
 273                         break;
 274                 }
 275 requeue:
 276                 ret = td_io_queue(td, io_u);
 277
 278                 switch (ret) {
 279                 case FIO_Q_COMPLETED:
 280                         if (io_u->error)
 281                                 ret = -io_u->error;
 282                         if (io_u->xfer_buflen != io_u->resid && io_u->resid) {
 283                                 int bytes = io_u->xfer_buflen - io_u->resid;
 284
 285                                 io_u->xfer_buflen = io_u->resid;
 286                                 io_u->xfer_buf += bytes;
 287                                 goto requeue;
 288                         }
 289                         ret = io_u_sync_complete(td, io_u, verify_io_u);
 290                         if (ret < 0)
 291                                 break;
 292                         continue;
 293                 case FIO_Q_QUEUED:
 294                         break;
 295                 case FIO_Q_BUSY:
 296                         requeue_io_u(td, &io_u);
 297                         ret = td_io_commit(td);
 298                         break;
 299                 default:
 300                         assert(ret < 0);
 301                         td_verror(td, -ret);
 302                         break;
 303                 }
 304
 305                 if (ret < 0 || td->error)
 306                         break;
 307
 308                 /*
 309                  * if we can queue more, do so. but check if there are
 310                  * completed io_u's first.
 311                  */
 312                 min_events = 0;
 313                 if (queue_full(td) || ret == FIO_Q_BUSY) {
 314                         min_events = 1;
 315
 316                         if (td->cur_depth > td->iodepth_low)
 317                                 min_events = td->cur_depth - td->iodepth_low;
 318                 }
 319
 320                 /*
 321                  * Reap required number of io units, if any, and do the
 322                  * verification on them through the callback handler
 323                  */
 324                 if (io_u_queued_complete(td, min_events, verify_io_u) < 0)
 325                         break;
 326         }
 327
 328         if (td->cur_depth)
 329                 cleanup_pending_aio(td);
 330
 331         td_set_runstate(td, TD_RUNNING);
 332 }
 333
 334 /*
 335  * Not really an io thread, all it does is burn CPU cycles in the specified
 336  * manner.
 337  */
 338 static void do_cpuio(struct thread_data *td)
 339 {
 340         struct timeval e;
 341         int split = 100 / td->cpuload;
 342         int i = 0;
 343
 344         while (!td->terminate) {
 345                 fio_gettime(&e, NULL);
 346
 347                 if (runtime_exceeded(td, &e))
 348                         break;
 349
 350                 if (!(i % split))
 351                         __usec_sleep(10000);
 352                 else
 353                         usec_sleep(td, 10000);
 354
 355                 i++;
 356         }
 357 }
 358
 359 /*
 360  * Main IO worker function. It retrieves io_u's to process and queues
 361  * and reaps them, checking for rate and errors along the way.
 362  */
 363 static void do_io(struct thread_data *td)
 364 {
 365         struct timeval s;
 366         unsigned long usec;
 367         int i, ret = 0;
 368
 369         td_set_runstate(td, TD_RUNNING);
 370
 371         while ((td->this_io_bytes[0] + td->this_io_bytes[1]) < td->io_size) {
 372                 struct timeval comp_time;
 373                 long bytes_done = 0;
 374                 int min_evts = 0;
 375                 struct io_u *io_u;
 376
 377                 if (td->terminate)
 378                         break;
 379
 380                 io_u = get_io_u(td);
 381                 if (!io_u)
 382                         break;
 383
 384                 memcpy(&s, &io_u->start_time, sizeof(s));
 385
 386                 if (runtime_exceeded(td, &s)) {
 387                         put_io_u(td, io_u);
 388                         break;
 389                 }
 390 requeue:
 391                 ret = td_io_queue(td, io_u);
 392
 393                 switch (ret) {
 394                 case FIO_Q_COMPLETED:
 395                         if (io_u->error) {
 396                                 ret = io_u->error;
 397                                 break;
 398                         }
 399                         if (io_u->xfer_buflen != io_u->resid && io_u->resid) {
 400                                 int bytes = io_u->xfer_buflen - io_u->resid;
 401
 402                                 io_u->xfer_buflen = io_u->resid;
 403                                 io_u->xfer_buf += bytes;
 404                                 goto requeue;
 405                         }
 406                         fio_gettime(&comp_time, NULL);
 407                         bytes_done = io_u_sync_complete(td, io_u, NULL);
 408                         if (bytes_done < 0)
 409                                 ret = bytes_done;
 410                         break;
 411                 case FIO_Q_QUEUED:
 412                         /*
 413                          * if the engine doesn't have a commit hook,
 414                          * the io_u is really queued. if it does have such
 415                          * a hook, it has to call io_u_queued() itself.
 416                          */
 417                         if (td->io_ops->commit == NULL)
 418                                 io_u_queued(td, io_u);
 419                         break;
 420                 case FIO_Q_BUSY:
 421                         requeue_io_u(td, &io_u);
 422                         ret = td_io_commit(td);
 423                         break;
 424                 default:
 425                         assert(ret < 0);
 426                         put_io_u(td, io_u);
 427                         break;
 428                 }
 429
 430                 if (ret < 0 || td->error)
 431                         break;
 432
 433                 /*
 434                  * See if we need to complete some commands
 435                  */
 436                 if (ret == FIO_Q_QUEUED || ret == FIO_Q_BUSY) {
 437                         min_evts = 0;
 438                         if (queue_full(td) || ret == FIO_Q_BUSY) {
 439                                 min_evts = 1;
 440
 441                                 if (td->cur_depth > td->iodepth_low)
 442                                         min_evts = td->cur_depth - td->iodepth_low;
 443                         }
 444
 445                         fio_gettime(&comp_time, NULL);
 446                         bytes_done = io_u_queued_complete(td, min_evts, NULL);
 447                         if (bytes_done < 0)
 448                                 break;
 449                 }
 450
 451                 if (!bytes_done)
 452                         continue;
 453
 454                 /*
 455                  * the rate is batched for now, it should work for batches
 456                  * of completions except the very first one which may look
 457                  * a little bursty
 458                  */
 459                 usec = utime_since(&s, &comp_time);
 460
 461                 rate_throttle(td, usec, bytes_done, td->ddir);
 462
 463                 if (check_min_rate(td, &comp_time)) {
 464                         if (exitall_on_terminate)
 465                                 terminate_threads(td->groupid, 0);
 466                         td_verror(td, ENODATA);
 467                         break;
 468                 }
 469
 470                 if (td->thinktime) {
 471                         unsigned long long b;
 472
 473                         b = td->io_blocks[0] + td->io_blocks[1];
 474                         if (!(b % td->thinktime_blocks)) {
 475                                 int left;
 476
 477                                 if (td->thinktime_spin)
 478                                         __usec_sleep(td->thinktime_spin);
 479
 480                                 left = td->thinktime - td->thinktime_spin;
 481                                 if (left)
 482                                         usec_sleep(td, left);
 483                         }
 484                 }
 485         }
 486
 487         if (!td->error) {
 488                 struct fio_file *f;
 489
 490                 if (td->cur_depth)
 491                         cleanup_pending_aio(td);
 492
 493                 if (should_fsync(td) && td->end_fsync) {
 494                         td_set_runstate(td, TD_FSYNCING);
 495                         for_each_file(td, f, i)
 496                                 fio_io_sync(td, f);
 497                 }
 498         }
 499 }
 500
 501 static void cleanup_io_u(struct thread_data *td)
 502 {
 503         struct list_head *entry, *n;
 504         struct io_u *io_u;
 505
 506         list_for_each_safe(entry, n, &td->io_u_freelist) {
 507                 io_u = list_entry(entry, struct io_u, list);
 508
 509                 list_del(&io_u->list);
 510                 free(io_u);
 511         }
 512
 513         free_io_mem(td);
 514 }
 515
 516 /*
 517  * "randomly" fill the buffer contents
 518  */
 519 static void fill_rand_buf(struct io_u *io_u, int max_bs)
 520 {
 521         int *ptr = io_u->buf;
 522
 523         while ((void *) ptr - io_u->buf < max_bs) {
 524                 *ptr = rand() * 0x9e370001;
 525                 ptr++;
 526         }
 527 }
 528
 529 static int init_io_u(struct thread_data *td)
 530 {
 531         struct io_u *io_u;
 532         unsigned int max_bs;
 533         int i, max_units;
 534         char *p;
 535
 536         if (td->io_ops->flags & FIO_CPUIO)
 537                 return 0;
 538
 539         if (td->io_ops->flags & FIO_SYNCIO)
 540                 max_units = 1;
 541         else
 542                 max_units = td->iodepth;
 543
 544         max_bs = max(td->max_bs[DDIR_READ], td->max_bs[DDIR_WRITE]);
 545         td->orig_buffer_size = max_bs * max_units;
 546
 547         if (td->mem_type == MEM_SHMHUGE || td->mem_type == MEM_MMAPHUGE)
 548                 td->orig_buffer_size = (td->orig_buffer_size + td->hugepage_size - 1) & ~(td->hugepage_size - 1);
 549         else
 550                 td->orig_buffer_size += page_mask;
 551
 552         if (allocate_io_mem(td))
 553                 return 1;
 554
 555         p = ALIGN(td->orig_buffer);
 556         for (i = 0; i < max_units; i++) {
 557                 io_u = malloc(sizeof(*io_u));
 558                 memset(io_u, 0, sizeof(*io_u));
 559                 INIT_LIST_HEAD(&io_u->list);
 560
 561                 io_u->buf = p + max_bs * i;
 562                 if (td_write(td) || td_rw(td))
 563                         fill_rand_buf(io_u, max_bs);
 564
 565                 io_u->index = i;
 566                 io_u->flags = IO_U_F_FREE;
 567                 list_add(&io_u->list, &td->io_u_freelist);
 568         }
 569
 570         return 0;
 571 }
 572
 573 static int switch_ioscheduler(struct thread_data *td)
 574 {
 575         char tmp[256], tmp2[128];
 576         FILE *f;
 577         int ret;
 578
 579         if (td->io_ops->flags & FIO_CPUIO)
 580                 return 0;
 581
 582         sprintf(tmp, "%s/queue/scheduler", td->sysfs_root);
 583
 584         f = fopen(tmp, "r+");
 585         if (!f) {
 586                 td_verror(td, errno);
 587                 return 1;
 588         }
 589
 590         /*
 591          * Set io scheduler.
 592          */
 593         ret = fwrite(td->ioscheduler, strlen(td->ioscheduler), 1, f);
 594         if (ferror(f) || ret != 1) {
 595                 td_verror(td, errno);
 596                 fclose(f);
 597                 return 1;
 598         }
 599
 600         rewind(f);
 601
 602         /*
 603          * Read back and check that the selected scheduler is now the default.
 604          */
 605         ret = fread(tmp, 1, sizeof(tmp), f);
 606         if (ferror(f) || ret < 0) {
 607                 td_verror(td, errno);
 608                 fclose(f);
 609                 return 1;
 610         }
 611
 612         sprintf(tmp2, "[%s]", td->ioscheduler);
 613         if (!strstr(tmp, tmp2)) {
 614                 log_err("fio: io scheduler %s not found\n", td->ioscheduler);
 615                 td_verror(td, EINVAL);
 616                 fclose(f);
 617                 return 1;
 618         }
 619
 620         fclose(f);
 621         return 0;
 622 }
 623
 624 static void clear_io_state(struct thread_data *td)
 625 {
 626         struct fio_file *f;
 627         int i;
 628
 629         td->ts.stat_io_bytes[0] = td->ts.stat_io_bytes[1] = 0;
 630         td->this_io_bytes[0] = td->this_io_bytes[1] = 0;
 631         td->zone_bytes = 0;
 632
 633         td->last_was_sync = 0;
 634
 635         for_each_file(td, f, i) {
 636                 f->last_completed_pos = 0;
 637
 638                 f->last_pos = 0;
 639                 if (td->io_ops->flags & FIO_SYNCIO)
 640                         lseek(f->fd, SEEK_SET, 0);
 641
 642                 if (f->file_map)
 643                         memset(f->file_map, 0, f->num_maps * sizeof(long));
 644         }
 645 }
 646
 647 /*
 648  * Entry point for the thread based jobs. The process based jobs end up
 649  * here as well, after a little setup.
 650  */
 651 static void *thread_main(void *data)
 652 {
 653         unsigned long long runtime[2];
 654         struct thread_data *td = data;
 655
 656         if (!td->use_thread)
 657                 setsid();
 658
 659         td->pid = getpid();
 660
 661         INIT_LIST_HEAD(&td->io_u_freelist);
 662         INIT_LIST_HEAD(&td->io_u_busylist);
 663         INIT_LIST_HEAD(&td->io_u_requeues);
 664         INIT_LIST_HEAD(&td->io_hist_list);
 665         INIT_LIST_HEAD(&td->io_log_list);
 666
 667         if (init_io_u(td))
 668                 goto err;
 669
 670         if (fio_setaffinity(td) == -1) {
 671                 td_verror(td, errno);
 672                 goto err;
 673         }
 674
 675         if (init_iolog(td))
 676                 goto err;
 677
 678         if (td->ioprio) {
 679                 if (ioprio_set(IOPRIO_WHO_PROCESS, 0, td->ioprio) == -1) {
 680                         td_verror(td, errno);
 681                         goto err;
 682                 }
 683         }
 684
 685         if (nice(td->nice) == -1) {
 686                 td_verror(td, errno);
 687                 goto err;
 688         }
 689
 690         if (init_random_state(td))
 691                 goto err;
 692
 693         if (td->ioscheduler && switch_ioscheduler(td))
 694                 goto err;
 695
 696         td_set_runstate(td, TD_INITIALIZED);
 697         fio_sem_up(&startup_sem);
 698         fio_sem_down(&td->mutex);
 699
 700         if (!td->create_serialize && setup_files(td))
 701                 goto err;
 702         if (open_files(td))
 703                 goto err;
 704
 705         /*
 706          * Do this late, as some IO engines would like to have the
 707          * files setup prior to initializing structures.
 708          */
 709         if (td_io_init(td))
 710                 goto err;
 711
 712         if (td->exec_prerun) {
 713                 if (system(td->exec_prerun) < 0)
 714                         goto err;
 715         }
 716
 717         fio_gettime(&td->epoch, NULL);
 718         getrusage(RUSAGE_SELF, &td->ts.ru_start);
 719
 720         runtime[0] = runtime[1] = 0;
 721         while (td->loops--) {
 722                 fio_gettime(&td->start, NULL);
 723                 memcpy(&td->ts.stat_sample_time, &td->start, sizeof(td->start));
 724
 725                 if (td->ratemin)
 726                         memcpy(&td->lastrate, &td->ts.stat_sample_time, sizeof(td->lastrate));
 727
 728                 clear_io_state(td);
 729                 prune_io_piece_log(td);
 730
 731                 if (td->io_ops->flags & FIO_CPUIO)
 732                         do_cpuio(td);
 733                 else
 734                         do_io(td);
 735
 736                 runtime[td->ddir] += utime_since_now(&td->start);
 737                 if (td_rw(td) && td->io_bytes[td->ddir ^ 1])
 738                         runtime[td->ddir ^ 1] = runtime[td->ddir];
 739
 740                 if (td->error || td->terminate)
 741                         break;
 742
 743                 if (td->verify == VERIFY_NONE)
 744                         continue;
 745
 746                 clear_io_state(td);
 747                 fio_gettime(&td->start, NULL);
 748
 749                 do_verify(td);
 750
 751                 runtime[DDIR_READ] += utime_since_now(&td->start);
 752
 753                 if (td->error || td->terminate)
 754                         break;
 755         }
 756
 757         update_rusage_stat(td);
 758         fio_gettime(&td->end_time, NULL);
 759         td->runtime[0] = runtime[0] / 1000;
 760         td->runtime[1] = runtime[1] / 1000;
 761
 762         if (td->ts.bw_log)
 763                 finish_log(td, td->ts.bw_log, "bw");
 764         if (td->ts.slat_log)
 765                 finish_log(td, td->ts.slat_log, "slat");
 766         if (td->ts.clat_log)
 767                 finish_log(td, td->ts.clat_log, "clat");
 768         if (td->write_iolog_file)
 769                 write_iolog_close(td);
 770         if (td->exec_postrun) {
 771                 if (system(td->exec_postrun) < 0)
 772                         log_err("fio: postrun %s failed\n", td->exec_postrun);
 773         }
 774
 775         if (exitall_on_terminate)
 776                 terminate_threads(td->groupid, 0);
 777
 778 err:
 779         if (td->error)
 780                 printf("fio: pid=%d, err=%d/%s\n", td->pid, td->error, td->verror);
 781         close_files(td);
 782         close_ioengine(td);
 783         cleanup_io_u(td);
 784         td_set_runstate(td, TD_EXITED);
 785         return (void *) (unsigned long) td->error;
 786 }
 787
 788 /*
 789  * We cannot pass the td data into a forked process, so attach the td and
 790  * pass it to the thread worker.
 791  */
 792 static int fork_main(int shmid, int offset)
 793 {
 794         struct thread_data *td;
 795         void *data, *ret;
 796
 797         data = shmat(shmid, NULL, 0);
 798         if (data == (void *) -1) {
 799                 int __err = errno;
 800
 801                 perror("shmat");
 802                 return __err;
 803         }
 804
 805         td = data + offset * sizeof(struct thread_data);
 806         ret = thread_main(td);
 807         shmdt(data);
 808         return (int) (unsigned long) ret;
 809 }
 810
 811 /*
 812  * Run over the job map and reap the threads that have exited, if any.
 813  */
 814 static void reap_threads(int *nr_running, int *t_rate, int *m_rate)
 815 {
 816         struct thread_data *td;
 817         int i, cputhreads, pending, status, ret;
 818
 819         /*
 820          * reap exited threads (TD_EXITED -> TD_REAPED)
 821          */
 822         pending = cputhreads = 0;
 823         for_each_td(td, i) {
 824                 /*
 825                  * ->io_ops is NULL for a thread that has closed its
 826                  * io engine
 827                  */
 828                 if (td->io_ops && td->io_ops->flags & FIO_CPUIO)
 829                         cputhreads++;
 830
 831                 if (td->runstate < TD_EXITED) {
 832                         /*
 833                          * check if someone quit or got killed in an unusual way
 834                          */
 835                         ret = waitpid(td->pid, &status, WNOHANG);
 836                         if (ret < 0)
 837                                 perror("waitpid");
 838                         else if ((ret == td->pid) && WIFSIGNALED(status)) {
 839                                 int sig = WTERMSIG(status);
 840
 841                                 log_err("fio: pid=%d, got signal=%d\n", td->pid, sig);
 842                                 td_set_runstate(td, TD_REAPED);
 843                                 goto reaped;
 844                         }
 845                 }
 846
 847                 if (td->runstate != TD_EXITED) {
 848                         if (td->runstate < TD_RUNNING)
 849                                 pending++;
 850
 851                         continue;
 852                 }
 853
 854                 if (td->error)
 855                         exit_value++;
 856
 857                 td_set_runstate(td, TD_REAPED);
 858
 859                 if (td->use_thread) {
 860                         long ret;
 861
 862                         if (pthread_join(td->thread, (void *) &ret))
 863                                 perror("thread_join");
 864                 } else {
 865                         int status;
 866
 867                         ret = waitpid(td->pid, &status, 0);
 868                         if (ret < 0)
 869                                 perror("waitpid");
 870                         else if (WIFEXITED(status) && WEXITSTATUS(status)) {
 871                                 if (!exit_value)
 872                                         exit_value++;
 873                         }
 874                 }
 875
 876 reaped:
 877                 (*nr_running)--;
 878                 (*m_rate) -= td->ratemin;
 879                 (*t_rate) -= td->rate;
 880         }
 881
 882         if (*nr_running == cputhreads && !pending)
 883                 terminate_threads(TERMINATE_ALL, 0);
 884 }
 885
 886 /*
 887  * Main function for kicking off and reaping jobs, as needed.
 888  */
 889 static void run_threads(void)
 890 {
 891         struct thread_data *td;
 892         unsigned long spent;
 893         int i, todo, nr_running, m_rate, t_rate, nr_started;
 894
 895         if (fio_pin_memory())
 896                 return;
 897
 898         if (!terse_output) {
 899                 printf("Starting %d thread%s\n", thread_number, thread_number > 1 ? "s" : "");
 900                 fflush(stdout);
 901         }
 902
 903         signal(SIGINT, sig_handler);
 904         signal(SIGALRM, sig_handler);
 905
 906         todo = thread_number;
 907         nr_running = 0;
 908         nr_started = 0;
 909         m_rate = t_rate = 0;
 910
 911         for_each_td(td, i) {
 912                 print_status_init(td->thread_number - 1);
 913
 914                 if (!td->create_serialize) {
 915                         init_disk_util(td);
 916                         continue;
 917                 }
 918
 919                 /*
 920                  * do file setup here so it happens sequentially,
 921                  * we don't want X number of threads getting their
 922                  * client data interspersed on disk
 923                  */
 924                 if (setup_files(td)) {
 925                         exit_value++;
 926                         if (td->error)
 927                                 log_err("fio: pid=%d, err=%d/%s\n", td->pid, td->error, td->verror);
 928                         td_set_runstate(td, TD_REAPED);
 929                         todo--;
 930                 }
 931
 932                 init_disk_util(td);
 933         }
 934
 935         while (todo) {
 936                 struct thread_data *map[MAX_JOBS];
 937                 struct timeval this_start;
 938                 int this_jobs = 0, left;
 939
 940                 /*
 941                  * create threads (TD_NOT_CREATED -> TD_CREATED)
 942                  */
 943                 for_each_td(td, i) {
 944                         if (td->runstate != TD_NOT_CREATED)
 945                                 continue;
 946
 947                         /*
 948                          * never got a chance to start, killed by other
 949                          * thread for some reason
 950                          */
 951                         if (td->terminate) {
 952                                 todo--;
 953                                 continue;
 954                         }
 955
 956                         if (td->start_delay) {
 957                                 spent = mtime_since_genesis();
 958
 959                                 if (td->start_delay * 1000 > spent)
 960                                         continue;
 961                         }
 962
 963                         if (td->stonewall && (nr_started || nr_running))
 964                                 break;
 965
 966                         /*
 967                          * Set state to created. Thread will transition
 968                          * to TD_INITIALIZED when it's done setting up.
 969                          */
 970                         td_set_runstate(td, TD_CREATED);
 971                         map[this_jobs++] = td;
 972                         fio_sem_init(&startup_sem, 1);
 973                         nr_started++;
 974
 975                         if (td->use_thread) {
 976                                 if (pthread_create(&td->thread, NULL, thread_main, td)) {
 977                                         perror("thread_create");
 978                                         nr_started--;
 979                                 }
 980                         } else {
 981                                 if (fork())
 982                                         fio_sem_down(&startup_sem);
 983                                 else {
 984                                         int ret = fork_main(shm_id, i);
 985
 986                                         exit(ret);
 987                                 }
 988                         }
 989                 }
 990
 991                 /*
 992                  * Wait for the started threads to transition to
 993                  * TD_INITIALIZED.
 994                  */
 995                 fio_gettime(&this_start, NULL);
 996                 left = this_jobs;
 997                 while (left && !fio_abort) {
 998                         if (mtime_since_now(&this_start) > JOB_START_TIMEOUT)
 999                                 break;
1000
1001                         usleep(100000);
1002
1003                         for (i = 0; i < this_jobs; i++) {
1004                                 td = map[i];
1005                                 if (!td)
1006                                         continue;
1007                                 if (td->runstate == TD_INITIALIZED) {
1008                                         map[i] = NULL;
1009                                         left--;
1010                                 } else if (td->runstate >= TD_EXITED) {
1011                                         map[i] = NULL;
1012                                         left--;
1013                                         todo--;
1014                                         nr_running++; /* work-around... */
1015                                 }
1016                         }
1017                 }
1018
1019                 if (left) {
1020                         log_err("fio: %d jobs failed to start\n", left);
1021                         for (i = 0; i < this_jobs; i++) {
1022                                 td = map[i];
1023                                 if (!td)
1024                                         continue;
1025                                 kill(td->pid, SIGTERM);
1026                         }
1027                         break;
1028                 }
1029
1030                 /*
1031                  * start created threads (TD_INITIALIZED -> TD_RUNNING).
1032                  */
1033                 for_each_td(td, i) {
1034                         if (td->runstate != TD_INITIALIZED)
1035                                 continue;
1036
1037                         td_set_runstate(td, TD_RUNNING);
1038                         nr_running++;
1039                         nr_started--;
1040                         m_rate += td->ratemin;
1041                         t_rate += td->rate;
1042                         todo--;
1043                         fio_sem_up(&td->mutex);
1044                 }
1045
1046                 reap_threads(&nr_running, &t_rate, &m_rate);
1047
1048                 if (todo)
1049                         usleep(100000);
1050         }
1051
1052         while (nr_running) {
1053                 reap_threads(&nr_running, &t_rate, &m_rate);
1054                 usleep(10000);
1055         }
1056
1057         update_io_ticks();
1058         fio_unpin_memory();
1059 }
1060
1061 int main(int argc, char *argv[])
1062 {
1063         long ps;
1064
1065         /*
1066          * We need locale for number printing, if it isn't set then just
1067          * go with the US format.
1068          */
1069         if (!getenv("LC_NUMERIC"))
1070                 setlocale(LC_NUMERIC, "en_US");
1071
1072         if (parse_options(argc, argv))
1073                 return 1;
1074
1075         if (!thread_number) {
1076                 log_err("Nothing to do\n");
1077                 return 1;
1078         }
1079
1080         ps = sysconf(_SC_PAGESIZE);
1081         if (ps < 0) {
1082                 log_err("Failed to get page size\n");
1083                 return 1;
1084         }
1085
1086         page_mask = ps - 1;
1087
1088         if (write_bw_log) {
1089                 setup_log(&agg_io_log[DDIR_READ]);
1090                 setup_log(&agg_io_log[DDIR_WRITE]);
1091         }
1092
1093         disk_util_timer_arm();
1094
1095         run_threads();
1096
1097         if (!fio_abort) {
1098                 show_run_stats();
1099                 if (write_bw_log) {
1100                         __finish_log(agg_io_log[DDIR_READ],"agg-read_bw.log");
1101                         __finish_log(agg_io_log[DDIR_WRITE],"agg-write_bw.log");
1102                 }
1103         }
1104
1105         return exit_value;
1106 }