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