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 "hash.h"
  39 #include "smalloc.h"
  40 #include "verify.h"
  41 #include "diskutil.h"
  42
  43 unsigned long page_mask;
  44 unsigned long page_size;
  45 #define ALIGN(buf)      \
  46         (char *) (((unsigned long) (buf) + page_mask) & ~page_mask)
  47
  48 int groupid = 0;
  49 int thread_number = 0;
  50 int nr_process = 0;
  51 int nr_thread = 0;
  52 int shm_id = 0;
  53 int temp_stall_ts;
  54 unsigned long done_secs = 0;
  55
  56 static struct fio_mutex *startup_mutex;
  57 static struct fio_mutex *writeout_mutex;
  58 static volatile int fio_abort;
  59 static int exit_value;
  60 static struct itimerval itimer;
  61 static pthread_t gtod_thread;
  62
  63 struct io_log *agg_io_log[2];
  64
  65 #define TERMINATE_ALL           (-1)
  66 #define JOB_START_TIMEOUT       (5 * 1000)
  67
  68 void td_set_runstate(struct thread_data *td, int runstate)
  69 {
  70         if (td->runstate == runstate)
  71                 return;
  72
  73         dprint(FD_PROCESS, "pid=%d: runstate %d -> %d\n", (int) td->pid,
  74                                                 td->runstate, runstate);
  75         td->runstate = runstate;
  76 }
  77
  78 static void terminate_threads(int group_id)
  79 {
  80         struct thread_data *td;
  81         int i;
  82
  83         dprint(FD_PROCESS, "terminate group_id=%d\n", group_id);
  84
  85         for_each_td(td, i) {
  86                 if (group_id == TERMINATE_ALL || groupid == td->groupid) {
  87                         dprint(FD_PROCESS, "setting terminate on %s/%d\n",
  88                                                 td->o.name, (int) td->pid);
  89                         td->terminate = 1;
  90                         td->o.start_delay = 0;
  91
  92                         /*
  93                          * if the thread is running, just let it exit
  94                          */
  95                         if (td->runstate < TD_RUNNING)
  96                                 kill(td->pid, SIGQUIT);
  97                         else {
  98                                 struct ioengine_ops *ops = td->io_ops;
  99
 100                                 if (ops && (ops->flags & FIO_SIGQUIT))
 101                                         kill(td->pid, SIGQUIT);
 102                         }
 103                 }
 104         }
 105 }
 106
 107 static void status_timer_arm(void)
 108 {
 109         itimer.it_value.tv_sec = 0;
 110         itimer.it_value.tv_usec = DISK_UTIL_MSEC * 1000;
 111         setitimer(ITIMER_REAL, &itimer, NULL);
 112 }
 113
 114 static void sig_alrm(int fio_unused sig)
 115 {
 116         if (threads) {
 117                 update_io_ticks();
 118                 print_thread_status();
 119                 status_timer_arm();
 120         }
 121 }
 122
 123 /*
 124  * Happens on thread runs with ctrl-c, ignore our own SIGQUIT
 125  */
 126 static void sig_quit(int sig)
 127 {
 128 }
 129
 130 static void sig_int(int sig)
 131 {
 132         if (threads) {
 133                 printf("\nfio: terminating on signal %d\n", sig);
 134                 fflush(stdout);
 135                 terminate_threads(TERMINATE_ALL);
 136         }
 137 }
 138
 139 static void sig_ill(int fio_unused sig)
 140 {
 141         if (!threads)
 142                 return;
 143
 144         log_err("fio: illegal instruction. your cpu does not support "
 145                 "the sse4.2 instruction for crc32c\n");
 146         terminate_threads(TERMINATE_ALL);
 147         exit(4);
 148 }
 149
 150 static void set_sig_handlers(void)
 151 {
 152         struct sigaction act;
 153
 154         memset(&act, 0, sizeof(act));
 155         act.sa_handler = sig_alrm;
 156         act.sa_flags = SA_RESTART;
 157         sigaction(SIGALRM, &act, NULL);
 158
 159         memset(&act, 0, sizeof(act));
 160         act.sa_handler = sig_int;
 161         act.sa_flags = SA_RESTART;
 162         sigaction(SIGINT, &act, NULL);
 163
 164         memset(&act, 0, sizeof(act));
 165         act.sa_handler = sig_ill;
 166         act.sa_flags = SA_RESTART;
 167         sigaction(SIGILL, &act, NULL);
 168
 169         memset(&act, 0, sizeof(act));
 170         act.sa_handler = sig_quit;
 171         act.sa_flags = SA_RESTART;
 172         sigaction(SIGQUIT, &act, NULL);
 173 }
 174
 175 /*
 176  * Check if we are above the minimum rate given.
 177  */
 178 static int __check_min_rate(struct thread_data *td, struct timeval *now,
 179                             enum td_ddir ddir)
 180 {
 181         unsigned long long bytes = 0;
 182         unsigned long iops = 0;
 183         unsigned long spent;
 184         unsigned long rate;
 185         unsigned int ratemin = 0;
 186         unsigned int rate_iops = 0;
 187         unsigned int rate_iops_min = 0;
 188
 189         if (!td->o.ratemin[ddir] && !td->o.rate_iops_min[ddir])
 190                 return 0;
 191
 192         /*
 193          * allow a 2 second settle period in the beginning
 194          */
 195         if (mtime_since(&td->start, now) < 2000)
 196                 return 0;
 197
 198         iops += td->io_blocks[ddir];
 199         bytes += td->this_io_bytes[ddir];
 200         ratemin += td->o.ratemin[ddir];
 201         rate_iops += td->o.rate_iops[ddir];
 202         rate_iops_min += td->o.rate_iops_min[ddir];
 203
 204         /*
 205          * if rate blocks is set, sample is running
 206          */
 207         if (td->rate_bytes[ddir] || td->rate_blocks[ddir]) {
 208                 spent = mtime_since(&td->lastrate[ddir], now);
 209                 if (spent < td->o.ratecycle)
 210                         return 0;
 211
 212                 if (td->o.rate[ddir]) {
 213                         /*
 214                          * check bandwidth specified rate
 215                          */
 216                         if (bytes < td->rate_bytes[ddir]) {
 217                                 log_err("%s: min rate %u not met\n", td->o.name,
 218                                                                 ratemin);
 219                                 return 1;
 220                         } else {
 221                                 rate = ((bytes - td->rate_bytes[ddir]) * 1000) / spent;
 222                                 if (rate < ratemin ||
 223                                     bytes < td->rate_bytes[ddir]) {
 224                                         log_err("%s: min rate %u not met, got"
 225                                                 " %luKiB/sec\n", td->o.name,
 226                                                         ratemin, rate);
 227                                         return 1;
 228                                 }
 229                         }
 230                 } else {
 231                         /*
 232                          * checks iops specified rate
 233                          */
 234                         if (iops < rate_iops) {
 235                                 log_err("%s: min iops rate %u not met\n",
 236                                                 td->o.name, rate_iops);
 237                                 return 1;
 238                         } else {
 239                                 rate = ((iops - td->rate_blocks[ddir]) * 1000) / spent;
 240                                 if (rate < rate_iops_min ||
 241                                     iops < td->rate_blocks[ddir]) {
 242                                         log_err("%s: min iops rate %u not met,"
 243                                                 " got %lu\n", td->o.name,
 244                                                         rate_iops_min, rate);
 245                                 }
 246                         }
 247                 }
 248         }
 249
 250         td->rate_bytes[ddir] = bytes;
 251         td->rate_blocks[ddir] = iops;
 252         memcpy(&td->lastrate[ddir], now, sizeof(*now));
 253         return 0;
 254 }
 255
 256 static int check_min_rate(struct thread_data *td, struct timeval *now,
 257                           unsigned long *bytes_done)
 258 {
 259         int ret = 0;
 260
 261         if (bytes_done[0])
 262                 ret |= __check_min_rate(td, now, 0);
 263         if (bytes_done[1])
 264                 ret |= __check_min_rate(td, now, 1);
 265
 266         return ret;
 267 }
 268
 269 static inline int runtime_exceeded(struct thread_data *td, struct timeval *t)
 270 {
 271         if (!td->o.timeout)
 272                 return 0;
 273         if (mtime_since(&td->epoch, t) >= td->o.timeout * 1000)
 274                 return 1;
 275
 276         return 0;
 277 }
 278
 279 /*
 280  * When job exits, we can cancel the in-flight IO if we are using async
 281  * io. Attempt to do so.
 282  */
 283 static void cleanup_pending_aio(struct thread_data *td)
 284 {
 285         struct flist_head *entry, *n;
 286         struct io_u *io_u;
 287         int r;
 288
 289         /*
 290          * get immediately available events, if any
 291          */
 292         r = io_u_queued_complete(td, 0, NULL);
 293         if (r < 0)
 294                 return;
 295
 296         /*
 297          * now cancel remaining active events
 298          */
 299         if (td->io_ops->cancel) {
 300                 flist_for_each_safe(entry, n, &td->io_u_busylist) {
 301                         io_u = flist_entry(entry, struct io_u, list);
 302
 303                         /*
 304                          * if the io_u isn't in flight, then that generally
 305                          * means someone leaked an io_u. complain but fix
 306                          * it up, so we don't stall here.
 307                          */
 308                         if ((io_u->flags & IO_U_F_FLIGHT) == 0) {
 309                                 log_err("fio: non-busy IO on busy list\n");
 310                                 put_io_u(td, io_u);
 311                         } else {
 312                                 r = td->io_ops->cancel(td, io_u);
 313                                 if (!r)
 314                                         put_io_u(td, io_u);
 315                         }
 316                 }
 317         }
 318
 319         if (td->cur_depth)
 320                 r = io_u_queued_complete(td, td->cur_depth, NULL);
 321 }
 322
 323 /*
 324  * Helper to handle the final sync of a file. Works just like the normal
 325  * io path, just does everything sync.
 326  */
 327 static int fio_io_sync(struct thread_data *td, struct fio_file *f)
 328 {
 329         struct io_u *io_u = __get_io_u(td);
 330         int ret;
 331
 332         if (!io_u)
 333                 return 1;
 334
 335         io_u->ddir = DDIR_SYNC;
 336         io_u->file = f;
 337
 338         if (td_io_prep(td, io_u)) {
 339                 put_io_u(td, io_u);
 340                 return 1;
 341         }
 342
 343 requeue:
 344         ret = td_io_queue(td, io_u);
 345         if (ret < 0) {
 346                 td_verror(td, io_u->error, "td_io_queue");
 347                 put_io_u(td, io_u);
 348                 return 1;
 349         } else if (ret == FIO_Q_QUEUED) {
 350                 if (io_u_queued_complete(td, 1, NULL) < 0)
 351                         return 1;
 352         } else if (ret == FIO_Q_COMPLETED) {
 353                 if (io_u->error) {
 354                         td_verror(td, io_u->error, "td_io_queue");
 355                         return 1;
 356                 }
 357
 358                 if (io_u_sync_complete(td, io_u, NULL) < 0)
 359                         return 1;
 360         } else if (ret == FIO_Q_BUSY) {
 361                 if (td_io_commit(td))
 362                         return 1;
 363                 goto requeue;
 364         }
 365
 366         return 0;
 367 }
 368
 369 static inline void update_tv_cache(struct thread_data *td)
 370 {
 371         if ((++td->tv_cache_nr & td->tv_cache_mask) == td->tv_cache_mask)
 372                 fio_gettime(&td->tv_cache, NULL);
 373 }
 374
 375 static int break_on_this_error(struct thread_data *td, int *retptr)
 376 {
 377         int ret = *retptr;
 378
 379         if (ret < 0 || td->error) {
 380                 int err;
 381
 382                 if (!td->o.continue_on_error)
 383                         return 1;
 384
 385                 if (ret < 0)
 386                         err = -ret;
 387                 else
 388                         err = td->error;
 389
 390                 update_error_count(td, err);
 391
 392                 if (td_non_fatal_error(err)) {
 393                         /*
 394                          * Continue with the I/Os in case of
 395                          * a non fatal error.
 396                          */
 397                         td_clear_error(td);
 398                         *retptr = 0;
 399                         return 0;
 400                 } else {
 401                         /*
 402                          * Stop the I/O in case of a fatal
 403                          * error.
 404                          */
 405                         return 1;
 406                 }
 407         }
 408
 409         return 0;
 410 }
 411
 412 /*
 413  * The main verify engine. Runs over the writes we previously submitted,
 414  * reads the blocks back in, and checks the crc/md5 of the data.
 415  */
 416 static void do_verify(struct thread_data *td)
 417 {
 418         struct fio_file *f;
 419         struct io_u *io_u;
 420         int ret, min_events;
 421         unsigned int i;
 422
 423         /*
 424          * sync io first and invalidate cache, to make sure we really
 425          * read from disk.
 426          */
 427         for_each_file(td, f, i) {
 428                 if (!fio_file_open(f))
 429                         continue;
 430                 if (fio_io_sync(td, f))
 431                         break;
 432                 if (file_invalidate_cache(td, f))
 433                         break;
 434         }
 435
 436         if (td->error)
 437                 return;
 438
 439         td_set_runstate(td, TD_VERIFYING);
 440
 441         io_u = NULL;
 442         while (!td->terminate) {
 443                 int ret2, full;
 444
 445                 update_tv_cache(td);
 446
 447                 if (runtime_exceeded(td, &td->tv_cache)) {
 448                         td->terminate = 1;
 449                         break;
 450                 }
 451
 452                 io_u = __get_io_u(td);
 453                 if (!io_u)
 454                         break;
 455
 456                 if (get_next_verify(td, io_u)) {
 457                         put_io_u(td, io_u);
 458                         break;
 459                 }
 460
 461                 if (td_io_prep(td, io_u)) {
 462                         put_io_u(td, io_u);
 463                         break;
 464                 }
 465
 466                 io_u->end_io = verify_io_u;
 467
 468                 ret = td_io_queue(td, io_u);
 469                 switch (ret) {
 470                 case FIO_Q_COMPLETED:
 471                         if (io_u->error) {
 472                                 ret = -io_u->error;
 473                                 clear_io_u(td, io_u);
 474                         } else if (io_u->resid) {
 475                                 int bytes = io_u->xfer_buflen - io_u->resid;
 476                                 struct fio_file *f = io_u->file;
 477
 478                                 /*
 479                                  * zero read, fail
 480                                  */
 481                                 if (!bytes) {
 482                                         td_verror(td, EIO, "full resid");
 483                                         put_io_u(td, io_u);
 484                                         break;
 485                                 }
 486
 487                                 io_u->xfer_buflen = io_u->resid;
 488                                 io_u->xfer_buf += bytes;
 489                                 io_u->offset += bytes;
 490
 491                                 td->ts.short_io_u[io_u->ddir]++;
 492
 493                                 if (io_u->offset == f->real_file_size)
 494                                         goto sync_done;
 495
 496                                 requeue_io_u(td, &io_u);
 497                         } else {
 498 sync_done:
 499                                 ret = io_u_sync_complete(td, io_u, NULL);
 500                                 if (ret < 0)
 501                                         break;
 502                         }
 503                         continue;
 504                 case FIO_Q_QUEUED:
 505                         break;
 506                 case FIO_Q_BUSY:
 507                         requeue_io_u(td, &io_u);
 508                         ret2 = td_io_commit(td);
 509                         if (ret2 < 0)
 510                                 ret = ret2;
 511                         break;
 512                 default:
 513                         assert(ret < 0);
 514                         td_verror(td, -ret, "td_io_queue");
 515                         break;
 516                 }
 517
 518                 if (break_on_this_error(td, &ret))
 519                         break;
 520
 521                 /*
 522                  * if we can queue more, do so. but check if there are
 523                  * completed io_u's first.
 524                  */
 525                 full = queue_full(td) || ret == FIO_Q_BUSY;
 526                 if (full || !td->o.iodepth_batch_complete) {
 527                         min_events = td->o.iodepth_batch_complete;
 528                         if (full && !min_events)
 529                                 min_events = 1;
 530
 531                         do {
 532                                 /*
 533                                  * Reap required number of io units, if any,
 534                                  * and do the verification on them through
 535                                  * the callback handler
 536                                  */
 537                                 if (io_u_queued_complete(td, min_events, NULL) < 0) {
 538                                         ret = -1;
 539                                         break;
 540                                 }
 541                         } while (full && (td->cur_depth > td->o.iodepth_low));
 542                 }
 543                 if (ret < 0)
 544                         break;
 545         }
 546
 547         if (!td->error) {
 548                 min_events = td->cur_depth;
 549
 550                 if (min_events)
 551                         ret = io_u_queued_complete(td, min_events, NULL);
 552         } else
 553                 cleanup_pending_aio(td);
 554
 555         td_set_runstate(td, TD_RUNNING);
 556 }
 557
 558 /*
 559  * Main IO worker function. It retrieves io_u's to process and queues
 560  * and reaps them, checking for rate and errors along the way.
 561  */
 562 static void do_io(struct thread_data *td)
 563 {
 564         unsigned int i;
 565         int ret = 0;
 566
 567         if (in_ramp_time(td))
 568                 td_set_runstate(td, TD_RAMP);
 569         else
 570                 td_set_runstate(td, TD_RUNNING);
 571
 572         while ((td->this_io_bytes[0] + td->this_io_bytes[1]) < td->o.size) {
 573                 struct timeval comp_time;
 574                 unsigned long bytes_done[2] = { 0, 0 };
 575                 int min_evts = 0;
 576                 struct io_u *io_u;
 577                 int ret2, full;
 578
 579                 if (td->terminate)
 580                         break;
 581
 582                 update_tv_cache(td);
 583
 584                 if (runtime_exceeded(td, &td->tv_cache)) {
 585                         td->terminate = 1;
 586                         break;
 587                 }
 588
 589                 io_u = get_io_u(td);
 590                 if (!io_u)
 591                         break;
 592
 593                 /*
 594                  * Add verification end_io handler, if asked to verify
 595                  * a previously written file.
 596                  */
 597                 if (td->o.verify != VERIFY_NONE && io_u->ddir == DDIR_READ) {
 598                         io_u->end_io = verify_io_u;
 599                         td_set_runstate(td, TD_VERIFYING);
 600                 } else if (in_ramp_time(td))
 601                         td_set_runstate(td, TD_RAMP);
 602                 else
 603                         td_set_runstate(td, TD_RUNNING);
 604
 605                 ret = td_io_queue(td, io_u);
 606                 switch (ret) {
 607                 case FIO_Q_COMPLETED:
 608                         if (io_u->error) {
 609                                 ret = -io_u->error;
 610                                 clear_io_u(td, io_u);
 611                         } else if (io_u->resid) {
 612                                 int bytes = io_u->xfer_buflen - io_u->resid;
 613                                 struct fio_file *f = io_u->file;
 614
 615                                 /*
 616                                  * zero read, fail
 617                                  */
 618                                 if (!bytes) {
 619                                         td_verror(td, EIO, "full resid");
 620                                         put_io_u(td, io_u);
 621                                         break;
 622                                 }
 623
 624                                 io_u->xfer_buflen = io_u->resid;
 625                                 io_u->xfer_buf += bytes;
 626                                 io_u->offset += bytes;
 627
 628                                 td->ts.short_io_u[io_u->ddir]++;
 629
 630                                 if (io_u->offset == f->real_file_size)
 631                                         goto sync_done;
 632
 633                                 requeue_io_u(td, &io_u);
 634                         } else {
 635 sync_done:
 636                                 if (__should_check_rate(td, 0) ||
 637                                     __should_check_rate(td, 1))
 638                                         fio_gettime(&comp_time, NULL);
 639
 640                                 ret = io_u_sync_complete(td, io_u, bytes_done);
 641                                 if (ret < 0)
 642                                         break;
 643                         }
 644                         break;
 645                 case FIO_Q_QUEUED:
 646                         /*
 647                          * if the engine doesn't have a commit hook,
 648                          * the io_u is really queued. if it does have such
 649                          * a hook, it has to call io_u_queued() itself.
 650                          */
 651                         if (td->io_ops->commit == NULL)
 652                                 io_u_queued(td, io_u);
 653                         break;
 654                 case FIO_Q_BUSY:
 655                         requeue_io_u(td, &io_u);
 656                         ret2 = td_io_commit(td);
 657                         if (ret2 < 0)
 658                                 ret = ret2;
 659                         break;
 660                 default:
 661                         assert(ret < 0);
 662                         put_io_u(td, io_u);
 663                         break;
 664                 }
 665
 666                 if (break_on_this_error(td, &ret))
 667                         break;
 668
 669                 /*
 670                  * See if we need to complete some commands
 671                  */
 672                 full = queue_full(td) || ret == FIO_Q_BUSY;
 673                 if (full || !td->o.iodepth_batch_complete) {
 674                         min_evts = td->o.iodepth_batch_complete;
 675                         if (full && !min_evts)
 676                                 min_evts = 1;
 677
 678                         if (__should_check_rate(td, 0) ||
 679                             __should_check_rate(td, 1))
 680                                 fio_gettime(&comp_time, NULL);
 681
 682                         do {
 683                                 ret = io_u_queued_complete(td, min_evts, bytes_done);
 684                                 if (ret < 0)
 685                                         break;
 686
 687                         } while (full && (td->cur_depth > td->o.iodepth_low));
 688                 }
 689
 690                 if (ret < 0)
 691                         break;
 692                 if (!(bytes_done[0] + bytes_done[1]))
 693                         continue;
 694
 695                 if (!in_ramp_time(td) && should_check_rate(td, bytes_done)) {
 696                         if (check_min_rate(td, &comp_time, bytes_done)) {
 697                                 if (exitall_on_terminate)
 698                                         terminate_threads(td->groupid);
 699                                 td_verror(td, EIO, "check_min_rate");
 700                                 break;
 701                         }
 702                 }
 703
 704                 if (td->o.thinktime) {
 705                         unsigned long long b;
 706
 707                         b = td->io_blocks[0] + td->io_blocks[1];
 708                         if (!(b % td->o.thinktime_blocks)) {
 709                                 int left;
 710
 711                                 if (td->o.thinktime_spin)
 712                                         usec_spin(td->o.thinktime_spin);
 713
 714                                 left = td->o.thinktime - td->o.thinktime_spin;
 715                                 if (left)
 716                                         usec_sleep(td, left);
 717                         }
 718                 }
 719         }
 720
 721         if (td->o.fill_device && td->error == ENOSPC) {
 722                 td->error = 0;
 723                 td->terminate = 1;
 724         }
 725         if (!td->error) {
 726                 struct fio_file *f;
 727
 728                 i = td->cur_depth;
 729                 if (i)
 730                         ret = io_u_queued_complete(td, i, NULL);
 731
 732                 if (should_fsync(td) && td->o.end_fsync) {
 733                         td_set_runstate(td, TD_FSYNCING);
 734
 735                         for_each_file(td, f, i) {
 736                                 if (!fio_file_open(f))
 737                                         continue;
 738                                 fio_io_sync(td, f);
 739                         }
 740                 }
 741         } else
 742                 cleanup_pending_aio(td);
 743
 744         /*
 745          * stop job if we failed doing any IO
 746          */
 747         if ((td->this_io_bytes[0] + td->this_io_bytes[1]) == 0)
 748                 td->done = 1;
 749 }
 750
 751 static void cleanup_io_u(struct thread_data *td)
 752 {
 753         struct flist_head *entry, *n;
 754         struct io_u *io_u;
 755
 756         flist_for_each_safe(entry, n, &td->io_u_freelist) {
 757                 io_u = flist_entry(entry, struct io_u, list);
 758
 759                 flist_del(&io_u->list);
 760                 free(io_u);
 761         }
 762
 763         free_io_mem(td);
 764 }
 765
 766 static int init_io_u(struct thread_data *td)
 767 {
 768         struct io_u *io_u;
 769         unsigned int max_bs;
 770         int cl_align, i, max_units;
 771         char *p;
 772
 773         max_units = td->o.iodepth;
 774         max_bs = max(td->o.max_bs[DDIR_READ], td->o.max_bs[DDIR_WRITE]);
 775         td->orig_buffer_size = (unsigned long long) max_bs
 776                                         * (unsigned long long) max_units;
 777
 778         if (td->o.mem_type == MEM_SHMHUGE || td->o.mem_type == MEM_MMAPHUGE) {
 779                 unsigned long bs;
 780
 781                 bs = td->orig_buffer_size + td->o.hugepage_size - 1;
 782                 td->orig_buffer_size = bs & ~(td->o.hugepage_size - 1);
 783         }
 784
 785         if (td->orig_buffer_size != (size_t) td->orig_buffer_size) {
 786                 log_err("fio: IO memory too large. Reduce max_bs or iodepth\n");
 787                 return 1;
 788         }
 789
 790         if (allocate_io_mem(td))
 791                 return 1;
 792
 793         if (td->o.odirect)
 794                 p = ALIGN(td->orig_buffer);
 795         else
 796                 p = td->orig_buffer;
 797
 798         cl_align = os_cache_line_size();
 799
 800         for (i = 0; i < max_units; i++) {
 801                 void *ptr;
 802
 803                 if (td->terminate)
 804                         return 1;
 805
 806                 if (posix_memalign(&ptr, cl_align, sizeof(*io_u))) {
 807                         log_err("fio: posix_memalign=%s\n", strerror(errno));
 808                         break;
 809                 }
 810
 811                 io_u = ptr;
 812                 memset(io_u, 0, sizeof(*io_u));
 813                 INIT_FLIST_HEAD(&io_u->list);
 814
 815                 if (!(td->io_ops->flags & FIO_NOIO)) {
 816                         io_u->buf = p + max_bs * i;
 817
 818                         if (td_write(td) && !td->o.refill_buffers)
 819                                 io_u_fill_buffer(td, io_u, max_bs);
 820                 }
 821
 822                 io_u->index = i;
 823                 io_u->flags = IO_U_F_FREE;
 824                 flist_add(&io_u->list, &td->io_u_freelist);
 825         }
 826
 827         return 0;
 828 }
 829
 830 static int switch_ioscheduler(struct thread_data *td)
 831 {
 832         char tmp[256], tmp2[128];
 833         FILE *f;
 834         int ret;
 835
 836         if (td->io_ops->flags & FIO_DISKLESSIO)
 837                 return 0;
 838
 839         sprintf(tmp, "%s/queue/scheduler", td->sysfs_root);
 840
 841         f = fopen(tmp, "r+");
 842         if (!f) {
 843                 if (errno == ENOENT) {
 844                         log_err("fio: os or kernel doesn't support IO scheduler"
 845                                 " switching\n");
 846                         return 0;
 847                 }
 848                 td_verror(td, errno, "fopen iosched");
 849                 return 1;
 850         }
 851
 852         /*
 853          * Set io scheduler.
 854          */
 855         ret = fwrite(td->o.ioscheduler, strlen(td->o.ioscheduler), 1, f);
 856         if (ferror(f) || ret != 1) {
 857                 td_verror(td, errno, "fwrite");
 858                 fclose(f);
 859                 return 1;
 860         }
 861
 862         rewind(f);
 863
 864         /*
 865          * Read back and check that the selected scheduler is now the default.
 866          */
 867         ret = fread(tmp, 1, sizeof(tmp), f);
 868         if (ferror(f) || ret < 0) {
 869                 td_verror(td, errno, "fread");
 870                 fclose(f);
 871                 return 1;
 872         }
 873
 874         sprintf(tmp2, "[%s]", td->o.ioscheduler);
 875         if (!strstr(tmp, tmp2)) {
 876                 log_err("fio: io scheduler %s not found\n", td->o.ioscheduler);
 877                 td_verror(td, EINVAL, "iosched_switch");
 878                 fclose(f);
 879                 return 1;
 880         }
 881
 882         fclose(f);
 883         return 0;
 884 }
 885
 886 static int keep_running(struct thread_data *td)
 887 {
 888         unsigned long long io_done;
 889
 890         if (td->done)
 891                 return 0;
 892         if (td->o.time_based)
 893                 return 1;
 894         if (td->o.loops) {
 895                 td->o.loops--;
 896                 return 1;
 897         }
 898
 899         io_done = td->io_bytes[DDIR_READ] + td->io_bytes[DDIR_WRITE]
 900                         + td->io_skip_bytes;
 901         if (io_done < td->o.size)
 902                 return 1;
 903
 904         return 0;
 905 }
 906
 907 static void reset_io_counters(struct thread_data *td)
 908 {
 909         td->ts.stat_io_bytes[0] = td->ts.stat_io_bytes[1] = 0;
 910         td->this_io_bytes[0] = td->this_io_bytes[1] = 0;
 911         td->zone_bytes = 0;
 912         td->rate_bytes[0] = td->rate_bytes[1] = 0;
 913         td->rate_blocks[0] = td->rate_blocks[1] = 0;
 914
 915         td->last_was_sync = 0;
 916
 917         /*
 918          * reset file done count if we are to start over
 919          */
 920         if (td->o.time_based || td->o.loops)
 921                 td->nr_done_files = 0;
 922
 923         /*
 924          * Set the same seed to get repeatable runs
 925          */
 926         td_fill_rand_seeds(td);
 927 }
 928
 929 void reset_all_stats(struct thread_data *td)
 930 {
 931         struct timeval tv;
 932         int i;
 933
 934         reset_io_counters(td);
 935
 936         for (i = 0; i < 2; i++) {
 937                 td->io_bytes[i] = 0;
 938                 td->io_blocks[i] = 0;
 939                 td->io_issues[i] = 0;
 940                 td->ts.total_io_u[i] = 0;
 941         }
 942
 943         fio_gettime(&tv, NULL);
 944         memcpy(&td->epoch, &tv, sizeof(tv));
 945         memcpy(&td->start, &tv, sizeof(tv));
 946 }
 947
 948 static void clear_io_state(struct thread_data *td)
 949 {
 950         struct fio_file *f;
 951         unsigned int i;
 952
 953         reset_io_counters(td);
 954
 955         close_files(td);
 956         for_each_file(td, f, i)
 957                 fio_file_clear_done(f);
 958 }
 959
 960 static int exec_string(const char *string)
 961 {
 962         int ret, newlen = strlen(string) + 1 + 8;
 963         char *str;
 964
 965         str = malloc(newlen);
 966         sprintf(str, "sh -c %s", string);
 967
 968         ret = system(str);
 969         if (ret == -1)
 970                 log_err("fio: exec of cmd <%s> failed\n", str);
 971
 972         free(str);
 973         return ret;
 974 }
 975
 976 /*
 977  * Entry point for the thread based jobs. The process based jobs end up
 978  * here as well, after a little setup.
 979  */
 980 static void *thread_main(void *data)
 981 {
 982         unsigned long long runtime[2], elapsed;
 983         struct thread_data *td = data;
 984         int clear_state;
 985
 986         if (!td->o.use_thread)
 987                 setsid();
 988
 989         td->pid = getpid();
 990
 991         dprint(FD_PROCESS, "jobs pid=%d started\n", (int) td->pid);
 992
 993         INIT_FLIST_HEAD(&td->io_u_freelist);
 994         INIT_FLIST_HEAD(&td->io_u_busylist);
 995         INIT_FLIST_HEAD(&td->io_u_requeues);
 996         INIT_FLIST_HEAD(&td->io_log_list);
 997         INIT_FLIST_HEAD(&td->io_hist_list);
 998         td->io_hist_tree = RB_ROOT;
 999
1000         td_set_runstate(td, TD_INITIALIZED);
1001         dprint(FD_MUTEX, "up startup_mutex\n");
1002         fio_mutex_up(startup_mutex);
1003         dprint(FD_MUTEX, "wait on td->mutex\n");
1004         fio_mutex_down(td->mutex);
1005         dprint(FD_MUTEX, "done waiting on td->mutex\n");
1006
1007         /*
1008          * the ->mutex mutex is now no longer used, close it to avoid
1009          * eating a file descriptor
1010          */
1011         fio_mutex_remove(td->mutex);
1012
1013         /*
1014          * May alter parameters that init_io_u() will use, so we need to
1015          * do this first.
1016          */
1017         if (init_iolog(td))
1018                 goto err;
1019
1020         if (init_io_u(td))
1021                 goto err;
1022
1023         if (td->o.cpumask_set && fio_setaffinity(td) == -1) {
1024                 td_verror(td, errno, "cpu_set_affinity");
1025                 goto err;
1026         }
1027
1028         /*
1029          * If we have a gettimeofday() thread, make sure we exclude that
1030          * thread from this job
1031          */
1032         if (td->o.gtod_cpu) {
1033                 fio_cpu_clear(&td->o.cpumask, td->o.gtod_cpu);
1034                 if (fio_setaffinity(td) == -1) {
1035                         td_verror(td, errno, "cpu_set_affinity");
1036                         goto err;
1037                 }
1038         }
1039
1040         if (td->ioprio_set) {
1041                 if (ioprio_set(IOPRIO_WHO_PROCESS, 0, td->ioprio) == -1) {
1042                         td_verror(td, errno, "ioprio_set");
1043                         goto err;
1044                 }
1045         }
1046
1047         if (nice(td->o.nice) == -1) {
1048                 td_verror(td, errno, "nice");
1049                 goto err;
1050         }
1051
1052         if (td->o.ioscheduler && switch_ioscheduler(td))
1053                 goto err;
1054
1055         if (!td->o.create_serialize && setup_files(td))
1056                 goto err;
1057
1058         if (td_io_init(td))
1059                 goto err;
1060
1061         if (init_random_map(td))
1062                 goto err;
1063
1064         if (td->o.exec_prerun) {
1065                 if (exec_string(td->o.exec_prerun))
1066                         goto err;
1067         }
1068
1069         if (td->o.pre_read) {
1070                 if (pre_read_files(td) < 0)
1071                         goto err;
1072         }
1073
1074         fio_gettime(&td->epoch, NULL);
1075         getrusage(RUSAGE_SELF, &td->ts.ru_start);
1076
1077         runtime[0] = runtime[1] = 0;
1078         clear_state = 0;
1079         while (keep_running(td)) {
1080                 fio_gettime(&td->start, NULL);
1081                 memcpy(&td->ts.stat_sample_time, &td->start, sizeof(td->start));
1082                 memcpy(&td->tv_cache, &td->start, sizeof(td->start));
1083
1084                 if (td->o.ratemin[0] || td->o.ratemin[1])
1085                         memcpy(&td->lastrate, &td->ts.stat_sample_time,
1086                                                         sizeof(td->lastrate));
1087
1088                 if (clear_state)
1089                         clear_io_state(td);
1090
1091                 prune_io_piece_log(td);
1092
1093                 do_io(td);
1094
1095                 clear_state = 1;
1096
1097                 if (td_read(td) && td->io_bytes[DDIR_READ]) {
1098                         elapsed = utime_since_now(&td->start);
1099                         runtime[DDIR_READ] += elapsed;
1100                 }
1101                 if (td_write(td) && td->io_bytes[DDIR_WRITE]) {
1102                         elapsed = utime_since_now(&td->start);
1103                         runtime[DDIR_WRITE] += elapsed;
1104                 }
1105
1106                 if (td->error || td->terminate)
1107                         break;
1108
1109                 if (!td->o.do_verify ||
1110                     td->o.verify == VERIFY_NONE ||
1111                     (td->io_ops->flags & FIO_UNIDIR))
1112                         continue;
1113
1114                 clear_io_state(td);
1115
1116                 fio_gettime(&td->start, NULL);
1117
1118                 do_verify(td);
1119
1120                 runtime[DDIR_READ] += utime_since_now(&td->start);
1121
1122                 if (td->error || td->terminate)
1123                         break;
1124         }
1125
1126         update_rusage_stat(td);
1127         td->ts.runtime[0] = (runtime[0] + 999) / 1000;
1128         td->ts.runtime[1] = (runtime[1] + 999) / 1000;
1129         td->ts.total_run_time = mtime_since_now(&td->epoch);
1130         td->ts.io_bytes[0] = td->io_bytes[0];
1131         td->ts.io_bytes[1] = td->io_bytes[1];
1132
1133         fio_mutex_down(writeout_mutex);
1134         if (td->ts.bw_log) {
1135                 if (td->o.bw_log_file) {
1136                         finish_log_named(td, td->ts.bw_log,
1137                                                 td->o.bw_log_file, "bw");
1138                 } else
1139                         finish_log(td, td->ts.bw_log, "bw");
1140         }
1141         if (td->ts.slat_log) {
1142                 if (td->o.lat_log_file) {
1143                         finish_log_named(td, td->ts.slat_log,
1144                                                 td->o.lat_log_file, "slat");
1145                 } else
1146                         finish_log(td, td->ts.slat_log, "slat");
1147         }
1148         if (td->ts.clat_log) {
1149                 if (td->o.lat_log_file) {
1150                         finish_log_named(td, td->ts.clat_log,
1151                                                 td->o.lat_log_file, "clat");
1152                 } else
1153                         finish_log(td, td->ts.clat_log, "clat");
1154         }
1155         fio_mutex_up(writeout_mutex);
1156         if (td->o.exec_postrun)
1157                 exec_string(td->o.exec_postrun);
1158
1159         if (exitall_on_terminate)
1160                 terminate_threads(td->groupid);
1161
1162 err:
1163         if (td->error)
1164                 printf("fio: pid=%d, err=%d/%s\n", (int) td->pid, td->error,
1165                                                         td->verror);
1166         close_and_free_files(td);
1167         close_ioengine(td);
1168         cleanup_io_u(td);
1169
1170         if (td->o.cpumask_set) {
1171                 int ret = fio_cpuset_exit(&td->o.cpumask);
1172
1173                 td_verror(td, ret, "fio_cpuset_exit");
1174         }
1175
1176         /*
1177          * do this very late, it will log file closing as well
1178          */
1179         if (td->o.write_iolog_file)
1180                 write_iolog_close(td);
1181
1182         options_mem_free(td);
1183         td_set_runstate(td, TD_EXITED);
1184         return (void *) (unsigned long) td->error;
1185 }
1186
1187 /*
1188  * We cannot pass the td data into a forked process, so attach the td and
1189  * pass it to the thread worker.
1190  */
1191 static int fork_main(int shmid, int offset)
1192 {
1193         struct thread_data *td;
1194         void *data, *ret;
1195
1196         data = shmat(shmid, NULL, 0);
1197         if (data == (void *) -1) {
1198                 int __err = errno;
1199
1200                 perror("shmat");
1201                 return __err;
1202         }
1203
1204         td = data + offset * sizeof(struct thread_data);
1205         ret = thread_main(td);
1206         shmdt(data);
1207         return (int) (unsigned long) ret;
1208 }
1209
1210 /*
1211  * Run over the job map and reap the threads that have exited, if any.
1212  */
1213 static void reap_threads(int *nr_running, int *t_rate, int *m_rate)
1214 {
1215         struct thread_data *td;
1216         int i, cputhreads, realthreads, pending, status, ret;
1217
1218         /*
1219          * reap exited threads (TD_EXITED -> TD_REAPED)
1220          */
1221         realthreads = pending = cputhreads = 0;
1222         for_each_td(td, i) {
1223                 int flags = 0;
1224
1225                 /*
1226                  * ->io_ops is NULL for a thread that has closed its
1227                  * io engine
1228                  */
1229                 if (td->io_ops && !strcmp(td->io_ops->name, "cpuio"))
1230                         cputhreads++;
1231                 else
1232                         realthreads++;
1233
1234                 if (!td->pid) {
1235                         pending++;
1236                         continue;
1237                 }
1238                 if (td->runstate == TD_REAPED)
1239                         continue;
1240                 if (td->o.use_thread) {
1241                         if (td->runstate == TD_EXITED) {
1242                                 td_set_runstate(td, TD_REAPED);
1243                                 goto reaped;
1244                         }
1245                         continue;
1246                 }
1247
1248                 flags = WNOHANG;
1249                 if (td->runstate == TD_EXITED)
1250                         flags = 0;
1251
1252                 /*
1253                  * check if someone quit or got killed in an unusual way
1254                  */
1255                 ret = waitpid(td->pid, &status, flags);
1256                 if (ret < 0) {
1257                         if (errno == ECHILD) {
1258                                 log_err("fio: pid=%d disappeared %d\n",
1259                                                 (int) td->pid, td->runstate);
1260                                 td_set_runstate(td, TD_REAPED);
1261                                 goto reaped;
1262                         }
1263                         perror("waitpid");
1264                 } else if (ret == td->pid) {
1265                         if (WIFSIGNALED(status)) {
1266                                 int sig = WTERMSIG(status);
1267
1268                                 if (sig != SIGQUIT)
1269                                         log_err("fio: pid=%d, got signal=%d\n",
1270                                                         (int) td->pid, sig);
1271                                 td_set_runstate(td, TD_REAPED);
1272                                 goto reaped;
1273                         }
1274                         if (WIFEXITED(status)) {
1275                                 if (WEXITSTATUS(status) && !td->error)
1276                                         td->error = WEXITSTATUS(status);
1277
1278                                 td_set_runstate(td, TD_REAPED);
1279                                 goto reaped;
1280                         }
1281                 }
1282
1283                 /*
1284                  * thread is not dead, continue
1285                  */
1286                 pending++;
1287                 continue;
1288 reaped:
1289                 (*nr_running)--;
1290                 (*m_rate) -= (td->o.ratemin[0] + td->o.ratemin[1]);
1291                 (*t_rate) -= (td->o.rate[0] + td->o.rate[1]);
1292                 if (!td->pid)
1293                         pending--;
1294
1295                 if (td->error)
1296                         exit_value++;
1297
1298                 done_secs += mtime_since_now(&td->epoch) / 1000;
1299         }
1300
1301         if (*nr_running == cputhreads && !pending && realthreads)
1302                 terminate_threads(TERMINATE_ALL);
1303 }
1304
1305 static void *gtod_thread_main(void *data)
1306 {
1307         fio_mutex_up(startup_mutex);
1308
1309         /*
1310          * As long as we have jobs around, update the clock. It would be nice
1311          * to have some way of NOT hammering that CPU with gettimeofday(),
1312          * but I'm not sure what to use outside of a simple CPU nop to relax
1313          * it - we don't want to lose precision.
1314          */
1315         while (threads) {
1316                 fio_gtod_update();
1317                 nop;
1318         }
1319
1320         return NULL;
1321 }
1322
1323 static int fio_start_gtod_thread(void)
1324 {
1325         int ret;
1326
1327         ret = pthread_create(&gtod_thread, NULL, gtod_thread_main, NULL);
1328         if (ret) {
1329                 log_err("Can't create gtod thread: %s\n", strerror(ret));
1330                 return 1;
1331         }
1332
1333         ret = pthread_detach(gtod_thread);
1334         if (ret) {
1335                 log_err("Can't detatch gtod thread: %s\n", strerror(ret));
1336                 return 1;
1337         }
1338
1339         dprint(FD_MUTEX, "wait on startup_mutex\n");
1340         fio_mutex_down(startup_mutex);
1341         dprint(FD_MUTEX, "done waiting on startup_mutex\n");
1342         return 0;
1343 }
1344
1345 /*
1346  * Main function for kicking off and reaping jobs, as needed.
1347  */
1348 static void run_threads(void)
1349 {
1350         struct thread_data *td;
1351         unsigned long spent;
1352         int i, todo, nr_running, m_rate, t_rate, nr_started;
1353
1354         if (fio_pin_memory())
1355                 return;
1356
1357         if (fio_gtod_offload && fio_start_gtod_thread())
1358                 return;
1359
1360         if (!terse_output) {
1361                 printf("Starting ");
1362                 if (nr_thread)
1363                         printf("%d thread%s", nr_thread,
1364                                                 nr_thread > 1 ? "s" : "");
1365                 if (nr_process) {
1366                         if (nr_thread)
1367                                 printf(" and ");
1368                         printf("%d process%s", nr_process,
1369                                                 nr_process > 1 ? "es" : "");
1370                 }
1371                 printf("\n");
1372                 fflush(stdout);
1373         }
1374
1375         set_sig_handlers();
1376
1377         todo = thread_number;
1378         nr_running = 0;
1379         nr_started = 0;
1380         m_rate = t_rate = 0;
1381
1382         for_each_td(td, i) {
1383                 print_status_init(td->thread_number - 1);
1384
1385                 if (!td->o.create_serialize) {
1386                         init_disk_util(td);
1387                         continue;
1388                 }
1389
1390                 /*
1391                  * do file setup here so it happens sequentially,
1392                  * we don't want X number of threads getting their
1393                  * client data interspersed on disk
1394                  */
1395                 if (setup_files(td)) {
1396                         exit_value++;
1397                         if (td->error)
1398                                 log_err("fio: pid=%d, err=%d/%s\n",
1399                                         (int) td->pid, td->error, td->verror);
1400                         td_set_runstate(td, TD_REAPED);
1401                         todo--;
1402                 } else {
1403                         struct fio_file *f;
1404                         unsigned int i;
1405
1406                         /*
1407                          * for sharing to work, each job must always open
1408                          * its own files. so close them, if we opened them
1409                          * for creation
1410                          */
1411                         for_each_file(td, f, i) {
1412                                 if (fio_file_open(f))
1413                                         td_io_close_file(td, f);
1414                         }
1415                 }
1416
1417                 init_disk_util(td);
1418         }
1419
1420         set_genesis_time();
1421
1422         while (todo) {
1423                 struct thread_data *map[MAX_JOBS];
1424                 struct timeval this_start;
1425                 int this_jobs = 0, left;
1426
1427                 /*
1428                  * create threads (TD_NOT_CREATED -> TD_CREATED)
1429                  */
1430                 for_each_td(td, i) {
1431                         if (td->runstate != TD_NOT_CREATED)
1432                                 continue;
1433
1434                         /*
1435                          * never got a chance to start, killed by other
1436                          * thread for some reason
1437                          */
1438                         if (td->terminate) {
1439                                 todo--;
1440                                 continue;
1441                         }
1442
1443                         if (td->o.start_delay) {
1444                                 spent = mtime_since_genesis();
1445
1446                                 if (td->o.start_delay * 1000 > spent)
1447                                         continue;
1448                         }
1449
1450                         if (td->o.stonewall && (nr_started || nr_running)) {
1451                                 dprint(FD_PROCESS, "%s: stonewall wait\n",
1452                                                         td->o.name);
1453                                 break;
1454                         }
1455
1456                         /*
1457                          * Set state to created. Thread will transition
1458                          * to TD_INITIALIZED when it's done setting up.
1459                          */
1460                         td_set_runstate(td, TD_CREATED);
1461                         map[this_jobs++] = td;
1462                         nr_started++;
1463
1464                         if (td->o.use_thread) {
1465                                 int ret;
1466
1467                                 dprint(FD_PROCESS, "will pthread_create\n");
1468                                 ret = pthread_create(&td->thread, NULL,
1469                                                         thread_main, td);
1470                                 if (ret) {
1471                                         log_err("pthread_create: %s\n",
1472                                                         strerror(ret));
1473                                         nr_started--;
1474                                         break;
1475                                 }
1476                                 ret = pthread_detach(td->thread);
1477                                 if (ret)
1478                                         log_err("pthread_detach: %s",
1479                                                         strerror(ret));
1480                         } else {
1481                                 pid_t pid;
1482                                 dprint(FD_PROCESS, "will fork\n");
1483                                 pid = fork();
1484                                 if (!pid) {
1485                                         int ret = fork_main(shm_id, i);
1486
1487                                         _exit(ret);
1488                                 } else if (i == fio_debug_jobno)
1489                                         *fio_debug_jobp = pid;
1490                         }
1491                         dprint(FD_MUTEX, "wait on startup_mutex\n");
1492                         fio_mutex_down(startup_mutex);
1493                         dprint(FD_MUTEX, "done waiting on startup_mutex\n");
1494                 }
1495
1496                 /*
1497                  * Wait for the started threads to transition to
1498                  * TD_INITIALIZED.
1499                  */
1500                 fio_gettime(&this_start, NULL);
1501                 left = this_jobs;
1502                 while (left && !fio_abort) {
1503                         if (mtime_since_now(&this_start) > JOB_START_TIMEOUT)
1504                                 break;
1505
1506                         usleep(100000);
1507
1508                         for (i = 0; i < this_jobs; i++) {
1509                                 td = map[i];
1510                                 if (!td)
1511                                         continue;
1512                                 if (td->runstate == TD_INITIALIZED) {
1513                                         map[i] = NULL;
1514                                         left--;
1515                                 } else if (td->runstate >= TD_EXITED) {
1516                                         map[i] = NULL;
1517                                         left--;
1518                                         todo--;
1519                                         nr_running++; /* work-around... */
1520                                 }
1521                         }
1522                 }
1523
1524                 if (left) {
1525                         log_err("fio: %d jobs failed to start\n", left);
1526                         for (i = 0; i < this_jobs; i++) {
1527                                 td = map[i];
1528                                 if (!td)
1529                                         continue;
1530                                 kill(td->pid, SIGTERM);
1531                         }
1532                         break;
1533                 }
1534
1535                 /*
1536                  * start created threads (TD_INITIALIZED -> TD_RUNNING).
1537                  */
1538                 for_each_td(td, i) {
1539                         if (td->runstate != TD_INITIALIZED)
1540                                 continue;
1541
1542                         if (in_ramp_time(td))
1543                                 td_set_runstate(td, TD_RAMP);
1544                         else
1545                                 td_set_runstate(td, TD_RUNNING);
1546                         nr_running++;
1547                         nr_started--;
1548                         m_rate += td->o.ratemin[0] + td->o.ratemin[1];
1549                         t_rate += td->o.rate[0] + td->o.rate[1];
1550                         todo--;
1551                         fio_mutex_up(td->mutex);
1552                 }
1553
1554                 reap_threads(&nr_running, &t_rate, &m_rate);
1555
1556                 if (todo)
1557                         usleep(100000);
1558         }
1559
1560         while (nr_running) {
1561                 reap_threads(&nr_running, &t_rate, &m_rate);
1562                 usleep(10000);
1563         }
1564
1565         update_io_ticks();
1566         fio_unpin_memory();
1567 }
1568
1569 int main(int argc, char *argv[])
1570 {
1571         long ps;
1572
1573         sinit();
1574
1575         /*
1576          * We need locale for number printing, if it isn't set then just
1577          * go with the US format.
1578          */
1579         if (!getenv("LC_NUMERIC"))
1580                 setlocale(LC_NUMERIC, "en_US");
1581
1582         if (parse_options(argc, argv))
1583                 return 1;
1584
1585         if (!thread_number)
1586                 return 0;
1587
1588         ps = sysconf(_SC_PAGESIZE);
1589         if (ps < 0) {
1590                 log_err("Failed to get page size\n");
1591                 return 1;
1592         }
1593
1594         page_size = ps;
1595         page_mask = ps - 1;
1596
1597         if (write_bw_log) {
1598                 setup_log(&agg_io_log[DDIR_READ]);
1599                 setup_log(&agg_io_log[DDIR_WRITE]);
1600         }
1601
1602         startup_mutex = fio_mutex_init(0);
1603         writeout_mutex = fio_mutex_init(1);
1604
1605         set_genesis_time();
1606
1607         status_timer_arm();
1608
1609         run_threads();
1610
1611         if (!fio_abort) {
1612                 show_run_stats();
1613                 if (write_bw_log) {
1614                         __finish_log(agg_io_log[DDIR_READ], "agg-read_bw.log");
1615                         __finish_log(agg_io_log[DDIR_WRITE],
1616                                         "agg-write_bw.log");
1617                 }
1618         }
1619
1620         fio_mutex_remove(startup_mutex);
1621         fio_mutex_remove(writeout_mutex);
1622         return exit_value;
1623 }