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 #include "cgroup.h"
  43 #include "profile.h"
  44 #include "lib/rand.h"
  45
  46 unsigned long page_mask;
  47 unsigned long page_size;
  48
  49 #define PAGE_ALIGN(buf) \
  50         (char *) (((unsigned long) (buf) + page_mask) & ~page_mask)
  51
  52 int groupid = 0;
  53 int thread_number = 0;
  54 int nr_process = 0;
  55 int nr_thread = 0;
  56 int shm_id = 0;
  57 int temp_stall_ts;
  58 unsigned long done_secs = 0;
  59
  60 static struct fio_mutex *startup_mutex;
  61 static struct fio_mutex *writeout_mutex;
  62 static volatile int fio_abort;
  63 static int exit_value;
  64 static struct itimerval itimer;
  65 static pthread_t gtod_thread;
  66 static struct flist_head *cgroup_list;
  67 static char *cgroup_mnt;
  68
  69 struct io_log *agg_io_log[2];
  70
  71 #define TERMINATE_ALL           (-1)
  72 #define JOB_START_TIMEOUT       (5 * 1000)
  73
  74 void td_set_runstate(struct thread_data *td, int runstate)
  75 {
  76         if (td->runstate == runstate)
  77                 return;
  78
  79         dprint(FD_PROCESS, "pid=%d: runstate %d -> %d\n", (int) td->pid,
  80                                                 td->runstate, runstate);
  81         td->runstate = runstate;
  82 }
  83
  84 static void terminate_threads(int group_id)
  85 {
  86         struct thread_data *td;
  87         int i;
  88
  89         dprint(FD_PROCESS, "terminate group_id=%d\n", group_id);
  90
  91         for_each_td(td, i) {
  92                 if (group_id == TERMINATE_ALL || groupid == td->groupid) {
  93                         dprint(FD_PROCESS, "setting terminate on %s/%d\n",
  94                                                 td->o.name, (int) td->pid);
  95                         td->terminate = 1;
  96                         td->o.start_delay = 0;
  97
  98                         /*
  99                          * if the thread is running, just let it exit
 100                          */
 101                         if (td->runstate < TD_RUNNING)
 102                                 kill(td->pid, SIGQUIT);
 103                         else {
 104                                 struct ioengine_ops *ops = td->io_ops;
 105
 106                                 if (ops && (ops->flags & FIO_SIGQUIT))
 107                                         kill(td->pid, SIGQUIT);
 108                         }
 109                 }
 110         }
 111 }
 112
 113 static void status_timer_arm(void)
 114 {
 115         itimer.it_value.tv_sec = 0;
 116         itimer.it_value.tv_usec = DISK_UTIL_MSEC * 1000;
 117         setitimer(ITIMER_REAL, &itimer, NULL);
 118 }
 119
 120 static void sig_alrm(int fio_unused sig)
 121 {
 122         if (threads) {
 123                 update_io_ticks();
 124                 print_thread_status();
 125                 status_timer_arm();
 126         }
 127 }
 128
 129 /*
 130  * Happens on thread runs with ctrl-c, ignore our own SIGQUIT
 131  */
 132 static void sig_quit(int sig)
 133 {
 134 }
 135
 136 static void sig_int(int sig)
 137 {
 138         if (threads) {
 139                 log_info("\nfio: terminating on signal %d\n", sig);
 140                 fflush(stdout);
 141                 terminate_threads(TERMINATE_ALL);
 142         }
 143 }
 144
 145 static void set_sig_handlers(void)
 146 {
 147         struct sigaction act;
 148
 149         memset(&act, 0, sizeof(act));
 150         act.sa_handler = sig_alrm;
 151         act.sa_flags = SA_RESTART;
 152         sigaction(SIGALRM, &act, NULL);
 153
 154         memset(&act, 0, sizeof(act));
 155         act.sa_handler = sig_int;
 156         act.sa_flags = SA_RESTART;
 157         sigaction(SIGINT, &act, NULL);
 158
 159         memset(&act, 0, sizeof(act));
 160         act.sa_handler = sig_quit;
 161         act.sa_flags = SA_RESTART;
 162         sigaction(SIGQUIT, &act, NULL);
 163 }
 164
 165 /*
 166  * Check if we are above the minimum rate given.
 167  */
 168 static int __check_min_rate(struct thread_data *td, struct timeval *now,
 169                             enum td_ddir ddir)
 170 {
 171         unsigned long long bytes = 0;
 172         unsigned long iops = 0;
 173         unsigned long spent;
 174         unsigned long rate;
 175         unsigned int ratemin = 0;
 176         unsigned int rate_iops = 0;
 177         unsigned int rate_iops_min = 0;
 178
 179         if (!td->o.ratemin[ddir] && !td->o.rate_iops_min[ddir])
 180                 return 0;
 181
 182         /*
 183          * allow a 2 second settle period in the beginning
 184          */
 185         if (mtime_since(&td->start, now) < 2000)
 186                 return 0;
 187
 188         iops += td->io_blocks[ddir];
 189         bytes += td->this_io_bytes[ddir];
 190         ratemin += td->o.ratemin[ddir];
 191         rate_iops += td->o.rate_iops[ddir];
 192         rate_iops_min += td->o.rate_iops_min[ddir];
 193
 194         /*
 195          * if rate blocks is set, sample is running
 196          */
 197         if (td->rate_bytes[ddir] || td->rate_blocks[ddir]) {
 198                 spent = mtime_since(&td->lastrate[ddir], now);
 199                 if (spent < td->o.ratecycle)
 200                         return 0;
 201
 202                 if (td->o.rate[ddir]) {
 203                         /*
 204                          * check bandwidth specified rate
 205                          */
 206                         if (bytes < td->rate_bytes[ddir]) {
 207                                 log_err("%s: min rate %u not met\n", td->o.name,
 208                                                                 ratemin);
 209                                 return 1;
 210                         } else {
 211                                 rate = ((bytes - td->rate_bytes[ddir]) * 1000) / spent;
 212                                 if (rate < ratemin ||
 213                                     bytes < td->rate_bytes[ddir]) {
 214                                         log_err("%s: min rate %u not met, got"
 215                                                 " %luKB/sec\n", td->o.name,
 216                                                         ratemin, rate);
 217                                         return 1;
 218                                 }
 219                         }
 220                 } else {
 221                         /*
 222                          * checks iops specified rate
 223                          */
 224                         if (iops < rate_iops) {
 225                                 log_err("%s: min iops rate %u not met\n",
 226                                                 td->o.name, rate_iops);
 227                                 return 1;
 228                         } else {
 229                                 rate = ((iops - td->rate_blocks[ddir]) * 1000) / spent;
 230                                 if (rate < rate_iops_min ||
 231                                     iops < td->rate_blocks[ddir]) {
 232                                         log_err("%s: min iops rate %u not met,"
 233                                                 " got %lu\n", td->o.name,
 234                                                         rate_iops_min, rate);
 235                                 }
 236                         }
 237                 }
 238         }
 239
 240         td->rate_bytes[ddir] = bytes;
 241         td->rate_blocks[ddir] = iops;
 242         memcpy(&td->lastrate[ddir], now, sizeof(*now));
 243         return 0;
 244 }
 245
 246 static int check_min_rate(struct thread_data *td, struct timeval *now,
 247                           unsigned long *bytes_done)
 248 {
 249         int ret = 0;
 250
 251         if (bytes_done[0])
 252                 ret |= __check_min_rate(td, now, 0);
 253         if (bytes_done[1])
 254                 ret |= __check_min_rate(td, now, 1);
 255
 256         return ret;
 257 }
 258
 259 static inline int runtime_exceeded(struct thread_data *td, struct timeval *t)
 260 {
 261         if (!td->o.timeout)
 262                 return 0;
 263         if (mtime_since(&td->epoch, t) >= td->o.timeout * 1000)
 264                 return 1;
 265
 266         return 0;
 267 }
 268
 269 /*
 270  * When job exits, we can cancel the in-flight IO if we are using async
 271  * io. Attempt to do so.
 272  */
 273 static void cleanup_pending_aio(struct thread_data *td)
 274 {
 275         struct flist_head *entry, *n;
 276         struct io_u *io_u;
 277         int r;
 278
 279         /*
 280          * get immediately available events, if any
 281          */
 282         r = io_u_queued_complete(td, 0, NULL);
 283         if (r < 0)
 284                 return;
 285
 286         /*
 287          * now cancel remaining active events
 288          */
 289         if (td->io_ops->cancel) {
 290                 flist_for_each_safe(entry, n, &td->io_u_busylist) {
 291                         io_u = flist_entry(entry, struct io_u, list);
 292
 293                         /*
 294                          * if the io_u isn't in flight, then that generally
 295                          * means someone leaked an io_u. complain but fix
 296                          * it up, so we don't stall here.
 297                          */
 298                         if ((io_u->flags & IO_U_F_FLIGHT) == 0) {
 299                                 log_err("fio: non-busy IO on busy list\n");
 300                                 put_io_u(td, io_u);
 301                         } else {
 302                                 r = td->io_ops->cancel(td, io_u);
 303                                 if (!r)
 304                                         put_io_u(td, io_u);
 305                         }
 306                 }
 307         }
 308
 309         if (td->cur_depth)
 310                 r = io_u_queued_complete(td, td->cur_depth, NULL);
 311 }
 312
 313 /*
 314  * Helper to handle the final sync of a file. Works just like the normal
 315  * io path, just does everything sync.
 316  */
 317 static int fio_io_sync(struct thread_data *td, struct fio_file *f)
 318 {
 319         struct io_u *io_u = __get_io_u(td);
 320         int ret;
 321
 322         if (!io_u)
 323                 return 1;
 324
 325         io_u->ddir = DDIR_SYNC;
 326         io_u->file = f;
 327
 328         if (td_io_prep(td, io_u)) {
 329                 put_io_u(td, io_u);
 330                 return 1;
 331         }
 332
 333 requeue:
 334         ret = td_io_queue(td, io_u);
 335         if (ret < 0) {
 336                 td_verror(td, io_u->error, "td_io_queue");
 337                 put_io_u(td, io_u);
 338                 return 1;
 339         } else if (ret == FIO_Q_QUEUED) {
 340                 if (io_u_queued_complete(td, 1, NULL) < 0)
 341                         return 1;
 342         } else if (ret == FIO_Q_COMPLETED) {
 343                 if (io_u->error) {
 344                         td_verror(td, io_u->error, "td_io_queue");
 345                         return 1;
 346                 }
 347
 348                 if (io_u_sync_complete(td, io_u, NULL) < 0)
 349                         return 1;
 350         } else if (ret == FIO_Q_BUSY) {
 351                 if (td_io_commit(td))
 352                         return 1;
 353                 goto requeue;
 354         }
 355
 356         return 0;
 357 }
 358
 359 static inline void update_tv_cache(struct thread_data *td)
 360 {
 361         if ((++td->tv_cache_nr & td->tv_cache_mask) == td->tv_cache_mask)
 362                 fio_gettime(&td->tv_cache, NULL);
 363 }
 364
 365 static int break_on_this_error(struct thread_data *td, int *retptr)
 366 {
 367         int ret = *retptr;
 368
 369         if (ret < 0 || td->error) {
 370                 int err;
 371
 372                 if (!td->o.continue_on_error)
 373                         return 1;
 374
 375                 if (ret < 0)
 376                         err = -ret;
 377                 else
 378                         err = td->error;
 379
 380                 if (td_non_fatal_error(err)) {
 381                         /*
 382                          * Continue with the I/Os in case of
 383                          * a non fatal error.
 384                          */
 385                         update_error_count(td, err);
 386                         td_clear_error(td);
 387                         *retptr = 0;
 388                         return 0;
 389                 } else if (td->o.fill_device && err == ENOSPC) {
 390                         /*
 391                          * We expect to hit this error if
 392                          * fill_device option is set.
 393                          */
 394                         td_clear_error(td);
 395                         td->terminate = 1;
 396                         return 1;
 397                 } else {
 398                         /*
 399                          * Stop the I/O in case of a fatal
 400                          * error.
 401                          */
 402                         update_error_count(td, err);
 403                         return 1;
 404                 }
 405         }
 406
 407         return 0;
 408 }
 409
 410 /*
 411  * The main verify engine. Runs over the writes we previously submitted,
 412  * reads the blocks back in, and checks the crc/md5 of the data.
 413  */
 414 static void do_verify(struct thread_data *td)
 415 {
 416         struct fio_file *f;
 417         struct io_u *io_u;
 418         int ret, min_events;
 419         unsigned int i;
 420
 421         dprint(FD_VERIFY, "starting loop\n");
 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                 if (td->o.verify_async)
 467                         io_u->end_io = verify_io_u_async;
 468                 else
 469                         io_u->end_io = verify_io_u;
 470
 471                 ret = td_io_queue(td, io_u);
 472                 switch (ret) {
 473                 case FIO_Q_COMPLETED:
 474                         if (io_u->error) {
 475                                 ret = -io_u->error;
 476                                 clear_io_u(td, io_u);
 477                         } else if (io_u->resid) {
 478                                 int bytes = io_u->xfer_buflen - io_u->resid;
 479                                 struct fio_file *f = io_u->file;
 480
 481                                 /*
 482                                  * zero read, fail
 483                                  */
 484                                 if (!bytes) {
 485                                         td_verror(td, EIO, "full resid");
 486                                         put_io_u(td, io_u);
 487                                         break;
 488                                 }
 489
 490                                 io_u->xfer_buflen = io_u->resid;
 491                                 io_u->xfer_buf += bytes;
 492                                 io_u->offset += bytes;
 493
 494                                 td->ts.short_io_u[io_u->ddir]++;
 495
 496                                 if (io_u->offset == f->real_file_size)
 497                                         goto sync_done;
 498
 499                                 requeue_io_u(td, &io_u);
 500                         } else {
 501 sync_done:
 502                                 ret = io_u_sync_complete(td, io_u, NULL);
 503                                 if (ret < 0)
 504                                         break;
 505                         }
 506                         continue;
 507                 case FIO_Q_QUEUED:
 508                         break;
 509                 case FIO_Q_BUSY:
 510                         requeue_io_u(td, &io_u);
 511                         ret2 = td_io_commit(td);
 512                         if (ret2 < 0)
 513                                 ret = ret2;
 514                         break;
 515                 default:
 516                         assert(ret < 0);
 517                         td_verror(td, -ret, "td_io_queue");
 518                         break;
 519                 }
 520
 521                 if (break_on_this_error(td, &ret))
 522                         break;
 523
 524                 /*
 525                  * if we can queue more, do so. but check if there are
 526                  * completed io_u's first.
 527                  */
 528                 full = queue_full(td) || ret == FIO_Q_BUSY;
 529                 if (full || !td->o.iodepth_batch_complete) {
 530                         min_events = min(td->o.iodepth_batch_complete,
 531                                          td->cur_depth);
 532                         if (full && !min_events)
 533                                 min_events = 1;
 534
 535                         do {
 536                                 /*
 537                                  * Reap required number of io units, if any,
 538                                  * and do the verification on them through
 539                                  * the callback handler
 540                                  */
 541                                 if (io_u_queued_complete(td, min_events, NULL) < 0) {
 542                                         ret = -1;
 543                                         break;
 544                                 }
 545                         } while (full && (td->cur_depth > td->o.iodepth_low));
 546                 }
 547                 if (ret < 0)
 548                         break;
 549         }
 550
 551         if (!td->error) {
 552                 min_events = td->cur_depth;
 553
 554                 if (min_events)
 555                         ret = io_u_queued_complete(td, min_events, NULL);
 556         } else
 557                 cleanup_pending_aio(td);
 558
 559         td_set_runstate(td, TD_RUNNING);
 560
 561         dprint(FD_VERIFY, "exiting loop\n");
 562 }
 563
 564 /*
 565  * Main IO worker function. It retrieves io_u's to process and queues
 566  * and reaps them, checking for rate and errors along the way.
 567  */
 568 static void do_io(struct thread_data *td)
 569 {
 570         unsigned int i;
 571         int ret = 0;
 572
 573         if (in_ramp_time(td))
 574                 td_set_runstate(td, TD_RAMP);
 575         else
 576                 td_set_runstate(td, TD_RUNNING);
 577
 578         while ( (td->o.read_iolog_file && !flist_empty(&td->io_log_list)) ||
 579                 ((td->this_io_bytes[0] + td->this_io_bytes[1]) < td->o.size) ) {
 580                 struct timeval comp_time;
 581                 unsigned long bytes_done[2] = { 0, 0 };
 582                 int min_evts = 0;
 583                 struct io_u *io_u;
 584                 int ret2, full;
 585
 586                 if (td->terminate)
 587                         break;
 588
 589                 update_tv_cache(td);
 590
 591                 if (runtime_exceeded(td, &td->tv_cache)) {
 592                         td->terminate = 1;
 593                         break;
 594                 }
 595
 596                 io_u = get_io_u(td);
 597                 if (!io_u)
 598                         break;
 599
 600                 /*
 601                  * Add verification end_io handler, if asked to verify
 602                  * a previously written file.
 603                  */
 604                 if (td->o.verify != VERIFY_NONE && io_u->ddir == DDIR_READ &&
 605                     !td_rw(td)) {
 606                         if (td->o.verify_async)
 607                                 io_u->end_io = verify_io_u_async;
 608                         else
 609                                 io_u->end_io = verify_io_u;
 610                         td_set_runstate(td, TD_VERIFYING);
 611                 } else if (in_ramp_time(td))
 612                         td_set_runstate(td, TD_RAMP);
 613                 else
 614                         td_set_runstate(td, TD_RUNNING);
 615
 616                 ret = td_io_queue(td, io_u);
 617                 switch (ret) {
 618                 case FIO_Q_COMPLETED:
 619                         if (io_u->error) {
 620                                 ret = -io_u->error;
 621                                 clear_io_u(td, io_u);
 622                         } else if (io_u->resid) {
 623                                 int bytes = io_u->xfer_buflen - io_u->resid;
 624                                 struct fio_file *f = io_u->file;
 625
 626                                 /*
 627                                  * zero read, fail
 628                                  */
 629                                 if (!bytes) {
 630                                         td_verror(td, EIO, "full resid");
 631                                         put_io_u(td, io_u);
 632                                         break;
 633                                 }
 634
 635                                 io_u->xfer_buflen = io_u->resid;
 636                                 io_u->xfer_buf += bytes;
 637                                 io_u->offset += bytes;
 638
 639                                 td->ts.short_io_u[io_u->ddir]++;
 640
 641                                 if (io_u->offset == f->real_file_size)
 642                                         goto sync_done;
 643
 644                                 requeue_io_u(td, &io_u);
 645                         } else {
 646 sync_done:
 647                                 if (__should_check_rate(td, 0) ||
 648                                     __should_check_rate(td, 1))
 649                                         fio_gettime(&comp_time, NULL);
 650
 651                                 ret = io_u_sync_complete(td, io_u, bytes_done);
 652                                 if (ret < 0)
 653                                         break;
 654                         }
 655                         break;
 656                 case FIO_Q_QUEUED:
 657                         /*
 658                          * if the engine doesn't have a commit hook,
 659                          * the io_u is really queued. if it does have such
 660                          * a hook, it has to call io_u_queued() itself.
 661                          */
 662                         if (td->io_ops->commit == NULL)
 663                                 io_u_queued(td, io_u);
 664                         break;
 665                 case FIO_Q_BUSY:
 666                         requeue_io_u(td, &io_u);
 667                         ret2 = td_io_commit(td);
 668                         if (ret2 < 0)
 669                                 ret = ret2;
 670                         break;
 671                 default:
 672                         assert(ret < 0);
 673                         put_io_u(td, io_u);
 674                         break;
 675                 }
 676
 677                 if (break_on_this_error(td, &ret))
 678                         break;
 679
 680                 /*
 681                  * See if we need to complete some commands
 682                  */
 683                 full = queue_full(td) || ret == FIO_Q_BUSY;
 684                 if (full || !td->o.iodepth_batch_complete) {
 685                         min_evts = min(td->o.iodepth_batch_complete,
 686                                         td->cur_depth);
 687                         if (full && !min_evts)
 688                                 min_evts = 1;
 689
 690                         if (__should_check_rate(td, 0) ||
 691                             __should_check_rate(td, 1))
 692                                 fio_gettime(&comp_time, NULL);
 693
 694                         do {
 695                                 ret = io_u_queued_complete(td, min_evts, bytes_done);
 696                                 if (ret < 0)
 697                                         break;
 698
 699                         } while (full && (td->cur_depth > td->o.iodepth_low));
 700                 }
 701
 702                 if (ret < 0)
 703                         break;
 704                 if (!(bytes_done[0] + bytes_done[1]))
 705                         continue;
 706
 707                 if (!in_ramp_time(td) && should_check_rate(td, bytes_done)) {
 708                         if (check_min_rate(td, &comp_time, bytes_done)) {
 709                                 if (exitall_on_terminate)
 710                                         terminate_threads(td->groupid);
 711                                 td_verror(td, EIO, "check_min_rate");
 712                                 break;
 713                         }
 714                 }
 715
 716                 if (td->o.thinktime) {
 717                         unsigned long long b;
 718
 719                         b = td->io_blocks[0] + td->io_blocks[1];
 720                         if (!(b % td->o.thinktime_blocks)) {
 721                                 int left;
 722
 723                                 if (td->o.thinktime_spin)
 724                                         usec_spin(td->o.thinktime_spin);
 725
 726                                 left = td->o.thinktime - td->o.thinktime_spin;
 727                                 if (left)
 728                                         usec_sleep(td, left);
 729                         }
 730                 }
 731         }
 732
 733         if (td->o.fill_device && td->error == ENOSPC) {
 734                 td->error = 0;
 735                 td->terminate = 1;
 736         }
 737         if (!td->error) {
 738                 struct fio_file *f;
 739
 740                 i = td->cur_depth;
 741                 if (i)
 742                         ret = io_u_queued_complete(td, i, NULL);
 743
 744                 if (should_fsync(td) && td->o.end_fsync) {
 745                         td_set_runstate(td, TD_FSYNCING);
 746
 747                         for_each_file(td, f, i) {
 748                                 if (!fio_file_open(f))
 749                                         continue;
 750                                 fio_io_sync(td, f);
 751                         }
 752                 }
 753         } else
 754                 cleanup_pending_aio(td);
 755
 756         /*
 757          * stop job if we failed doing any IO
 758          */
 759         if ((td->this_io_bytes[0] + td->this_io_bytes[1]) == 0)
 760                 td->done = 1;
 761 }
 762
 763 static void cleanup_io_u(struct thread_data *td)
 764 {
 765         struct flist_head *entry, *n;
 766         struct io_u *io_u;
 767
 768         flist_for_each_safe(entry, n, &td->io_u_freelist) {
 769                 io_u = flist_entry(entry, struct io_u, list);
 770
 771                 flist_del(&io_u->list);
 772                 free(io_u);
 773         }
 774
 775         free_io_mem(td);
 776 }
 777
 778 static int init_io_u(struct thread_data *td)
 779 {
 780         struct io_u *io_u;
 781         unsigned int max_bs;
 782         int cl_align, i, max_units;
 783         char *p;
 784
 785         max_units = td->o.iodepth;
 786         max_bs = max(td->o.max_bs[DDIR_READ], td->o.max_bs[DDIR_WRITE]);
 787         td->orig_buffer_size = (unsigned long long) max_bs
 788                                         * (unsigned long long) max_units;
 789
 790         if (td->o.mem_type == MEM_SHMHUGE || td->o.mem_type == MEM_MMAPHUGE) {
 791                 unsigned long bs;
 792
 793                 bs = td->orig_buffer_size + td->o.hugepage_size - 1;
 794                 td->orig_buffer_size = bs & ~(td->o.hugepage_size - 1);
 795         }
 796
 797         if (td->orig_buffer_size != (size_t) td->orig_buffer_size) {
 798                 log_err("fio: IO memory too large. Reduce max_bs or iodepth\n");
 799                 return 1;
 800         }
 801
 802         if (allocate_io_mem(td))
 803                 return 1;
 804
 805         if (td->o.odirect || td->o.mem_align)
 806                 p = PAGE_ALIGN(td->orig_buffer) + td->o.mem_align;
 807         else
 808                 p = td->orig_buffer;
 809
 810         cl_align = os_cache_line_size();
 811
 812         for (i = 0; i < max_units; i++) {
 813                 void *ptr;
 814
 815                 if (td->terminate)
 816                         return 1;
 817
 818                 if (posix_memalign(&ptr, cl_align, sizeof(*io_u))) {
 819                         log_err("fio: posix_memalign=%s\n", strerror(errno));
 820                         break;
 821                 }
 822
 823                 io_u = ptr;
 824                 memset(io_u, 0, sizeof(*io_u));
 825                 INIT_FLIST_HEAD(&io_u->list);
 826                 dprint(FD_MEM, "io_u alloc %p, index %u\n", io_u, i);
 827
 828                 if (!(td->io_ops->flags & FIO_NOIO)) {
 829                         io_u->buf = p + max_bs * i;
 830                         dprint(FD_MEM, "io_u %p, mem %p\n", io_u, io_u->buf);
 831
 832                         if (td_write(td) && !td->o.refill_buffers)
 833                                 io_u_fill_buffer(td, io_u, max_bs);
 834                         else if (td_write(td) && td->o.verify_pattern_bytes) {
 835                                 /*
 836                                  * Fill the buffer with the pattern if we are
 837                                  * going to be doing writes.
 838                                  */
 839                                 fill_pattern(td, io_u->buf, max_bs, io_u);
 840                         }
 841                 }
 842
 843                 io_u->index = i;
 844                 io_u->flags = IO_U_F_FREE;
 845                 flist_add(&io_u->list, &td->io_u_freelist);
 846         }
 847
 848         return 0;
 849 }
 850
 851 static int switch_ioscheduler(struct thread_data *td)
 852 {
 853         char tmp[256], tmp2[128];
 854         FILE *f;
 855         int ret;
 856
 857         if (td->io_ops->flags & FIO_DISKLESSIO)
 858                 return 0;
 859
 860         sprintf(tmp, "%s/queue/scheduler", td->sysfs_root);
 861
 862         f = fopen(tmp, "r+");
 863         if (!f) {
 864                 if (errno == ENOENT) {
 865                         log_err("fio: os or kernel doesn't support IO scheduler"
 866                                 " switching\n");
 867                         return 0;
 868                 }
 869                 td_verror(td, errno, "fopen iosched");
 870                 return 1;
 871         }
 872
 873         /*
 874          * Set io scheduler.
 875          */
 876         ret = fwrite(td->o.ioscheduler, strlen(td->o.ioscheduler), 1, f);
 877         if (ferror(f) || ret != 1) {
 878                 td_verror(td, errno, "fwrite");
 879                 fclose(f);
 880                 return 1;
 881         }
 882
 883         rewind(f);
 884
 885         /*
 886          * Read back and check that the selected scheduler is now the default.
 887          */
 888         ret = fread(tmp, 1, sizeof(tmp), f);
 889         if (ferror(f) || ret < 0) {
 890                 td_verror(td, errno, "fread");
 891                 fclose(f);
 892                 return 1;
 893         }
 894
 895         sprintf(tmp2, "[%s]", td->o.ioscheduler);
 896         if (!strstr(tmp, tmp2)) {
 897                 log_err("fio: io scheduler %s not found\n", td->o.ioscheduler);
 898                 td_verror(td, EINVAL, "iosched_switch");
 899                 fclose(f);
 900                 return 1;
 901         }
 902
 903         fclose(f);
 904         return 0;
 905 }
 906
 907 static int keep_running(struct thread_data *td)
 908 {
 909         unsigned long long io_done;
 910
 911         if (td->done)
 912                 return 0;
 913         if (td->o.time_based)
 914                 return 1;
 915         if (td->o.loops) {
 916                 td->o.loops--;
 917                 return 1;
 918         }
 919
 920         io_done = td->io_bytes[DDIR_READ] + td->io_bytes[DDIR_WRITE]
 921                         + td->io_skip_bytes;
 922         if (io_done < td->o.size)
 923                 return 1;
 924
 925         return 0;
 926 }
 927
 928 static void reset_io_counters(struct thread_data *td)
 929 {
 930         td->ts.stat_io_bytes[0] = td->ts.stat_io_bytes[1] = 0;
 931         td->this_io_bytes[0] = td->this_io_bytes[1] = 0;
 932         td->zone_bytes = 0;
 933         td->rate_bytes[0] = td->rate_bytes[1] = 0;
 934         td->rate_blocks[0] = td->rate_blocks[1] = 0;
 935
 936         td->last_was_sync = 0;
 937
 938         /*
 939          * reset file done count if we are to start over
 940          */
 941         if (td->o.time_based || td->o.loops)
 942                 td->nr_done_files = 0;
 943
 944         /*
 945          * Set the same seed to get repeatable runs
 946          */
 947         td_fill_rand_seeds(td);
 948 }
 949
 950 void reset_all_stats(struct thread_data *td)
 951 {
 952         struct timeval tv;
 953         int i;
 954
 955         reset_io_counters(td);
 956
 957         for (i = 0; i < 2; i++) {
 958                 td->io_bytes[i] = 0;
 959                 td->io_blocks[i] = 0;
 960                 td->io_issues[i] = 0;
 961                 td->ts.total_io_u[i] = 0;
 962         }
 963
 964         fio_gettime(&tv, NULL);
 965         memcpy(&td->epoch, &tv, sizeof(tv));
 966         memcpy(&td->start, &tv, sizeof(tv));
 967 }
 968
 969 static void clear_io_state(struct thread_data *td)
 970 {
 971         struct fio_file *f;
 972         unsigned int i;
 973
 974         reset_io_counters(td);
 975
 976         close_files(td);
 977         for_each_file(td, f, i)
 978                 fio_file_clear_done(f);
 979 }
 980
 981 static int exec_string(const char *string)
 982 {
 983         int ret, newlen = strlen(string) + 1 + 8;
 984         char *str;
 985
 986         str = malloc(newlen);
 987         sprintf(str, "sh -c %s", string);
 988
 989         ret = system(str);
 990         if (ret == -1)
 991                 log_err("fio: exec of cmd <%s> failed\n", str);
 992
 993         free(str);
 994         return ret;
 995 }
 996
 997 /*
 998  * Entry point for the thread based jobs. The process based jobs end up
 999  * here as well, after a little setup.
1000  */
1001 static void *thread_main(void *data)
1002 {
1003         unsigned long long runtime[2], elapsed;
1004         struct thread_data *td = data;
1005         pthread_condattr_t attr;
1006         int clear_state;
1007
1008         if (!td->o.use_thread)
1009                 setsid();
1010
1011         td->pid = getpid();
1012
1013         dprint(FD_PROCESS, "jobs pid=%d started\n", (int) td->pid);
1014
1015         INIT_FLIST_HEAD(&td->io_u_freelist);
1016         INIT_FLIST_HEAD(&td->io_u_busylist);
1017         INIT_FLIST_HEAD(&td->io_u_requeues);
1018         INIT_FLIST_HEAD(&td->io_log_list);
1019         INIT_FLIST_HEAD(&td->io_hist_list);
1020         INIT_FLIST_HEAD(&td->verify_list);
1021         pthread_mutex_init(&td->io_u_lock, NULL);
1022         td->io_hist_tree = RB_ROOT;
1023
1024         pthread_condattr_init(&attr);
1025         pthread_cond_init(&td->verify_cond, &attr);
1026         pthread_cond_init(&td->free_cond, &attr);
1027
1028         td_set_runstate(td, TD_INITIALIZED);
1029         dprint(FD_MUTEX, "up startup_mutex\n");
1030         fio_mutex_up(startup_mutex);
1031         dprint(FD_MUTEX, "wait on td->mutex\n");
1032         fio_mutex_down(td->mutex);
1033         dprint(FD_MUTEX, "done waiting on td->mutex\n");
1034
1035         /*
1036          * the ->mutex mutex is now no longer used, close it to avoid
1037          * eating a file descriptor
1038          */
1039         fio_mutex_remove(td->mutex);
1040
1041         if (td->o.uid != -1U && setuid(td->o.uid)) {
1042                 td_verror(td, errno, "setuid");
1043                 goto err;
1044         }
1045         if (td->o.gid != -1U && setgid(td->o.gid)) {
1046                 td_verror(td, errno, "setgid");
1047                 goto err;
1048         }
1049
1050         /*
1051          * May alter parameters that init_io_u() will use, so we need to
1052          * do this first.
1053          */
1054         if (init_iolog(td))
1055                 goto err;
1056
1057         if (init_io_u(td))
1058                 goto err;
1059
1060         if (td->o.verify_async && verify_async_init(td))
1061                 goto err;
1062
1063         if (td->o.cpumask_set && fio_setaffinity(td->pid, td->o.cpumask) == -1) {
1064                 td_verror(td, errno, "cpu_set_affinity");
1065                 goto err;
1066         }
1067
1068         /*
1069          * If we have a gettimeofday() thread, make sure we exclude that
1070          * thread from this job
1071          */
1072         if (td->o.gtod_cpu) {
1073                 fio_cpu_clear(&td->o.cpumask, td->o.gtod_cpu);
1074                 if (fio_setaffinity(td->pid, td->o.cpumask) == -1) {
1075                         td_verror(td, errno, "cpu_set_affinity");
1076                         goto err;
1077                 }
1078         }
1079
1080         if (td->ioprio_set) {
1081                 if (ioprio_set(IOPRIO_WHO_PROCESS, 0, td->ioprio) == -1) {
1082                         td_verror(td, errno, "ioprio_set");
1083                         goto err;
1084                 }
1085         }
1086
1087         if (td->o.cgroup_weight && cgroup_setup(td, cgroup_list, &cgroup_mnt))
1088                 goto err;
1089
1090         if (nice(td->o.nice) == -1) {
1091                 td_verror(td, errno, "nice");
1092                 goto err;
1093         }
1094
1095         if (td->o.ioscheduler && switch_ioscheduler(td))
1096                 goto err;
1097
1098         if (!td->o.create_serialize && setup_files(td))
1099                 goto err;
1100
1101         if (td_io_init(td))
1102                 goto err;
1103
1104         if (init_random_map(td))
1105                 goto err;
1106
1107         if (td->o.exec_prerun) {
1108                 if (exec_string(td->o.exec_prerun))
1109                         goto err;
1110         }
1111
1112         if (td->o.pre_read) {
1113                 if (pre_read_files(td) < 0)
1114                         goto err;
1115         }
1116
1117         fio_gettime(&td->epoch, NULL);
1118         getrusage(RUSAGE_SELF, &td->ts.ru_start);
1119
1120         runtime[0] = runtime[1] = 0;
1121         clear_state = 0;
1122         while (keep_running(td)) {
1123                 fio_gettime(&td->start, NULL);
1124                 memcpy(&td->ts.stat_sample_time[0], &td->start,
1125                                 sizeof(td->start));
1126                 memcpy(&td->ts.stat_sample_time[1], &td->start,
1127                                 sizeof(td->start));
1128                 memcpy(&td->tv_cache, &td->start, sizeof(td->start));
1129
1130                 if (td->o.ratemin[0] || td->o.ratemin[1])
1131                         memcpy(&td->lastrate, &td->ts.stat_sample_time,
1132                                                         sizeof(td->lastrate));
1133
1134                 if (clear_state)
1135                         clear_io_state(td);
1136
1137                 prune_io_piece_log(td);
1138
1139                 do_io(td);
1140
1141                 clear_state = 1;
1142
1143                 if (td_read(td) && td->io_bytes[DDIR_READ]) {
1144                         elapsed = utime_since_now(&td->start);
1145                         runtime[DDIR_READ] += elapsed;
1146                 }
1147                 if (td_write(td) && td->io_bytes[DDIR_WRITE]) {
1148                         elapsed = utime_since_now(&td->start);
1149                         runtime[DDIR_WRITE] += elapsed;
1150                 }
1151
1152                 if (td->error || td->terminate)
1153                         break;
1154
1155                 if (!td->o.do_verify ||
1156                     td->o.verify == VERIFY_NONE ||
1157                     (td->io_ops->flags & FIO_UNIDIR))
1158                         continue;
1159
1160                 clear_io_state(td);
1161
1162                 fio_gettime(&td->start, NULL);
1163
1164                 do_verify(td);
1165
1166                 runtime[DDIR_READ] += utime_since_now(&td->start);
1167
1168                 if (td->error || td->terminate)
1169                         break;
1170         }
1171
1172         update_rusage_stat(td);
1173         td->ts.runtime[0] = (runtime[0] + 999) / 1000;
1174         td->ts.runtime[1] = (runtime[1] + 999) / 1000;
1175         td->ts.total_run_time = mtime_since_now(&td->epoch);
1176         td->ts.io_bytes[0] = td->io_bytes[0];
1177         td->ts.io_bytes[1] = td->io_bytes[1];
1178
1179         fio_mutex_down(writeout_mutex);
1180         if (td->ts.bw_log) {
1181                 if (td->o.bw_log_file) {
1182                         finish_log_named(td, td->ts.bw_log,
1183                                                 td->o.bw_log_file, "bw");
1184                 } else
1185                         finish_log(td, td->ts.bw_log, "bw");
1186         }
1187         if (td->ts.lat_log) {
1188                 if (td->o.lat_log_file) {
1189                         finish_log_named(td, td->ts.lat_log,
1190                                                 td->o.lat_log_file, "lat");
1191                 } else
1192                         finish_log(td, td->ts.lat_log, "lat");
1193         }
1194         if (td->ts.slat_log) {
1195                 if (td->o.lat_log_file) {
1196                         finish_log_named(td, td->ts.slat_log,
1197                                                 td->o.lat_log_file, "slat");
1198                 } else
1199                         finish_log(td, td->ts.slat_log, "slat");
1200         }
1201         if (td->ts.clat_log) {
1202                 if (td->o.lat_log_file) {
1203                         finish_log_named(td, td->ts.clat_log,
1204                                                 td->o.lat_log_file, "clat");
1205                 } else
1206                         finish_log(td, td->ts.clat_log, "clat");
1207         }
1208         fio_mutex_up(writeout_mutex);
1209         if (td->o.exec_postrun)
1210                 exec_string(td->o.exec_postrun);
1211
1212         if (exitall_on_terminate)
1213                 terminate_threads(td->groupid);
1214
1215 err:
1216         if (td->error)
1217                 log_info("fio: pid=%d, err=%d/%s\n", (int) td->pid, td->error,
1218                                                         td->verror);
1219
1220         if (td->o.verify_async)
1221                 verify_async_exit(td);
1222
1223         close_and_free_files(td);
1224         close_ioengine(td);
1225         cleanup_io_u(td);
1226         cgroup_shutdown(td, &cgroup_mnt);
1227
1228         if (td->o.cpumask_set) {
1229                 int ret = fio_cpuset_exit(&td->o.cpumask);
1230
1231                 td_verror(td, ret, "fio_cpuset_exit");
1232         }
1233
1234         /*
1235          * do this very late, it will log file closing as well
1236          */
1237         if (td->o.write_iolog_file)
1238                 write_iolog_close(td);
1239
1240         options_mem_free(td);
1241         td_set_runstate(td, TD_EXITED);
1242         return (void *) (unsigned long) td->error;
1243 }
1244
1245 /*
1246  * We cannot pass the td data into a forked process, so attach the td and
1247  * pass it to the thread worker.
1248  */
1249 static int fork_main(int shmid, int offset)
1250 {
1251         struct thread_data *td;
1252         void *data, *ret;
1253
1254         data = shmat(shmid, NULL, 0);
1255         if (data == (void *) -1) {
1256                 int __err = errno;
1257
1258                 perror("shmat");
1259                 return __err;
1260         }
1261
1262         td = data + offset * sizeof(struct thread_data);
1263         ret = thread_main(td);
1264         shmdt(data);
1265         return (int) (unsigned long) ret;
1266 }
1267
1268 /*
1269  * Run over the job map and reap the threads that have exited, if any.
1270  */
1271 static void reap_threads(int *nr_running, int *t_rate, int *m_rate)
1272 {
1273         struct thread_data *td;
1274         int i, cputhreads, realthreads, pending, status, ret;
1275
1276         /*
1277          * reap exited threads (TD_EXITED -> TD_REAPED)
1278          */
1279         realthreads = pending = cputhreads = 0;
1280         for_each_td(td, i) {
1281                 int flags = 0;
1282
1283                 /*
1284                  * ->io_ops is NULL for a thread that has closed its
1285                  * io engine
1286                  */
1287                 if (td->io_ops && !strcmp(td->io_ops->name, "cpuio"))
1288                         cputhreads++;
1289                 else
1290                         realthreads++;
1291
1292                 if (!td->pid) {
1293                         pending++;
1294                         continue;
1295                 }
1296                 if (td->runstate == TD_REAPED)
1297                         continue;
1298                 if (td->o.use_thread) {
1299                         if (td->runstate == TD_EXITED) {
1300                                 td_set_runstate(td, TD_REAPED);
1301                                 goto reaped;
1302                         }
1303                         continue;
1304                 }
1305
1306                 flags = WNOHANG;
1307                 if (td->runstate == TD_EXITED)
1308                         flags = 0;
1309
1310                 /*
1311                  * check if someone quit or got killed in an unusual way
1312                  */
1313                 ret = waitpid(td->pid, &status, flags);
1314                 if (ret < 0) {
1315                         if (errno == ECHILD) {
1316                                 log_err("fio: pid=%d disappeared %d\n",
1317                                                 (int) td->pid, td->runstate);
1318                                 td_set_runstate(td, TD_REAPED);
1319                                 goto reaped;
1320                         }
1321                         perror("waitpid");
1322                 } else if (ret == td->pid) {
1323                         if (WIFSIGNALED(status)) {
1324                                 int sig = WTERMSIG(status);
1325
1326                                 if (sig != SIGQUIT)
1327                                         log_err("fio: pid=%d, got signal=%d\n",
1328                                                         (int) td->pid, sig);
1329                                 td_set_runstate(td, TD_REAPED);
1330                                 goto reaped;
1331                         }
1332                         if (WIFEXITED(status)) {
1333                                 if (WEXITSTATUS(status) && !td->error)
1334                                         td->error = WEXITSTATUS(status);
1335
1336                                 td_set_runstate(td, TD_REAPED);
1337                                 goto reaped;
1338                         }
1339                 }
1340
1341                 /*
1342                  * thread is not dead, continue
1343                  */
1344                 pending++;
1345                 continue;
1346 reaped:
1347                 (*nr_running)--;
1348                 (*m_rate) -= (td->o.ratemin[0] + td->o.ratemin[1]);
1349                 (*t_rate) -= (td->o.rate[0] + td->o.rate[1]);
1350                 if (!td->pid)
1351                         pending--;
1352
1353                 if (td->error)
1354                         exit_value++;
1355
1356                 done_secs += mtime_since_now(&td->epoch) / 1000;
1357         }
1358
1359         if (*nr_running == cputhreads && !pending && realthreads)
1360                 terminate_threads(TERMINATE_ALL);
1361 }
1362
1363 static void *gtod_thread_main(void *data)
1364 {
1365         fio_mutex_up(startup_mutex);
1366
1367         /*
1368          * As long as we have jobs around, update the clock. It would be nice
1369          * to have some way of NOT hammering that CPU with gettimeofday(),
1370          * but I'm not sure what to use outside of a simple CPU nop to relax
1371          * it - we don't want to lose precision.
1372          */
1373         while (threads) {
1374                 fio_gtod_update();
1375                 nop;
1376         }
1377
1378         return NULL;
1379 }
1380
1381 static int fio_start_gtod_thread(void)
1382 {
1383         pthread_attr_t attr;
1384         int ret;
1385
1386         pthread_attr_init(&attr);
1387         pthread_attr_setstacksize(&attr, PTHREAD_STACK_MIN);
1388         ret = pthread_create(&gtod_thread, &attr, gtod_thread_main, NULL);
1389         pthread_attr_destroy(&attr);
1390         if (ret) {
1391                 log_err("Can't create gtod thread: %s\n", strerror(ret));
1392                 return 1;
1393         }
1394
1395         ret = pthread_detach(gtod_thread);
1396         if (ret) {
1397                 log_err("Can't detatch gtod thread: %s\n", strerror(ret));
1398                 return 1;
1399         }
1400
1401         dprint(FD_MUTEX, "wait on startup_mutex\n");
1402         fio_mutex_down(startup_mutex);
1403         dprint(FD_MUTEX, "done waiting on startup_mutex\n");
1404         return 0;
1405 }
1406
1407 /*
1408  * Main function for kicking off and reaping jobs, as needed.
1409  */
1410 static void run_threads(void)
1411 {
1412         struct thread_data *td;
1413         unsigned long spent;
1414         int i, todo, nr_running, m_rate, t_rate, nr_started;
1415
1416         if (fio_pin_memory())
1417                 return;
1418
1419         if (fio_gtod_offload && fio_start_gtod_thread())
1420                 return;
1421
1422         if (!terse_output) {
1423                 log_info("Starting ");
1424                 if (nr_thread)
1425                         log_info("%d thread%s", nr_thread,
1426                                                 nr_thread > 1 ? "s" : "");
1427                 if (nr_process) {
1428                         if (nr_thread)
1429                                 printf(" and ");
1430                         log_info("%d process%s", nr_process,
1431                                                 nr_process > 1 ? "es" : "");
1432                 }
1433                 log_info("\n");
1434                 fflush(stdout);
1435         }
1436
1437         set_sig_handlers();
1438
1439         todo = thread_number;
1440         nr_running = 0;
1441         nr_started = 0;
1442         m_rate = t_rate = 0;
1443
1444         for_each_td(td, i) {
1445                 print_status_init(td->thread_number - 1);
1446
1447                 if (!td->o.create_serialize) {
1448                         init_disk_util(td);
1449                         continue;
1450                 }
1451
1452                 /*
1453                  * do file setup here so it happens sequentially,
1454                  * we don't want X number of threads getting their
1455                  * client data interspersed on disk
1456                  */
1457                 if (setup_files(td)) {
1458                         exit_value++;
1459                         if (td->error)
1460                                 log_err("fio: pid=%d, err=%d/%s\n",
1461                                         (int) td->pid, td->error, td->verror);
1462                         td_set_runstate(td, TD_REAPED);
1463                         todo--;
1464                 } else {
1465                         struct fio_file *f;
1466                         unsigned int i;
1467
1468                         /*
1469                          * for sharing to work, each job must always open
1470                          * its own files. so close them, if we opened them
1471                          * for creation
1472                          */
1473                         for_each_file(td, f, i) {
1474                                 if (fio_file_open(f))
1475                                         td_io_close_file(td, f);
1476                         }
1477                 }
1478
1479                 init_disk_util(td);
1480         }
1481
1482         set_genesis_time();
1483
1484         while (todo) {
1485                 struct thread_data *map[MAX_JOBS];
1486                 struct timeval this_start;
1487                 int this_jobs = 0, left;
1488
1489                 /*
1490                  * create threads (TD_NOT_CREATED -> TD_CREATED)
1491                  */
1492                 for_each_td(td, i) {
1493                         if (td->runstate != TD_NOT_CREATED)
1494                                 continue;
1495
1496                         /*
1497                          * never got a chance to start, killed by other
1498                          * thread for some reason
1499                          */
1500                         if (td->terminate) {
1501                                 todo--;
1502                                 continue;
1503                         }
1504
1505                         if (td->o.start_delay) {
1506                                 spent = mtime_since_genesis();
1507
1508                                 if (td->o.start_delay * 1000 > spent)
1509                                         continue;
1510                         }
1511
1512                         if (td->o.stonewall && (nr_started || nr_running)) {
1513                                 dprint(FD_PROCESS, "%s: stonewall wait\n",
1514                                                         td->o.name);
1515                                 break;
1516                         }
1517
1518                         /*
1519                          * Set state to created. Thread will transition
1520                          * to TD_INITIALIZED when it's done setting up.
1521                          */
1522                         td_set_runstate(td, TD_CREATED);
1523                         map[this_jobs++] = td;
1524                         nr_started++;
1525
1526                         if (td->o.use_thread) {
1527                                 int ret;
1528
1529                                 dprint(FD_PROCESS, "will pthread_create\n");
1530                                 ret = pthread_create(&td->thread, NULL,
1531                                                         thread_main, td);
1532                                 if (ret) {
1533                                         log_err("pthread_create: %s\n",
1534                                                         strerror(ret));
1535                                         nr_started--;
1536                                         break;
1537                                 }
1538                                 ret = pthread_detach(td->thread);
1539                                 if (ret)
1540                                         log_err("pthread_detach: %s",
1541                                                         strerror(ret));
1542                         } else {
1543                                 pid_t pid;
1544                                 dprint(FD_PROCESS, "will fork\n");
1545                                 pid = fork();
1546                                 if (!pid) {
1547                                         int ret = fork_main(shm_id, i);
1548
1549                                         _exit(ret);
1550                                 } else if (i == fio_debug_jobno)
1551                                         *fio_debug_jobp = pid;
1552                         }
1553                         dprint(FD_MUTEX, "wait on startup_mutex\n");
1554                         if (fio_mutex_down_timeout(startup_mutex, 10)) {
1555                                 log_err("fio: job startup hung? exiting.\n");
1556                                 terminate_threads(TERMINATE_ALL);
1557                                 fio_abort = 1;
1558                                 nr_started--;
1559                                 break;
1560                         }
1561                         dprint(FD_MUTEX, "done waiting on startup_mutex\n");
1562                 }
1563
1564                 /*
1565                  * Wait for the started threads to transition to
1566                  * TD_INITIALIZED.
1567                  */
1568                 fio_gettime(&this_start, NULL);
1569                 left = this_jobs;
1570                 while (left && !fio_abort) {
1571                         if (mtime_since_now(&this_start) > JOB_START_TIMEOUT)
1572                                 break;
1573
1574                         usleep(100000);
1575
1576                         for (i = 0; i < this_jobs; i++) {
1577                                 td = map[i];
1578                                 if (!td)
1579                                         continue;
1580                                 if (td->runstate == TD_INITIALIZED) {
1581                                         map[i] = NULL;
1582                                         left--;
1583                                 } else if (td->runstate >= TD_EXITED) {
1584                                         map[i] = NULL;
1585                                         left--;
1586                                         todo--;
1587                                         nr_running++; /* work-around... */
1588                                 }
1589                         }
1590                 }
1591
1592                 if (left) {
1593                         log_err("fio: %d jobs failed to start\n", left);
1594                         for (i = 0; i < this_jobs; i++) {
1595                                 td = map[i];
1596                                 if (!td)
1597                                         continue;
1598                                 kill(td->pid, SIGTERM);
1599                         }
1600                         break;
1601                 }
1602
1603                 /*
1604                  * start created threads (TD_INITIALIZED -> TD_RUNNING).
1605                  */
1606                 for_each_td(td, i) {
1607                         if (td->runstate != TD_INITIALIZED)
1608                                 continue;
1609
1610                         if (in_ramp_time(td))
1611                                 td_set_runstate(td, TD_RAMP);
1612                         else
1613                                 td_set_runstate(td, TD_RUNNING);
1614                         nr_running++;
1615                         nr_started--;
1616                         m_rate += td->o.ratemin[0] + td->o.ratemin[1];
1617                         t_rate += td->o.rate[0] + td->o.rate[1];
1618                         todo--;
1619                         fio_mutex_up(td->mutex);
1620                 }
1621
1622                 reap_threads(&nr_running, &t_rate, &m_rate);
1623
1624                 if (todo)
1625                         usleep(100000);
1626         }
1627
1628         while (nr_running) {
1629                 reap_threads(&nr_running, &t_rate, &m_rate);
1630                 usleep(10000);
1631         }
1632
1633         update_io_ticks();
1634         fio_unpin_memory();
1635 }
1636
1637 int main(int argc, char *argv[])
1638 {
1639         long ps;
1640
1641         sinit();
1642         init_rand(&__fio_rand_state);
1643
1644         /*
1645          * We need locale for number printing, if it isn't set then just
1646          * go with the US format.
1647          */
1648         if (!getenv("LC_NUMERIC"))
1649                 setlocale(LC_NUMERIC, "en_US");
1650
1651         ps = sysconf(_SC_PAGESIZE);
1652         if (ps < 0) {
1653                 log_err("Failed to get page size\n");
1654                 return 1;
1655         }
1656
1657         page_size = ps;
1658         page_mask = ps - 1;
1659
1660         fio_keywords_init();
1661
1662         if (parse_options(argc, argv))
1663                 return 1;
1664
1665         if (exec_profile && load_profile(exec_profile))
1666                 return 1;
1667
1668         if (!thread_number)
1669                 return 0;
1670
1671         if (write_bw_log) {
1672                 setup_log(&agg_io_log[DDIR_READ]);
1673                 setup_log(&agg_io_log[DDIR_WRITE]);
1674         }
1675
1676         startup_mutex = fio_mutex_init(0);
1677         writeout_mutex = fio_mutex_init(1);
1678
1679         set_genesis_time();
1680
1681         status_timer_arm();
1682
1683         cgroup_list = smalloc(sizeof(*cgroup_list));
1684         INIT_FLIST_HEAD(cgroup_list);
1685
1686         run_threads();
1687
1688         if (!fio_abort) {
1689                 show_run_stats();
1690                 if (write_bw_log) {
1691                         __finish_log(agg_io_log[DDIR_READ], "agg-read_bw.log");
1692                         __finish_log(agg_io_log[DDIR_WRITE],
1693                                         "agg-write_bw.log");
1694                 }
1695         }
1696
1697         cgroup_kill(cgroup_list);
1698         sfree(cgroup_list);
1699         sfree(cgroup_mnt);
1700
1701         fio_mutex_remove(startup_mutex);
1702         fio_mutex_remove(writeout_mutex);
1703         return exit_value;
1704 }