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
  44 unsigned long page_mask;
  45 unsigned long page_size;
  46
  47 #define PAGE_ALIGN(buf) \
  48         (char *) (((unsigned long) (buf) + page_mask) & ~page_mask)
  49
  50 int groupid = 0;
  51 int thread_number = 0;
  52 int nr_process = 0;
  53 int nr_thread = 0;
  54 int shm_id = 0;
  55 int temp_stall_ts;
  56 unsigned long done_secs = 0;
  57
  58 static struct fio_mutex *startup_mutex;
  59 static struct fio_mutex *writeout_mutex;
  60 static volatile int fio_abort;
  61 static int exit_value;
  62 static struct itimerval itimer;
  63 static pthread_t gtod_thread;
  64 static struct flist_head *cgroup_list;
  65
  66 struct io_log *agg_io_log[2];
  67
  68 #define TERMINATE_ALL           (-1)
  69 #define JOB_START_TIMEOUT       (5 * 1000)
  70
  71 void td_set_runstate(struct thread_data *td, int runstate)
  72 {
  73         if (td->runstate == runstate)
  74                 return;
  75
  76         dprint(FD_PROCESS, "pid=%d: runstate %d -> %d\n", (int) td->pid,
  77                                                 td->runstate, runstate);
  78         td->runstate = runstate;
  79 }
  80
  81 static void terminate_threads(int group_id)
  82 {
  83         struct thread_data *td;
  84         int i;
  85
  86         dprint(FD_PROCESS, "terminate group_id=%d\n", group_id);
  87
  88         for_each_td(td, i) {
  89                 if (group_id == TERMINATE_ALL || groupid == td->groupid) {
  90                         dprint(FD_PROCESS, "setting terminate on %s/%d\n",
  91                                                 td->o.name, (int) td->pid);
  92                         td->terminate = 1;
  93                         td->o.start_delay = 0;
  94
  95                         /*
  96                          * if the thread is running, just let it exit
  97                          */
  98                         if (td->runstate < TD_RUNNING)
  99                                 kill(td->pid, SIGQUIT);
 100                         else {
 101                                 struct ioengine_ops *ops = td->io_ops;
 102
 103                                 if (ops && (ops->flags & FIO_SIGQUIT))
 104                                         kill(td->pid, SIGQUIT);
 105                         }
 106                 }
 107         }
 108 }
 109
 110 static void status_timer_arm(void)
 111 {
 112         itimer.it_value.tv_sec = 0;
 113         itimer.it_value.tv_usec = DISK_UTIL_MSEC * 1000;
 114         setitimer(ITIMER_REAL, &itimer, NULL);
 115 }
 116
 117 static void sig_alrm(int fio_unused sig)
 118 {
 119         if (threads) {
 120                 update_io_ticks();
 121                 print_thread_status();
 122                 status_timer_arm();
 123         }
 124 }
 125
 126 /*
 127  * Happens on thread runs with ctrl-c, ignore our own SIGQUIT
 128  */
 129 static void sig_quit(int sig)
 130 {
 131 }
 132
 133 static void sig_int(int sig)
 134 {
 135         if (threads) {
 136                 printf("\nfio: terminating on signal %d\n", sig);
 137                 fflush(stdout);
 138                 terminate_threads(TERMINATE_ALL);
 139         }
 140 }
 141
 142 static void sig_ill(int fio_unused sig)
 143 {
 144         if (!threads)
 145                 return;
 146
 147         log_err("fio: illegal instruction. your cpu does not support "
 148                 "the sse4.2 instruction for crc32c\n");
 149         terminate_threads(TERMINATE_ALL);
 150         exit(4);
 151 }
 152
 153 static void set_sig_handlers(void)
 154 {
 155         struct sigaction act;
 156
 157         memset(&act, 0, sizeof(act));
 158         act.sa_handler = sig_alrm;
 159         act.sa_flags = SA_RESTART;
 160         sigaction(SIGALRM, &act, NULL);
 161
 162         memset(&act, 0, sizeof(act));
 163         act.sa_handler = sig_int;
 164         act.sa_flags = SA_RESTART;
 165         sigaction(SIGINT, &act, NULL);
 166
 167         memset(&act, 0, sizeof(act));
 168         act.sa_handler = sig_ill;
 169         act.sa_flags = SA_RESTART;
 170         sigaction(SIGILL, &act, NULL);
 171
 172         memset(&act, 0, sizeof(act));
 173         act.sa_handler = sig_quit;
 174         act.sa_flags = SA_RESTART;
 175         sigaction(SIGQUIT, &act, NULL);
 176 }
 177
 178 /*
 179  * Check if we are above the minimum rate given.
 180  */
 181 static int __check_min_rate(struct thread_data *td, struct timeval *now,
 182                             enum td_ddir ddir)
 183 {
 184         unsigned long long bytes = 0;
 185         unsigned long iops = 0;
 186         unsigned long spent;
 187         unsigned long rate;
 188         unsigned int ratemin = 0;
 189         unsigned int rate_iops = 0;
 190         unsigned int rate_iops_min = 0;
 191
 192         if (!td->o.ratemin[ddir] && !td->o.rate_iops_min[ddir])
 193                 return 0;
 194
 195         /*
 196          * allow a 2 second settle period in the beginning
 197          */
 198         if (mtime_since(&td->start, now) < 2000)
 199                 return 0;
 200
 201         iops += td->io_blocks[ddir];
 202         bytes += td->this_io_bytes[ddir];
 203         ratemin += td->o.ratemin[ddir];
 204         rate_iops += td->o.rate_iops[ddir];
 205         rate_iops_min += td->o.rate_iops_min[ddir];
 206
 207         /*
 208          * if rate blocks is set, sample is running
 209          */
 210         if (td->rate_bytes[ddir] || td->rate_blocks[ddir]) {
 211                 spent = mtime_since(&td->lastrate[ddir], now);
 212                 if (spent < td->o.ratecycle)
 213                         return 0;
 214
 215                 if (td->o.rate[ddir]) {
 216                         /*
 217                          * check bandwidth specified rate
 218                          */
 219                         if (bytes < td->rate_bytes[ddir]) {
 220                                 log_err("%s: min rate %u not met\n", td->o.name,
 221                                                                 ratemin);
 222                                 return 1;
 223                         } else {
 224                                 rate = ((bytes - td->rate_bytes[ddir]) * 1000) / spent;
 225                                 if (rate < ratemin ||
 226                                     bytes < td->rate_bytes[ddir]) {
 227                                         log_err("%s: min rate %u not met, got"
 228                                                 " %luKB/sec\n", td->o.name,
 229                                                         ratemin, rate);
 230                                         return 1;
 231                                 }
 232                         }
 233                 } else {
 234                         /*
 235                          * checks iops specified rate
 236                          */
 237                         if (iops < rate_iops) {
 238                                 log_err("%s: min iops rate %u not met\n",
 239                                                 td->o.name, rate_iops);
 240                                 return 1;
 241                         } else {
 242                                 rate = ((iops - td->rate_blocks[ddir]) * 1000) / spent;
 243                                 if (rate < rate_iops_min ||
 244                                     iops < td->rate_blocks[ddir]) {
 245                                         log_err("%s: min iops rate %u not met,"
 246                                                 " got %lu\n", td->o.name,
 247                                                         rate_iops_min, rate);
 248                                 }
 249                         }
 250                 }
 251         }
 252
 253         td->rate_bytes[ddir] = bytes;
 254         td->rate_blocks[ddir] = iops;
 255         memcpy(&td->lastrate[ddir], now, sizeof(*now));
 256         return 0;
 257 }
 258
 259 static int check_min_rate(struct thread_data *td, struct timeval *now,
 260                           unsigned long *bytes_done)
 261 {
 262         int ret = 0;
 263
 264         if (bytes_done[0])
 265                 ret |= __check_min_rate(td, now, 0);
 266         if (bytes_done[1])
 267                 ret |= __check_min_rate(td, now, 1);
 268
 269         return ret;
 270 }
 271
 272 static inline int runtime_exceeded(struct thread_data *td, struct timeval *t)
 273 {
 274         if (!td->o.timeout)
 275                 return 0;
 276         if (mtime_since(&td->epoch, t) >= td->o.timeout * 1000)
 277                 return 1;
 278
 279         return 0;
 280 }
 281
 282 /*
 283  * When job exits, we can cancel the in-flight IO if we are using async
 284  * io. Attempt to do so.
 285  */
 286 static void cleanup_pending_aio(struct thread_data *td)
 287 {
 288         struct flist_head *entry, *n;
 289         struct io_u *io_u;
 290         int r;
 291
 292         /*
 293          * get immediately available events, if any
 294          */
 295         r = io_u_queued_complete(td, 0, NULL);
 296         if (r < 0)
 297                 return;
 298
 299         /*
 300          * now cancel remaining active events
 301          */
 302         if (td->io_ops->cancel) {
 303                 flist_for_each_safe(entry, n, &td->io_u_busylist) {
 304                         io_u = flist_entry(entry, struct io_u, list);
 305
 306                         /*
 307                          * if the io_u isn't in flight, then that generally
 308                          * means someone leaked an io_u. complain but fix
 309                          * it up, so we don't stall here.
 310                          */
 311                         if ((io_u->flags & IO_U_F_FLIGHT) == 0) {
 312                                 log_err("fio: non-busy IO on busy list\n");
 313                                 put_io_u(td, io_u);
 314                         } else {
 315                                 r = td->io_ops->cancel(td, io_u);
 316                                 if (!r)
 317                                         put_io_u(td, io_u);
 318                         }
 319                 }
 320         }
 321
 322         if (td->cur_depth)
 323                 r = io_u_queued_complete(td, td->cur_depth, NULL);
 324 }
 325
 326 /*
 327  * Helper to handle the final sync of a file. Works just like the normal
 328  * io path, just does everything sync.
 329  */
 330 static int fio_io_sync(struct thread_data *td, struct fio_file *f)
 331 {
 332         struct io_u *io_u = __get_io_u(td);
 333         int ret;
 334
 335         if (!io_u)
 336                 return 1;
 337
 338         io_u->ddir = DDIR_SYNC;
 339         io_u->file = f;
 340
 341         if (td_io_prep(td, io_u)) {
 342                 put_io_u(td, io_u);
 343                 return 1;
 344         }
 345
 346 requeue:
 347         ret = td_io_queue(td, io_u);
 348         if (ret < 0) {
 349                 td_verror(td, io_u->error, "td_io_queue");
 350                 put_io_u(td, io_u);
 351                 return 1;
 352         } else if (ret == FIO_Q_QUEUED) {
 353                 if (io_u_queued_complete(td, 1, NULL) < 0)
 354                         return 1;
 355         } else if (ret == FIO_Q_COMPLETED) {
 356                 if (io_u->error) {
 357                         td_verror(td, io_u->error, "td_io_queue");
 358                         return 1;
 359                 }
 360
 361                 if (io_u_sync_complete(td, io_u, NULL) < 0)
 362                         return 1;
 363         } else if (ret == FIO_Q_BUSY) {
 364                 if (td_io_commit(td))
 365                         return 1;
 366                 goto requeue;
 367         }
 368
 369         return 0;
 370 }
 371
 372 static inline void update_tv_cache(struct thread_data *td)
 373 {
 374         if ((++td->tv_cache_nr & td->tv_cache_mask) == td->tv_cache_mask)
 375                 fio_gettime(&td->tv_cache, NULL);
 376 }
 377
 378 static int break_on_this_error(struct thread_data *td, int *retptr)
 379 {
 380         int ret = *retptr;
 381
 382         if (ret < 0 || td->error) {
 383                 int err;
 384
 385                 if (!td->o.continue_on_error)
 386                         return 1;
 387
 388                 if (ret < 0)
 389                         err = -ret;
 390                 else
 391                         err = td->error;
 392
 393                 if (td_non_fatal_error(err)) {
 394                         /*
 395                          * Continue with the I/Os in case of
 396                          * a non fatal error.
 397                          */
 398                         update_error_count(td, err);
 399                         td_clear_error(td);
 400                         *retptr = 0;
 401                         return 0;
 402                 } else if (td->o.fill_device && err == ENOSPC) {
 403                         /*
 404                          * We expect to hit this error if
 405                          * fill_device option is set.
 406                          */
 407                         td_clear_error(td);
 408                         td->terminate = 1;
 409                         return 1;
 410                 } else {
 411                         /*
 412                          * Stop the I/O in case of a fatal
 413                          * error.
 414                          */
 415                         update_error_count(td, err);
 416                         return 1;
 417                 }
 418         }
 419
 420         return 0;
 421 }
 422
 423 /*
 424  * The main verify engine. Runs over the writes we previously submitted,
 425  * reads the blocks back in, and checks the crc/md5 of the data.
 426  */
 427 static void do_verify(struct thread_data *td)
 428 {
 429         struct fio_file *f;
 430         struct io_u *io_u;
 431         int ret, min_events;
 432         unsigned int i;
 433
 434         /*
 435          * sync io first and invalidate cache, to make sure we really
 436          * read from disk.
 437          */
 438         for_each_file(td, f, i) {
 439                 if (!fio_file_open(f))
 440                         continue;
 441                 if (fio_io_sync(td, f))
 442                         break;
 443                 if (file_invalidate_cache(td, f))
 444                         break;
 445         }
 446
 447         if (td->error)
 448                 return;
 449
 450         td_set_runstate(td, TD_VERIFYING);
 451
 452         io_u = NULL;
 453         while (!td->terminate) {
 454                 int ret2, full;
 455
 456                 update_tv_cache(td);
 457
 458                 if (runtime_exceeded(td, &td->tv_cache)) {
 459                         td->terminate = 1;
 460                         break;
 461                 }
 462
 463                 io_u = __get_io_u(td);
 464                 if (!io_u)
 465                         break;
 466
 467                 if (get_next_verify(td, io_u)) {
 468                         put_io_u(td, io_u);
 469                         break;
 470                 }
 471
 472                 if (td_io_prep(td, io_u)) {
 473                         put_io_u(td, io_u);
 474                         break;
 475                 }
 476
 477                 if (td->o.verify_async)
 478                         io_u->end_io = verify_io_u_async;
 479                 else
 480                         io_u->end_io = verify_io_u;
 481
 482                 ret = td_io_queue(td, io_u);
 483                 switch (ret) {
 484                 case FIO_Q_COMPLETED:
 485                         if (io_u->error) {
 486                                 ret = -io_u->error;
 487                                 clear_io_u(td, io_u);
 488                         } else if (io_u->resid) {
 489                                 int bytes = io_u->xfer_buflen - io_u->resid;
 490                                 struct fio_file *f = io_u->file;
 491
 492                                 /*
 493                                  * zero read, fail
 494                                  */
 495                                 if (!bytes) {
 496                                         td_verror(td, EIO, "full resid");
 497                                         put_io_u(td, io_u);
 498                                         break;
 499                                 }
 500
 501                                 io_u->xfer_buflen = io_u->resid;
 502                                 io_u->xfer_buf += bytes;
 503                                 io_u->offset += bytes;
 504
 505                                 td->ts.short_io_u[io_u->ddir]++;
 506
 507                                 if (io_u->offset == f->real_file_size)
 508                                         goto sync_done;
 509
 510                                 requeue_io_u(td, &io_u);
 511                         } else {
 512 sync_done:
 513                                 ret = io_u_sync_complete(td, io_u, NULL);
 514                                 if (ret < 0)
 515                                         break;
 516                         }
 517                         continue;
 518                 case FIO_Q_QUEUED:
 519                         break;
 520                 case FIO_Q_BUSY:
 521                         requeue_io_u(td, &io_u);
 522                         ret2 = td_io_commit(td);
 523                         if (ret2 < 0)
 524                                 ret = ret2;
 525                         break;
 526                 default:
 527                         assert(ret < 0);
 528                         td_verror(td, -ret, "td_io_queue");
 529                         break;
 530                 }
 531
 532                 if (break_on_this_error(td, &ret))
 533                         break;
 534
 535                 /*
 536                  * if we can queue more, do so. but check if there are
 537                  * completed io_u's first.
 538                  */
 539                 full = queue_full(td) || ret == FIO_Q_BUSY;
 540                 if (full || !td->o.iodepth_batch_complete) {
 541                         min_events = min(td->o.iodepth_batch_complete,
 542                                          td->cur_depth);
 543                         if (full && !min_events)
 544                                 min_events = 1;
 545
 546                         do {
 547                                 /*
 548                                  * Reap required number of io units, if any,
 549                                  * and do the verification on them through
 550                                  * the callback handler
 551                                  */
 552                                 if (io_u_queued_complete(td, min_events, NULL) < 0) {
 553                                         ret = -1;
 554                                         break;
 555                                 }
 556                         } while (full && (td->cur_depth > td->o.iodepth_low));
 557                 }
 558                 if (ret < 0)
 559                         break;
 560         }
 561
 562         if (!td->error) {
 563                 min_events = td->cur_depth;
 564
 565                 if (min_events)
 566                         ret = io_u_queued_complete(td, min_events, NULL);
 567         } else
 568                 cleanup_pending_aio(td);
 569
 570         td_set_runstate(td, TD_RUNNING);
 571 }
 572
 573 /*
 574  * Main IO worker function. It retrieves io_u's to process and queues
 575  * and reaps them, checking for rate and errors along the way.
 576  */
 577 static void do_io(struct thread_data *td)
 578 {
 579         unsigned int i;
 580         int ret = 0;
 581
 582         if (in_ramp_time(td))
 583                 td_set_runstate(td, TD_RAMP);
 584         else
 585                 td_set_runstate(td, TD_RUNNING);
 586
 587         while ( (td->o.read_iolog_file && !flist_empty(&td->io_log_list)) ||
 588                 ((td->this_io_bytes[0] + td->this_io_bytes[1]) < td->o.size) ) {
 589                 struct timeval comp_time;
 590                 unsigned long bytes_done[2] = { 0, 0 };
 591                 int min_evts = 0;
 592                 struct io_u *io_u;
 593                 int ret2, full;
 594
 595                 if (td->terminate)
 596                         break;
 597
 598                 update_tv_cache(td);
 599
 600                 if (runtime_exceeded(td, &td->tv_cache)) {
 601                         td->terminate = 1;
 602                         break;
 603                 }
 604
 605                 io_u = get_io_u(td);
 606                 if (!io_u)
 607                         break;
 608
 609                 /*
 610                  * Add verification end_io handler, if asked to verify
 611                  * a previously written file.
 612                  */
 613                 if (td->o.verify != VERIFY_NONE && io_u->ddir == DDIR_READ &&
 614                     !td_rw(td)) {
 615                         if (td->o.verify_async)
 616                                 io_u->end_io = verify_io_u_async;
 617                         else
 618                                 io_u->end_io = verify_io_u;
 619                         td_set_runstate(td, TD_VERIFYING);
 620                 } else if (in_ramp_time(td))
 621                         td_set_runstate(td, TD_RAMP);
 622                 else
 623                         td_set_runstate(td, TD_RUNNING);
 624
 625                 ret = td_io_queue(td, io_u);
 626                 switch (ret) {
 627                 case FIO_Q_COMPLETED:
 628                         if (io_u->error) {
 629                                 ret = -io_u->error;
 630                                 clear_io_u(td, io_u);
 631                         } else if (io_u->resid) {
 632                                 int bytes = io_u->xfer_buflen - io_u->resid;
 633                                 struct fio_file *f = io_u->file;
 634
 635                                 /*
 636                                  * zero read, fail
 637                                  */
 638                                 if (!bytes) {
 639                                         td_verror(td, EIO, "full resid");
 640                                         put_io_u(td, io_u);
 641                                         break;
 642                                 }
 643
 644                                 io_u->xfer_buflen = io_u->resid;
 645                                 io_u->xfer_buf += bytes;
 646                                 io_u->offset += bytes;
 647
 648                                 td->ts.short_io_u[io_u->ddir]++;
 649
 650                                 if (io_u->offset == f->real_file_size)
 651                                         goto sync_done;
 652
 653                                 requeue_io_u(td, &io_u);
 654                         } else {
 655 sync_done:
 656                                 if (__should_check_rate(td, 0) ||
 657                                     __should_check_rate(td, 1))
 658                                         fio_gettime(&comp_time, NULL);
 659
 660                                 ret = io_u_sync_complete(td, io_u, bytes_done);
 661                                 if (ret < 0)
 662                                         break;
 663                         }
 664                         break;
 665                 case FIO_Q_QUEUED:
 666                         /*
 667                          * if the engine doesn't have a commit hook,
 668                          * the io_u is really queued. if it does have such
 669                          * a hook, it has to call io_u_queued() itself.
 670                          */
 671                         if (td->io_ops->commit == NULL)
 672                                 io_u_queued(td, io_u);
 673                         break;
 674                 case FIO_Q_BUSY:
 675                         requeue_io_u(td, &io_u);
 676                         ret2 = td_io_commit(td);
 677                         if (ret2 < 0)
 678                                 ret = ret2;
 679                         break;
 680                 default:
 681                         assert(ret < 0);
 682                         put_io_u(td, io_u);
 683                         break;
 684                 }
 685
 686                 if (break_on_this_error(td, &ret))
 687                         break;
 688
 689                 /*
 690                  * See if we need to complete some commands
 691                  */
 692                 full = queue_full(td) || ret == FIO_Q_BUSY;
 693                 if (full || !td->o.iodepth_batch_complete) {
 694                         min_evts = min(td->o.iodepth_batch_complete,
 695                                         td->cur_depth);
 696                         if (full && !min_evts)
 697                                 min_evts = 1;
 698
 699                         if (__should_check_rate(td, 0) ||
 700                             __should_check_rate(td, 1))
 701                                 fio_gettime(&comp_time, NULL);
 702
 703                         do {
 704                                 ret = io_u_queued_complete(td, min_evts, bytes_done);
 705                                 if (ret < 0)
 706                                         break;
 707
 708                         } while (full && (td->cur_depth > td->o.iodepth_low));
 709                 }
 710
 711                 if (ret < 0)
 712                         break;
 713                 if (!(bytes_done[0] + bytes_done[1]))
 714                         continue;
 715
 716                 if (!in_ramp_time(td) && should_check_rate(td, bytes_done)) {
 717                         if (check_min_rate(td, &comp_time, bytes_done)) {
 718                                 if (exitall_on_terminate)
 719                                         terminate_threads(td->groupid);
 720                                 td_verror(td, EIO, "check_min_rate");
 721                                 break;
 722                         }
 723                 }
 724
 725                 if (td->o.thinktime) {
 726                         unsigned long long b;
 727
 728                         b = td->io_blocks[0] + td->io_blocks[1];
 729                         if (!(b % td->o.thinktime_blocks)) {
 730                                 int left;
 731
 732                                 if (td->o.thinktime_spin)
 733                                         usec_spin(td->o.thinktime_spin);
 734
 735                                 left = td->o.thinktime - td->o.thinktime_spin;
 736                                 if (left)
 737                                         usec_sleep(td, left);
 738                         }
 739                 }
 740         }
 741
 742         if (td->o.fill_device && td->error == ENOSPC) {
 743                 td->error = 0;
 744                 td->terminate = 1;
 745         }
 746         if (!td->error) {
 747                 struct fio_file *f;
 748
 749                 i = td->cur_depth;
 750                 if (i)
 751                         ret = io_u_queued_complete(td, i, NULL);
 752
 753                 if (should_fsync(td) && td->o.end_fsync) {
 754                         td_set_runstate(td, TD_FSYNCING);
 755
 756                         for_each_file(td, f, i) {
 757                                 if (!fio_file_open(f))
 758                                         continue;
 759                                 fio_io_sync(td, f);
 760                         }
 761                 }
 762         } else
 763                 cleanup_pending_aio(td);
 764
 765         /*
 766          * stop job if we failed doing any IO
 767          */
 768         if ((td->this_io_bytes[0] + td->this_io_bytes[1]) == 0)
 769                 td->done = 1;
 770 }
 771
 772 static void cleanup_io_u(struct thread_data *td)
 773 {
 774         struct flist_head *entry, *n;
 775         struct io_u *io_u;
 776
 777         flist_for_each_safe(entry, n, &td->io_u_freelist) {
 778                 io_u = flist_entry(entry, struct io_u, list);
 779
 780                 flist_del(&io_u->list);
 781                 free(io_u);
 782         }
 783
 784         free_io_mem(td);
 785 }
 786
 787 static int init_io_u(struct thread_data *td)
 788 {
 789         struct io_u *io_u;
 790         unsigned int max_bs;
 791         int cl_align, i, max_units;
 792         char *p;
 793
 794         max_units = td->o.iodepth;
 795         max_bs = max(td->o.max_bs[DDIR_READ], td->o.max_bs[DDIR_WRITE]);
 796         td->orig_buffer_size = (unsigned long long) max_bs
 797                                         * (unsigned long long) max_units;
 798
 799         if (td->o.mem_type == MEM_SHMHUGE || td->o.mem_type == MEM_MMAPHUGE) {
 800                 unsigned long bs;
 801
 802                 bs = td->orig_buffer_size + td->o.hugepage_size - 1;
 803                 td->orig_buffer_size = bs & ~(td->o.hugepage_size - 1);
 804         }
 805
 806         if (td->orig_buffer_size != (size_t) td->orig_buffer_size) {
 807                 log_err("fio: IO memory too large. Reduce max_bs or iodepth\n");
 808                 return 1;
 809         }
 810
 811         if (allocate_io_mem(td))
 812                 return 1;
 813
 814         if (td->o.odirect || td->o.mem_align)
 815                 p = PAGE_ALIGN(td->orig_buffer) + td->o.mem_align;
 816         else
 817                 p = td->orig_buffer;
 818
 819         cl_align = os_cache_line_size();
 820
 821         for (i = 0; i < max_units; i++) {
 822                 void *ptr;
 823
 824                 if (td->terminate)
 825                         return 1;
 826
 827                 if (posix_memalign(&ptr, cl_align, sizeof(*io_u))) {
 828                         log_err("fio: posix_memalign=%s\n", strerror(errno));
 829                         break;
 830                 }
 831
 832                 io_u = ptr;
 833                 memset(io_u, 0, sizeof(*io_u));
 834                 INIT_FLIST_HEAD(&io_u->list);
 835                 dprint(FD_MEM, "io_u alloc %p, index %u\n", io_u, i);
 836
 837                 if (!(td->io_ops->flags & FIO_NOIO)) {
 838                         io_u->buf = p + max_bs * i;
 839                         dprint(FD_MEM, "io_u %p, mem %p\n", io_u, io_u->buf);
 840
 841                         if (td_write(td) && !td->o.refill_buffers)
 842                                 io_u_fill_buffer(td, io_u, max_bs);
 843                 }
 844
 845                 io_u->index = i;
 846                 io_u->flags = IO_U_F_FREE;
 847                 flist_add(&io_u->list, &td->io_u_freelist);
 848         }
 849
 850         return 0;
 851 }
 852
 853 static int switch_ioscheduler(struct thread_data *td)
 854 {
 855         char tmp[256], tmp2[128];
 856         FILE *f;
 857         int ret;
 858
 859         if (td->io_ops->flags & FIO_DISKLESSIO)
 860                 return 0;
 861
 862         sprintf(tmp, "%s/queue/scheduler", td->sysfs_root);
 863
 864         f = fopen(tmp, "r+");
 865         if (!f) {
 866                 if (errno == ENOENT) {
 867                         log_err("fio: os or kernel doesn't support IO scheduler"
 868                                 " switching\n");
 869                         return 0;
 870                 }
 871                 td_verror(td, errno, "fopen iosched");
 872                 return 1;
 873         }
 874
 875         /*
 876          * Set io scheduler.
 877          */
 878         ret = fwrite(td->o.ioscheduler, strlen(td->o.ioscheduler), 1, f);
 879         if (ferror(f) || ret != 1) {
 880                 td_verror(td, errno, "fwrite");
 881                 fclose(f);
 882                 return 1;
 883         }
 884
 885         rewind(f);
 886
 887         /*
 888          * Read back and check that the selected scheduler is now the default.
 889          */
 890         ret = fread(tmp, 1, sizeof(tmp), f);
 891         if (ferror(f) || ret < 0) {
 892                 td_verror(td, errno, "fread");
 893                 fclose(f);
 894                 return 1;
 895         }
 896
 897         sprintf(tmp2, "[%s]", td->o.ioscheduler);
 898         if (!strstr(tmp, tmp2)) {
 899                 log_err("fio: io scheduler %s not found\n", td->o.ioscheduler);
 900                 td_verror(td, EINVAL, "iosched_switch");
 901                 fclose(f);
 902                 return 1;
 903         }
 904
 905         fclose(f);
 906         return 0;
 907 }
 908
 909 static int keep_running(struct thread_data *td)
 910 {
 911         unsigned long long io_done;
 912
 913         if (td->done)
 914                 return 0;
 915         if (td->o.time_based)
 916                 return 1;
 917         if (td->o.loops) {
 918                 td->o.loops--;
 919                 return 1;
 920         }
 921
 922         io_done = td->io_bytes[DDIR_READ] + td->io_bytes[DDIR_WRITE]
 923                         + td->io_skip_bytes;
 924         if (io_done < td->o.size)
 925                 return 1;
 926
 927         return 0;
 928 }
 929
 930 static void reset_io_counters(struct thread_data *td)
 931 {
 932         td->ts.stat_io_bytes[0] = td->ts.stat_io_bytes[1] = 0;
 933         td->this_io_bytes[0] = td->this_io_bytes[1] = 0;
 934         td->zone_bytes = 0;
 935         td->rate_bytes[0] = td->rate_bytes[1] = 0;
 936         td->rate_blocks[0] = td->rate_blocks[1] = 0;
 937
 938         td->last_was_sync = 0;
 939
 940         /*
 941          * reset file done count if we are to start over
 942          */
 943         if (td->o.time_based || td->o.loops)
 944                 td->nr_done_files = 0;
 945
 946         /*
 947          * Set the same seed to get repeatable runs
 948          */
 949         td_fill_rand_seeds(td);
 950 }
 951
 952 void reset_all_stats(struct thread_data *td)
 953 {
 954         struct timeval tv;
 955         int i;
 956
 957         reset_io_counters(td);
 958
 959         for (i = 0; i < 2; i++) {
 960                 td->io_bytes[i] = 0;
 961                 td->io_blocks[i] = 0;
 962                 td->io_issues[i] = 0;
 963                 td->ts.total_io_u[i] = 0;
 964         }
 965
 966         fio_gettime(&tv, NULL);
 967         memcpy(&td->epoch, &tv, sizeof(tv));
 968         memcpy(&td->start, &tv, sizeof(tv));
 969 }
 970
 971 static void clear_io_state(struct thread_data *td)
 972 {
 973         struct fio_file *f;
 974         unsigned int i;
 975
 976         reset_io_counters(td);
 977
 978         close_files(td);
 979         for_each_file(td, f, i)
 980                 fio_file_clear_done(f);
 981 }
 982
 983 static int exec_string(const char *string)
 984 {
 985         int ret, newlen = strlen(string) + 1 + 8;
 986         char *str;
 987
 988         str = malloc(newlen);
 989         sprintf(str, "sh -c %s", string);
 990
 991         ret = system(str);
 992         if (ret == -1)
 993                 log_err("fio: exec of cmd <%s> failed\n", str);
 994
 995         free(str);
 996         return ret;
 997 }
 998
 999 /*
1000  * Entry point for the thread based jobs. The process based jobs end up
1001  * here as well, after a little setup.
1002  */
1003 static void *thread_main(void *data)
1004 {
1005         unsigned long long runtime[2], elapsed;
1006         struct thread_data *td = data;
1007         pthread_condattr_t attr;
1008         int clear_state;
1009
1010         if (!td->o.use_thread)
1011                 setsid();
1012
1013         td->pid = getpid();
1014
1015         dprint(FD_PROCESS, "jobs pid=%d started\n", (int) td->pid);
1016
1017         INIT_FLIST_HEAD(&td->io_u_freelist);
1018         INIT_FLIST_HEAD(&td->io_u_busylist);
1019         INIT_FLIST_HEAD(&td->io_u_requeues);
1020         INIT_FLIST_HEAD(&td->io_log_list);
1021         INIT_FLIST_HEAD(&td->io_hist_list);
1022         INIT_FLIST_HEAD(&td->verify_list);
1023         pthread_mutex_init(&td->io_u_lock, NULL);
1024         td->io_hist_tree = RB_ROOT;
1025
1026         pthread_condattr_init(&attr);
1027         pthread_cond_init(&td->verify_cond, &attr);
1028         pthread_cond_init(&td->free_cond, &attr);
1029
1030         td_set_runstate(td, TD_INITIALIZED);
1031         dprint(FD_MUTEX, "up startup_mutex\n");
1032         fio_mutex_up(startup_mutex);
1033         dprint(FD_MUTEX, "wait on td->mutex\n");
1034         fio_mutex_down(td->mutex);
1035         dprint(FD_MUTEX, "done waiting on td->mutex\n");
1036
1037         /*
1038          * the ->mutex mutex is now no longer used, close it to avoid
1039          * eating a file descriptor
1040          */
1041         fio_mutex_remove(td->mutex);
1042
1043         /*
1044          * May alter parameters that init_io_u() will use, so we need to
1045          * do this first.
1046          */
1047         if (init_iolog(td))
1048                 goto err;
1049
1050         if (init_io_u(td))
1051                 goto err;
1052
1053         if (td->o.verify_async && verify_async_init(td))
1054                 goto err;
1055
1056         if (td->o.cpumask_set && fio_setaffinity(td->pid, td->o.cpumask) == -1) {
1057                 td_verror(td, errno, "cpu_set_affinity");
1058                 goto err;
1059         }
1060
1061         /*
1062          * If we have a gettimeofday() thread, make sure we exclude that
1063          * thread from this job
1064          */
1065         if (td->o.gtod_cpu) {
1066                 fio_cpu_clear(&td->o.cpumask, td->o.gtod_cpu);
1067                 if (fio_setaffinity(td->pid, td->o.cpumask) == -1) {
1068                         td_verror(td, errno, "cpu_set_affinity");
1069                         goto err;
1070                 }
1071         }
1072
1073         if (td->ioprio_set) {
1074                 if (ioprio_set(IOPRIO_WHO_PROCESS, 0, td->ioprio) == -1) {
1075                         td_verror(td, errno, "ioprio_set");
1076                         goto err;
1077                 }
1078         }
1079
1080         if (td->o.cgroup_weight && cgroup_setup(td, cgroup_list))
1081                 goto err;
1082
1083         if (nice(td->o.nice) == -1) {
1084                 td_verror(td, errno, "nice");
1085                 goto err;
1086         }
1087
1088         if (td->o.ioscheduler && switch_ioscheduler(td))
1089                 goto err;
1090
1091         if (!td->o.create_serialize && setup_files(td))
1092                 goto err;
1093
1094         if (td_io_init(td))
1095                 goto err;
1096
1097         if (init_random_map(td))
1098                 goto err;
1099
1100         if (td->o.exec_prerun) {
1101                 if (exec_string(td->o.exec_prerun))
1102                         goto err;
1103         }
1104
1105         if (td->o.pre_read) {
1106                 if (pre_read_files(td) < 0)
1107                         goto err;
1108         }
1109
1110         fio_gettime(&td->epoch, NULL);
1111         getrusage(RUSAGE_SELF, &td->ts.ru_start);
1112
1113         runtime[0] = runtime[1] = 0;
1114         clear_state = 0;
1115         while (keep_running(td)) {
1116                 fio_gettime(&td->start, NULL);
1117                 memcpy(&td->ts.stat_sample_time[0], &td->start,
1118                                 sizeof(td->start));
1119                 memcpy(&td->ts.stat_sample_time[1], &td->start,
1120                                 sizeof(td->start));
1121                 memcpy(&td->tv_cache, &td->start, sizeof(td->start));
1122
1123                 if (td->o.ratemin[0] || td->o.ratemin[1])
1124                         memcpy(&td->lastrate, &td->ts.stat_sample_time,
1125                                                         sizeof(td->lastrate));
1126
1127                 if (clear_state)
1128                         clear_io_state(td);
1129
1130                 prune_io_piece_log(td);
1131
1132                 do_io(td);
1133
1134                 clear_state = 1;
1135
1136                 if (td_read(td) && td->io_bytes[DDIR_READ]) {
1137                         elapsed = utime_since_now(&td->start);
1138                         runtime[DDIR_READ] += elapsed;
1139                 }
1140                 if (td_write(td) && td->io_bytes[DDIR_WRITE]) {
1141                         elapsed = utime_since_now(&td->start);
1142                         runtime[DDIR_WRITE] += elapsed;
1143                 }
1144
1145                 if (td->error || td->terminate)
1146                         break;
1147
1148                 if (!td->o.do_verify ||
1149                     td->o.verify == VERIFY_NONE ||
1150                     (td->io_ops->flags & FIO_UNIDIR))
1151                         continue;
1152
1153                 clear_io_state(td);
1154
1155                 fio_gettime(&td->start, NULL);
1156
1157                 do_verify(td);
1158
1159                 runtime[DDIR_READ] += utime_since_now(&td->start);
1160
1161                 if (td->error || td->terminate)
1162                         break;
1163         }
1164
1165         update_rusage_stat(td);
1166         td->ts.runtime[0] = (runtime[0] + 999) / 1000;
1167         td->ts.runtime[1] = (runtime[1] + 999) / 1000;
1168         td->ts.total_run_time = mtime_since_now(&td->epoch);
1169         td->ts.io_bytes[0] = td->io_bytes[0];
1170         td->ts.io_bytes[1] = td->io_bytes[1];
1171
1172         fio_mutex_down(writeout_mutex);
1173         if (td->ts.bw_log) {
1174                 if (td->o.bw_log_file) {
1175                         finish_log_named(td, td->ts.bw_log,
1176                                                 td->o.bw_log_file, "bw");
1177                 } else
1178                         finish_log(td, td->ts.bw_log, "bw");
1179         }
1180         if (td->ts.slat_log) {
1181                 if (td->o.lat_log_file) {
1182                         finish_log_named(td, td->ts.slat_log,
1183                                                 td->o.lat_log_file, "slat");
1184                 } else
1185                         finish_log(td, td->ts.slat_log, "slat");
1186         }
1187         if (td->ts.clat_log) {
1188                 if (td->o.lat_log_file) {
1189                         finish_log_named(td, td->ts.clat_log,
1190                                                 td->o.lat_log_file, "clat");
1191                 } else
1192                         finish_log(td, td->ts.clat_log, "clat");
1193         }
1194         fio_mutex_up(writeout_mutex);
1195         if (td->o.exec_postrun)
1196                 exec_string(td->o.exec_postrun);
1197
1198         if (exitall_on_terminate)
1199                 terminate_threads(td->groupid);
1200
1201 err:
1202         if (td->error)
1203                 printf("fio: pid=%d, err=%d/%s\n", (int) td->pid, td->error,
1204                                                         td->verror);
1205
1206         if (td->o.verify_async)
1207                 verify_async_exit(td);
1208
1209         close_and_free_files(td);
1210         close_ioengine(td);
1211         cleanup_io_u(td);
1212         cgroup_shutdown(td);
1213
1214         if (td->o.cpumask_set) {
1215                 int ret = fio_cpuset_exit(&td->o.cpumask);
1216
1217                 td_verror(td, ret, "fio_cpuset_exit");
1218         }
1219
1220         /*
1221          * do this very late, it will log file closing as well
1222          */
1223         if (td->o.write_iolog_file)
1224                 write_iolog_close(td);
1225
1226         options_mem_free(td);
1227         td_set_runstate(td, TD_EXITED);
1228         return (void *) (unsigned long) td->error;
1229 }
1230
1231 /*
1232  * We cannot pass the td data into a forked process, so attach the td and
1233  * pass it to the thread worker.
1234  */
1235 static int fork_main(int shmid, int offset)
1236 {
1237         struct thread_data *td;
1238         void *data, *ret;
1239
1240         data = shmat(shmid, NULL, 0);
1241         if (data == (void *) -1) {
1242                 int __err = errno;
1243
1244                 perror("shmat");
1245                 return __err;
1246         }
1247
1248         td = data + offset * sizeof(struct thread_data);
1249         ret = thread_main(td);
1250         shmdt(data);
1251         return (int) (unsigned long) ret;
1252 }
1253
1254 /*
1255  * Run over the job map and reap the threads that have exited, if any.
1256  */
1257 static void reap_threads(int *nr_running, int *t_rate, int *m_rate)
1258 {
1259         struct thread_data *td;
1260         int i, cputhreads, realthreads, pending, status, ret;
1261
1262         /*
1263          * reap exited threads (TD_EXITED -> TD_REAPED)
1264          */
1265         realthreads = pending = cputhreads = 0;
1266         for_each_td(td, i) {
1267                 int flags = 0;
1268
1269                 /*
1270                  * ->io_ops is NULL for a thread that has closed its
1271                  * io engine
1272                  */
1273                 if (td->io_ops && !strcmp(td->io_ops->name, "cpuio"))
1274                         cputhreads++;
1275                 else
1276                         realthreads++;
1277
1278                 if (!td->pid) {
1279                         pending++;
1280                         continue;
1281                 }
1282                 if (td->runstate == TD_REAPED)
1283                         continue;
1284                 if (td->o.use_thread) {
1285                         if (td->runstate == TD_EXITED) {
1286                                 td_set_runstate(td, TD_REAPED);
1287                                 goto reaped;
1288                         }
1289                         continue;
1290                 }
1291
1292                 flags = WNOHANG;
1293                 if (td->runstate == TD_EXITED)
1294                         flags = 0;
1295
1296                 /*
1297                  * check if someone quit or got killed in an unusual way
1298                  */
1299                 ret = waitpid(td->pid, &status, flags);
1300                 if (ret < 0) {
1301                         if (errno == ECHILD) {
1302                                 log_err("fio: pid=%d disappeared %d\n",
1303                                                 (int) td->pid, td->runstate);
1304                                 td_set_runstate(td, TD_REAPED);
1305                                 goto reaped;
1306                         }
1307                         perror("waitpid");
1308                 } else if (ret == td->pid) {
1309                         if (WIFSIGNALED(status)) {
1310                                 int sig = WTERMSIG(status);
1311
1312                                 if (sig != SIGQUIT)
1313                                         log_err("fio: pid=%d, got signal=%d\n",
1314                                                         (int) td->pid, sig);
1315                                 td_set_runstate(td, TD_REAPED);
1316                                 goto reaped;
1317                         }
1318                         if (WIFEXITED(status)) {
1319                                 if (WEXITSTATUS(status) && !td->error)
1320                                         td->error = WEXITSTATUS(status);
1321
1322                                 td_set_runstate(td, TD_REAPED);
1323                                 goto reaped;
1324                         }
1325                 }
1326
1327                 /*
1328                  * thread is not dead, continue
1329                  */
1330                 pending++;
1331                 continue;
1332 reaped:
1333                 (*nr_running)--;
1334                 (*m_rate) -= (td->o.ratemin[0] + td->o.ratemin[1]);
1335                 (*t_rate) -= (td->o.rate[0] + td->o.rate[1]);
1336                 if (!td->pid)
1337                         pending--;
1338
1339                 if (td->error)
1340                         exit_value++;
1341
1342                 done_secs += mtime_since_now(&td->epoch) / 1000;
1343         }
1344
1345         if (*nr_running == cputhreads && !pending && realthreads)
1346                 terminate_threads(TERMINATE_ALL);
1347 }
1348
1349 static void *gtod_thread_main(void *data)
1350 {
1351         fio_mutex_up(startup_mutex);
1352
1353         /*
1354          * As long as we have jobs around, update the clock. It would be nice
1355          * to have some way of NOT hammering that CPU with gettimeofday(),
1356          * but I'm not sure what to use outside of a simple CPU nop to relax
1357          * it - we don't want to lose precision.
1358          */
1359         while (threads) {
1360                 fio_gtod_update();
1361                 nop;
1362         }
1363
1364         return NULL;
1365 }
1366
1367 static int fio_start_gtod_thread(void)
1368 {
1369         int ret;
1370
1371         ret = pthread_create(&gtod_thread, NULL, gtod_thread_main, NULL);
1372         if (ret) {
1373                 log_err("Can't create gtod thread: %s\n", strerror(ret));
1374                 return 1;
1375         }
1376
1377         ret = pthread_detach(gtod_thread);
1378         if (ret) {
1379                 log_err("Can't detatch gtod thread: %s\n", strerror(ret));
1380                 return 1;
1381         }
1382
1383         dprint(FD_MUTEX, "wait on startup_mutex\n");
1384         fio_mutex_down(startup_mutex);
1385         dprint(FD_MUTEX, "done waiting on startup_mutex\n");
1386         return 0;
1387 }
1388
1389 /*
1390  * Main function for kicking off and reaping jobs, as needed.
1391  */
1392 static void run_threads(void)
1393 {
1394         struct thread_data *td;
1395         unsigned long spent;
1396         int i, todo, nr_running, m_rate, t_rate, nr_started;
1397
1398         if (fio_pin_memory())
1399                 return;
1400
1401         if (fio_gtod_offload && fio_start_gtod_thread())
1402                 return;
1403
1404         if (!terse_output) {
1405                 printf("Starting ");
1406                 if (nr_thread)
1407                         printf("%d thread%s", nr_thread,
1408                                                 nr_thread > 1 ? "s" : "");
1409                 if (nr_process) {
1410                         if (nr_thread)
1411                                 printf(" and ");
1412                         printf("%d process%s", nr_process,
1413                                                 nr_process > 1 ? "es" : "");
1414                 }
1415                 printf("\n");
1416                 fflush(stdout);
1417         }
1418
1419         set_sig_handlers();
1420
1421         todo = thread_number;
1422         nr_running = 0;
1423         nr_started = 0;
1424         m_rate = t_rate = 0;
1425
1426         for_each_td(td, i) {
1427                 print_status_init(td->thread_number - 1);
1428
1429                 if (!td->o.create_serialize) {
1430                         init_disk_util(td);
1431                         continue;
1432                 }
1433
1434                 /*
1435                  * do file setup here so it happens sequentially,
1436                  * we don't want X number of threads getting their
1437                  * client data interspersed on disk
1438                  */
1439                 if (setup_files(td)) {
1440                         exit_value++;
1441                         if (td->error)
1442                                 log_err("fio: pid=%d, err=%d/%s\n",
1443                                         (int) td->pid, td->error, td->verror);
1444                         td_set_runstate(td, TD_REAPED);
1445                         todo--;
1446                 } else {
1447                         struct fio_file *f;
1448                         unsigned int i;
1449
1450                         /*
1451                          * for sharing to work, each job must always open
1452                          * its own files. so close them, if we opened them
1453                          * for creation
1454                          */
1455                         for_each_file(td, f, i) {
1456                                 if (fio_file_open(f))
1457                                         td_io_close_file(td, f);
1458                         }
1459                 }
1460
1461                 init_disk_util(td);
1462         }
1463
1464         set_genesis_time();
1465
1466         while (todo) {
1467                 struct thread_data *map[MAX_JOBS];
1468                 struct timeval this_start;
1469                 int this_jobs = 0, left;
1470
1471                 /*
1472                  * create threads (TD_NOT_CREATED -> TD_CREATED)
1473                  */
1474                 for_each_td(td, i) {
1475                         if (td->runstate != TD_NOT_CREATED)
1476                                 continue;
1477
1478                         /*
1479                          * never got a chance to start, killed by other
1480                          * thread for some reason
1481                          */
1482                         if (td->terminate) {
1483                                 todo--;
1484                                 continue;
1485                         }
1486
1487                         if (td->o.start_delay) {
1488                                 spent = mtime_since_genesis();
1489
1490                                 if (td->o.start_delay * 1000 > spent)
1491                                         continue;
1492                         }
1493
1494                         if (td->o.stonewall && (nr_started || nr_running)) {
1495                                 dprint(FD_PROCESS, "%s: stonewall wait\n",
1496                                                         td->o.name);
1497                                 break;
1498                         }
1499
1500                         /*
1501                          * Set state to created. Thread will transition
1502                          * to TD_INITIALIZED when it's done setting up.
1503                          */
1504                         td_set_runstate(td, TD_CREATED);
1505                         map[this_jobs++] = td;
1506                         nr_started++;
1507
1508                         if (td->o.use_thread) {
1509                                 int ret;
1510
1511                                 dprint(FD_PROCESS, "will pthread_create\n");
1512                                 ret = pthread_create(&td->thread, NULL,
1513                                                         thread_main, td);
1514                                 if (ret) {
1515                                         log_err("pthread_create: %s\n",
1516                                                         strerror(ret));
1517                                         nr_started--;
1518                                         break;
1519                                 }
1520                                 ret = pthread_detach(td->thread);
1521                                 if (ret)
1522                                         log_err("pthread_detach: %s",
1523                                                         strerror(ret));
1524                         } else {
1525                                 pid_t pid;
1526                                 dprint(FD_PROCESS, "will fork\n");
1527                                 pid = fork();
1528                                 if (!pid) {
1529                                         int ret = fork_main(shm_id, i);
1530
1531                                         _exit(ret);
1532                                 } else if (i == fio_debug_jobno)
1533                                         *fio_debug_jobp = pid;
1534                         }
1535                         dprint(FD_MUTEX, "wait on startup_mutex\n");
1536                         if (fio_mutex_down_timeout(startup_mutex, 10)) {
1537                                 log_err("fio: job startup hung? exiting.\n");
1538                                 terminate_threads(TERMINATE_ALL);
1539                                 fio_abort = 1;
1540                                 nr_started--;
1541                                 break;
1542                         }
1543                         dprint(FD_MUTEX, "done waiting on startup_mutex\n");
1544                 }
1545
1546                 /*
1547                  * Wait for the started threads to transition to
1548                  * TD_INITIALIZED.
1549                  */
1550                 fio_gettime(&this_start, NULL);
1551                 left = this_jobs;
1552                 while (left && !fio_abort) {
1553                         if (mtime_since_now(&this_start) > JOB_START_TIMEOUT)
1554                                 break;
1555
1556                         usleep(100000);
1557
1558                         for (i = 0; i < this_jobs; i++) {
1559                                 td = map[i];
1560                                 if (!td)
1561                                         continue;
1562                                 if (td->runstate == TD_INITIALIZED) {
1563                                         map[i] = NULL;
1564                                         left--;
1565                                 } else if (td->runstate >= TD_EXITED) {
1566                                         map[i] = NULL;
1567                                         left--;
1568                                         todo--;
1569                                         nr_running++; /* work-around... */
1570                                 }
1571                         }
1572                 }
1573
1574                 if (left) {
1575                         log_err("fio: %d jobs failed to start\n", left);
1576                         for (i = 0; i < this_jobs; i++) {
1577                                 td = map[i];
1578                                 if (!td)
1579                                         continue;
1580                                 kill(td->pid, SIGTERM);
1581                         }
1582                         break;
1583                 }
1584
1585                 /*
1586                  * start created threads (TD_INITIALIZED -> TD_RUNNING).
1587                  */
1588                 for_each_td(td, i) {
1589                         if (td->runstate != TD_INITIALIZED)
1590                                 continue;
1591
1592                         if (in_ramp_time(td))
1593                                 td_set_runstate(td, TD_RAMP);
1594                         else
1595                                 td_set_runstate(td, TD_RUNNING);
1596                         nr_running++;
1597                         nr_started--;
1598                         m_rate += td->o.ratemin[0] + td->o.ratemin[1];
1599                         t_rate += td->o.rate[0] + td->o.rate[1];
1600                         todo--;
1601                         fio_mutex_up(td->mutex);
1602                 }
1603
1604                 reap_threads(&nr_running, &t_rate, &m_rate);
1605
1606                 if (todo)
1607                         usleep(100000);
1608         }
1609
1610         while (nr_running) {
1611                 reap_threads(&nr_running, &t_rate, &m_rate);
1612                 usleep(10000);
1613         }
1614
1615         update_io_ticks();
1616         fio_unpin_memory();
1617 }
1618
1619 int main(int argc, char *argv[])
1620 {
1621         long ps;
1622
1623         sinit();
1624
1625         /*
1626          * We need locale for number printing, if it isn't set then just
1627          * go with the US format.
1628          */
1629         if (!getenv("LC_NUMERIC"))
1630                 setlocale(LC_NUMERIC, "en_US");
1631
1632         ps = sysconf(_SC_PAGESIZE);
1633         if (ps < 0) {
1634                 log_err("Failed to get page size\n");
1635                 return 1;
1636         }
1637
1638         page_size = ps;
1639         page_mask = ps - 1;
1640
1641         fio_keywords_init();
1642
1643         if (parse_options(argc, argv))
1644                 return 1;
1645
1646         if (!thread_number)
1647                 return 0;
1648
1649         if (write_bw_log) {
1650                 setup_log(&agg_io_log[DDIR_READ]);
1651                 setup_log(&agg_io_log[DDIR_WRITE]);
1652         }
1653
1654         startup_mutex = fio_mutex_init(0);
1655         writeout_mutex = fio_mutex_init(1);
1656
1657         set_genesis_time();
1658
1659         status_timer_arm();
1660
1661         cgroup_list = smalloc(sizeof(*cgroup_list));
1662         INIT_FLIST_HEAD(cgroup_list);
1663
1664         run_threads();
1665
1666         if (!fio_abort) {
1667                 show_run_stats();
1668                 if (write_bw_log) {
1669                         __finish_log(agg_io_log[DDIR_READ], "agg-read_bw.log");
1670                         __finish_log(agg_io_log[DDIR_WRITE],
1671                                         "agg-write_bw.log");
1672                 }
1673         }
1674
1675         cgroup_kill(cgroup_list);
1676         sfree(cgroup_list);
1677
1678         fio_mutex_remove(startup_mutex);
1679         fio_mutex_remove(writeout_mutex);
1680         return exit_value;
1681 }