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