fio.c

   1 /*
   2  * fio - the flexible io tester
   3  *
   4  * Copyright (C) 2005 Jens Axboe <axboe@suse.de>
   5  * Copyright (C) 2006 Jens Axboe <axboe@kernel.dk>
   6  *
   7  * The license below covers all files distributed with fio unless otherwise
   8  * noted in the file itself.
   9  *
  10  *  This program is free software; you can redistribute it and/or modify
  11  *  it under the terms of the GNU General Public License version 2 as
  12  *  published by the Free Software Foundation.
  13  *
  14  *  This program is distributed in the hope that it will be useful,
  15  *  but WITHOUT ANY WARRANTY; without even the implied warranty of
  16  *  MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
  17  *  GNU General Public License for more details.
  18  *
  19  *  You should have received a copy of the GNU General Public License
  20  *  along with this program; if not, write to the Free Software
  21  *  Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA  02111-1307  USA
  22  *
  23  */
  24 #include <unistd.h>
  25 #include <fcntl.h>
  26 #include <string.h>
  27 #include <signal.h>
  28 #include <time.h>
  29 #include <locale.h>
  30 #include <assert.h>
  31 #include <sys/stat.h>
  32 #include <sys/wait.h>
  33 #include <sys/ipc.h>
  34 #include <sys/shm.h>
  35 #include <sys/mman.h>
  36
  37 #include "fio.h"
  38 #include "hash.h"
  39 #include "smalloc.h"
  40 #include "verify.h"
  41 #include "diskutil.h"
  42
  43 unsigned long page_mask;
  44 unsigned long page_size;
  45 #define ALIGN(buf)      \
  46         (char *) (((unsigned long) (buf) + page_mask) & ~page_mask)
  47
  48 int groupid = 0;
  49 int thread_number = 0;
  50 int nr_process = 0;
  51 int nr_thread = 0;
  52 int shm_id = 0;
  53 int temp_stall_ts;
  54 unsigned long done_secs = 0;
  55
  56 static struct fio_mutex *startup_mutex;
  57 static struct fio_mutex *writeout_mutex;
  58 static volatile int fio_abort;
  59 static int exit_value;
  60 static struct itimerval itimer;
  61 static pthread_t gtod_thread;
  62
  63 struct io_log *agg_io_log[2];
  64
  65 #define TERMINATE_ALL           (-1)
  66 #define JOB_START_TIMEOUT       (5 * 1000)
  67
  68 void td_set_runstate(struct thread_data *td, int runstate)
  69 {
  70         if (td->runstate == runstate)
  71                 return;
  72
  73         dprint(FD_PROCESS, "pid=%d: runstate %d -> %d\n", (int) td->pid,
  74                                                 td->runstate, runstate);
  75         td->runstate = runstate;
  76 }
  77
  78 static void terminate_threads(int group_id)
  79 {
  80         struct thread_data *td;
  81         int i;
  82
  83         dprint(FD_PROCESS, "terminate group_id=%d\n", group_id);
  84
  85         for_each_td(td, i) {
  86                 if (group_id == TERMINATE_ALL || groupid == td->groupid) {
  87                         dprint(FD_PROCESS, "setting terminate on %s/%d\n",
  88                                                 td->o.name, (int) td->pid);
  89                         td->terminate = 1;
  90                         td->o.start_delay = 0;
  91
  92                         /*
  93                          * if the thread is running, just let it exit
  94                          */
  95                         if (td->runstate < TD_RUNNING)
  96                                 kill(td->pid, SIGQUIT);
  97                         else {
  98                                 struct ioengine_ops *ops = td->io_ops;
  99
 100                                 if (ops && (ops->flags & FIO_SIGQUIT))
 101                                         kill(td->pid, SIGQUIT);
 102                         }
 103                 }
 104         }
 105 }
 106
 107 static void status_timer_arm(void)
 108 {
 109         itimer.it_value.tv_sec = 0;
 110         itimer.it_value.tv_usec = DISK_UTIL_MSEC * 1000;
 111         setitimer(ITIMER_REAL, &itimer, NULL);
 112 }
 113
 114 static void sig_alrm(int fio_unused sig)
 115 {
 116         if (threads) {
 117                 update_io_ticks();
 118                 print_thread_status();
 119                 status_timer_arm();
 120         }
 121 }
 122
 123 /*
 124  * Happens on thread runs with ctrl-c, ignore our own SIGQUIT
 125  */
 126 static void sig_quit(int sig)
 127 {
 128 }
 129
 130 static void sig_int(int sig)
 131 {
 132         if (threads) {
 133                 printf("\nfio: terminating on signal %d\n", sig);
 134                 fflush(stdout);
 135                 terminate_threads(TERMINATE_ALL);
 136         }
 137 }
 138
 139 static void sig_ill(int fio_unused sig)
 140 {
 141         if (!threads)
 142                 return;
 143
 144         log_err("fio: illegal instruction. your cpu does not support "
 145                 "the sse4.2 instruction for crc32c\n");
 146         terminate_threads(TERMINATE_ALL);
 147         exit(4);
 148 }
 149
 150 static void set_sig_handlers(void)
 151 {
 152         struct sigaction act;
 153
 154         memset(&act, 0, sizeof(act));
 155         act.sa_handler = sig_alrm;
 156         act.sa_flags = SA_RESTART;
 157         sigaction(SIGALRM, &act, NULL);
 158
 159         memset(&act, 0, sizeof(act));
 160         act.sa_handler = sig_int;
 161         act.sa_flags = SA_RESTART;
 162         sigaction(SIGINT, &act, NULL);
 163
 164         memset(&act, 0, sizeof(act));
 165         act.sa_handler = sig_ill;
 166         act.sa_flags = SA_RESTART;
 167         sigaction(SIGILL, &act, NULL);
 168
 169         memset(&act, 0, sizeof(act));
 170         act.sa_handler = sig_quit;
 171         act.sa_flags = SA_RESTART;
 172         sigaction(SIGQUIT, &act, NULL);
 173 }
 174
 175 /*
 176  * Check if we are above the minimum rate given.
 177  */
 178 static int __check_min_rate(struct thread_data *td, struct timeval *now,
 179                             enum td_ddir ddir)
 180 {
 181         unsigned long long bytes = 0;
 182         unsigned long iops = 0;
 183         unsigned long spent;
 184         unsigned long rate;
 185         unsigned int ratemin = 0;
 186         unsigned int rate_iops = 0;
 187         unsigned int rate_iops_min = 0;
 188
 189         if (!td->o.ratemin[ddir] && !td->o.rate_iops_min[ddir])
 190                 return 0;
 191
 192         /*
 193          * allow a 2 second settle period in the beginning
 194          */
 195         if (mtime_since(&td->start, now) < 2000)
 196                 return 0;
 197
 198         iops += td->io_blocks[ddir];
 199         bytes += td->this_io_bytes[ddir];
 200         ratemin += td->o.ratemin[ddir];
 201         rate_iops += td->o.rate_iops[ddir];
 202         rate_iops_min += td->o.rate_iops_min[ddir];
 203
 204         /*
 205          * if rate blocks is set, sample is running
 206          */
 207         if (td->rate_bytes[ddir] || td->rate_blocks[ddir]) {
 208                 spent = mtime_since(&td->lastrate[ddir], now);
 209                 if (spent < td->o.ratecycle)
 210                         return 0;
 211
 212                 if (td->o.rate[ddir]) {
 213                         /*
 214                          * check bandwidth specified rate
 215                          */
 216                         if (bytes < td->rate_bytes[ddir]) {
 217                                 log_err("%s: min rate %u not met\n", td->o.name,
 218                                                                 ratemin);
 219                                 return 1;
 220                         } else {
 221                                 rate = ((bytes - td->rate_bytes[ddir]) * 1000) / spent;
 222                                 if (rate < ratemin ||
 223                                     bytes < td->rate_bytes[ddir]) {
 224                                         log_err("%s: min rate %u not met, got"
 225                                                 " %luKiB/sec\n", td->o.name,
 226                                                         ratemin, rate);
 227                                         return 1;
 228                                 }
 229                         }
 230                 } else {
 231                         /*
 232                          * checks iops specified rate
 233                          */
 234                         if (iops < rate_iops) {
 235                                 log_err("%s: min iops rate %u not met\n",
 236                                                 td->o.name, rate_iops);
 237                                 return 1;
 238                         } else {
 239                                 rate = ((iops - td->rate_blocks[ddir]) * 1000) / spent;
 240                                 if (rate < rate_iops_min ||
 241                                     iops < td->rate_blocks[ddir]) {
 242                                         log_err("%s: min iops rate %u not met,"
 243                                                 " got %lu\n", td->o.name,
 244                                                         rate_iops_min, rate);
 245                                 }
 246                         }
 247                 }
 248         }
 249
 250         td->rate_bytes[ddir] = bytes;
 251         td->rate_blocks[ddir] = iops;
 252         memcpy(&td->lastrate[ddir], now, sizeof(*now));
 253         return 0;
 254 }
 255
 256 static int check_min_rate(struct thread_data *td, struct timeval *now,
 257                           unsigned long *bytes_done)
 258 {
 259         int ret = 0;
 260
 261         if (bytes_done[0])
 262                 ret |= __check_min_rate(td, now, 0);
 263         if (bytes_done[1])
 264                 ret |= __check_min_rate(td, now, 1);
 265
 266         return ret;
 267 }
 268
 269 static inline int runtime_exceeded(struct thread_data *td, struct timeval *t)
 270 {
 271         if (!td->o.timeout)
 272                 return 0;
 273         if (mtime_since(&td->epoch, t) >= td->o.timeout * 1000)
 274                 return 1;
 275
 276         return 0;
 277 }
 278
 279 /*
 280  * When job exits, we can cancel the in-flight IO if we are using async
 281  * io. Attempt to do so.
 282  */
 283 static void cleanup_pending_aio(struct thread_data *td)
 284 {
 285         struct flist_head *entry, *n;
 286         struct io_u *io_u;
 287         int r;
 288
 289         /*
 290          * get immediately available events, if any
 291          */
 292         r = io_u_queued_complete(td, 0, NULL);
 293         if (r < 0)
 294                 return;
 295
 296         /*
 297          * now cancel remaining active events
 298          */
 299         if (td->io_ops->cancel) {
 300                 flist_for_each_safe(entry, n, &td->io_u_busylist) {
 301                         io_u = flist_entry(entry, struct io_u, list);
 302
 303                         /*
 304                          * if the io_u isn't in flight, then that generally
 305                          * means someone leaked an io_u. complain but fix
 306                          * it up, so we don't stall here.
 307                          */
 308                         if ((io_u->flags & IO_U_F_FLIGHT) == 0) {
 309                                 log_err("fio: non-busy IO on busy list\n");
 310                                 put_io_u(td, io_u);
 311                         } else {
 312                                 r = td->io_ops->cancel(td, io_u);
 313                                 if (!r)
 314                                         put_io_u(td, io_u);
 315                         }
 316                 }
 317         }
 318
 319         if (td->cur_depth)
 320                 r = io_u_queued_complete(td, td->cur_depth, NULL);
 321 }
 322
 323 /*
 324  * Helper to handle the final sync of a file. Works just like the normal
 325  * io path, just does everything sync.
 326  */
 327 static int fio_io_sync(struct thread_data *td, struct fio_file *f)
 328 {
 329         struct io_u *io_u = __get_io_u(td);
 330         int ret;
 331
 332         if (!io_u)
 333                 return 1;
 334
 335         io_u->ddir = DDIR_SYNC;
 336         io_u->file = f;
 337
 338         if (td_io_prep(td, io_u)) {
 339                 put_io_u(td, io_u);
 340                 return 1;
 341         }
 342
 343 requeue:
 344         ret = td_io_queue(td, io_u);
 345         if (ret < 0) {
 346                 td_verror(td, io_u->error, "td_io_queue");
 347                 put_io_u(td, io_u);
 348                 return 1;
 349         } else if (ret == FIO_Q_QUEUED) {
 350                 if (io_u_queued_complete(td, 1, NULL) < 0)
 351                         return 1;
 352         } else if (ret == FIO_Q_COMPLETED) {
 353                 if (io_u->error) {
 354                         td_verror(td, io_u->error, "td_io_queue");
 355                         return 1;
 356                 }
 357
 358                 if (io_u_sync_complete(td, io_u, NULL) < 0)
 359                         return 1;
 360         } else if (ret == FIO_Q_BUSY) {
 361                 if (td_io_commit(td))
 362                         return 1;
 363                 goto requeue;
 364         }
 365
 366         return 0;
 367 }
 368
 369 static inline void update_tv_cache(struct thread_data *td)
 370 {
 371         if ((++td->tv_cache_nr & td->tv_cache_mask) == td->tv_cache_mask)
 372                 fio_gettime(&td->tv_cache, NULL);
 373 }
 374
 375 /*
 376  * The main verify engine. Runs over the writes we previously submitted,
 377  * reads the blocks back in, and checks the crc/md5 of the data.
 378  */
 379 static void do_verify(struct thread_data *td)
 380 {
 381         struct fio_file *f;
 382         struct io_u *io_u;
 383         int ret, min_events;
 384         unsigned int i;
 385
 386         /*
 387          * sync io first and invalidate cache, to make sure we really
 388          * read from disk.
 389          */
 390         for_each_file(td, f, i) {
 391                 if (!fio_file_open(f))
 392                         continue;
 393                 if (fio_io_sync(td, f))
 394                         break;
 395                 if (file_invalidate_cache(td, f))
 396                         break;
 397         }
 398
 399         if (td->error)
 400                 return;
 401
 402         td_set_runstate(td, TD_VERIFYING);
 403
 404         io_u = NULL;
 405         while (!td->terminate) {
 406                 int ret2, full;
 407
 408                 io_u = __get_io_u(td);
 409                 if (!io_u)
 410                         break;
 411
 412                 update_tv_cache(td);
 413
 414                 if (runtime_exceeded(td, &td->tv_cache)) {
 415                         put_io_u(td, io_u);
 416                         td->terminate = 1;
 417                         break;
 418                 }
 419
 420                 if (get_next_verify(td, io_u)) {
 421                         put_io_u(td, io_u);
 422                         break;
 423                 }
 424
 425                 if (td_io_prep(td, io_u)) {
 426                         put_io_u(td, io_u);
 427                         break;
 428                 }
 429
 430                 io_u->end_io = verify_io_u;
 431
 432                 ret = td_io_queue(td, io_u);
 433                 switch (ret) {
 434                 case FIO_Q_COMPLETED:
 435                         if (io_u->error)
 436                                 ret = -io_u->error;
 437                         else if (io_u->resid) {
 438                                 int bytes = io_u->xfer_buflen - io_u->resid;
 439                                 struct fio_file *f = io_u->file;
 440
 441                                 /*
 442                                  * zero read, fail
 443                                  */
 444                                 if (!bytes) {
 445                                         td_verror(td, EIO, "full resid");
 446                                         put_io_u(td, io_u);
 447                                         break;
 448                                 }
 449
 450                                 io_u->xfer_buflen = io_u->resid;
 451                                 io_u->xfer_buf += bytes;
 452                                 io_u->offset += bytes;
 453
 454                                 td->ts.short_io_u[io_u->ddir]++;
 455
 456                                 if (io_u->offset == f->real_file_size)
 457                                         goto sync_done;
 458
 459                                 requeue_io_u(td, &io_u);
 460                         } else {
 461 sync_done:
 462                                 ret = io_u_sync_complete(td, io_u, NULL);
 463                                 if (ret < 0)
 464                                         break;
 465                         }
 466                         continue;
 467                 case FIO_Q_QUEUED:
 468                         break;
 469                 case FIO_Q_BUSY:
 470                         requeue_io_u(td, &io_u);
 471                         ret2 = td_io_commit(td);
 472                         if (ret2 < 0)
 473                                 ret = ret2;
 474                         break;
 475                 default:
 476                         assert(ret < 0);
 477                         td_verror(td, -ret, "td_io_queue");
 478                         break;
 479                 }
 480
 481                 if (ret < 0 || td->error)
 482                         break;
 483
 484                 /*
 485                  * if we can queue more, do so. but check if there are
 486                  * completed io_u's first.
 487                  */
 488                 full = queue_full(td) || ret == FIO_Q_BUSY;
 489                 if (full || !td->o.iodepth_batch_complete) {
 490                         min_events = td->o.iodepth_batch_complete;
 491                         if (full && !min_events)
 492                                 min_events = 1;
 493
 494                         do {
 495                                 /*
 496                                  * Reap required number of io units, if any,
 497                                  * and do the verification on them through
 498                                  * the callback handler
 499                                  */
 500                                 if (io_u_queued_complete(td, min_events, NULL) < 0) {
 501                                         ret = -1;
 502                                         break;
 503                                 }
 504                         } while (full && (td->cur_depth > td->o.iodepth_low));
 505                 }
 506                 if (ret < 0)
 507                         break;
 508         }
 509
 510         if (!td->error) {
 511                 min_events = td->cur_depth;
 512
 513                 if (min_events)
 514                         ret = io_u_queued_complete(td, min_events, NULL);
 515         } else
 516                 cleanup_pending_aio(td);
 517
 518         td_set_runstate(td, TD_RUNNING);
 519 }
 520
 521 /*
 522  * Main IO worker function. It retrieves io_u's to process and queues
 523  * and reaps them, checking for rate and errors along the way.
 524  */
 525 static void do_io(struct thread_data *td)
 526 {
 527         unsigned int i;
 528         int ret = 0;
 529
 530         if (in_ramp_time(td))
 531                 td_set_runstate(td, TD_RAMP);
 532         else
 533                 td_set_runstate(td, TD_RUNNING);
 534
 535         while ((td->this_io_bytes[0] + td->this_io_bytes[1]) < td->o.size) {
 536                 struct timeval comp_time;
 537                 unsigned long bytes_done[2] = { 0, 0 };
 538                 int min_evts = 0;
 539                 struct io_u *io_u;
 540                 int ret2, full;
 541
 542                 if (td->terminate)
 543                         break;
 544
 545                 io_u = get_io_u(td);
 546                 if (!io_u)
 547                         break;
 548
 549                 update_tv_cache(td);
 550
 551                 if (runtime_exceeded(td, &td->tv_cache)) {
 552                         put_io_u(td, io_u);
 553                         td->terminate = 1;
 554                         break;
 555                 }
 556
 557                 /*
 558                  * Add verification end_io handler, if asked to verify
 559                  * a previously written file.
 560                  */
 561                 if (td->o.verify != VERIFY_NONE && io_u->ddir == DDIR_READ) {
 562                         io_u->end_io = verify_io_u;
 563                         td_set_runstate(td, TD_VERIFYING);
 564                 } else if (in_ramp_time(td))
 565                         td_set_runstate(td, TD_RAMP);
 566                 else
 567                         td_set_runstate(td, TD_RUNNING);
 568
 569                 ret = td_io_queue(td, io_u);
 570                 switch (ret) {
 571                 case FIO_Q_COMPLETED:
 572                         if (io_u->error)
 573                                 ret = -io_u->error;
 574                         else if (io_u->resid) {
 575                                 int bytes = io_u->xfer_buflen - io_u->resid;
 576                                 struct fio_file *f = io_u->file;
 577
 578                                 /*
 579                                  * zero read, fail
 580                                  */
 581                                 if (!bytes) {
 582                                         td_verror(td, EIO, "full resid");
 583                                         put_io_u(td, io_u);
 584                                         break;
 585                                 }
 586
 587                                 io_u->xfer_buflen = io_u->resid;
 588                                 io_u->xfer_buf += bytes;
 589                                 io_u->offset += bytes;
 590
 591                                 td->ts.short_io_u[io_u->ddir]++;
 592
 593                                 if (io_u->offset == f->real_file_size)
 594                                         goto sync_done;
 595
 596                                 requeue_io_u(td, &io_u);
 597                         } else {
 598 sync_done:
 599                                 if (__should_check_rate(td, 0) ||
 600                                     __should_check_rate(td, 1))
 601                                         fio_gettime(&comp_time, NULL);
 602
 603                                 ret = io_u_sync_complete(td, io_u, bytes_done);
 604                                 if (ret < 0)
 605                                         break;
 606                         }
 607                         break;
 608                 case FIO_Q_QUEUED:
 609                         /*
 610                          * if the engine doesn't have a commit hook,
 611                          * the io_u is really queued. if it does have such
 612                          * a hook, it has to call io_u_queued() itself.
 613                          */
 614                         if (td->io_ops->commit == NULL)
 615                                 io_u_queued(td, io_u);
 616                         break;
 617                 case FIO_Q_BUSY:
 618                         requeue_io_u(td, &io_u);
 619                         ret2 = td_io_commit(td);
 620                         if (ret2 < 0)
 621                                 ret = ret2;
 622                         break;
 623                 default:
 624                         assert(ret < 0);
 625                         put_io_u(td, io_u);
 626                         break;
 627                 }
 628
 629                 if (ret < 0 || td->error)
 630                         break;
 631
 632                 /*
 633                  * See if we need to complete some commands
 634                  */
 635                 full = queue_full(td) || ret == FIO_Q_BUSY;
 636                 if (full || !td->o.iodepth_batch_complete) {
 637                         min_evts = td->o.iodepth_batch_complete;
 638                         if (full && !min_evts)
 639                                 min_evts = 1;
 640
 641                         if (__should_check_rate(td, 0) ||
 642                             __should_check_rate(td, 1))
 643                                 fio_gettime(&comp_time, NULL);
 644
 645                         do {
 646                                 ret = io_u_queued_complete(td, min_evts, bytes_done);
 647                                 if (ret < 0)
 648                                         break;
 649
 650                         } while (full && (td->cur_depth > td->o.iodepth_low));
 651                 }
 652
 653                 if (ret < 0)
 654                         break;
 655                 if (!(bytes_done[0] + bytes_done[1]))
 656                         continue;
 657
 658                 if (!in_ramp_time(td) && should_check_rate(td, bytes_done)) {
 659                         if (check_min_rate(td, &comp_time, bytes_done)) {
 660                                 if (exitall_on_terminate)
 661                                         terminate_threads(td->groupid);
 662                                 td_verror(td, EIO, "check_min_rate");
 663                                 break;
 664                         }
 665                 }
 666
 667                 if (td->o.thinktime) {
 668                         unsigned long long b;
 669
 670                         b = td->io_blocks[0] + td->io_blocks[1];
 671                         if (!(b % td->o.thinktime_blocks)) {
 672                                 int left;
 673
 674                                 if (td->o.thinktime_spin)
 675                                         usec_spin(td->o.thinktime_spin);
 676
 677                                 left = td->o.thinktime - td->o.thinktime_spin;
 678                                 if (left)
 679                                         usec_sleep(td, left);
 680                         }
 681                 }
 682         }
 683
 684         if (td->o.fill_device && td->error == ENOSPC) {
 685                 td->error = 0;
 686                 td->terminate = 1;
 687         }
 688         if (!td->error) {
 689                 struct fio_file *f;
 690
 691                 i = td->cur_depth;
 692                 if (i)
 693                         ret = io_u_queued_complete(td, i, NULL);
 694
 695                 if (should_fsync(td) && td->o.end_fsync) {
 696                         td_set_runstate(td, TD_FSYNCING);
 697
 698                         for_each_file(td, f, i) {
 699                                 if (!fio_file_open(f))
 700                                         continue;
 701                                 fio_io_sync(td, f);
 702                         }
 703                 }
 704         } else
 705                 cleanup_pending_aio(td);
 706
 707         /*
 708          * stop job if we failed doing any IO
 709          */
 710         if ((td->this_io_bytes[0] + td->this_io_bytes[1]) == 0)
 711                 td->done = 1;
 712 }
 713
 714 static void cleanup_io_u(struct thread_data *td)
 715 {
 716         struct flist_head *entry, *n;
 717         struct io_u *io_u;
 718
 719         flist_for_each_safe(entry, n, &td->io_u_freelist) {
 720                 io_u = flist_entry(entry, struct io_u, list);
 721
 722                 flist_del(&io_u->list);
 723                 free(io_u);
 724         }
 725
 726         free_io_mem(td);
 727 }
 728
 729 static int init_io_u(struct thread_data *td)
 730 {
 731         struct io_u *io_u;
 732         unsigned int max_bs;
 733         int cl_align, i, max_units;
 734         char *p;
 735
 736         max_units = td->o.iodepth;
 737         max_bs = max(td->o.max_bs[DDIR_READ], td->o.max_bs[DDIR_WRITE]);
 738         td->orig_buffer_size = (unsigned long long) max_bs
 739                                         * (unsigned long long) max_units;
 740
 741         if (td->o.mem_type == MEM_SHMHUGE || td->o.mem_type == MEM_MMAPHUGE) {
 742                 unsigned long bs;
 743
 744                 bs = td->orig_buffer_size + td->o.hugepage_size - 1;
 745                 td->orig_buffer_size = bs & ~(td->o.hugepage_size - 1);
 746         }
 747
 748         if (td->orig_buffer_size != (size_t) td->orig_buffer_size) {
 749                 log_err("fio: IO memory too large. Reduce max_bs or iodepth\n");
 750                 return 1;
 751         }
 752
 753         if (allocate_io_mem(td))
 754                 return 1;
 755
 756         if (td->o.odirect)
 757                 p = ALIGN(td->orig_buffer);
 758         else
 759                 p = td->orig_buffer;
 760
 761         cl_align = os_cache_line_size();
 762
 763         for (i = 0; i < max_units; i++) {
 764                 void *ptr;
 765
 766                 if (td->terminate)
 767                         return 1;
 768
 769                 if (posix_memalign(&ptr, cl_align, sizeof(*io_u))) {
 770                         log_err("fio: posix_memalign=%s\n", strerror(errno));
 771                         break;
 772                 }
 773
 774                 io_u = ptr;
 775                 memset(io_u, 0, sizeof(*io_u));
 776                 INIT_FLIST_HEAD(&io_u->list);
 777
 778                 if (!(td->io_ops->flags & FIO_NOIO)) {
 779                         io_u->buf = p + max_bs * i;
 780
 781                         if (td_write(td) && !td->o.refill_buffers)
 782                                 io_u_fill_buffer(td, io_u, max_bs);
 783                 }
 784
 785                 io_u->index = i;
 786                 io_u->flags = IO_U_F_FREE;
 787                 flist_add(&io_u->list, &td->io_u_freelist);
 788         }
 789
 790         return 0;
 791 }
 792
 793 static int switch_ioscheduler(struct thread_data *td)
 794 {
 795         char tmp[256], tmp2[128];
 796         FILE *f;
 797         int ret;
 798
 799         if (td->io_ops->flags & FIO_DISKLESSIO)
 800                 return 0;
 801
 802         sprintf(tmp, "%s/queue/scheduler", td->sysfs_root);
 803
 804         f = fopen(tmp, "r+");
 805         if (!f) {
 806                 if (errno == ENOENT) {
 807                         log_err("fio: os or kernel doesn't support IO scheduler"
 808                                 " switching\n");
 809                         return 0;
 810                 }
 811                 td_verror(td, errno, "fopen iosched");
 812                 return 1;
 813         }
 814
 815         /*
 816          * Set io scheduler.
 817          */
 818         ret = fwrite(td->o.ioscheduler, strlen(td->o.ioscheduler), 1, f);
 819         if (ferror(f) || ret != 1) {
 820                 td_verror(td, errno, "fwrite");
 821                 fclose(f);
 822                 return 1;
 823         }
 824
 825         rewind(f);
 826
 827         /*
 828          * Read back and check that the selected scheduler is now the default.
 829          */
 830         ret = fread(tmp, 1, sizeof(tmp), f);
 831         if (ferror(f) || ret < 0) {
 832                 td_verror(td, errno, "fread");
 833                 fclose(f);
 834                 return 1;
 835         }
 836
 837         sprintf(tmp2, "[%s]", td->o.ioscheduler);
 838         if (!strstr(tmp, tmp2)) {
 839                 log_err("fio: io scheduler %s not found\n", td->o.ioscheduler);
 840                 td_verror(td, EINVAL, "iosched_switch");
 841                 fclose(f);
 842                 return 1;
 843         }
 844
 845         fclose(f);
 846         return 0;
 847 }
 848
 849 static int keep_running(struct thread_data *td)
 850 {
 851         unsigned long long io_done;
 852
 853         if (td->done)
 854                 return 0;
 855         if (td->o.time_based)
 856                 return 1;
 857         if (td->o.loops) {
 858                 td->o.loops--;
 859                 return 1;
 860         }
 861
 862         io_done = td->io_bytes[DDIR_READ] + td->io_bytes[DDIR_WRITE]
 863                         + td->io_skip_bytes;
 864         if (io_done < td->o.size)
 865                 return 1;
 866
 867         return 0;
 868 }
 869
 870 static void reset_io_counters(struct thread_data *td)
 871 {
 872         td->ts.stat_io_bytes[0] = td->ts.stat_io_bytes[1] = 0;
 873         td->this_io_bytes[0] = td->this_io_bytes[1] = 0;
 874         td->zone_bytes = 0;
 875         td->rate_bytes[0] = td->rate_bytes[1] = 0;
 876         td->rate_blocks[0] = td->rate_blocks[1] = 0;
 877
 878         td->last_was_sync = 0;
 879
 880         /*
 881          * reset file done count if we are to start over
 882          */
 883         if (td->o.time_based || td->o.loops)
 884                 td->nr_done_files = 0;
 885
 886         /*
 887          * Set the same seed to get repeatable runs
 888          */
 889         td_fill_rand_seeds(td);
 890 }
 891
 892 void reset_all_stats(struct thread_data *td)
 893 {
 894         struct timeval tv;
 895         int i;
 896
 897         reset_io_counters(td);
 898
 899         for (i = 0; i < 2; i++) {
 900                 td->io_bytes[i] = 0;
 901                 td->io_blocks[i] = 0;
 902                 td->io_issues[i] = 0;
 903                 td->ts.total_io_u[i] = 0;
 904         }
 905
 906         fio_gettime(&tv, NULL);
 907         memcpy(&td->epoch, &tv, sizeof(tv));
 908         memcpy(&td->start, &tv, sizeof(tv));
 909 }
 910
 911 static void clear_io_state(struct thread_data *td)
 912 {
 913         struct fio_file *f;
 914         unsigned int i;
 915
 916         reset_io_counters(td);
 917
 918         close_files(td);
 919         for_each_file(td, f, i)
 920                 fio_file_clear_done(f);
 921 }
 922
 923 static int exec_string(const char *string)
 924 {
 925         int ret, newlen = strlen(string) + 1 + 8;
 926         char *str;
 927
 928         str = malloc(newlen);
 929         sprintf(str, "sh -c %s", string);
 930
 931         ret = system(str);
 932         if (ret == -1)
 933                 log_err("fio: exec of cmd <%s> failed\n", str);
 934
 935         free(str);
 936         return ret;
 937 }
 938
 939 /*
 940  * Entry point for the thread based jobs. The process based jobs end up
 941  * here as well, after a little setup.
 942  */
 943 static void *thread_main(void *data)
 944 {
 945         unsigned long long runtime[2], elapsed;
 946         struct thread_data *td = data;
 947         int clear_state;
 948
 949         if (!td->o.use_thread)
 950                 setsid();
 951
 952         td->pid = getpid();
 953
 954         dprint(FD_PROCESS, "jobs pid=%d started\n", (int) td->pid);
 955
 956         INIT_FLIST_HEAD(&td->io_u_freelist);
 957         INIT_FLIST_HEAD(&td->io_u_busylist);
 958         INIT_FLIST_HEAD(&td->io_u_requeues);
 959         INIT_FLIST_HEAD(&td->io_log_list);
 960         INIT_FLIST_HEAD(&td->io_hist_list);
 961         td->io_hist_tree = RB_ROOT;
 962
 963         td_set_runstate(td, TD_INITIALIZED);
 964         dprint(FD_MUTEX, "up startup_mutex\n");
 965         fio_mutex_up(startup_mutex);
 966         dprint(FD_MUTEX, "wait on td->mutex\n");
 967         fio_mutex_down(td->mutex);
 968         dprint(FD_MUTEX, "done waiting on td->mutex\n");
 969
 970         /*
 971          * the ->mutex mutex is now no longer used, close it to avoid
 972          * eating a file descriptor
 973          */
 974         fio_mutex_remove(td->mutex);
 975
 976         /*
 977          * May alter parameters that init_io_u() will use, so we need to
 978          * do this first.
 979          */
 980         if (init_iolog(td))
 981                 goto err;
 982
 983         if (init_io_u(td))
 984                 goto err;
 985
 986         if (td->o.cpumask_set && fio_setaffinity(td) == -1) {
 987                 td_verror(td, errno, "cpu_set_affinity");
 988                 goto err;
 989         }
 990
 991         /*
 992          * If we have a gettimeofday() thread, make sure we exclude that
 993          * thread from this job
 994          */
 995         if (td->o.gtod_cpu) {
 996                 fio_cpu_clear(&td->o.cpumask, td->o.gtod_cpu);
 997                 if (fio_setaffinity(td) == -1) {
 998                         td_verror(td, errno, "cpu_set_affinity");
 999                         goto err;
1000                 }
1001         }
1002
1003         if (td->ioprio_set) {
1004                 if (ioprio_set(IOPRIO_WHO_PROCESS, 0, td->ioprio) == -1) {
1005                         td_verror(td, errno, "ioprio_set");
1006                         goto err;
1007                 }
1008         }
1009
1010         if (nice(td->o.nice) == -1) {
1011                 td_verror(td, errno, "nice");
1012                 goto err;
1013         }
1014
1015         if (td->o.ioscheduler && switch_ioscheduler(td))
1016                 goto err;
1017
1018         if (!td->o.create_serialize && setup_files(td))
1019                 goto err;
1020
1021         if (td_io_init(td))
1022                 goto err;
1023
1024         if (init_random_map(td))
1025                 goto err;
1026
1027         if (td->o.exec_prerun) {
1028                 if (exec_string(td->o.exec_prerun))
1029                         goto err;
1030         }
1031
1032         if (td->o.pre_read) {
1033                 if (pre_read_files(td) < 0)
1034                         goto err;
1035         }
1036
1037         fio_gettime(&td->epoch, NULL);
1038         getrusage(RUSAGE_SELF, &td->ts.ru_start);
1039
1040         runtime[0] = runtime[1] = 0;
1041         clear_state = 0;
1042         while (keep_running(td)) {
1043                 fio_gettime(&td->start, NULL);
1044                 memcpy(&td->ts.stat_sample_time, &td->start, sizeof(td->start));
1045                 memcpy(&td->tv_cache, &td->start, sizeof(td->start));
1046
1047                 if (td->o.ratemin[0] || td->o.ratemin[1])
1048                         memcpy(&td->lastrate, &td->ts.stat_sample_time,
1049                                                         sizeof(td->lastrate));
1050
1051                 if (clear_state)
1052                         clear_io_state(td);
1053
1054                 prune_io_piece_log(td);
1055
1056                 do_io(td);
1057
1058                 clear_state = 1;
1059
1060                 if (td_read(td) && td->io_bytes[DDIR_READ]) {
1061                         elapsed = utime_since_now(&td->start);
1062                         runtime[DDIR_READ] += elapsed;
1063                 }
1064                 if (td_write(td) && td->io_bytes[DDIR_WRITE]) {
1065                         elapsed = utime_since_now(&td->start);
1066                         runtime[DDIR_WRITE] += elapsed;
1067                 }
1068
1069                 if (td->error || td->terminate)
1070                         break;
1071
1072                 if (!td->o.do_verify ||
1073                     td->o.verify == VERIFY_NONE ||
1074                     (td->io_ops->flags & FIO_UNIDIR))
1075                         continue;
1076
1077                 clear_io_state(td);
1078
1079                 fio_gettime(&td->start, NULL);
1080
1081                 do_verify(td);
1082
1083                 runtime[DDIR_READ] += utime_since_now(&td->start);
1084
1085                 if (td->error || td->terminate)
1086                         break;
1087         }
1088
1089         update_rusage_stat(td);
1090         td->ts.runtime[0] = (runtime[0] + 999) / 1000;
1091         td->ts.runtime[1] = (runtime[1] + 999) / 1000;
1092         td->ts.total_run_time = mtime_since_now(&td->epoch);
1093         td->ts.io_bytes[0] = td->io_bytes[0];
1094         td->ts.io_bytes[1] = td->io_bytes[1];
1095
1096         fio_mutex_down(writeout_mutex);
1097         if (td->ts.bw_log) {
1098                 if (td->o.bw_log_file) {
1099                         finish_log_named(td, td->ts.bw_log,
1100                                                 td->o.bw_log_file, "bw");
1101                 } else
1102                         finish_log(td, td->ts.bw_log, "bw");
1103         }
1104         if (td->ts.slat_log) {
1105                 if (td->o.lat_log_file) {
1106                         finish_log_named(td, td->ts.slat_log,
1107                                                 td->o.lat_log_file, "slat");
1108                 } else
1109                         finish_log(td, td->ts.slat_log, "slat");
1110         }
1111         if (td->ts.clat_log) {
1112                 if (td->o.lat_log_file) {
1113                         finish_log_named(td, td->ts.clat_log,
1114                                                 td->o.lat_log_file, "clat");
1115                 } else
1116                         finish_log(td, td->ts.clat_log, "clat");
1117         }
1118         fio_mutex_up(writeout_mutex);
1119         if (td->o.exec_postrun)
1120                 exec_string(td->o.exec_postrun);
1121
1122         if (exitall_on_terminate)
1123                 terminate_threads(td->groupid);
1124
1125 err:
1126         if (td->error)
1127                 printf("fio: pid=%d, err=%d/%s\n", (int) td->pid, td->error,
1128                                                         td->verror);
1129         close_and_free_files(td);
1130         close_ioengine(td);
1131         cleanup_io_u(td);
1132
1133         if (td->o.cpumask_set) {
1134                 int ret = fio_cpuset_exit(&td->o.cpumask);
1135
1136                 td_verror(td, ret, "fio_cpuset_exit");
1137         }
1138
1139         /*
1140          * do this very late, it will log file closing as well
1141          */
1142         if (td->o.write_iolog_file)
1143                 write_iolog_close(td);
1144
1145         options_mem_free(td);
1146         td_set_runstate(td, TD_EXITED);
1147         return (void *) (unsigned long) td->error;
1148 }
1149
1150 /*
1151  * We cannot pass the td data into a forked process, so attach the td and
1152  * pass it to the thread worker.
1153  */
1154 static int fork_main(int shmid, int offset)
1155 {
1156         struct thread_data *td;
1157         void *data, *ret;
1158
1159         data = shmat(shmid, NULL, 0);
1160         if (data == (void *) -1) {
1161                 int __err = errno;
1162
1163                 perror("shmat");
1164                 return __err;
1165         }
1166
1167         td = data + offset * sizeof(struct thread_data);
1168         ret = thread_main(td);
1169         shmdt(data);
1170         return (int) (unsigned long) ret;
1171 }
1172
1173 /*
1174  * Run over the job map and reap the threads that have exited, if any.
1175  */
1176 static void reap_threads(int *nr_running, int *t_rate, int *m_rate)
1177 {
1178         struct thread_data *td;
1179         int i, cputhreads, realthreads, pending, status, ret;
1180
1181         /*
1182          * reap exited threads (TD_EXITED -> TD_REAPED)
1183          */
1184         realthreads = pending = cputhreads = 0;
1185         for_each_td(td, i) {
1186                 int flags = 0;
1187
1188                 /*
1189                  * ->io_ops is NULL for a thread that has closed its
1190                  * io engine
1191                  */
1192                 if (td->io_ops && !strcmp(td->io_ops->name, "cpuio"))
1193                         cputhreads++;
1194                 else
1195                         realthreads++;
1196
1197                 if (!td->pid) {
1198                         pending++;
1199                         continue;
1200                 }
1201                 if (td->runstate == TD_REAPED)
1202                         continue;
1203                 if (td->o.use_thread) {
1204                         if (td->runstate == TD_EXITED) {
1205                                 td_set_runstate(td, TD_REAPED);
1206                                 goto reaped;
1207                         }
1208                         continue;
1209                 }
1210
1211                 flags = WNOHANG;
1212                 if (td->runstate == TD_EXITED)
1213                         flags = 0;
1214
1215                 /*
1216                  * check if someone quit or got killed in an unusual way
1217                  */
1218                 ret = waitpid(td->pid, &status, flags);
1219                 if (ret < 0) {
1220                         if (errno == ECHILD) {
1221                                 log_err("fio: pid=%d disappeared %d\n",
1222                                                 (int) td->pid, td->runstate);
1223                                 td_set_runstate(td, TD_REAPED);
1224                                 goto reaped;
1225                         }
1226                         perror("waitpid");
1227                 } else if (ret == td->pid) {
1228                         if (WIFSIGNALED(status)) {
1229                                 int sig = WTERMSIG(status);
1230
1231                                 if (sig != SIGQUIT)
1232                                         log_err("fio: pid=%d, got signal=%d\n",
1233                                                         (int) td->pid, sig);
1234                                 td_set_runstate(td, TD_REAPED);
1235                                 goto reaped;
1236                         }
1237                         if (WIFEXITED(status)) {
1238                                 if (WEXITSTATUS(status) && !td->error)
1239                                         td->error = WEXITSTATUS(status);
1240
1241                                 td_set_runstate(td, TD_REAPED);
1242                                 goto reaped;
1243                         }
1244                 }
1245
1246                 /*
1247                  * thread is not dead, continue
1248                  */
1249                 pending++;
1250                 continue;
1251 reaped:
1252                 (*nr_running)--;
1253                 (*m_rate) -= (td->o.ratemin[0] + td->o.ratemin[1]);
1254                 (*t_rate) -= (td->o.rate[0] + td->o.rate[1]);
1255                 if (!td->pid)
1256                         pending--;
1257
1258                 if (td->error)
1259                         exit_value++;
1260
1261                 done_secs += mtime_since_now(&td->epoch) / 1000;
1262         }
1263
1264         if (*nr_running == cputhreads && !pending && realthreads)
1265                 terminate_threads(TERMINATE_ALL);
1266 }
1267
1268 static void *gtod_thread_main(void *data)
1269 {
1270         fio_mutex_up(startup_mutex);
1271
1272         /*
1273          * As long as we have jobs around, update the clock. It would be nice
1274          * to have some way of NOT hammering that CPU with gettimeofday(),
1275          * but I'm not sure what to use outside of a simple CPU nop to relax
1276          * it - we don't want to lose precision.
1277          */
1278         while (threads) {
1279                 fio_gtod_update();
1280                 nop;
1281         }
1282
1283         return NULL;
1284 }
1285
1286 static int fio_start_gtod_thread(void)
1287 {
1288         int ret;
1289
1290         ret = pthread_create(&gtod_thread, NULL, gtod_thread_main, NULL);
1291         if (ret) {
1292                 log_err("Can't create gtod thread: %s\n", strerror(ret));
1293                 return 1;
1294         }
1295
1296         ret = pthread_detach(gtod_thread);
1297         if (ret) {
1298                 log_err("Can't detatch gtod thread: %s\n", strerror(ret));
1299                 return 1;
1300         }
1301
1302         dprint(FD_MUTEX, "wait on startup_mutex\n");
1303         fio_mutex_down(startup_mutex);
1304         dprint(FD_MUTEX, "done waiting on startup_mutex\n");
1305         return 0;
1306 }
1307
1308 /*
1309  * Main function for kicking off and reaping jobs, as needed.
1310  */
1311 static void run_threads(void)
1312 {
1313         struct thread_data *td;
1314         unsigned long spent;
1315         int i, todo, nr_running, m_rate, t_rate, nr_started;
1316
1317         if (fio_pin_memory())
1318                 return;
1319
1320         if (fio_gtod_offload && fio_start_gtod_thread())
1321                 return;
1322
1323         if (!terse_output) {
1324                 printf("Starting ");
1325                 if (nr_thread)
1326                         printf("%d thread%s", nr_thread,
1327                                                 nr_thread > 1 ? "s" : "");
1328                 if (nr_process) {
1329                         if (nr_thread)
1330                                 printf(" and ");
1331                         printf("%d process%s", nr_process,
1332                                                 nr_process > 1 ? "es" : "");
1333                 }
1334                 printf("\n");
1335                 fflush(stdout);
1336         }
1337
1338         set_sig_handlers();
1339
1340         todo = thread_number;
1341         nr_running = 0;
1342         nr_started = 0;
1343         m_rate = t_rate = 0;
1344
1345         for_each_td(td, i) {
1346                 print_status_init(td->thread_number - 1);
1347
1348                 if (!td->o.create_serialize) {
1349                         init_disk_util(td);
1350                         continue;
1351                 }
1352
1353                 /*
1354                  * do file setup here so it happens sequentially,
1355                  * we don't want X number of threads getting their
1356                  * client data interspersed on disk
1357                  */
1358                 if (setup_files(td)) {
1359                         exit_value++;
1360                         if (td->error)
1361                                 log_err("fio: pid=%d, err=%d/%s\n",
1362                                         (int) td->pid, td->error, td->verror);
1363                         td_set_runstate(td, TD_REAPED);
1364                         todo--;
1365                 } else {
1366                         struct fio_file *f;
1367                         unsigned int i;
1368
1369                         /*
1370                          * for sharing to work, each job must always open
1371                          * its own files. so close them, if we opened them
1372                          * for creation
1373                          */
1374                         for_each_file(td, f, i) {
1375                                 if (fio_file_open(f))
1376                                         td_io_close_file(td, f);
1377                                 else
1378                                         assert(f->fd == -1);
1379                         }
1380                 }
1381
1382                 init_disk_util(td);
1383         }
1384
1385         set_genesis_time();
1386
1387         while (todo) {
1388                 struct thread_data *map[MAX_JOBS];
1389                 struct timeval this_start;
1390                 int this_jobs = 0, left;
1391
1392                 /*
1393                  * create threads (TD_NOT_CREATED -> TD_CREATED)
1394                  */
1395                 for_each_td(td, i) {
1396                         if (td->runstate != TD_NOT_CREATED)
1397                                 continue;
1398
1399                         /*
1400                          * never got a chance to start, killed by other
1401                          * thread for some reason
1402                          */
1403                         if (td->terminate) {
1404                                 todo--;
1405                                 continue;
1406                         }
1407
1408                         if (td->o.start_delay) {
1409                                 spent = mtime_since_genesis();
1410
1411                                 if (td->o.start_delay * 1000 > spent)
1412                                         continue;
1413                         }
1414
1415                         if (td->o.stonewall && (nr_started || nr_running)) {
1416                                 dprint(FD_PROCESS, "%s: stonewall wait\n",
1417                                                         td->o.name);
1418                                 break;
1419                         }
1420
1421                         /*
1422                          * Set state to created. Thread will transition
1423                          * to TD_INITIALIZED when it's done setting up.
1424                          */
1425                         td_set_runstate(td, TD_CREATED);
1426                         map[this_jobs++] = td;
1427                         nr_started++;
1428
1429                         if (td->o.use_thread) {
1430                                 int ret;
1431
1432                                 dprint(FD_PROCESS, "will pthread_create\n");
1433                                 ret = pthread_create(&td->thread, NULL,
1434                                                         thread_main, td);
1435                                 if (ret) {
1436                                         log_err("pthread_create: %s\n",
1437                                                         strerror(ret));
1438                                         nr_started--;
1439                                         break;
1440                                 }
1441                                 ret = pthread_detach(td->thread);
1442                                 if (ret)
1443                                         log_err("pthread_detach: %s",
1444                                                         strerror(ret));
1445                         } else {
1446                                 pid_t pid;
1447                                 dprint(FD_PROCESS, "will fork\n");
1448                                 pid = fork();
1449                                 if (!pid) {
1450                                         int ret = fork_main(shm_id, i);
1451
1452                                         _exit(ret);
1453                                 } else if (i == fio_debug_jobno)
1454                                         *fio_debug_jobp = pid;
1455                         }
1456                         dprint(FD_MUTEX, "wait on startup_mutex\n");
1457                         fio_mutex_down(startup_mutex);
1458                         dprint(FD_MUTEX, "done waiting on startup_mutex\n");
1459                 }
1460
1461                 /*
1462                  * Wait for the started threads to transition to
1463                  * TD_INITIALIZED.
1464                  */
1465                 fio_gettime(&this_start, NULL);
1466                 left = this_jobs;
1467                 while (left && !fio_abort) {
1468                         if (mtime_since_now(&this_start) > JOB_START_TIMEOUT)
1469                                 break;
1470
1471                         usleep(100000);
1472
1473                         for (i = 0; i < this_jobs; i++) {
1474                                 td = map[i];
1475                                 if (!td)
1476                                         continue;
1477                                 if (td->runstate == TD_INITIALIZED) {
1478                                         map[i] = NULL;
1479                                         left--;
1480                                 } else if (td->runstate >= TD_EXITED) {
1481                                         map[i] = NULL;
1482                                         left--;
1483                                         todo--;
1484                                         nr_running++; /* work-around... */
1485                                 }
1486                         }
1487                 }
1488
1489                 if (left) {
1490                         log_err("fio: %d jobs failed to start\n", left);
1491                         for (i = 0; i < this_jobs; i++) {
1492                                 td = map[i];
1493                                 if (!td)
1494                                         continue;
1495                                 kill(td->pid, SIGTERM);
1496                         }
1497                         break;
1498                 }
1499
1500                 /*
1501                  * start created threads (TD_INITIALIZED -> TD_RUNNING).
1502                  */
1503                 for_each_td(td, i) {
1504                         if (td->runstate != TD_INITIALIZED)
1505                                 continue;
1506
1507                         if (in_ramp_time(td))
1508                                 td_set_runstate(td, TD_RAMP);
1509                         else
1510                                 td_set_runstate(td, TD_RUNNING);
1511                         nr_running++;
1512                         nr_started--;
1513                         m_rate += td->o.ratemin[0] + td->o.ratemin[1];
1514                         t_rate += td->o.rate[0] + td->o.rate[1];
1515                         todo--;
1516                         fio_mutex_up(td->mutex);
1517                 }
1518
1519                 reap_threads(&nr_running, &t_rate, &m_rate);
1520
1521                 if (todo)
1522                         usleep(100000);
1523         }
1524
1525         while (nr_running) {
1526                 reap_threads(&nr_running, &t_rate, &m_rate);
1527                 usleep(10000);
1528         }
1529
1530         update_io_ticks();
1531         fio_unpin_memory();
1532 }
1533
1534 int main(int argc, char *argv[])
1535 {
1536         long ps;
1537
1538         sinit();
1539
1540         /*
1541          * We need locale for number printing, if it isn't set then just
1542          * go with the US format.
1543          */
1544         if (!getenv("LC_NUMERIC"))
1545                 setlocale(LC_NUMERIC, "en_US");
1546
1547         if (parse_options(argc, argv))
1548                 return 1;
1549
1550         if (!thread_number)
1551                 return 0;
1552
1553         ps = sysconf(_SC_PAGESIZE);
1554         if (ps < 0) {
1555                 log_err("Failed to get page size\n");
1556                 return 1;
1557         }
1558
1559         page_size = ps;
1560         page_mask = ps - 1;
1561
1562         if (write_bw_log) {
1563                 setup_log(&agg_io_log[DDIR_READ]);
1564                 setup_log(&agg_io_log[DDIR_WRITE]);
1565         }
1566
1567         startup_mutex = fio_mutex_init(0);
1568         writeout_mutex = fio_mutex_init(1);
1569
1570         set_genesis_time();
1571
1572         status_timer_arm();
1573
1574         run_threads();
1575
1576         if (!fio_abort) {
1577                 show_run_stats();
1578                 if (write_bw_log) {
1579                         __finish_log(agg_io_log[DDIR_READ], "agg-read_bw.log");
1580                         __finish_log(agg_io_log[DDIR_WRITE],
1581                                         "agg-write_bw.log");
1582                 }
1583         }
1584
1585         fio_mutex_remove(startup_mutex);
1586         fio_mutex_remove(writeout_mutex);
1587         return exit_value;
1588 }