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