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