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