fio.c

   1 /*
   2  * fio - the flexible io tester
   3  *
   4  * Copyright (C) 2005 Jens Axboe <axboe@suse.de>
   5  * Copyright (C) 2006 Jens Axboe <axboe@kernel.dk>
   6  *
   7  * The license below covers all files distributed with fio unless otherwise
   8  * noted in the file itself.
   9  *
  10  *  This program is free software; you can redistribute it and/or modify
  11  *  it under the terms of the GNU General Public License version 2 as
  12  *  published by the Free Software Foundation.
  13  *
  14  *  This program is distributed in the hope that it will be useful,
  15  *  but WITHOUT ANY WARRANTY; without even the implied warranty of
  16  *  MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
  17  *  GNU General Public License for more details.
  18  *
  19  *  You should have received a copy of the GNU General Public License
  20  *  along with this program; if not, write to the Free Software
  21  *  Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA  02111-1307  USA
  22  *
  23  */
  24 #include <unistd.h>
  25 #include <fcntl.h>
  26 #include <string.h>
  27 #include <signal.h>
  28 #include <time.h>
  29 #include <locale.h>
  30 #include <assert.h>
  31 #include <sys/stat.h>
  32 #include <sys/wait.h>
  33 #include <sys/ipc.h>
  34 #include <sys/shm.h>
  35 #include <sys/mman.h>
  36
  37 #include "fio.h"
  38 #include "hash.h"
  39 #include "smalloc.h"
  40 #include "verify.h"
  41 #include "diskutil.h"
  42
  43 unsigned long page_mask;
  44 unsigned long page_size;
  45 #define ALIGN(buf)      \
  46         (char *) (((unsigned long) (buf) + page_mask) & ~page_mask)
  47
  48 int groupid = 0;
  49 int thread_number = 0;
  50 int nr_process = 0;
  51 int nr_thread = 0;
  52 int shm_id = 0;
  53 int temp_stall_ts;
  54 unsigned long done_secs = 0;
  55
  56 static struct fio_mutex *startup_mutex;
  57 static struct fio_mutex *writeout_mutex;
  58 static volatile int fio_abort;
  59 static int exit_value;
  60 static struct itimerval itimer;
  61 static pthread_t gtod_thread;
  62
  63 struct io_log *agg_io_log[2];
  64
  65 #define TERMINATE_ALL           (-1)
  66 #define JOB_START_TIMEOUT       (5 * 1000)
  67
  68 void td_set_runstate(struct thread_data *td, int runstate)
  69 {
  70         if (td->runstate == runstate)
  71                 return;
  72
  73         dprint(FD_PROCESS, "pid=%d: runstate %d -> %d\n", (int) td->pid,
  74                                                 td->runstate, runstate);
  75         td->runstate = runstate;
  76 }
  77
  78 static void terminate_threads(int group_id)
  79 {
  80         struct thread_data *td;
  81         int i;
  82
  83         dprint(FD_PROCESS, "terminate group_id=%d\n", group_id);
  84
  85         for_each_td(td, i) {
  86                 if (group_id == TERMINATE_ALL || groupid == td->groupid) {
  87                         dprint(FD_PROCESS, "setting terminate on %s/%d\n",
  88                                                 td->o.name, (int) td->pid);
  89                         td->terminate = 1;
  90                         td->o.start_delay = 0;
  91
  92                         /*
  93                          * if the thread is running, just let it exit
  94                          */
  95                         if (td->runstate < TD_RUNNING)
  96                                 kill(td->pid, SIGQUIT);
  97                         else {
  98                                 struct ioengine_ops *ops = td->io_ops;
  99
 100                                 if (ops && (ops->flags & FIO_SIGQUIT))
 101                                         kill(td->pid, SIGQUIT);
 102                         }
 103                 }
 104         }
 105 }
 106
 107 static void status_timer_arm(void)
 108 {
 109         itimer.it_value.tv_sec = 0;
 110         itimer.it_value.tv_usec = DISK_UTIL_MSEC * 1000;
 111         setitimer(ITIMER_REAL, &itimer, NULL);
 112 }
 113
 114 static void sig_alrm(int fio_unused sig)
 115 {
 116         if (threads) {
 117                 update_io_ticks();
 118                 print_thread_status();
 119                 status_timer_arm();
 120         }
 121 }
 122
 123 /*
 124  * Happens on thread runs with ctrl-c, ignore our own SIGQUIT
 125  */
 126 static void sig_quit(int sig)
 127 {
 128 }
 129
 130 static void sig_int(int sig)
 131 {
 132         if (threads) {
 133                 printf("\nfio: terminating on signal %d\n", sig);
 134                 fflush(stdout);
 135                 terminate_threads(TERMINATE_ALL);
 136         }
 137 }
 138
 139 static void sig_ill(int fio_unused sig)
 140 {
 141         if (!threads)
 142                 return;
 143
 144         log_err("fio: illegal instruction. your cpu does not support "
 145                 "the sse4.2 instruction for crc32c\n");
 146         terminate_threads(TERMINATE_ALL);
 147         exit(4);
 148 }
 149
 150 static void set_sig_handlers(void)
 151 {
 152         struct sigaction act;
 153
 154         memset(&act, 0, sizeof(act));
 155         act.sa_handler = sig_alrm;
 156         act.sa_flags = SA_RESTART;
 157         sigaction(SIGALRM, &act, NULL);
 158
 159         memset(&act, 0, sizeof(act));
 160         act.sa_handler = sig_int;
 161         act.sa_flags = SA_RESTART;
 162         sigaction(SIGINT, &act, NULL);
 163
 164         memset(&act, 0, sizeof(act));
 165         act.sa_handler = sig_ill;
 166         act.sa_flags = SA_RESTART;
 167         sigaction(SIGILL, &act, NULL);
 168
 169         memset(&act, 0, sizeof(act));
 170         act.sa_handler = sig_quit;
 171         act.sa_flags = SA_RESTART;
 172         sigaction(SIGQUIT, &act, NULL);
 173 }
 174
 175 /*
 176  * Check if we are above the minimum rate given.
 177  */
 178 static int __check_min_rate(struct thread_data *td, struct timeval *now,
 179                             enum td_ddir ddir)
 180 {
 181         unsigned long long bytes = 0;
 182         unsigned long iops = 0;
 183         unsigned long spent;
 184         unsigned long rate;
 185         unsigned int ratemin = 0;
 186         unsigned int rate_iops = 0;
 187         unsigned int rate_iops_min = 0;
 188
 189         if (!td->o.ratemin[ddir] && !td->o.rate_iops_min[ddir])
 190                 return 0;
 191
 192         /*
 193          * allow a 2 second settle period in the beginning
 194          */
 195         if (mtime_since(&td->start, now) < 2000)
 196                 return 0;
 197
 198         iops += td->io_blocks[ddir];
 199         bytes += td->this_io_bytes[ddir];
 200         ratemin += td->o.ratemin[ddir];
 201         rate_iops += td->o.rate_iops[ddir];
 202         rate_iops_min += td->o.rate_iops_min[ddir];
 203
 204         /*
 205          * if rate blocks is set, sample is running
 206          */
 207         if (td->rate_bytes[ddir] || td->rate_blocks[ddir]) {
 208                 spent = mtime_since(&td->lastrate[ddir], now);
 209                 if (spent < td->o.ratecycle)
 210                         return 0;
 211
 212                 if (td->o.rate[ddir]) {
 213                         /*
 214                          * check bandwidth specified rate
 215                          */
 216                         if (bytes < td->rate_bytes[ddir]) {
 217                                 log_err("%s: min rate %u not met\n", td->o.name,
 218                                                                 ratemin);
 219                                 return 1;
 220                         } else {
 221                                 rate = ((bytes - td->rate_bytes[ddir]) * 1000) / spent;
 222                                 if (rate < ratemin ||
 223                                     bytes < td->rate_bytes[ddir]) {
 224                                         log_err("%s: min rate %u not met, got"
 225                                                 " %luKiB/sec\n", td->o.name,
 226                                                         ratemin, rate);
 227                                         return 1;
 228                                 }
 229                         }
 230                 } else {
 231                         /*
 232                          * checks iops specified rate
 233                          */
 234                         if (iops < rate_iops) {
 235                                 log_err("%s: min iops rate %u not met\n",
 236                                                 td->o.name, rate_iops);
 237                                 return 1;
 238                         } else {
 239                                 rate = ((iops - td->rate_blocks[ddir]) * 1000) / spent;
 240                                 if (rate < rate_iops_min ||
 241                                     iops < td->rate_blocks[ddir]) {
 242                                         log_err("%s: min iops rate %u not met,"
 243                                                 " got %lu\n", td->o.name,
 244                                                         rate_iops_min, rate);
 245                                 }
 246                         }
 247                 }
 248         }
 249
 250         td->rate_bytes[ddir] = bytes;
 251         td->rate_blocks[ddir] = iops;
 252         memcpy(&td->lastrate[ddir], now, sizeof(*now));
 253         return 0;
 254 }
 255
 256 static int check_min_rate(struct thread_data *td, struct timeval *now,
 257                           unsigned long *bytes_done)
 258 {
 259         int ret = 0;
 260
 261         if (bytes_done[0])
 262                 ret |= __check_min_rate(td, now, 0);
 263         if (bytes_done[1])
 264                 ret |= __check_min_rate(td, now, 1);
 265
 266         return ret;
 267 }
 268
 269 static inline int runtime_exceeded(struct thread_data *td, struct timeval *t)
 270 {
 271         if (!td->o.timeout)
 272                 return 0;
 273         if (mtime_since(&td->epoch, t) >= td->o.timeout * 1000)
 274                 return 1;
 275
 276         return 0;
 277 }
 278
 279 /*
 280  * When job exits, we can cancel the in-flight IO if we are using async
 281  * io. Attempt to do so.
 282  */
 283 static void cleanup_pending_aio(struct thread_data *td)
 284 {
 285         struct flist_head *entry, *n;
 286         struct io_u *io_u;
 287         int r;
 288
 289         /*
 290          * get immediately available events, if any
 291          */
 292         r = io_u_queued_complete(td, 0, NULL);
 293         if (r < 0)
 294                 return;
 295
 296         /*
 297          * now cancel remaining active events
 298          */
 299         if (td->io_ops->cancel) {
 300                 flist_for_each_safe(entry, n, &td->io_u_busylist) {
 301                         io_u = flist_entry(entry, struct io_u, list);
 302
 303                         /*
 304                          * if the io_u isn't in flight, then that generally
 305                          * means someone leaked an io_u. complain but fix
 306                          * it up, so we don't stall here.
 307                          */
 308                         if ((io_u->flags & IO_U_F_FLIGHT) == 0) {
 309                                 log_err("fio: non-busy IO on busy list\n");
 310                                 put_io_u(td, io_u);
 311                         } else {
 312                                 r = td->io_ops->cancel(td, io_u);
 313                                 if (!r)
 314                                         put_io_u(td, io_u);
 315                         }
 316                 }
 317         }
 318
 319         if (td->cur_depth)
 320                 r = io_u_queued_complete(td, td->cur_depth, NULL);
 321 }
 322
 323 /*
 324  * Helper to handle the final sync of a file. Works just like the normal
 325  * io path, just does everything sync.
 326  */
 327 static int fio_io_sync(struct thread_data *td, struct fio_file *f)
 328 {
 329         struct io_u *io_u = __get_io_u(td);
 330         int ret;
 331
 332         if (!io_u)
 333                 return 1;
 334
 335         io_u->ddir = DDIR_SYNC;
 336         io_u->file = f;
 337
 338         if (td_io_prep(td, io_u)) {
 339                 put_io_u(td, io_u);
 340                 return 1;
 341         }
 342
 343 requeue:
 344         ret = td_io_queue(td, io_u);
 345         if (ret < 0) {
 346                 td_verror(td, io_u->error, "td_io_queue");
 347                 put_io_u(td, io_u);
 348                 return 1;
 349         } else if (ret == FIO_Q_QUEUED) {
 350                 if (io_u_queued_complete(td, 1, NULL) < 0)
 351                         return 1;
 352         } else if (ret == FIO_Q_COMPLETED) {
 353                 if (io_u->error) {
 354                         td_verror(td, io_u->error, "td_io_queue");
 355                         return 1;
 356                 }
 357
 358                 if (io_u_sync_complete(td, io_u, NULL) < 0)
 359                         return 1;
 360         } else if (ret == FIO_Q_BUSY) {
 361                 if (td_io_commit(td))
 362                         return 1;
 363                 goto requeue;
 364         }
 365
 366         return 0;
 367 }
 368
 369 static inline void update_tv_cache(struct thread_data *td)
 370 {
 371         if ((++td->tv_cache_nr & td->tv_cache_mask) == td->tv_cache_mask)
 372                 fio_gettime(&td->tv_cache, NULL);
 373 }
 374
 375 static int break_on_this_error(struct thread_data *td, int *retptr)
 376 {
 377         int ret = *retptr;
 378
 379         if (ret < 0 || td->error) {
 380                 int err;
 381
 382                 if (!td->o.continue_on_error)
 383                         return 1;
 384
 385                 if (ret < 0)
 386                         err = -ret;
 387                 else
 388                         err = td->error;
 389
 390                 update_error_count(td, err);
 391
 392                 if (td_non_fatal_error(err)) {
 393                         /*
 394                          * Continue with the I/Os in case of
 395                          * a non fatal error.
 396                          */
 397                         td_clear_error(td);
 398                         *retptr = 0;
 399                         return 0;
 400                 } else {
 401                         /*
 402                          * Stop the I/O in case of a fatal
 403                          * error.
 404                          */
 405                         return 1;
 406                 }
 407         }
 408
 409         return 0;
 410 }
 411
 412 /*
 413  * The main verify engine. Runs over the writes we previously submitted,
 414  * reads the blocks back in, and checks the crc/md5 of the data.
 415  */
 416 static void do_verify(struct thread_data *td)
 417 {
 418         struct fio_file *f;
 419         struct io_u *io_u;
 420         int ret, min_events;
 421         unsigned int i;
 422
 423         /*
 424          * sync io first and invalidate cache, to make sure we really
 425          * read from disk.
 426          */
 427         for_each_file(td, f, i) {
 428                 if (!fio_file_open(f))
 429                         continue;
 430                 if (fio_io_sync(td, f))
 431                         break;
 432                 if (file_invalidate_cache(td, f))
 433                         break;
 434         }
 435
 436         if (td->error)
 437                 return;
 438
 439         td_set_runstate(td, TD_VERIFYING);
 440
 441         io_u = NULL;
 442         while (!td->terminate) {
 443                 int ret2, full;
 444
 445                 io_u = __get_io_u(td);
 446                 if (!io_u)
 447                         break;
 448
 449                 update_tv_cache(td);
 450
 451                 if (runtime_exceeded(td, &td->tv_cache)) {
 452                         put_io_u(td, io_u);
 453                         td->terminate = 1;
 454                         break;
 455                 }
 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                 io_u = get_io_u(td);
 584                 if (!io_u)
 585                         break;
 586
 587                 update_tv_cache(td);
 588
 589                 if (runtime_exceeded(td, &td->tv_cache)) {
 590                         put_io_u(td, io_u);
 591                         td->terminate = 1;
 592                         break;
 593                 }
 594
 595                 /*
 596                  * Add verification end_io handler, if asked to verify
 597                  * a previously written file.
 598                  */
 599                 if (td->o.verify != VERIFY_NONE && io_u->ddir == DDIR_READ) {
 600                         io_u->end_io = verify_io_u;
 601                         td_set_runstate(td, TD_VERIFYING);
 602                 } else if (in_ramp_time(td))
 603                         td_set_runstate(td, TD_RAMP);
 604                 else
 605                         td_set_runstate(td, TD_RUNNING);
 606
 607                 ret = td_io_queue(td, io_u);
 608                 switch (ret) {
 609                 case FIO_Q_COMPLETED:
 610                         if (io_u->error) {
 611                                 ret = -io_u->error;
 612                                 clear_io_u(td, io_u);
 613                         } else if (io_u->resid) {
 614                                 int bytes = io_u->xfer_buflen - io_u->resid;
 615                                 struct fio_file *f = io_u->file;
 616
 617                                 /*
 618                                  * zero read, fail
 619                                  */
 620                                 if (!bytes) {
 621                                         td_verror(td, EIO, "full resid");
 622                                         put_io_u(td, io_u);
 623                                         break;
 624                                 }
 625
 626                                 io_u->xfer_buflen = io_u->resid;
 627                                 io_u->xfer_buf += bytes;
 628                                 io_u->offset += bytes;
 629
 630                                 td->ts.short_io_u[io_u->ddir]++;
 631
 632                                 if (io_u->offset == f->real_file_size)
 633                                         goto sync_done;
 634
 635                                 requeue_io_u(td, &io_u);
 636                         } else {
 637 sync_done:
 638                                 if (__should_check_rate(td, 0) ||
 639                                     __should_check_rate(td, 1))
 640                                         fio_gettime(&comp_time, NULL);
 641
 642                                 ret = io_u_sync_complete(td, io_u, bytes_done);
 643                                 if (ret < 0)
 644                                         break;
 645                         }
 646                         break;
 647                 case FIO_Q_QUEUED:
 648                         /*
 649                          * if the engine doesn't have a commit hook,
 650                          * the io_u is really queued. if it does have such
 651                          * a hook, it has to call io_u_queued() itself.
 652                          */
 653                         if (td->io_ops->commit == NULL)
 654                                 io_u_queued(td, io_u);
 655                         break;
 656                 case FIO_Q_BUSY:
 657                         requeue_io_u(td, &io_u);
 658                         ret2 = td_io_commit(td);
 659                         if (ret2 < 0)
 660                                 ret = ret2;
 661                         break;
 662                 default:
 663                         assert(ret < 0);
 664                         put_io_u(td, io_u);
 665                         break;
 666                 }
 667
 668                 if (break_on_this_error(td, &ret))
 669                         break;
 670
 671                 /*
 672                  * See if we need to complete some commands
 673                  */
 674                 full = queue_full(td) || ret == FIO_Q_BUSY;
 675                 if (full || !td->o.iodepth_batch_complete) {
 676                         min_evts = td->o.iodepth_batch_complete;
 677                         if (full && !min_evts)
 678                                 min_evts = 1;
 679
 680                         if (__should_check_rate(td, 0) ||
 681                             __should_check_rate(td, 1))
 682                                 fio_gettime(&comp_time, NULL);
 683
 684                         do {
 685                                 ret = io_u_queued_complete(td, min_evts, bytes_done);
 686                                 if (ret < 0)
 687                                         break;
 688
 689                         } while (full && (td->cur_depth > td->o.iodepth_low));
 690                 }
 691
 692                 if (ret < 0)
 693                         break;
 694                 if (!(bytes_done[0] + bytes_done[1]))
 695                         continue;
 696
 697                 if (!in_ramp_time(td) && should_check_rate(td, bytes_done)) {
 698                         if (check_min_rate(td, &comp_time, bytes_done)) {
 699                                 if (exitall_on_terminate)
 700                                         terminate_threads(td->groupid);
 701                                 td_verror(td, EIO, "check_min_rate");
 702                                 break;
 703                         }
 704                 }
 705
 706                 if (td->o.thinktime) {
 707                         unsigned long long b;
 708
 709                         b = td->io_blocks[0] + td->io_blocks[1];
 710                         if (!(b % td->o.thinktime_blocks)) {
 711                                 int left;
 712
 713                                 if (td->o.thinktime_spin)
 714                                         usec_spin(td->o.thinktime_spin);
 715
 716                                 left = td->o.thinktime - td->o.thinktime_spin;
 717                                 if (left)
 718                                         usec_sleep(td, left);
 719                         }
 720                 }
 721         }
 722
 723         if (td->o.fill_device && td->error == ENOSPC) {
 724                 td->error = 0;
 725                 td->terminate = 1;
 726         }
 727         if (!td->error) {
 728                 struct fio_file *f;
 729
 730                 i = td->cur_depth;
 731                 if (i)
 732                         ret = io_u_queued_complete(td, i, NULL);
 733
 734                 if (should_fsync(td) && td->o.end_fsync) {
 735                         td_set_runstate(td, TD_FSYNCING);
 736
 737                         for_each_file(td, f, i) {
 738                                 if (!fio_file_open(f))
 739                                         continue;
 740                                 fio_io_sync(td, f);
 741                         }
 742                 }
 743         } else
 744                 cleanup_pending_aio(td);
 745
 746         /*
 747          * stop job if we failed doing any IO
 748          */
 749         if ((td->this_io_bytes[0] + td->this_io_bytes[1]) == 0)
 750                 td->done = 1;
 751 }
 752
 753 static void cleanup_io_u(struct thread_data *td)
 754 {
 755         struct flist_head *entry, *n;
 756         struct io_u *io_u;
 757
 758         flist_for_each_safe(entry, n, &td->io_u_freelist) {
 759                 io_u = flist_entry(entry, struct io_u, list);
 760
 761                 flist_del(&io_u->list);
 762                 free(io_u);
 763         }
 764
 765         free_io_mem(td);
 766 }
 767
 768 static int init_io_u(struct thread_data *td)
 769 {
 770         struct io_u *io_u;
 771         unsigned int max_bs;
 772         int cl_align, i, max_units;
 773         char *p;
 774
 775         max_units = td->o.iodepth;
 776         max_bs = max(td->o.max_bs[DDIR_READ], td->o.max_bs[DDIR_WRITE]);
 777         td->orig_buffer_size = (unsigned long long) max_bs
 778                                         * (unsigned long long) max_units;
 779
 780         if (td->o.mem_type == MEM_SHMHUGE || td->o.mem_type == MEM_MMAPHUGE) {
 781                 unsigned long bs;
 782
 783                 bs = td->orig_buffer_size + td->o.hugepage_size - 1;
 784                 td->orig_buffer_size = bs & ~(td->o.hugepage_size - 1);
 785         }
 786
 787         if (td->orig_buffer_size != (size_t) td->orig_buffer_size) {
 788                 log_err("fio: IO memory too large. Reduce max_bs or iodepth\n");
 789                 return 1;
 790         }
 791
 792         if (allocate_io_mem(td))
 793                 return 1;
 794
 795         if (td->o.odirect)
 796                 p = ALIGN(td->orig_buffer);
 797         else
 798                 p = td->orig_buffer;
 799
 800         cl_align = os_cache_line_size();
 801
 802         for (i = 0; i < max_units; i++) {
 803                 void *ptr;
 804
 805                 if (td->terminate)
 806                         return 1;
 807
 808                 if (posix_memalign(&ptr, cl_align, sizeof(*io_u))) {
 809                         log_err("fio: posix_memalign=%s\n", strerror(errno));
 810                         break;
 811                 }
 812
 813                 io_u = ptr;
 814                 memset(io_u, 0, sizeof(*io_u));
 815                 INIT_FLIST_HEAD(&io_u->list);
 816
 817                 if (!(td->io_ops->flags & FIO_NOIO)) {
 818                         io_u->buf = p + max_bs * i;
 819
 820                         if (td_write(td) && !td->o.refill_buffers)
 821                                 io_u_fill_buffer(td, io_u, max_bs);
 822                 }
 823
 824                 io_u->index = i;
 825                 io_u->flags = IO_U_F_FREE;
 826                 flist_add(&io_u->list, &td->io_u_freelist);
 827         }
 828
 829         return 0;
 830 }
 831
 832 static int switch_ioscheduler(struct thread_data *td)
 833 {
 834         char tmp[256], tmp2[128];
 835         FILE *f;
 836         int ret;
 837
 838         if (td->io_ops->flags & FIO_DISKLESSIO)
 839                 return 0;
 840
 841         sprintf(tmp, "%s/queue/scheduler", td->sysfs_root);
 842
 843         f = fopen(tmp, "r+");
 844         if (!f) {
 845                 if (errno == ENOENT) {
 846                         log_err("fio: os or kernel doesn't support IO scheduler"
 847                                 " switching\n");
 848                         return 0;
 849                 }
 850                 td_verror(td, errno, "fopen iosched");
 851                 return 1;
 852         }
 853
 854         /*
 855          * Set io scheduler.
 856          */
 857         ret = fwrite(td->o.ioscheduler, strlen(td->o.ioscheduler), 1, f);
 858         if (ferror(f) || ret != 1) {
 859                 td_verror(td, errno, "fwrite");
 860                 fclose(f);
 861                 return 1;
 862         }
 863
 864         rewind(f);
 865
 866         /*
 867          * Read back and check that the selected scheduler is now the default.
 868          */
 869         ret = fread(tmp, 1, sizeof(tmp), f);
 870         if (ferror(f) || ret < 0) {
 871                 td_verror(td, errno, "fread");
 872                 fclose(f);
 873                 return 1;
 874         }
 875
 876         sprintf(tmp2, "[%s]", td->o.ioscheduler);
 877         if (!strstr(tmp, tmp2)) {
 878                 log_err("fio: io scheduler %s not found\n", td->o.ioscheduler);
 879                 td_verror(td, EINVAL, "iosched_switch");
 880                 fclose(f);
 881                 return 1;
 882         }
 883
 884         fclose(f);
 885         return 0;
 886 }
 887
 888 static int keep_running(struct thread_data *td)
 889 {
 890         unsigned long long io_done;
 891
 892         if (td->done)
 893                 return 0;
 894         if (td->o.time_based)
 895                 return 1;
 896         if (td->o.loops) {
 897                 td->o.loops--;
 898                 return 1;
 899         }
 900
 901         io_done = td->io_bytes[DDIR_READ] + td->io_bytes[DDIR_WRITE]
 902                         + td->io_skip_bytes;
 903         if (io_done < td->o.size)
 904                 return 1;
 905
 906         return 0;
 907 }
 908
 909 static void reset_io_counters(struct thread_data *td)
 910 {
 911         td->ts.stat_io_bytes[0] = td->ts.stat_io_bytes[1] = 0;
 912         td->this_io_bytes[0] = td->this_io_bytes[1] = 0;
 913         td->zone_bytes = 0;
 914         td->rate_bytes[0] = td->rate_bytes[1] = 0;
 915         td->rate_blocks[0] = td->rate_blocks[1] = 0;
 916
 917         td->last_was_sync = 0;
 918
 919         /*
 920          * reset file done count if we are to start over
 921          */
 922         if (td->o.time_based || td->o.loops)
 923                 td->nr_done_files = 0;
 924
 925         /*
 926          * Set the same seed to get repeatable runs
 927          */
 928         td_fill_rand_seeds(td);
 929 }
 930
 931 void reset_all_stats(struct thread_data *td)
 932 {
 933         struct timeval tv;
 934         int i;
 935
 936         reset_io_counters(td);
 937
 938         for (i = 0; i < 2; i++) {
 939                 td->io_bytes[i] = 0;
 940                 td->io_blocks[i] = 0;
 941                 td->io_issues[i] = 0;
 942                 td->ts.total_io_u[i] = 0;
 943         }
 944
 945         fio_gettime(&tv, NULL);
 946         memcpy(&td->epoch, &tv, sizeof(tv));
 947         memcpy(&td->start, &tv, sizeof(tv));
 948 }
 949
 950 static void clear_io_state(struct thread_data *td)
 951 {
 952         struct fio_file *f;
 953         unsigned int i;
 954
 955         reset_io_counters(td);
 956
 957         close_files(td);
 958         for_each_file(td, f, i)
 959                 fio_file_clear_done(f);
 960 }
 961
 962 static int exec_string(const char *string)
 963 {
 964         int ret, newlen = strlen(string) + 1 + 8;
 965         char *str;
 966
 967         str = malloc(newlen);
 968         sprintf(str, "sh -c %s", string);
 969
 970         ret = system(str);
 971         if (ret == -1)
 972                 log_err("fio: exec of cmd <%s> failed\n", str);
 973
 974         free(str);
 975         return ret;
 976 }
 977
 978 /*
 979  * Entry point for the thread based jobs. The process based jobs end up
 980  * here as well, after a little setup.
 981  */
 982 static void *thread_main(void *data)
 983 {
 984         unsigned long long runtime[2], elapsed;
 985         struct thread_data *td = data;
 986         int clear_state;
 987
 988         if (!td->o.use_thread)
 989                 setsid();
 990
 991         td->pid = getpid();
 992
 993         dprint(FD_PROCESS, "jobs pid=%d started\n", (int) td->pid);
 994
 995         INIT_FLIST_HEAD(&td->io_u_freelist);
 996         INIT_FLIST_HEAD(&td->io_u_busylist);
 997         INIT_FLIST_HEAD(&td->io_u_requeues);
 998         INIT_FLIST_HEAD(&td->io_log_list);
 999         INIT_FLIST_HEAD(&td->io_hist_list);
1000         td->io_hist_tree = RB_ROOT;
1001
1002         td_set_runstate(td, TD_INITIALIZED);
1003         dprint(FD_MUTEX, "up startup_mutex\n");
1004         fio_mutex_up(startup_mutex);
1005         dprint(FD_MUTEX, "wait on td->mutex\n");
1006         fio_mutex_down(td->mutex);
1007         dprint(FD_MUTEX, "done waiting on td->mutex\n");
1008
1009         /*
1010          * the ->mutex mutex is now no longer used, close it to avoid
1011          * eating a file descriptor
1012          */
1013         fio_mutex_remove(td->mutex);
1014
1015         /*
1016          * May alter parameters that init_io_u() will use, so we need to
1017          * do this first.
1018          */
1019         if (init_iolog(td))
1020                 goto err;
1021
1022         if (init_io_u(td))
1023                 goto err;
1024
1025         if (td->o.cpumask_set && fio_setaffinity(td) == -1) {
1026                 td_verror(td, errno, "cpu_set_affinity");
1027                 goto err;
1028         }
1029
1030         /*
1031          * If we have a gettimeofday() thread, make sure we exclude that
1032          * thread from this job
1033          */
1034         if (td->o.gtod_cpu) {
1035                 fio_cpu_clear(&td->o.cpumask, td->o.gtod_cpu);
1036                 if (fio_setaffinity(td) == -1) {
1037                         td_verror(td, errno, "cpu_set_affinity");
1038                         goto err;
1039                 }
1040         }
1041
1042         if (td->ioprio_set) {
1043                 if (ioprio_set(IOPRIO_WHO_PROCESS, 0, td->ioprio) == -1) {
1044                         td_verror(td, errno, "ioprio_set");
1045                         goto err;
1046                 }
1047         }
1048
1049         if (nice(td->o.nice) == -1) {
1050                 td_verror(td, errno, "nice");
1051                 goto err;
1052         }
1053
1054         if (td->o.ioscheduler && switch_ioscheduler(td))
1055                 goto err;
1056
1057         if (!td->o.create_serialize && setup_files(td))
1058                 goto err;
1059
1060         if (td_io_init(td))
1061                 goto err;
1062
1063         if (init_random_map(td))
1064                 goto err;
1065
1066         if (td->o.exec_prerun) {
1067                 if (exec_string(td->o.exec_prerun))
1068                         goto err;
1069         }
1070
1071         if (td->o.pre_read) {
1072                 if (pre_read_files(td) < 0)
1073                         goto err;
1074         }
1075
1076         fio_gettime(&td->epoch, NULL);
1077         getrusage(RUSAGE_SELF, &td->ts.ru_start);
1078
1079         runtime[0] = runtime[1] = 0;
1080         clear_state = 0;
1081         while (keep_running(td)) {
1082                 fio_gettime(&td->start, NULL);
1083                 memcpy(&td->ts.stat_sample_time, &td->start, sizeof(td->start));
1084                 memcpy(&td->tv_cache, &td->start, sizeof(td->start));
1085
1086                 if (td->o.ratemin[0] || td->o.ratemin[1])
1087                         memcpy(&td->lastrate, &td->ts.stat_sample_time,
1088                                                         sizeof(td->lastrate));
1089
1090                 if (clear_state)
1091                         clear_io_state(td);
1092
1093                 prune_io_piece_log(td);
1094
1095                 do_io(td);
1096
1097                 clear_state = 1;
1098
1099                 if (td_read(td) && td->io_bytes[DDIR_READ]) {
1100                         elapsed = utime_since_now(&td->start);
1101                         runtime[DDIR_READ] += elapsed;
1102                 }
1103                 if (td_write(td) && td->io_bytes[DDIR_WRITE]) {
1104                         elapsed = utime_since_now(&td->start);
1105                         runtime[DDIR_WRITE] += elapsed;
1106                 }
1107
1108                 if (td->error || td->terminate)
1109                         break;
1110
1111                 if (!td->o.do_verify ||
1112                     td->o.verify == VERIFY_NONE ||
1113                     (td->io_ops->flags & FIO_UNIDIR))
1114                         continue;
1115
1116                 clear_io_state(td);
1117
1118                 fio_gettime(&td->start, NULL);
1119
1120                 do_verify(td);
1121
1122                 runtime[DDIR_READ] += utime_since_now(&td->start);
1123
1124                 if (td->error || td->terminate)
1125                         break;
1126         }
1127
1128         update_rusage_stat(td);
1129         td->ts.runtime[0] = (runtime[0] + 999) / 1000;
1130         td->ts.runtime[1] = (runtime[1] + 999) / 1000;
1131         td->ts.total_run_time = mtime_since_now(&td->epoch);
1132         td->ts.io_bytes[0] = td->io_bytes[0];
1133         td->ts.io_bytes[1] = td->io_bytes[1];
1134
1135         fio_mutex_down(writeout_mutex);
1136         if (td->ts.bw_log) {
1137                 if (td->o.bw_log_file) {
1138                         finish_log_named(td, td->ts.bw_log,
1139                                                 td->o.bw_log_file, "bw");
1140                 } else
1141                         finish_log(td, td->ts.bw_log, "bw");
1142         }
1143         if (td->ts.slat_log) {
1144                 if (td->o.lat_log_file) {
1145                         finish_log_named(td, td->ts.slat_log,
1146                                                 td->o.lat_log_file, "slat");
1147                 } else
1148                         finish_log(td, td->ts.slat_log, "slat");
1149         }
1150         if (td->ts.clat_log) {
1151                 if (td->o.lat_log_file) {
1152                         finish_log_named(td, td->ts.clat_log,
1153                                                 td->o.lat_log_file, "clat");
1154                 } else
1155                         finish_log(td, td->ts.clat_log, "clat");
1156         }
1157         fio_mutex_up(writeout_mutex);
1158         if (td->o.exec_postrun)
1159                 exec_string(td->o.exec_postrun);
1160
1161         if (exitall_on_terminate)
1162                 terminate_threads(td->groupid);
1163
1164 err:
1165         if (td->error)
1166                 printf("fio: pid=%d, err=%d/%s\n", (int) td->pid, td->error,
1167                                                         td->verror);
1168         close_and_free_files(td);
1169         close_ioengine(td);
1170         cleanup_io_u(td);
1171
1172         if (td->o.cpumask_set) {
1173                 int ret = fio_cpuset_exit(&td->o.cpumask);
1174
1175                 td_verror(td, ret, "fio_cpuset_exit");
1176         }
1177
1178         /*
1179          * do this very late, it will log file closing as well
1180          */
1181         if (td->o.write_iolog_file)
1182                 write_iolog_close(td);
1183
1184         options_mem_free(td);
1185         td_set_runstate(td, TD_EXITED);
1186         return (void *) (unsigned long) td->error;
1187 }
1188
1189 /*
1190  * We cannot pass the td data into a forked process, so attach the td and
1191  * pass it to the thread worker.
1192  */
1193 static int fork_main(int shmid, int offset)
1194 {
1195         struct thread_data *td;
1196         void *data, *ret;
1197
1198         data = shmat(shmid, NULL, 0);
1199         if (data == (void *) -1) {
1200                 int __err = errno;
1201
1202                 perror("shmat");
1203                 return __err;
1204         }
1205
1206         td = data + offset * sizeof(struct thread_data);
1207         ret = thread_main(td);
1208         shmdt(data);
1209         return (int) (unsigned long) ret;
1210 }
1211
1212 /*
1213  * Run over the job map and reap the threads that have exited, if any.
1214  */
1215 static void reap_threads(int *nr_running, int *t_rate, int *m_rate)
1216 {
1217         struct thread_data *td;
1218         int i, cputhreads, realthreads, pending, status, ret;
1219
1220         /*
1221          * reap exited threads (TD_EXITED -> TD_REAPED)
1222          */
1223         realthreads = pending = cputhreads = 0;
1224         for_each_td(td, i) {
1225                 int flags = 0;
1226
1227                 /*
1228                  * ->io_ops is NULL for a thread that has closed its
1229                  * io engine
1230                  */
1231                 if (td->io_ops && !strcmp(td->io_ops->name, "cpuio"))
1232                         cputhreads++;
1233                 else
1234                         realthreads++;
1235
1236                 if (!td->pid) {
1237                         pending++;
1238                         continue;
1239                 }
1240                 if (td->runstate == TD_REAPED)
1241                         continue;
1242                 if (td->o.use_thread) {
1243                         if (td->runstate == TD_EXITED) {
1244                                 td_set_runstate(td, TD_REAPED);
1245                                 goto reaped;
1246                         }
1247                         continue;
1248                 }
1249
1250                 flags = WNOHANG;
1251                 if (td->runstate == TD_EXITED)
1252                         flags = 0;
1253
1254                 /*
1255                  * check if someone quit or got killed in an unusual way
1256                  */
1257                 ret = waitpid(td->pid, &status, flags);
1258                 if (ret < 0) {
1259                         if (errno == ECHILD) {
1260                                 log_err("fio: pid=%d disappeared %d\n",
1261                                                 (int) td->pid, td->runstate);
1262                                 td_set_runstate(td, TD_REAPED);
1263                                 goto reaped;
1264                         }
1265                         perror("waitpid");
1266                 } else if (ret == td->pid) {
1267                         if (WIFSIGNALED(status)) {
1268                                 int sig = WTERMSIG(status);
1269
1270                                 if (sig != SIGQUIT)
1271                                         log_err("fio: pid=%d, got signal=%d\n",
1272                                                         (int) td->pid, sig);
1273                                 td_set_runstate(td, TD_REAPED);
1274                                 goto reaped;
1275                         }
1276                         if (WIFEXITED(status)) {
1277                                 if (WEXITSTATUS(status) && !td->error)
1278                                         td->error = WEXITSTATUS(status);
1279
1280                                 td_set_runstate(td, TD_REAPED);
1281                                 goto reaped;
1282                         }
1283                 }
1284
1285                 /*
1286                  * thread is not dead, continue
1287                  */
1288                 pending++;
1289                 continue;
1290 reaped:
1291                 (*nr_running)--;
1292                 (*m_rate) -= (td->o.ratemin[0] + td->o.ratemin[1]);
1293                 (*t_rate) -= (td->o.rate[0] + td->o.rate[1]);
1294                 if (!td->pid)
1295                         pending--;
1296
1297                 if (td->error)
1298                         exit_value++;
1299
1300                 done_secs += mtime_since_now(&td->epoch) / 1000;
1301         }
1302
1303         if (*nr_running == cputhreads && !pending && realthreads)
1304                 terminate_threads(TERMINATE_ALL);
1305 }
1306
1307 static void *gtod_thread_main(void *data)
1308 {
1309         fio_mutex_up(startup_mutex);
1310
1311         /*
1312          * As long as we have jobs around, update the clock. It would be nice
1313          * to have some way of NOT hammering that CPU with gettimeofday(),
1314          * but I'm not sure what to use outside of a simple CPU nop to relax
1315          * it - we don't want to lose precision.
1316          */
1317         while (threads) {
1318                 fio_gtod_update();
1319                 nop;
1320         }
1321
1322         return NULL;
1323 }
1324
1325 static int fio_start_gtod_thread(void)
1326 {
1327         int ret;
1328
1329         ret = pthread_create(&gtod_thread, NULL, gtod_thread_main, NULL);
1330         if (ret) {
1331                 log_err("Can't create gtod thread: %s\n", strerror(ret));
1332                 return 1;
1333         }
1334
1335         ret = pthread_detach(gtod_thread);
1336         if (ret) {
1337                 log_err("Can't detatch gtod thread: %s\n", strerror(ret));
1338                 return 1;
1339         }
1340
1341         dprint(FD_MUTEX, "wait on startup_mutex\n");
1342         fio_mutex_down(startup_mutex);
1343         dprint(FD_MUTEX, "done waiting on startup_mutex\n");
1344         return 0;
1345 }
1346
1347 /*
1348  * Main function for kicking off and reaping jobs, as needed.
1349  */
1350 static void run_threads(void)
1351 {
1352         struct thread_data *td;
1353         unsigned long spent;
1354         int i, todo, nr_running, m_rate, t_rate, nr_started;
1355
1356         if (fio_pin_memory())
1357                 return;
1358
1359         if (fio_gtod_offload && fio_start_gtod_thread())
1360                 return;
1361
1362         if (!terse_output) {
1363                 printf("Starting ");
1364                 if (nr_thread)
1365                         printf("%d thread%s", nr_thread,
1366                                                 nr_thread > 1 ? "s" : "");
1367                 if (nr_process) {
1368                         if (nr_thread)
1369                                 printf(" and ");
1370                         printf("%d process%s", nr_process,
1371                                                 nr_process > 1 ? "es" : "");
1372                 }
1373                 printf("\n");
1374                 fflush(stdout);
1375         }
1376
1377         set_sig_handlers();
1378
1379         todo = thread_number;
1380         nr_running = 0;
1381         nr_started = 0;
1382         m_rate = t_rate = 0;
1383
1384         for_each_td(td, i) {
1385                 print_status_init(td->thread_number - 1);
1386
1387                 if (!td->o.create_serialize) {
1388                         init_disk_util(td);
1389                         continue;
1390                 }
1391
1392                 /*
1393                  * do file setup here so it happens sequentially,
1394                  * we don't want X number of threads getting their
1395                  * client data interspersed on disk
1396                  */
1397                 if (setup_files(td)) {
1398                         exit_value++;
1399                         if (td->error)
1400                                 log_err("fio: pid=%d, err=%d/%s\n",
1401                                         (int) td->pid, td->error, td->verror);
1402                         td_set_runstate(td, TD_REAPED);
1403                         todo--;
1404                 } else {
1405                         struct fio_file *f;
1406                         unsigned int i;
1407
1408                         /*
1409                          * for sharing to work, each job must always open
1410                          * its own files. so close them, if we opened them
1411                          * for creation
1412                          */
1413                         for_each_file(td, f, i) {
1414                                 if (fio_file_open(f))
1415                                         td_io_close_file(td, f);
1416                         }
1417                 }
1418
1419                 init_disk_util(td);
1420         }
1421
1422         set_genesis_time();
1423
1424         while (todo) {
1425                 struct thread_data *map[MAX_JOBS];
1426                 struct timeval this_start;
1427                 int this_jobs = 0, left;
1428
1429                 /*
1430                  * create threads (TD_NOT_CREATED -> TD_CREATED)
1431                  */
1432                 for_each_td(td, i) {
1433                         if (td->runstate != TD_NOT_CREATED)
1434                                 continue;
1435
1436                         /*
1437                          * never got a chance to start, killed by other
1438                          * thread for some reason
1439                          */
1440                         if (td->terminate) {
1441                                 todo--;
1442                                 continue;
1443                         }
1444
1445                         if (td->o.start_delay) {
1446                                 spent = mtime_since_genesis();
1447
1448                                 if (td->o.start_delay * 1000 > spent)
1449                                         continue;
1450                         }
1451
1452                         if (td->o.stonewall && (nr_started || nr_running)) {
1453                                 dprint(FD_PROCESS, "%s: stonewall wait\n",
1454                                                         td->o.name);
1455                                 break;
1456                         }
1457
1458                         /*
1459                          * Set state to created. Thread will transition
1460                          * to TD_INITIALIZED when it's done setting up.
1461                          */
1462                         td_set_runstate(td, TD_CREATED);
1463                         map[this_jobs++] = td;
1464                         nr_started++;
1465
1466                         if (td->o.use_thread) {
1467                                 int ret;
1468
1469                                 dprint(FD_PROCESS, "will pthread_create\n");
1470                                 ret = pthread_create(&td->thread, NULL,
1471                                                         thread_main, td);
1472                                 if (ret) {
1473                                         log_err("pthread_create: %s\n",
1474                                                         strerror(ret));
1475                                         nr_started--;
1476                                         break;
1477                                 }
1478                                 ret = pthread_detach(td->thread);
1479                                 if (ret)
1480                                         log_err("pthread_detach: %s",
1481                                                         strerror(ret));
1482                         } else {
1483                                 pid_t pid;
1484                                 dprint(FD_PROCESS, "will fork\n");
1485                                 pid = fork();
1486                                 if (!pid) {
1487                                         int ret = fork_main(shm_id, i);
1488
1489                                         _exit(ret);
1490                                 } else if (i == fio_debug_jobno)
1491                                         *fio_debug_jobp = pid;
1492                         }
1493                         dprint(FD_MUTEX, "wait on startup_mutex\n");
1494                         fio_mutex_down(startup_mutex);
1495                         dprint(FD_MUTEX, "done waiting on startup_mutex\n");
1496                 }
1497
1498                 /*
1499                  * Wait for the started threads to transition to
1500                  * TD_INITIALIZED.
1501                  */
1502                 fio_gettime(&this_start, NULL);
1503                 left = this_jobs;
1504                 while (left && !fio_abort) {
1505                         if (mtime_since_now(&this_start) > JOB_START_TIMEOUT)
1506                                 break;
1507
1508                         usleep(100000);
1509
1510                         for (i = 0; i < this_jobs; i++) {
1511                                 td = map[i];
1512                                 if (!td)
1513                                         continue;
1514                                 if (td->runstate == TD_INITIALIZED) {
1515                                         map[i] = NULL;
1516                                         left--;
1517                                 } else if (td->runstate >= TD_EXITED) {
1518                                         map[i] = NULL;
1519                                         left--;
1520                                         todo--;
1521                                         nr_running++; /* work-around... */
1522                                 }
1523                         }
1524                 }
1525
1526                 if (left) {
1527                         log_err("fio: %d jobs failed to start\n", left);
1528                         for (i = 0; i < this_jobs; i++) {
1529                                 td = map[i];
1530                                 if (!td)
1531                                         continue;
1532                                 kill(td->pid, SIGTERM);
1533                         }
1534                         break;
1535                 }
1536
1537                 /*
1538                  * start created threads (TD_INITIALIZED -> TD_RUNNING).
1539                  */
1540                 for_each_td(td, i) {
1541                         if (td->runstate != TD_INITIALIZED)
1542                                 continue;
1543
1544                         if (in_ramp_time(td))
1545                                 td_set_runstate(td, TD_RAMP);
1546                         else
1547                                 td_set_runstate(td, TD_RUNNING);
1548                         nr_running++;
1549                         nr_started--;
1550                         m_rate += td->o.ratemin[0] + td->o.ratemin[1];
1551                         t_rate += td->o.rate[0] + td->o.rate[1];
1552                         todo--;
1553                         fio_mutex_up(td->mutex);
1554                 }
1555
1556                 reap_threads(&nr_running, &t_rate, &m_rate);
1557
1558                 if (todo)
1559                         usleep(100000);
1560         }
1561
1562         while (nr_running) {
1563                 reap_threads(&nr_running, &t_rate, &m_rate);
1564                 usleep(10000);
1565         }
1566
1567         update_io_ticks();
1568         fio_unpin_memory();
1569 }
1570
1571 int main(int argc, char *argv[])
1572 {
1573         long ps;
1574
1575         sinit();
1576
1577         /*
1578          * We need locale for number printing, if it isn't set then just
1579          * go with the US format.
1580          */
1581         if (!getenv("LC_NUMERIC"))
1582                 setlocale(LC_NUMERIC, "en_US");
1583
1584         if (parse_options(argc, argv))
1585                 return 1;
1586
1587         if (!thread_number)
1588                 return 0;
1589
1590         ps = sysconf(_SC_PAGESIZE);
1591         if (ps < 0) {
1592                 log_err("Failed to get page size\n");
1593                 return 1;
1594         }
1595
1596         page_size = ps;
1597         page_mask = ps - 1;
1598
1599         if (write_bw_log) {
1600                 setup_log(&agg_io_log[DDIR_READ]);
1601                 setup_log(&agg_io_log[DDIR_WRITE]);
1602         }
1603
1604         startup_mutex = fio_mutex_init(0);
1605         writeout_mutex = fio_mutex_init(1);
1606
1607         set_genesis_time();
1608
1609         status_timer_arm();
1610
1611         run_threads();
1612
1613         if (!fio_abort) {
1614                 show_run_stats();
1615                 if (write_bw_log) {
1616                         __finish_log(agg_io_log[DDIR_READ], "agg-read_bw.log");
1617                         __finish_log(agg_io_log[DDIR_WRITE],
1618                                         "agg-write_bw.log");
1619                 }
1620         }
1621
1622         fio_mutex_remove(startup_mutex);
1623         fio_mutex_remove(writeout_mutex);
1624         return exit_value;
1625 }