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                 if (td_non_fatal_error(err)) {
 392                         /*
 393                          * Continue with the I/Os in case of
 394                          * a non fatal error.
 395                          */
 396                         update_error_count(td, err);
 397                         td_clear_error(td);
 398                         *retptr = 0;
 399                         return 0;
 400                 } else if (td->o.fill_device && err == ENOSPC) {
 401                         /*
 402                          * We expect to hit this error if
 403                          * fill_device option is set.
 404                          */
 405                         td_clear_error(td);
 406                         td->terminate = 1;
 407                         return 1;
 408                 } else {
 409                         /*
 410                          * Stop the I/O in case of a fatal
 411                          * error.
 412                          */
 413                         update_error_count(td, err);
 414                         return 1;
 415                 }
 416         }
 417
 418         return 0;
 419 }
 420
 421 /*
 422  * The main verify engine. Runs over the writes we previously submitted,
 423  * reads the blocks back in, and checks the crc/md5 of the data.
 424  */
 425 static void do_verify(struct thread_data *td)
 426 {
 427         struct fio_file *f;
 428         struct io_u *io_u;
 429         int ret, min_events;
 430         unsigned int i;
 431
 432         /*
 433          * sync io first and invalidate cache, to make sure we really
 434          * read from disk.
 435          */
 436         for_each_file(td, f, i) {
 437                 if (!fio_file_open(f))
 438                         continue;
 439                 if (fio_io_sync(td, f))
 440                         break;
 441                 if (file_invalidate_cache(td, f))
 442                         break;
 443         }
 444
 445         if (td->error)
 446                 return;
 447
 448         td_set_runstate(td, TD_VERIFYING);
 449
 450         io_u = NULL;
 451         while (!td->terminate) {
 452                 int ret2, full;
 453
 454                 update_tv_cache(td);
 455
 456                 if (runtime_exceeded(td, &td->tv_cache)) {
 457                         td->terminate = 1;
 458                         break;
 459                 }
 460
 461                 io_u = __get_io_u(td);
 462                 if (!io_u)
 463                         break;
 464
 465                 if (get_next_verify(td, io_u)) {
 466                         put_io_u(td, io_u);
 467                         break;
 468                 }
 469
 470                 if (td_io_prep(td, io_u)) {
 471                         put_io_u(td, io_u);
 472                         break;
 473                 }
 474
 475                 io_u->end_io = verify_io_u;
 476
 477                 ret = td_io_queue(td, io_u);
 478                 switch (ret) {
 479                 case FIO_Q_COMPLETED:
 480                         if (io_u->error) {
 481                                 ret = -io_u->error;
 482                                 clear_io_u(td, io_u);
 483                         } else if (io_u->resid) {
 484                                 int bytes = io_u->xfer_buflen - io_u->resid;
 485                                 struct fio_file *f = io_u->file;
 486
 487                                 /*
 488                                  * zero read, fail
 489                                  */
 490                                 if (!bytes) {
 491                                         td_verror(td, EIO, "full resid");
 492                                         put_io_u(td, io_u);
 493                                         break;
 494                                 }
 495
 496                                 io_u->xfer_buflen = io_u->resid;
 497                                 io_u->xfer_buf += bytes;
 498                                 io_u->offset += bytes;
 499
 500                                 td->ts.short_io_u[io_u->ddir]++;
 501
 502                                 if (io_u->offset == f->real_file_size)
 503                                         goto sync_done;
 504
 505                                 requeue_io_u(td, &io_u);
 506                         } else {
 507 sync_done:
 508                                 ret = io_u_sync_complete(td, io_u, NULL);
 509                                 if (ret < 0)
 510                                         break;
 511                         }
 512                         continue;
 513                 case FIO_Q_QUEUED:
 514                         break;
 515                 case FIO_Q_BUSY:
 516                         requeue_io_u(td, &io_u);
 517                         ret2 = td_io_commit(td);
 518                         if (ret2 < 0)
 519                                 ret = ret2;
 520                         break;
 521                 default:
 522                         assert(ret < 0);
 523                         td_verror(td, -ret, "td_io_queue");
 524                         break;
 525                 }
 526
 527                 if (break_on_this_error(td, &ret))
 528                         break;
 529
 530                 /*
 531                  * if we can queue more, do so. but check if there are
 532                  * completed io_u's first.
 533                  */
 534                 full = queue_full(td) || ret == FIO_Q_BUSY;
 535                 if (full || !td->o.iodepth_batch_complete) {
 536                         min_events = td->o.iodepth_batch_complete;
 537                         if (full && !min_events)
 538                                 min_events = 1;
 539
 540                         do {
 541                                 /*
 542                                  * Reap required number of io units, if any,
 543                                  * and do the verification on them through
 544                                  * the callback handler
 545                                  */
 546                                 if (io_u_queued_complete(td, min_events, NULL) < 0) {
 547                                         ret = -1;
 548                                         break;
 549                                 }
 550                         } while (full && (td->cur_depth > td->o.iodepth_low));
 551                 }
 552                 if (ret < 0)
 553                         break;
 554         }
 555
 556         if (!td->error) {
 557                 min_events = td->cur_depth;
 558
 559                 if (min_events)
 560                         ret = io_u_queued_complete(td, min_events, NULL);
 561         } else
 562                 cleanup_pending_aio(td);
 563
 564         td_set_runstate(td, TD_RUNNING);
 565 }
 566
 567 /*
 568  * Main IO worker function. It retrieves io_u's to process and queues
 569  * and reaps them, checking for rate and errors along the way.
 570  */
 571 static void do_io(struct thread_data *td)
 572 {
 573         unsigned int i;
 574         int ret = 0;
 575
 576         if (in_ramp_time(td))
 577                 td_set_runstate(td, TD_RAMP);
 578         else
 579                 td_set_runstate(td, TD_RUNNING);
 580
 581         while ((td->this_io_bytes[0] + td->this_io_bytes[1]) < td->o.size) {
 582                 struct timeval comp_time;
 583                 unsigned long bytes_done[2] = { 0, 0 };
 584                 int min_evts = 0;
 585                 struct io_u *io_u;
 586                 int ret2, full;
 587
 588                 if (td->terminate)
 589                         break;
 590
 591                 update_tv_cache(td);
 592
 593                 if (runtime_exceeded(td, &td->tv_cache)) {
 594                         td->terminate = 1;
 595                         break;
 596                 }
 597
 598                 io_u = get_io_u(td);
 599                 if (!io_u)
 600                         break;
 601
 602                 /*
 603                  * Add verification end_io handler, if asked to verify
 604                  * a previously written file.
 605                  */
 606                 if (td->o.verify != VERIFY_NONE && io_u->ddir == DDIR_READ) {
 607                         io_u->end_io = verify_io_u;
 608                         td_set_runstate(td, TD_VERIFYING);
 609                 } else if (in_ramp_time(td))
 610                         td_set_runstate(td, TD_RAMP);
 611                 else
 612                         td_set_runstate(td, TD_RUNNING);
 613
 614                 ret = td_io_queue(td, io_u);
 615                 switch (ret) {
 616                 case FIO_Q_COMPLETED:
 617                         if (io_u->error) {
 618                                 ret = -io_u->error;
 619                                 clear_io_u(td, io_u);
 620                         } else if (io_u->resid) {
 621                                 int bytes = io_u->xfer_buflen - io_u->resid;
 622                                 struct fio_file *f = io_u->file;
 623
 624                                 /*
 625                                  * zero read, fail
 626                                  */
 627                                 if (!bytes) {
 628                                         td_verror(td, EIO, "full resid");
 629                                         put_io_u(td, io_u);
 630                                         break;
 631                                 }
 632
 633                                 io_u->xfer_buflen = io_u->resid;
 634                                 io_u->xfer_buf += bytes;
 635                                 io_u->offset += bytes;
 636
 637                                 td->ts.short_io_u[io_u->ddir]++;
 638
 639                                 if (io_u->offset == f->real_file_size)
 640                                         goto sync_done;
 641
 642                                 requeue_io_u(td, &io_u);
 643                         } else {
 644 sync_done:
 645                                 if (__should_check_rate(td, 0) ||
 646                                     __should_check_rate(td, 1))
 647                                         fio_gettime(&comp_time, NULL);
 648
 649                                 ret = io_u_sync_complete(td, io_u, bytes_done);
 650                                 if (ret < 0)
 651                                         break;
 652                         }
 653                         break;
 654                 case FIO_Q_QUEUED:
 655                         /*
 656                          * if the engine doesn't have a commit hook,
 657                          * the io_u is really queued. if it does have such
 658                          * a hook, it has to call io_u_queued() itself.
 659                          */
 660                         if (td->io_ops->commit == NULL)
 661                                 io_u_queued(td, io_u);
 662                         break;
 663                 case FIO_Q_BUSY:
 664                         requeue_io_u(td, &io_u);
 665                         ret2 = td_io_commit(td);
 666                         if (ret2 < 0)
 667                                 ret = ret2;
 668                         break;
 669                 default:
 670                         assert(ret < 0);
 671                         put_io_u(td, io_u);
 672                         break;
 673                 }
 674
 675                 if (break_on_this_error(td, &ret))
 676                         break;
 677
 678                 /*
 679                  * See if we need to complete some commands
 680                  */
 681                 full = queue_full(td) || ret == FIO_Q_BUSY;
 682                 if (full || !td->o.iodepth_batch_complete) {
 683                         min_evts = td->o.iodepth_batch_complete;
 684                         if (full && !min_evts)
 685                                 min_evts = 1;
 686
 687                         if (__should_check_rate(td, 0) ||
 688                             __should_check_rate(td, 1))
 689                                 fio_gettime(&comp_time, NULL);
 690
 691                         do {
 692                                 ret = io_u_queued_complete(td, min_evts, bytes_done);
 693                                 if (ret < 0)
 694                                         break;
 695
 696                         } while (full && (td->cur_depth > td->o.iodepth_low));
 697                 }
 698
 699                 if (ret < 0)
 700                         break;
 701                 if (!(bytes_done[0] + bytes_done[1]))
 702                         continue;
 703
 704                 if (!in_ramp_time(td) && should_check_rate(td, bytes_done)) {
 705                         if (check_min_rate(td, &comp_time, bytes_done)) {
 706                                 if (exitall_on_terminate)
 707                                         terminate_threads(td->groupid);
 708                                 td_verror(td, EIO, "check_min_rate");
 709                                 break;
 710                         }
 711                 }
 712
 713                 if (td->o.thinktime) {
 714                         unsigned long long b;
 715
 716                         b = td->io_blocks[0] + td->io_blocks[1];
 717                         if (!(b % td->o.thinktime_blocks)) {
 718                                 int left;
 719
 720                                 if (td->o.thinktime_spin)
 721                                         usec_spin(td->o.thinktime_spin);
 722
 723                                 left = td->o.thinktime - td->o.thinktime_spin;
 724                                 if (left)
 725                                         usec_sleep(td, left);
 726                         }
 727                 }
 728         }
 729
 730         if (td->o.fill_device && td->error == ENOSPC) {
 731                 td->error = 0;
 732                 td->terminate = 1;
 733         }
 734         if (!td->error) {
 735                 struct fio_file *f;
 736
 737                 i = td->cur_depth;
 738                 if (i)
 739                         ret = io_u_queued_complete(td, i, NULL);
 740
 741                 if (should_fsync(td) && td->o.end_fsync) {
 742                         td_set_runstate(td, TD_FSYNCING);
 743
 744                         for_each_file(td, f, i) {
 745                                 if (!fio_file_open(f))
 746                                         continue;
 747                                 fio_io_sync(td, f);
 748                         }
 749                 }
 750         } else
 751                 cleanup_pending_aio(td);
 752
 753         /*
 754          * stop job if we failed doing any IO
 755          */
 756         if ((td->this_io_bytes[0] + td->this_io_bytes[1]) == 0)
 757                 td->done = 1;
 758 }
 759
 760 static void cleanup_io_u(struct thread_data *td)
 761 {
 762         struct flist_head *entry, *n;
 763         struct io_u *io_u;
 764
 765         flist_for_each_safe(entry, n, &td->io_u_freelist) {
 766                 io_u = flist_entry(entry, struct io_u, list);
 767
 768                 flist_del(&io_u->list);
 769                 free(io_u);
 770         }
 771
 772         free_io_mem(td);
 773 }
 774
 775 static int init_io_u(struct thread_data *td)
 776 {
 777         struct io_u *io_u;
 778         unsigned int max_bs;
 779         int cl_align, i, max_units;
 780         char *p;
 781
 782         max_units = td->o.iodepth;
 783         max_bs = max(td->o.max_bs[DDIR_READ], td->o.max_bs[DDIR_WRITE]);
 784         td->orig_buffer_size = (unsigned long long) max_bs
 785                                         * (unsigned long long) max_units;
 786
 787         if (td->o.mem_type == MEM_SHMHUGE || td->o.mem_type == MEM_MMAPHUGE) {
 788                 unsigned long bs;
 789
 790                 bs = td->orig_buffer_size + td->o.hugepage_size - 1;
 791                 td->orig_buffer_size = bs & ~(td->o.hugepage_size - 1);
 792         }
 793
 794         if (td->orig_buffer_size != (size_t) td->orig_buffer_size) {
 795                 log_err("fio: IO memory too large. Reduce max_bs or iodepth\n");
 796                 return 1;
 797         }
 798
 799         if (allocate_io_mem(td))
 800                 return 1;
 801
 802         if (td->o.odirect || td->o.mem_align)
 803                 p = PAGE_ALIGN(td->orig_buffer) + td->o.mem_align;
 804         else
 805                 p = td->orig_buffer;
 806
 807         cl_align = os_cache_line_size();
 808
 809         for (i = 0; i < max_units; i++) {
 810                 void *ptr;
 811
 812                 if (td->terminate)
 813                         return 1;
 814
 815                 if (posix_memalign(&ptr, cl_align, sizeof(*io_u))) {
 816                         log_err("fio: posix_memalign=%s\n", strerror(errno));
 817                         break;
 818                 }
 819
 820                 io_u = ptr;
 821                 memset(io_u, 0, sizeof(*io_u));
 822                 INIT_FLIST_HEAD(&io_u->list);
 823                 dprint(FD_MEM, "io_u alloc %p, index %u\n", io_u, i);
 824
 825                 if (!(td->io_ops->flags & FIO_NOIO)) {
 826                         io_u->buf = p + max_bs * i;
 827                         dprint(FD_MEM, "io_u %p, mem %p\n", io_u, io_u->buf);
 828
 829                         if (td_write(td) && !td->o.refill_buffers)
 830                                 io_u_fill_buffer(td, io_u, max_bs);
 831                 }
 832
 833                 io_u->index = i;
 834                 io_u->flags = IO_U_F_FREE;
 835                 flist_add(&io_u->list, &td->io_u_freelist);
 836         }
 837
 838         return 0;
 839 }
 840
 841 static int switch_ioscheduler(struct thread_data *td)
 842 {
 843         char tmp[256], tmp2[128];
 844         FILE *f;
 845         int ret;
 846
 847         if (td->io_ops->flags & FIO_DISKLESSIO)
 848                 return 0;
 849
 850         sprintf(tmp, "%s/queue/scheduler", td->sysfs_root);
 851
 852         f = fopen(tmp, "r+");
 853         if (!f) {
 854                 if (errno == ENOENT) {
 855                         log_err("fio: os or kernel doesn't support IO scheduler"
 856                                 " switching\n");
 857                         return 0;
 858                 }
 859                 td_verror(td, errno, "fopen iosched");
 860                 return 1;
 861         }
 862
 863         /*
 864          * Set io scheduler.
 865          */
 866         ret = fwrite(td->o.ioscheduler, strlen(td->o.ioscheduler), 1, f);
 867         if (ferror(f) || ret != 1) {
 868                 td_verror(td, errno, "fwrite");
 869                 fclose(f);
 870                 return 1;
 871         }
 872
 873         rewind(f);
 874
 875         /*
 876          * Read back and check that the selected scheduler is now the default.
 877          */
 878         ret = fread(tmp, 1, sizeof(tmp), f);
 879         if (ferror(f) || ret < 0) {
 880                 td_verror(td, errno, "fread");
 881                 fclose(f);
 882                 return 1;
 883         }
 884
 885         sprintf(tmp2, "[%s]", td->o.ioscheduler);
 886         if (!strstr(tmp, tmp2)) {
 887                 log_err("fio: io scheduler %s not found\n", td->o.ioscheduler);
 888                 td_verror(td, EINVAL, "iosched_switch");
 889                 fclose(f);
 890                 return 1;
 891         }
 892
 893         fclose(f);
 894         return 0;
 895 }
 896
 897 static int keep_running(struct thread_data *td)
 898 {
 899         unsigned long long io_done;
 900
 901         if (td->done)
 902                 return 0;
 903         if (td->o.time_based)
 904                 return 1;
 905         if (td->o.loops) {
 906                 td->o.loops--;
 907                 return 1;
 908         }
 909
 910         io_done = td->io_bytes[DDIR_READ] + td->io_bytes[DDIR_WRITE]
 911                         + td->io_skip_bytes;
 912         if (io_done < td->o.size)
 913                 return 1;
 914
 915         return 0;
 916 }
 917
 918 static void reset_io_counters(struct thread_data *td)
 919 {
 920         td->ts.stat_io_bytes[0] = td->ts.stat_io_bytes[1] = 0;
 921         td->this_io_bytes[0] = td->this_io_bytes[1] = 0;
 922         td->zone_bytes = 0;
 923         td->rate_bytes[0] = td->rate_bytes[1] = 0;
 924         td->rate_blocks[0] = td->rate_blocks[1] = 0;
 925
 926         td->last_was_sync = 0;
 927
 928         /*
 929          * reset file done count if we are to start over
 930          */
 931         if (td->o.time_based || td->o.loops)
 932                 td->nr_done_files = 0;
 933
 934         /*
 935          * Set the same seed to get repeatable runs
 936          */
 937         td_fill_rand_seeds(td);
 938 }
 939
 940 void reset_all_stats(struct thread_data *td)
 941 {
 942         struct timeval tv;
 943         int i;
 944
 945         reset_io_counters(td);
 946
 947         for (i = 0; i < 2; i++) {
 948                 td->io_bytes[i] = 0;
 949                 td->io_blocks[i] = 0;
 950                 td->io_issues[i] = 0;
 951                 td->ts.total_io_u[i] = 0;
 952         }
 953
 954         fio_gettime(&tv, NULL);
 955         memcpy(&td->epoch, &tv, sizeof(tv));
 956         memcpy(&td->start, &tv, sizeof(tv));
 957 }
 958
 959 static void clear_io_state(struct thread_data *td)
 960 {
 961         struct fio_file *f;
 962         unsigned int i;
 963
 964         reset_io_counters(td);
 965
 966         close_files(td);
 967         for_each_file(td, f, i)
 968                 fio_file_clear_done(f);
 969 }
 970
 971 static int exec_string(const char *string)
 972 {
 973         int ret, newlen = strlen(string) + 1 + 8;
 974         char *str;
 975
 976         str = malloc(newlen);
 977         sprintf(str, "sh -c %s", string);
 978
 979         ret = system(str);
 980         if (ret == -1)
 981                 log_err("fio: exec of cmd <%s> failed\n", str);
 982
 983         free(str);
 984         return ret;
 985 }
 986
 987 /*
 988  * Entry point for the thread based jobs. The process based jobs end up
 989  * here as well, after a little setup.
 990  */
 991 static void *thread_main(void *data)
 992 {
 993         unsigned long long runtime[2], elapsed;
 994         struct thread_data *td = data;
 995         int clear_state;
 996
 997         if (!td->o.use_thread)
 998                 setsid();
 999
1000         td->pid = getpid();
1001
1002         dprint(FD_PROCESS, "jobs pid=%d started\n", (int) td->pid);
1003
1004         INIT_FLIST_HEAD(&td->io_u_freelist);
1005         INIT_FLIST_HEAD(&td->io_u_busylist);
1006         INIT_FLIST_HEAD(&td->io_u_requeues);
1007         INIT_FLIST_HEAD(&td->io_log_list);
1008         INIT_FLIST_HEAD(&td->io_hist_list);
1009         td->io_hist_tree = RB_ROOT;
1010
1011         td_set_runstate(td, TD_INITIALIZED);
1012         dprint(FD_MUTEX, "up startup_mutex\n");
1013         fio_mutex_up(startup_mutex);
1014         dprint(FD_MUTEX, "wait on td->mutex\n");
1015         fio_mutex_down(td->mutex);
1016         dprint(FD_MUTEX, "done waiting on td->mutex\n");
1017
1018         /*
1019          * the ->mutex mutex is now no longer used, close it to avoid
1020          * eating a file descriptor
1021          */
1022         fio_mutex_remove(td->mutex);
1023
1024         /*
1025          * May alter parameters that init_io_u() will use, so we need to
1026          * do this first.
1027          */
1028         if (init_iolog(td))
1029                 goto err;
1030
1031         if (init_io_u(td))
1032                 goto err;
1033
1034         if (td->o.cpumask_set && fio_setaffinity(td) == -1) {
1035                 td_verror(td, errno, "cpu_set_affinity");
1036                 goto err;
1037         }
1038
1039         /*
1040          * If we have a gettimeofday() thread, make sure we exclude that
1041          * thread from this job
1042          */
1043         if (td->o.gtod_cpu) {
1044                 fio_cpu_clear(&td->o.cpumask, td->o.gtod_cpu);
1045                 if (fio_setaffinity(td) == -1) {
1046                         td_verror(td, errno, "cpu_set_affinity");
1047                         goto err;
1048                 }
1049         }
1050
1051         if (td->ioprio_set) {
1052                 if (ioprio_set(IOPRIO_WHO_PROCESS, 0, td->ioprio) == -1) {
1053                         td_verror(td, errno, "ioprio_set");
1054                         goto err;
1055                 }
1056         }
1057
1058         if (nice(td->o.nice) == -1) {
1059                 td_verror(td, errno, "nice");
1060                 goto err;
1061         }
1062
1063         if (td->o.ioscheduler && switch_ioscheduler(td))
1064                 goto err;
1065
1066         if (!td->o.create_serialize && setup_files(td))
1067                 goto err;
1068
1069         if (td_io_init(td))
1070                 goto err;
1071
1072         if (init_random_map(td))
1073                 goto err;
1074
1075         if (td->o.exec_prerun) {
1076                 if (exec_string(td->o.exec_prerun))
1077                         goto err;
1078         }
1079
1080         if (td->o.pre_read) {
1081                 if (pre_read_files(td) < 0)
1082                         goto err;
1083         }
1084
1085         fio_gettime(&td->epoch, NULL);
1086         getrusage(RUSAGE_SELF, &td->ts.ru_start);
1087
1088         runtime[0] = runtime[1] = 0;
1089         clear_state = 0;
1090         while (keep_running(td)) {
1091                 fio_gettime(&td->start, NULL);
1092                 memcpy(&td->ts.stat_sample_time, &td->start, sizeof(td->start));
1093                 memcpy(&td->tv_cache, &td->start, sizeof(td->start));
1094
1095                 if (td->o.ratemin[0] || td->o.ratemin[1])
1096                         memcpy(&td->lastrate, &td->ts.stat_sample_time,
1097                                                         sizeof(td->lastrate));
1098
1099                 if (clear_state)
1100                         clear_io_state(td);
1101
1102                 prune_io_piece_log(td);
1103
1104                 do_io(td);
1105
1106                 clear_state = 1;
1107
1108                 if (td_read(td) && td->io_bytes[DDIR_READ]) {
1109                         elapsed = utime_since_now(&td->start);
1110                         runtime[DDIR_READ] += elapsed;
1111                 }
1112                 if (td_write(td) && td->io_bytes[DDIR_WRITE]) {
1113                         elapsed = utime_since_now(&td->start);
1114                         runtime[DDIR_WRITE] += elapsed;
1115                 }
1116
1117                 if (td->error || td->terminate)
1118                         break;
1119
1120                 if (!td->o.do_verify ||
1121                     td->o.verify == VERIFY_NONE ||
1122                     (td->io_ops->flags & FIO_UNIDIR))
1123                         continue;
1124
1125                 clear_io_state(td);
1126
1127                 fio_gettime(&td->start, NULL);
1128
1129                 do_verify(td);
1130
1131                 runtime[DDIR_READ] += utime_since_now(&td->start);
1132
1133                 if (td->error || td->terminate)
1134                         break;
1135         }
1136
1137         update_rusage_stat(td);
1138         td->ts.runtime[0] = (runtime[0] + 999) / 1000;
1139         td->ts.runtime[1] = (runtime[1] + 999) / 1000;
1140         td->ts.total_run_time = mtime_since_now(&td->epoch);
1141         td->ts.io_bytes[0] = td->io_bytes[0];
1142         td->ts.io_bytes[1] = td->io_bytes[1];
1143
1144         fio_mutex_down(writeout_mutex);
1145         if (td->ts.bw_log) {
1146                 if (td->o.bw_log_file) {
1147                         finish_log_named(td, td->ts.bw_log,
1148                                                 td->o.bw_log_file, "bw");
1149                 } else
1150                         finish_log(td, td->ts.bw_log, "bw");
1151         }
1152         if (td->ts.slat_log) {
1153                 if (td->o.lat_log_file) {
1154                         finish_log_named(td, td->ts.slat_log,
1155                                                 td->o.lat_log_file, "slat");
1156                 } else
1157                         finish_log(td, td->ts.slat_log, "slat");
1158         }
1159         if (td->ts.clat_log) {
1160                 if (td->o.lat_log_file) {
1161                         finish_log_named(td, td->ts.clat_log,
1162                                                 td->o.lat_log_file, "clat");
1163                 } else
1164                         finish_log(td, td->ts.clat_log, "clat");
1165         }
1166         fio_mutex_up(writeout_mutex);
1167         if (td->o.exec_postrun)
1168                 exec_string(td->o.exec_postrun);
1169
1170         if (exitall_on_terminate)
1171                 terminate_threads(td->groupid);
1172
1173 err:
1174         if (td->error)
1175                 printf("fio: pid=%d, err=%d/%s\n", (int) td->pid, td->error,
1176                                                         td->verror);
1177         close_and_free_files(td);
1178         close_ioengine(td);
1179         cleanup_io_u(td);
1180
1181         if (td->o.cpumask_set) {
1182                 int ret = fio_cpuset_exit(&td->o.cpumask);
1183
1184                 td_verror(td, ret, "fio_cpuset_exit");
1185         }
1186
1187         /*
1188          * do this very late, it will log file closing as well
1189          */
1190         if (td->o.write_iolog_file)
1191                 write_iolog_close(td);
1192
1193         options_mem_free(td);
1194         td_set_runstate(td, TD_EXITED);
1195         return (void *) (unsigned long) td->error;
1196 }
1197
1198 /*
1199  * We cannot pass the td data into a forked process, so attach the td and
1200  * pass it to the thread worker.
1201  */
1202 static int fork_main(int shmid, int offset)
1203 {
1204         struct thread_data *td;
1205         void *data, *ret;
1206
1207         data = shmat(shmid, NULL, 0);
1208         if (data == (void *) -1) {
1209                 int __err = errno;
1210
1211                 perror("shmat");
1212                 return __err;
1213         }
1214
1215         td = data + offset * sizeof(struct thread_data);
1216         ret = thread_main(td);
1217         shmdt(data);
1218         return (int) (unsigned long) ret;
1219 }
1220
1221 /*
1222  * Run over the job map and reap the threads that have exited, if any.
1223  */
1224 static void reap_threads(int *nr_running, int *t_rate, int *m_rate)
1225 {
1226         struct thread_data *td;
1227         int i, cputhreads, realthreads, pending, status, ret;
1228
1229         /*
1230          * reap exited threads (TD_EXITED -> TD_REAPED)
1231          */
1232         realthreads = pending = cputhreads = 0;
1233         for_each_td(td, i) {
1234                 int flags = 0;
1235
1236                 /*
1237                  * ->io_ops is NULL for a thread that has closed its
1238                  * io engine
1239                  */
1240                 if (td->io_ops && !strcmp(td->io_ops->name, "cpuio"))
1241                         cputhreads++;
1242                 else
1243                         realthreads++;
1244
1245                 if (!td->pid) {
1246                         pending++;
1247                         continue;
1248                 }
1249                 if (td->runstate == TD_REAPED)
1250                         continue;
1251                 if (td->o.use_thread) {
1252                         if (td->runstate == TD_EXITED) {
1253                                 td_set_runstate(td, TD_REAPED);
1254                                 goto reaped;
1255                         }
1256                         continue;
1257                 }
1258
1259                 flags = WNOHANG;
1260                 if (td->runstate == TD_EXITED)
1261                         flags = 0;
1262
1263                 /*
1264                  * check if someone quit or got killed in an unusual way
1265                  */
1266                 ret = waitpid(td->pid, &status, flags);
1267                 if (ret < 0) {
1268                         if (errno == ECHILD) {
1269                                 log_err("fio: pid=%d disappeared %d\n",
1270                                                 (int) td->pid, td->runstate);
1271                                 td_set_runstate(td, TD_REAPED);
1272                                 goto reaped;
1273                         }
1274                         perror("waitpid");
1275                 } else if (ret == td->pid) {
1276                         if (WIFSIGNALED(status)) {
1277                                 int sig = WTERMSIG(status);
1278
1279                                 if (sig != SIGQUIT)
1280                                         log_err("fio: pid=%d, got signal=%d\n",
1281                                                         (int) td->pid, sig);
1282                                 td_set_runstate(td, TD_REAPED);
1283                                 goto reaped;
1284                         }
1285                         if (WIFEXITED(status)) {
1286                                 if (WEXITSTATUS(status) && !td->error)
1287                                         td->error = WEXITSTATUS(status);
1288
1289                                 td_set_runstate(td, TD_REAPED);
1290                                 goto reaped;
1291                         }
1292                 }
1293
1294                 /*
1295                  * thread is not dead, continue
1296                  */
1297                 pending++;
1298                 continue;
1299 reaped:
1300                 (*nr_running)--;
1301                 (*m_rate) -= (td->o.ratemin[0] + td->o.ratemin[1]);
1302                 (*t_rate) -= (td->o.rate[0] + td->o.rate[1]);
1303                 if (!td->pid)
1304                         pending--;
1305
1306                 if (td->error)
1307                         exit_value++;
1308
1309                 done_secs += mtime_since_now(&td->epoch) / 1000;
1310         }
1311
1312         if (*nr_running == cputhreads && !pending && realthreads)
1313                 terminate_threads(TERMINATE_ALL);
1314 }
1315
1316 static void *gtod_thread_main(void *data)
1317 {
1318         fio_mutex_up(startup_mutex);
1319
1320         /*
1321          * As long as we have jobs around, update the clock. It would be nice
1322          * to have some way of NOT hammering that CPU with gettimeofday(),
1323          * but I'm not sure what to use outside of a simple CPU nop to relax
1324          * it - we don't want to lose precision.
1325          */
1326         while (threads) {
1327                 fio_gtod_update();
1328                 nop;
1329         }
1330
1331         return NULL;
1332 }
1333
1334 static int fio_start_gtod_thread(void)
1335 {
1336         int ret;
1337
1338         ret = pthread_create(&gtod_thread, NULL, gtod_thread_main, NULL);
1339         if (ret) {
1340                 log_err("Can't create gtod thread: %s\n", strerror(ret));
1341                 return 1;
1342         }
1343
1344         ret = pthread_detach(gtod_thread);
1345         if (ret) {
1346                 log_err("Can't detatch gtod thread: %s\n", strerror(ret));
1347                 return 1;
1348         }
1349
1350         dprint(FD_MUTEX, "wait on startup_mutex\n");
1351         fio_mutex_down(startup_mutex);
1352         dprint(FD_MUTEX, "done waiting on startup_mutex\n");
1353         return 0;
1354 }
1355
1356 /*
1357  * Main function for kicking off and reaping jobs, as needed.
1358  */
1359 static void run_threads(void)
1360 {
1361         struct thread_data *td;
1362         unsigned long spent;
1363         int i, todo, nr_running, m_rate, t_rate, nr_started;
1364
1365         if (fio_pin_memory())
1366                 return;
1367
1368         if (fio_gtod_offload && fio_start_gtod_thread())
1369                 return;
1370
1371         if (!terse_output) {
1372                 printf("Starting ");
1373                 if (nr_thread)
1374                         printf("%d thread%s", nr_thread,
1375                                                 nr_thread > 1 ? "s" : "");
1376                 if (nr_process) {
1377                         if (nr_thread)
1378                                 printf(" and ");
1379                         printf("%d process%s", nr_process,
1380                                                 nr_process > 1 ? "es" : "");
1381                 }
1382                 printf("\n");
1383                 fflush(stdout);
1384         }
1385
1386         set_sig_handlers();
1387
1388         todo = thread_number;
1389         nr_running = 0;
1390         nr_started = 0;
1391         m_rate = t_rate = 0;
1392
1393         for_each_td(td, i) {
1394                 print_status_init(td->thread_number - 1);
1395
1396                 if (!td->o.create_serialize) {
1397                         init_disk_util(td);
1398                         continue;
1399                 }
1400
1401                 /*
1402                  * do file setup here so it happens sequentially,
1403                  * we don't want X number of threads getting their
1404                  * client data interspersed on disk
1405                  */
1406                 if (setup_files(td)) {
1407                         exit_value++;
1408                         if (td->error)
1409                                 log_err("fio: pid=%d, err=%d/%s\n",
1410                                         (int) td->pid, td->error, td->verror);
1411                         td_set_runstate(td, TD_REAPED);
1412                         todo--;
1413                 } else {
1414                         struct fio_file *f;
1415                         unsigned int i;
1416
1417                         /*
1418                          * for sharing to work, each job must always open
1419                          * its own files. so close them, if we opened them
1420                          * for creation
1421                          */
1422                         for_each_file(td, f, i) {
1423                                 if (fio_file_open(f))
1424                                         td_io_close_file(td, f);
1425                         }
1426                 }
1427
1428                 init_disk_util(td);
1429         }
1430
1431         set_genesis_time();
1432
1433         while (todo) {
1434                 struct thread_data *map[MAX_JOBS];
1435                 struct timeval this_start;
1436                 int this_jobs = 0, left;
1437
1438                 /*
1439                  * create threads (TD_NOT_CREATED -> TD_CREATED)
1440                  */
1441                 for_each_td(td, i) {
1442                         if (td->runstate != TD_NOT_CREATED)
1443                                 continue;
1444
1445                         /*
1446                          * never got a chance to start, killed by other
1447                          * thread for some reason
1448                          */
1449                         if (td->terminate) {
1450                                 todo--;
1451                                 continue;
1452                         }
1453
1454                         if (td->o.start_delay) {
1455                                 spent = mtime_since_genesis();
1456
1457                                 if (td->o.start_delay * 1000 > spent)
1458                                         continue;
1459                         }
1460
1461                         if (td->o.stonewall && (nr_started || nr_running)) {
1462                                 dprint(FD_PROCESS, "%s: stonewall wait\n",
1463                                                         td->o.name);
1464                                 break;
1465                         }
1466
1467                         /*
1468                          * Set state to created. Thread will transition
1469                          * to TD_INITIALIZED when it's done setting up.
1470                          */
1471                         td_set_runstate(td, TD_CREATED);
1472                         map[this_jobs++] = td;
1473                         nr_started++;
1474
1475                         if (td->o.use_thread) {
1476                                 int ret;
1477
1478                                 dprint(FD_PROCESS, "will pthread_create\n");
1479                                 ret = pthread_create(&td->thread, NULL,
1480                                                         thread_main, td);
1481                                 if (ret) {
1482                                         log_err("pthread_create: %s\n",
1483                                                         strerror(ret));
1484                                         nr_started--;
1485                                         break;
1486                                 }
1487                                 ret = pthread_detach(td->thread);
1488                                 if (ret)
1489                                         log_err("pthread_detach: %s",
1490                                                         strerror(ret));
1491                         } else {
1492                                 pid_t pid;
1493                                 dprint(FD_PROCESS, "will fork\n");
1494                                 pid = fork();
1495                                 if (!pid) {
1496                                         int ret = fork_main(shm_id, i);
1497
1498                                         _exit(ret);
1499                                 } else if (i == fio_debug_jobno)
1500                                         *fio_debug_jobp = pid;
1501                         }
1502                         dprint(FD_MUTEX, "wait on startup_mutex\n");
1503                         if (fio_mutex_down_timeout(startup_mutex, 10)) {
1504                                 log_err("fio: job startup hung? exiting.\n");
1505                                 terminate_threads(TERMINATE_ALL);
1506                                 fio_abort = 1;
1507                                 nr_started--;
1508                                 break;
1509                         }
1510                         dprint(FD_MUTEX, "done waiting on startup_mutex\n");
1511                 }
1512
1513                 /*
1514                  * Wait for the started threads to transition to
1515                  * TD_INITIALIZED.
1516                  */
1517                 fio_gettime(&this_start, NULL);
1518                 left = this_jobs;
1519                 while (left && !fio_abort) {
1520                         if (mtime_since_now(&this_start) > JOB_START_TIMEOUT)
1521                                 break;
1522
1523                         usleep(100000);
1524
1525                         for (i = 0; i < this_jobs; i++) {
1526                                 td = map[i];
1527                                 if (!td)
1528                                         continue;
1529                                 if (td->runstate == TD_INITIALIZED) {
1530                                         map[i] = NULL;
1531                                         left--;
1532                                 } else if (td->runstate >= TD_EXITED) {
1533                                         map[i] = NULL;
1534                                         left--;
1535                                         todo--;
1536                                         nr_running++; /* work-around... */
1537                                 }
1538                         }
1539                 }
1540
1541                 if (left) {
1542                         log_err("fio: %d jobs failed to start\n", left);
1543                         for (i = 0; i < this_jobs; i++) {
1544                                 td = map[i];
1545                                 if (!td)
1546                                         continue;
1547                                 kill(td->pid, SIGTERM);
1548                         }
1549                         break;
1550                 }
1551
1552                 /*
1553                  * start created threads (TD_INITIALIZED -> TD_RUNNING).
1554                  */
1555                 for_each_td(td, i) {
1556                         if (td->runstate != TD_INITIALIZED)
1557                                 continue;
1558
1559                         if (in_ramp_time(td))
1560                                 td_set_runstate(td, TD_RAMP);
1561                         else
1562                                 td_set_runstate(td, TD_RUNNING);
1563                         nr_running++;
1564                         nr_started--;
1565                         m_rate += td->o.ratemin[0] + td->o.ratemin[1];
1566                         t_rate += td->o.rate[0] + td->o.rate[1];
1567                         todo--;
1568                         fio_mutex_up(td->mutex);
1569                 }
1570
1571                 reap_threads(&nr_running, &t_rate, &m_rate);
1572
1573                 if (todo)
1574                         usleep(100000);
1575         }
1576
1577         while (nr_running) {
1578                 reap_threads(&nr_running, &t_rate, &m_rate);
1579                 usleep(10000);
1580         }
1581
1582         update_io_ticks();
1583         fio_unpin_memory();
1584 }
1585
1586 int main(int argc, char *argv[])
1587 {
1588         long ps;
1589
1590         sinit();
1591
1592         /*
1593          * We need locale for number printing, if it isn't set then just
1594          * go with the US format.
1595          */
1596         if (!getenv("LC_NUMERIC"))
1597                 setlocale(LC_NUMERIC, "en_US");
1598
1599         if (parse_options(argc, argv))
1600                 return 1;
1601
1602         if (!thread_number)
1603                 return 0;
1604
1605         ps = sysconf(_SC_PAGESIZE);
1606         if (ps < 0) {
1607                 log_err("Failed to get page size\n");
1608                 return 1;
1609         }
1610
1611         page_size = ps;
1612         page_mask = ps - 1;
1613
1614         if (write_bw_log) {
1615                 setup_log(&agg_io_log[DDIR_READ]);
1616                 setup_log(&agg_io_log[DDIR_WRITE]);
1617         }
1618
1619         startup_mutex = fio_mutex_init(0);
1620         writeout_mutex = fio_mutex_init(1);
1621
1622         set_genesis_time();
1623
1624         status_timer_arm();
1625
1626         run_threads();
1627
1628         if (!fio_abort) {
1629                 show_run_stats();
1630                 if (write_bw_log) {
1631                         __finish_log(agg_io_log[DDIR_READ], "agg-read_bw.log");
1632                         __finish_log(agg_io_log[DDIR_WRITE],
1633                                         "agg-write_bw.log");
1634                 }
1635         }
1636
1637         fio_mutex_remove(startup_mutex);
1638         fio_mutex_remove(writeout_mutex);
1639         return exit_value;
1640 }