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                                                 " %luKB/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                 if (td->o.verify_async)
 476                         io_u->end_io = verify_io_u_async;
 477                 else
 478                         io_u->end_io = verify_io_u;
 479
 480                 ret = td_io_queue(td, io_u);
 481                 switch (ret) {
 482                 case FIO_Q_COMPLETED:
 483                         if (io_u->error) {
 484                                 ret = -io_u->error;
 485                                 clear_io_u(td, io_u);
 486                         } else if (io_u->resid) {
 487                                 int bytes = io_u->xfer_buflen - io_u->resid;
 488                                 struct fio_file *f = io_u->file;
 489
 490                                 /*
 491                                  * zero read, fail
 492                                  */
 493                                 if (!bytes) {
 494                                         td_verror(td, EIO, "full resid");
 495                                         put_io_u(td, io_u);
 496                                         break;
 497                                 }
 498
 499                                 io_u->xfer_buflen = io_u->resid;
 500                                 io_u->xfer_buf += bytes;
 501                                 io_u->offset += bytes;
 502
 503                                 td->ts.short_io_u[io_u->ddir]++;
 504
 505                                 if (io_u->offset == f->real_file_size)
 506                                         goto sync_done;
 507
 508                                 requeue_io_u(td, &io_u);
 509                         } else {
 510 sync_done:
 511                                 ret = io_u_sync_complete(td, io_u, NULL);
 512                                 if (ret < 0)
 513                                         break;
 514                         }
 515                         continue;
 516                 case FIO_Q_QUEUED:
 517                         break;
 518                 case FIO_Q_BUSY:
 519                         requeue_io_u(td, &io_u);
 520                         ret2 = td_io_commit(td);
 521                         if (ret2 < 0)
 522                                 ret = ret2;
 523                         break;
 524                 default:
 525                         assert(ret < 0);
 526                         td_verror(td, -ret, "td_io_queue");
 527                         break;
 528                 }
 529
 530                 if (break_on_this_error(td, &ret))
 531                         break;
 532
 533                 /*
 534                  * if we can queue more, do so. but check if there are
 535                  * completed io_u's first.
 536                  */
 537                 full = queue_full(td) || ret == FIO_Q_BUSY;
 538                 if (full || !td->o.iodepth_batch_complete) {
 539                         min_events = td->o.iodepth_batch_complete;
 540                         if (full && !min_events)
 541                                 min_events = 1;
 542
 543                         do {
 544                                 /*
 545                                  * Reap required number of io units, if any,
 546                                  * and do the verification on them through
 547                                  * the callback handler
 548                                  */
 549                                 if (io_u_queued_complete(td, min_events, NULL) < 0) {
 550                                         ret = -1;
 551                                         break;
 552                                 }
 553                         } while (full && (td->cur_depth > td->o.iodepth_low));
 554                 }
 555                 if (ret < 0)
 556                         break;
 557         }
 558
 559         if (!td->error) {
 560                 min_events = td->cur_depth;
 561
 562                 if (min_events)
 563                         ret = io_u_queued_complete(td, min_events, NULL);
 564         } else
 565                 cleanup_pending_aio(td);
 566
 567         td_set_runstate(td, TD_RUNNING);
 568 }
 569
 570 /*
 571  * Main IO worker function. It retrieves io_u's to process and queues
 572  * and reaps them, checking for rate and errors along the way.
 573  */
 574 static void do_io(struct thread_data *td)
 575 {
 576         unsigned int i;
 577         int ret = 0;
 578
 579         if (in_ramp_time(td))
 580                 td_set_runstate(td, TD_RAMP);
 581         else
 582                 td_set_runstate(td, TD_RUNNING);
 583
 584         while ( (td->o.read_iolog_file && !flist_empty(&td->io_log_list)) ||
 585                 ((td->this_io_bytes[0] + td->this_io_bytes[1]) < td->o.size) ) {
 586                 struct timeval comp_time;
 587                 unsigned long bytes_done[2] = { 0, 0 };
 588                 int min_evts = 0;
 589                 struct io_u *io_u;
 590                 int ret2, full;
 591
 592                 if (td->terminate)
 593                         break;
 594
 595                 update_tv_cache(td);
 596
 597                 if (runtime_exceeded(td, &td->tv_cache)) {
 598                         td->terminate = 1;
 599                         break;
 600                 }
 601
 602                 io_u = get_io_u(td);
 603                 if (!io_u)
 604                         break;
 605
 606                 /*
 607                  * Add verification end_io handler, if asked to verify
 608                  * a previously written file.
 609                  */
 610                 if (td->o.verify != VERIFY_NONE && io_u->ddir == DDIR_READ &&
 611                     !td_rw(td)) {
 612                         if (td->o.verify_async)
 613                                 io_u->end_io = verify_io_u_async;
 614                         else
 615                                 io_u->end_io = verify_io_u;
 616                         td_set_runstate(td, TD_VERIFYING);
 617                 } else if (in_ramp_time(td))
 618                         td_set_runstate(td, TD_RAMP);
 619                 else
 620                         td_set_runstate(td, TD_RUNNING);
 621
 622                 ret = td_io_queue(td, io_u);
 623                 switch (ret) {
 624                 case FIO_Q_COMPLETED:
 625                         if (io_u->error) {
 626                                 ret = -io_u->error;
 627                                 clear_io_u(td, io_u);
 628                         } else if (io_u->resid) {
 629                                 int bytes = io_u->xfer_buflen - io_u->resid;
 630                                 struct fio_file *f = io_u->file;
 631
 632                                 /*
 633                                  * zero read, fail
 634                                  */
 635                                 if (!bytes) {
 636                                         td_verror(td, EIO, "full resid");
 637                                         put_io_u(td, io_u);
 638                                         break;
 639                                 }
 640
 641                                 io_u->xfer_buflen = io_u->resid;
 642                                 io_u->xfer_buf += bytes;
 643                                 io_u->offset += bytes;
 644
 645                                 td->ts.short_io_u[io_u->ddir]++;
 646
 647                                 if (io_u->offset == f->real_file_size)
 648                                         goto sync_done;
 649
 650                                 requeue_io_u(td, &io_u);
 651                         } else {
 652 sync_done:
 653                                 if (__should_check_rate(td, 0) ||
 654                                     __should_check_rate(td, 1))
 655                                         fio_gettime(&comp_time, NULL);
 656
 657                                 ret = io_u_sync_complete(td, io_u, bytes_done);
 658                                 if (ret < 0)
 659                                         break;
 660                         }
 661                         break;
 662                 case FIO_Q_QUEUED:
 663                         /*
 664                          * if the engine doesn't have a commit hook,
 665                          * the io_u is really queued. if it does have such
 666                          * a hook, it has to call io_u_queued() itself.
 667                          */
 668                         if (td->io_ops->commit == NULL)
 669                                 io_u_queued(td, io_u);
 670                         break;
 671                 case FIO_Q_BUSY:
 672                         requeue_io_u(td, &io_u);
 673                         ret2 = td_io_commit(td);
 674                         if (ret2 < 0)
 675                                 ret = ret2;
 676                         break;
 677                 default:
 678                         assert(ret < 0);
 679                         put_io_u(td, io_u);
 680                         break;
 681                 }
 682
 683                 if (break_on_this_error(td, &ret))
 684                         break;
 685
 686                 /*
 687                  * See if we need to complete some commands
 688                  */
 689                 full = queue_full(td) || ret == FIO_Q_BUSY;
 690                 if (full || !td->o.iodepth_batch_complete) {
 691                         min_evts = td->o.iodepth_batch_complete;
 692                         if (full && !min_evts)
 693                                 min_evts = 1;
 694
 695                         if (__should_check_rate(td, 0) ||
 696                             __should_check_rate(td, 1))
 697                                 fio_gettime(&comp_time, NULL);
 698
 699                         do {
 700                                 ret = io_u_queued_complete(td, min_evts, bytes_done);
 701                                 if (ret < 0)
 702                                         break;
 703
 704                         } while (full && (td->cur_depth > td->o.iodepth_low));
 705                 }
 706
 707                 if (ret < 0)
 708                         break;
 709                 if (!(bytes_done[0] + bytes_done[1]))
 710                         continue;
 711
 712                 if (!in_ramp_time(td) && should_check_rate(td, bytes_done)) {
 713                         if (check_min_rate(td, &comp_time, bytes_done)) {
 714                                 if (exitall_on_terminate)
 715                                         terminate_threads(td->groupid);
 716                                 td_verror(td, EIO, "check_min_rate");
 717                                 break;
 718                         }
 719                 }
 720
 721                 if (td->o.thinktime) {
 722                         unsigned long long b;
 723
 724                         b = td->io_blocks[0] + td->io_blocks[1];
 725                         if (!(b % td->o.thinktime_blocks)) {
 726                                 int left;
 727
 728                                 if (td->o.thinktime_spin)
 729                                         usec_spin(td->o.thinktime_spin);
 730
 731                                 left = td->o.thinktime - td->o.thinktime_spin;
 732                                 if (left)
 733                                         usec_sleep(td, left);
 734                         }
 735                 }
 736         }
 737
 738         if (td->o.fill_device && td->error == ENOSPC) {
 739                 td->error = 0;
 740                 td->terminate = 1;
 741         }
 742         if (!td->error) {
 743                 struct fio_file *f;
 744
 745                 i = td->cur_depth;
 746                 if (i)
 747                         ret = io_u_queued_complete(td, i, NULL);
 748
 749                 if (should_fsync(td) && td->o.end_fsync) {
 750                         td_set_runstate(td, TD_FSYNCING);
 751
 752                         for_each_file(td, f, i) {
 753                                 if (!fio_file_open(f))
 754                                         continue;
 755                                 fio_io_sync(td, f);
 756                         }
 757                 }
 758         } else
 759                 cleanup_pending_aio(td);
 760
 761         /*
 762          * stop job if we failed doing any IO
 763          */
 764         if ((td->this_io_bytes[0] + td->this_io_bytes[1]) == 0)
 765                 td->done = 1;
 766 }
 767
 768 static void cleanup_io_u(struct thread_data *td)
 769 {
 770         struct flist_head *entry, *n;
 771         struct io_u *io_u;
 772
 773         flist_for_each_safe(entry, n, &td->io_u_freelist) {
 774                 io_u = flist_entry(entry, struct io_u, list);
 775
 776                 flist_del(&io_u->list);
 777                 free(io_u);
 778         }
 779
 780         free_io_mem(td);
 781 }
 782
 783 static int init_io_u(struct thread_data *td)
 784 {
 785         struct io_u *io_u;
 786         unsigned int max_bs;
 787         int cl_align, i, max_units;
 788         char *p;
 789
 790         max_units = td->o.iodepth;
 791         max_bs = max(td->o.max_bs[DDIR_READ], td->o.max_bs[DDIR_WRITE]);
 792         td->orig_buffer_size = (unsigned long long) max_bs
 793                                         * (unsigned long long) max_units;
 794
 795         if (td->o.mem_type == MEM_SHMHUGE || td->o.mem_type == MEM_MMAPHUGE) {
 796                 unsigned long bs;
 797
 798                 bs = td->orig_buffer_size + td->o.hugepage_size - 1;
 799                 td->orig_buffer_size = bs & ~(td->o.hugepage_size - 1);
 800         }
 801
 802         if (td->orig_buffer_size != (size_t) td->orig_buffer_size) {
 803                 log_err("fio: IO memory too large. Reduce max_bs or iodepth\n");
 804                 return 1;
 805         }
 806
 807         if (allocate_io_mem(td))
 808                 return 1;
 809
 810         if (td->o.odirect || td->o.mem_align)
 811                 p = PAGE_ALIGN(td->orig_buffer) + td->o.mem_align;
 812         else
 813                 p = td->orig_buffer;
 814
 815         cl_align = os_cache_line_size();
 816
 817         for (i = 0; i < max_units; i++) {
 818                 void *ptr;
 819
 820                 if (td->terminate)
 821                         return 1;
 822
 823                 if (posix_memalign(&ptr, cl_align, sizeof(*io_u))) {
 824                         log_err("fio: posix_memalign=%s\n", strerror(errno));
 825                         break;
 826                 }
 827
 828                 io_u = ptr;
 829                 memset(io_u, 0, sizeof(*io_u));
 830                 INIT_FLIST_HEAD(&io_u->list);
 831                 dprint(FD_MEM, "io_u alloc %p, index %u\n", io_u, i);
 832
 833                 if (!(td->io_ops->flags & FIO_NOIO)) {
 834                         io_u->buf = p + max_bs * i;
 835                         dprint(FD_MEM, "io_u %p, mem %p\n", io_u, io_u->buf);
 836
 837                         if (td_write(td) && !td->o.refill_buffers)
 838                                 io_u_fill_buffer(td, io_u, max_bs);
 839                 }
 840
 841                 io_u->index = i;
 842                 io_u->flags = IO_U_F_FREE;
 843                 flist_add(&io_u->list, &td->io_u_freelist);
 844         }
 845
 846         return 0;
 847 }
 848
 849 static int switch_ioscheduler(struct thread_data *td)
 850 {
 851         char tmp[256], tmp2[128];
 852         FILE *f;
 853         int ret;
 854
 855         if (td->io_ops->flags & FIO_DISKLESSIO)
 856                 return 0;
 857
 858         sprintf(tmp, "%s/queue/scheduler", td->sysfs_root);
 859
 860         f = fopen(tmp, "r+");
 861         if (!f) {
 862                 if (errno == ENOENT) {
 863                         log_err("fio: os or kernel doesn't support IO scheduler"
 864                                 " switching\n");
 865                         return 0;
 866                 }
 867                 td_verror(td, errno, "fopen iosched");
 868                 return 1;
 869         }
 870
 871         /*
 872          * Set io scheduler.
 873          */
 874         ret = fwrite(td->o.ioscheduler, strlen(td->o.ioscheduler), 1, f);
 875         if (ferror(f) || ret != 1) {
 876                 td_verror(td, errno, "fwrite");
 877                 fclose(f);
 878                 return 1;
 879         }
 880
 881         rewind(f);
 882
 883         /*
 884          * Read back and check that the selected scheduler is now the default.
 885          */
 886         ret = fread(tmp, 1, sizeof(tmp), f);
 887         if (ferror(f) || ret < 0) {
 888                 td_verror(td, errno, "fread");
 889                 fclose(f);
 890                 return 1;
 891         }
 892
 893         sprintf(tmp2, "[%s]", td->o.ioscheduler);
 894         if (!strstr(tmp, tmp2)) {
 895                 log_err("fio: io scheduler %s not found\n", td->o.ioscheduler);
 896                 td_verror(td, EINVAL, "iosched_switch");
 897                 fclose(f);
 898                 return 1;
 899         }
 900
 901         fclose(f);
 902         return 0;
 903 }
 904
 905 static int keep_running(struct thread_data *td)
 906 {
 907         unsigned long long io_done;
 908
 909         if (td->done)
 910                 return 0;
 911         if (td->o.time_based)
 912                 return 1;
 913         if (td->o.loops) {
 914                 td->o.loops--;
 915                 return 1;
 916         }
 917
 918         io_done = td->io_bytes[DDIR_READ] + td->io_bytes[DDIR_WRITE]
 919                         + td->io_skip_bytes;
 920         if (io_done < td->o.size)
 921                 return 1;
 922
 923         return 0;
 924 }
 925
 926 static void reset_io_counters(struct thread_data *td)
 927 {
 928         td->ts.stat_io_bytes[0] = td->ts.stat_io_bytes[1] = 0;
 929         td->this_io_bytes[0] = td->this_io_bytes[1] = 0;
 930         td->zone_bytes = 0;
 931         td->rate_bytes[0] = td->rate_bytes[1] = 0;
 932         td->rate_blocks[0] = td->rate_blocks[1] = 0;
 933
 934         td->last_was_sync = 0;
 935
 936         /*
 937          * reset file done count if we are to start over
 938          */
 939         if (td->o.time_based || td->o.loops)
 940                 td->nr_done_files = 0;
 941
 942         /*
 943          * Set the same seed to get repeatable runs
 944          */
 945         td_fill_rand_seeds(td);
 946 }
 947
 948 void reset_all_stats(struct thread_data *td)
 949 {
 950         struct timeval tv;
 951         int i;
 952
 953         reset_io_counters(td);
 954
 955         for (i = 0; i < 2; i++) {
 956                 td->io_bytes[i] = 0;
 957                 td->io_blocks[i] = 0;
 958                 td->io_issues[i] = 0;
 959                 td->ts.total_io_u[i] = 0;
 960         }
 961
 962         fio_gettime(&tv, NULL);
 963         memcpy(&td->epoch, &tv, sizeof(tv));
 964         memcpy(&td->start, &tv, sizeof(tv));
 965 }
 966
 967 static void clear_io_state(struct thread_data *td)
 968 {
 969         struct fio_file *f;
 970         unsigned int i;
 971
 972         reset_io_counters(td);
 973
 974         close_files(td);
 975         for_each_file(td, f, i)
 976                 fio_file_clear_done(f);
 977 }
 978
 979 static int exec_string(const char *string)
 980 {
 981         int ret, newlen = strlen(string) + 1 + 8;
 982         char *str;
 983
 984         str = malloc(newlen);
 985         sprintf(str, "sh -c %s", string);
 986
 987         ret = system(str);
 988         if (ret == -1)
 989                 log_err("fio: exec of cmd <%s> failed\n", str);
 990
 991         free(str);
 992         return ret;
 993 }
 994
 995 /*
 996  * Entry point for the thread based jobs. The process based jobs end up
 997  * here as well, after a little setup.
 998  */
 999 static void *thread_main(void *data)
1000 {
1001         unsigned long long runtime[2], elapsed;
1002         struct thread_data *td = data;
1003         pthread_condattr_t attr;
1004         int clear_state;
1005
1006         if (!td->o.use_thread)
1007                 setsid();
1008
1009         td->pid = getpid();
1010
1011         dprint(FD_PROCESS, "jobs pid=%d started\n", (int) td->pid);
1012
1013         INIT_FLIST_HEAD(&td->io_u_freelist);
1014         INIT_FLIST_HEAD(&td->io_u_busylist);
1015         INIT_FLIST_HEAD(&td->io_u_requeues);
1016         INIT_FLIST_HEAD(&td->io_log_list);
1017         INIT_FLIST_HEAD(&td->io_hist_list);
1018         INIT_FLIST_HEAD(&td->verify_list);
1019         pthread_mutex_init(&td->io_u_lock, NULL);
1020         td->io_hist_tree = RB_ROOT;
1021
1022         pthread_condattr_init(&attr);
1023         pthread_cond_init(&td->verify_cond, &attr);
1024         pthread_cond_init(&td->free_cond, &attr);
1025
1026         td_set_runstate(td, TD_INITIALIZED);
1027         dprint(FD_MUTEX, "up startup_mutex\n");
1028         fio_mutex_up(startup_mutex);
1029         dprint(FD_MUTEX, "wait on td->mutex\n");
1030         fio_mutex_down(td->mutex);
1031         dprint(FD_MUTEX, "done waiting on td->mutex\n");
1032
1033         /*
1034          * the ->mutex mutex is now no longer used, close it to avoid
1035          * eating a file descriptor
1036          */
1037         fio_mutex_remove(td->mutex);
1038
1039         /*
1040          * May alter parameters that init_io_u() will use, so we need to
1041          * do this first.
1042          */
1043         if (init_iolog(td))
1044                 goto err;
1045
1046         if (init_io_u(td))
1047                 goto err;
1048
1049         if (td->o.verify_async && verify_async_init(td))
1050                 goto err;
1051
1052         if (td->o.cpumask_set && fio_setaffinity(td->pid, td->o.cpumask) == -1) {
1053                 td_verror(td, errno, "cpu_set_affinity");
1054                 goto err;
1055         }
1056
1057         /*
1058          * If we have a gettimeofday() thread, make sure we exclude that
1059          * thread from this job
1060          */
1061         if (td->o.gtod_cpu) {
1062                 fio_cpu_clear(&td->o.cpumask, td->o.gtod_cpu);
1063                 if (fio_setaffinity(td->pid, td->o.cpumask) == -1) {
1064                         td_verror(td, errno, "cpu_set_affinity");
1065                         goto err;
1066                 }
1067         }
1068
1069         if (td->ioprio_set) {
1070                 if (ioprio_set(IOPRIO_WHO_PROCESS, 0, td->ioprio) == -1) {
1071                         td_verror(td, errno, "ioprio_set");
1072                         goto err;
1073                 }
1074         }
1075
1076         if (nice(td->o.nice) == -1) {
1077                 td_verror(td, errno, "nice");
1078                 goto err;
1079         }
1080
1081         if (td->o.ioscheduler && switch_ioscheduler(td))
1082                 goto err;
1083
1084         if (!td->o.create_serialize && setup_files(td))
1085                 goto err;
1086
1087         if (td_io_init(td))
1088                 goto err;
1089
1090         if (init_random_map(td))
1091                 goto err;
1092
1093         if (td->o.exec_prerun) {
1094                 if (exec_string(td->o.exec_prerun))
1095                         goto err;
1096         }
1097
1098         if (td->o.pre_read) {
1099                 if (pre_read_files(td) < 0)
1100                         goto err;
1101         }
1102
1103         fio_gettime(&td->epoch, NULL);
1104         getrusage(RUSAGE_SELF, &td->ts.ru_start);
1105
1106         runtime[0] = runtime[1] = 0;
1107         clear_state = 0;
1108         while (keep_running(td)) {
1109                 fio_gettime(&td->start, NULL);
1110                 memcpy(&td->ts.stat_sample_time[0], &td->start,
1111                                 sizeof(td->start));
1112                 memcpy(&td->ts.stat_sample_time[1], &td->start,
1113                                 sizeof(td->start));
1114                 memcpy(&td->tv_cache, &td->start, sizeof(td->start));
1115
1116                 if (td->o.ratemin[0] || td->o.ratemin[1])
1117                         memcpy(&td->lastrate, &td->ts.stat_sample_time,
1118                                                         sizeof(td->lastrate));
1119
1120                 if (clear_state)
1121                         clear_io_state(td);
1122
1123                 prune_io_piece_log(td);
1124
1125                 do_io(td);
1126
1127                 clear_state = 1;
1128
1129                 if (td_read(td) && td->io_bytes[DDIR_READ]) {
1130                         elapsed = utime_since_now(&td->start);
1131                         runtime[DDIR_READ] += elapsed;
1132                 }
1133                 if (td_write(td) && td->io_bytes[DDIR_WRITE]) {
1134                         elapsed = utime_since_now(&td->start);
1135                         runtime[DDIR_WRITE] += elapsed;
1136                 }
1137
1138                 if (td->error || td->terminate)
1139                         break;
1140
1141                 if (!td->o.do_verify ||
1142                     td->o.verify == VERIFY_NONE ||
1143                     (td->io_ops->flags & FIO_UNIDIR))
1144                         continue;
1145
1146                 clear_io_state(td);
1147
1148                 fio_gettime(&td->start, NULL);
1149
1150                 do_verify(td);
1151
1152                 runtime[DDIR_READ] += utime_since_now(&td->start);
1153
1154                 if (td->error || td->terminate)
1155                         break;
1156         }
1157
1158         update_rusage_stat(td);
1159         td->ts.runtime[0] = (runtime[0] + 999) / 1000;
1160         td->ts.runtime[1] = (runtime[1] + 999) / 1000;
1161         td->ts.total_run_time = mtime_since_now(&td->epoch);
1162         td->ts.io_bytes[0] = td->io_bytes[0];
1163         td->ts.io_bytes[1] = td->io_bytes[1];
1164
1165         fio_mutex_down(writeout_mutex);
1166         if (td->ts.bw_log) {
1167                 if (td->o.bw_log_file) {
1168                         finish_log_named(td, td->ts.bw_log,
1169                                                 td->o.bw_log_file, "bw");
1170                 } else
1171                         finish_log(td, td->ts.bw_log, "bw");
1172         }
1173         if (td->ts.slat_log) {
1174                 if (td->o.lat_log_file) {
1175                         finish_log_named(td, td->ts.slat_log,
1176                                                 td->o.lat_log_file, "slat");
1177                 } else
1178                         finish_log(td, td->ts.slat_log, "slat");
1179         }
1180         if (td->ts.clat_log) {
1181                 if (td->o.lat_log_file) {
1182                         finish_log_named(td, td->ts.clat_log,
1183                                                 td->o.lat_log_file, "clat");
1184                 } else
1185                         finish_log(td, td->ts.clat_log, "clat");
1186         }
1187         fio_mutex_up(writeout_mutex);
1188         if (td->o.exec_postrun)
1189                 exec_string(td->o.exec_postrun);
1190
1191         if (exitall_on_terminate)
1192                 terminate_threads(td->groupid);
1193
1194 err:
1195         if (td->error)
1196                 printf("fio: pid=%d, err=%d/%s\n", (int) td->pid, td->error,
1197                                                         td->verror);
1198         close_and_free_files(td);
1199         close_ioengine(td);
1200         cleanup_io_u(td);
1201
1202         if (td->o.cpumask_set) {
1203                 int ret = fio_cpuset_exit(&td->o.cpumask);
1204
1205                 td_verror(td, ret, "fio_cpuset_exit");
1206         }
1207
1208         if (td->o.verify_async)
1209                 verify_async_exit(td);
1210
1211         /*
1212          * do this very late, it will log file closing as well
1213          */
1214         if (td->o.write_iolog_file)
1215                 write_iolog_close(td);
1216
1217         options_mem_free(td);
1218         td_set_runstate(td, TD_EXITED);
1219         return (void *) (unsigned long) td->error;
1220 }
1221
1222 /*
1223  * We cannot pass the td data into a forked process, so attach the td and
1224  * pass it to the thread worker.
1225  */
1226 static int fork_main(int shmid, int offset)
1227 {
1228         struct thread_data *td;
1229         void *data, *ret;
1230
1231         data = shmat(shmid, NULL, 0);
1232         if (data == (void *) -1) {
1233                 int __err = errno;
1234
1235                 perror("shmat");
1236                 return __err;
1237         }
1238
1239         td = data + offset * sizeof(struct thread_data);
1240         ret = thread_main(td);
1241         shmdt(data);
1242         return (int) (unsigned long) ret;
1243 }
1244
1245 /*
1246  * Run over the job map and reap the threads that have exited, if any.
1247  */
1248 static void reap_threads(int *nr_running, int *t_rate, int *m_rate)
1249 {
1250         struct thread_data *td;
1251         int i, cputhreads, realthreads, pending, status, ret;
1252
1253         /*
1254          * reap exited threads (TD_EXITED -> TD_REAPED)
1255          */
1256         realthreads = pending = cputhreads = 0;
1257         for_each_td(td, i) {
1258                 int flags = 0;
1259
1260                 /*
1261                  * ->io_ops is NULL for a thread that has closed its
1262                  * io engine
1263                  */
1264                 if (td->io_ops && !strcmp(td->io_ops->name, "cpuio"))
1265                         cputhreads++;
1266                 else
1267                         realthreads++;
1268
1269                 if (!td->pid) {
1270                         pending++;
1271                         continue;
1272                 }
1273                 if (td->runstate == TD_REAPED)
1274                         continue;
1275                 if (td->o.use_thread) {
1276                         if (td->runstate == TD_EXITED) {
1277                                 td_set_runstate(td, TD_REAPED);
1278                                 goto reaped;
1279                         }
1280                         continue;
1281                 }
1282
1283                 flags = WNOHANG;
1284                 if (td->runstate == TD_EXITED)
1285                         flags = 0;
1286
1287                 /*
1288                  * check if someone quit or got killed in an unusual way
1289                  */
1290                 ret = waitpid(td->pid, &status, flags);
1291                 if (ret < 0) {
1292                         if (errno == ECHILD) {
1293                                 log_err("fio: pid=%d disappeared %d\n",
1294                                                 (int) td->pid, td->runstate);
1295                                 td_set_runstate(td, TD_REAPED);
1296                                 goto reaped;
1297                         }
1298                         perror("waitpid");
1299                 } else if (ret == td->pid) {
1300                         if (WIFSIGNALED(status)) {
1301                                 int sig = WTERMSIG(status);
1302
1303                                 if (sig != SIGQUIT)
1304                                         log_err("fio: pid=%d, got signal=%d\n",
1305                                                         (int) td->pid, sig);
1306                                 td_set_runstate(td, TD_REAPED);
1307                                 goto reaped;
1308                         }
1309                         if (WIFEXITED(status)) {
1310                                 if (WEXITSTATUS(status) && !td->error)
1311                                         td->error = WEXITSTATUS(status);
1312
1313                                 td_set_runstate(td, TD_REAPED);
1314                                 goto reaped;
1315                         }
1316                 }
1317
1318                 /*
1319                  * thread is not dead, continue
1320                  */
1321                 pending++;
1322                 continue;
1323 reaped:
1324                 (*nr_running)--;
1325                 (*m_rate) -= (td->o.ratemin[0] + td->o.ratemin[1]);
1326                 (*t_rate) -= (td->o.rate[0] + td->o.rate[1]);
1327                 if (!td->pid)
1328                         pending--;
1329
1330                 if (td->error)
1331                         exit_value++;
1332
1333                 done_secs += mtime_since_now(&td->epoch) / 1000;
1334         }
1335
1336         if (*nr_running == cputhreads && !pending && realthreads)
1337                 terminate_threads(TERMINATE_ALL);
1338 }
1339
1340 static void *gtod_thread_main(void *data)
1341 {
1342         fio_mutex_up(startup_mutex);
1343
1344         /*
1345          * As long as we have jobs around, update the clock. It would be nice
1346          * to have some way of NOT hammering that CPU with gettimeofday(),
1347          * but I'm not sure what to use outside of a simple CPU nop to relax
1348          * it - we don't want to lose precision.
1349          */
1350         while (threads) {
1351                 fio_gtod_update();
1352                 nop;
1353         }
1354
1355         return NULL;
1356 }
1357
1358 static int fio_start_gtod_thread(void)
1359 {
1360         int ret;
1361
1362         ret = pthread_create(&gtod_thread, NULL, gtod_thread_main, NULL);
1363         if (ret) {
1364                 log_err("Can't create gtod thread: %s\n", strerror(ret));
1365                 return 1;
1366         }
1367
1368         ret = pthread_detach(gtod_thread);
1369         if (ret) {
1370                 log_err("Can't detatch gtod thread: %s\n", strerror(ret));
1371                 return 1;
1372         }
1373
1374         dprint(FD_MUTEX, "wait on startup_mutex\n");
1375         fio_mutex_down(startup_mutex);
1376         dprint(FD_MUTEX, "done waiting on startup_mutex\n");
1377         return 0;
1378 }
1379
1380 /*
1381  * Main function for kicking off and reaping jobs, as needed.
1382  */
1383 static void run_threads(void)
1384 {
1385         struct thread_data *td;
1386         unsigned long spent;
1387         int i, todo, nr_running, m_rate, t_rate, nr_started;
1388
1389         if (fio_pin_memory())
1390                 return;
1391
1392         if (fio_gtod_offload && fio_start_gtod_thread())
1393                 return;
1394
1395         if (!terse_output) {
1396                 printf("Starting ");
1397                 if (nr_thread)
1398                         printf("%d thread%s", nr_thread,
1399                                                 nr_thread > 1 ? "s" : "");
1400                 if (nr_process) {
1401                         if (nr_thread)
1402                                 printf(" and ");
1403                         printf("%d process%s", nr_process,
1404                                                 nr_process > 1 ? "es" : "");
1405                 }
1406                 printf("\n");
1407                 fflush(stdout);
1408         }
1409
1410         set_sig_handlers();
1411
1412         todo = thread_number;
1413         nr_running = 0;
1414         nr_started = 0;
1415         m_rate = t_rate = 0;
1416
1417         for_each_td(td, i) {
1418                 print_status_init(td->thread_number - 1);
1419
1420                 if (!td->o.create_serialize) {
1421                         init_disk_util(td);
1422                         continue;
1423                 }
1424
1425                 /*
1426                  * do file setup here so it happens sequentially,
1427                  * we don't want X number of threads getting their
1428                  * client data interspersed on disk
1429                  */
1430                 if (setup_files(td)) {
1431                         exit_value++;
1432                         if (td->error)
1433                                 log_err("fio: pid=%d, err=%d/%s\n",
1434                                         (int) td->pid, td->error, td->verror);
1435                         td_set_runstate(td, TD_REAPED);
1436                         todo--;
1437                 } else {
1438                         struct fio_file *f;
1439                         unsigned int i;
1440
1441                         /*
1442                          * for sharing to work, each job must always open
1443                          * its own files. so close them, if we opened them
1444                          * for creation
1445                          */
1446                         for_each_file(td, f, i) {
1447                                 if (fio_file_open(f))
1448                                         td_io_close_file(td, f);
1449                         }
1450                 }
1451
1452                 init_disk_util(td);
1453         }
1454
1455         set_genesis_time();
1456
1457         while (todo) {
1458                 struct thread_data *map[MAX_JOBS];
1459                 struct timeval this_start;
1460                 int this_jobs = 0, left;
1461
1462                 /*
1463                  * create threads (TD_NOT_CREATED -> TD_CREATED)
1464                  */
1465                 for_each_td(td, i) {
1466                         if (td->runstate != TD_NOT_CREATED)
1467                                 continue;
1468
1469                         /*
1470                          * never got a chance to start, killed by other
1471                          * thread for some reason
1472                          */
1473                         if (td->terminate) {
1474                                 todo--;
1475                                 continue;
1476                         }
1477
1478                         if (td->o.start_delay) {
1479                                 spent = mtime_since_genesis();
1480
1481                                 if (td->o.start_delay * 1000 > spent)
1482                                         continue;
1483                         }
1484
1485                         if (td->o.stonewall && (nr_started || nr_running)) {
1486                                 dprint(FD_PROCESS, "%s: stonewall wait\n",
1487                                                         td->o.name);
1488                                 break;
1489                         }
1490
1491                         /*
1492                          * Set state to created. Thread will transition
1493                          * to TD_INITIALIZED when it's done setting up.
1494                          */
1495                         td_set_runstate(td, TD_CREATED);
1496                         map[this_jobs++] = td;
1497                         nr_started++;
1498
1499                         if (td->o.use_thread) {
1500                                 int ret;
1501
1502                                 dprint(FD_PROCESS, "will pthread_create\n");
1503                                 ret = pthread_create(&td->thread, NULL,
1504                                                         thread_main, td);
1505                                 if (ret) {
1506                                         log_err("pthread_create: %s\n",
1507                                                         strerror(ret));
1508                                         nr_started--;
1509                                         break;
1510                                 }
1511                                 ret = pthread_detach(td->thread);
1512                                 if (ret)
1513                                         log_err("pthread_detach: %s",
1514                                                         strerror(ret));
1515                         } else {
1516                                 pid_t pid;
1517                                 dprint(FD_PROCESS, "will fork\n");
1518                                 pid = fork();
1519                                 if (!pid) {
1520                                         int ret = fork_main(shm_id, i);
1521
1522                                         _exit(ret);
1523                                 } else if (i == fio_debug_jobno)
1524                                         *fio_debug_jobp = pid;
1525                         }
1526                         dprint(FD_MUTEX, "wait on startup_mutex\n");
1527                         if (fio_mutex_down_timeout(startup_mutex, 10)) {
1528                                 log_err("fio: job startup hung? exiting.\n");
1529                                 terminate_threads(TERMINATE_ALL);
1530                                 fio_abort = 1;
1531                                 nr_started--;
1532                                 break;
1533                         }
1534                         dprint(FD_MUTEX, "done waiting on startup_mutex\n");
1535                 }
1536
1537                 /*
1538                  * Wait for the started threads to transition to
1539                  * TD_INITIALIZED.
1540                  */
1541                 fio_gettime(&this_start, NULL);
1542                 left = this_jobs;
1543                 while (left && !fio_abort) {
1544                         if (mtime_since_now(&this_start) > JOB_START_TIMEOUT)
1545                                 break;
1546
1547                         usleep(100000);
1548
1549                         for (i = 0; i < this_jobs; i++) {
1550                                 td = map[i];
1551                                 if (!td)
1552                                         continue;
1553                                 if (td->runstate == TD_INITIALIZED) {
1554                                         map[i] = NULL;
1555                                         left--;
1556                                 } else if (td->runstate >= TD_EXITED) {
1557                                         map[i] = NULL;
1558                                         left--;
1559                                         todo--;
1560                                         nr_running++; /* work-around... */
1561                                 }
1562                         }
1563                 }
1564
1565                 if (left) {
1566                         log_err("fio: %d jobs failed to start\n", left);
1567                         for (i = 0; i < this_jobs; i++) {
1568                                 td = map[i];
1569                                 if (!td)
1570                                         continue;
1571                                 kill(td->pid, SIGTERM);
1572                         }
1573                         break;
1574                 }
1575
1576                 /*
1577                  * start created threads (TD_INITIALIZED -> TD_RUNNING).
1578                  */
1579                 for_each_td(td, i) {
1580                         if (td->runstate != TD_INITIALIZED)
1581                                 continue;
1582
1583                         if (in_ramp_time(td))
1584                                 td_set_runstate(td, TD_RAMP);
1585                         else
1586                                 td_set_runstate(td, TD_RUNNING);
1587                         nr_running++;
1588                         nr_started--;
1589                         m_rate += td->o.ratemin[0] + td->o.ratemin[1];
1590                         t_rate += td->o.rate[0] + td->o.rate[1];
1591                         todo--;
1592                         fio_mutex_up(td->mutex);
1593                 }
1594
1595                 reap_threads(&nr_running, &t_rate, &m_rate);
1596
1597                 if (todo)
1598                         usleep(100000);
1599         }
1600
1601         while (nr_running) {
1602                 reap_threads(&nr_running, &t_rate, &m_rate);
1603                 usleep(10000);
1604         }
1605
1606         update_io_ticks();
1607         fio_unpin_memory();
1608 }
1609
1610 int main(int argc, char *argv[])
1611 {
1612         long ps;
1613
1614         sinit();
1615
1616         /*
1617          * We need locale for number printing, if it isn't set then just
1618          * go with the US format.
1619          */
1620         if (!getenv("LC_NUMERIC"))
1621                 setlocale(LC_NUMERIC, "en_US");
1622
1623         ps = sysconf(_SC_PAGESIZE);
1624         if (ps < 0) {
1625                 log_err("Failed to get page size\n");
1626                 return 1;
1627         }
1628
1629         page_size = ps;
1630         page_mask = ps - 1;
1631
1632         fio_keywords_init();
1633
1634         if (parse_options(argc, argv))
1635                 return 1;
1636
1637         if (!thread_number)
1638                 return 0;
1639
1640         if (write_bw_log) {
1641                 setup_log(&agg_io_log[DDIR_READ]);
1642                 setup_log(&agg_io_log[DDIR_WRITE]);
1643         }
1644
1645         startup_mutex = fio_mutex_init(0);
1646         writeout_mutex = fio_mutex_init(1);
1647
1648         set_genesis_time();
1649
1650         status_timer_arm();
1651
1652         run_threads();
1653
1654         if (!fio_abort) {
1655                 show_run_stats();
1656                 if (write_bw_log) {
1657                         __finish_log(agg_io_log[DDIR_READ], "agg-read_bw.log");
1658                         __finish_log(agg_io_log[DDIR_WRITE],
1659                                         "agg-write_bw.log");
1660                 }
1661         }
1662
1663         fio_mutex_remove(startup_mutex);
1664         fio_mutex_remove(writeout_mutex);
1665         return exit_value;
1666 }