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