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 #include "cgroup.h"
  43 #include "profile.h"
  44
  45 unsigned long page_mask;
  46 unsigned long page_size;
  47
  48 #define PAGE_ALIGN(buf) \
  49         (char *) (((unsigned long) (buf) + page_mask) & ~page_mask)
  50
  51 int groupid = 0;
  52 int thread_number = 0;
  53 int nr_process = 0;
  54 int nr_thread = 0;
  55 int shm_id = 0;
  56 int temp_stall_ts;
  57 unsigned long done_secs = 0;
  58
  59 static struct fio_mutex *startup_mutex;
  60 static struct fio_mutex *writeout_mutex;
  61 static volatile int fio_abort;
  62 static int exit_value;
  63 static struct itimerval itimer;
  64 static pthread_t gtod_thread;
  65 static struct flist_head *cgroup_list;
  66 static char *cgroup_mnt;
  67
  68 struct io_log *agg_io_log[2];
  69
  70 #define TERMINATE_ALL           (-1)
  71 #define JOB_START_TIMEOUT       (5 * 1000)
  72
  73 void td_set_runstate(struct thread_data *td, int runstate)
  74 {
  75         if (td->runstate == runstate)
  76                 return;
  77
  78         dprint(FD_PROCESS, "pid=%d: runstate %d -> %d\n", (int) td->pid,
  79                                                 td->runstate, runstate);
  80         td->runstate = runstate;
  81 }
  82
  83 static void terminate_threads(int group_id)
  84 {
  85         struct thread_data *td;
  86         int i;
  87
  88         dprint(FD_PROCESS, "terminate group_id=%d\n", group_id);
  89
  90         for_each_td(td, i) {
  91                 if (group_id == TERMINATE_ALL || groupid == td->groupid) {
  92                         dprint(FD_PROCESS, "setting terminate on %s/%d\n",
  93                                                 td->o.name, (int) td->pid);
  94                         td->terminate = 1;
  95                         td->o.start_delay = 0;
  96
  97                         /*
  98                          * if the thread is running, just let it exit
  99                          */
 100                         if (td->runstate < TD_RUNNING)
 101                                 kill(td->pid, SIGQUIT);
 102                         else {
 103                                 struct ioengine_ops *ops = td->io_ops;
 104
 105                                 if (ops && (ops->flags & FIO_SIGQUIT))
 106                                         kill(td->pid, SIGQUIT);
 107                         }
 108                 }
 109         }
 110 }
 111
 112 static void status_timer_arm(void)
 113 {
 114         itimer.it_value.tv_sec = 0;
 115         itimer.it_value.tv_usec = DISK_UTIL_MSEC * 1000;
 116         setitimer(ITIMER_REAL, &itimer, NULL);
 117 }
 118
 119 static void sig_alrm(int fio_unused sig)
 120 {
 121         if (threads) {
 122                 update_io_ticks();
 123                 print_thread_status();
 124                 status_timer_arm();
 125         }
 126 }
 127
 128 /*
 129  * Happens on thread runs with ctrl-c, ignore our own SIGQUIT
 130  */
 131 static void sig_quit(int sig)
 132 {
 133 }
 134
 135 static void sig_int(int sig)
 136 {
 137         if (threads) {
 138                 log_info("\nfio: terminating on signal %d\n", sig);
 139                 fflush(stdout);
 140                 terminate_threads(TERMINATE_ALL);
 141         }
 142 }
 143
 144 static void sig_ill(int fio_unused sig)
 145 {
 146         if (!threads)
 147                 return;
 148
 149         log_err("fio: illegal instruction. your cpu does not support "
 150                 "the sse4.2 instruction for crc32c\n");
 151         terminate_threads(TERMINATE_ALL);
 152         exit(4);
 153 }
 154
 155 static void set_sig_handlers(void)
 156 {
 157         struct sigaction act;
 158
 159         memset(&act, 0, sizeof(act));
 160         act.sa_handler = sig_alrm;
 161         act.sa_flags = SA_RESTART;
 162         sigaction(SIGALRM, &act, NULL);
 163
 164         memset(&act, 0, sizeof(act));
 165         act.sa_handler = sig_int;
 166         act.sa_flags = SA_RESTART;
 167         sigaction(SIGINT, &act, NULL);
 168
 169         memset(&act, 0, sizeof(act));
 170         act.sa_handler = sig_ill;
 171         act.sa_flags = SA_RESTART;
 172         sigaction(SIGILL, &act, NULL);
 173
 174         memset(&act, 0, sizeof(act));
 175         act.sa_handler = sig_quit;
 176         act.sa_flags = SA_RESTART;
 177         sigaction(SIGQUIT, &act, NULL);
 178 }
 179
 180 /*
 181  * Check if we are above the minimum rate given.
 182  */
 183 static int __check_min_rate(struct thread_data *td, struct timeval *now,
 184                             enum td_ddir ddir)
 185 {
 186         unsigned long long bytes = 0;
 187         unsigned long iops = 0;
 188         unsigned long spent;
 189         unsigned long rate;
 190         unsigned int ratemin = 0;
 191         unsigned int rate_iops = 0;
 192         unsigned int rate_iops_min = 0;
 193
 194         if (!td->o.ratemin[ddir] && !td->o.rate_iops_min[ddir])
 195                 return 0;
 196
 197         /*
 198          * allow a 2 second settle period in the beginning
 199          */
 200         if (mtime_since(&td->start, now) < 2000)
 201                 return 0;
 202
 203         iops += td->io_blocks[ddir];
 204         bytes += td->this_io_bytes[ddir];
 205         ratemin += td->o.ratemin[ddir];
 206         rate_iops += td->o.rate_iops[ddir];
 207         rate_iops_min += td->o.rate_iops_min[ddir];
 208
 209         /*
 210          * if rate blocks is set, sample is running
 211          */
 212         if (td->rate_bytes[ddir] || td->rate_blocks[ddir]) {
 213                 spent = mtime_since(&td->lastrate[ddir], now);
 214                 if (spent < td->o.ratecycle)
 215                         return 0;
 216
 217                 if (td->o.rate[ddir]) {
 218                         /*
 219                          * check bandwidth specified rate
 220                          */
 221                         if (bytes < td->rate_bytes[ddir]) {
 222                                 log_err("%s: min rate %u not met\n", td->o.name,
 223                                                                 ratemin);
 224                                 return 1;
 225                         } else {
 226                                 rate = ((bytes - td->rate_bytes[ddir]) * 1000) / spent;
 227                                 if (rate < ratemin ||
 228                                     bytes < td->rate_bytes[ddir]) {
 229                                         log_err("%s: min rate %u not met, got"
 230                                                 " %luKB/sec\n", td->o.name,
 231                                                         ratemin, rate);
 232                                         return 1;
 233                                 }
 234                         }
 235                 } else {
 236                         /*
 237                          * checks iops specified rate
 238                          */
 239                         if (iops < rate_iops) {
 240                                 log_err("%s: min iops rate %u not met\n",
 241                                                 td->o.name, rate_iops);
 242                                 return 1;
 243                         } else {
 244                                 rate = ((iops - td->rate_blocks[ddir]) * 1000) / spent;
 245                                 if (rate < rate_iops_min ||
 246                                     iops < td->rate_blocks[ddir]) {
 247                                         log_err("%s: min iops rate %u not met,"
 248                                                 " got %lu\n", td->o.name,
 249                                                         rate_iops_min, rate);
 250                                 }
 251                         }
 252                 }
 253         }
 254
 255         td->rate_bytes[ddir] = bytes;
 256         td->rate_blocks[ddir] = iops;
 257         memcpy(&td->lastrate[ddir], now, sizeof(*now));
 258         return 0;
 259 }
 260
 261 static int check_min_rate(struct thread_data *td, struct timeval *now,
 262                           unsigned long *bytes_done)
 263 {
 264         int ret = 0;
 265
 266         if (bytes_done[0])
 267                 ret |= __check_min_rate(td, now, 0);
 268         if (bytes_done[1])
 269                 ret |= __check_min_rate(td, now, 1);
 270
 271         return ret;
 272 }
 273
 274 static inline int runtime_exceeded(struct thread_data *td, struct timeval *t)
 275 {
 276         if (!td->o.timeout)
 277                 return 0;
 278         if (mtime_since(&td->epoch, t) >= td->o.timeout * 1000)
 279                 return 1;
 280
 281         return 0;
 282 }
 283
 284 /*
 285  * When job exits, we can cancel the in-flight IO if we are using async
 286  * io. Attempt to do so.
 287  */
 288 static void cleanup_pending_aio(struct thread_data *td)
 289 {
 290         struct flist_head *entry, *n;
 291         struct io_u *io_u;
 292         int r;
 293
 294         /*
 295          * get immediately available events, if any
 296          */
 297         r = io_u_queued_complete(td, 0, NULL);
 298         if (r < 0)
 299                 return;
 300
 301         /*
 302          * now cancel remaining active events
 303          */
 304         if (td->io_ops->cancel) {
 305                 flist_for_each_safe(entry, n, &td->io_u_busylist) {
 306                         io_u = flist_entry(entry, struct io_u, list);
 307
 308                         /*
 309                          * if the io_u isn't in flight, then that generally
 310                          * means someone leaked an io_u. complain but fix
 311                          * it up, so we don't stall here.
 312                          */
 313                         if ((io_u->flags & IO_U_F_FLIGHT) == 0) {
 314                                 log_err("fio: non-busy IO on busy list\n");
 315                                 put_io_u(td, io_u);
 316                         } else {
 317                                 r = td->io_ops->cancel(td, io_u);
 318                                 if (!r)
 319                                         put_io_u(td, io_u);
 320                         }
 321                 }
 322         }
 323
 324         if (td->cur_depth)
 325                 r = io_u_queued_complete(td, td->cur_depth, NULL);
 326 }
 327
 328 /*
 329  * Helper to handle the final sync of a file. Works just like the normal
 330  * io path, just does everything sync.
 331  */
 332 static int fio_io_sync(struct thread_data *td, struct fio_file *f)
 333 {
 334         struct io_u *io_u = __get_io_u(td);
 335         int ret;
 336
 337         if (!io_u)
 338                 return 1;
 339
 340         io_u->ddir = DDIR_SYNC;
 341         io_u->file = f;
 342
 343         if (td_io_prep(td, io_u)) {
 344                 put_io_u(td, io_u);
 345                 return 1;
 346         }
 347
 348 requeue:
 349         ret = td_io_queue(td, io_u);
 350         if (ret < 0) {
 351                 td_verror(td, io_u->error, "td_io_queue");
 352                 put_io_u(td, io_u);
 353                 return 1;
 354         } else if (ret == FIO_Q_QUEUED) {
 355                 if (io_u_queued_complete(td, 1, NULL) < 0)
 356                         return 1;
 357         } else if (ret == FIO_Q_COMPLETED) {
 358                 if (io_u->error) {
 359                         td_verror(td, io_u->error, "td_io_queue");
 360                         return 1;
 361                 }
 362
 363                 if (io_u_sync_complete(td, io_u, NULL) < 0)
 364                         return 1;
 365         } else if (ret == FIO_Q_BUSY) {
 366                 if (td_io_commit(td))
 367                         return 1;
 368                 goto requeue;
 369         }
 370
 371         return 0;
 372 }
 373
 374 static inline void update_tv_cache(struct thread_data *td)
 375 {
 376         if ((++td->tv_cache_nr & td->tv_cache_mask) == td->tv_cache_mask)
 377                 fio_gettime(&td->tv_cache, NULL);
 378 }
 379
 380 static int break_on_this_error(struct thread_data *td, int *retptr)
 381 {
 382         int ret = *retptr;
 383
 384         if (ret < 0 || td->error) {
 385                 int err;
 386
 387                 if (!td->o.continue_on_error)
 388                         return 1;
 389
 390                 if (ret < 0)
 391                         err = -ret;
 392                 else
 393                         err = td->error;
 394
 395                 if (td_non_fatal_error(err)) {
 396                         /*
 397                          * Continue with the I/Os in case of
 398                          * a non fatal error.
 399                          */
 400                         update_error_count(td, err);
 401                         td_clear_error(td);
 402                         *retptr = 0;
 403                         return 0;
 404                 } else if (td->o.fill_device && err == ENOSPC) {
 405                         /*
 406                          * We expect to hit this error if
 407                          * fill_device option is set.
 408                          */
 409                         td_clear_error(td);
 410                         td->terminate = 1;
 411                         return 1;
 412                 } else {
 413                         /*
 414                          * Stop the I/O in case of a fatal
 415                          * error.
 416                          */
 417                         update_error_count(td, err);
 418                         return 1;
 419                 }
 420         }
 421
 422         return 0;
 423 }
 424
 425 /*
 426  * The main verify engine. Runs over the writes we previously submitted,
 427  * reads the blocks back in, and checks the crc/md5 of the data.
 428  */
 429 static void do_verify(struct thread_data *td)
 430 {
 431         struct fio_file *f;
 432         struct io_u *io_u;
 433         int ret, min_events;
 434         unsigned int i;
 435
 436         dprint(FD_VERIFY, "starting loop\n");
 437
 438         /*
 439          * sync io first and invalidate cache, to make sure we really
 440          * read from disk.
 441          */
 442         for_each_file(td, f, i) {
 443                 if (!fio_file_open(f))
 444                         continue;
 445                 if (fio_io_sync(td, f))
 446                         break;
 447                 if (file_invalidate_cache(td, f))
 448                         break;
 449         }
 450
 451         if (td->error)
 452                 return;
 453
 454         td_set_runstate(td, TD_VERIFYING);
 455
 456         io_u = NULL;
 457         while (!td->terminate) {
 458                 int ret2, full;
 459
 460                 update_tv_cache(td);
 461
 462                 if (runtime_exceeded(td, &td->tv_cache)) {
 463                         td->terminate = 1;
 464                         break;
 465                 }
 466
 467                 io_u = __get_io_u(td);
 468                 if (!io_u)
 469                         break;
 470
 471                 if (get_next_verify(td, io_u)) {
 472                         put_io_u(td, io_u);
 473                         break;
 474                 }
 475
 476                 if (td_io_prep(td, io_u)) {
 477                         put_io_u(td, io_u);
 478                         break;
 479                 }
 480
 481                 if (td->o.verify_async)
 482                         io_u->end_io = verify_io_u_async;
 483                 else
 484                         io_u->end_io = verify_io_u;
 485
 486                 ret = td_io_queue(td, io_u);
 487                 switch (ret) {
 488                 case FIO_Q_COMPLETED:
 489                         if (io_u->error) {
 490                                 ret = -io_u->error;
 491                                 clear_io_u(td, io_u);
 492                         } else if (io_u->resid) {
 493                                 int bytes = io_u->xfer_buflen - io_u->resid;
 494                                 struct fio_file *f = io_u->file;
 495
 496                                 /*
 497                                  * zero read, fail
 498                                  */
 499                                 if (!bytes) {
 500                                         td_verror(td, EIO, "full resid");
 501                                         put_io_u(td, io_u);
 502                                         break;
 503                                 }
 504
 505                                 io_u->xfer_buflen = io_u->resid;
 506                                 io_u->xfer_buf += bytes;
 507                                 io_u->offset += bytes;
 508
 509                                 td->ts.short_io_u[io_u->ddir]++;
 510
 511                                 if (io_u->offset == f->real_file_size)
 512                                         goto sync_done;
 513
 514                                 requeue_io_u(td, &io_u);
 515                         } else {
 516 sync_done:
 517                                 ret = io_u_sync_complete(td, io_u, NULL);
 518                                 if (ret < 0)
 519                                         break;
 520                         }
 521                         continue;
 522                 case FIO_Q_QUEUED:
 523                         break;
 524                 case FIO_Q_BUSY:
 525                         requeue_io_u(td, &io_u);
 526                         ret2 = td_io_commit(td);
 527                         if (ret2 < 0)
 528                                 ret = ret2;
 529                         break;
 530                 default:
 531                         assert(ret < 0);
 532                         td_verror(td, -ret, "td_io_queue");
 533                         break;
 534                 }
 535
 536                 if (break_on_this_error(td, &ret))
 537                         break;
 538
 539                 /*
 540                  * if we can queue more, do so. but check if there are
 541                  * completed io_u's first.
 542                  */
 543                 full = queue_full(td) || ret == FIO_Q_BUSY;
 544                 if (full || !td->o.iodepth_batch_complete) {
 545                         min_events = min(td->o.iodepth_batch_complete,
 546                                          td->cur_depth);
 547                         if (full && !min_events)
 548                                 min_events = 1;
 549
 550                         do {
 551                                 /*
 552                                  * Reap required number of io units, if any,
 553                                  * and do the verification on them through
 554                                  * the callback handler
 555                                  */
 556                                 if (io_u_queued_complete(td, min_events, NULL) < 0) {
 557                                         ret = -1;
 558                                         break;
 559                                 }
 560                         } while (full && (td->cur_depth > td->o.iodepth_low));
 561                 }
 562                 if (ret < 0)
 563                         break;
 564         }
 565
 566         if (!td->error) {
 567                 min_events = td->cur_depth;
 568
 569                 if (min_events)
 570                         ret = io_u_queued_complete(td, min_events, NULL);
 571         } else
 572                 cleanup_pending_aio(td);
 573
 574         td_set_runstate(td, TD_RUNNING);
 575
 576         dprint(FD_VERIFY, "exiting loop\n");
 577 }
 578
 579 /*
 580  * Main IO worker function. It retrieves io_u's to process and queues
 581  * and reaps them, checking for rate and errors along the way.
 582  */
 583 static void do_io(struct thread_data *td)
 584 {
 585         unsigned int i;
 586         int ret = 0;
 587
 588         if (in_ramp_time(td))
 589                 td_set_runstate(td, TD_RAMP);
 590         else
 591                 td_set_runstate(td, TD_RUNNING);
 592
 593         while ( (td->o.read_iolog_file && !flist_empty(&td->io_log_list)) ||
 594                 ((td->this_io_bytes[0] + td->this_io_bytes[1]) < td->o.size) ) {
 595                 struct timeval comp_time;
 596                 unsigned long bytes_done[2] = { 0, 0 };
 597                 int min_evts = 0;
 598                 struct io_u *io_u;
 599                 int ret2, full;
 600
 601                 if (td->terminate)
 602                         break;
 603
 604                 update_tv_cache(td);
 605
 606                 if (runtime_exceeded(td, &td->tv_cache)) {
 607                         td->terminate = 1;
 608                         break;
 609                 }
 610
 611                 io_u = get_io_u(td);
 612                 if (!io_u)
 613                         break;
 614
 615                 /*
 616                  * Add verification end_io handler, if asked to verify
 617                  * a previously written file.
 618                  */
 619                 if (td->o.verify != VERIFY_NONE && io_u->ddir == DDIR_READ &&
 620                     !td_rw(td)) {
 621                         if (td->o.verify_async)
 622                                 io_u->end_io = verify_io_u_async;
 623                         else
 624                                 io_u->end_io = verify_io_u;
 625                         td_set_runstate(td, TD_VERIFYING);
 626                 } else if (in_ramp_time(td))
 627                         td_set_runstate(td, TD_RAMP);
 628                 else
 629                         td_set_runstate(td, TD_RUNNING);
 630
 631                 ret = td_io_queue(td, io_u);
 632                 switch (ret) {
 633                 case FIO_Q_COMPLETED:
 634                         if (io_u->error) {
 635                                 ret = -io_u->error;
 636                                 clear_io_u(td, io_u);
 637                         } else if (io_u->resid) {
 638                                 int bytes = io_u->xfer_buflen - io_u->resid;
 639                                 struct fio_file *f = io_u->file;
 640
 641                                 /*
 642                                  * zero read, fail
 643                                  */
 644                                 if (!bytes) {
 645                                         td_verror(td, EIO, "full resid");
 646                                         put_io_u(td, io_u);
 647                                         break;
 648                                 }
 649
 650                                 io_u->xfer_buflen = io_u->resid;
 651                                 io_u->xfer_buf += bytes;
 652                                 io_u->offset += bytes;
 653
 654                                 td->ts.short_io_u[io_u->ddir]++;
 655
 656                                 if (io_u->offset == f->real_file_size)
 657                                         goto sync_done;
 658
 659                                 requeue_io_u(td, &io_u);
 660                         } else {
 661 sync_done:
 662                                 if (__should_check_rate(td, 0) ||
 663                                     __should_check_rate(td, 1))
 664                                         fio_gettime(&comp_time, NULL);
 665
 666                                 ret = io_u_sync_complete(td, io_u, bytes_done);
 667                                 if (ret < 0)
 668                                         break;
 669                         }
 670                         break;
 671                 case FIO_Q_QUEUED:
 672                         /*
 673                          * if the engine doesn't have a commit hook,
 674                          * the io_u is really queued. if it does have such
 675                          * a hook, it has to call io_u_queued() itself.
 676                          */
 677                         if (td->io_ops->commit == NULL)
 678                                 io_u_queued(td, io_u);
 679                         break;
 680                 case FIO_Q_BUSY:
 681                         requeue_io_u(td, &io_u);
 682                         ret2 = td_io_commit(td);
 683                         if (ret2 < 0)
 684                                 ret = ret2;
 685                         break;
 686                 default:
 687                         assert(ret < 0);
 688                         put_io_u(td, io_u);
 689                         break;
 690                 }
 691
 692                 if (break_on_this_error(td, &ret))
 693                         break;
 694
 695                 /*
 696                  * See if we need to complete some commands
 697                  */
 698                 full = queue_full(td) || ret == FIO_Q_BUSY;
 699                 if (full || !td->o.iodepth_batch_complete) {
 700                         min_evts = min(td->o.iodepth_batch_complete,
 701                                         td->cur_depth);
 702                         if (full && !min_evts)
 703                                 min_evts = 1;
 704
 705                         if (__should_check_rate(td, 0) ||
 706                             __should_check_rate(td, 1))
 707                                 fio_gettime(&comp_time, NULL);
 708
 709                         do {
 710                                 ret = io_u_queued_complete(td, min_evts, bytes_done);
 711                                 if (ret < 0)
 712                                         break;
 713
 714                         } while (full && (td->cur_depth > td->o.iodepth_low));
 715                 }
 716
 717                 if (ret < 0)
 718                         break;
 719                 if (!(bytes_done[0] + bytes_done[1]))
 720                         continue;
 721
 722                 if (!in_ramp_time(td) && should_check_rate(td, bytes_done)) {
 723                         if (check_min_rate(td, &comp_time, bytes_done)) {
 724                                 if (exitall_on_terminate)
 725                                         terminate_threads(td->groupid);
 726                                 td_verror(td, EIO, "check_min_rate");
 727                                 break;
 728                         }
 729                 }
 730
 731                 if (td->o.thinktime) {
 732                         unsigned long long b;
 733
 734                         b = td->io_blocks[0] + td->io_blocks[1];
 735                         if (!(b % td->o.thinktime_blocks)) {
 736                                 int left;
 737
 738                                 if (td->o.thinktime_spin)
 739                                         usec_spin(td->o.thinktime_spin);
 740
 741                                 left = td->o.thinktime - td->o.thinktime_spin;
 742                                 if (left)
 743                                         usec_sleep(td, left);
 744                         }
 745                 }
 746         }
 747
 748         if (td->o.fill_device && td->error == ENOSPC) {
 749                 td->error = 0;
 750                 td->terminate = 1;
 751         }
 752         if (!td->error) {
 753                 struct fio_file *f;
 754
 755                 i = td->cur_depth;
 756                 if (i)
 757                         ret = io_u_queued_complete(td, i, NULL);
 758
 759                 if (should_fsync(td) && td->o.end_fsync) {
 760                         td_set_runstate(td, TD_FSYNCING);
 761
 762                         for_each_file(td, f, i) {
 763                                 if (!fio_file_open(f))
 764                                         continue;
 765                                 fio_io_sync(td, f);
 766                         }
 767                 }
 768         } else
 769                 cleanup_pending_aio(td);
 770
 771         /*
 772          * stop job if we failed doing any IO
 773          */
 774         if ((td->this_io_bytes[0] + td->this_io_bytes[1]) == 0)
 775                 td->done = 1;
 776 }
 777
 778 static void cleanup_io_u(struct thread_data *td)
 779 {
 780         struct flist_head *entry, *n;
 781         struct io_u *io_u;
 782
 783         flist_for_each_safe(entry, n, &td->io_u_freelist) {
 784                 io_u = flist_entry(entry, struct io_u, list);
 785
 786                 flist_del(&io_u->list);
 787                 free(io_u);
 788         }
 789
 790         free_io_mem(td);
 791 }
 792
 793 static int init_io_u(struct thread_data *td)
 794 {
 795         struct io_u *io_u;
 796         unsigned int max_bs;
 797         int cl_align, i, max_units;
 798         char *p;
 799
 800         max_units = td->o.iodepth;
 801         max_bs = max(td->o.max_bs[DDIR_READ], td->o.max_bs[DDIR_WRITE]);
 802         td->orig_buffer_size = (unsigned long long) max_bs
 803                                         * (unsigned long long) max_units;
 804
 805         if (td->o.mem_type == MEM_SHMHUGE || td->o.mem_type == MEM_MMAPHUGE) {
 806                 unsigned long bs;
 807
 808                 bs = td->orig_buffer_size + td->o.hugepage_size - 1;
 809                 td->orig_buffer_size = bs & ~(td->o.hugepage_size - 1);
 810         }
 811
 812         if (td->orig_buffer_size != (size_t) td->orig_buffer_size) {
 813                 log_err("fio: IO memory too large. Reduce max_bs or iodepth\n");
 814                 return 1;
 815         }
 816
 817         if (allocate_io_mem(td))
 818                 return 1;
 819
 820         if (td->o.odirect || td->o.mem_align)
 821                 p = PAGE_ALIGN(td->orig_buffer) + td->o.mem_align;
 822         else
 823                 p = td->orig_buffer;
 824
 825         cl_align = os_cache_line_size();
 826
 827         for (i = 0; i < max_units; i++) {
 828                 void *ptr;
 829
 830                 if (td->terminate)
 831                         return 1;
 832
 833                 if (posix_memalign(&ptr, cl_align, sizeof(*io_u))) {
 834                         log_err("fio: posix_memalign=%s\n", strerror(errno));
 835                         break;
 836                 }
 837
 838                 io_u = ptr;
 839                 memset(io_u, 0, sizeof(*io_u));
 840                 INIT_FLIST_HEAD(&io_u->list);
 841                 dprint(FD_MEM, "io_u alloc %p, index %u\n", io_u, i);
 842
 843                 if (!(td->io_ops->flags & FIO_NOIO)) {
 844                         io_u->buf = p + max_bs * i;
 845                         dprint(FD_MEM, "io_u %p, mem %p\n", io_u, io_u->buf);
 846
 847                         if (td_write(td) && !td->o.refill_buffers)
 848                                 io_u_fill_buffer(td, io_u, max_bs);
 849                 }
 850
 851                 io_u->index = i;
 852                 io_u->flags = IO_U_F_FREE;
 853                 flist_add(&io_u->list, &td->io_u_freelist);
 854         }
 855
 856         return 0;
 857 }
 858
 859 static int switch_ioscheduler(struct thread_data *td)
 860 {
 861         char tmp[256], tmp2[128];
 862         FILE *f;
 863         int ret;
 864
 865         if (td->io_ops->flags & FIO_DISKLESSIO)
 866                 return 0;
 867
 868         sprintf(tmp, "%s/queue/scheduler", td->sysfs_root);
 869
 870         f = fopen(tmp, "r+");
 871         if (!f) {
 872                 if (errno == ENOENT) {
 873                         log_err("fio: os or kernel doesn't support IO scheduler"
 874                                 " switching\n");
 875                         return 0;
 876                 }
 877                 td_verror(td, errno, "fopen iosched");
 878                 return 1;
 879         }
 880
 881         /*
 882          * Set io scheduler.
 883          */
 884         ret = fwrite(td->o.ioscheduler, strlen(td->o.ioscheduler), 1, f);
 885         if (ferror(f) || ret != 1) {
 886                 td_verror(td, errno, "fwrite");
 887                 fclose(f);
 888                 return 1;
 889         }
 890
 891         rewind(f);
 892
 893         /*
 894          * Read back and check that the selected scheduler is now the default.
 895          */
 896         ret = fread(tmp, 1, sizeof(tmp), f);
 897         if (ferror(f) || ret < 0) {
 898                 td_verror(td, errno, "fread");
 899                 fclose(f);
 900                 return 1;
 901         }
 902
 903         sprintf(tmp2, "[%s]", td->o.ioscheduler);
 904         if (!strstr(tmp, tmp2)) {
 905                 log_err("fio: io scheduler %s not found\n", td->o.ioscheduler);
 906                 td_verror(td, EINVAL, "iosched_switch");
 907                 fclose(f);
 908                 return 1;
 909         }
 910
 911         fclose(f);
 912         return 0;
 913 }
 914
 915 static int keep_running(struct thread_data *td)
 916 {
 917         unsigned long long io_done;
 918
 919         if (td->done)
 920                 return 0;
 921         if (td->o.time_based)
 922                 return 1;
 923         if (td->o.loops) {
 924                 td->o.loops--;
 925                 return 1;
 926         }
 927
 928         io_done = td->io_bytes[DDIR_READ] + td->io_bytes[DDIR_WRITE]
 929                         + td->io_skip_bytes;
 930         if (io_done < td->o.size)
 931                 return 1;
 932
 933         return 0;
 934 }
 935
 936 static void reset_io_counters(struct thread_data *td)
 937 {
 938         td->ts.stat_io_bytes[0] = td->ts.stat_io_bytes[1] = 0;
 939         td->this_io_bytes[0] = td->this_io_bytes[1] = 0;
 940         td->zone_bytes = 0;
 941         td->rate_bytes[0] = td->rate_bytes[1] = 0;
 942         td->rate_blocks[0] = td->rate_blocks[1] = 0;
 943
 944         td->last_was_sync = 0;
 945
 946         /*
 947          * reset file done count if we are to start over
 948          */
 949         if (td->o.time_based || td->o.loops)
 950                 td->nr_done_files = 0;
 951
 952         /*
 953          * Set the same seed to get repeatable runs
 954          */
 955         td_fill_rand_seeds(td);
 956 }
 957
 958 void reset_all_stats(struct thread_data *td)
 959 {
 960         struct timeval tv;
 961         int i;
 962
 963         reset_io_counters(td);
 964
 965         for (i = 0; i < 2; i++) {
 966                 td->io_bytes[i] = 0;
 967                 td->io_blocks[i] = 0;
 968                 td->io_issues[i] = 0;
 969                 td->ts.total_io_u[i] = 0;
 970         }
 971
 972         fio_gettime(&tv, NULL);
 973         memcpy(&td->epoch, &tv, sizeof(tv));
 974         memcpy(&td->start, &tv, sizeof(tv));
 975 }
 976
 977 static void clear_io_state(struct thread_data *td)
 978 {
 979         struct fio_file *f;
 980         unsigned int i;
 981
 982         reset_io_counters(td);
 983
 984         close_files(td);
 985         for_each_file(td, f, i)
 986                 fio_file_clear_done(f);
 987 }
 988
 989 static int exec_string(const char *string)
 990 {
 991         int ret, newlen = strlen(string) + 1 + 8;
 992         char *str;
 993
 994         str = malloc(newlen);
 995         sprintf(str, "sh -c %s", string);
 996
 997         ret = system(str);
 998         if (ret == -1)
 999                 log_err("fio: exec of cmd <%s> failed\n", str);
1000
1001         free(str);
1002         return ret;
1003 }
1004
1005 /*
1006  * Entry point for the thread based jobs. The process based jobs end up
1007  * here as well, after a little setup.
1008  */
1009 static void *thread_main(void *data)
1010 {
1011         unsigned long long runtime[2], elapsed;
1012         struct thread_data *td = data;
1013         pthread_condattr_t attr;
1014         int clear_state;
1015
1016         if (!td->o.use_thread)
1017                 setsid();
1018
1019         td->pid = getpid();
1020
1021         dprint(FD_PROCESS, "jobs pid=%d started\n", (int) td->pid);
1022
1023         INIT_FLIST_HEAD(&td->io_u_freelist);
1024         INIT_FLIST_HEAD(&td->io_u_busylist);
1025         INIT_FLIST_HEAD(&td->io_u_requeues);
1026         INIT_FLIST_HEAD(&td->io_log_list);
1027         INIT_FLIST_HEAD(&td->io_hist_list);
1028         INIT_FLIST_HEAD(&td->verify_list);
1029         pthread_mutex_init(&td->io_u_lock, NULL);
1030         td->io_hist_tree = RB_ROOT;
1031
1032         pthread_condattr_init(&attr);
1033         pthread_cond_init(&td->verify_cond, &attr);
1034         pthread_cond_init(&td->free_cond, &attr);
1035
1036         td_set_runstate(td, TD_INITIALIZED);
1037         dprint(FD_MUTEX, "up startup_mutex\n");
1038         fio_mutex_up(startup_mutex);
1039         dprint(FD_MUTEX, "wait on td->mutex\n");
1040         fio_mutex_down(td->mutex);
1041         dprint(FD_MUTEX, "done waiting on td->mutex\n");
1042
1043         /*
1044          * the ->mutex mutex is now no longer used, close it to avoid
1045          * eating a file descriptor
1046          */
1047         fio_mutex_remove(td->mutex);
1048
1049         if (td->o.uid != -1U && setuid(td->o.uid)) {
1050                 td_verror(td, errno, "setuid");
1051                 goto err;
1052         }
1053         if (td->o.gid != -1U && setgid(td->o.gid)) {
1054                 td_verror(td, errno, "setgid");
1055                 goto err;
1056         }
1057
1058         /*
1059          * May alter parameters that init_io_u() will use, so we need to
1060          * do this first.
1061          */
1062         if (init_iolog(td))
1063                 goto err;
1064
1065         if (init_io_u(td))
1066                 goto err;
1067
1068         if (td->o.verify_async && verify_async_init(td))
1069                 goto err;
1070
1071         if (td->o.cpumask_set && fio_setaffinity(td->pid, td->o.cpumask) == -1) {
1072                 td_verror(td, errno, "cpu_set_affinity");
1073                 goto err;
1074         }
1075
1076         /*
1077          * If we have a gettimeofday() thread, make sure we exclude that
1078          * thread from this job
1079          */
1080         if (td->o.gtod_cpu) {
1081                 fio_cpu_clear(&td->o.cpumask, td->o.gtod_cpu);
1082                 if (fio_setaffinity(td->pid, td->o.cpumask) == -1) {
1083                         td_verror(td, errno, "cpu_set_affinity");
1084                         goto err;
1085                 }
1086         }
1087
1088         if (td->ioprio_set) {
1089                 if (ioprio_set(IOPRIO_WHO_PROCESS, 0, td->ioprio) == -1) {
1090                         td_verror(td, errno, "ioprio_set");
1091                         goto err;
1092                 }
1093         }
1094
1095         if (td->o.cgroup_weight && cgroup_setup(td, cgroup_list, &cgroup_mnt))
1096                 goto err;
1097
1098         if (nice(td->o.nice) == -1) {
1099                 td_verror(td, errno, "nice");
1100                 goto err;
1101         }
1102
1103         if (td->o.ioscheduler && switch_ioscheduler(td))
1104                 goto err;
1105
1106         if (!td->o.create_serialize && setup_files(td))
1107                 goto err;
1108
1109         if (td_io_init(td))
1110                 goto err;
1111
1112         if (init_random_map(td))
1113                 goto err;
1114
1115         if (td->o.exec_prerun) {
1116                 if (exec_string(td->o.exec_prerun))
1117                         goto err;
1118         }
1119
1120         if (td->o.pre_read) {
1121                 if (pre_read_files(td) < 0)
1122                         goto err;
1123         }
1124
1125         fio_gettime(&td->epoch, NULL);
1126         getrusage(RUSAGE_SELF, &td->ts.ru_start);
1127
1128         runtime[0] = runtime[1] = 0;
1129         clear_state = 0;
1130         while (keep_running(td)) {
1131                 fio_gettime(&td->start, NULL);
1132                 memcpy(&td->ts.stat_sample_time[0], &td->start,
1133                                 sizeof(td->start));
1134                 memcpy(&td->ts.stat_sample_time[1], &td->start,
1135                                 sizeof(td->start));
1136                 memcpy(&td->tv_cache, &td->start, sizeof(td->start));
1137
1138                 if (td->o.ratemin[0] || td->o.ratemin[1])
1139                         memcpy(&td->lastrate, &td->ts.stat_sample_time,
1140                                                         sizeof(td->lastrate));
1141
1142                 if (clear_state)
1143                         clear_io_state(td);
1144
1145                 prune_io_piece_log(td);
1146
1147                 do_io(td);
1148
1149                 clear_state = 1;
1150
1151                 if (td_read(td) && td->io_bytes[DDIR_READ]) {
1152                         elapsed = utime_since_now(&td->start);
1153                         runtime[DDIR_READ] += elapsed;
1154                 }
1155                 if (td_write(td) && td->io_bytes[DDIR_WRITE]) {
1156                         elapsed = utime_since_now(&td->start);
1157                         runtime[DDIR_WRITE] += elapsed;
1158                 }
1159
1160                 if (td->error || td->terminate)
1161                         break;
1162
1163                 if (!td->o.do_verify ||
1164                     td->o.verify == VERIFY_NONE ||
1165                     (td->io_ops->flags & FIO_UNIDIR))
1166                         continue;
1167
1168                 clear_io_state(td);
1169
1170                 fio_gettime(&td->start, NULL);
1171
1172                 do_verify(td);
1173
1174                 runtime[DDIR_READ] += utime_since_now(&td->start);
1175
1176                 if (td->error || td->terminate)
1177                         break;
1178         }
1179
1180         update_rusage_stat(td);
1181         td->ts.runtime[0] = (runtime[0] + 999) / 1000;
1182         td->ts.runtime[1] = (runtime[1] + 999) / 1000;
1183         td->ts.total_run_time = mtime_since_now(&td->epoch);
1184         td->ts.io_bytes[0] = td->io_bytes[0];
1185         td->ts.io_bytes[1] = td->io_bytes[1];
1186
1187         fio_mutex_down(writeout_mutex);
1188         if (td->ts.bw_log) {
1189                 if (td->o.bw_log_file) {
1190                         finish_log_named(td, td->ts.bw_log,
1191                                                 td->o.bw_log_file, "bw");
1192                 } else
1193                         finish_log(td, td->ts.bw_log, "bw");
1194         }
1195         if (td->ts.slat_log) {
1196                 if (td->o.lat_log_file) {
1197                         finish_log_named(td, td->ts.slat_log,
1198                                                 td->o.lat_log_file, "slat");
1199                 } else
1200                         finish_log(td, td->ts.slat_log, "slat");
1201         }
1202         if (td->ts.clat_log) {
1203                 if (td->o.lat_log_file) {
1204                         finish_log_named(td, td->ts.clat_log,
1205                                                 td->o.lat_log_file, "clat");
1206                 } else
1207                         finish_log(td, td->ts.clat_log, "clat");
1208         }
1209         fio_mutex_up(writeout_mutex);
1210         if (td->o.exec_postrun)
1211                 exec_string(td->o.exec_postrun);
1212
1213         if (exitall_on_terminate)
1214                 terminate_threads(td->groupid);
1215
1216 err:
1217         if (td->error)
1218                 log_info("fio: pid=%d, err=%d/%s\n", (int) td->pid, td->error,
1219                                                         td->verror);
1220
1221         if (td->o.verify_async)
1222                 verify_async_exit(td);
1223
1224         close_and_free_files(td);
1225         close_ioengine(td);
1226         cleanup_io_u(td);
1227         cgroup_shutdown(td, &cgroup_mnt);
1228
1229         if (td->o.cpumask_set) {
1230                 int ret = fio_cpuset_exit(&td->o.cpumask);
1231
1232                 td_verror(td, ret, "fio_cpuset_exit");
1233         }
1234
1235         /*
1236          * do this very late, it will log file closing as well
1237          */
1238         if (td->o.write_iolog_file)
1239                 write_iolog_close(td);
1240
1241         options_mem_free(td);
1242         td_set_runstate(td, TD_EXITED);
1243         return (void *) (unsigned long) td->error;
1244 }
1245
1246 /*
1247  * We cannot pass the td data into a forked process, so attach the td and
1248  * pass it to the thread worker.
1249  */
1250 static int fork_main(int shmid, int offset)
1251 {
1252         struct thread_data *td;
1253         void *data, *ret;
1254
1255         data = shmat(shmid, NULL, 0);
1256         if (data == (void *) -1) {
1257                 int __err = errno;
1258
1259                 perror("shmat");
1260                 return __err;
1261         }
1262
1263         td = data + offset * sizeof(struct thread_data);
1264         ret = thread_main(td);
1265         shmdt(data);
1266         return (int) (unsigned long) ret;
1267 }
1268
1269 /*
1270  * Run over the job map and reap the threads that have exited, if any.
1271  */
1272 static void reap_threads(int *nr_running, int *t_rate, int *m_rate)
1273 {
1274         struct thread_data *td;
1275         int i, cputhreads, realthreads, pending, status, ret;
1276
1277         /*
1278          * reap exited threads (TD_EXITED -> TD_REAPED)
1279          */
1280         realthreads = pending = cputhreads = 0;
1281         for_each_td(td, i) {
1282                 int flags = 0;
1283
1284                 /*
1285                  * ->io_ops is NULL for a thread that has closed its
1286                  * io engine
1287                  */
1288                 if (td->io_ops && !strcmp(td->io_ops->name, "cpuio"))
1289                         cputhreads++;
1290                 else
1291                         realthreads++;
1292
1293                 if (!td->pid) {
1294                         pending++;
1295                         continue;
1296                 }
1297                 if (td->runstate == TD_REAPED)
1298                         continue;
1299                 if (td->o.use_thread) {
1300                         if (td->runstate == TD_EXITED) {
1301                                 td_set_runstate(td, TD_REAPED);
1302                                 goto reaped;
1303                         }
1304                         continue;
1305                 }
1306
1307                 flags = WNOHANG;
1308                 if (td->runstate == TD_EXITED)
1309                         flags = 0;
1310
1311                 /*
1312                  * check if someone quit or got killed in an unusual way
1313                  */
1314                 ret = waitpid(td->pid, &status, flags);
1315                 if (ret < 0) {
1316                         if (errno == ECHILD) {
1317                                 log_err("fio: pid=%d disappeared %d\n",
1318                                                 (int) td->pid, td->runstate);
1319                                 td_set_runstate(td, TD_REAPED);
1320                                 goto reaped;
1321                         }
1322                         perror("waitpid");
1323                 } else if (ret == td->pid) {
1324                         if (WIFSIGNALED(status)) {
1325                                 int sig = WTERMSIG(status);
1326
1327                                 if (sig != SIGQUIT)
1328                                         log_err("fio: pid=%d, got signal=%d\n",
1329                                                         (int) td->pid, sig);
1330                                 td_set_runstate(td, TD_REAPED);
1331                                 goto reaped;
1332                         }
1333                         if (WIFEXITED(status)) {
1334                                 if (WEXITSTATUS(status) && !td->error)
1335                                         td->error = WEXITSTATUS(status);
1336
1337                                 td_set_runstate(td, TD_REAPED);
1338                                 goto reaped;
1339                         }
1340                 }
1341
1342                 /*
1343                  * thread is not dead, continue
1344                  */
1345                 pending++;
1346                 continue;
1347 reaped:
1348                 (*nr_running)--;
1349                 (*m_rate) -= (td->o.ratemin[0] + td->o.ratemin[1]);
1350                 (*t_rate) -= (td->o.rate[0] + td->o.rate[1]);
1351                 if (!td->pid)
1352                         pending--;
1353
1354                 if (td->error)
1355                         exit_value++;
1356
1357                 done_secs += mtime_since_now(&td->epoch) / 1000;
1358         }
1359
1360         if (*nr_running == cputhreads && !pending && realthreads)
1361                 terminate_threads(TERMINATE_ALL);
1362 }
1363
1364 static void *gtod_thread_main(void *data)
1365 {
1366         fio_mutex_up(startup_mutex);
1367
1368         /*
1369          * As long as we have jobs around, update the clock. It would be nice
1370          * to have some way of NOT hammering that CPU with gettimeofday(),
1371          * but I'm not sure what to use outside of a simple CPU nop to relax
1372          * it - we don't want to lose precision.
1373          */
1374         while (threads) {
1375                 fio_gtod_update();
1376                 nop;
1377         }
1378
1379         return NULL;
1380 }
1381
1382 static int fio_start_gtod_thread(void)
1383 {
1384         int ret;
1385
1386         ret = pthread_create(&gtod_thread, NULL, gtod_thread_main, NULL);
1387         if (ret) {
1388                 log_err("Can't create gtod thread: %s\n", strerror(ret));
1389                 return 1;
1390         }
1391
1392         ret = pthread_detach(gtod_thread);
1393         if (ret) {
1394                 log_err("Can't detatch gtod thread: %s\n", strerror(ret));
1395                 return 1;
1396         }
1397
1398         dprint(FD_MUTEX, "wait on startup_mutex\n");
1399         fio_mutex_down(startup_mutex);
1400         dprint(FD_MUTEX, "done waiting on startup_mutex\n");
1401         return 0;
1402 }
1403
1404 /*
1405  * Main function for kicking off and reaping jobs, as needed.
1406  */
1407 static void run_threads(void)
1408 {
1409         struct thread_data *td;
1410         unsigned long spent;
1411         int i, todo, nr_running, m_rate, t_rate, nr_started;
1412
1413         if (fio_pin_memory())
1414                 return;
1415
1416         if (fio_gtod_offload && fio_start_gtod_thread())
1417                 return;
1418
1419         if (!terse_output) {
1420                 log_info("Starting ");
1421                 if (nr_thread)
1422                         log_info("%d thread%s", nr_thread,
1423                                                 nr_thread > 1 ? "s" : "");
1424                 if (nr_process) {
1425                         if (nr_thread)
1426                                 printf(" and ");
1427                         log_info("%d process%s", nr_process,
1428                                                 nr_process > 1 ? "es" : "");
1429                 }
1430                 log_info("\n");
1431                 fflush(stdout);
1432         }
1433
1434         set_sig_handlers();
1435
1436         todo = thread_number;
1437         nr_running = 0;
1438         nr_started = 0;
1439         m_rate = t_rate = 0;
1440
1441         for_each_td(td, i) {
1442                 print_status_init(td->thread_number - 1);
1443
1444                 if (!td->o.create_serialize) {
1445                         init_disk_util(td);
1446                         continue;
1447                 }
1448
1449                 /*
1450                  * do file setup here so it happens sequentially,
1451                  * we don't want X number of threads getting their
1452                  * client data interspersed on disk
1453                  */
1454                 if (setup_files(td)) {
1455                         exit_value++;
1456                         if (td->error)
1457                                 log_err("fio: pid=%d, err=%d/%s\n",
1458                                         (int) td->pid, td->error, td->verror);
1459                         td_set_runstate(td, TD_REAPED);
1460                         todo--;
1461                 } else {
1462                         struct fio_file *f;
1463                         unsigned int i;
1464
1465                         /*
1466                          * for sharing to work, each job must always open
1467                          * its own files. so close them, if we opened them
1468                          * for creation
1469                          */
1470                         for_each_file(td, f, i) {
1471                                 if (fio_file_open(f))
1472                                         td_io_close_file(td, f);
1473                         }
1474                 }
1475
1476                 init_disk_util(td);
1477         }
1478
1479         set_genesis_time();
1480
1481         while (todo) {
1482                 struct thread_data *map[MAX_JOBS];
1483                 struct timeval this_start;
1484                 int this_jobs = 0, left;
1485
1486                 /*
1487                  * create threads (TD_NOT_CREATED -> TD_CREATED)
1488                  */
1489                 for_each_td(td, i) {
1490                         if (td->runstate != TD_NOT_CREATED)
1491                                 continue;
1492
1493                         /*
1494                          * never got a chance to start, killed by other
1495                          * thread for some reason
1496                          */
1497                         if (td->terminate) {
1498                                 todo--;
1499                                 continue;
1500                         }
1501
1502                         if (td->o.start_delay) {
1503                                 spent = mtime_since_genesis();
1504
1505                                 if (td->o.start_delay * 1000 > spent)
1506                                         continue;
1507                         }
1508
1509                         if (td->o.stonewall && (nr_started || nr_running)) {
1510                                 dprint(FD_PROCESS, "%s: stonewall wait\n",
1511                                                         td->o.name);
1512                                 break;
1513                         }
1514
1515                         /*
1516                          * Set state to created. Thread will transition
1517                          * to TD_INITIALIZED when it's done setting up.
1518                          */
1519                         td_set_runstate(td, TD_CREATED);
1520                         map[this_jobs++] = td;
1521                         nr_started++;
1522
1523                         if (td->o.use_thread) {
1524                                 int ret;
1525
1526                                 dprint(FD_PROCESS, "will pthread_create\n");
1527                                 ret = pthread_create(&td->thread, NULL,
1528                                                         thread_main, td);
1529                                 if (ret) {
1530                                         log_err("pthread_create: %s\n",
1531                                                         strerror(ret));
1532                                         nr_started--;
1533                                         break;
1534                                 }
1535                                 ret = pthread_detach(td->thread);
1536                                 if (ret)
1537                                         log_err("pthread_detach: %s",
1538                                                         strerror(ret));
1539                         } else {
1540                                 pid_t pid;
1541                                 dprint(FD_PROCESS, "will fork\n");
1542                                 pid = fork();
1543                                 if (!pid) {
1544                                         int ret = fork_main(shm_id, i);
1545
1546                                         _exit(ret);
1547                                 } else if (i == fio_debug_jobno)
1548                                         *fio_debug_jobp = pid;
1549                         }
1550                         dprint(FD_MUTEX, "wait on startup_mutex\n");
1551                         if (fio_mutex_down_timeout(startup_mutex, 10)) {
1552                                 log_err("fio: job startup hung? exiting.\n");
1553                                 terminate_threads(TERMINATE_ALL);
1554                                 fio_abort = 1;
1555                                 nr_started--;
1556                                 break;
1557                         }
1558                         dprint(FD_MUTEX, "done waiting on startup_mutex\n");
1559                 }
1560
1561                 /*
1562                  * Wait for the started threads to transition to
1563                  * TD_INITIALIZED.
1564                  */
1565                 fio_gettime(&this_start, NULL);
1566                 left = this_jobs;
1567                 while (left && !fio_abort) {
1568                         if (mtime_since_now(&this_start) > JOB_START_TIMEOUT)
1569                                 break;
1570
1571                         usleep(100000);
1572
1573                         for (i = 0; i < this_jobs; i++) {
1574                                 td = map[i];
1575                                 if (!td)
1576                                         continue;
1577                                 if (td->runstate == TD_INITIALIZED) {
1578                                         map[i] = NULL;
1579                                         left--;
1580                                 } else if (td->runstate >= TD_EXITED) {
1581                                         map[i] = NULL;
1582                                         left--;
1583                                         todo--;
1584                                         nr_running++; /* work-around... */
1585                                 }
1586                         }
1587                 }
1588
1589                 if (left) {
1590                         log_err("fio: %d jobs failed to start\n", left);
1591                         for (i = 0; i < this_jobs; i++) {
1592                                 td = map[i];
1593                                 if (!td)
1594                                         continue;
1595                                 kill(td->pid, SIGTERM);
1596                         }
1597                         break;
1598                 }
1599
1600                 /*
1601                  * start created threads (TD_INITIALIZED -> TD_RUNNING).
1602                  */
1603                 for_each_td(td, i) {
1604                         if (td->runstate != TD_INITIALIZED)
1605                                 continue;
1606
1607                         if (in_ramp_time(td))
1608                                 td_set_runstate(td, TD_RAMP);
1609                         else
1610                                 td_set_runstate(td, TD_RUNNING);
1611                         nr_running++;
1612                         nr_started--;
1613                         m_rate += td->o.ratemin[0] + td->o.ratemin[1];
1614                         t_rate += td->o.rate[0] + td->o.rate[1];
1615                         todo--;
1616                         fio_mutex_up(td->mutex);
1617                 }
1618
1619                 reap_threads(&nr_running, &t_rate, &m_rate);
1620
1621                 if (todo)
1622                         usleep(100000);
1623         }
1624
1625         while (nr_running) {
1626                 reap_threads(&nr_running, &t_rate, &m_rate);
1627                 usleep(10000);
1628         }
1629
1630         update_io_ticks();
1631         fio_unpin_memory();
1632 }
1633
1634 int main(int argc, char *argv[])
1635 {
1636         long ps;
1637
1638         sinit();
1639
1640         /*
1641          * We need locale for number printing, if it isn't set then just
1642          * go with the US format.
1643          */
1644         if (!getenv("LC_NUMERIC"))
1645                 setlocale(LC_NUMERIC, "en_US");
1646
1647         ps = sysconf(_SC_PAGESIZE);
1648         if (ps < 0) {
1649                 log_err("Failed to get page size\n");
1650                 return 1;
1651         }
1652
1653         page_size = ps;
1654         page_mask = ps - 1;
1655
1656         fio_keywords_init();
1657
1658         if (parse_options(argc, argv))
1659                 return 1;
1660
1661         if (exec_profile && load_profile(exec_profile))
1662                 return 1;
1663
1664         if (!thread_number)
1665                 return 0;
1666
1667         if (write_bw_log) {
1668                 setup_log(&agg_io_log[DDIR_READ]);
1669                 setup_log(&agg_io_log[DDIR_WRITE]);
1670         }
1671
1672         startup_mutex = fio_mutex_init(0);
1673         writeout_mutex = fio_mutex_init(1);
1674
1675         set_genesis_time();
1676
1677         status_timer_arm();
1678
1679         cgroup_list = smalloc(sizeof(*cgroup_list));
1680         INIT_FLIST_HEAD(cgroup_list);
1681
1682         run_threads();
1683
1684         if (!fio_abort) {
1685                 show_run_stats();
1686                 if (write_bw_log) {
1687                         __finish_log(agg_io_log[DDIR_READ], "agg-read_bw.log");
1688                         __finish_log(agg_io_log[DDIR_WRITE],
1689                                         "agg-write_bw.log");
1690                 }
1691         }
1692
1693         cgroup_kill(cgroup_list);
1694         sfree(cgroup_list);
1695         if (cgroup_mnt)
1696                 sfree(cgroup_mnt);
1697
1698         fio_mutex_remove(startup_mutex);
1699         fio_mutex_remove(writeout_mutex);
1700         return exit_value;
1701 }