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 "trim.h"
  42 #include "diskutil.h"
  43 #include "cgroup.h"
  44 #include "profile.h"
  45 #include "lib/rand.h"
  46
  47 unsigned long page_mask;
  48 unsigned long page_size;
  49
  50 #define PAGE_ALIGN(buf) \
  51         (char *) (((unsigned long) (buf) + page_mask) & ~page_mask)
  52
  53 int groupid = 0;
  54 int thread_number = 0;
  55 int nr_process = 0;
  56 int nr_thread = 0;
  57 int shm_id = 0;
  58 int temp_stall_ts;
  59 unsigned long done_secs = 0;
  60
  61 static struct fio_mutex *startup_mutex;
  62 static struct fio_mutex *writeout_mutex;
  63 static volatile int fio_abort;
  64 static int exit_value;
  65 static struct itimerval itimer;
  66 static pthread_t gtod_thread;
  67 static struct flist_head *cgroup_list;
  68 static char *cgroup_mnt;
  69
  70 struct io_log *agg_io_log[2];
  71
  72 #define TERMINATE_ALL           (-1)
  73 #define JOB_START_TIMEOUT       (5 * 1000)
  74
  75 void td_set_runstate(struct thread_data *td, int runstate)
  76 {
  77         if (td->runstate == runstate)
  78                 return;
  79
  80         dprint(FD_PROCESS, "pid=%d: runstate %d -> %d\n", (int) td->pid,
  81                                                 td->runstate, runstate);
  82         td->runstate = runstate;
  83 }
  84
  85 static void terminate_threads(int group_id)
  86 {
  87         struct thread_data *td;
  88         int i;
  89
  90         dprint(FD_PROCESS, "terminate group_id=%d\n", group_id);
  91
  92         for_each_td(td, i) {
  93                 if (group_id == TERMINATE_ALL || groupid == td->groupid) {
  94                         dprint(FD_PROCESS, "setting terminate on %s/%d\n",
  95                                                 td->o.name, (int) td->pid);
  96                         td->terminate = 1;
  97                         td->o.start_delay = 0;
  98
  99                         /*
 100                          * if the thread is running, just let it exit
 101                          */
 102                         if (td->runstate < TD_RUNNING)
 103                                 kill(td->pid, SIGQUIT);
 104                         else {
 105                                 struct ioengine_ops *ops = td->io_ops;
 106
 107                                 if (ops && (ops->flags & FIO_SIGQUIT))
 108                                         kill(td->pid, SIGQUIT);
 109                         }
 110                 }
 111         }
 112 }
 113
 114 static void status_timer_arm(void)
 115 {
 116         itimer.it_value.tv_sec = 0;
 117         itimer.it_value.tv_usec = DISK_UTIL_MSEC * 1000;
 118         setitimer(ITIMER_REAL, &itimer, NULL);
 119 }
 120
 121 static void sig_alrm(int fio_unused sig)
 122 {
 123         if (threads) {
 124                 update_io_ticks();
 125                 print_thread_status();
 126                 status_timer_arm();
 127         }
 128 }
 129
 130 /*
 131  * Happens on thread runs with ctrl-c, ignore our own SIGQUIT
 132  */
 133 static void sig_quit(int sig)
 134 {
 135 }
 136
 137 static void sig_int(int sig)
 138 {
 139         if (threads) {
 140                 log_info("\nfio: terminating on signal %d\n", sig);
 141                 fflush(stdout);
 142                 terminate_threads(TERMINATE_ALL);
 143         }
 144 }
 145
 146 static void set_sig_handlers(void)
 147 {
 148         struct sigaction act;
 149
 150         memset(&act, 0, sizeof(act));
 151         act.sa_handler = sig_alrm;
 152         act.sa_flags = SA_RESTART;
 153         sigaction(SIGALRM, &act, NULL);
 154
 155         memset(&act, 0, sizeof(act));
 156         act.sa_handler = sig_int;
 157         act.sa_flags = SA_RESTART;
 158         sigaction(SIGINT, &act, NULL);
 159
 160         memset(&act, 0, sizeof(act));
 161         act.sa_handler = sig_quit;
 162         act.sa_flags = SA_RESTART;
 163         sigaction(SIGQUIT, &act, NULL);
 164 }
 165
 166 /*
 167  * Check if we are above the minimum rate given.
 168  */
 169 static int __check_min_rate(struct thread_data *td, struct timeval *now,
 170                             enum td_ddir ddir)
 171 {
 172         unsigned long long bytes = 0;
 173         unsigned long iops = 0;
 174         unsigned long spent;
 175         unsigned long rate;
 176         unsigned int ratemin = 0;
 177         unsigned int rate_iops = 0;
 178         unsigned int rate_iops_min = 0;
 179
 180         assert(ddir_rw(ddir));
 181
 182         if (!td->o.ratemin[ddir] && !td->o.rate_iops_min[ddir])
 183                 return 0;
 184
 185         /*
 186          * allow a 2 second settle period in the beginning
 187          */
 188         if (mtime_since(&td->start, now) < 2000)
 189                 return 0;
 190
 191         iops += td->io_blocks[ddir];
 192         bytes += td->this_io_bytes[ddir];
 193         ratemin += td->o.ratemin[ddir];
 194         rate_iops += td->o.rate_iops[ddir];
 195         rate_iops_min += td->o.rate_iops_min[ddir];
 196
 197         /*
 198          * if rate blocks is set, sample is running
 199          */
 200         if (td->rate_bytes[ddir] || td->rate_blocks[ddir]) {
 201                 spent = mtime_since(&td->lastrate[ddir], now);
 202                 if (spent < td->o.ratecycle)
 203                         return 0;
 204
 205                 if (td->o.rate[ddir]) {
 206                         /*
 207                          * check bandwidth specified rate
 208                          */
 209                         if (bytes < td->rate_bytes[ddir]) {
 210                                 log_err("%s: min rate %u not met\n", td->o.name,
 211                                                                 ratemin);
 212                                 return 1;
 213                         } else {
 214                                 rate = ((bytes - td->rate_bytes[ddir]) * 1000) / spent;
 215                                 if (rate < ratemin ||
 216                                     bytes < td->rate_bytes[ddir]) {
 217                                         log_err("%s: min rate %u not met, got"
 218                                                 " %luKB/sec\n", td->o.name,
 219                                                         ratemin, rate);
 220                                         return 1;
 221                                 }
 222                         }
 223                 } else {
 224                         /*
 225                          * checks iops specified rate
 226                          */
 227                         if (iops < rate_iops) {
 228                                 log_err("%s: min iops rate %u not met\n",
 229                                                 td->o.name, rate_iops);
 230                                 return 1;
 231                         } else {
 232                                 rate = ((iops - td->rate_blocks[ddir]) * 1000) / spent;
 233                                 if (rate < rate_iops_min ||
 234                                     iops < td->rate_blocks[ddir]) {
 235                                         log_err("%s: min iops rate %u not met,"
 236                                                 " got %lu\n", td->o.name,
 237                                                         rate_iops_min, rate);
 238                                 }
 239                         }
 240                 }
 241         }
 242
 243         td->rate_bytes[ddir] = bytes;
 244         td->rate_blocks[ddir] = iops;
 245         memcpy(&td->lastrate[ddir], now, sizeof(*now));
 246         return 0;
 247 }
 248
 249 static int check_min_rate(struct thread_data *td, struct timeval *now,
 250                           unsigned long *bytes_done)
 251 {
 252         int ret = 0;
 253
 254         if (bytes_done[0])
 255                 ret |= __check_min_rate(td, now, 0);
 256         if (bytes_done[1])
 257                 ret |= __check_min_rate(td, now, 1);
 258
 259         return ret;
 260 }
 261
 262 static inline int runtime_exceeded(struct thread_data *td, struct timeval *t)
 263 {
 264         if (!td->o.timeout)
 265                 return 0;
 266         if (mtime_since(&td->epoch, t) >= td->o.timeout * 1000)
 267                 return 1;
 268
 269         return 0;
 270 }
 271
 272 /*
 273  * When job exits, we can cancel the in-flight IO if we are using async
 274  * io. Attempt to do so.
 275  */
 276 static void cleanup_pending_aio(struct thread_data *td)
 277 {
 278         struct flist_head *entry, *n;
 279         struct io_u *io_u;
 280         int r;
 281
 282         /*
 283          * get immediately available events, if any
 284          */
 285         r = io_u_queued_complete(td, 0, NULL);
 286         if (r < 0)
 287                 return;
 288
 289         /*
 290          * now cancel remaining active events
 291          */
 292         if (td->io_ops->cancel) {
 293                 flist_for_each_safe(entry, n, &td->io_u_busylist) {
 294                         io_u = flist_entry(entry, struct io_u, list);
 295
 296                         /*
 297                          * if the io_u isn't in flight, then that generally
 298                          * means someone leaked an io_u. complain but fix
 299                          * it up, so we don't stall here.
 300                          */
 301                         if ((io_u->flags & IO_U_F_FLIGHT) == 0) {
 302                                 log_err("fio: non-busy IO on busy list\n");
 303                                 put_io_u(td, io_u);
 304                         } else {
 305                                 r = td->io_ops->cancel(td, io_u);
 306                                 if (!r)
 307                                         put_io_u(td, io_u);
 308                         }
 309                 }
 310         }
 311
 312         if (td->cur_depth)
 313                 r = io_u_queued_complete(td, td->cur_depth, NULL);
 314 }
 315
 316 /*
 317  * Helper to handle the final sync of a file. Works just like the normal
 318  * io path, just does everything sync.
 319  */
 320 static int fio_io_sync(struct thread_data *td, struct fio_file *f)
 321 {
 322         struct io_u *io_u = __get_io_u(td);
 323         int ret;
 324
 325         if (!io_u)
 326                 return 1;
 327
 328         io_u->ddir = DDIR_SYNC;
 329         io_u->file = f;
 330
 331         if (td_io_prep(td, io_u)) {
 332                 put_io_u(td, io_u);
 333                 return 1;
 334         }
 335
 336 requeue:
 337         ret = td_io_queue(td, io_u);
 338         if (ret < 0) {
 339                 td_verror(td, io_u->error, "td_io_queue");
 340                 put_io_u(td, io_u);
 341                 return 1;
 342         } else if (ret == FIO_Q_QUEUED) {
 343                 if (io_u_queued_complete(td, 1, NULL) < 0)
 344                         return 1;
 345         } else if (ret == FIO_Q_COMPLETED) {
 346                 if (io_u->error) {
 347                         td_verror(td, io_u->error, "td_io_queue");
 348                         return 1;
 349                 }
 350
 351                 if (io_u_sync_complete(td, io_u, NULL) < 0)
 352                         return 1;
 353         } else if (ret == FIO_Q_BUSY) {
 354                 if (td_io_commit(td))
 355                         return 1;
 356                 goto requeue;
 357         }
 358
 359         return 0;
 360 }
 361
 362 static inline void update_tv_cache(struct thread_data *td)
 363 {
 364         if ((++td->tv_cache_nr & td->tv_cache_mask) == td->tv_cache_mask)
 365                 fio_gettime(&td->tv_cache, NULL);
 366 }
 367
 368 static int break_on_this_error(struct thread_data *td, int *retptr)
 369 {
 370         int ret = *retptr;
 371
 372         if (ret < 0 || td->error) {
 373                 int err;
 374
 375                 if (!td->o.continue_on_error)
 376                         return 1;
 377
 378                 if (ret < 0)
 379                         err = -ret;
 380                 else
 381                         err = td->error;
 382
 383                 if (td_non_fatal_error(err)) {
 384                         /*
 385                          * Continue with the I/Os in case of
 386                          * a non fatal error.
 387                          */
 388                         update_error_count(td, err);
 389                         td_clear_error(td);
 390                         *retptr = 0;
 391                         return 0;
 392                 } else if (td->o.fill_device && err == ENOSPC) {
 393                         /*
 394                          * We expect to hit this error if
 395                          * fill_device option is set.
 396                          */
 397                         td_clear_error(td);
 398                         td->terminate = 1;
 399                         return 1;
 400                 } else {
 401                         /*
 402                          * Stop the I/O in case of a fatal
 403                          * error.
 404                          */
 405                         update_error_count(td, err);
 406                         return 1;
 407                 }
 408         }
 409
 410         return 0;
 411 }
 412
 413 /*
 414  * The main verify engine. Runs over the writes we previously submitted,
 415  * reads the blocks back in, and checks the crc/md5 of the data.
 416  */
 417 static void do_verify(struct thread_data *td)
 418 {
 419         struct fio_file *f;
 420         struct io_u *io_u;
 421         int ret, min_events;
 422         unsigned int i;
 423
 424         dprint(FD_VERIFY, "starting loop\n");
 425
 426         /*
 427          * sync io first and invalidate cache, to make sure we really
 428          * read from disk.
 429          */
 430         for_each_file(td, f, i) {
 431                 if (!fio_file_open(f))
 432                         continue;
 433                 if (fio_io_sync(td, f))
 434                         break;
 435                 if (file_invalidate_cache(td, f))
 436                         break;
 437         }
 438
 439         if (td->error)
 440                 return;
 441
 442         td_set_runstate(td, TD_VERIFYING);
 443
 444         io_u = NULL;
 445         while (!td->terminate) {
 446                 int ret2, full;
 447
 448                 update_tv_cache(td);
 449
 450                 if (runtime_exceeded(td, &td->tv_cache)) {
 451                         td->terminate = 1;
 452                         break;
 453                 }
 454
 455                 io_u = __get_io_u(td);
 456                 if (!io_u)
 457                         break;
 458
 459                 if (get_next_verify(td, io_u)) {
 460                         put_io_u(td, io_u);
 461                         break;
 462                 }
 463
 464                 if (td_io_prep(td, io_u)) {
 465                         put_io_u(td, io_u);
 466                         break;
 467                 }
 468
 469                 if (td->o.verify_async)
 470                         io_u->end_io = verify_io_u_async;
 471                 else
 472                         io_u->end_io = verify_io_u;
 473
 474                 ret = td_io_queue(td, io_u);
 475                 switch (ret) {
 476                 case FIO_Q_COMPLETED:
 477                         if (io_u->error) {
 478                                 ret = -io_u->error;
 479                                 clear_io_u(td, io_u);
 480                         } else if (io_u->resid) {
 481                                 int bytes = io_u->xfer_buflen - io_u->resid;
 482                                 struct fio_file *f = io_u->file;
 483
 484                                 /*
 485                                  * zero read, fail
 486                                  */
 487                                 if (!bytes) {
 488                                         td_verror(td, EIO, "full resid");
 489                                         put_io_u(td, io_u);
 490                                         break;
 491                                 }
 492
 493                                 io_u->xfer_buflen = io_u->resid;
 494                                 io_u->xfer_buf += bytes;
 495                                 io_u->offset += bytes;
 496
 497                                 if (ddir_rw(io_u->ddir))
 498                                         td->ts.short_io_u[io_u->ddir]++;
 499
 500                                 if (io_u->offset == f->real_file_size)
 501                                         goto sync_done;
 502
 503                                 requeue_io_u(td, &io_u);
 504                         } else {
 505 sync_done:
 506                                 ret = io_u_sync_complete(td, io_u, NULL);
 507                                 if (ret < 0)
 508                                         break;
 509                         }
 510                         continue;
 511                 case FIO_Q_QUEUED:
 512                         break;
 513                 case FIO_Q_BUSY:
 514                         requeue_io_u(td, &io_u);
 515                         ret2 = td_io_commit(td);
 516                         if (ret2 < 0)
 517                                 ret = ret2;
 518                         break;
 519                 default:
 520                         assert(ret < 0);
 521                         td_verror(td, -ret, "td_io_queue");
 522                         break;
 523                 }
 524
 525                 if (break_on_this_error(td, &ret))
 526                         break;
 527
 528                 /*
 529                  * if we can queue more, do so. but check if there are
 530                  * completed io_u's first.
 531                  */
 532                 full = queue_full(td) || ret == FIO_Q_BUSY;
 533                 if (full || !td->o.iodepth_batch_complete) {
 534                         min_events = min(td->o.iodepth_batch_complete,
 535                                          td->cur_depth);
 536                         if (full && !min_events)
 537                                 min_events = 1;
 538
 539                         do {
 540                                 /*
 541                                  * Reap required number of io units, if any,
 542                                  * and do the verification on them through
 543                                  * the callback handler
 544                                  */
 545                                 if (io_u_queued_complete(td, min_events, NULL) < 0) {
 546                                         ret = -1;
 547                                         break;
 548                                 }
 549                         } while (full && (td->cur_depth > td->o.iodepth_low));
 550                 }
 551                 if (ret < 0)
 552                         break;
 553         }
 554
 555         if (!td->error) {
 556                 min_events = td->cur_depth;
 557
 558                 if (min_events)
 559                         ret = io_u_queued_complete(td, min_events, NULL);
 560         } else
 561                 cleanup_pending_aio(td);
 562
 563         td_set_runstate(td, TD_RUNNING);
 564
 565         dprint(FD_VERIFY, "exiting loop\n");
 566 }
 567
 568 /*
 569  * Main IO worker function. It retrieves io_u's to process and queues
 570  * and reaps them, checking for rate and errors along the way.
 571  */
 572 static void do_io(struct thread_data *td)
 573 {
 574         unsigned int i;
 575         int ret = 0;
 576
 577         if (in_ramp_time(td))
 578                 td_set_runstate(td, TD_RAMP);
 579         else
 580                 td_set_runstate(td, TD_RUNNING);
 581
 582         while ( (td->o.read_iolog_file && !flist_empty(&td->io_log_list)) ||
 583                 (!flist_empty(&td->trim_list)) ||
 584                 ((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                                 if (ddir_rw(io_u->ddir))
 645                                         td->ts.short_io_u[io_u->ddir]++;
 646
 647                                 if (io_u->offset == f->real_file_size)
 648                                         goto sync_done;
 649
 650                                 requeue_io_u(td, &io_u);
 651                         } else {
 652 sync_done:
 653                                 if (__should_check_rate(td, 0) ||
 654                                     __should_check_rate(td, 1))
 655                                         fio_gettime(&comp_time, NULL);
 656
 657                                 ret = io_u_sync_complete(td, io_u, bytes_done);
 658                                 if (ret < 0)
 659                                         break;
 660                         }
 661                         break;
 662                 case FIO_Q_QUEUED:
 663                         /*
 664                          * if the engine doesn't have a commit hook,
 665                          * the io_u is really queued. if it does have such
 666                          * a hook, it has to call io_u_queued() itself.
 667                          */
 668                         if (td->io_ops->commit == NULL)
 669                                 io_u_queued(td, io_u);
 670                         break;
 671                 case FIO_Q_BUSY:
 672                         requeue_io_u(td, &io_u);
 673                         ret2 = td_io_commit(td);
 674                         if (ret2 < 0)
 675                                 ret = ret2;
 676                         break;
 677                 default:
 678                         assert(ret < 0);
 679                         put_io_u(td, io_u);
 680                         break;
 681                 }
 682
 683                 if (break_on_this_error(td, &ret))
 684                         break;
 685
 686                 /*
 687                  * See if we need to complete some commands
 688                  */
 689                 full = queue_full(td) || ret == FIO_Q_BUSY;
 690                 if (full || !td->o.iodepth_batch_complete) {
 691                         min_evts = min(td->o.iodepth_batch_complete,
 692                                         td->cur_depth);
 693                         if (full && !min_evts)
 694                                 min_evts = 1;
 695
 696                         if (__should_check_rate(td, 0) ||
 697                             __should_check_rate(td, 1))
 698                                 fio_gettime(&comp_time, NULL);
 699
 700                         do {
 701                                 ret = io_u_queued_complete(td, min_evts, bytes_done);
 702                                 if (ret < 0)
 703                                         break;
 704
 705                         } while (full && (td->cur_depth > td->o.iodepth_low));
 706                 }
 707
 708                 if (ret < 0)
 709                         break;
 710                 if (!(bytes_done[0] + bytes_done[1]))
 711                         continue;
 712
 713                 if (!in_ramp_time(td) && should_check_rate(td, bytes_done)) {
 714                         if (check_min_rate(td, &comp_time, bytes_done)) {
 715                                 if (exitall_on_terminate)
 716                                         terminate_threads(td->groupid);
 717                                 td_verror(td, EIO, "check_min_rate");
 718                                 break;
 719                         }
 720                 }
 721
 722                 if (td->o.thinktime) {
 723                         unsigned long long b;
 724
 725                         b = td->io_blocks[0] + td->io_blocks[1];
 726                         if (!(b % td->o.thinktime_blocks)) {
 727                                 int left;
 728
 729                                 if (td->o.thinktime_spin)
 730                                         usec_spin(td->o.thinktime_spin);
 731
 732                                 left = td->o.thinktime - td->o.thinktime_spin;
 733                                 if (left)
 734                                         usec_sleep(td, left);
 735                         }
 736                 }
 737         }
 738
 739         assert(!td->trim_entries);
 740
 741         if (td->o.fill_device && td->error == ENOSPC) {
 742                 td->error = 0;
 743                 td->terminate = 1;
 744         }
 745         if (!td->error) {
 746                 struct fio_file *f;
 747
 748                 i = td->cur_depth;
 749                 if (i)
 750                         ret = io_u_queued_complete(td, i, NULL);
 751
 752                 if (should_fsync(td) && td->o.end_fsync) {
 753                         td_set_runstate(td, TD_FSYNCING);
 754
 755                         for_each_file(td, f, i) {
 756                                 if (!fio_file_open(f))
 757                                         continue;
 758                                 fio_io_sync(td, f);
 759                         }
 760                 }
 761         } else
 762                 cleanup_pending_aio(td);
 763
 764         /*
 765          * stop job if we failed doing any IO
 766          */
 767         if ((td->this_io_bytes[0] + td->this_io_bytes[1]) == 0)
 768                 td->done = 1;
 769 }
 770
 771 static void cleanup_io_u(struct thread_data *td)
 772 {
 773         struct flist_head *entry, *n;
 774         struct io_u *io_u;
 775
 776         flist_for_each_safe(entry, n, &td->io_u_freelist) {
 777                 io_u = flist_entry(entry, struct io_u, list);
 778
 779                 flist_del(&io_u->list);
 780                 free(io_u);
 781         }
 782
 783         free_io_mem(td);
 784 }
 785
 786 static int init_io_u(struct thread_data *td)
 787 {
 788         struct io_u *io_u;
 789         unsigned int max_bs;
 790         int cl_align, i, max_units;
 791         char *p;
 792
 793         max_units = td->o.iodepth;
 794         max_bs = max(td->o.max_bs[DDIR_READ], td->o.max_bs[DDIR_WRITE]);
 795         td->orig_buffer_size = (unsigned long long) max_bs
 796                                         * (unsigned long long) max_units;
 797
 798         if (td->o.mem_type == MEM_SHMHUGE || td->o.mem_type == MEM_MMAPHUGE) {
 799                 unsigned long bs;
 800
 801                 bs = td->orig_buffer_size + td->o.hugepage_size - 1;
 802                 td->orig_buffer_size = bs & ~(td->o.hugepage_size - 1);
 803         }
 804
 805         if (td->orig_buffer_size != (size_t) td->orig_buffer_size) {
 806                 log_err("fio: IO memory too large. Reduce max_bs or iodepth\n");
 807                 return 1;
 808         }
 809
 810         if (allocate_io_mem(td))
 811                 return 1;
 812
 813         if (td->o.odirect || td->o.mem_align)
 814                 p = PAGE_ALIGN(td->orig_buffer) + td->o.mem_align;
 815         else
 816                 p = td->orig_buffer;
 817
 818         cl_align = os_cache_line_size();
 819
 820         for (i = 0; i < max_units; i++) {
 821                 void *ptr;
 822
 823                 if (td->terminate)
 824                         return 1;
 825
 826                 if (posix_memalign(&ptr, cl_align, sizeof(*io_u))) {
 827                         log_err("fio: posix_memalign=%s\n", strerror(errno));
 828                         break;
 829                 }
 830
 831                 io_u = ptr;
 832                 memset(io_u, 0, sizeof(*io_u));
 833                 INIT_FLIST_HEAD(&io_u->list);
 834                 dprint(FD_MEM, "io_u alloc %p, index %u\n", io_u, i);
 835
 836                 if (!(td->io_ops->flags & FIO_NOIO)) {
 837                         io_u->buf = p + max_bs * i;
 838                         dprint(FD_MEM, "io_u %p, mem %p\n", io_u, io_u->buf);
 839
 840                         if (td_write(td) && !td->o.refill_buffers)
 841                                 io_u_fill_buffer(td, io_u, max_bs);
 842                         else if (td_write(td) && td->o.verify_pattern_bytes) {
 843                                 /*
 844                                  * Fill the buffer with the pattern if we are
 845                                  * going to be doing writes.
 846                                  */
 847                                 fill_pattern(td, io_u->buf, max_bs, io_u);
 848                         }
 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         INIT_FLIST_HEAD(&td->trim_list);
1030         pthread_mutex_init(&td->io_u_lock, NULL);
1031         td->io_hist_tree = RB_ROOT;
1032
1033         pthread_condattr_init(&attr);
1034         pthread_cond_init(&td->verify_cond, &attr);
1035         pthread_cond_init(&td->free_cond, &attr);
1036
1037         td_set_runstate(td, TD_INITIALIZED);
1038         dprint(FD_MUTEX, "up startup_mutex\n");
1039         fio_mutex_up(startup_mutex);
1040         dprint(FD_MUTEX, "wait on td->mutex\n");
1041         fio_mutex_down(td->mutex);
1042         dprint(FD_MUTEX, "done waiting on td->mutex\n");
1043
1044         /*
1045          * the ->mutex mutex is now no longer used, close it to avoid
1046          * eating a file descriptor
1047          */
1048         fio_mutex_remove(td->mutex);
1049
1050         if (td->o.uid != -1U && setuid(td->o.uid)) {
1051                 td_verror(td, errno, "setuid");
1052                 goto err;
1053         }
1054         if (td->o.gid != -1U && setgid(td->o.gid)) {
1055                 td_verror(td, errno, "setgid");
1056                 goto err;
1057         }
1058
1059         /*
1060          * May alter parameters that init_io_u() will use, so we need to
1061          * do this first.
1062          */
1063         if (init_iolog(td))
1064                 goto err;
1065
1066         if (init_io_u(td))
1067                 goto err;
1068
1069         if (td->o.verify_async && verify_async_init(td))
1070                 goto err;
1071
1072         if (td->o.cpumask_set && fio_setaffinity(td->pid, td->o.cpumask) == -1) {
1073                 td_verror(td, errno, "cpu_set_affinity");
1074                 goto err;
1075         }
1076
1077         /*
1078          * If we have a gettimeofday() thread, make sure we exclude that
1079          * thread from this job
1080          */
1081         if (td->o.gtod_cpu) {
1082                 fio_cpu_clear(&td->o.cpumask, td->o.gtod_cpu);
1083                 if (fio_setaffinity(td->pid, td->o.cpumask) == -1) {
1084                         td_verror(td, errno, "cpu_set_affinity");
1085                         goto err;
1086                 }
1087         }
1088
1089         if (td->ioprio_set) {
1090                 if (ioprio_set(IOPRIO_WHO_PROCESS, 0, td->ioprio) == -1) {
1091                         td_verror(td, errno, "ioprio_set");
1092                         goto err;
1093                 }
1094         }
1095
1096         if (td->o.cgroup_weight && cgroup_setup(td, cgroup_list, &cgroup_mnt))
1097                 goto err;
1098
1099         if (nice(td->o.nice) == -1) {
1100                 td_verror(td, errno, "nice");
1101                 goto err;
1102         }
1103
1104         if (td->o.ioscheduler && switch_ioscheduler(td))
1105                 goto err;
1106
1107         if (!td->o.create_serialize && setup_files(td))
1108                 goto err;
1109
1110         if (td_io_init(td))
1111                 goto err;
1112
1113         if (init_random_map(td))
1114                 goto err;
1115
1116         if (td->o.exec_prerun) {
1117                 if (exec_string(td->o.exec_prerun))
1118                         goto err;
1119         }
1120
1121         if (td->o.pre_read) {
1122                 if (pre_read_files(td) < 0)
1123                         goto err;
1124         }
1125
1126         fio_gettime(&td->epoch, NULL);
1127         getrusage(RUSAGE_SELF, &td->ts.ru_start);
1128
1129         runtime[0] = runtime[1] = 0;
1130         clear_state = 0;
1131         while (keep_running(td)) {
1132                 fio_gettime(&td->start, NULL);
1133                 memcpy(&td->ts.stat_sample_time[0], &td->start,
1134                                 sizeof(td->start));
1135                 memcpy(&td->ts.stat_sample_time[1], &td->start,
1136                                 sizeof(td->start));
1137                 memcpy(&td->tv_cache, &td->start, sizeof(td->start));
1138
1139                 if (td->o.ratemin[0] || td->o.ratemin[1])
1140                         memcpy(&td->lastrate, &td->ts.stat_sample_time,
1141                                                         sizeof(td->lastrate));
1142
1143                 if (clear_state)
1144                         clear_io_state(td);
1145
1146                 prune_io_piece_log(td);
1147
1148                 do_io(td);
1149
1150                 clear_state = 1;
1151
1152                 if (td_read(td) && td->io_bytes[DDIR_READ]) {
1153                         elapsed = utime_since_now(&td->start);
1154                         runtime[DDIR_READ] += elapsed;
1155                 }
1156                 if (td_write(td) && td->io_bytes[DDIR_WRITE]) {
1157                         elapsed = utime_since_now(&td->start);
1158                         runtime[DDIR_WRITE] += elapsed;
1159                 }
1160
1161                 if (td->error || td->terminate)
1162                         break;
1163
1164                 if (!td->o.do_verify ||
1165                     td->o.verify == VERIFY_NONE ||
1166                     (td->io_ops->flags & FIO_UNIDIR))
1167                         continue;
1168
1169                 clear_io_state(td);
1170
1171                 fio_gettime(&td->start, NULL);
1172
1173                 do_verify(td);
1174
1175                 runtime[DDIR_READ] += utime_since_now(&td->start);
1176
1177                 if (td->error || td->terminate)
1178                         break;
1179         }
1180
1181         update_rusage_stat(td);
1182         td->ts.runtime[0] = (runtime[0] + 999) / 1000;
1183         td->ts.runtime[1] = (runtime[1] + 999) / 1000;
1184         td->ts.total_run_time = mtime_since_now(&td->epoch);
1185         td->ts.io_bytes[0] = td->io_bytes[0];
1186         td->ts.io_bytes[1] = td->io_bytes[1];
1187
1188         fio_mutex_down(writeout_mutex);
1189         if (td->ts.bw_log) {
1190                 if (td->o.bw_log_file) {
1191                         finish_log_named(td, td->ts.bw_log,
1192                                                 td->o.bw_log_file, "bw");
1193                 } else
1194                         finish_log(td, td->ts.bw_log, "bw");
1195         }
1196         if (td->ts.lat_log) {
1197                 if (td->o.lat_log_file) {
1198                         finish_log_named(td, td->ts.lat_log,
1199                                                 td->o.lat_log_file, "lat");
1200                 } else
1201                         finish_log(td, td->ts.lat_log, "lat");
1202         }
1203         if (td->ts.slat_log) {
1204                 if (td->o.lat_log_file) {
1205                         finish_log_named(td, td->ts.slat_log,
1206                                                 td->o.lat_log_file, "slat");
1207                 } else
1208                         finish_log(td, td->ts.slat_log, "slat");
1209         }
1210         if (td->ts.clat_log) {
1211                 if (td->o.lat_log_file) {
1212                         finish_log_named(td, td->ts.clat_log,
1213                                                 td->o.lat_log_file, "clat");
1214                 } else
1215                         finish_log(td, td->ts.clat_log, "clat");
1216         }
1217         fio_mutex_up(writeout_mutex);
1218         if (td->o.exec_postrun)
1219                 exec_string(td->o.exec_postrun);
1220
1221         if (exitall_on_terminate)
1222                 terminate_threads(td->groupid);
1223
1224 err:
1225         if (td->error)
1226                 log_info("fio: pid=%d, err=%d/%s\n", (int) td->pid, td->error,
1227                                                         td->verror);
1228
1229         if (td->o.verify_async)
1230                 verify_async_exit(td);
1231
1232         close_and_free_files(td);
1233         close_ioengine(td);
1234         cleanup_io_u(td);
1235         cgroup_shutdown(td, &cgroup_mnt);
1236
1237         if (td->o.cpumask_set) {
1238                 int ret = fio_cpuset_exit(&td->o.cpumask);
1239
1240                 td_verror(td, ret, "fio_cpuset_exit");
1241         }
1242
1243         /*
1244          * do this very late, it will log file closing as well
1245          */
1246         if (td->o.write_iolog_file)
1247                 write_iolog_close(td);
1248
1249         options_mem_free(td);
1250         td_set_runstate(td, TD_EXITED);
1251         return (void *) (unsigned long) td->error;
1252 }
1253
1254 /*
1255  * We cannot pass the td data into a forked process, so attach the td and
1256  * pass it to the thread worker.
1257  */
1258 static int fork_main(int shmid, int offset)
1259 {
1260         struct thread_data *td;
1261         void *data, *ret;
1262
1263         data = shmat(shmid, NULL, 0);
1264         if (data == (void *) -1) {
1265                 int __err = errno;
1266
1267                 perror("shmat");
1268                 return __err;
1269         }
1270
1271         td = data + offset * sizeof(struct thread_data);
1272         ret = thread_main(td);
1273         shmdt(data);
1274         return (int) (unsigned long) ret;
1275 }
1276
1277 /*
1278  * Run over the job map and reap the threads that have exited, if any.
1279  */
1280 static void reap_threads(int *nr_running, int *t_rate, int *m_rate)
1281 {
1282         struct thread_data *td;
1283         int i, cputhreads, realthreads, pending, status, ret;
1284
1285         /*
1286          * reap exited threads (TD_EXITED -> TD_REAPED)
1287          */
1288         realthreads = pending = cputhreads = 0;
1289         for_each_td(td, i) {
1290                 int flags = 0;
1291
1292                 /*
1293                  * ->io_ops is NULL for a thread that has closed its
1294                  * io engine
1295                  */
1296                 if (td->io_ops && !strcmp(td->io_ops->name, "cpuio"))
1297                         cputhreads++;
1298                 else
1299                         realthreads++;
1300
1301                 if (!td->pid) {
1302                         pending++;
1303                         continue;
1304                 }
1305                 if (td->runstate == TD_REAPED)
1306                         continue;
1307                 if (td->o.use_thread) {
1308                         if (td->runstate == TD_EXITED) {
1309                                 td_set_runstate(td, TD_REAPED);
1310                                 goto reaped;
1311                         }
1312                         continue;
1313                 }
1314
1315                 flags = WNOHANG;
1316                 if (td->runstate == TD_EXITED)
1317                         flags = 0;
1318
1319                 /*
1320                  * check if someone quit or got killed in an unusual way
1321                  */
1322                 ret = waitpid(td->pid, &status, flags);
1323                 if (ret < 0) {
1324                         if (errno == ECHILD) {
1325                                 log_err("fio: pid=%d disappeared %d\n",
1326                                                 (int) td->pid, td->runstate);
1327                                 td_set_runstate(td, TD_REAPED);
1328                                 goto reaped;
1329                         }
1330                         perror("waitpid");
1331                 } else if (ret == td->pid) {
1332                         if (WIFSIGNALED(status)) {
1333                                 int sig = WTERMSIG(status);
1334
1335                                 if (sig != SIGQUIT)
1336                                         log_err("fio: pid=%d, got signal=%d\n",
1337                                                         (int) td->pid, sig);
1338                                 td_set_runstate(td, TD_REAPED);
1339                                 goto reaped;
1340                         }
1341                         if (WIFEXITED(status)) {
1342                                 if (WEXITSTATUS(status) && !td->error)
1343                                         td->error = WEXITSTATUS(status);
1344
1345                                 td_set_runstate(td, TD_REAPED);
1346                                 goto reaped;
1347                         }
1348                 }
1349
1350                 /*
1351                  * thread is not dead, continue
1352                  */
1353                 pending++;
1354                 continue;
1355 reaped:
1356                 (*nr_running)--;
1357                 (*m_rate) -= (td->o.ratemin[0] + td->o.ratemin[1]);
1358                 (*t_rate) -= (td->o.rate[0] + td->o.rate[1]);
1359                 if (!td->pid)
1360                         pending--;
1361
1362                 if (td->error)
1363                         exit_value++;
1364
1365                 done_secs += mtime_since_now(&td->epoch) / 1000;
1366         }
1367
1368         if (*nr_running == cputhreads && !pending && realthreads)
1369                 terminate_threads(TERMINATE_ALL);
1370 }
1371
1372 static void *gtod_thread_main(void *data)
1373 {
1374         fio_mutex_up(startup_mutex);
1375
1376         /*
1377          * As long as we have jobs around, update the clock. It would be nice
1378          * to have some way of NOT hammering that CPU with gettimeofday(),
1379          * but I'm not sure what to use outside of a simple CPU nop to relax
1380          * it - we don't want to lose precision.
1381          */
1382         while (threads) {
1383                 fio_gtod_update();
1384                 nop;
1385         }
1386
1387         return NULL;
1388 }
1389
1390 static int fio_start_gtod_thread(void)
1391 {
1392         pthread_attr_t attr;
1393         int ret;
1394
1395         pthread_attr_init(&attr);
1396         pthread_attr_setstacksize(&attr, PTHREAD_STACK_MIN);
1397         ret = pthread_create(&gtod_thread, &attr, gtod_thread_main, NULL);
1398         pthread_attr_destroy(&attr);
1399         if (ret) {
1400                 log_err("Can't create gtod thread: %s\n", strerror(ret));
1401                 return 1;
1402         }
1403
1404         ret = pthread_detach(gtod_thread);
1405         if (ret) {
1406                 log_err("Can't detatch gtod thread: %s\n", strerror(ret));
1407                 return 1;
1408         }
1409
1410         dprint(FD_MUTEX, "wait on startup_mutex\n");
1411         fio_mutex_down(startup_mutex);
1412         dprint(FD_MUTEX, "done waiting on startup_mutex\n");
1413         return 0;
1414 }
1415
1416 /*
1417  * Main function for kicking off and reaping jobs, as needed.
1418  */
1419 static void run_threads(void)
1420 {
1421         struct thread_data *td;
1422         unsigned long spent;
1423         int i, todo, nr_running, m_rate, t_rate, nr_started;
1424
1425         if (fio_pin_memory())
1426                 return;
1427
1428         if (fio_gtod_offload && fio_start_gtod_thread())
1429                 return;
1430
1431         if (!terse_output) {
1432                 log_info("Starting ");
1433                 if (nr_thread)
1434                         log_info("%d thread%s", nr_thread,
1435                                                 nr_thread > 1 ? "s" : "");
1436                 if (nr_process) {
1437                         if (nr_thread)
1438                                 printf(" and ");
1439                         log_info("%d process%s", nr_process,
1440                                                 nr_process > 1 ? "es" : "");
1441                 }
1442                 log_info("\n");
1443                 fflush(stdout);
1444         }
1445
1446         set_sig_handlers();
1447
1448         todo = thread_number;
1449         nr_running = 0;
1450         nr_started = 0;
1451         m_rate = t_rate = 0;
1452
1453         for_each_td(td, i) {
1454                 print_status_init(td->thread_number - 1);
1455
1456                 if (!td->o.create_serialize) {
1457                         init_disk_util(td);
1458                         continue;
1459                 }
1460
1461                 /*
1462                  * do file setup here so it happens sequentially,
1463                  * we don't want X number of threads getting their
1464                  * client data interspersed on disk
1465                  */
1466                 if (setup_files(td)) {
1467                         exit_value++;
1468                         if (td->error)
1469                                 log_err("fio: pid=%d, err=%d/%s\n",
1470                                         (int) td->pid, td->error, td->verror);
1471                         td_set_runstate(td, TD_REAPED);
1472                         todo--;
1473                 } else {
1474                         struct fio_file *f;
1475                         unsigned int i;
1476
1477                         /*
1478                          * for sharing to work, each job must always open
1479                          * its own files. so close them, if we opened them
1480                          * for creation
1481                          */
1482                         for_each_file(td, f, i) {
1483                                 if (fio_file_open(f))
1484                                         td_io_close_file(td, f);
1485                         }
1486                 }
1487
1488                 init_disk_util(td);
1489         }
1490
1491         set_genesis_time();
1492
1493         while (todo) {
1494                 struct thread_data *map[MAX_JOBS];
1495                 struct timeval this_start;
1496                 int this_jobs = 0, left;
1497
1498                 /*
1499                  * create threads (TD_NOT_CREATED -> TD_CREATED)
1500                  */
1501                 for_each_td(td, i) {
1502                         if (td->runstate != TD_NOT_CREATED)
1503                                 continue;
1504
1505                         /*
1506                          * never got a chance to start, killed by other
1507                          * thread for some reason
1508                          */
1509                         if (td->terminate) {
1510                                 todo--;
1511                                 continue;
1512                         }
1513
1514                         if (td->o.start_delay) {
1515                                 spent = mtime_since_genesis();
1516
1517                                 if (td->o.start_delay * 1000 > spent)
1518                                         continue;
1519                         }
1520
1521                         if (td->o.stonewall && (nr_started || nr_running)) {
1522                                 dprint(FD_PROCESS, "%s: stonewall wait\n",
1523                                                         td->o.name);
1524                                 break;
1525                         }
1526
1527                         /*
1528                          * Set state to created. Thread will transition
1529                          * to TD_INITIALIZED when it's done setting up.
1530                          */
1531                         td_set_runstate(td, TD_CREATED);
1532                         map[this_jobs++] = td;
1533                         nr_started++;
1534
1535                         if (td->o.use_thread) {
1536                                 int ret;
1537
1538                                 dprint(FD_PROCESS, "will pthread_create\n");
1539                                 ret = pthread_create(&td->thread, NULL,
1540                                                         thread_main, td);
1541                                 if (ret) {
1542                                         log_err("pthread_create: %s\n",
1543                                                         strerror(ret));
1544                                         nr_started--;
1545                                         break;
1546                                 }
1547                                 ret = pthread_detach(td->thread);
1548                                 if (ret)
1549                                         log_err("pthread_detach: %s",
1550                                                         strerror(ret));
1551                         } else {
1552                                 pid_t pid;
1553                                 dprint(FD_PROCESS, "will fork\n");
1554                                 pid = fork();
1555                                 if (!pid) {
1556                                         int ret = fork_main(shm_id, i);
1557
1558                                         _exit(ret);
1559                                 } else if (i == fio_debug_jobno)
1560                                         *fio_debug_jobp = pid;
1561                         }
1562                         dprint(FD_MUTEX, "wait on startup_mutex\n");
1563                         if (fio_mutex_down_timeout(startup_mutex, 10)) {
1564                                 log_err("fio: job startup hung? exiting.\n");
1565                                 terminate_threads(TERMINATE_ALL);
1566                                 fio_abort = 1;
1567                                 nr_started--;
1568                                 break;
1569                         }
1570                         dprint(FD_MUTEX, "done waiting on startup_mutex\n");
1571                 }
1572
1573                 /*
1574                  * Wait for the started threads to transition to
1575                  * TD_INITIALIZED.
1576                  */
1577                 fio_gettime(&this_start, NULL);
1578                 left = this_jobs;
1579                 while (left && !fio_abort) {
1580                         if (mtime_since_now(&this_start) > JOB_START_TIMEOUT)
1581                                 break;
1582
1583                         usleep(100000);
1584
1585                         for (i = 0; i < this_jobs; i++) {
1586                                 td = map[i];
1587                                 if (!td)
1588                                         continue;
1589                                 if (td->runstate == TD_INITIALIZED) {
1590                                         map[i] = NULL;
1591                                         left--;
1592                                 } else if (td->runstate >= TD_EXITED) {
1593                                         map[i] = NULL;
1594                                         left--;
1595                                         todo--;
1596                                         nr_running++; /* work-around... */
1597                                 }
1598                         }
1599                 }
1600
1601                 if (left) {
1602                         log_err("fio: %d jobs failed to start\n", left);
1603                         for (i = 0; i < this_jobs; i++) {
1604                                 td = map[i];
1605                                 if (!td)
1606                                         continue;
1607                                 kill(td->pid, SIGTERM);
1608                         }
1609                         break;
1610                 }
1611
1612                 /*
1613                  * start created threads (TD_INITIALIZED -> TD_RUNNING).
1614                  */
1615                 for_each_td(td, i) {
1616                         if (td->runstate != TD_INITIALIZED)
1617                                 continue;
1618
1619                         if (in_ramp_time(td))
1620                                 td_set_runstate(td, TD_RAMP);
1621                         else
1622                                 td_set_runstate(td, TD_RUNNING);
1623                         nr_running++;
1624                         nr_started--;
1625                         m_rate += td->o.ratemin[0] + td->o.ratemin[1];
1626                         t_rate += td->o.rate[0] + td->o.rate[1];
1627                         todo--;
1628                         fio_mutex_up(td->mutex);
1629                 }
1630
1631                 reap_threads(&nr_running, &t_rate, &m_rate);
1632
1633                 if (todo)
1634                         usleep(100000);
1635         }
1636
1637         while (nr_running) {
1638                 reap_threads(&nr_running, &t_rate, &m_rate);
1639                 usleep(10000);
1640         }
1641
1642         update_io_ticks();
1643         fio_unpin_memory();
1644 }
1645
1646 int main(int argc, char *argv[])
1647 {
1648         long ps;
1649
1650         sinit();
1651         init_rand(&__fio_rand_state);
1652
1653         /*
1654          * We need locale for number printing, if it isn't set then just
1655          * go with the US format.
1656          */
1657         if (!getenv("LC_NUMERIC"))
1658                 setlocale(LC_NUMERIC, "en_US");
1659
1660         ps = sysconf(_SC_PAGESIZE);
1661         if (ps < 0) {
1662                 log_err("Failed to get page size\n");
1663                 return 1;
1664         }
1665
1666         page_size = ps;
1667         page_mask = ps - 1;
1668
1669         fio_keywords_init();
1670
1671         if (parse_options(argc, argv))
1672                 return 1;
1673
1674         if (exec_profile && load_profile(exec_profile))
1675                 return 1;
1676
1677         if (!thread_number)
1678                 return 0;
1679
1680         if (write_bw_log) {
1681                 setup_log(&agg_io_log[DDIR_READ]);
1682                 setup_log(&agg_io_log[DDIR_WRITE]);
1683         }
1684
1685         startup_mutex = fio_mutex_init(0);
1686         writeout_mutex = fio_mutex_init(1);
1687
1688         set_genesis_time();
1689
1690         status_timer_arm();
1691
1692         cgroup_list = smalloc(sizeof(*cgroup_list));
1693         INIT_FLIST_HEAD(cgroup_list);
1694
1695         run_threads();
1696
1697         if (!fio_abort) {
1698                 show_run_stats();
1699                 if (write_bw_log) {
1700                         __finish_log(agg_io_log[DDIR_READ], "agg-read_bw.log");
1701                         __finish_log(agg_io_log[DDIR_WRITE],
1702                                         "agg-write_bw.log");
1703                 }
1704         }
1705
1706         cgroup_kill(cgroup_list);
1707         sfree(cgroup_list);
1708         sfree(cgroup_mnt);
1709
1710         fio_mutex_remove(startup_mutex);
1711         fio_mutex_remove(writeout_mutex);
1712         return exit_value;
1713 }