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         /*
 437          * sync io first and invalidate cache, to make sure we really
 438          * read from disk.
 439          */
 440         for_each_file(td, f, i) {
 441                 if (!fio_file_open(f))
 442                         continue;
 443                 if (fio_io_sync(td, f))
 444                         break;
 445                 if (file_invalidate_cache(td, f))
 446                         break;
 447         }
 448
 449         if (td->error)
 450                 return;
 451
 452         td_set_runstate(td, TD_VERIFYING);
 453
 454         io_u = NULL;
 455         while (!td->terminate) {
 456                 int ret2, full;
 457
 458                 update_tv_cache(td);
 459
 460                 if (runtime_exceeded(td, &td->tv_cache)) {
 461                         td->terminate = 1;
 462                         break;
 463                 }
 464
 465                 io_u = __get_io_u(td);
 466                 if (!io_u)
 467                         break;
 468
 469                 if (get_next_verify(td, io_u)) {
 470                         put_io_u(td, io_u);
 471                         break;
 472                 }
 473
 474                 if (td_io_prep(td, io_u)) {
 475                         put_io_u(td, io_u);
 476                         break;
 477                 }
 478
 479                 if (td->o.verify_async)
 480                         io_u->end_io = verify_io_u_async;
 481                 else
 482                         io_u->end_io = verify_io_u;
 483
 484                 ret = td_io_queue(td, io_u);
 485                 switch (ret) {
 486                 case FIO_Q_COMPLETED:
 487                         if (io_u->error) {
 488                                 ret = -io_u->error;
 489                                 clear_io_u(td, io_u);
 490                         } else if (io_u->resid) {
 491                                 int bytes = io_u->xfer_buflen - io_u->resid;
 492                                 struct fio_file *f = io_u->file;
 493
 494                                 /*
 495                                  * zero read, fail
 496                                  */
 497                                 if (!bytes) {
 498                                         td_verror(td, EIO, "full resid");
 499                                         put_io_u(td, io_u);
 500                                         break;
 501                                 }
 502
 503                                 io_u->xfer_buflen = io_u->resid;
 504                                 io_u->xfer_buf += bytes;
 505                                 io_u->offset += bytes;
 506
 507                                 td->ts.short_io_u[io_u->ddir]++;
 508
 509                                 if (io_u->offset == f->real_file_size)
 510                                         goto sync_done;
 511
 512                                 requeue_io_u(td, &io_u);
 513                         } else {
 514 sync_done:
 515                                 ret = io_u_sync_complete(td, io_u, NULL);
 516                                 if (ret < 0)
 517                                         break;
 518                         }
 519                         continue;
 520                 case FIO_Q_QUEUED:
 521                         break;
 522                 case FIO_Q_BUSY:
 523                         requeue_io_u(td, &io_u);
 524                         ret2 = td_io_commit(td);
 525                         if (ret2 < 0)
 526                                 ret = ret2;
 527                         break;
 528                 default:
 529                         assert(ret < 0);
 530                         td_verror(td, -ret, "td_io_queue");
 531                         break;
 532                 }
 533
 534                 if (break_on_this_error(td, &ret))
 535                         break;
 536
 537                 /*
 538                  * if we can queue more, do so. but check if there are
 539                  * completed io_u's first.
 540                  */
 541                 full = queue_full(td) || ret == FIO_Q_BUSY;
 542                 if (full || !td->o.iodepth_batch_complete) {
 543                         min_events = min(td->o.iodepth_batch_complete,
 544                                          td->cur_depth);
 545                         if (full && !min_events)
 546                                 min_events = 1;
 547
 548                         do {
 549                                 /*
 550                                  * Reap required number of io units, if any,
 551                                  * and do the verification on them through
 552                                  * the callback handler
 553                                  */
 554                                 if (io_u_queued_complete(td, min_events, NULL) < 0) {
 555                                         ret = -1;
 556                                         break;
 557                                 }
 558                         } while (full && (td->cur_depth > td->o.iodepth_low));
 559                 }
 560                 if (ret < 0)
 561                         break;
 562         }
 563
 564         if (!td->error) {
 565                 min_events = td->cur_depth;
 566
 567                 if (min_events)
 568                         ret = io_u_queued_complete(td, min_events, NULL);
 569         } else
 570                 cleanup_pending_aio(td);
 571
 572         td_set_runstate(td, TD_RUNNING);
 573 }
 574
 575 /*
 576  * Main IO worker function. It retrieves io_u's to process and queues
 577  * and reaps them, checking for rate and errors along the way.
 578  */
 579 static void do_io(struct thread_data *td)
 580 {
 581         unsigned int i;
 582         int ret = 0;
 583
 584         if (in_ramp_time(td))
 585                 td_set_runstate(td, TD_RAMP);
 586         else
 587                 td_set_runstate(td, TD_RUNNING);
 588
 589         while ( (td->o.read_iolog_file && !flist_empty(&td->io_log_list)) ||
 590                 ((td->this_io_bytes[0] + td->this_io_bytes[1]) < td->o.size) ) {
 591                 struct timeval comp_time;
 592                 unsigned long bytes_done[2] = { 0, 0 };
 593                 int min_evts = 0;
 594                 struct io_u *io_u;
 595                 int ret2, full;
 596
 597                 if (td->terminate)
 598                         break;
 599
 600                 update_tv_cache(td);
 601
 602                 if (runtime_exceeded(td, &td->tv_cache)) {
 603                         td->terminate = 1;
 604                         break;
 605                 }
 606
 607                 io_u = get_io_u(td);
 608                 if (!io_u)
 609                         break;
 610
 611                 /*
 612                  * Add verification end_io handler, if asked to verify
 613                  * a previously written file.
 614                  */
 615                 if (td->o.verify != VERIFY_NONE && io_u->ddir == DDIR_READ &&
 616                     !td_rw(td)) {
 617                         if (td->o.verify_async)
 618                                 io_u->end_io = verify_io_u_async;
 619                         else
 620                                 io_u->end_io = verify_io_u;
 621                         td_set_runstate(td, TD_VERIFYING);
 622                 } else if (in_ramp_time(td))
 623                         td_set_runstate(td, TD_RAMP);
 624                 else
 625                         td_set_runstate(td, TD_RUNNING);
 626
 627                 ret = td_io_queue(td, io_u);
 628                 switch (ret) {
 629                 case FIO_Q_COMPLETED:
 630                         if (io_u->error) {
 631                                 ret = -io_u->error;
 632                                 clear_io_u(td, io_u);
 633                         } else if (io_u->resid) {
 634                                 int bytes = io_u->xfer_buflen - io_u->resid;
 635                                 struct fio_file *f = io_u->file;
 636
 637                                 /*
 638                                  * zero read, fail
 639                                  */
 640                                 if (!bytes) {
 641                                         td_verror(td, EIO, "full resid");
 642                                         put_io_u(td, io_u);
 643                                         break;
 644                                 }
 645
 646                                 io_u->xfer_buflen = io_u->resid;
 647                                 io_u->xfer_buf += bytes;
 648                                 io_u->offset += bytes;
 649
 650                                 td->ts.short_io_u[io_u->ddir]++;
 651
 652                                 if (io_u->offset == f->real_file_size)
 653                                         goto sync_done;
 654
 655                                 requeue_io_u(td, &io_u);
 656                         } else {
 657 sync_done:
 658                                 if (__should_check_rate(td, 0) ||
 659                                     __should_check_rate(td, 1))
 660                                         fio_gettime(&comp_time, NULL);
 661
 662                                 ret = io_u_sync_complete(td, io_u, bytes_done);
 663                                 if (ret < 0)
 664                                         break;
 665                         }
 666                         break;
 667                 case FIO_Q_QUEUED:
 668                         /*
 669                          * if the engine doesn't have a commit hook,
 670                          * the io_u is really queued. if it does have such
 671                          * a hook, it has to call io_u_queued() itself.
 672                          */
 673                         if (td->io_ops->commit == NULL)
 674                                 io_u_queued(td, io_u);
 675                         break;
 676                 case FIO_Q_BUSY:
 677                         requeue_io_u(td, &io_u);
 678                         ret2 = td_io_commit(td);
 679                         if (ret2 < 0)
 680                                 ret = ret2;
 681                         break;
 682                 default:
 683                         assert(ret < 0);
 684                         put_io_u(td, io_u);
 685                         break;
 686                 }
 687
 688                 if (break_on_this_error(td, &ret))
 689                         break;
 690
 691                 /*
 692                  * See if we need to complete some commands
 693                  */
 694                 full = queue_full(td) || ret == FIO_Q_BUSY;
 695                 if (full || !td->o.iodepth_batch_complete) {
 696                         min_evts = min(td->o.iodepth_batch_complete,
 697                                         td->cur_depth);
 698                         if (full && !min_evts)
 699                                 min_evts = 1;
 700
 701                         if (__should_check_rate(td, 0) ||
 702                             __should_check_rate(td, 1))
 703                                 fio_gettime(&comp_time, NULL);
 704
 705                         do {
 706                                 ret = io_u_queued_complete(td, min_evts, bytes_done);
 707                                 if (ret < 0)
 708                                         break;
 709
 710                         } while (full && (td->cur_depth > td->o.iodepth_low));
 711                 }
 712
 713                 if (ret < 0)
 714                         break;
 715                 if (!(bytes_done[0] + bytes_done[1]))
 716                         continue;
 717
 718                 if (!in_ramp_time(td) && should_check_rate(td, bytes_done)) {
 719                         if (check_min_rate(td, &comp_time, bytes_done)) {
 720                                 if (exitall_on_terminate)
 721                                         terminate_threads(td->groupid);
 722                                 td_verror(td, EIO, "check_min_rate");
 723                                 break;
 724                         }
 725                 }
 726
 727                 if (td->o.thinktime) {
 728                         unsigned long long b;
 729
 730                         b = td->io_blocks[0] + td->io_blocks[1];
 731                         if (!(b % td->o.thinktime_blocks)) {
 732                                 int left;
 733
 734                                 if (td->o.thinktime_spin)
 735                                         usec_spin(td->o.thinktime_spin);
 736
 737                                 left = td->o.thinktime - td->o.thinktime_spin;
 738                                 if (left)
 739                                         usec_sleep(td, left);
 740                         }
 741                 }
 742         }
 743
 744         if (td->o.fill_device && td->error == ENOSPC) {
 745                 td->error = 0;
 746                 td->terminate = 1;
 747         }
 748         if (!td->error) {
 749                 struct fio_file *f;
 750
 751                 i = td->cur_depth;
 752                 if (i)
 753                         ret = io_u_queued_complete(td, i, NULL);
 754
 755                 if (should_fsync(td) && td->o.end_fsync) {
 756                         td_set_runstate(td, TD_FSYNCING);
 757
 758                         for_each_file(td, f, i) {
 759                                 if (!fio_file_open(f))
 760                                         continue;
 761                                 fio_io_sync(td, f);
 762                         }
 763                 }
 764         } else
 765                 cleanup_pending_aio(td);
 766
 767         /*
 768          * stop job if we failed doing any IO
 769          */
 770         if ((td->this_io_bytes[0] + td->this_io_bytes[1]) == 0)
 771                 td->done = 1;
 772 }
 773
 774 static void cleanup_io_u(struct thread_data *td)
 775 {
 776         struct flist_head *entry, *n;
 777         struct io_u *io_u;
 778
 779         flist_for_each_safe(entry, n, &td->io_u_freelist) {
 780                 io_u = flist_entry(entry, struct io_u, list);
 781
 782                 flist_del(&io_u->list);
 783                 free(io_u);
 784         }
 785
 786         free_io_mem(td);
 787 }
 788
 789 static int init_io_u(struct thread_data *td)
 790 {
 791         struct io_u *io_u;
 792         unsigned int max_bs;
 793         int cl_align, i, max_units;
 794         char *p;
 795
 796         max_units = td->o.iodepth;
 797         max_bs = max(td->o.max_bs[DDIR_READ], td->o.max_bs[DDIR_WRITE]);
 798         td->orig_buffer_size = (unsigned long long) max_bs
 799                                         * (unsigned long long) max_units;
 800
 801         if (td->o.mem_type == MEM_SHMHUGE || td->o.mem_type == MEM_MMAPHUGE) {
 802                 unsigned long bs;
 803
 804                 bs = td->orig_buffer_size + td->o.hugepage_size - 1;
 805                 td->orig_buffer_size = bs & ~(td->o.hugepage_size - 1);
 806         }
 807
 808         if (td->orig_buffer_size != (size_t) td->orig_buffer_size) {
 809                 log_err("fio: IO memory too large. Reduce max_bs or iodepth\n");
 810                 return 1;
 811         }
 812
 813         if (allocate_io_mem(td))
 814                 return 1;
 815
 816         if (td->o.odirect || td->o.mem_align)
 817                 p = PAGE_ALIGN(td->orig_buffer) + td->o.mem_align;
 818         else
 819                 p = td->orig_buffer;
 820
 821         cl_align = os_cache_line_size();
 822
 823         for (i = 0; i < max_units; i++) {
 824                 void *ptr;
 825
 826                 if (td->terminate)
 827                         return 1;
 828
 829                 if (posix_memalign(&ptr, cl_align, sizeof(*io_u))) {
 830                         log_err("fio: posix_memalign=%s\n", strerror(errno));
 831                         break;
 832                 }
 833
 834                 io_u = ptr;
 835                 memset(io_u, 0, sizeof(*io_u));
 836                 INIT_FLIST_HEAD(&io_u->list);
 837                 dprint(FD_MEM, "io_u alloc %p, index %u\n", io_u, i);
 838
 839                 if (!(td->io_ops->flags & FIO_NOIO)) {
 840                         io_u->buf = p + max_bs * i;
 841                         dprint(FD_MEM, "io_u %p, mem %p\n", io_u, io_u->buf);
 842
 843                         if (td_write(td) && !td->o.refill_buffers)
 844                                 io_u_fill_buffer(td, io_u, max_bs);
 845                 }
 846
 847                 io_u->index = i;
 848                 io_u->flags = IO_U_F_FREE;
 849                 flist_add(&io_u->list, &td->io_u_freelist);
 850         }
 851
 852         return 0;
 853 }
 854
 855 static int switch_ioscheduler(struct thread_data *td)
 856 {
 857         char tmp[256], tmp2[128];
 858         FILE *f;
 859         int ret;
 860
 861         if (td->io_ops->flags & FIO_DISKLESSIO)
 862                 return 0;
 863
 864         sprintf(tmp, "%s/queue/scheduler", td->sysfs_root);
 865
 866         f = fopen(tmp, "r+");
 867         if (!f) {
 868                 if (errno == ENOENT) {
 869                         log_err("fio: os or kernel doesn't support IO scheduler"
 870                                 " switching\n");
 871                         return 0;
 872                 }
 873                 td_verror(td, errno, "fopen iosched");
 874                 return 1;
 875         }
 876
 877         /*
 878          * Set io scheduler.
 879          */
 880         ret = fwrite(td->o.ioscheduler, strlen(td->o.ioscheduler), 1, f);
 881         if (ferror(f) || ret != 1) {
 882                 td_verror(td, errno, "fwrite");
 883                 fclose(f);
 884                 return 1;
 885         }
 886
 887         rewind(f);
 888
 889         /*
 890          * Read back and check that the selected scheduler is now the default.
 891          */
 892         ret = fread(tmp, 1, sizeof(tmp), f);
 893         if (ferror(f) || ret < 0) {
 894                 td_verror(td, errno, "fread");
 895                 fclose(f);
 896                 return 1;
 897         }
 898
 899         sprintf(tmp2, "[%s]", td->o.ioscheduler);
 900         if (!strstr(tmp, tmp2)) {
 901                 log_err("fio: io scheduler %s not found\n", td->o.ioscheduler);
 902                 td_verror(td, EINVAL, "iosched_switch");
 903                 fclose(f);
 904                 return 1;
 905         }
 906
 907         fclose(f);
 908         return 0;
 909 }
 910
 911 static int keep_running(struct thread_data *td)
 912 {
 913         unsigned long long io_done;
 914
 915         if (td->done)
 916                 return 0;
 917         if (td->o.time_based)
 918                 return 1;
 919         if (td->o.loops) {
 920                 td->o.loops--;
 921                 return 1;
 922         }
 923
 924         io_done = td->io_bytes[DDIR_READ] + td->io_bytes[DDIR_WRITE]
 925                         + td->io_skip_bytes;
 926         if (io_done < td->o.size)
 927                 return 1;
 928
 929         return 0;
 930 }
 931
 932 static void reset_io_counters(struct thread_data *td)
 933 {
 934         td->ts.stat_io_bytes[0] = td->ts.stat_io_bytes[1] = 0;
 935         td->this_io_bytes[0] = td->this_io_bytes[1] = 0;
 936         td->zone_bytes = 0;
 937         td->rate_bytes[0] = td->rate_bytes[1] = 0;
 938         td->rate_blocks[0] = td->rate_blocks[1] = 0;
 939
 940         td->last_was_sync = 0;
 941
 942         /*
 943          * reset file done count if we are to start over
 944          */
 945         if (td->o.time_based || td->o.loops)
 946                 td->nr_done_files = 0;
 947
 948         /*
 949          * Set the same seed to get repeatable runs
 950          */
 951         td_fill_rand_seeds(td);
 952 }
 953
 954 void reset_all_stats(struct thread_data *td)
 955 {
 956         struct timeval tv;
 957         int i;
 958
 959         reset_io_counters(td);
 960
 961         for (i = 0; i < 2; i++) {
 962                 td->io_bytes[i] = 0;
 963                 td->io_blocks[i] = 0;
 964                 td->io_issues[i] = 0;
 965                 td->ts.total_io_u[i] = 0;
 966         }
 967
 968         fio_gettime(&tv, NULL);
 969         memcpy(&td->epoch, &tv, sizeof(tv));
 970         memcpy(&td->start, &tv, sizeof(tv));
 971 }
 972
 973 static void clear_io_state(struct thread_data *td)
 974 {
 975         struct fio_file *f;
 976         unsigned int i;
 977
 978         reset_io_counters(td);
 979
 980         close_files(td);
 981         for_each_file(td, f, i)
 982                 fio_file_clear_done(f);
 983 }
 984
 985 static int exec_string(const char *string)
 986 {
 987         int ret, newlen = strlen(string) + 1 + 8;
 988         char *str;
 989
 990         str = malloc(newlen);
 991         sprintf(str, "sh -c %s", string);
 992
 993         ret = system(str);
 994         if (ret == -1)
 995                 log_err("fio: exec of cmd <%s> failed\n", str);
 996
 997         free(str);
 998         return ret;
 999 }
1000
1001 /*
1002  * Entry point for the thread based jobs. The process based jobs end up
1003  * here as well, after a little setup.
1004  */
1005 static void *thread_main(void *data)
1006 {
1007         unsigned long long runtime[2], elapsed;
1008         struct thread_data *td = data;
1009         pthread_condattr_t attr;
1010         int clear_state;
1011
1012         if (!td->o.use_thread)
1013                 setsid();
1014
1015         td->pid = getpid();
1016
1017         dprint(FD_PROCESS, "jobs pid=%d started\n", (int) td->pid);
1018
1019         INIT_FLIST_HEAD(&td->io_u_freelist);
1020         INIT_FLIST_HEAD(&td->io_u_busylist);
1021         INIT_FLIST_HEAD(&td->io_u_requeues);
1022         INIT_FLIST_HEAD(&td->io_log_list);
1023         INIT_FLIST_HEAD(&td->io_hist_list);
1024         INIT_FLIST_HEAD(&td->verify_list);
1025         pthread_mutex_init(&td->io_u_lock, NULL);
1026         td->io_hist_tree = RB_ROOT;
1027
1028         pthread_condattr_init(&attr);
1029         pthread_cond_init(&td->verify_cond, &attr);
1030         pthread_cond_init(&td->free_cond, &attr);
1031
1032         td_set_runstate(td, TD_INITIALIZED);
1033         dprint(FD_MUTEX, "up startup_mutex\n");
1034         fio_mutex_up(startup_mutex);
1035         dprint(FD_MUTEX, "wait on td->mutex\n");
1036         fio_mutex_down(td->mutex);
1037         dprint(FD_MUTEX, "done waiting on td->mutex\n");
1038
1039         /*
1040          * the ->mutex mutex is now no longer used, close it to avoid
1041          * eating a file descriptor
1042          */
1043         fio_mutex_remove(td->mutex);
1044
1045         if (td->o.uid != -1U && setuid(td->o.uid)) {
1046                 td_verror(td, errno, "setuid");
1047                 goto err;
1048         }
1049         if (td->o.gid != -1U && setgid(td->o.gid)) {
1050                 td_verror(td, errno, "setgid");
1051                 goto err;
1052         }
1053
1054         /*
1055          * May alter parameters that init_io_u() will use, so we need to
1056          * do this first.
1057          */
1058         if (init_iolog(td))
1059                 goto err;
1060
1061         if (init_io_u(td))
1062                 goto err;
1063
1064         if (td->o.verify_async && verify_async_init(td))
1065                 goto err;
1066
1067         if (td->o.cpumask_set && fio_setaffinity(td->pid, td->o.cpumask) == -1) {
1068                 td_verror(td, errno, "cpu_set_affinity");
1069                 goto err;
1070         }
1071
1072         /*
1073          * If we have a gettimeofday() thread, make sure we exclude that
1074          * thread from this job
1075          */
1076         if (td->o.gtod_cpu) {
1077                 fio_cpu_clear(&td->o.cpumask, td->o.gtod_cpu);
1078                 if (fio_setaffinity(td->pid, td->o.cpumask) == -1) {
1079                         td_verror(td, errno, "cpu_set_affinity");
1080                         goto err;
1081                 }
1082         }
1083
1084         if (td->ioprio_set) {
1085                 if (ioprio_set(IOPRIO_WHO_PROCESS, 0, td->ioprio) == -1) {
1086                         td_verror(td, errno, "ioprio_set");
1087                         goto err;
1088                 }
1089         }
1090
1091         if (td->o.cgroup_weight && cgroup_setup(td, cgroup_list, &cgroup_mnt))
1092                 goto err;
1093
1094         if (nice(td->o.nice) == -1) {
1095                 td_verror(td, errno, "nice");
1096                 goto err;
1097         }
1098
1099         if (td->o.ioscheduler && switch_ioscheduler(td))
1100                 goto err;
1101
1102         if (!td->o.create_serialize && setup_files(td))
1103                 goto err;
1104
1105         if (td_io_init(td))
1106                 goto err;
1107
1108         if (init_random_map(td))
1109                 goto err;
1110
1111         if (td->o.exec_prerun) {
1112                 if (exec_string(td->o.exec_prerun))
1113                         goto err;
1114         }
1115
1116         if (td->o.pre_read) {
1117                 if (pre_read_files(td) < 0)
1118                         goto err;
1119         }
1120
1121         fio_gettime(&td->epoch, NULL);
1122         getrusage(RUSAGE_SELF, &td->ts.ru_start);
1123
1124         runtime[0] = runtime[1] = 0;
1125         clear_state = 0;
1126         while (keep_running(td)) {
1127                 fio_gettime(&td->start, NULL);
1128                 memcpy(&td->ts.stat_sample_time[0], &td->start,
1129                                 sizeof(td->start));
1130                 memcpy(&td->ts.stat_sample_time[1], &td->start,
1131                                 sizeof(td->start));
1132                 memcpy(&td->tv_cache, &td->start, sizeof(td->start));
1133
1134                 if (td->o.ratemin[0] || td->o.ratemin[1])
1135                         memcpy(&td->lastrate, &td->ts.stat_sample_time,
1136                                                         sizeof(td->lastrate));
1137
1138                 if (clear_state)
1139                         clear_io_state(td);
1140
1141                 prune_io_piece_log(td);
1142
1143                 do_io(td);
1144
1145                 clear_state = 1;
1146
1147                 if (td_read(td) && td->io_bytes[DDIR_READ]) {
1148                         elapsed = utime_since_now(&td->start);
1149                         runtime[DDIR_READ] += elapsed;
1150                 }
1151                 if (td_write(td) && td->io_bytes[DDIR_WRITE]) {
1152                         elapsed = utime_since_now(&td->start);
1153                         runtime[DDIR_WRITE] += elapsed;
1154                 }
1155
1156                 if (td->error || td->terminate)
1157                         break;
1158
1159                 if (!td->o.do_verify ||
1160                     td->o.verify == VERIFY_NONE ||
1161                     (td->io_ops->flags & FIO_UNIDIR))
1162                         continue;
1163
1164                 clear_io_state(td);
1165
1166                 fio_gettime(&td->start, NULL);
1167
1168                 do_verify(td);
1169
1170                 runtime[DDIR_READ] += utime_since_now(&td->start);
1171
1172                 if (td->error || td->terminate)
1173                         break;
1174         }
1175
1176         update_rusage_stat(td);
1177         td->ts.runtime[0] = (runtime[0] + 999) / 1000;
1178         td->ts.runtime[1] = (runtime[1] + 999) / 1000;
1179         td->ts.total_run_time = mtime_since_now(&td->epoch);
1180         td->ts.io_bytes[0] = td->io_bytes[0];
1181         td->ts.io_bytes[1] = td->io_bytes[1];
1182
1183         fio_mutex_down(writeout_mutex);
1184         if (td->ts.bw_log) {
1185                 if (td->o.bw_log_file) {
1186                         finish_log_named(td, td->ts.bw_log,
1187                                                 td->o.bw_log_file, "bw");
1188                 } else
1189                         finish_log(td, td->ts.bw_log, "bw");
1190         }
1191         if (td->ts.slat_log) {
1192                 if (td->o.lat_log_file) {
1193                         finish_log_named(td, td->ts.slat_log,
1194                                                 td->o.lat_log_file, "slat");
1195                 } else
1196                         finish_log(td, td->ts.slat_log, "slat");
1197         }
1198         if (td->ts.clat_log) {
1199                 if (td->o.lat_log_file) {
1200                         finish_log_named(td, td->ts.clat_log,
1201                                                 td->o.lat_log_file, "clat");
1202                 } else
1203                         finish_log(td, td->ts.clat_log, "clat");
1204         }
1205         fio_mutex_up(writeout_mutex);
1206         if (td->o.exec_postrun)
1207                 exec_string(td->o.exec_postrun);
1208
1209         if (exitall_on_terminate)
1210                 terminate_threads(td->groupid);
1211
1212 err:
1213         if (td->error)
1214                 log_info("fio: pid=%d, err=%d/%s\n", (int) td->pid, td->error,
1215                                                         td->verror);
1216
1217         if (td->o.verify_async)
1218                 verify_async_exit(td);
1219
1220         close_and_free_files(td);
1221         close_ioengine(td);
1222         cleanup_io_u(td);
1223         cgroup_shutdown(td, &cgroup_mnt);
1224
1225         if (td->o.cpumask_set) {
1226                 int ret = fio_cpuset_exit(&td->o.cpumask);
1227
1228                 td_verror(td, ret, "fio_cpuset_exit");
1229         }
1230
1231         /*
1232          * do this very late, it will log file closing as well
1233          */
1234         if (td->o.write_iolog_file)
1235                 write_iolog_close(td);
1236
1237         options_mem_free(td);
1238         td_set_runstate(td, TD_EXITED);
1239         return (void *) (unsigned long) td->error;
1240 }
1241
1242 /*
1243  * We cannot pass the td data into a forked process, so attach the td and
1244  * pass it to the thread worker.
1245  */
1246 static int fork_main(int shmid, int offset)
1247 {
1248         struct thread_data *td;
1249         void *data, *ret;
1250
1251         data = shmat(shmid, NULL, 0);
1252         if (data == (void *) -1) {
1253                 int __err = errno;
1254
1255                 perror("shmat");
1256                 return __err;
1257         }
1258
1259         td = data + offset * sizeof(struct thread_data);
1260         ret = thread_main(td);
1261         shmdt(data);
1262         return (int) (unsigned long) ret;
1263 }
1264
1265 /*
1266  * Run over the job map and reap the threads that have exited, if any.
1267  */
1268 static void reap_threads(int *nr_running, int *t_rate, int *m_rate)
1269 {
1270         struct thread_data *td;
1271         int i, cputhreads, realthreads, pending, status, ret;
1272
1273         /*
1274          * reap exited threads (TD_EXITED -> TD_REAPED)
1275          */
1276         realthreads = pending = cputhreads = 0;
1277         for_each_td(td, i) {
1278                 int flags = 0;
1279
1280                 /*
1281                  * ->io_ops is NULL for a thread that has closed its
1282                  * io engine
1283                  */
1284                 if (td->io_ops && !strcmp(td->io_ops->name, "cpuio"))
1285                         cputhreads++;
1286                 else
1287                         realthreads++;
1288
1289                 if (!td->pid) {
1290                         pending++;
1291                         continue;
1292                 }
1293                 if (td->runstate == TD_REAPED)
1294                         continue;
1295                 if (td->o.use_thread) {
1296                         if (td->runstate == TD_EXITED) {
1297                                 td_set_runstate(td, TD_REAPED);
1298                                 goto reaped;
1299                         }
1300                         continue;
1301                 }
1302
1303                 flags = WNOHANG;
1304                 if (td->runstate == TD_EXITED)
1305                         flags = 0;
1306
1307                 /*
1308                  * check if someone quit or got killed in an unusual way
1309                  */
1310                 ret = waitpid(td->pid, &status, flags);
1311                 if (ret < 0) {
1312                         if (errno == ECHILD) {
1313                                 log_err("fio: pid=%d disappeared %d\n",
1314                                                 (int) td->pid, td->runstate);
1315                                 td_set_runstate(td, TD_REAPED);
1316                                 goto reaped;
1317                         }
1318                         perror("waitpid");
1319                 } else if (ret == td->pid) {
1320                         if (WIFSIGNALED(status)) {
1321                                 int sig = WTERMSIG(status);
1322
1323                                 if (sig != SIGQUIT)
1324                                         log_err("fio: pid=%d, got signal=%d\n",
1325                                                         (int) td->pid, sig);
1326                                 td_set_runstate(td, TD_REAPED);
1327                                 goto reaped;
1328                         }
1329                         if (WIFEXITED(status)) {
1330                                 if (WEXITSTATUS(status) && !td->error)
1331                                         td->error = WEXITSTATUS(status);
1332
1333                                 td_set_runstate(td, TD_REAPED);
1334                                 goto reaped;
1335                         }
1336                 }
1337
1338                 /*
1339                  * thread is not dead, continue
1340                  */
1341                 pending++;
1342                 continue;
1343 reaped:
1344                 (*nr_running)--;
1345                 (*m_rate) -= (td->o.ratemin[0] + td->o.ratemin[1]);
1346                 (*t_rate) -= (td->o.rate[0] + td->o.rate[1]);
1347                 if (!td->pid)
1348                         pending--;
1349
1350                 if (td->error)
1351                         exit_value++;
1352
1353                 done_secs += mtime_since_now(&td->epoch) / 1000;
1354         }
1355
1356         if (*nr_running == cputhreads && !pending && realthreads)
1357                 terminate_threads(TERMINATE_ALL);
1358 }
1359
1360 static void *gtod_thread_main(void *data)
1361 {
1362         fio_mutex_up(startup_mutex);
1363
1364         /*
1365          * As long as we have jobs around, update the clock. It would be nice
1366          * to have some way of NOT hammering that CPU with gettimeofday(),
1367          * but I'm not sure what to use outside of a simple CPU nop to relax
1368          * it - we don't want to lose precision.
1369          */
1370         while (threads) {
1371                 fio_gtod_update();
1372                 nop;
1373         }
1374
1375         return NULL;
1376 }
1377
1378 static int fio_start_gtod_thread(void)
1379 {
1380         int ret;
1381
1382         ret = pthread_create(&gtod_thread, NULL, gtod_thread_main, NULL);
1383         if (ret) {
1384                 log_err("Can't create gtod thread: %s\n", strerror(ret));
1385                 return 1;
1386         }
1387
1388         ret = pthread_detach(gtod_thread);
1389         if (ret) {
1390                 log_err("Can't detatch gtod thread: %s\n", strerror(ret));
1391                 return 1;
1392         }
1393
1394         dprint(FD_MUTEX, "wait on startup_mutex\n");
1395         fio_mutex_down(startup_mutex);
1396         dprint(FD_MUTEX, "done waiting on startup_mutex\n");
1397         return 0;
1398 }
1399
1400 /*
1401  * Main function for kicking off and reaping jobs, as needed.
1402  */
1403 static void run_threads(void)
1404 {
1405         struct thread_data *td;
1406         unsigned long spent;
1407         int i, todo, nr_running, m_rate, t_rate, nr_started;
1408
1409         if (fio_pin_memory())
1410                 return;
1411
1412         if (fio_gtod_offload && fio_start_gtod_thread())
1413                 return;
1414
1415         if (!terse_output) {
1416                 log_info("Starting ");
1417                 if (nr_thread)
1418                         log_info("%d thread%s", nr_thread,
1419                                                 nr_thread > 1 ? "s" : "");
1420                 if (nr_process) {
1421                         if (nr_thread)
1422                                 printf(" and ");
1423                         log_info("%d process%s", nr_process,
1424                                                 nr_process > 1 ? "es" : "");
1425                 }
1426                 log_info("\n");
1427                 fflush(stdout);
1428         }
1429
1430         set_sig_handlers();
1431
1432         todo = thread_number;
1433         nr_running = 0;
1434         nr_started = 0;
1435         m_rate = t_rate = 0;
1436
1437         for_each_td(td, i) {
1438                 print_status_init(td->thread_number - 1);
1439
1440                 if (!td->o.create_serialize) {
1441                         init_disk_util(td);
1442                         continue;
1443                 }
1444
1445                 /*
1446                  * do file setup here so it happens sequentially,
1447                  * we don't want X number of threads getting their
1448                  * client data interspersed on disk
1449                  */
1450                 if (setup_files(td)) {
1451                         exit_value++;
1452                         if (td->error)
1453                                 log_err("fio: pid=%d, err=%d/%s\n",
1454                                         (int) td->pid, td->error, td->verror);
1455                         td_set_runstate(td, TD_REAPED);
1456                         todo--;
1457                 } else {
1458                         struct fio_file *f;
1459                         unsigned int i;
1460
1461                         /*
1462                          * for sharing to work, each job must always open
1463                          * its own files. so close them, if we opened them
1464                          * for creation
1465                          */
1466                         for_each_file(td, f, i) {
1467                                 if (fio_file_open(f))
1468                                         td_io_close_file(td, f);
1469                         }
1470                 }
1471
1472                 init_disk_util(td);
1473         }
1474
1475         set_genesis_time();
1476
1477         while (todo) {
1478                 struct thread_data *map[MAX_JOBS];
1479                 struct timeval this_start;
1480                 int this_jobs = 0, left;
1481
1482                 /*
1483                  * create threads (TD_NOT_CREATED -> TD_CREATED)
1484                  */
1485                 for_each_td(td, i) {
1486                         if (td->runstate != TD_NOT_CREATED)
1487                                 continue;
1488
1489                         /*
1490                          * never got a chance to start, killed by other
1491                          * thread for some reason
1492                          */
1493                         if (td->terminate) {
1494                                 todo--;
1495                                 continue;
1496                         }
1497
1498                         if (td->o.start_delay) {
1499                                 spent = mtime_since_genesis();
1500
1501                                 if (td->o.start_delay * 1000 > spent)
1502                                         continue;
1503                         }
1504
1505                         if (td->o.stonewall && (nr_started || nr_running)) {
1506                                 dprint(FD_PROCESS, "%s: stonewall wait\n",
1507                                                         td->o.name);
1508                                 break;
1509                         }
1510
1511                         /*
1512                          * Set state to created. Thread will transition
1513                          * to TD_INITIALIZED when it's done setting up.
1514                          */
1515                         td_set_runstate(td, TD_CREATED);
1516                         map[this_jobs++] = td;
1517                         nr_started++;
1518
1519                         if (td->o.use_thread) {
1520                                 int ret;
1521
1522                                 dprint(FD_PROCESS, "will pthread_create\n");
1523                                 ret = pthread_create(&td->thread, NULL,
1524                                                         thread_main, td);
1525                                 if (ret) {
1526                                         log_err("pthread_create: %s\n",
1527                                                         strerror(ret));
1528                                         nr_started--;
1529                                         break;
1530                                 }
1531                                 ret = pthread_detach(td->thread);
1532                                 if (ret)
1533                                         log_err("pthread_detach: %s",
1534                                                         strerror(ret));
1535                         } else {
1536                                 pid_t pid;
1537                                 dprint(FD_PROCESS, "will fork\n");
1538                                 pid = fork();
1539                                 if (!pid) {
1540                                         int ret = fork_main(shm_id, i);
1541
1542                                         _exit(ret);
1543                                 } else if (i == fio_debug_jobno)
1544                                         *fio_debug_jobp = pid;
1545                         }
1546                         dprint(FD_MUTEX, "wait on startup_mutex\n");
1547                         if (fio_mutex_down_timeout(startup_mutex, 10)) {
1548                                 log_err("fio: job startup hung? exiting.\n");
1549                                 terminate_threads(TERMINATE_ALL);
1550                                 fio_abort = 1;
1551                                 nr_started--;
1552                                 break;
1553                         }
1554                         dprint(FD_MUTEX, "done waiting on startup_mutex\n");
1555                 }
1556
1557                 /*
1558                  * Wait for the started threads to transition to
1559                  * TD_INITIALIZED.
1560                  */
1561                 fio_gettime(&this_start, NULL);
1562                 left = this_jobs;
1563                 while (left && !fio_abort) {
1564                         if (mtime_since_now(&this_start) > JOB_START_TIMEOUT)
1565                                 break;
1566
1567                         usleep(100000);
1568
1569                         for (i = 0; i < this_jobs; i++) {
1570                                 td = map[i];
1571                                 if (!td)
1572                                         continue;
1573                                 if (td->runstate == TD_INITIALIZED) {
1574                                         map[i] = NULL;
1575                                         left--;
1576                                 } else if (td->runstate >= TD_EXITED) {
1577                                         map[i] = NULL;
1578                                         left--;
1579                                         todo--;
1580                                         nr_running++; /* work-around... */
1581                                 }
1582                         }
1583                 }
1584
1585                 if (left) {
1586                         log_err("fio: %d jobs failed to start\n", left);
1587                         for (i = 0; i < this_jobs; i++) {
1588                                 td = map[i];
1589                                 if (!td)
1590                                         continue;
1591                                 kill(td->pid, SIGTERM);
1592                         }
1593                         break;
1594                 }
1595
1596                 /*
1597                  * start created threads (TD_INITIALIZED -> TD_RUNNING).
1598                  */
1599                 for_each_td(td, i) {
1600                         if (td->runstate != TD_INITIALIZED)
1601                                 continue;
1602
1603                         if (in_ramp_time(td))
1604                                 td_set_runstate(td, TD_RAMP);
1605                         else
1606                                 td_set_runstate(td, TD_RUNNING);
1607                         nr_running++;
1608                         nr_started--;
1609                         m_rate += td->o.ratemin[0] + td->o.ratemin[1];
1610                         t_rate += td->o.rate[0] + td->o.rate[1];
1611                         todo--;
1612                         fio_mutex_up(td->mutex);
1613                 }
1614
1615                 reap_threads(&nr_running, &t_rate, &m_rate);
1616
1617                 if (todo)
1618                         usleep(100000);
1619         }
1620
1621         while (nr_running) {
1622                 reap_threads(&nr_running, &t_rate, &m_rate);
1623                 usleep(10000);
1624         }
1625
1626         update_io_ticks();
1627         fio_unpin_memory();
1628 }
1629
1630 int main(int argc, char *argv[])
1631 {
1632         long ps;
1633
1634         sinit();
1635
1636         /*
1637          * We need locale for number printing, if it isn't set then just
1638          * go with the US format.
1639          */
1640         if (!getenv("LC_NUMERIC"))
1641                 setlocale(LC_NUMERIC, "en_US");
1642
1643         ps = sysconf(_SC_PAGESIZE);
1644         if (ps < 0) {
1645                 log_err("Failed to get page size\n");
1646                 return 1;
1647         }
1648
1649         page_size = ps;
1650         page_mask = ps - 1;
1651
1652         fio_keywords_init();
1653
1654         if (parse_options(argc, argv))
1655                 return 1;
1656
1657         if (exec_profile && load_profile(exec_profile))
1658                 return 1;
1659
1660         if (!thread_number)
1661                 return 0;
1662
1663         if (write_bw_log) {
1664                 setup_log(&agg_io_log[DDIR_READ]);
1665                 setup_log(&agg_io_log[DDIR_WRITE]);
1666         }
1667
1668         startup_mutex = fio_mutex_init(0);
1669         writeout_mutex = fio_mutex_init(1);
1670
1671         set_genesis_time();
1672
1673         status_timer_arm();
1674
1675         cgroup_list = smalloc(sizeof(*cgroup_list));
1676         INIT_FLIST_HEAD(cgroup_list);
1677
1678         run_threads();
1679
1680         if (!fio_abort) {
1681                 show_run_stats();
1682                 if (write_bw_log) {
1683                         __finish_log(agg_io_log[DDIR_READ], "agg-read_bw.log");
1684                         __finish_log(agg_io_log[DDIR_WRITE],
1685                                         "agg-write_bw.log");
1686                 }
1687         }
1688
1689         cgroup_kill(cgroup_list);
1690         sfree(cgroup_list);
1691         if (cgroup_mnt)
1692                 sfree(cgroup_mnt);
1693
1694         fio_mutex_remove(startup_mutex);
1695         fio_mutex_remove(writeout_mutex);
1696         return exit_value;
1697 }