backend.c

   1 /*
   2  * fio - the flexible io tester
   3  *
   4  * Copyright (C) 2005 Jens Axboe <axboe@suse.de>
   5  * Copyright (C) 2006-2012 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 <limits.h>
  28 #include <signal.h>
  29 #include <time.h>
  30 #include <locale.h>
  31 #include <assert.h>
  32 #include <time.h>
  33 #include <inttypes.h>
  34 #include <sys/stat.h>
  35 #include <sys/wait.h>
  36 #include <sys/ipc.h>
  37 #include <sys/mman.h>
  38
  39 #include "fio.h"
  40 #ifndef FIO_NO_HAVE_SHM_H
  41 #include <sys/shm.h>
  42 #endif
  43 #include "hash.h"
  44 #include "smalloc.h"
  45 #include "verify.h"
  46 #include "trim.h"
  47 #include "diskutil.h"
  48 #include "cgroup.h"
  49 #include "profile.h"
  50 #include "lib/rand.h"
  51 #include "memalign.h"
  52 #include "server.h"
  53 #include "lib/getrusage.h"
  54 #include "idletime.h"
  55 #include "err.h"
  56 #include "lib/tp.h"
  57
  58 static pthread_t helper_thread;
  59 static pthread_mutex_t helper_lock;
  60 pthread_cond_t helper_cond;
  61 int helper_do_stat = 0;
  62
  63 static struct fio_mutex *startup_mutex;
  64 static struct flist_head *cgroup_list;
  65 static char *cgroup_mnt;
  66 static int exit_value;
  67 static volatile int fio_abort;
  68 static unsigned int nr_process = 0;
  69 static unsigned int nr_thread = 0;
  70
  71 struct io_log *agg_io_log[DDIR_RWDIR_CNT];
  72
  73 int groupid = 0;
  74 unsigned int thread_number = 0;
  75 unsigned int stat_number = 0;
  76 int shm_id = 0;
  77 int temp_stall_ts;
  78 unsigned long done_secs = 0;
  79 volatile int helper_exit = 0;
  80
  81 #define PAGE_ALIGN(buf) \
  82         (char *) (((uintptr_t) (buf) + page_mask) & ~page_mask)
  83
  84 #define JOB_START_TIMEOUT       (5 * 1000)
  85
  86 static void sig_int(int sig)
  87 {
  88         if (threads) {
  89                 if (is_backend)
  90                         fio_server_got_signal(sig);
  91                 else {
  92                         log_info("\nfio: terminating on signal %d\n", sig);
  93                         log_info_flush();
  94                         exit_value = 128;
  95                 }
  96
  97                 fio_terminate_threads(TERMINATE_ALL);
  98         }
  99 }
 100
 101 static void sig_show_status(int sig)
 102 {
 103         show_running_run_stats();
 104 }
 105
 106 static void set_sig_handlers(void)
 107 {
 108         struct sigaction act;
 109
 110         memset(&act, 0, sizeof(act));
 111         act.sa_handler = sig_int;
 112         act.sa_flags = SA_RESTART;
 113         sigaction(SIGINT, &act, NULL);
 114
 115         memset(&act, 0, sizeof(act));
 116         act.sa_handler = sig_int;
 117         act.sa_flags = SA_RESTART;
 118         sigaction(SIGTERM, &act, NULL);
 119
 120 /* Windows uses SIGBREAK as a quit signal from other applications */
 121 #ifdef WIN32
 122         memset(&act, 0, sizeof(act));
 123         act.sa_handler = sig_int;
 124         act.sa_flags = SA_RESTART;
 125         sigaction(SIGBREAK, &act, NULL);
 126 #endif
 127
 128         memset(&act, 0, sizeof(act));
 129         act.sa_handler = sig_show_status;
 130         act.sa_flags = SA_RESTART;
 131         sigaction(SIGUSR1, &act, NULL);
 132
 133         if (is_backend) {
 134                 memset(&act, 0, sizeof(act));
 135                 act.sa_handler = sig_int;
 136                 act.sa_flags = SA_RESTART;
 137                 sigaction(SIGPIPE, &act, NULL);
 138         }
 139 }
 140
 141 /*
 142  * Check if we are above the minimum rate given.
 143  */
 144 static int __check_min_rate(struct thread_data *td, struct timeval *now,
 145                             enum fio_ddir ddir)
 146 {
 147         unsigned long long bytes = 0;
 148         unsigned long iops = 0;
 149         unsigned long spent;
 150         unsigned long rate;
 151         unsigned int ratemin = 0;
 152         unsigned int rate_iops = 0;
 153         unsigned int rate_iops_min = 0;
 154
 155         assert(ddir_rw(ddir));
 156
 157         if (!td->o.ratemin[ddir] && !td->o.rate_iops_min[ddir])
 158                 return 0;
 159
 160         /*
 161          * allow a 2 second settle period in the beginning
 162          */
 163         if (mtime_since(&td->start, now) < 2000)
 164                 return 0;
 165
 166         iops += td->this_io_blocks[ddir];
 167         bytes += td->this_io_bytes[ddir];
 168         ratemin += td->o.ratemin[ddir];
 169         rate_iops += td->o.rate_iops[ddir];
 170         rate_iops_min += td->o.rate_iops_min[ddir];
 171
 172         /*
 173          * if rate blocks is set, sample is running
 174          */
 175         if (td->rate_bytes[ddir] || td->rate_blocks[ddir]) {
 176                 spent = mtime_since(&td->lastrate[ddir], now);
 177                 if (spent < td->o.ratecycle)
 178                         return 0;
 179
 180                 if (td->o.rate[ddir]) {
 181                         /*
 182                          * check bandwidth specified rate
 183                          */
 184                         if (bytes < td->rate_bytes[ddir]) {
 185                                 log_err("%s: min rate %u not met\n", td->o.name,
 186                                                                 ratemin);
 187                                 return 1;
 188                         } else {
 189                                 if (spent)
 190                                         rate = ((bytes - td->rate_bytes[ddir]) * 1000) / spent;
 191                                 else
 192                                         rate = 0;
 193
 194                                 if (rate < ratemin ||
 195                                     bytes < td->rate_bytes[ddir]) {
 196                                         log_err("%s: min rate %u not met, got"
 197                                                 " %luKB/sec\n", td->o.name,
 198                                                         ratemin, rate);
 199                                         return 1;
 200                                 }
 201                         }
 202                 } else {
 203                         /*
 204                          * checks iops specified rate
 205                          */
 206                         if (iops < rate_iops) {
 207                                 log_err("%s: min iops rate %u not met\n",
 208                                                 td->o.name, rate_iops);
 209                                 return 1;
 210                         } else {
 211                                 if (spent)
 212                                         rate = ((iops - td->rate_blocks[ddir]) * 1000) / spent;
 213                                 else
 214                                         rate = 0;
 215
 216                                 if (rate < rate_iops_min ||
 217                                     iops < td->rate_blocks[ddir]) {
 218                                         log_err("%s: min iops rate %u not met,"
 219                                                 " got %lu\n", td->o.name,
 220                                                         rate_iops_min, rate);
 221                                 }
 222                         }
 223                 }
 224         }
 225
 226         td->rate_bytes[ddir] = bytes;
 227         td->rate_blocks[ddir] = iops;
 228         memcpy(&td->lastrate[ddir], now, sizeof(*now));
 229         return 0;
 230 }
 231
 232 static int check_min_rate(struct thread_data *td, struct timeval *now,
 233                           uint64_t *bytes_done)
 234 {
 235         int ret = 0;
 236
 237         if (bytes_done[DDIR_READ])
 238                 ret |= __check_min_rate(td, now, DDIR_READ);
 239         if (bytes_done[DDIR_WRITE])
 240                 ret |= __check_min_rate(td, now, DDIR_WRITE);
 241         if (bytes_done[DDIR_TRIM])
 242                 ret |= __check_min_rate(td, now, DDIR_TRIM);
 243
 244         return ret;
 245 }
 246
 247 /*
 248  * When job exits, we can cancel the in-flight IO if we are using async
 249  * io. Attempt to do so.
 250  */
 251 static void cleanup_pending_aio(struct thread_data *td)
 252 {
 253         int r;
 254
 255         /*
 256          * get immediately available events, if any
 257          */
 258         r = io_u_queued_complete(td, 0, NULL);
 259         if (r < 0)
 260                 return;
 261
 262         /*
 263          * now cancel remaining active events
 264          */
 265         if (td->io_ops->cancel) {
 266                 struct io_u *io_u;
 267                 int i;
 268
 269                 io_u_qiter(&td->io_u_all, io_u, i) {
 270                         if (io_u->flags & IO_U_F_FLIGHT) {
 271                                 r = td->io_ops->cancel(td, io_u);
 272                                 if (!r)
 273                                         put_io_u(td, io_u);
 274                         }
 275                 }
 276         }
 277
 278         if (td->cur_depth)
 279                 r = io_u_queued_complete(td, td->cur_depth, NULL);
 280 }
 281
 282 /*
 283  * Helper to handle the final sync of a file. Works just like the normal
 284  * io path, just does everything sync.
 285  */
 286 static int fio_io_sync(struct thread_data *td, struct fio_file *f)
 287 {
 288         struct io_u *io_u = __get_io_u(td);
 289         int ret;
 290
 291         if (!io_u)
 292                 return 1;
 293
 294         io_u->ddir = DDIR_SYNC;
 295         io_u->file = f;
 296
 297         if (td_io_prep(td, io_u)) {
 298                 put_io_u(td, io_u);
 299                 return 1;
 300         }
 301
 302 requeue:
 303         ret = td_io_queue(td, io_u);
 304         if (ret < 0) {
 305                 td_verror(td, io_u->error, "td_io_queue");
 306                 put_io_u(td, io_u);
 307                 return 1;
 308         } else if (ret == FIO_Q_QUEUED) {
 309                 if (io_u_queued_complete(td, 1, NULL) < 0)
 310                         return 1;
 311         } else if (ret == FIO_Q_COMPLETED) {
 312                 if (io_u->error) {
 313                         td_verror(td, io_u->error, "td_io_queue");
 314                         return 1;
 315                 }
 316
 317                 if (io_u_sync_complete(td, io_u, NULL) < 0)
 318                         return 1;
 319         } else if (ret == FIO_Q_BUSY) {
 320                 if (td_io_commit(td))
 321                         return 1;
 322                 goto requeue;
 323         }
 324
 325         return 0;
 326 }
 327
 328 static int fio_file_fsync(struct thread_data *td, struct fio_file *f)
 329 {
 330         int ret;
 331
 332         if (fio_file_open(f))
 333                 return fio_io_sync(td, f);
 334
 335         if (td_io_open_file(td, f))
 336                 return 1;
 337
 338         ret = fio_io_sync(td, f);
 339         td_io_close_file(td, f);
 340         return ret;
 341 }
 342
 343 static inline void __update_tv_cache(struct thread_data *td)
 344 {
 345         fio_gettime(&td->tv_cache, NULL);
 346 }
 347
 348 static inline void update_tv_cache(struct thread_data *td)
 349 {
 350         if ((++td->tv_cache_nr & td->tv_cache_mask) == td->tv_cache_mask)
 351                 __update_tv_cache(td);
 352 }
 353
 354 static inline int runtime_exceeded(struct thread_data *td, struct timeval *t)
 355 {
 356         if (in_ramp_time(td))
 357                 return 0;
 358         if (!td->o.timeout)
 359                 return 0;
 360         if (utime_since(&td->epoch, t) >= td->o.timeout)
 361                 return 1;
 362
 363         return 0;
 364 }
 365
 366 static int break_on_this_error(struct thread_data *td, enum fio_ddir ddir,
 367                                int *retptr)
 368 {
 369         int ret = *retptr;
 370
 371         if (ret < 0 || td->error) {
 372                 int err = td->error;
 373                 enum error_type_bit eb;
 374
 375                 if (ret < 0)
 376                         err = -ret;
 377
 378                 eb = td_error_type(ddir, err);
 379                 if (!(td->o.continue_on_error & (1 << eb)))
 380                         return 1;
 381
 382                 if (td_non_fatal_error(td, eb, err)) {
 383                         /*
 384                          * Continue with the I/Os in case of
 385                          * a non fatal error.
 386                          */
 387                         update_error_count(td, err);
 388                         td_clear_error(td);
 389                         *retptr = 0;
 390                         return 0;
 391                 } else if (td->o.fill_device && err == ENOSPC) {
 392                         /*
 393                          * We expect to hit this error if
 394                          * fill_device option is set.
 395                          */
 396                         td_clear_error(td);
 397                         fio_mark_td_terminate(td);
 398                         return 1;
 399                 } else {
 400                         /*
 401                          * Stop the I/O in case of a fatal
 402                          * error.
 403                          */
 404                         update_error_count(td, err);
 405                         return 1;
 406                 }
 407         }
 408
 409         return 0;
 410 }
 411
 412 static void check_update_rusage(struct thread_data *td)
 413 {
 414         if (td->update_rusage) {
 415                 td->update_rusage = 0;
 416                 update_rusage_stat(td);
 417                 fio_mutex_up(td->rusage_sem);
 418         }
 419 }
 420
 421 static int wait_for_completions(struct thread_data *td, struct timeval *time,
 422                                 uint64_t *bytes_done)
 423 {
 424         const int full = queue_full(td);
 425         int min_evts = 0;
 426         int ret;
 427
 428         /*
 429          * if the queue is full, we MUST reap at least 1 event
 430          */
 431         min_evts = min(td->o.iodepth_batch_complete, td->cur_depth);
 432         if (full && !min_evts)
 433                 min_evts = 1;
 434
 435         if (time && (__should_check_rate(td, DDIR_READ) ||
 436             __should_check_rate(td, DDIR_WRITE) ||
 437             __should_check_rate(td, DDIR_TRIM)))
 438                 fio_gettime(time, NULL);
 439
 440         do {
 441                 ret = io_u_queued_complete(td, min_evts, bytes_done);
 442                 if (ret < 0)
 443                         break;
 444         } while (full && (td->cur_depth > td->o.iodepth_low));
 445
 446         return ret;
 447 }
 448
 449 /*
 450  * The main verify engine. Runs over the writes we previously submitted,
 451  * reads the blocks back in, and checks the crc/md5 of the data.
 452  */
 453 static void do_verify(struct thread_data *td, uint64_t verify_bytes)
 454 {
 455         uint64_t bytes_done[DDIR_RWDIR_CNT] = { 0, 0, 0 };
 456         struct fio_file *f;
 457         struct io_u *io_u;
 458         int ret, min_events;
 459         unsigned int i;
 460
 461         dprint(FD_VERIFY, "starting loop\n");
 462
 463         /*
 464          * sync io first and invalidate cache, to make sure we really
 465          * read from disk.
 466          */
 467         for_each_file(td, f, i) {
 468                 if (!fio_file_open(f))
 469                         continue;
 470                 if (fio_io_sync(td, f))
 471                         break;
 472                 if (file_invalidate_cache(td, f))
 473                         break;
 474         }
 475
 476         check_update_rusage(td);
 477
 478         if (td->error)
 479                 return;
 480
 481         td_set_runstate(td, TD_VERIFYING);
 482
 483         io_u = NULL;
 484         while (!td->terminate) {
 485                 enum fio_ddir ddir;
 486                 int ret2, full;
 487
 488                 update_tv_cache(td);
 489                 check_update_rusage(td);
 490
 491                 if (runtime_exceeded(td, &td->tv_cache)) {
 492                         __update_tv_cache(td);
 493                         if (runtime_exceeded(td, &td->tv_cache)) {
 494                                 fio_mark_td_terminate(td);
 495                                 break;
 496                         }
 497                 }
 498
 499                 if (flow_threshold_exceeded(td))
 500                         continue;
 501
 502                 if (!td->o.experimental_verify) {
 503                         io_u = __get_io_u(td);
 504                         if (!io_u)
 505                                 break;
 506
 507                         if (get_next_verify(td, io_u)) {
 508                                 put_io_u(td, io_u);
 509                                 break;
 510                         }
 511
 512                         if (td_io_prep(td, io_u)) {
 513                                 put_io_u(td, io_u);
 514                                 break;
 515                         }
 516                 } else {
 517                         if (ddir_rw_sum(bytes_done) + td->o.rw_min_bs > verify_bytes)
 518                                 break;
 519
 520                         while ((io_u = get_io_u(td)) != NULL) {
 521                                 if (IS_ERR(io_u)) {
 522                                         io_u = NULL;
 523                                         ret = FIO_Q_BUSY;
 524                                         goto reap;
 525                                 }
 526
 527                                 /*
 528                                  * We are only interested in the places where
 529                                  * we wrote or trimmed IOs. Turn those into
 530                                  * reads for verification purposes.
 531                                  */
 532                                 if (io_u->ddir == DDIR_READ) {
 533                                         /*
 534                                          * Pretend we issued it for rwmix
 535                                          * accounting
 536                                          */
 537                                         td->io_issues[DDIR_READ]++;
 538                                         put_io_u(td, io_u);
 539                                         continue;
 540                                 } else if (io_u->ddir == DDIR_TRIM) {
 541                                         io_u->ddir = DDIR_READ;
 542                                         io_u->flags |= IO_U_F_TRIMMED;
 543                                         break;
 544                                 } else if (io_u->ddir == DDIR_WRITE) {
 545                                         io_u->ddir = DDIR_READ;
 546                                         break;
 547                                 } else {
 548                                         put_io_u(td, io_u);
 549                                         continue;
 550                                 }
 551                         }
 552
 553                         if (!io_u)
 554                                 break;
 555                 }
 556
 557                 if (verify_state_should_stop(td, io_u)) {
 558                         put_io_u(td, io_u);
 559                         break;
 560                 }
 561
 562                 if (td->o.verify_async)
 563                         io_u->end_io = verify_io_u_async;
 564                 else
 565                         io_u->end_io = verify_io_u;
 566
 567                 ddir = io_u->ddir;
 568                 if (!td->o.disable_slat)
 569                         fio_gettime(&io_u->start_time, NULL);
 570
 571                 ret = td_io_queue(td, io_u);
 572                 switch (ret) {
 573                 case FIO_Q_COMPLETED:
 574                         if (io_u->error) {
 575                                 ret = -io_u->error;
 576                                 clear_io_u(td, io_u);
 577                         } else if (io_u->resid) {
 578                                 int bytes = io_u->xfer_buflen - io_u->resid;
 579
 580                                 /*
 581                                  * zero read, fail
 582                                  */
 583                                 if (!bytes) {
 584                                         td_verror(td, EIO, "full resid");
 585                                         put_io_u(td, io_u);
 586                                         break;
 587                                 }
 588
 589                                 io_u->xfer_buflen = io_u->resid;
 590                                 io_u->xfer_buf += bytes;
 591                                 io_u->offset += bytes;
 592
 593                                 if (ddir_rw(io_u->ddir))
 594                                         td->ts.short_io_u[io_u->ddir]++;
 595
 596                                 f = io_u->file;
 597                                 if (io_u->offset == f->real_file_size)
 598                                         goto sync_done;
 599
 600                                 requeue_io_u(td, &io_u);
 601                         } else {
 602 sync_done:
 603                                 ret = io_u_sync_complete(td, io_u, bytes_done);
 604                                 if (ret < 0)
 605                                         break;
 606                         }
 607                         continue;
 608                 case FIO_Q_QUEUED:
 609                         break;
 610                 case FIO_Q_BUSY:
 611                         requeue_io_u(td, &io_u);
 612                         ret2 = td_io_commit(td);
 613                         if (ret2 < 0)
 614                                 ret = ret2;
 615                         break;
 616                 default:
 617                         assert(ret < 0);
 618                         td_verror(td, -ret, "td_io_queue");
 619                         break;
 620                 }
 621
 622                 if (break_on_this_error(td, ddir, &ret))
 623                         break;
 624
 625                 /*
 626                  * if we can queue more, do so. but check if there are
 627                  * completed io_u's first. Note that we can get BUSY even
 628                  * without IO queued, if the system is resource starved.
 629                  */
 630 reap:
 631                 full = queue_full(td) || (ret == FIO_Q_BUSY && td->cur_depth);
 632                 if (full || !td->o.iodepth_batch_complete)
 633                         ret = wait_for_completions(td, NULL, bytes_done);
 634
 635                 if (ret < 0)
 636                         break;
 637         }
 638
 639         check_update_rusage(td);
 640
 641         if (!td->error) {
 642                 min_events = td->cur_depth;
 643
 644                 if (min_events)
 645                         ret = io_u_queued_complete(td, min_events, NULL);
 646         } else
 647                 cleanup_pending_aio(td);
 648
 649         td_set_runstate(td, TD_RUNNING);
 650
 651         dprint(FD_VERIFY, "exiting loop\n");
 652 }
 653
 654 static unsigned int exceeds_number_ios(struct thread_data *td)
 655 {
 656         unsigned long long number_ios;
 657
 658         if (!td->o.number_ios)
 659                 return 0;
 660
 661         number_ios = ddir_rw_sum(td->this_io_blocks);
 662         number_ios += td->io_u_queued + td->io_u_in_flight;
 663
 664         return number_ios >= td->o.number_ios;
 665 }
 666
 667 static int io_issue_bytes_exceeded(struct thread_data *td)
 668 {
 669         unsigned long long bytes, limit;
 670
 671         if (td_rw(td))
 672                 bytes = td->io_issue_bytes[DDIR_READ] + td->io_issue_bytes[DDIR_WRITE];
 673         else if (td_write(td))
 674                 bytes = td->io_issue_bytes[DDIR_WRITE];
 675         else if (td_read(td))
 676                 bytes = td->io_issue_bytes[DDIR_READ];
 677         else
 678                 bytes = td->io_issue_bytes[DDIR_TRIM];
 679
 680         if (td->o.io_limit)
 681                 limit = td->o.io_limit;
 682         else
 683                 limit = td->o.size;
 684
 685         return bytes >= limit || exceeds_number_ios(td);
 686 }
 687
 688 static int io_complete_bytes_exceeded(struct thread_data *td)
 689 {
 690         unsigned long long bytes, limit;
 691
 692         if (td_rw(td))
 693                 bytes = td->this_io_bytes[DDIR_READ] + td->this_io_bytes[DDIR_WRITE];
 694         else if (td_write(td))
 695                 bytes = td->this_io_bytes[DDIR_WRITE];
 696         else if (td_read(td))
 697                 bytes = td->this_io_bytes[DDIR_READ];
 698         else
 699                 bytes = td->this_io_bytes[DDIR_TRIM];
 700
 701         if (td->o.io_limit)
 702                 limit = td->o.io_limit;
 703         else
 704                 limit = td->o.size;
 705
 706         return bytes >= limit || exceeds_number_ios(td);
 707 }
 708
 709 /*
 710  * Main IO worker function. It retrieves io_u's to process and queues
 711  * and reaps them, checking for rate and errors along the way.
 712  *
 713  * Returns number of bytes written and trimmed.
 714  */
 715 static uint64_t do_io(struct thread_data *td)
 716 {
 717         uint64_t bytes_done[DDIR_RWDIR_CNT] = { 0, 0, 0 };
 718         unsigned int i;
 719         int ret = 0;
 720         uint64_t total_bytes, bytes_issued = 0;
 721
 722         if (in_ramp_time(td))
 723                 td_set_runstate(td, TD_RAMP);
 724         else
 725                 td_set_runstate(td, TD_RUNNING);
 726
 727         lat_target_init(td);
 728
 729         /*
 730          * If verify_backlog is enabled, we'll run the verify in this
 731          * handler as well. For that case, we may need up to twice the
 732          * amount of bytes.
 733          */
 734         total_bytes = td->o.size;
 735         if (td->o.verify != VERIFY_NONE &&
 736            (td_write(td) && td->o.verify_backlog))
 737                 total_bytes += td->o.size;
 738
 739         while ((td->o.read_iolog_file && !flist_empty(&td->io_log_list)) ||
 740                 (!flist_empty(&td->trim_list)) || !io_issue_bytes_exceeded(td) ||
 741                 td->o.time_based) {
 742                 struct timeval comp_time;
 743                 struct io_u *io_u;
 744                 int ret2, full;
 745                 enum fio_ddir ddir;
 746
 747                 check_update_rusage(td);
 748
 749                 if (td->terminate || td->done)
 750                         break;
 751
 752                 update_tv_cache(td);
 753
 754                 if (runtime_exceeded(td, &td->tv_cache)) {
 755                         __update_tv_cache(td);
 756                         if (runtime_exceeded(td, &td->tv_cache)) {
 757                                 fio_mark_td_terminate(td);
 758                                 break;
 759                         }
 760                 }
 761
 762                 if (flow_threshold_exceeded(td))
 763                         continue;
 764
 765                 if (bytes_issued >= total_bytes)
 766                         break;
 767
 768                 io_u = get_io_u(td);
 769                 if (IS_ERR_OR_NULL(io_u)) {
 770                         int err = PTR_ERR(io_u);
 771
 772                         io_u = NULL;
 773                         if (err == -EBUSY) {
 774                                 ret = FIO_Q_BUSY;
 775                                 goto reap;
 776                         }
 777                         if (td->o.latency_target)
 778                                 goto reap;
 779                         break;
 780                 }
 781
 782                 ddir = io_u->ddir;
 783
 784                 /*
 785                  * Add verification end_io handler if:
 786                  *      - Asked to verify (!td_rw(td))
 787                  *      - Or the io_u is from our verify list (mixed write/ver)
 788                  */
 789                 if (td->o.verify != VERIFY_NONE && io_u->ddir == DDIR_READ &&
 790                     ((io_u->flags & IO_U_F_VER_LIST) || !td_rw(td))) {
 791
 792                         if (!td->o.verify_pattern_bytes) {
 793                                 io_u->rand_seed = __rand(&td->verify_state);
 794                                 if (sizeof(int) != sizeof(long *))
 795                                         io_u->rand_seed *= __rand(&td->verify_state);
 796                         }
 797
 798                         if (verify_state_should_stop(td, io_u)) {
 799                                 put_io_u(td, io_u);
 800                                 break;
 801                         }
 802
 803                         if (td->o.verify_async)
 804                                 io_u->end_io = verify_io_u_async;
 805                         else
 806                                 io_u->end_io = verify_io_u;
 807                         td_set_runstate(td, TD_VERIFYING);
 808                 } else if (in_ramp_time(td))
 809                         td_set_runstate(td, TD_RAMP);
 810                 else
 811                         td_set_runstate(td, TD_RUNNING);
 812
 813                 /*
 814                  * Always log IO before it's issued, so we know the specific
 815                  * order of it. The logged unit will track when the IO has
 816                  * completed.
 817                  */
 818                 if (td_write(td) && io_u->ddir == DDIR_WRITE &&
 819                     td->o.do_verify &&
 820                     td->o.verify != VERIFY_NONE &&
 821                     !td->o.experimental_verify)
 822                         log_io_piece(td, io_u);
 823
 824                 ret = td_io_queue(td, io_u);
 825                 switch (ret) {
 826                 case FIO_Q_COMPLETED:
 827                         if (io_u->error) {
 828                                 ret = -io_u->error;
 829                                 unlog_io_piece(td, io_u);
 830                                 clear_io_u(td, io_u);
 831                         } else if (io_u->resid) {
 832                                 int bytes = io_u->xfer_buflen - io_u->resid;
 833                                 struct fio_file *f = io_u->file;
 834
 835                                 bytes_issued += bytes;
 836
 837                                 trim_io_piece(td, io_u);
 838
 839                                 /*
 840                                  * zero read, fail
 841                                  */
 842                                 if (!bytes) {
 843                                         unlog_io_piece(td, io_u);
 844                                         td_verror(td, EIO, "full resid");
 845                                         put_io_u(td, io_u);
 846                                         break;
 847                                 }
 848
 849                                 io_u->xfer_buflen = io_u->resid;
 850                                 io_u->xfer_buf += bytes;
 851                                 io_u->offset += bytes;
 852
 853                                 if (ddir_rw(io_u->ddir))
 854                                         td->ts.short_io_u[io_u->ddir]++;
 855
 856                                 if (io_u->offset == f->real_file_size)
 857                                         goto sync_done;
 858
 859                                 requeue_io_u(td, &io_u);
 860                         } else {
 861 sync_done:
 862                                 if (__should_check_rate(td, DDIR_READ) ||
 863                                     __should_check_rate(td, DDIR_WRITE) ||
 864                                     __should_check_rate(td, DDIR_TRIM))
 865                                         fio_gettime(&comp_time, NULL);
 866
 867                                 ret = io_u_sync_complete(td, io_u, bytes_done);
 868                                 if (ret < 0)
 869                                         break;
 870                                 bytes_issued += io_u->xfer_buflen;
 871                         }
 872                         break;
 873                 case FIO_Q_QUEUED:
 874                         /*
 875                          * if the engine doesn't have a commit hook,
 876                          * the io_u is really queued. if it does have such
 877                          * a hook, it has to call io_u_queued() itself.
 878                          */
 879                         if (td->io_ops->commit == NULL)
 880                                 io_u_queued(td, io_u);
 881                         bytes_issued += io_u->xfer_buflen;
 882                         break;
 883                 case FIO_Q_BUSY:
 884                         unlog_io_piece(td, io_u);
 885                         requeue_io_u(td, &io_u);
 886                         ret2 = td_io_commit(td);
 887                         if (ret2 < 0)
 888                                 ret = ret2;
 889                         break;
 890                 default:
 891                         assert(ret < 0);
 892                         put_io_u(td, io_u);
 893                         break;
 894                 }
 895
 896                 if (break_on_this_error(td, ddir, &ret))
 897                         break;
 898
 899                 /*
 900                  * See if we need to complete some commands. Note that we
 901                  * can get BUSY even without IO queued, if the system is
 902                  * resource starved.
 903                  */
 904 reap:
 905                 full = queue_full(td) || (ret == FIO_Q_BUSY && td->cur_depth);
 906                 if (full || !td->o.iodepth_batch_complete)
 907                         ret = wait_for_completions(td, &comp_time, bytes_done);
 908                 if (ret < 0)
 909                         break;
 910                 if (!ddir_rw_sum(bytes_done) && !(td->io_ops->flags & FIO_NOIO))
 911                         continue;
 912
 913                 if (!in_ramp_time(td) && should_check_rate(td, bytes_done)) {
 914                         if (check_min_rate(td, &comp_time, bytes_done)) {
 915                                 if (exitall_on_terminate)
 916                                         fio_terminate_threads(td->groupid);
 917                                 td_verror(td, EIO, "check_min_rate");
 918                                 break;
 919                         }
 920                 }
 921                 if (!in_ramp_time(td) && td->o.latency_target)
 922                         lat_target_check(td);
 923
 924                 if (td->o.thinktime) {
 925                         unsigned long long b;
 926
 927                         b = ddir_rw_sum(td->io_blocks);
 928                         if (!(b % td->o.thinktime_blocks)) {
 929                                 int left;
 930
 931                                 io_u_quiesce(td);
 932
 933                                 if (td->o.thinktime_spin)
 934                                         usec_spin(td->o.thinktime_spin);
 935
 936                                 left = td->o.thinktime - td->o.thinktime_spin;
 937                                 if (left)
 938                                         usec_sleep(td, left);
 939                         }
 940                 }
 941         }
 942
 943         check_update_rusage(td);
 944
 945         if (td->trim_entries)
 946                 log_err("fio: %lu trim entries leaked?\n", td->trim_entries);
 947
 948         if (td->o.fill_device && td->error == ENOSPC) {
 949                 td->error = 0;
 950                 fio_mark_td_terminate(td);
 951         }
 952         if (!td->error) {
 953                 struct fio_file *f;
 954
 955                 i = td->cur_depth;
 956                 if (i) {
 957                         ret = io_u_queued_complete(td, i, bytes_done);
 958                         if (td->o.fill_device && td->error == ENOSPC)
 959                                 td->error = 0;
 960                 }
 961
 962                 if (should_fsync(td) && td->o.end_fsync) {
 963                         td_set_runstate(td, TD_FSYNCING);
 964
 965                         for_each_file(td, f, i) {
 966                                 if (!fio_file_fsync(td, f))
 967                                         continue;
 968
 969                                 log_err("fio: end_fsync failed for file %s\n",
 970                                                                 f->file_name);
 971                         }
 972                 }
 973         } else
 974                 cleanup_pending_aio(td);
 975
 976         /*
 977          * stop job if we failed doing any IO
 978          */
 979         if (!ddir_rw_sum(td->this_io_bytes))
 980                 td->done = 1;
 981
 982         return bytes_done[DDIR_WRITE] + bytes_done[DDIR_TRIM];
 983 }
 984
 985 static void cleanup_io_u(struct thread_data *td)
 986 {
 987         struct io_u *io_u;
 988
 989         while ((io_u = io_u_qpop(&td->io_u_freelist)) != NULL) {
 990
 991                 if (td->io_ops->io_u_free)
 992                         td->io_ops->io_u_free(td, io_u);
 993
 994                 fio_memfree(io_u, sizeof(*io_u));
 995         }
 996
 997         free_io_mem(td);
 998
 999         io_u_rexit(&td->io_u_requeues);
1000         io_u_qexit(&td->io_u_freelist);
1001         io_u_qexit(&td->io_u_all);
1002
1003         if (td->last_write_comp)
1004                 sfree(td->last_write_comp);
1005 }
1006
1007 static int init_io_u(struct thread_data *td)
1008 {
1009         struct io_u *io_u;
1010         unsigned int max_bs, min_write;
1011         int cl_align, i, max_units;
1012         int data_xfer = 1, err;
1013         char *p;
1014
1015         max_units = td->o.iodepth;
1016         max_bs = td_max_bs(td);
1017         min_write = td->o.min_bs[DDIR_WRITE];
1018         td->orig_buffer_size = (unsigned long long) max_bs
1019                                         * (unsigned long long) max_units;
1020
1021         if ((td->io_ops->flags & FIO_NOIO) || !(td_read(td) || td_write(td)))
1022                 data_xfer = 0;
1023
1024         err = 0;
1025         err += io_u_rinit(&td->io_u_requeues, td->o.iodepth);
1026         err += io_u_qinit(&td->io_u_freelist, td->o.iodepth);
1027         err += io_u_qinit(&td->io_u_all, td->o.iodepth);
1028
1029         if (err) {
1030                 log_err("fio: failed setting up IO queues\n");
1031                 return 1;
1032         }
1033
1034         /*
1035          * if we may later need to do address alignment, then add any
1036          * possible adjustment here so that we don't cause a buffer
1037          * overflow later. this adjustment may be too much if we get
1038          * lucky and the allocator gives us an aligned address.
1039          */
1040         if (td->o.odirect || td->o.mem_align || td->o.oatomic ||
1041             (td->io_ops->flags & FIO_RAWIO))
1042                 td->orig_buffer_size += page_mask + td->o.mem_align;
1043
1044         if (td->o.mem_type == MEM_SHMHUGE || td->o.mem_type == MEM_MMAPHUGE) {
1045                 unsigned long bs;
1046
1047                 bs = td->orig_buffer_size + td->o.hugepage_size - 1;
1048                 td->orig_buffer_size = bs & ~(td->o.hugepage_size - 1);
1049         }
1050
1051         if (td->orig_buffer_size != (size_t) td->orig_buffer_size) {
1052                 log_err("fio: IO memory too large. Reduce max_bs or iodepth\n");
1053                 return 1;
1054         }
1055
1056         if (data_xfer && allocate_io_mem(td))
1057                 return 1;
1058
1059         if (td->o.odirect || td->o.mem_align || td->o.oatomic ||
1060             (td->io_ops->flags & FIO_RAWIO))
1061                 p = PAGE_ALIGN(td->orig_buffer) + td->o.mem_align;
1062         else
1063                 p = td->orig_buffer;
1064
1065         cl_align = os_cache_line_size();
1066
1067         for (i = 0; i < max_units; i++) {
1068                 void *ptr;
1069
1070                 if (td->terminate)
1071                         return 1;
1072
1073                 ptr = fio_memalign(cl_align, sizeof(*io_u));
1074                 if (!ptr) {
1075                         log_err("fio: unable to allocate aligned memory\n");
1076                         break;
1077                 }
1078
1079                 io_u = ptr;
1080                 memset(io_u, 0, sizeof(*io_u));
1081                 INIT_FLIST_HEAD(&io_u->verify_list);
1082                 dprint(FD_MEM, "io_u alloc %p, index %u\n", io_u, i);
1083
1084                 if (data_xfer) {
1085                         io_u->buf = p;
1086                         dprint(FD_MEM, "io_u %p, mem %p\n", io_u, io_u->buf);
1087
1088                         if (td_write(td))
1089                                 io_u_fill_buffer(td, io_u, min_write, max_bs);
1090                         if (td_write(td) && td->o.verify_pattern_bytes) {
1091                                 /*
1092                                  * Fill the buffer with the pattern if we are
1093                                  * going to be doing writes.
1094                                  */
1095                                 fill_verify_pattern(td, io_u->buf, max_bs, io_u, 0, 0);
1096                         }
1097                 }
1098
1099                 io_u->index = i;
1100                 io_u->flags = IO_U_F_FREE;
1101                 io_u_qpush(&td->io_u_freelist, io_u);
1102
1103                 /*
1104                  * io_u never leaves this stack, used for iteration of all
1105                  * io_u buffers.
1106                  */
1107                 io_u_qpush(&td->io_u_all, io_u);
1108
1109                 if (td->io_ops->io_u_init) {
1110                         int ret = td->io_ops->io_u_init(td, io_u);
1111
1112                         if (ret) {
1113                                 log_err("fio: failed to init engine data: %d\n", ret);
1114                                 return 1;
1115                         }
1116                 }
1117
1118                 p += max_bs;
1119         }
1120
1121         if (td->o.verify != VERIFY_NONE) {
1122                 td->last_write_comp = scalloc(max_units, sizeof(uint64_t));
1123                 if (!td->last_write_comp) {
1124                         log_err("fio: failed to alloc write comp data\n");
1125                         return 1;
1126                 }
1127         }
1128
1129         return 0;
1130 }
1131
1132 static int switch_ioscheduler(struct thread_data *td)
1133 {
1134         char tmp[256], tmp2[128];
1135         FILE *f;
1136         int ret;
1137
1138         if (td->io_ops->flags & FIO_DISKLESSIO)
1139                 return 0;
1140
1141         sprintf(tmp, "%s/queue/scheduler", td->sysfs_root);
1142
1143         f = fopen(tmp, "r+");
1144         if (!f) {
1145                 if (errno == ENOENT) {
1146                         log_err("fio: os or kernel doesn't support IO scheduler"
1147                                 " switching\n");
1148                         return 0;
1149                 }
1150                 td_verror(td, errno, "fopen iosched");
1151                 return 1;
1152         }
1153
1154         /*
1155          * Set io scheduler.
1156          */
1157         ret = fwrite(td->o.ioscheduler, strlen(td->o.ioscheduler), 1, f);
1158         if (ferror(f) || ret != 1) {
1159                 td_verror(td, errno, "fwrite");
1160                 fclose(f);
1161                 return 1;
1162         }
1163
1164         rewind(f);
1165
1166         /*
1167          * Read back and check that the selected scheduler is now the default.
1168          */
1169         ret = fread(tmp, sizeof(tmp), 1, f);
1170         if (ferror(f) || ret < 0) {
1171                 td_verror(td, errno, "fread");
1172                 fclose(f);
1173                 return 1;
1174         }
1175         tmp[sizeof(tmp) - 1] = '\0';
1176
1177
1178         sprintf(tmp2, "[%s]", td->o.ioscheduler);
1179         if (!strstr(tmp, tmp2)) {
1180                 log_err("fio: io scheduler %s not found\n", td->o.ioscheduler);
1181                 td_verror(td, EINVAL, "iosched_switch");
1182                 fclose(f);
1183                 return 1;
1184         }
1185
1186         fclose(f);
1187         return 0;
1188 }
1189
1190 static int keep_running(struct thread_data *td)
1191 {
1192         unsigned long long limit;
1193
1194         if (td->done)
1195                 return 0;
1196         if (td->o.time_based)
1197                 return 1;
1198         if (td->o.loops) {
1199                 td->o.loops--;
1200                 return 1;
1201         }
1202         if (exceeds_number_ios(td))
1203                 return 0;
1204
1205         if (td->o.io_limit)
1206                 limit = td->o.io_limit;
1207         else
1208                 limit = td->o.size;
1209
1210         if (limit != -1ULL && ddir_rw_sum(td->io_bytes) < limit) {
1211                 uint64_t diff;
1212
1213                 /*
1214                  * If the difference is less than the minimum IO size, we
1215                  * are done.
1216                  */
1217                 diff = limit - ddir_rw_sum(td->io_bytes);
1218                 if (diff < td_max_bs(td))
1219                         return 0;
1220
1221                 if (fio_files_done(td))
1222                         return 0;
1223
1224                 return 1;
1225         }
1226
1227         return 0;
1228 }
1229
1230 static int exec_string(struct thread_options *o, const char *string, const char *mode)
1231 {
1232         int ret, newlen = strlen(string) + strlen(o->name) + strlen(mode) + 9 + 1;
1233         char *str;
1234
1235         str = malloc(newlen);
1236         sprintf(str, "%s &> %s.%s.txt", string, o->name, mode);
1237
1238         log_info("%s : Saving output of %s in %s.%s.txt\n",o->name, mode, o->name, mode);
1239         ret = system(str);
1240         if (ret == -1)
1241                 log_err("fio: exec of cmd <%s> failed\n", str);
1242
1243         free(str);
1244         return ret;
1245 }
1246
1247 /*
1248  * Dry run to compute correct state of numberio for verification.
1249  */
1250 static uint64_t do_dry_run(struct thread_data *td)
1251 {
1252         uint64_t bytes_done[DDIR_RWDIR_CNT] = { 0, 0, 0 };
1253
1254         td_set_runstate(td, TD_RUNNING);
1255
1256         while ((td->o.read_iolog_file && !flist_empty(&td->io_log_list)) ||
1257                 (!flist_empty(&td->trim_list)) || !io_complete_bytes_exceeded(td)) {
1258                 struct io_u *io_u;
1259                 int ret;
1260
1261                 if (td->terminate || td->done)
1262                         break;
1263
1264                 io_u = get_io_u(td);
1265                 if (!io_u)
1266                         break;
1267
1268                 io_u->flags |= IO_U_F_FLIGHT;
1269                 io_u->error = 0;
1270                 io_u->resid = 0;
1271                 if (ddir_rw(acct_ddir(io_u)))
1272                         td->io_issues[acct_ddir(io_u)]++;
1273                 if (ddir_rw(io_u->ddir)) {
1274                         io_u_mark_depth(td, 1);
1275                         td->ts.total_io_u[io_u->ddir]++;
1276                 }
1277
1278                 if (td_write(td) && io_u->ddir == DDIR_WRITE &&
1279                     td->o.do_verify &&
1280                     td->o.verify != VERIFY_NONE &&
1281                     !td->o.experimental_verify)
1282                         log_io_piece(td, io_u);
1283
1284                 ret = io_u_sync_complete(td, io_u, bytes_done);
1285                 (void) ret;
1286         }
1287
1288         return bytes_done[DDIR_WRITE] + bytes_done[DDIR_TRIM];
1289 }
1290
1291 /*
1292  * Entry point for the thread based jobs. The process based jobs end up
1293  * here as well, after a little setup.
1294  */
1295 static void *thread_main(void *data)
1296 {
1297         unsigned long long elapsed;
1298         struct thread_data *td = data;
1299         struct thread_options *o = &td->o;
1300         pthread_condattr_t attr;
1301         int clear_state;
1302         int ret;
1303
1304         if (!o->use_thread) {
1305                 setsid();
1306                 td->pid = getpid();
1307         } else
1308                 td->pid = gettid();
1309
1310         fio_local_clock_init(o->use_thread);
1311
1312         dprint(FD_PROCESS, "jobs pid=%d started\n", (int) td->pid);
1313
1314         if (is_backend)
1315                 fio_server_send_start(td);
1316
1317         INIT_FLIST_HEAD(&td->io_log_list);
1318         INIT_FLIST_HEAD(&td->io_hist_list);
1319         INIT_FLIST_HEAD(&td->verify_list);
1320         INIT_FLIST_HEAD(&td->trim_list);
1321         INIT_FLIST_HEAD(&td->next_rand_list);
1322         pthread_mutex_init(&td->io_u_lock, NULL);
1323         td->io_hist_tree = RB_ROOT;
1324
1325         pthread_condattr_init(&attr);
1326         pthread_cond_init(&td->verify_cond, &attr);
1327         pthread_cond_init(&td->free_cond, &attr);
1328
1329         td_set_runstate(td, TD_INITIALIZED);
1330         dprint(FD_MUTEX, "up startup_mutex\n");
1331         fio_mutex_up(startup_mutex);
1332         dprint(FD_MUTEX, "wait on td->mutex\n");
1333         fio_mutex_down(td->mutex);
1334         dprint(FD_MUTEX, "done waiting on td->mutex\n");
1335
1336         /*
1337          * A new gid requires privilege, so we need to do this before setting
1338          * the uid.
1339          */
1340         if (o->gid != -1U && setgid(o->gid)) {
1341                 td_verror(td, errno, "setgid");
1342                 goto err;
1343         }
1344         if (o->uid != -1U && setuid(o->uid)) {
1345                 td_verror(td, errno, "setuid");
1346                 goto err;
1347         }
1348
1349         /*
1350          * If we have a gettimeofday() thread, make sure we exclude that
1351          * thread from this job
1352          */
1353         if (o->gtod_cpu)
1354                 fio_cpu_clear(&o->cpumask, o->gtod_cpu);
1355
1356         /*
1357          * Set affinity first, in case it has an impact on the memory
1358          * allocations.
1359          */
1360         if (fio_option_is_set(o, cpumask)) {
1361                 if (o->cpus_allowed_policy == FIO_CPUS_SPLIT) {
1362                         ret = fio_cpus_split(&o->cpumask, td->thread_number - 1);
1363                         if (!ret) {
1364                                 log_err("fio: no CPUs set\n");
1365                                 log_err("fio: Try increasing number of available CPUs\n");
1366                                 td_verror(td, EINVAL, "cpus_split");
1367                                 goto err;
1368                         }
1369                 }
1370                 ret = fio_setaffinity(td->pid, o->cpumask);
1371                 if (ret == -1) {
1372                         td_verror(td, errno, "cpu_set_affinity");
1373                         goto err;
1374                 }
1375         }
1376
1377 #ifdef CONFIG_LIBNUMA
1378         /* numa node setup */
1379         if (fio_option_is_set(o, numa_cpunodes) ||
1380             fio_option_is_set(o, numa_memnodes)) {
1381                 struct bitmask *mask;
1382
1383                 if (numa_available() < 0) {
1384                         td_verror(td, errno, "Does not support NUMA API\n");
1385                         goto err;
1386                 }
1387
1388                 if (fio_option_is_set(o, numa_cpunodes)) {
1389                         mask = numa_parse_nodestring(o->numa_cpunodes);
1390                         ret = numa_run_on_node_mask(mask);
1391                         numa_free_nodemask(mask);
1392                         if (ret == -1) {
1393                                 td_verror(td, errno, \
1394                                         "numa_run_on_node_mask failed\n");
1395                                 goto err;
1396                         }
1397                 }
1398
1399                 if (fio_option_is_set(o, numa_memnodes)) {
1400                         mask = NULL;
1401                         if (o->numa_memnodes)
1402                                 mask = numa_parse_nodestring(o->numa_memnodes);
1403
1404                         switch (o->numa_mem_mode) {
1405                         case MPOL_INTERLEAVE:
1406                                 numa_set_interleave_mask(mask);
1407                                 break;
1408                         case MPOL_BIND:
1409                                 numa_set_membind(mask);
1410                                 break;
1411                         case MPOL_LOCAL:
1412                                 numa_set_localalloc();
1413                                 break;
1414                         case MPOL_PREFERRED:
1415                                 numa_set_preferred(o->numa_mem_prefer_node);
1416                                 break;
1417                         case MPOL_DEFAULT:
1418                         default:
1419                                 break;
1420                         }
1421
1422                         if (mask)
1423                                 numa_free_nodemask(mask);
1424
1425                 }
1426         }
1427 #endif
1428
1429         if (fio_pin_memory(td))
1430                 goto err;
1431
1432         /*
1433          * May alter parameters that init_io_u() will use, so we need to
1434          * do this first.
1435          */
1436         if (init_iolog(td))
1437                 goto err;
1438
1439         if (init_io_u(td))
1440                 goto err;
1441
1442         if (o->verify_async && verify_async_init(td))
1443                 goto err;
1444
1445         if (fio_option_is_set(o, ioprio) ||
1446             fio_option_is_set(o, ioprio_class)) {
1447                 ret = ioprio_set(IOPRIO_WHO_PROCESS, 0, o->ioprio_class, o->ioprio);
1448                 if (ret == -1) {
1449                         td_verror(td, errno, "ioprio_set");
1450                         goto err;
1451                 }
1452         }
1453
1454         if (o->cgroup && cgroup_setup(td, cgroup_list, &cgroup_mnt))
1455                 goto err;
1456
1457         errno = 0;
1458         if (nice(o->nice) == -1 && errno != 0) {
1459                 td_verror(td, errno, "nice");
1460                 goto err;
1461         }
1462
1463         if (o->ioscheduler && switch_ioscheduler(td))
1464                 goto err;
1465
1466         if (!o->create_serialize && setup_files(td))
1467                 goto err;
1468
1469         if (td_io_init(td))
1470                 goto err;
1471
1472         if (init_random_map(td))
1473                 goto err;
1474
1475         if (o->exec_prerun && exec_string(o, o->exec_prerun, (const char *)"prerun"))
1476                 goto err;
1477
1478         if (o->pre_read) {
1479                 if (pre_read_files(td) < 0)
1480                         goto err;
1481         }
1482
1483         if (td->flags & TD_F_COMPRESS_LOG)
1484                 tp_init(&td->tp_data);
1485
1486         fio_verify_init(td);
1487
1488         fio_gettime(&td->epoch, NULL);
1489         fio_getrusage(&td->ru_start);
1490         clear_state = 0;
1491         while (keep_running(td)) {
1492                 uint64_t verify_bytes;
1493
1494                 fio_gettime(&td->start, NULL);
1495                 memcpy(&td->bw_sample_time, &td->start, sizeof(td->start));
1496                 memcpy(&td->iops_sample_time, &td->start, sizeof(td->start));
1497                 memcpy(&td->tv_cache, &td->start, sizeof(td->start));
1498
1499                 if (o->ratemin[DDIR_READ] || o->ratemin[DDIR_WRITE] ||
1500                                 o->ratemin[DDIR_TRIM]) {
1501                         memcpy(&td->lastrate[DDIR_READ], &td->bw_sample_time,
1502                                                 sizeof(td->bw_sample_time));
1503                         memcpy(&td->lastrate[DDIR_WRITE], &td->bw_sample_time,
1504                                                 sizeof(td->bw_sample_time));
1505                         memcpy(&td->lastrate[DDIR_TRIM], &td->bw_sample_time,
1506                                                 sizeof(td->bw_sample_time));
1507                 }
1508
1509                 if (clear_state)
1510                         clear_io_state(td);
1511
1512                 prune_io_piece_log(td);
1513
1514                 if (td->o.verify_only && (td_write(td) || td_rw(td)))
1515                         verify_bytes = do_dry_run(td);
1516                 else
1517                         verify_bytes = do_io(td);
1518
1519                 clear_state = 1;
1520
1521                 fio_mutex_down(stat_mutex);
1522                 if (td_read(td) && td->io_bytes[DDIR_READ]) {
1523                         elapsed = mtime_since_now(&td->start);
1524                         td->ts.runtime[DDIR_READ] += elapsed;
1525                 }
1526                 if (td_write(td) && td->io_bytes[DDIR_WRITE]) {
1527                         elapsed = mtime_since_now(&td->start);
1528                         td->ts.runtime[DDIR_WRITE] += elapsed;
1529                 }
1530                 if (td_trim(td) && td->io_bytes[DDIR_TRIM]) {
1531                         elapsed = mtime_since_now(&td->start);
1532                         td->ts.runtime[DDIR_TRIM] += elapsed;
1533                 }
1534                 fio_gettime(&td->start, NULL);
1535                 fio_mutex_up(stat_mutex);
1536
1537                 if (td->error || td->terminate)
1538                         break;
1539
1540                 if (!o->do_verify ||
1541                     o->verify == VERIFY_NONE ||
1542                     (td->io_ops->flags & FIO_UNIDIR))
1543                         continue;
1544
1545                 clear_io_state(td);
1546
1547                 fio_gettime(&td->start, NULL);
1548
1549                 do_verify(td, verify_bytes);
1550
1551                 fio_mutex_down(stat_mutex);
1552                 td->ts.runtime[DDIR_READ] += mtime_since_now(&td->start);
1553                 fio_gettime(&td->start, NULL);
1554                 fio_mutex_up(stat_mutex);
1555
1556                 if (td->error || td->terminate)
1557                         break;
1558         }
1559
1560         update_rusage_stat(td);
1561         td->ts.total_run_time = mtime_since_now(&td->epoch);
1562         td->ts.io_bytes[DDIR_READ] = td->io_bytes[DDIR_READ];
1563         td->ts.io_bytes[DDIR_WRITE] = td->io_bytes[DDIR_WRITE];
1564         td->ts.io_bytes[DDIR_TRIM] = td->io_bytes[DDIR_TRIM];
1565
1566         if (td->o.verify_state_save && !(td->flags & TD_F_VSTATE_SAVED) &&
1567             (td->o.verify != VERIFY_NONE && td_write(td))) {
1568                 struct all_io_list *state;
1569                 size_t sz;
1570
1571                 state = get_all_io_list(td->thread_number, &sz);
1572                 if (state) {
1573                         __verify_save_state(state, "local");
1574                         free(state);
1575                 }
1576         }
1577
1578         fio_unpin_memory(td);
1579
1580         fio_writeout_logs(td);
1581
1582         if (td->flags & TD_F_COMPRESS_LOG)
1583                 tp_exit(&td->tp_data);
1584
1585         if (o->exec_postrun)
1586                 exec_string(o, o->exec_postrun, (const char *)"postrun");
1587
1588         if (exitall_on_terminate)
1589                 fio_terminate_threads(td->groupid);
1590
1591 err:
1592         if (td->error)
1593                 log_info("fio: pid=%d, err=%d/%s\n", (int) td->pid, td->error,
1594                                                         td->verror);
1595
1596         if (o->verify_async)
1597                 verify_async_exit(td);
1598
1599         close_and_free_files(td);
1600         cleanup_io_u(td);
1601         close_ioengine(td);
1602         cgroup_shutdown(td, &cgroup_mnt);
1603         verify_free_state(td);
1604
1605         if (fio_option_is_set(o, cpumask)) {
1606                 ret = fio_cpuset_exit(&o->cpumask);
1607                 if (ret)
1608                         td_verror(td, ret, "fio_cpuset_exit");
1609         }
1610
1611         /*
1612          * do this very late, it will log file closing as well
1613          */
1614         if (o->write_iolog_file)
1615                 write_iolog_close(td);
1616
1617         fio_mutex_remove(td->mutex);
1618         td->mutex = NULL;
1619
1620         td_set_runstate(td, TD_EXITED);
1621
1622         /*
1623          * Do this last after setting our runstate to exited, so we
1624          * know that the stat thread is signaled.
1625          */
1626         check_update_rusage(td);
1627
1628         return (void *) (uintptr_t) td->error;
1629 }
1630
1631
1632 /*
1633  * We cannot pass the td data into a forked process, so attach the td and
1634  * pass it to the thread worker.
1635  */
1636 static int fork_main(int shmid, int offset)
1637 {
1638         struct thread_data *td;
1639         void *data, *ret;
1640
1641 #if !defined(__hpux) && !defined(CONFIG_NO_SHM)
1642         data = shmat(shmid, NULL, 0);
1643         if (data == (void *) -1) {
1644                 int __err = errno;
1645
1646                 perror("shmat");
1647                 return __err;
1648         }
1649 #else
1650         /*
1651          * HP-UX inherits shm mappings?
1652          */
1653         data = threads;
1654 #endif
1655
1656         td = data + offset * sizeof(struct thread_data);
1657         ret = thread_main(td);
1658         shmdt(data);
1659         return (int) (uintptr_t) ret;
1660 }
1661
1662 static void dump_td_info(struct thread_data *td)
1663 {
1664         log_err("fio: job '%s' hasn't exited in %lu seconds, it appears to "
1665                 "be stuck. Doing forceful exit of this job.\n", td->o.name,
1666                         (unsigned long) time_since_now(&td->terminate_time));
1667 }
1668
1669 /*
1670  * Run over the job map and reap the threads that have exited, if any.
1671  */
1672 static void reap_threads(unsigned int *nr_running, unsigned int *t_rate,
1673                          unsigned int *m_rate)
1674 {
1675         struct thread_data *td;
1676         unsigned int cputhreads, realthreads, pending;
1677         int i, status, ret;
1678
1679         /*
1680          * reap exited threads (TD_EXITED -> TD_REAPED)
1681          */
1682         realthreads = pending = cputhreads = 0;
1683         for_each_td(td, i) {
1684                 int flags = 0;
1685
1686                 /*
1687                  * ->io_ops is NULL for a thread that has closed its
1688                  * io engine
1689                  */
1690                 if (td->io_ops && !strcmp(td->io_ops->name, "cpuio"))
1691                         cputhreads++;
1692                 else
1693                         realthreads++;
1694
1695                 if (!td->pid) {
1696                         pending++;
1697                         continue;
1698                 }
1699                 if (td->runstate == TD_REAPED)
1700                         continue;
1701                 if (td->o.use_thread) {
1702                         if (td->runstate == TD_EXITED) {
1703                                 td_set_runstate(td, TD_REAPED);
1704                                 goto reaped;
1705                         }
1706                         continue;
1707                 }
1708
1709                 flags = WNOHANG;
1710                 if (td->runstate == TD_EXITED)
1711                         flags = 0;
1712
1713                 /*
1714                  * check if someone quit or got killed in an unusual way
1715                  */
1716                 ret = waitpid(td->pid, &status, flags);
1717                 if (ret < 0) {
1718                         if (errno == ECHILD) {
1719                                 log_err("fio: pid=%d disappeared %d\n",
1720                                                 (int) td->pid, td->runstate);
1721                                 td->sig = ECHILD;
1722                                 td_set_runstate(td, TD_REAPED);
1723                                 goto reaped;
1724                         }
1725                         perror("waitpid");
1726                 } else if (ret == td->pid) {
1727                         if (WIFSIGNALED(status)) {
1728                                 int sig = WTERMSIG(status);
1729
1730                                 if (sig != SIGTERM && sig != SIGUSR2)
1731                                         log_err("fio: pid=%d, got signal=%d\n",
1732                                                         (int) td->pid, sig);
1733                                 td->sig = sig;
1734                                 td_set_runstate(td, TD_REAPED);
1735                                 goto reaped;
1736                         }
1737                         if (WIFEXITED(status)) {
1738                                 if (WEXITSTATUS(status) && !td->error)
1739                                         td->error = WEXITSTATUS(status);
1740
1741                                 td_set_runstate(td, TD_REAPED);
1742                                 goto reaped;
1743                         }
1744                 }
1745
1746                 /*
1747                  * If the job is stuck, do a forceful timeout of it and
1748                  * move on.
1749                  */
1750                 if (td->terminate &&
1751                     time_since_now(&td->terminate_time) >= FIO_REAP_TIMEOUT) {
1752                         dump_td_info(td);
1753                         td_set_runstate(td, TD_REAPED);
1754                         goto reaped;
1755                 }
1756
1757                 /*
1758                  * thread is not dead, continue
1759                  */
1760                 pending++;
1761                 continue;
1762 reaped:
1763                 (*nr_running)--;
1764                 (*m_rate) -= ddir_rw_sum(td->o.ratemin);
1765                 (*t_rate) -= ddir_rw_sum(td->o.rate);
1766                 if (!td->pid)
1767                         pending--;
1768
1769                 if (td->error)
1770                         exit_value++;
1771
1772                 done_secs += mtime_since_now(&td->epoch) / 1000;
1773                 profile_td_exit(td);
1774         }
1775
1776         if (*nr_running == cputhreads && !pending && realthreads)
1777                 fio_terminate_threads(TERMINATE_ALL);
1778 }
1779
1780 static int __check_trigger_file(void)
1781 {
1782         struct stat sb;
1783
1784         if (!trigger_file)
1785                 return 0;
1786
1787         if (stat(trigger_file, &sb))
1788                 return 0;
1789
1790         if (unlink(trigger_file) < 0)
1791                 log_err("fio: failed to unlink %s: %s\n", trigger_file,
1792                                                         strerror(errno));
1793
1794         return 1;
1795 }
1796
1797 static int trigger_timedout(void)
1798 {
1799         if (trigger_timeout)
1800                 return time_since_genesis() >= trigger_timeout;
1801
1802         return 0;
1803 }
1804
1805 void exec_trigger(const char *cmd)
1806 {
1807         int ret;
1808
1809         if (!cmd)
1810                 return;
1811
1812         ret = system(cmd);
1813         if (ret == -1)
1814                 log_err("fio: failed executing %s trigger\n", cmd);
1815 }
1816
1817 void check_trigger_file(void)
1818 {
1819         if (__check_trigger_file() || trigger_timedout()) {
1820                 if (nr_clients)
1821                         fio_clients_send_trigger(trigger_remote_cmd);
1822                 else {
1823                         verify_save_state();
1824                         fio_terminate_threads(TERMINATE_ALL);
1825                         exec_trigger(trigger_cmd);
1826                 }
1827         }
1828 }
1829
1830 static int fio_verify_load_state(struct thread_data *td)
1831 {
1832         int ret;
1833
1834         if (!td->o.verify_state)
1835                 return 0;
1836
1837         if (is_backend) {
1838                 void *data;
1839
1840                 ret = fio_server_get_verify_state(td->o.name,
1841                                         td->thread_number - 1, &data);
1842                 if (!ret)
1843                         verify_convert_assign_state(td, data);
1844         } else
1845                 ret = verify_load_state(td, "local");
1846
1847         return ret;
1848 }
1849
1850 static void do_usleep(unsigned int usecs)
1851 {
1852         check_for_running_stats();
1853         check_trigger_file();
1854         usleep(usecs);
1855 }
1856
1857 /*
1858  * Main function for kicking off and reaping jobs, as needed.
1859  */
1860 static void run_threads(void)
1861 {
1862         struct thread_data *td;
1863         unsigned int i, todo, nr_running, m_rate, t_rate, nr_started;
1864         uint64_t spent;
1865
1866         if (fio_gtod_offload && fio_start_gtod_thread())
1867                 return;
1868
1869         fio_idle_prof_init();
1870
1871         set_sig_handlers();
1872
1873         nr_thread = nr_process = 0;
1874         for_each_td(td, i) {
1875                 if (td->o.use_thread)
1876                         nr_thread++;
1877                 else
1878                         nr_process++;
1879         }
1880
1881         if (output_format == FIO_OUTPUT_NORMAL) {
1882                 log_info("Starting ");
1883                 if (nr_thread)
1884                         log_info("%d thread%s", nr_thread,
1885                                                 nr_thread > 1 ? "s" : "");
1886                 if (nr_process) {
1887                         if (nr_thread)
1888                                 log_info(" and ");
1889                         log_info("%d process%s", nr_process,
1890                                                 nr_process > 1 ? "es" : "");
1891                 }
1892                 log_info("\n");
1893                 log_info_flush();
1894         }
1895
1896         todo = thread_number;
1897         nr_running = 0;
1898         nr_started = 0;
1899         m_rate = t_rate = 0;
1900
1901         for_each_td(td, i) {
1902                 print_status_init(td->thread_number - 1);
1903
1904                 if (!td->o.create_serialize)
1905                         continue;
1906
1907                 if (fio_verify_load_state(td))
1908                         goto reap;
1909
1910                 /*
1911                  * do file setup here so it happens sequentially,
1912                  * we don't want X number of threads getting their
1913                  * client data interspersed on disk
1914                  */
1915                 if (setup_files(td)) {
1916 reap:
1917                         exit_value++;
1918                         if (td->error)
1919                                 log_err("fio: pid=%d, err=%d/%s\n",
1920                                         (int) td->pid, td->error, td->verror);
1921                         td_set_runstate(td, TD_REAPED);
1922                         todo--;
1923                 } else {
1924                         struct fio_file *f;
1925                         unsigned int j;
1926
1927                         /*
1928                          * for sharing to work, each job must always open
1929                          * its own files. so close them, if we opened them
1930                          * for creation
1931                          */
1932                         for_each_file(td, f, j) {
1933                                 if (fio_file_open(f))
1934                                         td_io_close_file(td, f);
1935                         }
1936                 }
1937         }
1938
1939         /* start idle threads before io threads start to run */
1940         fio_idle_prof_start();
1941
1942         set_genesis_time();
1943
1944         while (todo) {
1945                 struct thread_data *map[REAL_MAX_JOBS];
1946                 struct timeval this_start;
1947                 int this_jobs = 0, left;
1948
1949                 /*
1950                  * create threads (TD_NOT_CREATED -> TD_CREATED)
1951                  */
1952                 for_each_td(td, i) {
1953                         if (td->runstate != TD_NOT_CREATED)
1954                                 continue;
1955
1956                         /*
1957                          * never got a chance to start, killed by other
1958                          * thread for some reason
1959                          */
1960                         if (td->terminate) {
1961                                 todo--;
1962                                 continue;
1963                         }
1964
1965                         if (td->o.start_delay) {
1966                                 spent = utime_since_genesis();
1967
1968                                 if (td->o.start_delay > spent)
1969                                         continue;
1970                         }
1971
1972                         if (td->o.stonewall && (nr_started || nr_running)) {
1973                                 dprint(FD_PROCESS, "%s: stonewall wait\n",
1974                                                         td->o.name);
1975                                 break;
1976                         }
1977
1978                         init_disk_util(td);
1979
1980                         td->rusage_sem = fio_mutex_init(FIO_MUTEX_LOCKED);
1981                         td->update_rusage = 0;
1982
1983                         /*
1984                          * Set state to created. Thread will transition
1985                          * to TD_INITIALIZED when it's done setting up.
1986                          */
1987                         td_set_runstate(td, TD_CREATED);
1988                         map[this_jobs++] = td;
1989                         nr_started++;
1990
1991                         if (td->o.use_thread) {
1992                                 int ret;
1993
1994                                 dprint(FD_PROCESS, "will pthread_create\n");
1995                                 ret = pthread_create(&td->thread, NULL,
1996                                                         thread_main, td);
1997                                 if (ret) {
1998                                         log_err("pthread_create: %s\n",
1999                                                         strerror(ret));
2000                                         nr_started--;
2001                                         break;
2002                                 }
2003                                 ret = pthread_detach(td->thread);
2004                                 if (ret)
2005                                         log_err("pthread_detach: %s",
2006                                                         strerror(ret));
2007                         } else {
2008                                 pid_t pid;
2009                                 dprint(FD_PROCESS, "will fork\n");
2010                                 pid = fork();
2011                                 if (!pid) {
2012                                         int ret = fork_main(shm_id, i);
2013
2014                                         _exit(ret);
2015                                 } else if (i == fio_debug_jobno)
2016                                         *fio_debug_jobp = pid;
2017                         }
2018                         dprint(FD_MUTEX, "wait on startup_mutex\n");
2019                         if (fio_mutex_down_timeout(startup_mutex, 10)) {
2020                                 log_err("fio: job startup hung? exiting.\n");
2021                                 fio_terminate_threads(TERMINATE_ALL);
2022                                 fio_abort = 1;
2023                                 nr_started--;
2024                                 break;
2025                         }
2026                         dprint(FD_MUTEX, "done waiting on startup_mutex\n");
2027                 }
2028
2029                 /*
2030                  * Wait for the started threads to transition to
2031                  * TD_INITIALIZED.
2032                  */
2033                 fio_gettime(&this_start, NULL);
2034                 left = this_jobs;
2035                 while (left && !fio_abort) {
2036                         if (mtime_since_now(&this_start) > JOB_START_TIMEOUT)
2037                                 break;
2038
2039                         do_usleep(100000);
2040
2041                         for (i = 0; i < this_jobs; i++) {
2042                                 td = map[i];
2043                                 if (!td)
2044                                         continue;
2045                                 if (td->runstate == TD_INITIALIZED) {
2046                                         map[i] = NULL;
2047                                         left--;
2048                                 } else if (td->runstate >= TD_EXITED) {
2049                                         map[i] = NULL;
2050                                         left--;
2051                                         todo--;
2052                                         nr_running++; /* work-around... */
2053                                 }
2054                         }
2055                 }
2056
2057                 if (left) {
2058                         log_err("fio: %d job%s failed to start\n", left,
2059                                         left > 1 ? "s" : "");
2060                         for (i = 0; i < this_jobs; i++) {
2061                                 td = map[i];
2062                                 if (!td)
2063                                         continue;
2064                                 kill(td->pid, SIGTERM);
2065                         }
2066                         break;
2067                 }
2068
2069                 /*
2070                  * start created threads (TD_INITIALIZED -> TD_RUNNING).
2071                  */
2072                 for_each_td(td, i) {
2073                         if (td->runstate != TD_INITIALIZED)
2074                                 continue;
2075
2076                         if (in_ramp_time(td))
2077                                 td_set_runstate(td, TD_RAMP);
2078                         else
2079                                 td_set_runstate(td, TD_RUNNING);
2080                         nr_running++;
2081                         nr_started--;
2082                         m_rate += ddir_rw_sum(td->o.ratemin);
2083                         t_rate += ddir_rw_sum(td->o.rate);
2084                         todo--;
2085                         fio_mutex_up(td->mutex);
2086                 }
2087
2088                 reap_threads(&nr_running, &t_rate, &m_rate);
2089
2090                 if (todo)
2091                         do_usleep(100000);
2092         }
2093
2094         while (nr_running) {
2095                 reap_threads(&nr_running, &t_rate, &m_rate);
2096                 do_usleep(10000);
2097         }
2098
2099         fio_idle_prof_stop();
2100
2101         update_io_ticks();
2102 }
2103
2104 static void wait_for_helper_thread_exit(void)
2105 {
2106         void *ret;
2107
2108         helper_exit = 1;
2109         pthread_cond_signal(&helper_cond);
2110         pthread_join(helper_thread, &ret);
2111 }
2112
2113 static void free_disk_util(void)
2114 {
2115         disk_util_prune_entries();
2116
2117         pthread_cond_destroy(&helper_cond);
2118 }
2119
2120 static void *helper_thread_main(void *data)
2121 {
2122         int ret = 0;
2123
2124         fio_mutex_up(startup_mutex);
2125
2126         while (!ret) {
2127                 uint64_t sec = DISK_UTIL_MSEC / 1000;
2128                 uint64_t nsec = (DISK_UTIL_MSEC % 1000) * 1000000;
2129                 struct timespec ts;
2130                 struct timeval tv;
2131
2132                 gettimeofday(&tv, NULL);
2133                 ts.tv_sec = tv.tv_sec + sec;
2134                 ts.tv_nsec = (tv.tv_usec * 1000) + nsec;
2135
2136                 if (ts.tv_nsec >= 1000000000ULL) {
2137                         ts.tv_nsec -= 1000000000ULL;
2138                         ts.tv_sec++;
2139                 }
2140
2141                 pthread_cond_timedwait(&helper_cond, &helper_lock, &ts);
2142
2143                 ret = update_io_ticks();
2144
2145                 if (helper_do_stat) {
2146                         helper_do_stat = 0;
2147                         __show_running_run_stats();
2148                 }
2149
2150                 if (!is_backend)
2151                         print_thread_status();
2152         }
2153
2154         return NULL;
2155 }
2156
2157 static int create_helper_thread(void)
2158 {
2159         int ret;
2160
2161         setup_disk_util();
2162
2163         pthread_cond_init(&helper_cond, NULL);
2164         pthread_mutex_init(&helper_lock, NULL);
2165
2166         ret = pthread_create(&helper_thread, NULL, helper_thread_main, NULL);
2167         if (ret) {
2168                 log_err("Can't create helper thread: %s\n", strerror(ret));
2169                 return 1;
2170         }
2171
2172         dprint(FD_MUTEX, "wait on startup_mutex\n");
2173         fio_mutex_down(startup_mutex);
2174         dprint(FD_MUTEX, "done waiting on startup_mutex\n");
2175         return 0;
2176 }
2177
2178 int fio_backend(void)
2179 {
2180         struct thread_data *td;
2181         int i;
2182
2183         if (exec_profile) {
2184                 if (load_profile(exec_profile))
2185                         return 1;
2186                 free(exec_profile);
2187                 exec_profile = NULL;
2188         }
2189         if (!thread_number)
2190                 return 0;
2191
2192         if (write_bw_log) {
2193                 struct log_params p = {
2194                         .log_type = IO_LOG_TYPE_BW,
2195                 };
2196
2197                 setup_log(&agg_io_log[DDIR_READ], &p, "agg-read_bw.log");
2198                 setup_log(&agg_io_log[DDIR_WRITE], &p, "agg-write_bw.log");
2199                 setup_log(&agg_io_log[DDIR_TRIM], &p, "agg-trim_bw.log");
2200         }
2201
2202         startup_mutex = fio_mutex_init(FIO_MUTEX_LOCKED);
2203         if (startup_mutex == NULL)
2204                 return 1;
2205
2206         set_genesis_time();
2207         stat_init();
2208         create_helper_thread();
2209
2210         cgroup_list = smalloc(sizeof(*cgroup_list));
2211         INIT_FLIST_HEAD(cgroup_list);
2212
2213         run_threads();
2214
2215         wait_for_helper_thread_exit();
2216
2217         if (!fio_abort) {
2218                 __show_run_stats();
2219                 if (write_bw_log) {
2220                         for (i = 0; i < DDIR_RWDIR_CNT; i++) {
2221                                 struct io_log *log = agg_io_log[i];
2222
2223                                 flush_log(log);
2224                                 free_log(log);
2225                         }
2226                 }
2227         }
2228
2229         for_each_td(td, i) {
2230                 fio_options_free(td);
2231                 if (td->rusage_sem) {
2232                         fio_mutex_remove(td->rusage_sem);
2233                         td->rusage_sem = NULL;
2234                 }
2235         }
2236
2237         free_disk_util();
2238         cgroup_kill(cgroup_list);
2239         sfree(cgroup_list);
2240         sfree(cgroup_mnt);
2241
2242         fio_mutex_remove(startup_mutex);
2243         stat_exit();
2244         return exit_value;
2245 }