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
 569                 ret = td_io_queue(td, io_u);
 570                 switch (ret) {
 571                 case FIO_Q_COMPLETED:
 572                         if (io_u->error) {
 573                                 ret = -io_u->error;
 574                                 clear_io_u(td, io_u);
 575                         } else if (io_u->resid) {
 576                                 int bytes = io_u->xfer_buflen - io_u->resid;
 577
 578                                 /*
 579                                  * zero read, fail
 580                                  */
 581                                 if (!bytes) {
 582                                         td_verror(td, EIO, "full resid");
 583                                         put_io_u(td, io_u);
 584                                         break;
 585                                 }
 586
 587                                 io_u->xfer_buflen = io_u->resid;
 588                                 io_u->xfer_buf += bytes;
 589                                 io_u->offset += bytes;
 590
 591                                 if (ddir_rw(io_u->ddir))
 592                                         td->ts.short_io_u[io_u->ddir]++;
 593
 594                                 f = io_u->file;
 595                                 if (io_u->offset == f->real_file_size)
 596                                         goto sync_done;
 597
 598                                 requeue_io_u(td, &io_u);
 599                         } else {
 600 sync_done:
 601                                 ret = io_u_sync_complete(td, io_u, bytes_done);
 602                                 if (ret < 0)
 603                                         break;
 604                         }
 605                         continue;
 606                 case FIO_Q_QUEUED:
 607                         break;
 608                 case FIO_Q_BUSY:
 609                         requeue_io_u(td, &io_u);
 610                         ret2 = td_io_commit(td);
 611                         if (ret2 < 0)
 612                                 ret = ret2;
 613                         break;
 614                 default:
 615                         assert(ret < 0);
 616                         td_verror(td, -ret, "td_io_queue");
 617                         break;
 618                 }
 619
 620                 if (break_on_this_error(td, ddir, &ret))
 621                         break;
 622
 623                 /*
 624                  * if we can queue more, do so. but check if there are
 625                  * completed io_u's first. Note that we can get BUSY even
 626                  * without IO queued, if the system is resource starved.
 627                  */
 628 reap:
 629                 full = queue_full(td) || (ret == FIO_Q_BUSY && td->cur_depth);
 630                 if (full || !td->o.iodepth_batch_complete)
 631                         ret = wait_for_completions(td, NULL, bytes_done);
 632
 633                 if (ret < 0)
 634                         break;
 635         }
 636
 637         check_update_rusage(td);
 638
 639         if (!td->error) {
 640                 min_events = td->cur_depth;
 641
 642                 if (min_events)
 643                         ret = io_u_queued_complete(td, min_events, NULL);
 644         } else
 645                 cleanup_pending_aio(td);
 646
 647         td_set_runstate(td, TD_RUNNING);
 648
 649         dprint(FD_VERIFY, "exiting loop\n");
 650 }
 651
 652 static unsigned int exceeds_number_ios(struct thread_data *td)
 653 {
 654         unsigned long long number_ios;
 655
 656         if (!td->o.number_ios)
 657                 return 0;
 658
 659         number_ios = ddir_rw_sum(td->this_io_blocks);
 660         number_ios += td->io_u_queued + td->io_u_in_flight;
 661
 662         return number_ios >= td->o.number_ios;
 663 }
 664
 665 static int io_issue_bytes_exceeded(struct thread_data *td)
 666 {
 667         unsigned long long bytes, limit;
 668
 669         if (td_rw(td))
 670                 bytes = td->io_issue_bytes[DDIR_READ] + td->io_issue_bytes[DDIR_WRITE];
 671         else if (td_write(td))
 672                 bytes = td->io_issue_bytes[DDIR_WRITE];
 673         else if (td_read(td))
 674                 bytes = td->io_issue_bytes[DDIR_READ];
 675         else
 676                 bytes = td->io_issue_bytes[DDIR_TRIM];
 677
 678         if (td->o.io_limit)
 679                 limit = td->o.io_limit;
 680         else
 681                 limit = td->o.size;
 682
 683         return bytes >= limit || exceeds_number_ios(td);
 684 }
 685
 686 static int io_complete_bytes_exceeded(struct thread_data *td)
 687 {
 688         unsigned long long bytes, limit;
 689
 690         if (td_rw(td))
 691                 bytes = td->this_io_bytes[DDIR_READ] + td->this_io_bytes[DDIR_WRITE];
 692         else if (td_write(td))
 693                 bytes = td->this_io_bytes[DDIR_WRITE];
 694         else if (td_read(td))
 695                 bytes = td->this_io_bytes[DDIR_READ];
 696         else
 697                 bytes = td->this_io_bytes[DDIR_TRIM];
 698
 699         if (td->o.io_limit)
 700                 limit = td->o.io_limit;
 701         else
 702                 limit = td->o.size;
 703
 704         return bytes >= limit || exceeds_number_ios(td);
 705 }
 706
 707 /*
 708  * Main IO worker function. It retrieves io_u's to process and queues
 709  * and reaps them, checking for rate and errors along the way.
 710  *
 711  * Returns number of bytes written and trimmed.
 712  */
 713 static uint64_t do_io(struct thread_data *td)
 714 {
 715         uint64_t bytes_done[DDIR_RWDIR_CNT] = { 0, 0, 0 };
 716         unsigned int i;
 717         int ret = 0;
 718         uint64_t total_bytes, bytes_issued = 0;
 719
 720         if (in_ramp_time(td))
 721                 td_set_runstate(td, TD_RAMP);
 722         else
 723                 td_set_runstate(td, TD_RUNNING);
 724
 725         lat_target_init(td);
 726
 727         /*
 728          * If verify_backlog is enabled, we'll run the verify in this
 729          * handler as well. For that case, we may need up to twice the
 730          * amount of bytes.
 731          */
 732         total_bytes = td->o.size;
 733         if (td->o.verify != VERIFY_NONE &&
 734            (td_write(td) && td->o.verify_backlog))
 735                 total_bytes += td->o.size;
 736
 737         while ((td->o.read_iolog_file && !flist_empty(&td->io_log_list)) ||
 738                 (!flist_empty(&td->trim_list)) || !io_issue_bytes_exceeded(td) ||
 739                 td->o.time_based) {
 740                 struct timeval comp_time;
 741                 struct io_u *io_u;
 742                 int ret2, full;
 743                 enum fio_ddir ddir;
 744
 745                 check_update_rusage(td);
 746
 747                 if (td->terminate || td->done)
 748                         break;
 749
 750                 update_tv_cache(td);
 751
 752                 if (runtime_exceeded(td, &td->tv_cache)) {
 753                         __update_tv_cache(td);
 754                         if (runtime_exceeded(td, &td->tv_cache)) {
 755                                 fio_mark_td_terminate(td);
 756                                 break;
 757                         }
 758                 }
 759
 760                 if (flow_threshold_exceeded(td))
 761                         continue;
 762
 763                 if (bytes_issued >= total_bytes)
 764                         break;
 765
 766                 io_u = get_io_u(td);
 767                 if (IS_ERR_OR_NULL(io_u)) {
 768                         int err = PTR_ERR(io_u);
 769
 770                         io_u = NULL;
 771                         if (err == -EBUSY) {
 772                                 ret = FIO_Q_BUSY;
 773                                 goto reap;
 774                         }
 775                         if (td->o.latency_target)
 776                                 goto reap;
 777                         break;
 778                 }
 779
 780                 ddir = io_u->ddir;
 781
 782                 /*
 783                  * Add verification end_io handler if:
 784                  *      - Asked to verify (!td_rw(td))
 785                  *      - Or the io_u is from our verify list (mixed write/ver)
 786                  */
 787                 if (td->o.verify != VERIFY_NONE && io_u->ddir == DDIR_READ &&
 788                     ((io_u->flags & IO_U_F_VER_LIST) || !td_rw(td))) {
 789
 790                         if (!td->o.verify_pattern_bytes) {
 791                                 io_u->rand_seed = __rand(&td->verify_state);
 792                                 if (sizeof(int) != sizeof(long *))
 793                                         io_u->rand_seed *= __rand(&td->verify_state);
 794                         }
 795
 796                         if (verify_state_should_stop(td, io_u)) {
 797                                 put_io_u(td, io_u);
 798                                 break;
 799                         }
 800
 801                         if (td->o.verify_async)
 802                                 io_u->end_io = verify_io_u_async;
 803                         else
 804                                 io_u->end_io = verify_io_u;
 805                         td_set_runstate(td, TD_VERIFYING);
 806                 } else if (in_ramp_time(td))
 807                         td_set_runstate(td, TD_RAMP);
 808                 else
 809                         td_set_runstate(td, TD_RUNNING);
 810
 811                 /*
 812                  * Always log IO before it's issued, so we know the specific
 813                  * order of it. The logged unit will track when the IO has
 814                  * completed.
 815                  */
 816                 if (td_write(td) && io_u->ddir == DDIR_WRITE &&
 817                     td->o.do_verify &&
 818                     td->o.verify != VERIFY_NONE &&
 819                     !td->o.experimental_verify)
 820                         log_io_piece(td, io_u);
 821
 822                 ret = td_io_queue(td, io_u);
 823                 switch (ret) {
 824                 case FIO_Q_COMPLETED:
 825                         if (io_u->error) {
 826                                 ret = -io_u->error;
 827                                 unlog_io_piece(td, io_u);
 828                                 clear_io_u(td, io_u);
 829                         } else if (io_u->resid) {
 830                                 int bytes = io_u->xfer_buflen - io_u->resid;
 831                                 struct fio_file *f = io_u->file;
 832
 833                                 bytes_issued += bytes;
 834
 835                                 trim_io_piece(td, io_u);
 836
 837                                 /*
 838                                  * zero read, fail
 839                                  */
 840                                 if (!bytes) {
 841                                         unlog_io_piece(td, io_u);
 842                                         td_verror(td, EIO, "full resid");
 843                                         put_io_u(td, io_u);
 844                                         break;
 845                                 }
 846
 847                                 io_u->xfer_buflen = io_u->resid;
 848                                 io_u->xfer_buf += bytes;
 849                                 io_u->offset += bytes;
 850
 851                                 if (ddir_rw(io_u->ddir))
 852                                         td->ts.short_io_u[io_u->ddir]++;
 853
 854                                 if (io_u->offset == f->real_file_size)
 855                                         goto sync_done;
 856
 857                                 requeue_io_u(td, &io_u);
 858                         } else {
 859 sync_done:
 860                                 if (__should_check_rate(td, DDIR_READ) ||
 861                                     __should_check_rate(td, DDIR_WRITE) ||
 862                                     __should_check_rate(td, DDIR_TRIM))
 863                                         fio_gettime(&comp_time, NULL);
 864
 865                                 ret = io_u_sync_complete(td, io_u, bytes_done);
 866                                 if (ret < 0)
 867                                         break;
 868                                 bytes_issued += io_u->xfer_buflen;
 869                         }
 870                         break;
 871                 case FIO_Q_QUEUED:
 872                         /*
 873                          * if the engine doesn't have a commit hook,
 874                          * the io_u is really queued. if it does have such
 875                          * a hook, it has to call io_u_queued() itself.
 876                          */
 877                         if (td->io_ops->commit == NULL)
 878                                 io_u_queued(td, io_u);
 879                         bytes_issued += io_u->xfer_buflen;
 880                         break;
 881                 case FIO_Q_BUSY:
 882                         unlog_io_piece(td, io_u);
 883                         requeue_io_u(td, &io_u);
 884                         ret2 = td_io_commit(td);
 885                         if (ret2 < 0)
 886                                 ret = ret2;
 887                         break;
 888                 default:
 889                         assert(ret < 0);
 890                         put_io_u(td, io_u);
 891                         break;
 892                 }
 893
 894                 if (break_on_this_error(td, ddir, &ret))
 895                         break;
 896
 897                 /*
 898                  * See if we need to complete some commands. Note that we
 899                  * can get BUSY even without IO queued, if the system is
 900                  * resource starved.
 901                  */
 902 reap:
 903                 full = queue_full(td) || (ret == FIO_Q_BUSY && td->cur_depth);
 904                 if (full || !td->o.iodepth_batch_complete)
 905                         ret = wait_for_completions(td, &comp_time, bytes_done);
 906                 if (ret < 0)
 907                         break;
 908                 if (!ddir_rw_sum(bytes_done) && !(td->io_ops->flags & FIO_NOIO))
 909                         continue;
 910
 911                 if (!in_ramp_time(td) && should_check_rate(td, bytes_done)) {
 912                         if (check_min_rate(td, &comp_time, bytes_done)) {
 913                                 if (exitall_on_terminate)
 914                                         fio_terminate_threads(td->groupid);
 915                                 td_verror(td, EIO, "check_min_rate");
 916                                 break;
 917                         }
 918                 }
 919                 if (!in_ramp_time(td) && td->o.latency_target)
 920                         lat_target_check(td);
 921
 922                 if (td->o.thinktime) {
 923                         unsigned long long b;
 924
 925                         b = ddir_rw_sum(td->io_blocks);
 926                         if (!(b % td->o.thinktime_blocks)) {
 927                                 int left;
 928
 929                                 io_u_quiesce(td);
 930
 931                                 if (td->o.thinktime_spin)
 932                                         usec_spin(td->o.thinktime_spin);
 933
 934                                 left = td->o.thinktime - td->o.thinktime_spin;
 935                                 if (left)
 936                                         usec_sleep(td, left);
 937                         }
 938                 }
 939         }
 940
 941         check_update_rusage(td);
 942
 943         if (td->trim_entries)
 944                 log_err("fio: %lu trim entries leaked?\n", td->trim_entries);
 945
 946         if (td->o.fill_device && td->error == ENOSPC) {
 947                 td->error = 0;
 948                 fio_mark_td_terminate(td);
 949         }
 950         if (!td->error) {
 951                 struct fio_file *f;
 952
 953                 i = td->cur_depth;
 954                 if (i) {
 955                         ret = io_u_queued_complete(td, i, bytes_done);
 956                         if (td->o.fill_device && td->error == ENOSPC)
 957                                 td->error = 0;
 958                 }
 959
 960                 if (should_fsync(td) && td->o.end_fsync) {
 961                         td_set_runstate(td, TD_FSYNCING);
 962
 963                         for_each_file(td, f, i) {
 964                                 if (!fio_file_fsync(td, f))
 965                                         continue;
 966
 967                                 log_err("fio: end_fsync failed for file %s\n",
 968                                                                 f->file_name);
 969                         }
 970                 }
 971         } else
 972                 cleanup_pending_aio(td);
 973
 974         /*
 975          * stop job if we failed doing any IO
 976          */
 977         if (!ddir_rw_sum(td->this_io_bytes))
 978                 td->done = 1;
 979
 980         return bytes_done[DDIR_WRITE] + bytes_done[DDIR_TRIM];
 981 }
 982
 983 static void cleanup_io_u(struct thread_data *td)
 984 {
 985         struct io_u *io_u;
 986
 987         while ((io_u = io_u_qpop(&td->io_u_freelist)) != NULL) {
 988
 989                 if (td->io_ops->io_u_free)
 990                         td->io_ops->io_u_free(td, io_u);
 991
 992                 fio_memfree(io_u, sizeof(*io_u));
 993         }
 994
 995         free_io_mem(td);
 996
 997         io_u_rexit(&td->io_u_requeues);
 998         io_u_qexit(&td->io_u_freelist);
 999         io_u_qexit(&td->io_u_all);
1000
1001         if (td->last_write_comp)
1002                 sfree(td->last_write_comp);
1003 }
1004
1005 static int init_io_u(struct thread_data *td)
1006 {
1007         struct io_u *io_u;
1008         unsigned int max_bs, min_write;
1009         int cl_align, i, max_units;
1010         int data_xfer = 1, err;
1011         char *p;
1012
1013         max_units = td->o.iodepth;
1014         max_bs = td_max_bs(td);
1015         min_write = td->o.min_bs[DDIR_WRITE];
1016         td->orig_buffer_size = (unsigned long long) max_bs
1017                                         * (unsigned long long) max_units;
1018
1019         if ((td->io_ops->flags & FIO_NOIO) || !(td_read(td) || td_write(td)))
1020                 data_xfer = 0;
1021
1022         err = 0;
1023         err += io_u_rinit(&td->io_u_requeues, td->o.iodepth);
1024         err += io_u_qinit(&td->io_u_freelist, td->o.iodepth);
1025         err += io_u_qinit(&td->io_u_all, td->o.iodepth);
1026
1027         if (err) {
1028                 log_err("fio: failed setting up IO queues\n");
1029                 return 1;
1030         }
1031
1032         /*
1033          * if we may later need to do address alignment, then add any
1034          * possible adjustment here so that we don't cause a buffer
1035          * overflow later. this adjustment may be too much if we get
1036          * lucky and the allocator gives us an aligned address.
1037          */
1038         if (td->o.odirect || td->o.mem_align || td->o.oatomic ||
1039             (td->io_ops->flags & FIO_RAWIO))
1040                 td->orig_buffer_size += page_mask + td->o.mem_align;
1041
1042         if (td->o.mem_type == MEM_SHMHUGE || td->o.mem_type == MEM_MMAPHUGE) {
1043                 unsigned long bs;
1044
1045                 bs = td->orig_buffer_size + td->o.hugepage_size - 1;
1046                 td->orig_buffer_size = bs & ~(td->o.hugepage_size - 1);
1047         }
1048
1049         if (td->orig_buffer_size != (size_t) td->orig_buffer_size) {
1050                 log_err("fio: IO memory too large. Reduce max_bs or iodepth\n");
1051                 return 1;
1052         }
1053
1054         if (data_xfer && allocate_io_mem(td))
1055                 return 1;
1056
1057         if (td->o.odirect || td->o.mem_align || td->o.oatomic ||
1058             (td->io_ops->flags & FIO_RAWIO))
1059                 p = PAGE_ALIGN(td->orig_buffer) + td->o.mem_align;
1060         else
1061                 p = td->orig_buffer;
1062
1063         cl_align = os_cache_line_size();
1064
1065         for (i = 0; i < max_units; i++) {
1066                 void *ptr;
1067
1068                 if (td->terminate)
1069                         return 1;
1070
1071                 ptr = fio_memalign(cl_align, sizeof(*io_u));
1072                 if (!ptr) {
1073                         log_err("fio: unable to allocate aligned memory\n");
1074                         break;
1075                 }
1076
1077                 io_u = ptr;
1078                 memset(io_u, 0, sizeof(*io_u));
1079                 INIT_FLIST_HEAD(&io_u->verify_list);
1080                 dprint(FD_MEM, "io_u alloc %p, index %u\n", io_u, i);
1081
1082                 if (data_xfer) {
1083                         io_u->buf = p;
1084                         dprint(FD_MEM, "io_u %p, mem %p\n", io_u, io_u->buf);
1085
1086                         if (td_write(td))
1087                                 io_u_fill_buffer(td, io_u, min_write, max_bs);
1088                         if (td_write(td) && td->o.verify_pattern_bytes) {
1089                                 /*
1090                                  * Fill the buffer with the pattern if we are
1091                                  * going to be doing writes.
1092                                  */
1093                                 fill_verify_pattern(td, io_u->buf, max_bs, io_u, 0, 0);
1094                         }
1095                 }
1096
1097                 io_u->index = i;
1098                 io_u->flags = IO_U_F_FREE;
1099                 io_u_qpush(&td->io_u_freelist, io_u);
1100
1101                 /*
1102                  * io_u never leaves this stack, used for iteration of all
1103                  * io_u buffers.
1104                  */
1105                 io_u_qpush(&td->io_u_all, io_u);
1106
1107                 if (td->io_ops->io_u_init) {
1108                         int ret = td->io_ops->io_u_init(td, io_u);
1109
1110                         if (ret) {
1111                                 log_err("fio: failed to init engine data: %d\n", ret);
1112                                 return 1;
1113                         }
1114                 }
1115
1116                 p += max_bs;
1117         }
1118
1119         if (td->o.verify != VERIFY_NONE) {
1120                 td->last_write_comp = scalloc(max_units, sizeof(uint64_t));
1121                 if (!td->last_write_comp) {
1122                         log_err("fio: failed to alloc write comp data\n");
1123                         return 1;
1124                 }
1125         }
1126
1127         return 0;
1128 }
1129
1130 static int switch_ioscheduler(struct thread_data *td)
1131 {
1132         char tmp[256], tmp2[128];
1133         FILE *f;
1134         int ret;
1135
1136         if (td->io_ops->flags & FIO_DISKLESSIO)
1137                 return 0;
1138
1139         sprintf(tmp, "%s/queue/scheduler", td->sysfs_root);
1140
1141         f = fopen(tmp, "r+");
1142         if (!f) {
1143                 if (errno == ENOENT) {
1144                         log_err("fio: os or kernel doesn't support IO scheduler"
1145                                 " switching\n");
1146                         return 0;
1147                 }
1148                 td_verror(td, errno, "fopen iosched");
1149                 return 1;
1150         }
1151
1152         /*
1153          * Set io scheduler.
1154          */
1155         ret = fwrite(td->o.ioscheduler, strlen(td->o.ioscheduler), 1, f);
1156         if (ferror(f) || ret != 1) {
1157                 td_verror(td, errno, "fwrite");
1158                 fclose(f);
1159                 return 1;
1160         }
1161
1162         rewind(f);
1163
1164         /*
1165          * Read back and check that the selected scheduler is now the default.
1166          */
1167         ret = fread(tmp, sizeof(tmp), 1, f);
1168         if (ferror(f) || ret < 0) {
1169                 td_verror(td, errno, "fread");
1170                 fclose(f);
1171                 return 1;
1172         }
1173         tmp[sizeof(tmp) - 1] = '\0';
1174
1175
1176         sprintf(tmp2, "[%s]", td->o.ioscheduler);
1177         if (!strstr(tmp, tmp2)) {
1178                 log_err("fio: io scheduler %s not found\n", td->o.ioscheduler);
1179                 td_verror(td, EINVAL, "iosched_switch");
1180                 fclose(f);
1181                 return 1;
1182         }
1183
1184         fclose(f);
1185         return 0;
1186 }
1187
1188 static int keep_running(struct thread_data *td)
1189 {
1190         unsigned long long limit;
1191
1192         if (td->done)
1193                 return 0;
1194         if (td->o.time_based)
1195                 return 1;
1196         if (td->o.loops) {
1197                 td->o.loops--;
1198                 return 1;
1199         }
1200         if (exceeds_number_ios(td))
1201                 return 0;
1202
1203         if (td->o.io_limit)
1204                 limit = td->o.io_limit;
1205         else
1206                 limit = td->o.size;
1207
1208         if (limit != -1ULL && ddir_rw_sum(td->io_bytes) < limit) {
1209                 uint64_t diff;
1210
1211                 /*
1212                  * If the difference is less than the minimum IO size, we
1213                  * are done.
1214                  */
1215                 diff = limit - ddir_rw_sum(td->io_bytes);
1216                 if (diff < td_max_bs(td))
1217                         return 0;
1218
1219                 if (fio_files_done(td))
1220                         return 0;
1221
1222                 return 1;
1223         }
1224
1225         return 0;
1226 }
1227
1228 static int exec_string(struct thread_options *o, const char *string, const char *mode)
1229 {
1230         int ret, newlen = strlen(string) + strlen(o->name) + strlen(mode) + 9 + 1;
1231         char *str;
1232
1233         str = malloc(newlen);
1234         sprintf(str, "%s &> %s.%s.txt", string, o->name, mode);
1235
1236         log_info("%s : Saving output of %s in %s.%s.txt\n",o->name, mode, o->name, mode);
1237         ret = system(str);
1238         if (ret == -1)
1239                 log_err("fio: exec of cmd <%s> failed\n", str);
1240
1241         free(str);
1242         return ret;
1243 }
1244
1245 /*
1246  * Dry run to compute correct state of numberio for verification.
1247  */
1248 static uint64_t do_dry_run(struct thread_data *td)
1249 {
1250         uint64_t bytes_done[DDIR_RWDIR_CNT] = { 0, 0, 0 };
1251
1252         td_set_runstate(td, TD_RUNNING);
1253
1254         while ((td->o.read_iolog_file && !flist_empty(&td->io_log_list)) ||
1255                 (!flist_empty(&td->trim_list)) || !io_complete_bytes_exceeded(td)) {
1256                 struct io_u *io_u;
1257                 int ret;
1258
1259                 if (td->terminate || td->done)
1260                         break;
1261
1262                 io_u = get_io_u(td);
1263                 if (!io_u)
1264                         break;
1265
1266                 io_u->flags |= IO_U_F_FLIGHT;
1267                 io_u->error = 0;
1268                 io_u->resid = 0;
1269                 if (ddir_rw(acct_ddir(io_u)))
1270                         td->io_issues[acct_ddir(io_u)]++;
1271                 if (ddir_rw(io_u->ddir)) {
1272                         io_u_mark_depth(td, 1);
1273                         td->ts.total_io_u[io_u->ddir]++;
1274                 }
1275
1276                 if (td_write(td) && io_u->ddir == DDIR_WRITE &&
1277                     td->o.do_verify &&
1278                     td->o.verify != VERIFY_NONE &&
1279                     !td->o.experimental_verify)
1280                         log_io_piece(td, io_u);
1281
1282                 ret = io_u_sync_complete(td, io_u, bytes_done);
1283                 (void) ret;
1284         }
1285
1286         return bytes_done[DDIR_WRITE] + bytes_done[DDIR_TRIM];
1287 }
1288
1289 /*
1290  * Entry point for the thread based jobs. The process based jobs end up
1291  * here as well, after a little setup.
1292  */
1293 static void *thread_main(void *data)
1294 {
1295         unsigned long long elapsed;
1296         struct thread_data *td = data;
1297         struct thread_options *o = &td->o;
1298         pthread_condattr_t attr;
1299         int clear_state;
1300         int ret;
1301
1302         if (!o->use_thread) {
1303                 setsid();
1304                 td->pid = getpid();
1305         } else
1306                 td->pid = gettid();
1307
1308         fio_local_clock_init(o->use_thread);
1309
1310         dprint(FD_PROCESS, "jobs pid=%d started\n", (int) td->pid);
1311
1312         if (is_backend)
1313                 fio_server_send_start(td);
1314
1315         INIT_FLIST_HEAD(&td->io_log_list);
1316         INIT_FLIST_HEAD(&td->io_hist_list);
1317         INIT_FLIST_HEAD(&td->verify_list);
1318         INIT_FLIST_HEAD(&td->trim_list);
1319         INIT_FLIST_HEAD(&td->next_rand_list);
1320         pthread_mutex_init(&td->io_u_lock, NULL);
1321         td->io_hist_tree = RB_ROOT;
1322
1323         pthread_condattr_init(&attr);
1324         pthread_cond_init(&td->verify_cond, &attr);
1325         pthread_cond_init(&td->free_cond, &attr);
1326
1327         td_set_runstate(td, TD_INITIALIZED);
1328         dprint(FD_MUTEX, "up startup_mutex\n");
1329         fio_mutex_up(startup_mutex);
1330         dprint(FD_MUTEX, "wait on td->mutex\n");
1331         fio_mutex_down(td->mutex);
1332         dprint(FD_MUTEX, "done waiting on td->mutex\n");
1333
1334         /*
1335          * A new gid requires privilege, so we need to do this before setting
1336          * the uid.
1337          */
1338         if (o->gid != -1U && setgid(o->gid)) {
1339                 td_verror(td, errno, "setgid");
1340                 goto err;
1341         }
1342         if (o->uid != -1U && setuid(o->uid)) {
1343                 td_verror(td, errno, "setuid");
1344                 goto err;
1345         }
1346
1347         /*
1348          * If we have a gettimeofday() thread, make sure we exclude that
1349          * thread from this job
1350          */
1351         if (o->gtod_cpu)
1352                 fio_cpu_clear(&o->cpumask, o->gtod_cpu);
1353
1354         /*
1355          * Set affinity first, in case it has an impact on the memory
1356          * allocations.
1357          */
1358         if (fio_option_is_set(o, cpumask)) {
1359                 if (o->cpus_allowed_policy == FIO_CPUS_SPLIT) {
1360                         ret = fio_cpus_split(&o->cpumask, td->thread_number - 1);
1361                         if (!ret) {
1362                                 log_err("fio: no CPUs set\n");
1363                                 log_err("fio: Try increasing number of available CPUs\n");
1364                                 td_verror(td, EINVAL, "cpus_split");
1365                                 goto err;
1366                         }
1367                 }
1368                 ret = fio_setaffinity(td->pid, o->cpumask);
1369                 if (ret == -1) {
1370                         td_verror(td, errno, "cpu_set_affinity");
1371                         goto err;
1372                 }
1373         }
1374
1375 #ifdef CONFIG_LIBNUMA
1376         /* numa node setup */
1377         if (fio_option_is_set(o, numa_cpunodes) ||
1378             fio_option_is_set(o, numa_memnodes)) {
1379                 struct bitmask *mask;
1380
1381                 if (numa_available() < 0) {
1382                         td_verror(td, errno, "Does not support NUMA API\n");
1383                         goto err;
1384                 }
1385
1386                 if (fio_option_is_set(o, numa_cpunodes)) {
1387                         mask = numa_parse_nodestring(o->numa_cpunodes);
1388                         ret = numa_run_on_node_mask(mask);
1389                         numa_free_nodemask(mask);
1390                         if (ret == -1) {
1391                                 td_verror(td, errno, \
1392                                         "numa_run_on_node_mask failed\n");
1393                                 goto err;
1394                         }
1395                 }
1396
1397                 if (fio_option_is_set(o, numa_memnodes)) {
1398                         mask = NULL;
1399                         if (o->numa_memnodes)
1400                                 mask = numa_parse_nodestring(o->numa_memnodes);
1401
1402                         switch (o->numa_mem_mode) {
1403                         case MPOL_INTERLEAVE:
1404                                 numa_set_interleave_mask(mask);
1405                                 break;
1406                         case MPOL_BIND:
1407                                 numa_set_membind(mask);
1408                                 break;
1409                         case MPOL_LOCAL:
1410                                 numa_set_localalloc();
1411                                 break;
1412                         case MPOL_PREFERRED:
1413                                 numa_set_preferred(o->numa_mem_prefer_node);
1414                                 break;
1415                         case MPOL_DEFAULT:
1416                         default:
1417                                 break;
1418                         }
1419
1420                         if (mask)
1421                                 numa_free_nodemask(mask);
1422
1423                 }
1424         }
1425 #endif
1426
1427         if (fio_pin_memory(td))
1428                 goto err;
1429
1430         /*
1431          * May alter parameters that init_io_u() will use, so we need to
1432          * do this first.
1433          */
1434         if (init_iolog(td))
1435                 goto err;
1436
1437         if (init_io_u(td))
1438                 goto err;
1439
1440         if (o->verify_async && verify_async_init(td))
1441                 goto err;
1442
1443         if (fio_option_is_set(o, ioprio) ||
1444             fio_option_is_set(o, ioprio_class)) {
1445                 ret = ioprio_set(IOPRIO_WHO_PROCESS, 0, o->ioprio_class, o->ioprio);
1446                 if (ret == -1) {
1447                         td_verror(td, errno, "ioprio_set");
1448                         goto err;
1449                 }
1450         }
1451
1452         if (o->cgroup && cgroup_setup(td, cgroup_list, &cgroup_mnt))
1453                 goto err;
1454
1455         errno = 0;
1456         if (nice(o->nice) == -1 && errno != 0) {
1457                 td_verror(td, errno, "nice");
1458                 goto err;
1459         }
1460
1461         if (o->ioscheduler && switch_ioscheduler(td))
1462                 goto err;
1463
1464         if (!o->create_serialize && setup_files(td))
1465                 goto err;
1466
1467         if (td_io_init(td))
1468                 goto err;
1469
1470         if (init_random_map(td))
1471                 goto err;
1472
1473         if (o->exec_prerun && exec_string(o, o->exec_prerun, (const char *)"prerun"))
1474                 goto err;
1475
1476         if (o->pre_read) {
1477                 if (pre_read_files(td) < 0)
1478                         goto err;
1479         }
1480
1481         if (td->flags & TD_F_COMPRESS_LOG)
1482                 tp_init(&td->tp_data);
1483
1484         fio_verify_init(td);
1485
1486         fio_gettime(&td->epoch, NULL);
1487         fio_getrusage(&td->ru_start);
1488         clear_state = 0;
1489         while (keep_running(td)) {
1490                 uint64_t verify_bytes;
1491
1492                 fio_gettime(&td->start, NULL);
1493                 memcpy(&td->bw_sample_time, &td->start, sizeof(td->start));
1494                 memcpy(&td->iops_sample_time, &td->start, sizeof(td->start));
1495                 memcpy(&td->tv_cache, &td->start, sizeof(td->start));
1496
1497                 if (o->ratemin[DDIR_READ] || o->ratemin[DDIR_WRITE] ||
1498                                 o->ratemin[DDIR_TRIM]) {
1499                         memcpy(&td->lastrate[DDIR_READ], &td->bw_sample_time,
1500                                                 sizeof(td->bw_sample_time));
1501                         memcpy(&td->lastrate[DDIR_WRITE], &td->bw_sample_time,
1502                                                 sizeof(td->bw_sample_time));
1503                         memcpy(&td->lastrate[DDIR_TRIM], &td->bw_sample_time,
1504                                                 sizeof(td->bw_sample_time));
1505                 }
1506
1507                 if (clear_state)
1508                         clear_io_state(td);
1509
1510                 prune_io_piece_log(td);
1511
1512                 if (td->o.verify_only && (td_write(td) || td_rw(td)))
1513                         verify_bytes = do_dry_run(td);
1514                 else
1515                         verify_bytes = do_io(td);
1516
1517                 clear_state = 1;
1518
1519                 fio_mutex_down(stat_mutex);
1520                 if (td_read(td) && td->io_bytes[DDIR_READ]) {
1521                         elapsed = mtime_since_now(&td->start);
1522                         td->ts.runtime[DDIR_READ] += elapsed;
1523                 }
1524                 if (td_write(td) && td->io_bytes[DDIR_WRITE]) {
1525                         elapsed = mtime_since_now(&td->start);
1526                         td->ts.runtime[DDIR_WRITE] += elapsed;
1527                 }
1528                 if (td_trim(td) && td->io_bytes[DDIR_TRIM]) {
1529                         elapsed = mtime_since_now(&td->start);
1530                         td->ts.runtime[DDIR_TRIM] += elapsed;
1531                 }
1532                 fio_gettime(&td->start, NULL);
1533                 fio_mutex_up(stat_mutex);
1534
1535                 if (td->error || td->terminate)
1536                         break;
1537
1538                 if (!o->do_verify ||
1539                     o->verify == VERIFY_NONE ||
1540                     (td->io_ops->flags & FIO_UNIDIR))
1541                         continue;
1542
1543                 clear_io_state(td);
1544
1545                 fio_gettime(&td->start, NULL);
1546
1547                 do_verify(td, verify_bytes);
1548
1549                 fio_mutex_down(stat_mutex);
1550                 td->ts.runtime[DDIR_READ] += mtime_since_now(&td->start);
1551                 fio_gettime(&td->start, NULL);
1552                 fio_mutex_up(stat_mutex);
1553
1554                 if (td->error || td->terminate)
1555                         break;
1556         }
1557
1558         update_rusage_stat(td);
1559         td->ts.total_run_time = mtime_since_now(&td->epoch);
1560         td->ts.io_bytes[DDIR_READ] = td->io_bytes[DDIR_READ];
1561         td->ts.io_bytes[DDIR_WRITE] = td->io_bytes[DDIR_WRITE];
1562         td->ts.io_bytes[DDIR_TRIM] = td->io_bytes[DDIR_TRIM];
1563
1564         if (td->o.verify_state_save && !(td->flags & TD_F_VSTATE_SAVED) &&
1565             (td->o.verify != VERIFY_NONE && td_write(td))) {
1566                 struct all_io_list *state;
1567                 size_t sz;
1568
1569                 state = get_all_io_list(td->thread_number, &sz);
1570                 if (state) {
1571                         __verify_save_state(state, "local");
1572                         free(state);
1573                 }
1574         }
1575
1576         fio_unpin_memory(td);
1577
1578         fio_writeout_logs(td);
1579
1580         if (td->flags & TD_F_COMPRESS_LOG)
1581                 tp_exit(&td->tp_data);
1582
1583         if (o->exec_postrun)
1584                 exec_string(o, o->exec_postrun, (const char *)"postrun");
1585
1586         if (exitall_on_terminate)
1587                 fio_terminate_threads(td->groupid);
1588
1589 err:
1590         if (td->error)
1591                 log_info("fio: pid=%d, err=%d/%s\n", (int) td->pid, td->error,
1592                                                         td->verror);
1593
1594         if (o->verify_async)
1595                 verify_async_exit(td);
1596
1597         close_and_free_files(td);
1598         cleanup_io_u(td);
1599         close_ioengine(td);
1600         cgroup_shutdown(td, &cgroup_mnt);
1601         verify_free_state(td);
1602
1603         if (fio_option_is_set(o, cpumask)) {
1604                 ret = fio_cpuset_exit(&o->cpumask);
1605                 if (ret)
1606                         td_verror(td, ret, "fio_cpuset_exit");
1607         }
1608
1609         /*
1610          * do this very late, it will log file closing as well
1611          */
1612         if (o->write_iolog_file)
1613                 write_iolog_close(td);
1614
1615         fio_mutex_remove(td->mutex);
1616         td->mutex = NULL;
1617
1618         td_set_runstate(td, TD_EXITED);
1619
1620         /*
1621          * Do this last after setting our runstate to exited, so we
1622          * know that the stat thread is signaled.
1623          */
1624         check_update_rusage(td);
1625
1626         return (void *) (uintptr_t) td->error;
1627 }
1628
1629
1630 /*
1631  * We cannot pass the td data into a forked process, so attach the td and
1632  * pass it to the thread worker.
1633  */
1634 static int fork_main(int shmid, int offset)
1635 {
1636         struct thread_data *td;
1637         void *data, *ret;
1638
1639 #if !defined(__hpux) && !defined(CONFIG_NO_SHM)
1640         data = shmat(shmid, NULL, 0);
1641         if (data == (void *) -1) {
1642                 int __err = errno;
1643
1644                 perror("shmat");
1645                 return __err;
1646         }
1647 #else
1648         /*
1649          * HP-UX inherits shm mappings?
1650          */
1651         data = threads;
1652 #endif
1653
1654         td = data + offset * sizeof(struct thread_data);
1655         ret = thread_main(td);
1656         shmdt(data);
1657         return (int) (uintptr_t) ret;
1658 }
1659
1660 static void dump_td_info(struct thread_data *td)
1661 {
1662         log_err("fio: job '%s' hasn't exited in %lu seconds, it appears to "
1663                 "be stuck. Doing forceful exit of this job.\n", td->o.name,
1664                         (unsigned long) time_since_now(&td->terminate_time));
1665 }
1666
1667 /*
1668  * Run over the job map and reap the threads that have exited, if any.
1669  */
1670 static void reap_threads(unsigned int *nr_running, unsigned int *t_rate,
1671                          unsigned int *m_rate)
1672 {
1673         struct thread_data *td;
1674         unsigned int cputhreads, realthreads, pending;
1675         int i, status, ret;
1676
1677         /*
1678          * reap exited threads (TD_EXITED -> TD_REAPED)
1679          */
1680         realthreads = pending = cputhreads = 0;
1681         for_each_td(td, i) {
1682                 int flags = 0;
1683
1684                 /*
1685                  * ->io_ops is NULL for a thread that has closed its
1686                  * io engine
1687                  */
1688                 if (td->io_ops && !strcmp(td->io_ops->name, "cpuio"))
1689                         cputhreads++;
1690                 else
1691                         realthreads++;
1692
1693                 if (!td->pid) {
1694                         pending++;
1695                         continue;
1696                 }
1697                 if (td->runstate == TD_REAPED)
1698                         continue;
1699                 if (td->o.use_thread) {
1700                         if (td->runstate == TD_EXITED) {
1701                                 td_set_runstate(td, TD_REAPED);
1702                                 goto reaped;
1703                         }
1704                         continue;
1705                 }
1706
1707                 flags = WNOHANG;
1708                 if (td->runstate == TD_EXITED)
1709                         flags = 0;
1710
1711                 /*
1712                  * check if someone quit or got killed in an unusual way
1713                  */
1714                 ret = waitpid(td->pid, &status, flags);
1715                 if (ret < 0) {
1716                         if (errno == ECHILD) {
1717                                 log_err("fio: pid=%d disappeared %d\n",
1718                                                 (int) td->pid, td->runstate);
1719                                 td->sig = ECHILD;
1720                                 td_set_runstate(td, TD_REAPED);
1721                                 goto reaped;
1722                         }
1723                         perror("waitpid");
1724                 } else if (ret == td->pid) {
1725                         if (WIFSIGNALED(status)) {
1726                                 int sig = WTERMSIG(status);
1727
1728                                 if (sig != SIGTERM && sig != SIGUSR2)
1729                                         log_err("fio: pid=%d, got signal=%d\n",
1730                                                         (int) td->pid, sig);
1731                                 td->sig = sig;
1732                                 td_set_runstate(td, TD_REAPED);
1733                                 goto reaped;
1734                         }
1735                         if (WIFEXITED(status)) {
1736                                 if (WEXITSTATUS(status) && !td->error)
1737                                         td->error = WEXITSTATUS(status);
1738
1739                                 td_set_runstate(td, TD_REAPED);
1740                                 goto reaped;
1741                         }
1742                 }
1743
1744                 /*
1745                  * If the job is stuck, do a forceful timeout of it and
1746                  * move on.
1747                  */
1748                 if (td->terminate &&
1749                     time_since_now(&td->terminate_time) >= FIO_REAP_TIMEOUT) {
1750                         dump_td_info(td);
1751                         td_set_runstate(td, TD_REAPED);
1752                         goto reaped;
1753                 }
1754
1755                 /*
1756                  * thread is not dead, continue
1757                  */
1758                 pending++;
1759                 continue;
1760 reaped:
1761                 (*nr_running)--;
1762                 (*m_rate) -= ddir_rw_sum(td->o.ratemin);
1763                 (*t_rate) -= ddir_rw_sum(td->o.rate);
1764                 if (!td->pid)
1765                         pending--;
1766
1767                 if (td->error)
1768                         exit_value++;
1769
1770                 done_secs += mtime_since_now(&td->epoch) / 1000;
1771                 profile_td_exit(td);
1772         }
1773
1774         if (*nr_running == cputhreads && !pending && realthreads)
1775                 fio_terminate_threads(TERMINATE_ALL);
1776 }
1777
1778 static int __check_trigger_file(void)
1779 {
1780         struct stat sb;
1781
1782         if (!trigger_file)
1783                 return 0;
1784
1785         if (stat(trigger_file, &sb))
1786                 return 0;
1787
1788         if (unlink(trigger_file) < 0)
1789                 log_err("fio: failed to unlink %s: %s\n", trigger_file,
1790                                                         strerror(errno));
1791
1792         return 1;
1793 }
1794
1795 static int trigger_timedout(void)
1796 {
1797         if (trigger_timeout)
1798                 return time_since_genesis() >= trigger_timeout;
1799
1800         return 0;
1801 }
1802
1803 void exec_trigger(const char *cmd)
1804 {
1805         int ret;
1806
1807         if (!cmd)
1808                 return;
1809
1810         ret = system(cmd);
1811         if (ret == -1)
1812                 log_err("fio: failed executing %s trigger\n", cmd);
1813 }
1814
1815 void check_trigger_file(void)
1816 {
1817         if (__check_trigger_file() || trigger_timedout()) {
1818                 if (nr_clients)
1819                         fio_clients_send_trigger(trigger_remote_cmd);
1820                 else {
1821                         verify_save_state();
1822                         fio_terminate_threads(TERMINATE_ALL);
1823                         exec_trigger(trigger_cmd);
1824                 }
1825         }
1826 }
1827
1828 static int fio_verify_load_state(struct thread_data *td)
1829 {
1830         int ret;
1831
1832         if (!td->o.verify_state)
1833                 return 0;
1834
1835         if (is_backend) {
1836                 void *data;
1837
1838                 ret = fio_server_get_verify_state(td->o.name,
1839                                         td->thread_number - 1, &data);
1840                 if (!ret)
1841                         verify_convert_assign_state(td, data);
1842         } else
1843                 ret = verify_load_state(td, "local");
1844
1845         return ret;
1846 }
1847
1848 static void do_usleep(unsigned int usecs)
1849 {
1850         check_for_running_stats();
1851         check_trigger_file();
1852         usleep(usecs);
1853 }
1854
1855 /*
1856  * Main function for kicking off and reaping jobs, as needed.
1857  */
1858 static void run_threads(void)
1859 {
1860         struct thread_data *td;
1861         unsigned int i, todo, nr_running, m_rate, t_rate, nr_started;
1862         uint64_t spent;
1863
1864         if (fio_gtod_offload && fio_start_gtod_thread())
1865                 return;
1866
1867         fio_idle_prof_init();
1868
1869         set_sig_handlers();
1870
1871         nr_thread = nr_process = 0;
1872         for_each_td(td, i) {
1873                 if (td->o.use_thread)
1874                         nr_thread++;
1875                 else
1876                         nr_process++;
1877         }
1878
1879         if (output_format == FIO_OUTPUT_NORMAL) {
1880                 log_info("Starting ");
1881                 if (nr_thread)
1882                         log_info("%d thread%s", nr_thread,
1883                                                 nr_thread > 1 ? "s" : "");
1884                 if (nr_process) {
1885                         if (nr_thread)
1886                                 log_info(" and ");
1887                         log_info("%d process%s", nr_process,
1888                                                 nr_process > 1 ? "es" : "");
1889                 }
1890                 log_info("\n");
1891                 log_info_flush();
1892         }
1893
1894         todo = thread_number;
1895         nr_running = 0;
1896         nr_started = 0;
1897         m_rate = t_rate = 0;
1898
1899         for_each_td(td, i) {
1900                 print_status_init(td->thread_number - 1);
1901
1902                 if (!td->o.create_serialize)
1903                         continue;
1904
1905                 if (fio_verify_load_state(td))
1906                         goto reap;
1907
1908                 /*
1909                  * do file setup here so it happens sequentially,
1910                  * we don't want X number of threads getting their
1911                  * client data interspersed on disk
1912                  */
1913                 if (setup_files(td)) {
1914 reap:
1915                         exit_value++;
1916                         if (td->error)
1917                                 log_err("fio: pid=%d, err=%d/%s\n",
1918                                         (int) td->pid, td->error, td->verror);
1919                         td_set_runstate(td, TD_REAPED);
1920                         todo--;
1921                 } else {
1922                         struct fio_file *f;
1923                         unsigned int j;
1924
1925                         /*
1926                          * for sharing to work, each job must always open
1927                          * its own files. so close them, if we opened them
1928                          * for creation
1929                          */
1930                         for_each_file(td, f, j) {
1931                                 if (fio_file_open(f))
1932                                         td_io_close_file(td, f);
1933                         }
1934                 }
1935         }
1936
1937         /* start idle threads before io threads start to run */
1938         fio_idle_prof_start();
1939
1940         set_genesis_time();
1941
1942         while (todo) {
1943                 struct thread_data *map[REAL_MAX_JOBS];
1944                 struct timeval this_start;
1945                 int this_jobs = 0, left;
1946
1947                 /*
1948                  * create threads (TD_NOT_CREATED -> TD_CREATED)
1949                  */
1950                 for_each_td(td, i) {
1951                         if (td->runstate != TD_NOT_CREATED)
1952                                 continue;
1953
1954                         /*
1955                          * never got a chance to start, killed by other
1956                          * thread for some reason
1957                          */
1958                         if (td->terminate) {
1959                                 todo--;
1960                                 continue;
1961                         }
1962
1963                         if (td->o.start_delay) {
1964                                 spent = utime_since_genesis();
1965
1966                                 if (td->o.start_delay > spent)
1967                                         continue;
1968                         }
1969
1970                         if (td->o.stonewall && (nr_started || nr_running)) {
1971                                 dprint(FD_PROCESS, "%s: stonewall wait\n",
1972                                                         td->o.name);
1973                                 break;
1974                         }
1975
1976                         init_disk_util(td);
1977
1978                         td->rusage_sem = fio_mutex_init(FIO_MUTEX_LOCKED);
1979                         td->update_rusage = 0;
1980
1981                         /*
1982                          * Set state to created. Thread will transition
1983                          * to TD_INITIALIZED when it's done setting up.
1984                          */
1985                         td_set_runstate(td, TD_CREATED);
1986                         map[this_jobs++] = td;
1987                         nr_started++;
1988
1989                         if (td->o.use_thread) {
1990                                 int ret;
1991
1992                                 dprint(FD_PROCESS, "will pthread_create\n");
1993                                 ret = pthread_create(&td->thread, NULL,
1994                                                         thread_main, td);
1995                                 if (ret) {
1996                                         log_err("pthread_create: %s\n",
1997                                                         strerror(ret));
1998                                         nr_started--;
1999                                         break;
2000                                 }
2001                                 ret = pthread_detach(td->thread);
2002                                 if (ret)
2003                                         log_err("pthread_detach: %s",
2004                                                         strerror(ret));
2005                         } else {
2006                                 pid_t pid;
2007                                 dprint(FD_PROCESS, "will fork\n");
2008                                 pid = fork();
2009                                 if (!pid) {
2010                                         int ret = fork_main(shm_id, i);
2011
2012                                         _exit(ret);
2013                                 } else if (i == fio_debug_jobno)
2014                                         *fio_debug_jobp = pid;
2015                         }
2016                         dprint(FD_MUTEX, "wait on startup_mutex\n");
2017                         if (fio_mutex_down_timeout(startup_mutex, 10)) {
2018                                 log_err("fio: job startup hung? exiting.\n");
2019                                 fio_terminate_threads(TERMINATE_ALL);
2020                                 fio_abort = 1;
2021                                 nr_started--;
2022                                 break;
2023                         }
2024                         dprint(FD_MUTEX, "done waiting on startup_mutex\n");
2025                 }
2026
2027                 /*
2028                  * Wait for the started threads to transition to
2029                  * TD_INITIALIZED.
2030                  */
2031                 fio_gettime(&this_start, NULL);
2032                 left = this_jobs;
2033                 while (left && !fio_abort) {
2034                         if (mtime_since_now(&this_start) > JOB_START_TIMEOUT)
2035                                 break;
2036
2037                         do_usleep(100000);
2038
2039                         for (i = 0; i < this_jobs; i++) {
2040                                 td = map[i];
2041                                 if (!td)
2042                                         continue;
2043                                 if (td->runstate == TD_INITIALIZED) {
2044                                         map[i] = NULL;
2045                                         left--;
2046                                 } else if (td->runstate >= TD_EXITED) {
2047                                         map[i] = NULL;
2048                                         left--;
2049                                         todo--;
2050                                         nr_running++; /* work-around... */
2051                                 }
2052                         }
2053                 }
2054
2055                 if (left) {
2056                         log_err("fio: %d job%s failed to start\n", left,
2057                                         left > 1 ? "s" : "");
2058                         for (i = 0; i < this_jobs; i++) {
2059                                 td = map[i];
2060                                 if (!td)
2061                                         continue;
2062                                 kill(td->pid, SIGTERM);
2063                         }
2064                         break;
2065                 }
2066
2067                 /*
2068                  * start created threads (TD_INITIALIZED -> TD_RUNNING).
2069                  */
2070                 for_each_td(td, i) {
2071                         if (td->runstate != TD_INITIALIZED)
2072                                 continue;
2073
2074                         if (in_ramp_time(td))
2075                                 td_set_runstate(td, TD_RAMP);
2076                         else
2077                                 td_set_runstate(td, TD_RUNNING);
2078                         nr_running++;
2079                         nr_started--;
2080                         m_rate += ddir_rw_sum(td->o.ratemin);
2081                         t_rate += ddir_rw_sum(td->o.rate);
2082                         todo--;
2083                         fio_mutex_up(td->mutex);
2084                 }
2085
2086                 reap_threads(&nr_running, &t_rate, &m_rate);
2087
2088                 if (todo)
2089                         do_usleep(100000);
2090         }
2091
2092         while (nr_running) {
2093                 reap_threads(&nr_running, &t_rate, &m_rate);
2094                 do_usleep(10000);
2095         }
2096
2097         fio_idle_prof_stop();
2098
2099         update_io_ticks();
2100 }
2101
2102 static void wait_for_helper_thread_exit(void)
2103 {
2104         void *ret;
2105
2106         helper_exit = 1;
2107         pthread_cond_signal(&helper_cond);
2108         pthread_join(helper_thread, &ret);
2109 }
2110
2111 static void free_disk_util(void)
2112 {
2113         disk_util_prune_entries();
2114
2115         pthread_cond_destroy(&helper_cond);
2116 }
2117
2118 static void *helper_thread_main(void *data)
2119 {
2120         int ret = 0;
2121
2122         fio_mutex_up(startup_mutex);
2123
2124         while (!ret) {
2125                 uint64_t sec = DISK_UTIL_MSEC / 1000;
2126                 uint64_t nsec = (DISK_UTIL_MSEC % 1000) * 1000000;
2127                 struct timespec ts;
2128                 struct timeval tv;
2129
2130                 gettimeofday(&tv, NULL);
2131                 ts.tv_sec = tv.tv_sec + sec;
2132                 ts.tv_nsec = (tv.tv_usec * 1000) + nsec;
2133
2134                 if (ts.tv_nsec >= 1000000000ULL) {
2135                         ts.tv_nsec -= 1000000000ULL;
2136                         ts.tv_sec++;
2137                 }
2138
2139                 pthread_cond_timedwait(&helper_cond, &helper_lock, &ts);
2140
2141                 ret = update_io_ticks();
2142
2143                 if (helper_do_stat) {
2144                         helper_do_stat = 0;
2145                         __show_running_run_stats();
2146                 }
2147
2148                 if (!is_backend)
2149                         print_thread_status();
2150         }
2151
2152         return NULL;
2153 }
2154
2155 static int create_helper_thread(void)
2156 {
2157         int ret;
2158
2159         setup_disk_util();
2160
2161         pthread_cond_init(&helper_cond, NULL);
2162         pthread_mutex_init(&helper_lock, NULL);
2163
2164         ret = pthread_create(&helper_thread, NULL, helper_thread_main, NULL);
2165         if (ret) {
2166                 log_err("Can't create helper thread: %s\n", strerror(ret));
2167                 return 1;
2168         }
2169
2170         dprint(FD_MUTEX, "wait on startup_mutex\n");
2171         fio_mutex_down(startup_mutex);
2172         dprint(FD_MUTEX, "done waiting on startup_mutex\n");
2173         return 0;
2174 }
2175
2176 int fio_backend(void)
2177 {
2178         struct thread_data *td;
2179         int i;
2180
2181         if (exec_profile) {
2182                 if (load_profile(exec_profile))
2183                         return 1;
2184                 free(exec_profile);
2185                 exec_profile = NULL;
2186         }
2187         if (!thread_number)
2188                 return 0;
2189
2190         if (write_bw_log) {
2191                 struct log_params p = {
2192                         .log_type = IO_LOG_TYPE_BW,
2193                 };
2194
2195                 setup_log(&agg_io_log[DDIR_READ], &p, "agg-read_bw.log");
2196                 setup_log(&agg_io_log[DDIR_WRITE], &p, "agg-write_bw.log");
2197                 setup_log(&agg_io_log[DDIR_TRIM], &p, "agg-trim_bw.log");
2198         }
2199
2200         startup_mutex = fio_mutex_init(FIO_MUTEX_LOCKED);
2201         if (startup_mutex == NULL)
2202                 return 1;
2203
2204         set_genesis_time();
2205         stat_init();
2206         create_helper_thread();
2207
2208         cgroup_list = smalloc(sizeof(*cgroup_list));
2209         INIT_FLIST_HEAD(cgroup_list);
2210
2211         run_threads();
2212
2213         wait_for_helper_thread_exit();
2214
2215         if (!fio_abort) {
2216                 __show_run_stats();
2217                 if (write_bw_log) {
2218                         for (i = 0; i < DDIR_RWDIR_CNT; i++) {
2219                                 struct io_log *log = agg_io_log[i];
2220
2221                                 flush_log(log);
2222                                 free_log(log);
2223                         }
2224                 }
2225         }
2226
2227         for_each_td(td, i) {
2228                 fio_options_free(td);
2229                 if (td->rusage_sem) {
2230                         fio_mutex_remove(td->rusage_sem);
2231                         td->rusage_sem = NULL;
2232                 }
2233         }
2234
2235         free_disk_util();
2236         cgroup_kill(cgroup_list);
2237         sfree(cgroup_list);
2238         sfree(cgroup_mnt);
2239
2240         fio_mutex_remove(startup_mutex);
2241         stat_exit();
2242         return exit_value;
2243 }