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