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 /*
 448  * The main verify engine. Runs over the writes we previously submitted,
 449  * reads the blocks back in, and checks the crc/md5 of the data.
 450  */
 451 static void do_verify(struct thread_data *td, uint64_t verify_bytes)
 452 {
 453         struct fio_file *f;
 454         struct io_u *io_u;
 455         int ret, min_events;
 456         unsigned int i;
 457
 458         dprint(FD_VERIFY, "starting loop\n");
 459
 460         /*
 461          * sync io first and invalidate cache, to make sure we really
 462          * read from disk.
 463          */
 464         for_each_file(td, f, i) {
 465                 if (!fio_file_open(f))
 466                         continue;
 467                 if (fio_io_sync(td, f))
 468                         break;
 469                 if (file_invalidate_cache(td, f))
 470                         break;
 471         }
 472
 473         check_update_rusage(td);
 474
 475         if (td->error)
 476                 return;
 477
 478         td_set_runstate(td, TD_VERIFYING);
 479
 480         io_u = NULL;
 481         while (!td->terminate) {
 482                 enum fio_ddir ddir;
 483                 int ret2, full;
 484
 485                 update_tv_cache(td);
 486                 check_update_rusage(td);
 487
 488                 if (runtime_exceeded(td, &td->tv_cache)) {
 489                         __update_tv_cache(td);
 490                         if (runtime_exceeded(td, &td->tv_cache)) {
 491                                 fio_mark_td_terminate(td);
 492                                 break;
 493                         }
 494                 }
 495
 496                 if (flow_threshold_exceeded(td))
 497                         continue;
 498
 499                 if (!td->o.experimental_verify) {
 500                         io_u = __get_io_u(td);
 501                         if (!io_u)
 502                                 break;
 503
 504                         if (get_next_verify(td, io_u)) {
 505                                 put_io_u(td, io_u);
 506                                 break;
 507                         }
 508
 509                         if (td_io_prep(td, io_u)) {
 510                                 put_io_u(td, io_u);
 511                                 break;
 512                         }
 513                 } else {
 514                         if (ddir_rw_sum(td->bytes_done) + td->o.rw_min_bs > verify_bytes)
 515                                 break;
 516
 517                         while ((io_u = get_io_u(td)) != NULL) {
 518                                 if (IS_ERR(io_u)) {
 519                                         io_u = NULL;
 520                                         ret = FIO_Q_BUSY;
 521                                         goto reap;
 522                                 }
 523
 524                                 /*
 525                                  * We are only interested in the places where
 526                                  * we wrote or trimmed IOs. Turn those into
 527                                  * reads for verification purposes.
 528                                  */
 529                                 if (io_u->ddir == DDIR_READ) {
 530                                         /*
 531                                          * Pretend we issued it for rwmix
 532                                          * accounting
 533                                          */
 534                                         td->io_issues[DDIR_READ]++;
 535                                         put_io_u(td, io_u);
 536                                         continue;
 537                                 } else if (io_u->ddir == DDIR_TRIM) {
 538                                         io_u->ddir = DDIR_READ;
 539                                         io_u->flags |= IO_U_F_TRIMMED;
 540                                         break;
 541                                 } else if (io_u->ddir == DDIR_WRITE) {
 542                                         io_u->ddir = DDIR_READ;
 543                                         break;
 544                                 } else {
 545                                         put_io_u(td, io_u);
 546                                         continue;
 547                                 }
 548                         }
 549
 550                         if (!io_u)
 551                                 break;
 552                 }
 553
 554                 if (verify_state_should_stop(td, io_u)) {
 555                         put_io_u(td, io_u);
 556                         break;
 557                 }
 558
 559                 if (td->o.verify_async)
 560                         io_u->end_io = verify_io_u_async;
 561                 else
 562                         io_u->end_io = verify_io_u;
 563
 564                 ddir = io_u->ddir;
 565                 if (!td->o.disable_slat)
 566                         fio_gettime(&io_u->start_time, NULL);
 567
 568                 ret = td_io_queue(td, io_u);
 569                 switch (ret) {
 570                 case FIO_Q_COMPLETED:
 571                         if (io_u->error) {
 572                                 ret = -io_u->error;
 573                                 clear_io_u(td, io_u);
 574                         } else if (io_u->resid) {
 575                                 int bytes = io_u->xfer_buflen - io_u->resid;
 576
 577                                 /*
 578                                  * zero read, fail
 579                                  */
 580                                 if (!bytes) {
 581                                         td_verror(td, EIO, "full resid");
 582                                         put_io_u(td, io_u);
 583                                         break;
 584                                 }
 585
 586                                 io_u->xfer_buflen = io_u->resid;
 587                                 io_u->xfer_buf += bytes;
 588                                 io_u->offset += bytes;
 589
 590                                 if (ddir_rw(io_u->ddir))
 591                                         td->ts.short_io_u[io_u->ddir]++;
 592
 593                                 f = io_u->file;
 594                                 if (io_u->offset == f->real_file_size)
 595                                         goto sync_done;
 596
 597                                 requeue_io_u(td, &io_u);
 598                         } else {
 599 sync_done:
 600                                 ret = io_u_sync_complete(td, io_u);
 601                                 if (ret < 0)
 602                                         break;
 603                         }
 604                         continue;
 605                 case FIO_Q_QUEUED:
 606                         break;
 607                 case FIO_Q_BUSY:
 608                         requeue_io_u(td, &io_u);
 609                         ret2 = td_io_commit(td);
 610                         if (ret2 < 0)
 611                                 ret = ret2;
 612                         break;
 613                 default:
 614                         assert(ret < 0);
 615                         td_verror(td, -ret, "td_io_queue");
 616                         break;
 617                 }
 618
 619                 if (break_on_this_error(td, ddir, &ret))
 620                         break;
 621
 622                 /*
 623                  * if we can queue more, do so. but check if there are
 624                  * completed io_u's first. Note that we can get BUSY even
 625                  * without IO queued, if the system is resource starved.
 626                  */
 627 reap:
 628                 full = queue_full(td) || (ret == FIO_Q_BUSY && td->cur_depth);
 629                 if (full || !td->o.iodepth_batch_complete)
 630                         ret = wait_for_completions(td, NULL);
 631
 632                 if (ret < 0)
 633                         break;
 634         }
 635
 636         check_update_rusage(td);
 637
 638         if (!td->error) {
 639                 min_events = td->cur_depth;
 640
 641                 if (min_events)
 642                         ret = io_u_queued_complete(td, min_events);
 643         } else
 644                 cleanup_pending_aio(td);
 645
 646         td_set_runstate(td, TD_RUNNING);
 647
 648         dprint(FD_VERIFY, "exiting loop\n");
 649 }
 650
 651 static unsigned int exceeds_number_ios(struct thread_data *td)
 652 {
 653         unsigned long long number_ios;
 654
 655         if (!td->o.number_ios)
 656                 return 0;
 657
 658         number_ios = ddir_rw_sum(td->io_blocks);
 659         number_ios += td->io_u_queued + td->io_u_in_flight;
 660
 661         return number_ios >= (td->o.number_ios * td->loops);
 662 }
 663
 664 static int io_issue_bytes_exceeded(struct thread_data *td)
 665 {
 666         unsigned long long bytes, limit;
 667
 668         if (td_rw(td))
 669                 bytes = td->io_issue_bytes[DDIR_READ] + td->io_issue_bytes[DDIR_WRITE];
 670         else if (td_write(td))
 671                 bytes = td->io_issue_bytes[DDIR_WRITE];
 672         else if (td_read(td))
 673                 bytes = td->io_issue_bytes[DDIR_READ];
 674         else
 675                 bytes = td->io_issue_bytes[DDIR_TRIM];
 676
 677         if (td->o.io_limit)
 678                 limit = td->o.io_limit;
 679         else
 680                 limit = td->o.size;
 681
 682         limit *= td->loops;
 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         limit *= td->loops;
 705         return bytes >= limit || exceeds_number_ios(td);
 706 }
 707
 708 /*
 709  * Main IO worker function. It retrieves io_u's to process and queues
 710  * and reaps them, checking for rate and errors along the way.
 711  *
 712  * Returns number of bytes written and trimmed.
 713  */
 714 static uint64_t do_io(struct thread_data *td)
 715 {
 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         total_bytes = td->o.size;
 728         /*
 729         * Allow random overwrite workloads to write up to io_limit
 730         * before starting verification phase as 'size' doesn't apply.
 731         */
 732         if (td_write(td) && td_random(td) && td->o.norandommap)
 733                 total_bytes = max(total_bytes, (uint64_t) td->o.io_limit);
 734         /*
 735          * If verify_backlog is enabled, we'll run the verify in this
 736          * handler as well. For that case, we may need up to twice the
 737          * amount of bytes.
 738          */
 739         if (td->o.verify != VERIFY_NONE &&
 740            (td_write(td) && td->o.verify_backlog))
 741                 total_bytes += td->o.size;
 742
 743         /* In trimwrite mode, each byte is trimmed and then written, so
 744          * allow total_bytes to be twice as big */
 745         if (td_trimwrite(td))
 746                 total_bytes += td->total_io_size;
 747
 748         while ((td->o.read_iolog_file && !flist_empty(&td->io_log_list)) ||
 749                 (!flist_empty(&td->trim_list)) || !io_issue_bytes_exceeded(td) ||
 750                 td->o.time_based) {
 751                 struct timeval comp_time;
 752                 struct io_u *io_u;
 753                 int ret2, full;
 754                 enum fio_ddir ddir;
 755
 756                 check_update_rusage(td);
 757
 758                 if (td->terminate || td->done)
 759                         break;
 760
 761                 update_tv_cache(td);
 762
 763                 if (runtime_exceeded(td, &td->tv_cache)) {
 764                         __update_tv_cache(td);
 765                         if (runtime_exceeded(td, &td->tv_cache)) {
 766                                 fio_mark_td_terminate(td);
 767                                 break;
 768                         }
 769                 }
 770
 771                 if (flow_threshold_exceeded(td))
 772                         continue;
 773
 774                 if (bytes_issued >= total_bytes)
 775                         break;
 776
 777                 io_u = get_io_u(td);
 778                 if (IS_ERR_OR_NULL(io_u)) {
 779                         int err = PTR_ERR(io_u);
 780
 781                         io_u = NULL;
 782                         if (err == -EBUSY) {
 783                                 ret = FIO_Q_BUSY;
 784                                 goto reap;
 785                         }
 786                         if (td->o.latency_target)
 787                                 goto reap;
 788                         break;
 789                 }
 790
 791                 ddir = io_u->ddir;
 792
 793                 /*
 794                  * Add verification end_io handler if:
 795                  *      - Asked to verify (!td_rw(td))
 796                  *      - Or the io_u is from our verify list (mixed write/ver)
 797                  */
 798                 if (td->o.verify != VERIFY_NONE && io_u->ddir == DDIR_READ &&
 799                     ((io_u->flags & IO_U_F_VER_LIST) || !td_rw(td))) {
 800
 801                         if (!td->o.verify_pattern_bytes) {
 802                                 io_u->rand_seed = __rand(&td->verify_state);
 803                                 if (sizeof(int) != sizeof(long *))
 804                                         io_u->rand_seed *= __rand(&td->verify_state);
 805                         }
 806
 807                         if (verify_state_should_stop(td, io_u)) {
 808                                 put_io_u(td, io_u);
 809                                 break;
 810                         }
 811
 812                         if (td->o.verify_async)
 813                                 io_u->end_io = verify_io_u_async;
 814                         else
 815                                 io_u->end_io = verify_io_u;
 816                         td_set_runstate(td, TD_VERIFYING);
 817                 } else if (in_ramp_time(td))
 818                         td_set_runstate(td, TD_RAMP);
 819                 else
 820                         td_set_runstate(td, TD_RUNNING);
 821
 822                 /*
 823                  * Always log IO before it's issued, so we know the specific
 824                  * order of it. The logged unit will track when the IO has
 825                  * completed.
 826                  */
 827                 if (td_write(td) && io_u->ddir == DDIR_WRITE &&
 828                     td->o.do_verify &&
 829                     td->o.verify != VERIFY_NONE &&
 830                     !td->o.experimental_verify)
 831                         log_io_piece(td, io_u);
 832
 833                 ret = td_io_queue(td, io_u);
 834                 switch (ret) {
 835                 case FIO_Q_COMPLETED:
 836                         if (io_u->error) {
 837                                 ret = -io_u->error;
 838                                 unlog_io_piece(td, io_u);
 839                                 clear_io_u(td, io_u);
 840                         } else if (io_u->resid) {
 841                                 int bytes = io_u->xfer_buflen - io_u->resid;
 842                                 struct fio_file *f = io_u->file;
 843
 844                                 bytes_issued += bytes;
 845
 846                                 trim_io_piece(td, io_u);
 847
 848                                 /*
 849                                  * zero read, fail
 850                                  */
 851                                 if (!bytes) {
 852                                         unlog_io_piece(td, io_u);
 853                                         td_verror(td, EIO, "full resid");
 854                                         put_io_u(td, io_u);
 855                                         break;
 856                                 }
 857
 858                                 io_u->xfer_buflen = io_u->resid;
 859                                 io_u->xfer_buf += bytes;
 860                                 io_u->offset += bytes;
 861
 862                                 if (ddir_rw(io_u->ddir))
 863                                         td->ts.short_io_u[io_u->ddir]++;
 864
 865                                 if (io_u->offset == f->real_file_size)
 866                                         goto sync_done;
 867
 868                                 requeue_io_u(td, &io_u);
 869                         } else {
 870 sync_done:
 871                                 if (__should_check_rate(td, DDIR_READ) ||
 872                                     __should_check_rate(td, DDIR_WRITE) ||
 873                                     __should_check_rate(td, DDIR_TRIM))
 874                                         fio_gettime(&comp_time, NULL);
 875
 876                                 ret = io_u_sync_complete(td, io_u);
 877                                 if (ret < 0)
 878                                         break;
 879                                 bytes_issued += io_u->xfer_buflen;
 880                         }
 881                         break;
 882                 case FIO_Q_QUEUED:
 883                         /*
 884                          * if the engine doesn't have a commit hook,
 885                          * the io_u is really queued. if it does have such
 886                          * a hook, it has to call io_u_queued() itself.
 887                          */
 888                         if (td->io_ops->commit == NULL)
 889                                 io_u_queued(td, io_u);
 890                         bytes_issued += io_u->xfer_buflen;
 891                         break;
 892                 case FIO_Q_BUSY:
 893                         unlog_io_piece(td, io_u);
 894                         requeue_io_u(td, &io_u);
 895                         ret2 = td_io_commit(td);
 896                         if (ret2 < 0)
 897                                 ret = ret2;
 898                         break;
 899                 default:
 900                         assert(ret < 0);
 901                         put_io_u(td, io_u);
 902                         break;
 903                 }
 904
 905                 if (break_on_this_error(td, ddir, &ret))
 906                         break;
 907
 908                 /*
 909                  * See if we need to complete some commands. Note that we
 910                  * can get BUSY even without IO queued, if the system is
 911                  * resource starved.
 912                  */
 913 reap:
 914                 full = queue_full(td) || (ret == FIO_Q_BUSY && td->cur_depth);
 915                 if (full || !td->o.iodepth_batch_complete)
 916                         ret = wait_for_completions(td, &comp_time);
 917                 if (ret < 0)
 918                         break;
 919                 if (!ddir_rw_sum(td->bytes_done) &&
 920                     !(td->io_ops->flags & FIO_NOIO))
 921                         continue;
 922
 923                 if (!in_ramp_time(td) && should_check_rate(td)) {
 924                         if (check_min_rate(td, &comp_time)) {
 925                                 if (exitall_on_terminate)
 926                                         fio_terminate_threads(td->groupid);
 927                                 td_verror(td, EIO, "check_min_rate");
 928                                 break;
 929                         }
 930                 }
 931                 if (!in_ramp_time(td) && td->o.latency_target)
 932                         lat_target_check(td);
 933
 934                 if (td->o.thinktime) {
 935                         unsigned long long b;
 936
 937                         b = ddir_rw_sum(td->io_blocks);
 938                         if (!(b % td->o.thinktime_blocks)) {
 939                                 int left;
 940
 941                                 io_u_quiesce(td);
 942
 943                                 if (td->o.thinktime_spin)
 944                                         usec_spin(td->o.thinktime_spin);
 945
 946                                 left = td->o.thinktime - td->o.thinktime_spin;
 947                                 if (left)
 948                                         usec_sleep(td, left);
 949                         }
 950                 }
 951         }
 952
 953         check_update_rusage(td);
 954
 955         if (td->trim_entries)
 956                 log_err("fio: %lu trim entries leaked?\n", td->trim_entries);
 957
 958         if (td->o.fill_device && td->error == ENOSPC) {
 959                 td->error = 0;
 960                 fio_mark_td_terminate(td);
 961         }
 962         if (!td->error) {
 963                 struct fio_file *f;
 964
 965                 i = td->cur_depth;
 966                 if (i) {
 967                         ret = io_u_queued_complete(td, i);
 968                         if (td->o.fill_device && td->error == ENOSPC)
 969                                 td->error = 0;
 970                 }
 971
 972                 if (should_fsync(td) && td->o.end_fsync) {
 973                         td_set_runstate(td, TD_FSYNCING);
 974
 975                         for_each_file(td, f, i) {
 976                                 if (!fio_file_fsync(td, f))
 977                                         continue;
 978
 979                                 log_err("fio: end_fsync failed for file %s\n",
 980                                                                 f->file_name);
 981                         }
 982                 }
 983         } else
 984                 cleanup_pending_aio(td);
 985
 986         /*
 987          * stop job if we failed doing any IO
 988          */
 989         if (!ddir_rw_sum(td->this_io_bytes))
 990                 td->done = 1;
 991
 992         return td->bytes_done[DDIR_WRITE] + td->bytes_done[DDIR_TRIM];
 993 }
 994
 995 static void cleanup_io_u(struct thread_data *td)
 996 {
 997         struct io_u *io_u;
 998
 999         while ((io_u = io_u_qpop(&td->io_u_freelist)) != NULL) {
1000
1001                 if (td->io_ops->io_u_free)
1002                         td->io_ops->io_u_free(td, io_u);
1003
1004                 fio_memfree(io_u, sizeof(*io_u));
1005         }
1006
1007         free_io_mem(td);
1008
1009         io_u_rexit(&td->io_u_requeues);
1010         io_u_qexit(&td->io_u_freelist);
1011         io_u_qexit(&td->io_u_all);
1012
1013         if (td->last_write_comp)
1014                 sfree(td->last_write_comp);
1015 }
1016
1017 static int init_io_u(struct thread_data *td)
1018 {
1019         struct io_u *io_u;
1020         unsigned int max_bs, min_write;
1021         int cl_align, i, max_units;
1022         int data_xfer = 1, err;
1023         char *p;
1024
1025         max_units = td->o.iodepth;
1026         max_bs = td_max_bs(td);
1027         min_write = td->o.min_bs[DDIR_WRITE];
1028         td->orig_buffer_size = (unsigned long long) max_bs
1029                                         * (unsigned long long) max_units;
1030
1031         if ((td->io_ops->flags & FIO_NOIO) || !(td_read(td) || td_write(td)))
1032                 data_xfer = 0;
1033
1034         err = 0;
1035         err += io_u_rinit(&td->io_u_requeues, td->o.iodepth);
1036         err += io_u_qinit(&td->io_u_freelist, td->o.iodepth);
1037         err += io_u_qinit(&td->io_u_all, td->o.iodepth);
1038
1039         if (err) {
1040                 log_err("fio: failed setting up IO queues\n");
1041                 return 1;
1042         }
1043
1044         /*
1045          * if we may later need to do address alignment, then add any
1046          * possible adjustment here so that we don't cause a buffer
1047          * overflow later. this adjustment may be too much if we get
1048          * lucky and the allocator gives us an aligned address.
1049          */
1050         if (td->o.odirect || td->o.mem_align || td->o.oatomic ||
1051             (td->io_ops->flags & FIO_RAWIO))
1052                 td->orig_buffer_size += page_mask + td->o.mem_align;
1053
1054         if (td->o.mem_type == MEM_SHMHUGE || td->o.mem_type == MEM_MMAPHUGE) {
1055                 unsigned long bs;
1056
1057                 bs = td->orig_buffer_size + td->o.hugepage_size - 1;
1058                 td->orig_buffer_size = bs & ~(td->o.hugepage_size - 1);
1059         }
1060
1061         if (td->orig_buffer_size != (size_t) td->orig_buffer_size) {
1062                 log_err("fio: IO memory too large. Reduce max_bs or iodepth\n");
1063                 return 1;
1064         }
1065
1066         if (data_xfer && allocate_io_mem(td))
1067                 return 1;
1068
1069         if (td->o.odirect || td->o.mem_align || td->o.oatomic ||
1070             (td->io_ops->flags & FIO_RAWIO))
1071                 p = PAGE_ALIGN(td->orig_buffer) + td->o.mem_align;
1072         else
1073                 p = td->orig_buffer;
1074
1075         cl_align = os_cache_line_size();
1076
1077         for (i = 0; i < max_units; i++) {
1078                 void *ptr;
1079
1080                 if (td->terminate)
1081                         return 1;
1082
1083                 ptr = fio_memalign(cl_align, sizeof(*io_u));
1084                 if (!ptr) {
1085                         log_err("fio: unable to allocate aligned memory\n");
1086                         break;
1087                 }
1088
1089                 io_u = ptr;
1090                 memset(io_u, 0, sizeof(*io_u));
1091                 INIT_FLIST_HEAD(&io_u->verify_list);
1092                 dprint(FD_MEM, "io_u alloc %p, index %u\n", io_u, i);
1093
1094                 if (data_xfer) {
1095                         io_u->buf = p;
1096                         dprint(FD_MEM, "io_u %p, mem %p\n", io_u, io_u->buf);
1097
1098                         if (td_write(td))
1099                                 io_u_fill_buffer(td, io_u, min_write, max_bs);
1100                         if (td_write(td) && td->o.verify_pattern_bytes) {
1101                                 /*
1102                                  * Fill the buffer with the pattern if we are
1103                                  * going to be doing writes.
1104                                  */
1105                                 fill_verify_pattern(td, io_u->buf, max_bs, io_u, 0, 0);
1106                         }
1107                 }
1108
1109                 io_u->index = i;
1110                 io_u->flags = IO_U_F_FREE;
1111                 io_u_qpush(&td->io_u_freelist, io_u);
1112
1113                 /*
1114                  * io_u never leaves this stack, used for iteration of all
1115                  * io_u buffers.
1116                  */
1117                 io_u_qpush(&td->io_u_all, io_u);
1118
1119                 if (td->io_ops->io_u_init) {
1120                         int ret = td->io_ops->io_u_init(td, io_u);
1121
1122                         if (ret) {
1123                                 log_err("fio: failed to init engine data: %d\n", ret);
1124                                 return 1;
1125                         }
1126                 }
1127
1128                 p += max_bs;
1129         }
1130
1131         if (td->o.verify != VERIFY_NONE) {
1132                 td->last_write_comp = scalloc(max_units, sizeof(uint64_t));
1133                 if (!td->last_write_comp) {
1134                         log_err("fio: failed to alloc write comp data\n");
1135                         return 1;
1136                 }
1137         }
1138
1139         return 0;
1140 }
1141
1142 static int switch_ioscheduler(struct thread_data *td)
1143 {
1144         char tmp[256], tmp2[128];
1145         FILE *f;
1146         int ret;
1147
1148         if (td->io_ops->flags & FIO_DISKLESSIO)
1149                 return 0;
1150
1151         sprintf(tmp, "%s/queue/scheduler", td->sysfs_root);
1152
1153         f = fopen(tmp, "r+");
1154         if (!f) {
1155                 if (errno == ENOENT) {
1156                         log_err("fio: os or kernel doesn't support IO scheduler"
1157                                 " switching\n");
1158                         return 0;
1159                 }
1160                 td_verror(td, errno, "fopen iosched");
1161                 return 1;
1162         }
1163
1164         /*
1165          * Set io scheduler.
1166          */
1167         ret = fwrite(td->o.ioscheduler, strlen(td->o.ioscheduler), 1, f);
1168         if (ferror(f) || ret != 1) {
1169                 td_verror(td, errno, "fwrite");
1170                 fclose(f);
1171                 return 1;
1172         }
1173
1174         rewind(f);
1175
1176         /*
1177          * Read back and check that the selected scheduler is now the default.
1178          */
1179         ret = fread(tmp, sizeof(tmp), 1, f);
1180         if (ferror(f) || ret < 0) {
1181                 td_verror(td, errno, "fread");
1182                 fclose(f);
1183                 return 1;
1184         }
1185         tmp[sizeof(tmp) - 1] = '\0';
1186
1187
1188         sprintf(tmp2, "[%s]", td->o.ioscheduler);
1189         if (!strstr(tmp, tmp2)) {
1190                 log_err("fio: io scheduler %s not found\n", td->o.ioscheduler);
1191                 td_verror(td, EINVAL, "iosched_switch");
1192                 fclose(f);
1193                 return 1;
1194         }
1195
1196         fclose(f);
1197         return 0;
1198 }
1199
1200 static int keep_running(struct thread_data *td)
1201 {
1202         unsigned long long limit;
1203
1204         if (td->done)
1205                 return 0;
1206         if (td->o.time_based)
1207                 return 1;
1208         if (td->o.loops) {
1209                 td->o.loops--;
1210                 return 1;
1211         }
1212         if (exceeds_number_ios(td))
1213                 return 0;
1214
1215         if (td->o.io_limit)
1216                 limit = td->o.io_limit;
1217         else
1218                 limit = td->o.size;
1219
1220         if (limit != -1ULL && ddir_rw_sum(td->io_bytes) < limit) {
1221                 uint64_t diff;
1222
1223                 /*
1224                  * If the difference is less than the minimum IO size, we
1225                  * are done.
1226                  */
1227                 diff = limit - ddir_rw_sum(td->io_bytes);
1228                 if (diff < td_max_bs(td))
1229                         return 0;
1230
1231                 if (fio_files_done(td))
1232                         return 0;
1233
1234                 return 1;
1235         }
1236
1237         return 0;
1238 }
1239
1240 static int exec_string(struct thread_options *o, const char *string, const char *mode)
1241 {
1242         int ret, newlen = strlen(string) + strlen(o->name) + strlen(mode) + 9 + 1;
1243         char *str;
1244
1245         str = malloc(newlen);
1246         sprintf(str, "%s &> %s.%s.txt", string, o->name, mode);
1247
1248         log_info("%s : Saving output of %s in %s.%s.txt\n",o->name, mode, o->name, mode);
1249         ret = system(str);
1250         if (ret == -1)
1251                 log_err("fio: exec of cmd <%s> failed\n", str);
1252
1253         free(str);
1254         return ret;
1255 }
1256
1257 /*
1258  * Dry run to compute correct state of numberio for verification.
1259  */
1260 static uint64_t do_dry_run(struct thread_data *td)
1261 {
1262         td_set_runstate(td, TD_RUNNING);
1263
1264         while ((td->o.read_iolog_file && !flist_empty(&td->io_log_list)) ||
1265                 (!flist_empty(&td->trim_list)) || !io_complete_bytes_exceeded(td)) {
1266                 struct io_u *io_u;
1267                 int ret;
1268
1269                 if (td->terminate || td->done)
1270                         break;
1271
1272                 io_u = get_io_u(td);
1273                 if (!io_u)
1274                         break;
1275
1276                 io_u->flags |= IO_U_F_FLIGHT;
1277                 io_u->error = 0;
1278                 io_u->resid = 0;
1279                 if (ddir_rw(acct_ddir(io_u)))
1280                         td->io_issues[acct_ddir(io_u)]++;
1281                 if (ddir_rw(io_u->ddir)) {
1282                         io_u_mark_depth(td, 1);
1283                         td->ts.total_io_u[io_u->ddir]++;
1284                 }
1285
1286                 if (td_write(td) && io_u->ddir == DDIR_WRITE &&
1287                     td->o.do_verify &&
1288                     td->o.verify != VERIFY_NONE &&
1289                     !td->o.experimental_verify)
1290                         log_io_piece(td, io_u);
1291
1292                 ret = io_u_sync_complete(td, io_u);
1293                 (void) ret;
1294         }
1295
1296         return td->bytes_done[DDIR_WRITE] + td->bytes_done[DDIR_TRIM];
1297 }
1298
1299 /*
1300  * Entry point for the thread based jobs. The process based jobs end up
1301  * here as well, after a little setup.
1302  */
1303 static void *thread_main(void *data)
1304 {
1305         unsigned long long elapsed;
1306         struct thread_data *td = data;
1307         struct thread_options *o = &td->o;
1308         pthread_condattr_t attr;
1309         int clear_state;
1310         int ret;
1311
1312         if (!o->use_thread) {
1313                 setsid();
1314                 td->pid = getpid();
1315         } else
1316                 td->pid = gettid();
1317
1318         fio_local_clock_init(o->use_thread);
1319
1320         dprint(FD_PROCESS, "jobs pid=%d started\n", (int) td->pid);
1321
1322         if (is_backend)
1323                 fio_server_send_start(td);
1324
1325         INIT_FLIST_HEAD(&td->io_log_list);
1326         INIT_FLIST_HEAD(&td->io_hist_list);
1327         INIT_FLIST_HEAD(&td->verify_list);
1328         INIT_FLIST_HEAD(&td->trim_list);
1329         INIT_FLIST_HEAD(&td->next_rand_list);
1330         pthread_mutex_init(&td->io_u_lock, NULL);
1331         td->io_hist_tree = RB_ROOT;
1332
1333         pthread_condattr_init(&attr);
1334         pthread_cond_init(&td->verify_cond, &attr);
1335         pthread_cond_init(&td->free_cond, &attr);
1336
1337         td_set_runstate(td, TD_INITIALIZED);
1338         dprint(FD_MUTEX, "up startup_mutex\n");
1339         fio_mutex_up(startup_mutex);
1340         dprint(FD_MUTEX, "wait on td->mutex\n");
1341         fio_mutex_down(td->mutex);
1342         dprint(FD_MUTEX, "done waiting on td->mutex\n");
1343
1344         /*
1345          * A new gid requires privilege, so we need to do this before setting
1346          * the uid.
1347          */
1348         if (o->gid != -1U && setgid(o->gid)) {
1349                 td_verror(td, errno, "setgid");
1350                 goto err;
1351         }
1352         if (o->uid != -1U && setuid(o->uid)) {
1353                 td_verror(td, errno, "setuid");
1354                 goto err;
1355         }
1356
1357         /*
1358          * If we have a gettimeofday() thread, make sure we exclude that
1359          * thread from this job
1360          */
1361         if (o->gtod_cpu)
1362                 fio_cpu_clear(&o->cpumask, o->gtod_cpu);
1363
1364         /*
1365          * Set affinity first, in case it has an impact on the memory
1366          * allocations.
1367          */
1368         if (fio_option_is_set(o, cpumask)) {
1369                 if (o->cpus_allowed_policy == FIO_CPUS_SPLIT) {
1370                         ret = fio_cpus_split(&o->cpumask, td->thread_number - 1);
1371                         if (!ret) {
1372                                 log_err("fio: no CPUs set\n");
1373                                 log_err("fio: Try increasing number of available CPUs\n");
1374                                 td_verror(td, EINVAL, "cpus_split");
1375                                 goto err;
1376                         }
1377                 }
1378                 ret = fio_setaffinity(td->pid, o->cpumask);
1379                 if (ret == -1) {
1380                         td_verror(td, errno, "cpu_set_affinity");
1381                         goto err;
1382                 }
1383         }
1384
1385 #ifdef CONFIG_LIBNUMA
1386         /* numa node setup */
1387         if (fio_option_is_set(o, numa_cpunodes) ||
1388             fio_option_is_set(o, numa_memnodes)) {
1389                 struct bitmask *mask;
1390
1391                 if (numa_available() < 0) {
1392                         td_verror(td, errno, "Does not support NUMA API\n");
1393                         goto err;
1394                 }
1395
1396                 if (fio_option_is_set(o, numa_cpunodes)) {
1397                         mask = numa_parse_nodestring(o->numa_cpunodes);
1398                         ret = numa_run_on_node_mask(mask);
1399                         numa_free_nodemask(mask);
1400                         if (ret == -1) {
1401                                 td_verror(td, errno, \
1402                                         "numa_run_on_node_mask failed\n");
1403                                 goto err;
1404                         }
1405                 }
1406
1407                 if (fio_option_is_set(o, numa_memnodes)) {
1408                         mask = NULL;
1409                         if (o->numa_memnodes)
1410                                 mask = numa_parse_nodestring(o->numa_memnodes);
1411
1412                         switch (o->numa_mem_mode) {
1413                         case MPOL_INTERLEAVE:
1414                                 numa_set_interleave_mask(mask);
1415                                 break;
1416                         case MPOL_BIND:
1417                                 numa_set_membind(mask);
1418                                 break;
1419                         case MPOL_LOCAL:
1420                                 numa_set_localalloc();
1421                                 break;
1422                         case MPOL_PREFERRED:
1423                                 numa_set_preferred(o->numa_mem_prefer_node);
1424                                 break;
1425                         case MPOL_DEFAULT:
1426                         default:
1427                                 break;
1428                         }
1429
1430                         if (mask)
1431                                 numa_free_nodemask(mask);
1432
1433                 }
1434         }
1435 #endif
1436
1437         if (fio_pin_memory(td))
1438                 goto err;
1439
1440         /*
1441          * May alter parameters that init_io_u() will use, so we need to
1442          * do this first.
1443          */
1444         if (init_iolog(td))
1445                 goto err;
1446
1447         if (init_io_u(td))
1448                 goto err;
1449
1450         if (o->verify_async && verify_async_init(td))
1451                 goto err;
1452
1453         if (fio_option_is_set(o, ioprio) ||
1454             fio_option_is_set(o, ioprio_class)) {
1455                 ret = ioprio_set(IOPRIO_WHO_PROCESS, 0, o->ioprio_class, o->ioprio);
1456                 if (ret == -1) {
1457                         td_verror(td, errno, "ioprio_set");
1458                         goto err;
1459                 }
1460         }
1461
1462         if (o->cgroup && cgroup_setup(td, cgroup_list, &cgroup_mnt))
1463                 goto err;
1464
1465         errno = 0;
1466         if (nice(o->nice) == -1 && errno != 0) {
1467                 td_verror(td, errno, "nice");
1468                 goto err;
1469         }
1470
1471         if (o->ioscheduler && switch_ioscheduler(td))
1472                 goto err;
1473
1474         if (!o->create_serialize && setup_files(td))
1475                 goto err;
1476
1477         if (td_io_init(td))
1478                 goto err;
1479
1480         if (init_random_map(td))
1481                 goto err;
1482
1483         if (o->exec_prerun && exec_string(o, o->exec_prerun, (const char *)"prerun"))
1484                 goto err;
1485
1486         if (o->pre_read) {
1487                 if (pre_read_files(td) < 0)
1488                         goto err;
1489         }
1490
1491         if (td->flags & TD_F_COMPRESS_LOG)
1492                 tp_init(&td->tp_data);
1493
1494         fio_verify_init(td);
1495
1496         fio_gettime(&td->epoch, NULL);
1497         fio_getrusage(&td->ru_start);
1498         clear_state = 0;
1499         while (keep_running(td)) {
1500                 uint64_t verify_bytes;
1501
1502                 fio_gettime(&td->start, NULL);
1503                 memcpy(&td->bw_sample_time, &td->start, sizeof(td->start));
1504                 memcpy(&td->iops_sample_time, &td->start, sizeof(td->start));
1505                 memcpy(&td->tv_cache, &td->start, sizeof(td->start));
1506
1507                 if (o->ratemin[DDIR_READ] || o->ratemin[DDIR_WRITE] ||
1508                                 o->ratemin[DDIR_TRIM]) {
1509                         memcpy(&td->lastrate[DDIR_READ], &td->bw_sample_time,
1510                                                 sizeof(td->bw_sample_time));
1511                         memcpy(&td->lastrate[DDIR_WRITE], &td->bw_sample_time,
1512                                                 sizeof(td->bw_sample_time));
1513                         memcpy(&td->lastrate[DDIR_TRIM], &td->bw_sample_time,
1514                                                 sizeof(td->bw_sample_time));
1515                 }
1516
1517                 if (clear_state)
1518                         clear_io_state(td);
1519
1520                 prune_io_piece_log(td);
1521
1522                 if (td->o.verify_only && (td_write(td) || td_rw(td)))
1523                         verify_bytes = do_dry_run(td);
1524                 else
1525                         verify_bytes = do_io(td);
1526
1527                 clear_state = 1;
1528
1529                 /*
1530                  * Make sure we've successfully updated the rusage stats
1531                  * before waiting on the stat mutex. Otherwise we could have
1532                  * the stat thread holding stat mutex and waiting for
1533                  * the rusage_sem, which would never get upped because
1534                  * this thread is waiting for the stat mutex.
1535                  */
1536                 check_update_rusage(td);
1537
1538                 fio_mutex_down(stat_mutex);
1539                 if (td_read(td) && td->io_bytes[DDIR_READ]) {
1540                         elapsed = mtime_since_now(&td->start);
1541                         td->ts.runtime[DDIR_READ] += elapsed;
1542                 }
1543                 if (td_write(td) && td->io_bytes[DDIR_WRITE]) {
1544                         elapsed = mtime_since_now(&td->start);
1545                         td->ts.runtime[DDIR_WRITE] += elapsed;
1546                 }
1547                 if (td_trim(td) && td->io_bytes[DDIR_TRIM]) {
1548                         elapsed = mtime_since_now(&td->start);
1549                         td->ts.runtime[DDIR_TRIM] += elapsed;
1550                 }
1551                 fio_gettime(&td->start, NULL);
1552                 fio_mutex_up(stat_mutex);
1553
1554                 if (td->error || td->terminate)
1555                         break;
1556
1557                 if (!o->do_verify ||
1558                     o->verify == VERIFY_NONE ||
1559                     (td->io_ops->flags & FIO_UNIDIR))
1560                         continue;
1561
1562                 clear_io_state(td);
1563
1564                 fio_gettime(&td->start, NULL);
1565
1566                 do_verify(td, verify_bytes);
1567
1568                 /*
1569                  * See comment further up for why this is done here.
1570                  */
1571                 check_update_rusage(td);
1572
1573                 fio_mutex_down(stat_mutex);
1574                 td->ts.runtime[DDIR_READ] += mtime_since_now(&td->start);
1575                 fio_gettime(&td->start, NULL);
1576                 fio_mutex_up(stat_mutex);
1577
1578                 if (td->error || td->terminate)
1579                         break;
1580         }
1581
1582         update_rusage_stat(td);
1583         td->ts.total_run_time = mtime_since_now(&td->epoch);
1584         td->ts.io_bytes[DDIR_READ] = td->io_bytes[DDIR_READ];
1585         td->ts.io_bytes[DDIR_WRITE] = td->io_bytes[DDIR_WRITE];
1586         td->ts.io_bytes[DDIR_TRIM] = td->io_bytes[DDIR_TRIM];
1587
1588         if (td->o.verify_state_save && !(td->flags & TD_F_VSTATE_SAVED) &&
1589             (td->o.verify != VERIFY_NONE && td_write(td))) {
1590                 struct all_io_list *state;
1591                 size_t sz;
1592
1593                 state = get_all_io_list(td->thread_number, &sz);
1594                 if (state) {
1595                         __verify_save_state(state, "local");
1596                         free(state);
1597                 }
1598         }
1599
1600         fio_unpin_memory(td);
1601
1602         fio_writeout_logs(td);
1603
1604         if (td->flags & TD_F_COMPRESS_LOG)
1605                 tp_exit(&td->tp_data);
1606
1607         if (o->exec_postrun)
1608                 exec_string(o, o->exec_postrun, (const char *)"postrun");
1609
1610         if (exitall_on_terminate)
1611                 fio_terminate_threads(td->groupid);
1612
1613 err:
1614         if (td->error)
1615                 log_info("fio: pid=%d, err=%d/%s\n", (int) td->pid, td->error,
1616                                                         td->verror);
1617
1618         if (o->verify_async)
1619                 verify_async_exit(td);
1620
1621         close_and_free_files(td);
1622         cleanup_io_u(td);
1623         close_ioengine(td);
1624         cgroup_shutdown(td, &cgroup_mnt);
1625         verify_free_state(td);
1626
1627         if (fio_option_is_set(o, cpumask)) {
1628                 ret = fio_cpuset_exit(&o->cpumask);
1629                 if (ret)
1630                         td_verror(td, ret, "fio_cpuset_exit");
1631         }
1632
1633         /*
1634          * do this very late, it will log file closing as well
1635          */
1636         if (o->write_iolog_file)
1637                 write_iolog_close(td);
1638
1639         fio_mutex_remove(td->mutex);
1640         td->mutex = NULL;
1641
1642         td_set_runstate(td, TD_EXITED);
1643
1644         /*
1645          * Do this last after setting our runstate to exited, so we
1646          * know that the stat thread is signaled.
1647          */
1648         check_update_rusage(td);
1649
1650         return (void *) (uintptr_t) td->error;
1651 }
1652
1653
1654 /*
1655  * We cannot pass the td data into a forked process, so attach the td and
1656  * pass it to the thread worker.
1657  */
1658 static int fork_main(int shmid, int offset)
1659 {
1660         struct thread_data *td;
1661         void *data, *ret;
1662
1663 #if !defined(__hpux) && !defined(CONFIG_NO_SHM)
1664         data = shmat(shmid, NULL, 0);
1665         if (data == (void *) -1) {
1666                 int __err = errno;
1667
1668                 perror("shmat");
1669                 return __err;
1670         }
1671 #else
1672         /*
1673          * HP-UX inherits shm mappings?
1674          */
1675         data = threads;
1676 #endif
1677
1678         td = data + offset * sizeof(struct thread_data);
1679         ret = thread_main(td);
1680         shmdt(data);
1681         return (int) (uintptr_t) ret;
1682 }
1683
1684 static void dump_td_info(struct thread_data *td)
1685 {
1686         log_err("fio: job '%s' hasn't exited in %lu seconds, it appears to "
1687                 "be stuck. Doing forceful exit of this job.\n", td->o.name,
1688                         (unsigned long) time_since_now(&td->terminate_time));
1689 }
1690
1691 /*
1692  * Run over the job map and reap the threads that have exited, if any.
1693  */
1694 static void reap_threads(unsigned int *nr_running, unsigned int *t_rate,
1695                          unsigned int *m_rate)
1696 {
1697         struct thread_data *td;
1698         unsigned int cputhreads, realthreads, pending;
1699         int i, status, ret;
1700
1701         /*
1702          * reap exited threads (TD_EXITED -> TD_REAPED)
1703          */
1704         realthreads = pending = cputhreads = 0;
1705         for_each_td(td, i) {
1706                 int flags = 0;
1707
1708                 /*
1709                  * ->io_ops is NULL for a thread that has closed its
1710                  * io engine
1711                  */
1712                 if (td->io_ops && !strcmp(td->io_ops->name, "cpuio"))
1713                         cputhreads++;
1714                 else
1715                         realthreads++;
1716
1717                 if (!td->pid) {
1718                         pending++;
1719                         continue;
1720                 }
1721                 if (td->runstate == TD_REAPED)
1722                         continue;
1723                 if (td->o.use_thread) {
1724                         if (td->runstate == TD_EXITED) {
1725                                 td_set_runstate(td, TD_REAPED);
1726                                 goto reaped;
1727                         }
1728                         continue;
1729                 }
1730
1731                 flags = WNOHANG;
1732                 if (td->runstate == TD_EXITED)
1733                         flags = 0;
1734
1735                 /*
1736                  * check if someone quit or got killed in an unusual way
1737                  */
1738                 ret = waitpid(td->pid, &status, flags);
1739                 if (ret < 0) {
1740                         if (errno == ECHILD) {
1741                                 log_err("fio: pid=%d disappeared %d\n",
1742                                                 (int) td->pid, td->runstate);
1743                                 td->sig = ECHILD;
1744                                 td_set_runstate(td, TD_REAPED);
1745                                 goto reaped;
1746                         }
1747                         perror("waitpid");
1748                 } else if (ret == td->pid) {
1749                         if (WIFSIGNALED(status)) {
1750                                 int sig = WTERMSIG(status);
1751
1752                                 if (sig != SIGTERM && sig != SIGUSR2)
1753                                         log_err("fio: pid=%d, got signal=%d\n",
1754                                                         (int) td->pid, sig);
1755                                 td->sig = sig;
1756                                 td_set_runstate(td, TD_REAPED);
1757                                 goto reaped;
1758                         }
1759                         if (WIFEXITED(status)) {
1760                                 if (WEXITSTATUS(status) && !td->error)
1761                                         td->error = WEXITSTATUS(status);
1762
1763                                 td_set_runstate(td, TD_REAPED);
1764                                 goto reaped;
1765                         }
1766                 }
1767
1768                 /*
1769                  * If the job is stuck, do a forceful timeout of it and
1770                  * move on.
1771                  */
1772                 if (td->terminate &&
1773                     time_since_now(&td->terminate_time) >= FIO_REAP_TIMEOUT) {
1774                         dump_td_info(td);
1775                         td_set_runstate(td, TD_REAPED);
1776                         goto reaped;
1777                 }
1778
1779                 /*
1780                  * thread is not dead, continue
1781                  */
1782                 pending++;
1783                 continue;
1784 reaped:
1785                 (*nr_running)--;
1786                 (*m_rate) -= ddir_rw_sum(td->o.ratemin);
1787                 (*t_rate) -= ddir_rw_sum(td->o.rate);
1788                 if (!td->pid)
1789                         pending--;
1790
1791                 if (td->error)
1792                         exit_value++;
1793
1794                 done_secs += mtime_since_now(&td->epoch) / 1000;
1795                 profile_td_exit(td);
1796         }
1797
1798         if (*nr_running == cputhreads && !pending && realthreads)
1799                 fio_terminate_threads(TERMINATE_ALL);
1800 }
1801
1802 static int __check_trigger_file(void)
1803 {
1804         struct stat sb;
1805
1806         if (!trigger_file)
1807                 return 0;
1808
1809         if (stat(trigger_file, &sb))
1810                 return 0;
1811
1812         if (unlink(trigger_file) < 0)
1813                 log_err("fio: failed to unlink %s: %s\n", trigger_file,
1814                                                         strerror(errno));
1815
1816         return 1;
1817 }
1818
1819 static int trigger_timedout(void)
1820 {
1821         if (trigger_timeout)
1822                 return time_since_genesis() >= trigger_timeout;
1823
1824         return 0;
1825 }
1826
1827 void exec_trigger(const char *cmd)
1828 {
1829         int ret;
1830
1831         if (!cmd)
1832                 return;
1833
1834         ret = system(cmd);
1835         if (ret == -1)
1836                 log_err("fio: failed executing %s trigger\n", cmd);
1837 }
1838
1839 void check_trigger_file(void)
1840 {
1841         if (__check_trigger_file() || trigger_timedout()) {
1842                 if (nr_clients)
1843                         fio_clients_send_trigger(trigger_remote_cmd);
1844                 else {
1845                         verify_save_state();
1846                         fio_terminate_threads(TERMINATE_ALL);
1847                         exec_trigger(trigger_cmd);
1848                 }
1849         }
1850 }
1851
1852 static int fio_verify_load_state(struct thread_data *td)
1853 {
1854         int ret;
1855
1856         if (!td->o.verify_state)
1857                 return 0;
1858
1859         if (is_backend) {
1860                 void *data;
1861
1862                 ret = fio_server_get_verify_state(td->o.name,
1863                                         td->thread_number - 1, &data);
1864                 if (!ret)
1865                         verify_convert_assign_state(td, data);
1866         } else
1867                 ret = verify_load_state(td, "local");
1868
1869         return ret;
1870 }
1871
1872 static void do_usleep(unsigned int usecs)
1873 {
1874         check_for_running_stats();
1875         check_trigger_file();
1876         usleep(usecs);
1877 }
1878
1879 /*
1880  * Main function for kicking off and reaping jobs, as needed.
1881  */
1882 static void run_threads(void)
1883 {
1884         struct thread_data *td;
1885         unsigned int i, todo, nr_running, m_rate, t_rate, nr_started;
1886         uint64_t spent;
1887
1888         if (fio_gtod_offload && fio_start_gtod_thread())
1889                 return;
1890
1891         fio_idle_prof_init();
1892
1893         set_sig_handlers();
1894
1895         nr_thread = nr_process = 0;
1896         for_each_td(td, i) {
1897                 if (td->o.use_thread)
1898                         nr_thread++;
1899                 else
1900                         nr_process++;
1901         }
1902
1903         if (output_format == FIO_OUTPUT_NORMAL) {
1904                 log_info("Starting ");
1905                 if (nr_thread)
1906                         log_info("%d thread%s", nr_thread,
1907                                                 nr_thread > 1 ? "s" : "");
1908                 if (nr_process) {
1909                         if (nr_thread)
1910                                 log_info(" and ");
1911                         log_info("%d process%s", nr_process,
1912                                                 nr_process > 1 ? "es" : "");
1913                 }
1914                 log_info("\n");
1915                 log_info_flush();
1916         }
1917
1918         todo = thread_number;
1919         nr_running = 0;
1920         nr_started = 0;
1921         m_rate = t_rate = 0;
1922
1923         for_each_td(td, i) {
1924                 print_status_init(td->thread_number - 1);
1925
1926                 if (!td->o.create_serialize)
1927                         continue;
1928
1929                 if (fio_verify_load_state(td))
1930                         goto reap;
1931
1932                 /*
1933                  * do file setup here so it happens sequentially,
1934                  * we don't want X number of threads getting their
1935                  * client data interspersed on disk
1936                  */
1937                 if (setup_files(td)) {
1938 reap:
1939                         exit_value++;
1940                         if (td->error)
1941                                 log_err("fio: pid=%d, err=%d/%s\n",
1942                                         (int) td->pid, td->error, td->verror);
1943                         td_set_runstate(td, TD_REAPED);
1944                         todo--;
1945                 } else {
1946                         struct fio_file *f;
1947                         unsigned int j;
1948
1949                         /*
1950                          * for sharing to work, each job must always open
1951                          * its own files. so close them, if we opened them
1952                          * for creation
1953                          */
1954                         for_each_file(td, f, j) {
1955                                 if (fio_file_open(f))
1956                                         td_io_close_file(td, f);
1957                         }
1958                 }
1959         }
1960
1961         /* start idle threads before io threads start to run */
1962         fio_idle_prof_start();
1963
1964         set_genesis_time();
1965
1966         while (todo) {
1967                 struct thread_data *map[REAL_MAX_JOBS];
1968                 struct timeval this_start;
1969                 int this_jobs = 0, left;
1970
1971                 /*
1972                  * create threads (TD_NOT_CREATED -> TD_CREATED)
1973                  */
1974                 for_each_td(td, i) {
1975                         if (td->runstate != TD_NOT_CREATED)
1976                                 continue;
1977
1978                         /*
1979                          * never got a chance to start, killed by other
1980                          * thread for some reason
1981                          */
1982                         if (td->terminate) {
1983                                 todo--;
1984                                 continue;
1985                         }
1986
1987                         if (td->o.start_delay) {
1988                                 spent = utime_since_genesis();
1989
1990                                 if (td->o.start_delay > spent)
1991                                         continue;
1992                         }
1993
1994                         if (td->o.stonewall && (nr_started || nr_running)) {
1995                                 dprint(FD_PROCESS, "%s: stonewall wait\n",
1996                                                         td->o.name);
1997                                 break;
1998                         }
1999
2000                         init_disk_util(td);
2001
2002                         td->rusage_sem = fio_mutex_init(FIO_MUTEX_LOCKED);
2003                         td->update_rusage = 0;
2004
2005                         /*
2006                          * Set state to created. Thread will transition
2007                          * to TD_INITIALIZED when it's done setting up.
2008                          */
2009                         td_set_runstate(td, TD_CREATED);
2010                         map[this_jobs++] = td;
2011                         nr_started++;
2012
2013                         if (td->o.use_thread) {
2014                                 int ret;
2015
2016                                 dprint(FD_PROCESS, "will pthread_create\n");
2017                                 ret = pthread_create(&td->thread, NULL,
2018                                                         thread_main, td);
2019                                 if (ret) {
2020                                         log_err("pthread_create: %s\n",
2021                                                         strerror(ret));
2022                                         nr_started--;
2023                                         break;
2024                                 }
2025                                 ret = pthread_detach(td->thread);
2026                                 if (ret)
2027                                         log_err("pthread_detach: %s",
2028                                                         strerror(ret));
2029                         } else {
2030                                 pid_t pid;
2031                                 dprint(FD_PROCESS, "will fork\n");
2032                                 pid = fork();
2033                                 if (!pid) {
2034                                         int ret = fork_main(shm_id, i);
2035
2036                                         _exit(ret);
2037                                 } else if (i == fio_debug_jobno)
2038                                         *fio_debug_jobp = pid;
2039                         }
2040                         dprint(FD_MUTEX, "wait on startup_mutex\n");
2041                         if (fio_mutex_down_timeout(startup_mutex, 10)) {
2042                                 log_err("fio: job startup hung? exiting.\n");
2043                                 fio_terminate_threads(TERMINATE_ALL);
2044                                 fio_abort = 1;
2045                                 nr_started--;
2046                                 break;
2047                         }
2048                         dprint(FD_MUTEX, "done waiting on startup_mutex\n");
2049                 }
2050
2051                 /*
2052                  * Wait for the started threads to transition to
2053                  * TD_INITIALIZED.
2054                  */
2055                 fio_gettime(&this_start, NULL);
2056                 left = this_jobs;
2057                 while (left && !fio_abort) {
2058                         if (mtime_since_now(&this_start) > JOB_START_TIMEOUT)
2059                                 break;
2060
2061                         do_usleep(100000);
2062
2063                         for (i = 0; i < this_jobs; i++) {
2064                                 td = map[i];
2065                                 if (!td)
2066                                         continue;
2067                                 if (td->runstate == TD_INITIALIZED) {
2068                                         map[i] = NULL;
2069                                         left--;
2070                                 } else if (td->runstate >= TD_EXITED) {
2071                                         map[i] = NULL;
2072                                         left--;
2073                                         todo--;
2074                                         nr_running++; /* work-around... */
2075                                 }
2076                         }
2077                 }
2078
2079                 if (left) {
2080                         log_err("fio: %d job%s failed to start\n", left,
2081                                         left > 1 ? "s" : "");
2082                         for (i = 0; i < this_jobs; i++) {
2083                                 td = map[i];
2084                                 if (!td)
2085                                         continue;
2086                                 kill(td->pid, SIGTERM);
2087                         }
2088                         break;
2089                 }
2090
2091                 /*
2092                  * start created threads (TD_INITIALIZED -> TD_RUNNING).
2093                  */
2094                 for_each_td(td, i) {
2095                         if (td->runstate != TD_INITIALIZED)
2096                                 continue;
2097
2098                         if (in_ramp_time(td))
2099                                 td_set_runstate(td, TD_RAMP);
2100                         else
2101                                 td_set_runstate(td, TD_RUNNING);
2102                         nr_running++;
2103                         nr_started--;
2104                         m_rate += ddir_rw_sum(td->o.ratemin);
2105                         t_rate += ddir_rw_sum(td->o.rate);
2106                         todo--;
2107                         fio_mutex_up(td->mutex);
2108                 }
2109
2110                 reap_threads(&nr_running, &t_rate, &m_rate);
2111
2112                 if (todo)
2113                         do_usleep(100000);
2114         }
2115
2116         while (nr_running) {
2117                 reap_threads(&nr_running, &t_rate, &m_rate);
2118                 do_usleep(10000);
2119         }
2120
2121         fio_idle_prof_stop();
2122
2123         update_io_ticks();
2124 }
2125
2126 static void wait_for_helper_thread_exit(void)
2127 {
2128         void *ret;
2129
2130         helper_exit = 1;
2131         pthread_cond_signal(&helper_cond);
2132         pthread_join(helper_thread, &ret);
2133 }
2134
2135 static void free_disk_util(void)
2136 {
2137         disk_util_prune_entries();
2138
2139         pthread_cond_destroy(&helper_cond);
2140 }
2141
2142 static void *helper_thread_main(void *data)
2143 {
2144         int ret = 0;
2145
2146         fio_mutex_up(startup_mutex);
2147
2148         while (!ret) {
2149                 uint64_t sec = DISK_UTIL_MSEC / 1000;
2150                 uint64_t nsec = (DISK_UTIL_MSEC % 1000) * 1000000;
2151                 struct timespec ts;
2152                 struct timeval tv;
2153
2154                 gettimeofday(&tv, NULL);
2155                 ts.tv_sec = tv.tv_sec + sec;
2156                 ts.tv_nsec = (tv.tv_usec * 1000) + nsec;
2157
2158                 if (ts.tv_nsec >= 1000000000ULL) {
2159                         ts.tv_nsec -= 1000000000ULL;
2160                         ts.tv_sec++;
2161                 }
2162
2163                 pthread_cond_timedwait(&helper_cond, &helper_lock, &ts);
2164
2165                 ret = update_io_ticks();
2166
2167                 if (helper_do_stat) {
2168                         helper_do_stat = 0;
2169                         __show_running_run_stats();
2170                 }
2171
2172                 if (!is_backend)
2173                         print_thread_status();
2174         }
2175
2176         return NULL;
2177 }
2178
2179 static int create_helper_thread(void)
2180 {
2181         int ret;
2182
2183         setup_disk_util();
2184
2185         pthread_cond_init(&helper_cond, NULL);
2186         pthread_mutex_init(&helper_lock, NULL);
2187
2188         ret = pthread_create(&helper_thread, NULL, helper_thread_main, NULL);
2189         if (ret) {
2190                 log_err("Can't create helper thread: %s\n", strerror(ret));
2191                 return 1;
2192         }
2193
2194         dprint(FD_MUTEX, "wait on startup_mutex\n");
2195         fio_mutex_down(startup_mutex);
2196         dprint(FD_MUTEX, "done waiting on startup_mutex\n");
2197         return 0;
2198 }
2199
2200 int fio_backend(void)
2201 {
2202         struct thread_data *td;
2203         int i;
2204
2205         if (exec_profile) {
2206                 if (load_profile(exec_profile))
2207                         return 1;
2208                 free(exec_profile);
2209                 exec_profile = NULL;
2210         }
2211         if (!thread_number)
2212                 return 0;
2213
2214         if (write_bw_log) {
2215                 struct log_params p = {
2216                         .log_type = IO_LOG_TYPE_BW,
2217                 };
2218
2219                 setup_log(&agg_io_log[DDIR_READ], &p, "agg-read_bw.log");
2220                 setup_log(&agg_io_log[DDIR_WRITE], &p, "agg-write_bw.log");
2221                 setup_log(&agg_io_log[DDIR_TRIM], &p, "agg-trim_bw.log");
2222         }
2223
2224         startup_mutex = fio_mutex_init(FIO_MUTEX_LOCKED);
2225         if (startup_mutex == NULL)
2226                 return 1;
2227
2228         set_genesis_time();
2229         stat_init();
2230         create_helper_thread();
2231
2232         cgroup_list = smalloc(sizeof(*cgroup_list));
2233         INIT_FLIST_HEAD(cgroup_list);
2234
2235         run_threads();
2236
2237         wait_for_helper_thread_exit();
2238
2239         if (!fio_abort) {
2240                 __show_run_stats();
2241                 if (write_bw_log) {
2242                         for (i = 0; i < DDIR_RWDIR_CNT; i++) {
2243                                 struct io_log *log = agg_io_log[i];
2244
2245                                 flush_log(log);
2246                                 free_log(log);
2247                         }
2248                 }
2249         }
2250
2251         for_each_td(td, i) {
2252                 fio_options_free(td);
2253                 if (td->rusage_sem) {
2254                         fio_mutex_remove(td->rusage_sem);
2255                         td->rusage_sem = NULL;
2256                 }
2257         }
2258
2259         free_disk_util();
2260         cgroup_kill(cgroup_list);
2261         sfree(cgroup_list);
2262         sfree(cgroup_mnt);
2263
2264         fio_mutex_remove(startup_mutex);
2265         stat_exit();
2266         return exit_value;
2267 }