backend.c

   1 /*
   2  * fio - the flexible io tester
   3  *
   4  * Copyright (C) 2005 Jens Axboe <axboe@suse.de>
   5  * Copyright (C) 2006-2012 Jens Axboe <axboe@kernel.dk>
   6  *
   7  * The license below covers all files distributed with fio unless otherwise
   8  * noted in the file itself.
   9  *
  10  *  This program is free software; you can redistribute it and/or modify
  11  *  it under the terms of the GNU General Public License version 2 as
  12  *  published by the Free Software Foundation.
  13  *
  14  *  This program is distributed in the hope that it will be useful,
  15  *  but WITHOUT ANY WARRANTY; without even the implied warranty of
  16  *  MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
  17  *  GNU General Public License for more details.
  18  *
  19  *  You should have received a copy of the GNU General Public License
  20  *  along with this program; if not, write to the Free Software
  21  *  Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA  02111-1307  USA
  22  *
  23  */
  24 #include <unistd.h>
  25 #include <fcntl.h>
  26 #include <string.h>
  27 #include <limits.h>
  28 #include <signal.h>
  29 #include <time.h>
  30 #include <locale.h>
  31 #include <assert.h>
  32 #include <time.h>
  33 #include <inttypes.h>
  34 #include <sys/stat.h>
  35 #include <sys/wait.h>
  36 #include <sys/ipc.h>
  37 #include <sys/mman.h>
  38
  39 #include "fio.h"
  40 #ifndef FIO_NO_HAVE_SHM_H
  41 #include <sys/shm.h>
  42 #endif
  43 #include "hash.h"
  44 #include "smalloc.h"
  45 #include "verify.h"
  46 #include "trim.h"
  47 #include "diskutil.h"
  48 #include "cgroup.h"
  49 #include "profile.h"
  50 #include "lib/rand.h"
  51 #include "memalign.h"
  52 #include "server.h"
  53 #include "lib/getrusage.h"
  54 #include "idletime.h"
  55 #include "err.h"
  56 #include "lib/tp.h"
  57
  58 static pthread_t helper_thread;
  59 static pthread_mutex_t helper_lock;
  60 pthread_cond_t helper_cond;
  61 int helper_do_stat = 0;
  62
  63 static struct fio_mutex *startup_mutex;
  64 static struct flist_head *cgroup_list;
  65 static char *cgroup_mnt;
  66 static int exit_value;
  67 static volatile int fio_abort;
  68 static unsigned int nr_process = 0;
  69 static unsigned int nr_thread = 0;
  70
  71 struct io_log *agg_io_log[DDIR_RWDIR_CNT];
  72
  73 int groupid = 0;
  74 unsigned int thread_number = 0;
  75 unsigned int stat_number = 0;
  76 int shm_id = 0;
  77 int temp_stall_ts;
  78 unsigned long done_secs = 0;
  79 volatile int helper_exit = 0;
  80
  81 #define PAGE_ALIGN(buf) \
  82         (char *) (((uintptr_t) (buf) + page_mask) & ~page_mask)
  83
  84 #define JOB_START_TIMEOUT       (5 * 1000)
  85
  86 static void sig_int(int sig)
  87 {
  88         if (threads) {
  89                 if (is_backend)
  90                         fio_server_got_signal(sig);
  91                 else {
  92                         log_info("\nfio: terminating on signal %d\n", sig);
  93                         log_info_flush();
  94                         exit_value = 128;
  95                 }
  96
  97                 fio_terminate_threads(TERMINATE_ALL);
  98         }
  99 }
 100
 101 static void sig_show_status(int sig)
 102 {
 103         show_running_run_stats();
 104 }
 105
 106 static void set_sig_handlers(void)
 107 {
 108         struct sigaction act;
 109
 110         memset(&act, 0, sizeof(act));
 111         act.sa_handler = sig_int;
 112         act.sa_flags = SA_RESTART;
 113         sigaction(SIGINT, &act, NULL);
 114
 115         memset(&act, 0, sizeof(act));
 116         act.sa_handler = sig_int;
 117         act.sa_flags = SA_RESTART;
 118         sigaction(SIGTERM, &act, NULL);
 119
 120 /* Windows uses SIGBREAK as a quit signal from other applications */
 121 #ifdef WIN32
 122         memset(&act, 0, sizeof(act));
 123         act.sa_handler = sig_int;
 124         act.sa_flags = SA_RESTART;
 125         sigaction(SIGBREAK, &act, NULL);
 126 #endif
 127
 128         memset(&act, 0, sizeof(act));
 129         act.sa_handler = sig_show_status;
 130         act.sa_flags = SA_RESTART;
 131         sigaction(SIGUSR1, &act, NULL);
 132
 133         if (is_backend) {
 134                 memset(&act, 0, sizeof(act));
 135                 act.sa_handler = sig_int;
 136                 act.sa_flags = SA_RESTART;
 137                 sigaction(SIGPIPE, &act, NULL);
 138         }
 139 }
 140
 141 /*
 142  * Check if we are above the minimum rate given.
 143  */
 144 static int __check_min_rate(struct thread_data *td, struct timeval *now,
 145                             enum fio_ddir ddir)
 146 {
 147         unsigned long long bytes = 0;
 148         unsigned long iops = 0;
 149         unsigned long spent;
 150         unsigned long rate;
 151         unsigned int ratemin = 0;
 152         unsigned int rate_iops = 0;
 153         unsigned int rate_iops_min = 0;
 154
 155         assert(ddir_rw(ddir));
 156
 157         if (!td->o.ratemin[ddir] && !td->o.rate_iops_min[ddir])
 158                 return 0;
 159
 160         /*
 161          * allow a 2 second settle period in the beginning
 162          */
 163         if (mtime_since(&td->start, now) < 2000)
 164                 return 0;
 165
 166         iops += td->this_io_blocks[ddir];
 167         bytes += td->this_io_bytes[ddir];
 168         ratemin += td->o.ratemin[ddir];
 169         rate_iops += td->o.rate_iops[ddir];
 170         rate_iops_min += td->o.rate_iops_min[ddir];
 171
 172         /*
 173          * if rate blocks is set, sample is running
 174          */
 175         if (td->rate_bytes[ddir] || td->rate_blocks[ddir]) {
 176                 spent = mtime_since(&td->lastrate[ddir], now);
 177                 if (spent < td->o.ratecycle)
 178                         return 0;
 179
 180                 if (td->o.rate[ddir]) {
 181                         /*
 182                          * check bandwidth specified rate
 183                          */
 184                         if (bytes < td->rate_bytes[ddir]) {
 185                                 log_err("%s: min rate %u not met\n", td->o.name,
 186                                                                 ratemin);
 187                                 return 1;
 188                         } else {
 189                                 if (spent)
 190                                         rate = ((bytes - td->rate_bytes[ddir]) * 1000) / spent;
 191                                 else
 192                                         rate = 0;
 193
 194                                 if (rate < ratemin ||
 195                                     bytes < td->rate_bytes[ddir]) {
 196                                         log_err("%s: min rate %u not met, got"
 197                                                 " %luKB/sec\n", td->o.name,
 198                                                         ratemin, rate);
 199                                         return 1;
 200                                 }
 201                         }
 202                 } else {
 203                         /*
 204                          * checks iops specified rate
 205                          */
 206                         if (iops < rate_iops) {
 207                                 log_err("%s: min iops rate %u not met\n",
 208                                                 td->o.name, rate_iops);
 209                                 return 1;
 210                         } else {
 211                                 if (spent)
 212                                         rate = ((iops - td->rate_blocks[ddir]) * 1000) / spent;
 213                                 else
 214                                         rate = 0;
 215
 216                                 if (rate < rate_iops_min ||
 217                                     iops < td->rate_blocks[ddir]) {
 218                                         log_err("%s: min iops rate %u not met,"
 219                                                 " got %lu\n", td->o.name,
 220                                                         rate_iops_min, rate);
 221                                 }
 222                         }
 223                 }
 224         }
 225
 226         td->rate_bytes[ddir] = bytes;
 227         td->rate_blocks[ddir] = iops;
 228         memcpy(&td->lastrate[ddir], now, sizeof(*now));
 229         return 0;
 230 }
 231
 232 static int check_min_rate(struct thread_data *td, struct timeval *now,
 233                           uint64_t *bytes_done)
 234 {
 235         int ret = 0;
 236
 237         if (bytes_done[DDIR_READ])
 238                 ret |= __check_min_rate(td, now, DDIR_READ);
 239         if (bytes_done[DDIR_WRITE])
 240                 ret |= __check_min_rate(td, now, DDIR_WRITE);
 241         if (bytes_done[DDIR_TRIM])
 242                 ret |= __check_min_rate(td, now, DDIR_TRIM);
 243
 244         return ret;
 245 }
 246
 247 /*
 248  * When job exits, we can cancel the in-flight IO if we are using async
 249  * io. Attempt to do so.
 250  */
 251 static void cleanup_pending_aio(struct thread_data *td)
 252 {
 253         int r;
 254
 255         /*
 256          * get immediately available events, if any
 257          */
 258         r = io_u_queued_complete(td, 0, NULL);
 259         if (r < 0)
 260                 return;
 261
 262         /*
 263          * now cancel remaining active events
 264          */
 265         if (td->io_ops->cancel) {
 266                 struct io_u *io_u;
 267                 int i;
 268
 269                 io_u_qiter(&td->io_u_all, io_u, i) {
 270                         if (io_u->flags & IO_U_F_FLIGHT) {
 271                                 r = td->io_ops->cancel(td, io_u);
 272                                 if (!r)
 273                                         put_io_u(td, io_u);
 274                         }
 275                 }
 276         }
 277
 278         if (td->cur_depth)
 279                 r = io_u_queued_complete(td, td->cur_depth, NULL);
 280 }
 281
 282 /*
 283  * Helper to handle the final sync of a file. Works just like the normal
 284  * io path, just does everything sync.
 285  */
 286 static int fio_io_sync(struct thread_data *td, struct fio_file *f)
 287 {
 288         struct io_u *io_u = __get_io_u(td);
 289         int ret;
 290
 291         if (!io_u)
 292                 return 1;
 293
 294         io_u->ddir = DDIR_SYNC;
 295         io_u->file = f;
 296
 297         if (td_io_prep(td, io_u)) {
 298                 put_io_u(td, io_u);
 299                 return 1;
 300         }
 301
 302 requeue:
 303         ret = td_io_queue(td, io_u);
 304         if (ret < 0) {
 305                 td_verror(td, io_u->error, "td_io_queue");
 306                 put_io_u(td, io_u);
 307                 return 1;
 308         } else if (ret == FIO_Q_QUEUED) {
 309                 if (io_u_queued_complete(td, 1, NULL) < 0)
 310                         return 1;
 311         } else if (ret == FIO_Q_COMPLETED) {
 312                 if (io_u->error) {
 313                         td_verror(td, io_u->error, "td_io_queue");
 314                         return 1;
 315                 }
 316
 317                 if (io_u_sync_complete(td, io_u, NULL) < 0)
 318                         return 1;
 319         } else if (ret == FIO_Q_BUSY) {
 320                 if (td_io_commit(td))
 321                         return 1;
 322                 goto requeue;
 323         }
 324
 325         return 0;
 326 }
 327
 328 static int fio_file_fsync(struct thread_data *td, struct fio_file *f)
 329 {
 330         int ret;
 331
 332         if (fio_file_open(f))
 333                 return fio_io_sync(td, f);
 334
 335         if (td_io_open_file(td, f))
 336                 return 1;
 337
 338         ret = fio_io_sync(td, f);
 339         td_io_close_file(td, f);
 340         return ret;
 341 }
 342
 343 static inline void __update_tv_cache(struct thread_data *td)
 344 {
 345         fio_gettime(&td->tv_cache, NULL);
 346 }
 347
 348 static inline void update_tv_cache(struct thread_data *td)
 349 {
 350         if ((++td->tv_cache_nr & td->tv_cache_mask) == td->tv_cache_mask)
 351                 __update_tv_cache(td);
 352 }
 353
 354 static inline int runtime_exceeded(struct thread_data *td, struct timeval *t)
 355 {
 356         if (in_ramp_time(td))
 357                 return 0;
 358         if (!td->o.timeout)
 359                 return 0;
 360         if (utime_since(&td->epoch, t) >= td->o.timeout)
 361                 return 1;
 362
 363         return 0;
 364 }
 365
 366 static int break_on_this_error(struct thread_data *td, enum fio_ddir ddir,
 367                                int *retptr)
 368 {
 369         int ret = *retptr;
 370
 371         if (ret < 0 || td->error) {
 372                 int err = td->error;
 373                 enum error_type_bit eb;
 374
 375                 if (ret < 0)
 376                         err = -ret;
 377
 378                 eb = td_error_type(ddir, err);
 379                 if (!(td->o.continue_on_error & (1 << eb)))
 380                         return 1;
 381
 382                 if (td_non_fatal_error(td, eb, err)) {
 383                         /*
 384                          * Continue with the I/Os in case of
 385                          * a non fatal error.
 386                          */
 387                         update_error_count(td, err);
 388                         td_clear_error(td);
 389                         *retptr = 0;
 390                         return 0;
 391                 } else if (td->o.fill_device && err == ENOSPC) {
 392                         /*
 393                          * We expect to hit this error if
 394                          * fill_device option is set.
 395                          */
 396                         td_clear_error(td);
 397                         fio_mark_td_terminate(td);
 398                         return 1;
 399                 } else {
 400                         /*
 401                          * Stop the I/O in case of a fatal
 402                          * error.
 403                          */
 404                         update_error_count(td, err);
 405                         return 1;
 406                 }
 407         }
 408
 409         return 0;
 410 }
 411
 412 static void check_update_rusage(struct thread_data *td)
 413 {
 414         if (td->update_rusage) {
 415                 td->update_rusage = 0;
 416                 update_rusage_stat(td);
 417                 fio_mutex_up(td->rusage_sem);
 418         }
 419 }
 420
 421 static int wait_for_completions(struct thread_data *td, struct timeval *time,
 422                                 uint64_t *bytes_done)
 423 {
 424         const int full = queue_full(td);
 425         int min_evts = 0;
 426         int ret;
 427
 428         /*
 429          * if the queue is full, we MUST reap at least 1 event
 430          */
 431         min_evts = min(td->o.iodepth_batch_complete, td->cur_depth);
 432         if (full && !min_evts)
 433                 min_evts = 1;
 434
 435         if (time && (__should_check_rate(td, DDIR_READ) ||
 436             __should_check_rate(td, DDIR_WRITE) ||
 437             __should_check_rate(td, DDIR_TRIM)))
 438                 fio_gettime(time, NULL);
 439
 440         do {
 441                 ret = io_u_queued_complete(td, min_evts, bytes_done);
 442                 if (ret < 0)
 443                         break;
 444         } while (full && (td->cur_depth > td->o.iodepth_low));
 445
 446         return ret;
 447 }
 448
 449 /*
 450  * The main verify engine. Runs over the writes we previously submitted,
 451  * reads the blocks back in, and checks the crc/md5 of the data.
 452  */
 453 static void do_verify(struct thread_data *td, uint64_t verify_bytes)
 454 {
 455         uint64_t bytes_done[DDIR_RWDIR_CNT] = { 0, 0, 0 };
 456         struct fio_file *f;
 457         struct io_u *io_u;
 458         int ret, min_events;
 459         unsigned int i;
 460
 461         dprint(FD_VERIFY, "starting loop\n");
 462
 463         /*
 464          * sync io first and invalidate cache, to make sure we really
 465          * read from disk.
 466          */
 467         for_each_file(td, f, i) {
 468                 if (!fio_file_open(f))
 469                         continue;
 470                 if (fio_io_sync(td, f))
 471                         break;
 472                 if (file_invalidate_cache(td, f))
 473                         break;
 474         }
 475
 476         check_update_rusage(td);
 477
 478         if (td->error)
 479                 return;
 480
 481         td_set_runstate(td, TD_VERIFYING);
 482
 483         io_u = NULL;
 484         while (!td->terminate) {
 485                 enum fio_ddir ddir;
 486                 int ret2, full;
 487
 488                 update_tv_cache(td);
 489                 check_update_rusage(td);
 490
 491                 if (runtime_exceeded(td, &td->tv_cache)) {
 492                         __update_tv_cache(td);
 493                         if (runtime_exceeded(td, &td->tv_cache)) {
 494                                 fio_mark_td_terminate(td);
 495                                 break;
 496                         }
 497                 }
 498
 499                 if (flow_threshold_exceeded(td))
 500                         continue;
 501
 502                 if (!td->o.experimental_verify) {
 503                         io_u = __get_io_u(td);
 504                         if (!io_u)
 505                                 break;
 506
 507                         if (get_next_verify(td, io_u)) {
 508                                 put_io_u(td, io_u);
 509                                 break;
 510                         }
 511
 512                         if (td_io_prep(td, io_u)) {
 513                                 put_io_u(td, io_u);
 514                                 break;
 515                         }
 516                 } else {
 517                         if (ddir_rw_sum(bytes_done) + td->o.rw_min_bs > verify_bytes)
 518                                 break;
 519
 520                         while ((io_u = get_io_u(td)) != NULL) {
 521                                 if (IS_ERR(io_u)) {
 522                                         io_u = NULL;
 523                                         ret = FIO_Q_BUSY;
 524                                         goto reap;
 525                                 }
 526
 527                                 /*
 528                                  * We are only interested in the places where
 529                                  * we wrote or trimmed IOs. Turn those into
 530                                  * reads for verification purposes.
 531                                  */
 532                                 if (io_u->ddir == DDIR_READ) {
 533                                         /*
 534                                          * Pretend we issued it for rwmix
 535                                          * accounting
 536                                          */
 537                                         td->io_issues[DDIR_READ]++;
 538                                         put_io_u(td, io_u);
 539                                         continue;
 540                                 } else if (io_u->ddir == DDIR_TRIM) {
 541                                         io_u->ddir = DDIR_READ;
 542                                         io_u->flags |= IO_U_F_TRIMMED;
 543                                         break;
 544                                 } else if (io_u->ddir == DDIR_WRITE) {
 545                                         io_u->ddir = DDIR_READ;
 546                                         break;
 547                                 } else {
 548                                         put_io_u(td, io_u);
 549                                         continue;
 550                                 }
 551                         }
 552
 553                         if (!io_u)
 554                                 break;
 555                 }
 556
 557                 if (verify_state_should_stop(td, io_u)) {
 558                         put_io_u(td, io_u);
 559                         break;
 560                 }
 561
 562                 if (td->o.verify_async)
 563                         io_u->end_io = verify_io_u_async;
 564                 else
 565                         io_u->end_io = verify_io_u;
 566
 567                 ddir = io_u->ddir;
 568                 if (!td->o.disable_slat)
 569                         fio_gettime(&io_u->start_time, NULL);
 570
 571                 ret = td_io_queue(td, io_u);
 572                 switch (ret) {
 573                 case FIO_Q_COMPLETED:
 574                         if (io_u->error) {
 575                                 ret = -io_u->error;
 576                                 clear_io_u(td, io_u);
 577                         } else if (io_u->resid) {
 578                                 int bytes = io_u->xfer_buflen - io_u->resid;
 579
 580                                 /*
 581                                  * zero read, fail
 582                                  */
 583                                 if (!bytes) {
 584                                         td_verror(td, EIO, "full resid");
 585                                         put_io_u(td, io_u);
 586                                         break;
 587                                 }
 588
 589                                 io_u->xfer_buflen = io_u->resid;
 590                                 io_u->xfer_buf += bytes;
 591                                 io_u->offset += bytes;
 592
 593                                 if (ddir_rw(io_u->ddir))
 594                                         td->ts.short_io_u[io_u->ddir]++;
 595
 596                                 f = io_u->file;
 597                                 if (io_u->offset == f->real_file_size)
 598                                         goto sync_done;
 599
 600                                 requeue_io_u(td, &io_u);
 601                         } else {
 602 sync_done:
 603                                 ret = io_u_sync_complete(td, io_u, bytes_done);
 604                                 if (ret < 0)
 605                                         break;
 606                         }
 607                         continue;
 608                 case FIO_Q_QUEUED:
 609                         break;
 610                 case FIO_Q_BUSY:
 611                         requeue_io_u(td, &io_u);
 612                         ret2 = td_io_commit(td);
 613                         if (ret2 < 0)
 614                                 ret = ret2;
 615                         break;
 616                 default:
 617                         assert(ret < 0);
 618                         td_verror(td, -ret, "td_io_queue");
 619                         break;
 620                 }
 621
 622                 if (break_on_this_error(td, ddir, &ret))
 623                         break;
 624
 625                 /*
 626                  * if we can queue more, do so. but check if there are
 627                  * completed io_u's first. Note that we can get BUSY even
 628                  * without IO queued, if the system is resource starved.
 629                  */
 630 reap:
 631                 full = queue_full(td) || (ret == FIO_Q_BUSY && td->cur_depth);
 632                 if (full || !td->o.iodepth_batch_complete)
 633                         ret = wait_for_completions(td, NULL, bytes_done);
 634
 635                 if (ret < 0)
 636                         break;
 637         }
 638
 639         check_update_rusage(td);
 640
 641         if (!td->error) {
 642                 min_events = td->cur_depth;
 643
 644                 if (min_events)
 645                         ret = io_u_queued_complete(td, min_events, NULL);
 646         } else
 647                 cleanup_pending_aio(td);
 648
 649         td_set_runstate(td, TD_RUNNING);
 650
 651         dprint(FD_VERIFY, "exiting loop\n");
 652 }
 653
 654 static unsigned int exceeds_number_ios(struct thread_data *td)
 655 {
 656         unsigned long long number_ios;
 657
 658         if (!td->o.number_ios)
 659                 return 0;
 660
 661         number_ios = ddir_rw_sum(td->this_io_blocks);
 662         number_ios += td->io_u_queued + td->io_u_in_flight;
 663
 664         return number_ios >= td->o.number_ios;
 665 }
 666
 667 static int io_issue_bytes_exceeded(struct thread_data *td)
 668 {
 669         unsigned long long bytes, limit;
 670
 671         if (td_rw(td))
 672                 bytes = td->io_issue_bytes[DDIR_READ] + td->io_issue_bytes[DDIR_WRITE];
 673         else if (td_write(td))
 674                 bytes = td->io_issue_bytes[DDIR_WRITE];
 675         else if (td_read(td))
 676                 bytes = td->io_issue_bytes[DDIR_READ];
 677         else
 678                 bytes = td->io_issue_bytes[DDIR_TRIM];
 679
 680         if (td->o.io_limit)
 681                 limit = td->o.io_limit;
 682         else
 683                 limit = td->o.size;
 684
 685         return bytes >= limit || exceeds_number_ios(td);
 686 }
 687
 688 static int io_complete_bytes_exceeded(struct thread_data *td)
 689 {
 690         unsigned long long bytes, limit;
 691
 692         if (td_rw(td))
 693                 bytes = td->this_io_bytes[DDIR_READ] + td->this_io_bytes[DDIR_WRITE];
 694         else if (td_write(td))
 695                 bytes = td->this_io_bytes[DDIR_WRITE];
 696         else if (td_read(td))
 697                 bytes = td->this_io_bytes[DDIR_READ];
 698         else
 699                 bytes = td->this_io_bytes[DDIR_TRIM];
 700
 701         if (td->o.io_limit)
 702                 limit = td->o.io_limit;
 703         else
 704                 limit = td->o.size;
 705
 706         return bytes >= limit || exceeds_number_ios(td);
 707 }
 708
 709 /*
 710  * Main IO worker function. It retrieves io_u's to process and queues
 711  * and reaps them, checking for rate and errors along the way.
 712  *
 713  * Returns number of bytes written and trimmed.
 714  */
 715 static uint64_t do_io(struct thread_data *td)
 716 {
 717         uint64_t bytes_done[DDIR_RWDIR_CNT] = { 0, 0, 0 };
 718         unsigned int i;
 719         int ret = 0;
 720         uint64_t total_bytes, bytes_issued = 0;
 721
 722         if (in_ramp_time(td))
 723                 td_set_runstate(td, TD_RAMP);
 724         else
 725                 td_set_runstate(td, TD_RUNNING);
 726
 727         lat_target_init(td);
 728
 729         total_bytes = td->o.size;
 730         /*
 731         * Allow random overwrite workloads to write up to io_limit
 732         * before starting verification phase as 'size' doesn't apply.
 733         */
 734         if (td_write(td) && td_random(td) && td->o.norandommap)
 735                 total_bytes = max(total_bytes, (uint64_t) td->o.io_limit);
 736         /*
 737          * If verify_backlog is enabled, we'll run the verify in this
 738          * handler as well. For that case, we may need up to twice the
 739          * amount of bytes.
 740          */
 741         if (td->o.verify != VERIFY_NONE &&
 742            (td_write(td) && td->o.verify_backlog))
 743                 total_bytes += td->o.size;
 744
 745         while ((td->o.read_iolog_file && !flist_empty(&td->io_log_list)) ||
 746                 (!flist_empty(&td->trim_list)) || !io_issue_bytes_exceeded(td) ||
 747                 td->o.time_based) {
 748                 struct timeval comp_time;
 749                 struct io_u *io_u;
 750                 int ret2, full;
 751                 enum fio_ddir ddir;
 752
 753                 check_update_rusage(td);
 754
 755                 if (td->terminate || td->done)
 756                         break;
 757
 758                 update_tv_cache(td);
 759
 760                 if (runtime_exceeded(td, &td->tv_cache)) {
 761                         __update_tv_cache(td);
 762                         if (runtime_exceeded(td, &td->tv_cache)) {
 763                                 fio_mark_td_terminate(td);
 764                                 break;
 765                         }
 766                 }
 767
 768                 if (flow_threshold_exceeded(td))
 769                         continue;
 770
 771                 if (bytes_issued >= total_bytes)
 772                         break;
 773
 774                 io_u = get_io_u(td);
 775                 if (IS_ERR_OR_NULL(io_u)) {
 776                         int err = PTR_ERR(io_u);
 777
 778                         io_u = NULL;
 779                         if (err == -EBUSY) {
 780                                 ret = FIO_Q_BUSY;
 781                                 goto reap;
 782                         }
 783                         if (td->o.latency_target)
 784                                 goto reap;
 785                         break;
 786                 }
 787
 788                 ddir = io_u->ddir;
 789
 790                 /*
 791                  * Add verification end_io handler if:
 792                  *      - Asked to verify (!td_rw(td))
 793                  *      - Or the io_u is from our verify list (mixed write/ver)
 794                  */
 795                 if (td->o.verify != VERIFY_NONE && io_u->ddir == DDIR_READ &&
 796                     ((io_u->flags & IO_U_F_VER_LIST) || !td_rw(td))) {
 797
 798                         if (!td->o.verify_pattern_bytes) {
 799                                 io_u->rand_seed = __rand(&td->verify_state);
 800                                 if (sizeof(int) != sizeof(long *))
 801                                         io_u->rand_seed *= __rand(&td->verify_state);
 802                         }
 803
 804                         if (verify_state_should_stop(td, io_u)) {
 805                                 put_io_u(td, io_u);
 806                                 break;
 807                         }
 808
 809                         if (td->o.verify_async)
 810                                 io_u->end_io = verify_io_u_async;
 811                         else
 812                                 io_u->end_io = verify_io_u;
 813                         td_set_runstate(td, TD_VERIFYING);
 814                 } else if (in_ramp_time(td))
 815                         td_set_runstate(td, TD_RAMP);
 816                 else
 817                         td_set_runstate(td, TD_RUNNING);
 818
 819                 /*
 820                  * Always log IO before it's issued, so we know the specific
 821                  * order of it. The logged unit will track when the IO has
 822                  * completed.
 823                  */
 824                 if (td_write(td) && io_u->ddir == DDIR_WRITE &&
 825                     td->o.do_verify &&
 826                     td->o.verify != VERIFY_NONE &&
 827                     !td->o.experimental_verify)
 828                         log_io_piece(td, io_u);
 829
 830                 ret = td_io_queue(td, io_u);
 831                 switch (ret) {
 832                 case FIO_Q_COMPLETED:
 833                         if (io_u->error) {
 834                                 ret = -io_u->error;
 835                                 unlog_io_piece(td, io_u);
 836                                 clear_io_u(td, io_u);
 837                         } else if (io_u->resid) {
 838                                 int bytes = io_u->xfer_buflen - io_u->resid;
 839                                 struct fio_file *f = io_u->file;
 840
 841                                 bytes_issued += bytes;
 842
 843                                 trim_io_piece(td, io_u);
 844
 845                                 /*
 846                                  * zero read, fail
 847                                  */
 848                                 if (!bytes) {
 849                                         unlog_io_piece(td, io_u);
 850                                         td_verror(td, EIO, "full resid");
 851                                         put_io_u(td, io_u);
 852                                         break;
 853                                 }
 854
 855                                 io_u->xfer_buflen = io_u->resid;
 856                                 io_u->xfer_buf += bytes;
 857                                 io_u->offset += bytes;
 858
 859                                 if (ddir_rw(io_u->ddir))
 860                                         td->ts.short_io_u[io_u->ddir]++;
 861
 862                                 if (io_u->offset == f->real_file_size)
 863                                         goto sync_done;
 864
 865                                 requeue_io_u(td, &io_u);
 866                         } else {
 867 sync_done:
 868                                 if (__should_check_rate(td, DDIR_READ) ||
 869                                     __should_check_rate(td, DDIR_WRITE) ||
 870                                     __should_check_rate(td, DDIR_TRIM))
 871                                         fio_gettime(&comp_time, NULL);
 872
 873                                 ret = io_u_sync_complete(td, io_u, bytes_done);
 874                                 if (ret < 0)
 875                                         break;
 876                                 bytes_issued += io_u->xfer_buflen;
 877                         }
 878                         break;
 879                 case FIO_Q_QUEUED:
 880                         /*
 881                          * if the engine doesn't have a commit hook,
 882                          * the io_u is really queued. if it does have such
 883                          * a hook, it has to call io_u_queued() itself.
 884                          */
 885                         if (td->io_ops->commit == NULL)
 886                                 io_u_queued(td, io_u);
 887                         bytes_issued += io_u->xfer_buflen;
 888                         break;
 889                 case FIO_Q_BUSY:
 890                         unlog_io_piece(td, io_u);
 891                         requeue_io_u(td, &io_u);
 892                         ret2 = td_io_commit(td);
 893                         if (ret2 < 0)
 894                                 ret = ret2;
 895                         break;
 896                 default:
 897                         assert(ret < 0);
 898                         put_io_u(td, io_u);
 899                         break;
 900                 }
 901
 902                 if (break_on_this_error(td, ddir, &ret))
 903                         break;
 904
 905                 /*
 906                  * See if we need to complete some commands. Note that we
 907                  * can get BUSY even without IO queued, if the system is
 908                  * resource starved.
 909                  */
 910 reap:
 911                 full = queue_full(td) || (ret == FIO_Q_BUSY && td->cur_depth);
 912                 if (full || !td->o.iodepth_batch_complete)
 913                         ret = wait_for_completions(td, &comp_time, bytes_done);
 914                 if (ret < 0)
 915                         break;
 916                 if (!ddir_rw_sum(bytes_done) && !(td->io_ops->flags & FIO_NOIO))
 917                         continue;
 918
 919                 if (!in_ramp_time(td) && should_check_rate(td, bytes_done)) {
 920                         if (check_min_rate(td, &comp_time, bytes_done)) {
 921                                 if (exitall_on_terminate)
 922                                         fio_terminate_threads(td->groupid);
 923                                 td_verror(td, EIO, "check_min_rate");
 924                                 break;
 925                         }
 926                 }
 927                 if (!in_ramp_time(td) && td->o.latency_target)
 928                         lat_target_check(td);
 929
 930                 if (td->o.thinktime) {
 931                         unsigned long long b;
 932
 933                         b = ddir_rw_sum(td->io_blocks);
 934                         if (!(b % td->o.thinktime_blocks)) {
 935                                 int left;
 936
 937                                 io_u_quiesce(td);
 938
 939                                 if (td->o.thinktime_spin)
 940                                         usec_spin(td->o.thinktime_spin);
 941
 942                                 left = td->o.thinktime - td->o.thinktime_spin;
 943                                 if (left)
 944                                         usec_sleep(td, left);
 945                         }
 946                 }
 947         }
 948
 949         check_update_rusage(td);
 950
 951         if (td->trim_entries)
 952                 log_err("fio: %lu trim entries leaked?\n", td->trim_entries);
 953
 954         if (td->o.fill_device && td->error == ENOSPC) {
 955                 td->error = 0;
 956                 fio_mark_td_terminate(td);
 957         }
 958         if (!td->error) {
 959                 struct fio_file *f;
 960
 961                 i = td->cur_depth;
 962                 if (i) {
 963                         ret = io_u_queued_complete(td, i, bytes_done);
 964                         if (td->o.fill_device && td->error == ENOSPC)
 965                                 td->error = 0;
 966                 }
 967
 968                 if (should_fsync(td) && td->o.end_fsync) {
 969                         td_set_runstate(td, TD_FSYNCING);
 970
 971                         for_each_file(td, f, i) {
 972                                 if (!fio_file_fsync(td, f))
 973                                         continue;
 974
 975                                 log_err("fio: end_fsync failed for file %s\n",
 976                                                                 f->file_name);
 977                         }
 978                 }
 979         } else
 980                 cleanup_pending_aio(td);
 981
 982         /*
 983          * stop job if we failed doing any IO
 984          */
 985         if (!ddir_rw_sum(td->this_io_bytes))
 986                 td->done = 1;
 987
 988         return bytes_done[DDIR_WRITE] + bytes_done[DDIR_TRIM];
 989 }
 990
 991 static void cleanup_io_u(struct thread_data *td)
 992 {
 993         struct io_u *io_u;
 994
 995         while ((io_u = io_u_qpop(&td->io_u_freelist)) != NULL) {
 996
 997                 if (td->io_ops->io_u_free)
 998                         td->io_ops->io_u_free(td, io_u);
 999
1000                 fio_memfree(io_u, sizeof(*io_u));
1001         }
1002
1003         free_io_mem(td);
1004
1005         io_u_rexit(&td->io_u_requeues);
1006         io_u_qexit(&td->io_u_freelist);
1007         io_u_qexit(&td->io_u_all);
1008
1009         if (td->last_write_comp)
1010                 sfree(td->last_write_comp);
1011 }
1012
1013 static int init_io_u(struct thread_data *td)
1014 {
1015         struct io_u *io_u;
1016         unsigned int max_bs, min_write;
1017         int cl_align, i, max_units;
1018         int data_xfer = 1, err;
1019         char *p;
1020
1021         max_units = td->o.iodepth;
1022         max_bs = td_max_bs(td);
1023         min_write = td->o.min_bs[DDIR_WRITE];
1024         td->orig_buffer_size = (unsigned long long) max_bs
1025                                         * (unsigned long long) max_units;
1026
1027         if ((td->io_ops->flags & FIO_NOIO) || !(td_read(td) || td_write(td)))
1028                 data_xfer = 0;
1029
1030         err = 0;
1031         err += io_u_rinit(&td->io_u_requeues, td->o.iodepth);
1032         err += io_u_qinit(&td->io_u_freelist, td->o.iodepth);
1033         err += io_u_qinit(&td->io_u_all, td->o.iodepth);
1034
1035         if (err) {
1036                 log_err("fio: failed setting up IO queues\n");
1037                 return 1;
1038         }
1039
1040         /*
1041          * if we may later need to do address alignment, then add any
1042          * possible adjustment here so that we don't cause a buffer
1043          * overflow later. this adjustment may be too much if we get
1044          * lucky and the allocator gives us an aligned address.
1045          */
1046         if (td->o.odirect || td->o.mem_align || td->o.oatomic ||
1047             (td->io_ops->flags & FIO_RAWIO))
1048                 td->orig_buffer_size += page_mask + td->o.mem_align;
1049
1050         if (td->o.mem_type == MEM_SHMHUGE || td->o.mem_type == MEM_MMAPHUGE) {
1051                 unsigned long bs;
1052
1053                 bs = td->orig_buffer_size + td->o.hugepage_size - 1;
1054                 td->orig_buffer_size = bs & ~(td->o.hugepage_size - 1);
1055         }
1056
1057         if (td->orig_buffer_size != (size_t) td->orig_buffer_size) {
1058                 log_err("fio: IO memory too large. Reduce max_bs or iodepth\n");
1059                 return 1;
1060         }
1061
1062         if (data_xfer && allocate_io_mem(td))
1063                 return 1;
1064
1065         if (td->o.odirect || td->o.mem_align || td->o.oatomic ||
1066             (td->io_ops->flags & FIO_RAWIO))
1067                 p = PAGE_ALIGN(td->orig_buffer) + td->o.mem_align;
1068         else
1069                 p = td->orig_buffer;
1070
1071         cl_align = os_cache_line_size();
1072
1073         for (i = 0; i < max_units; i++) {
1074                 void *ptr;
1075
1076                 if (td->terminate)
1077                         return 1;
1078
1079                 ptr = fio_memalign(cl_align, sizeof(*io_u));
1080                 if (!ptr) {
1081                         log_err("fio: unable to allocate aligned memory\n");
1082                         break;
1083                 }
1084
1085                 io_u = ptr;
1086                 memset(io_u, 0, sizeof(*io_u));
1087                 INIT_FLIST_HEAD(&io_u->verify_list);
1088                 dprint(FD_MEM, "io_u alloc %p, index %u\n", io_u, i);
1089
1090                 if (data_xfer) {
1091                         io_u->buf = p;
1092                         dprint(FD_MEM, "io_u %p, mem %p\n", io_u, io_u->buf);
1093
1094                         if (td_write(td))
1095                                 io_u_fill_buffer(td, io_u, min_write, max_bs);
1096                         if (td_write(td) && td->o.verify_pattern_bytes) {
1097                                 /*
1098                                  * Fill the buffer with the pattern if we are
1099                                  * going to be doing writes.
1100                                  */
1101                                 fill_verify_pattern(td, io_u->buf, max_bs, io_u, 0, 0);
1102                         }
1103                 }
1104
1105                 io_u->index = i;
1106                 io_u->flags = IO_U_F_FREE;
1107                 io_u_qpush(&td->io_u_freelist, io_u);
1108
1109                 /*
1110                  * io_u never leaves this stack, used for iteration of all
1111                  * io_u buffers.
1112                  */
1113                 io_u_qpush(&td->io_u_all, io_u);
1114
1115                 if (td->io_ops->io_u_init) {
1116                         int ret = td->io_ops->io_u_init(td, io_u);
1117
1118                         if (ret) {
1119                                 log_err("fio: failed to init engine data: %d\n", ret);
1120                                 return 1;
1121                         }
1122                 }
1123
1124                 p += max_bs;
1125         }
1126
1127         if (td->o.verify != VERIFY_NONE) {
1128                 td->last_write_comp = scalloc(max_units, sizeof(uint64_t));
1129                 if (!td->last_write_comp) {
1130                         log_err("fio: failed to alloc write comp data\n");
1131                         return 1;
1132                 }
1133         }
1134
1135         return 0;
1136 }
1137
1138 static int switch_ioscheduler(struct thread_data *td)
1139 {
1140         char tmp[256], tmp2[128];
1141         FILE *f;
1142         int ret;
1143
1144         if (td->io_ops->flags & FIO_DISKLESSIO)
1145                 return 0;
1146
1147         sprintf(tmp, "%s/queue/scheduler", td->sysfs_root);
1148
1149         f = fopen(tmp, "r+");
1150         if (!f) {
1151                 if (errno == ENOENT) {
1152                         log_err("fio: os or kernel doesn't support IO scheduler"
1153                                 " switching\n");
1154                         return 0;
1155                 }
1156                 td_verror(td, errno, "fopen iosched");
1157                 return 1;
1158         }
1159
1160         /*
1161          * Set io scheduler.
1162          */
1163         ret = fwrite(td->o.ioscheduler, strlen(td->o.ioscheduler), 1, f);
1164         if (ferror(f) || ret != 1) {
1165                 td_verror(td, errno, "fwrite");
1166                 fclose(f);
1167                 return 1;
1168         }
1169
1170         rewind(f);
1171
1172         /*
1173          * Read back and check that the selected scheduler is now the default.
1174          */
1175         ret = fread(tmp, sizeof(tmp), 1, f);
1176         if (ferror(f) || ret < 0) {
1177                 td_verror(td, errno, "fread");
1178                 fclose(f);
1179                 return 1;
1180         }
1181         tmp[sizeof(tmp) - 1] = '\0';
1182
1183
1184         sprintf(tmp2, "[%s]", td->o.ioscheduler);
1185         if (!strstr(tmp, tmp2)) {
1186                 log_err("fio: io scheduler %s not found\n", td->o.ioscheduler);
1187                 td_verror(td, EINVAL, "iosched_switch");
1188                 fclose(f);
1189                 return 1;
1190         }
1191
1192         fclose(f);
1193         return 0;
1194 }
1195
1196 static int keep_running(struct thread_data *td)
1197 {
1198         unsigned long long limit;
1199
1200         if (td->done)
1201                 return 0;
1202         if (td->o.time_based)
1203                 return 1;
1204         if (td->o.loops) {
1205                 td->o.loops--;
1206                 return 1;
1207         }
1208         if (exceeds_number_ios(td))
1209                 return 0;
1210
1211         if (td->o.io_limit)
1212                 limit = td->o.io_limit;
1213         else
1214                 limit = td->o.size;
1215
1216         if (limit != -1ULL && ddir_rw_sum(td->io_bytes) < limit) {
1217                 uint64_t diff;
1218
1219                 /*
1220                  * If the difference is less than the minimum IO size, we
1221                  * are done.
1222                  */
1223                 diff = limit - ddir_rw_sum(td->io_bytes);
1224                 if (diff < td_max_bs(td))
1225                         return 0;
1226
1227                 if (fio_files_done(td))
1228                         return 0;
1229
1230                 return 1;
1231         }
1232
1233         return 0;
1234 }
1235
1236 static int exec_string(struct thread_options *o, const char *string, const char *mode)
1237 {
1238         int ret, newlen = strlen(string) + strlen(o->name) + strlen(mode) + 9 + 1;
1239         char *str;
1240
1241         str = malloc(newlen);
1242         sprintf(str, "%s &> %s.%s.txt", string, o->name, mode);
1243
1244         log_info("%s : Saving output of %s in %s.%s.txt\n",o->name, mode, o->name, mode);
1245         ret = system(str);
1246         if (ret == -1)
1247                 log_err("fio: exec of cmd <%s> failed\n", str);
1248
1249         free(str);
1250         return ret;
1251 }
1252
1253 /*
1254  * Dry run to compute correct state of numberio for verification.
1255  */
1256 static uint64_t do_dry_run(struct thread_data *td)
1257 {
1258         uint64_t bytes_done[DDIR_RWDIR_CNT] = { 0, 0, 0 };
1259
1260         td_set_runstate(td, TD_RUNNING);
1261
1262         while ((td->o.read_iolog_file && !flist_empty(&td->io_log_list)) ||
1263                 (!flist_empty(&td->trim_list)) || !io_complete_bytes_exceeded(td)) {
1264                 struct io_u *io_u;
1265                 int ret;
1266
1267                 if (td->terminate || td->done)
1268                         break;
1269
1270                 io_u = get_io_u(td);
1271                 if (!io_u)
1272                         break;
1273
1274                 io_u->flags |= IO_U_F_FLIGHT;
1275                 io_u->error = 0;
1276                 io_u->resid = 0;
1277                 if (ddir_rw(acct_ddir(io_u)))
1278                         td->io_issues[acct_ddir(io_u)]++;
1279                 if (ddir_rw(io_u->ddir)) {
1280                         io_u_mark_depth(td, 1);
1281                         td->ts.total_io_u[io_u->ddir]++;
1282                 }
1283
1284                 if (td_write(td) && io_u->ddir == DDIR_WRITE &&
1285                     td->o.do_verify &&
1286                     td->o.verify != VERIFY_NONE &&
1287                     !td->o.experimental_verify)
1288                         log_io_piece(td, io_u);
1289
1290                 ret = io_u_sync_complete(td, io_u, bytes_done);
1291                 (void) ret;
1292         }
1293
1294         return bytes_done[DDIR_WRITE] + bytes_done[DDIR_TRIM];
1295 }
1296
1297 /*
1298  * Entry point for the thread based jobs. The process based jobs end up
1299  * here as well, after a little setup.
1300  */
1301 static void *thread_main(void *data)
1302 {
1303         unsigned long long elapsed;
1304         struct thread_data *td = data;
1305         struct thread_options *o = &td->o;
1306         pthread_condattr_t attr;
1307         int clear_state;
1308         int ret;
1309
1310         if (!o->use_thread) {
1311                 setsid();
1312                 td->pid = getpid();
1313         } else
1314                 td->pid = gettid();
1315
1316         fio_local_clock_init(o->use_thread);
1317
1318         dprint(FD_PROCESS, "jobs pid=%d started\n", (int) td->pid);
1319
1320         if (is_backend)
1321                 fio_server_send_start(td);
1322
1323         INIT_FLIST_HEAD(&td->io_log_list);
1324         INIT_FLIST_HEAD(&td->io_hist_list);
1325         INIT_FLIST_HEAD(&td->verify_list);
1326         INIT_FLIST_HEAD(&td->trim_list);
1327         INIT_FLIST_HEAD(&td->next_rand_list);
1328         pthread_mutex_init(&td->io_u_lock, NULL);
1329         td->io_hist_tree = RB_ROOT;
1330
1331         pthread_condattr_init(&attr);
1332         pthread_cond_init(&td->verify_cond, &attr);
1333         pthread_cond_init(&td->free_cond, &attr);
1334
1335         td_set_runstate(td, TD_INITIALIZED);
1336         dprint(FD_MUTEX, "up startup_mutex\n");
1337         fio_mutex_up(startup_mutex);
1338         dprint(FD_MUTEX, "wait on td->mutex\n");
1339         fio_mutex_down(td->mutex);
1340         dprint(FD_MUTEX, "done waiting on td->mutex\n");
1341
1342         /*
1343          * A new gid requires privilege, so we need to do this before setting
1344          * the uid.
1345          */
1346         if (o->gid != -1U && setgid(o->gid)) {
1347                 td_verror(td, errno, "setgid");
1348                 goto err;
1349         }
1350         if (o->uid != -1U && setuid(o->uid)) {
1351                 td_verror(td, errno, "setuid");
1352                 goto err;
1353         }
1354
1355         /*
1356          * If we have a gettimeofday() thread, make sure we exclude that
1357          * thread from this job
1358          */
1359         if (o->gtod_cpu)
1360                 fio_cpu_clear(&o->cpumask, o->gtod_cpu);
1361
1362         /*
1363          * Set affinity first, in case it has an impact on the memory
1364          * allocations.
1365          */
1366         if (fio_option_is_set(o, cpumask)) {
1367                 if (o->cpus_allowed_policy == FIO_CPUS_SPLIT) {
1368                         ret = fio_cpus_split(&o->cpumask, td->thread_number - 1);
1369                         if (!ret) {
1370                                 log_err("fio: no CPUs set\n");
1371                                 log_err("fio: Try increasing number of available CPUs\n");
1372                                 td_verror(td, EINVAL, "cpus_split");
1373                                 goto err;
1374                         }
1375                 }
1376                 ret = fio_setaffinity(td->pid, o->cpumask);
1377                 if (ret == -1) {
1378                         td_verror(td, errno, "cpu_set_affinity");
1379                         goto err;
1380                 }
1381         }
1382
1383 #ifdef CONFIG_LIBNUMA
1384         /* numa node setup */
1385         if (fio_option_is_set(o, numa_cpunodes) ||
1386             fio_option_is_set(o, numa_memnodes)) {
1387                 struct bitmask *mask;
1388
1389                 if (numa_available() < 0) {
1390                         td_verror(td, errno, "Does not support NUMA API\n");
1391                         goto err;
1392                 }
1393
1394                 if (fio_option_is_set(o, numa_cpunodes)) {
1395                         mask = numa_parse_nodestring(o->numa_cpunodes);
1396                         ret = numa_run_on_node_mask(mask);
1397                         numa_free_nodemask(mask);
1398                         if (ret == -1) {
1399                                 td_verror(td, errno, \
1400                                         "numa_run_on_node_mask failed\n");
1401                                 goto err;
1402                         }
1403                 }
1404
1405                 if (fio_option_is_set(o, numa_memnodes)) {
1406                         mask = NULL;
1407                         if (o->numa_memnodes)
1408                                 mask = numa_parse_nodestring(o->numa_memnodes);
1409
1410                         switch (o->numa_mem_mode) {
1411                         case MPOL_INTERLEAVE:
1412                                 numa_set_interleave_mask(mask);
1413                                 break;
1414                         case MPOL_BIND:
1415                                 numa_set_membind(mask);
1416                                 break;
1417                         case MPOL_LOCAL:
1418                                 numa_set_localalloc();
1419                                 break;
1420                         case MPOL_PREFERRED:
1421                                 numa_set_preferred(o->numa_mem_prefer_node);
1422                                 break;
1423                         case MPOL_DEFAULT:
1424                         default:
1425                                 break;
1426                         }
1427
1428                         if (mask)
1429                                 numa_free_nodemask(mask);
1430
1431                 }
1432         }
1433 #endif
1434
1435         if (fio_pin_memory(td))
1436                 goto err;
1437
1438         /*
1439          * May alter parameters that init_io_u() will use, so we need to
1440          * do this first.
1441          */
1442         if (init_iolog(td))
1443                 goto err;
1444
1445         if (init_io_u(td))
1446                 goto err;
1447
1448         if (o->verify_async && verify_async_init(td))
1449                 goto err;
1450
1451         if (fio_option_is_set(o, ioprio) ||
1452             fio_option_is_set(o, ioprio_class)) {
1453                 ret = ioprio_set(IOPRIO_WHO_PROCESS, 0, o->ioprio_class, o->ioprio);
1454                 if (ret == -1) {
1455                         td_verror(td, errno, "ioprio_set");
1456                         goto err;
1457                 }
1458         }
1459
1460         if (o->cgroup && cgroup_setup(td, cgroup_list, &cgroup_mnt))
1461                 goto err;
1462
1463         errno = 0;
1464         if (nice(o->nice) == -1 && errno != 0) {
1465                 td_verror(td, errno, "nice");
1466                 goto err;
1467         }
1468
1469         if (o->ioscheduler && switch_ioscheduler(td))
1470                 goto err;
1471
1472         if (!o->create_serialize && setup_files(td))
1473                 goto err;
1474
1475         if (td_io_init(td))
1476                 goto err;
1477
1478         if (init_random_map(td))
1479                 goto err;
1480
1481         if (o->exec_prerun && exec_string(o, o->exec_prerun, (const char *)"prerun"))
1482                 goto err;
1483
1484         if (o->pre_read) {
1485                 if (pre_read_files(td) < 0)
1486                         goto err;
1487         }
1488
1489         if (td->flags & TD_F_COMPRESS_LOG)
1490                 tp_init(&td->tp_data);
1491
1492         fio_verify_init(td);
1493
1494         fio_gettime(&td->epoch, NULL);
1495         fio_getrusage(&td->ru_start);
1496         clear_state = 0;
1497         while (keep_running(td)) {
1498                 uint64_t verify_bytes;
1499
1500                 fio_gettime(&td->start, NULL);
1501                 memcpy(&td->bw_sample_time, &td->start, sizeof(td->start));
1502                 memcpy(&td->iops_sample_time, &td->start, sizeof(td->start));
1503                 memcpy(&td->tv_cache, &td->start, sizeof(td->start));
1504
1505                 if (o->ratemin[DDIR_READ] || o->ratemin[DDIR_WRITE] ||
1506                                 o->ratemin[DDIR_TRIM]) {
1507                         memcpy(&td->lastrate[DDIR_READ], &td->bw_sample_time,
1508                                                 sizeof(td->bw_sample_time));
1509                         memcpy(&td->lastrate[DDIR_WRITE], &td->bw_sample_time,
1510                                                 sizeof(td->bw_sample_time));
1511                         memcpy(&td->lastrate[DDIR_TRIM], &td->bw_sample_time,
1512                                                 sizeof(td->bw_sample_time));
1513                 }
1514
1515                 if (clear_state)
1516                         clear_io_state(td);
1517
1518                 prune_io_piece_log(td);
1519
1520                 if (td->o.verify_only && (td_write(td) || td_rw(td)))
1521                         verify_bytes = do_dry_run(td);
1522                 else
1523                         verify_bytes = do_io(td);
1524
1525                 clear_state = 1;
1526
1527                 fio_mutex_down(stat_mutex);
1528                 if (td_read(td) && td->io_bytes[DDIR_READ]) {
1529                         elapsed = mtime_since_now(&td->start);
1530                         td->ts.runtime[DDIR_READ] += elapsed;
1531                 }
1532                 if (td_write(td) && td->io_bytes[DDIR_WRITE]) {
1533                         elapsed = mtime_since_now(&td->start);
1534                         td->ts.runtime[DDIR_WRITE] += elapsed;
1535                 }
1536                 if (td_trim(td) && td->io_bytes[DDIR_TRIM]) {
1537                         elapsed = mtime_since_now(&td->start);
1538                         td->ts.runtime[DDIR_TRIM] += elapsed;
1539                 }
1540                 fio_gettime(&td->start, NULL);
1541                 fio_mutex_up(stat_mutex);
1542
1543                 if (td->error || td->terminate)
1544                         break;
1545
1546                 if (!o->do_verify ||
1547                     o->verify == VERIFY_NONE ||
1548                     (td->io_ops->flags & FIO_UNIDIR))
1549                         continue;
1550
1551                 clear_io_state(td);
1552
1553                 fio_gettime(&td->start, NULL);
1554
1555                 do_verify(td, verify_bytes);
1556
1557                 fio_mutex_down(stat_mutex);
1558                 td->ts.runtime[DDIR_READ] += mtime_since_now(&td->start);
1559                 fio_gettime(&td->start, NULL);
1560                 fio_mutex_up(stat_mutex);
1561
1562                 if (td->error || td->terminate)
1563                         break;
1564         }
1565
1566         update_rusage_stat(td);
1567         td->ts.total_run_time = mtime_since_now(&td->epoch);
1568         td->ts.io_bytes[DDIR_READ] = td->io_bytes[DDIR_READ];
1569         td->ts.io_bytes[DDIR_WRITE] = td->io_bytes[DDIR_WRITE];
1570         td->ts.io_bytes[DDIR_TRIM] = td->io_bytes[DDIR_TRIM];
1571
1572         if (td->o.verify_state_save && !(td->flags & TD_F_VSTATE_SAVED) &&
1573             (td->o.verify != VERIFY_NONE && td_write(td))) {
1574                 struct all_io_list *state;
1575                 size_t sz;
1576
1577                 state = get_all_io_list(td->thread_number, &sz);
1578                 if (state) {
1579                         __verify_save_state(state, "local");
1580                         free(state);
1581                 }
1582         }
1583
1584         fio_unpin_memory(td);
1585
1586         fio_writeout_logs(td);
1587
1588         if (td->flags & TD_F_COMPRESS_LOG)
1589                 tp_exit(&td->tp_data);
1590
1591         if (o->exec_postrun)
1592                 exec_string(o, o->exec_postrun, (const char *)"postrun");
1593
1594         if (exitall_on_terminate)
1595                 fio_terminate_threads(td->groupid);
1596
1597 err:
1598         if (td->error)
1599                 log_info("fio: pid=%d, err=%d/%s\n", (int) td->pid, td->error,
1600                                                         td->verror);
1601
1602         if (o->verify_async)
1603                 verify_async_exit(td);
1604
1605         close_and_free_files(td);
1606         cleanup_io_u(td);
1607         close_ioengine(td);
1608         cgroup_shutdown(td, &cgroup_mnt);
1609         verify_free_state(td);
1610
1611         if (fio_option_is_set(o, cpumask)) {
1612                 ret = fio_cpuset_exit(&o->cpumask);
1613                 if (ret)
1614                         td_verror(td, ret, "fio_cpuset_exit");
1615         }
1616
1617         /*
1618          * do this very late, it will log file closing as well
1619          */
1620         if (o->write_iolog_file)
1621                 write_iolog_close(td);
1622
1623         fio_mutex_remove(td->mutex);
1624         td->mutex = NULL;
1625
1626         td_set_runstate(td, TD_EXITED);
1627
1628         /*
1629          * Do this last after setting our runstate to exited, so we
1630          * know that the stat thread is signaled.
1631          */
1632         check_update_rusage(td);
1633
1634         return (void *) (uintptr_t) td->error;
1635 }
1636
1637
1638 /*
1639  * We cannot pass the td data into a forked process, so attach the td and
1640  * pass it to the thread worker.
1641  */
1642 static int fork_main(int shmid, int offset)
1643 {
1644         struct thread_data *td;
1645         void *data, *ret;
1646
1647 #if !defined(__hpux) && !defined(CONFIG_NO_SHM)
1648         data = shmat(shmid, NULL, 0);
1649         if (data == (void *) -1) {
1650                 int __err = errno;
1651
1652                 perror("shmat");
1653                 return __err;
1654         }
1655 #else
1656         /*
1657          * HP-UX inherits shm mappings?
1658          */
1659         data = threads;
1660 #endif
1661
1662         td = data + offset * sizeof(struct thread_data);
1663         ret = thread_main(td);
1664         shmdt(data);
1665         return (int) (uintptr_t) ret;
1666 }
1667
1668 static void dump_td_info(struct thread_data *td)
1669 {
1670         log_err("fio: job '%s' hasn't exited in %lu seconds, it appears to "
1671                 "be stuck. Doing forceful exit of this job.\n", td->o.name,
1672                         (unsigned long) time_since_now(&td->terminate_time));
1673 }
1674
1675 /*
1676  * Run over the job map and reap the threads that have exited, if any.
1677  */
1678 static void reap_threads(unsigned int *nr_running, unsigned int *t_rate,
1679                          unsigned int *m_rate)
1680 {
1681         struct thread_data *td;
1682         unsigned int cputhreads, realthreads, pending;
1683         int i, status, ret;
1684
1685         /*
1686          * reap exited threads (TD_EXITED -> TD_REAPED)
1687          */
1688         realthreads = pending = cputhreads = 0;
1689         for_each_td(td, i) {
1690                 int flags = 0;
1691
1692                 /*
1693                  * ->io_ops is NULL for a thread that has closed its
1694                  * io engine
1695                  */
1696                 if (td->io_ops && !strcmp(td->io_ops->name, "cpuio"))
1697                         cputhreads++;
1698                 else
1699                         realthreads++;
1700
1701                 if (!td->pid) {
1702                         pending++;
1703                         continue;
1704                 }
1705                 if (td->runstate == TD_REAPED)
1706                         continue;
1707                 if (td->o.use_thread) {
1708                         if (td->runstate == TD_EXITED) {
1709                                 td_set_runstate(td, TD_REAPED);
1710                                 goto reaped;
1711                         }
1712                         continue;
1713                 }
1714
1715                 flags = WNOHANG;
1716                 if (td->runstate == TD_EXITED)
1717                         flags = 0;
1718
1719                 /*
1720                  * check if someone quit or got killed in an unusual way
1721                  */
1722                 ret = waitpid(td->pid, &status, flags);
1723                 if (ret < 0) {
1724                         if (errno == ECHILD) {
1725                                 log_err("fio: pid=%d disappeared %d\n",
1726                                                 (int) td->pid, td->runstate);
1727                                 td->sig = ECHILD;
1728                                 td_set_runstate(td, TD_REAPED);
1729                                 goto reaped;
1730                         }
1731                         perror("waitpid");
1732                 } else if (ret == td->pid) {
1733                         if (WIFSIGNALED(status)) {
1734                                 int sig = WTERMSIG(status);
1735
1736                                 if (sig != SIGTERM && sig != SIGUSR2)
1737                                         log_err("fio: pid=%d, got signal=%d\n",
1738                                                         (int) td->pid, sig);
1739                                 td->sig = sig;
1740                                 td_set_runstate(td, TD_REAPED);
1741                                 goto reaped;
1742                         }
1743                         if (WIFEXITED(status)) {
1744                                 if (WEXITSTATUS(status) && !td->error)
1745                                         td->error = WEXITSTATUS(status);
1746
1747                                 td_set_runstate(td, TD_REAPED);
1748                                 goto reaped;
1749                         }
1750                 }
1751
1752                 /*
1753                  * If the job is stuck, do a forceful timeout of it and
1754                  * move on.
1755                  */
1756                 if (td->terminate &&
1757                     time_since_now(&td->terminate_time) >= FIO_REAP_TIMEOUT) {
1758                         dump_td_info(td);
1759                         td_set_runstate(td, TD_REAPED);
1760                         goto reaped;
1761                 }
1762
1763                 /*
1764                  * thread is not dead, continue
1765                  */
1766                 pending++;
1767                 continue;
1768 reaped:
1769                 (*nr_running)--;
1770                 (*m_rate) -= ddir_rw_sum(td->o.ratemin);
1771                 (*t_rate) -= ddir_rw_sum(td->o.rate);
1772                 if (!td->pid)
1773                         pending--;
1774
1775                 if (td->error)
1776                         exit_value++;
1777
1778                 done_secs += mtime_since_now(&td->epoch) / 1000;
1779                 profile_td_exit(td);
1780         }
1781
1782         if (*nr_running == cputhreads && !pending && realthreads)
1783                 fio_terminate_threads(TERMINATE_ALL);
1784 }
1785
1786 static int __check_trigger_file(void)
1787 {
1788         struct stat sb;
1789
1790         if (!trigger_file)
1791                 return 0;
1792
1793         if (stat(trigger_file, &sb))
1794                 return 0;
1795
1796         if (unlink(trigger_file) < 0)
1797                 log_err("fio: failed to unlink %s: %s\n", trigger_file,
1798                                                         strerror(errno));
1799
1800         return 1;
1801 }
1802
1803 static int trigger_timedout(void)
1804 {
1805         if (trigger_timeout)
1806                 return time_since_genesis() >= trigger_timeout;
1807
1808         return 0;
1809 }
1810
1811 void exec_trigger(const char *cmd)
1812 {
1813         int ret;
1814
1815         if (!cmd)
1816                 return;
1817
1818         ret = system(cmd);
1819         if (ret == -1)
1820                 log_err("fio: failed executing %s trigger\n", cmd);
1821 }
1822
1823 void check_trigger_file(void)
1824 {
1825         if (__check_trigger_file() || trigger_timedout()) {
1826                 if (nr_clients)
1827                         fio_clients_send_trigger(trigger_remote_cmd);
1828                 else {
1829                         verify_save_state();
1830                         fio_terminate_threads(TERMINATE_ALL);
1831                         exec_trigger(trigger_cmd);
1832                 }
1833         }
1834 }
1835
1836 static int fio_verify_load_state(struct thread_data *td)
1837 {
1838         int ret;
1839
1840         if (!td->o.verify_state)
1841                 return 0;
1842
1843         if (is_backend) {
1844                 void *data;
1845
1846                 ret = fio_server_get_verify_state(td->o.name,
1847                                         td->thread_number - 1, &data);
1848                 if (!ret)
1849                         verify_convert_assign_state(td, data);
1850         } else
1851                 ret = verify_load_state(td, "local");
1852
1853         return ret;
1854 }
1855
1856 static void do_usleep(unsigned int usecs)
1857 {
1858         check_for_running_stats();
1859         check_trigger_file();
1860         usleep(usecs);
1861 }
1862
1863 /*
1864  * Main function for kicking off and reaping jobs, as needed.
1865  */
1866 static void run_threads(void)
1867 {
1868         struct thread_data *td;
1869         unsigned int i, todo, nr_running, m_rate, t_rate, nr_started;
1870         uint64_t spent;
1871
1872         if (fio_gtod_offload && fio_start_gtod_thread())
1873                 return;
1874
1875         fio_idle_prof_init();
1876
1877         set_sig_handlers();
1878
1879         nr_thread = nr_process = 0;
1880         for_each_td(td, i) {
1881                 if (td->o.use_thread)
1882                         nr_thread++;
1883                 else
1884                         nr_process++;
1885         }
1886
1887         if (output_format == FIO_OUTPUT_NORMAL) {
1888                 log_info("Starting ");
1889                 if (nr_thread)
1890                         log_info("%d thread%s", nr_thread,
1891                                                 nr_thread > 1 ? "s" : "");
1892                 if (nr_process) {
1893                         if (nr_thread)
1894                                 log_info(" and ");
1895                         log_info("%d process%s", nr_process,
1896                                                 nr_process > 1 ? "es" : "");
1897                 }
1898                 log_info("\n");
1899                 log_info_flush();
1900         }
1901
1902         todo = thread_number;
1903         nr_running = 0;
1904         nr_started = 0;
1905         m_rate = t_rate = 0;
1906
1907         for_each_td(td, i) {
1908                 print_status_init(td->thread_number - 1);
1909
1910                 if (!td->o.create_serialize)
1911                         continue;
1912
1913                 if (fio_verify_load_state(td))
1914                         goto reap;
1915
1916                 /*
1917                  * do file setup here so it happens sequentially,
1918                  * we don't want X number of threads getting their
1919                  * client data interspersed on disk
1920                  */
1921                 if (setup_files(td)) {
1922 reap:
1923                         exit_value++;
1924                         if (td->error)
1925                                 log_err("fio: pid=%d, err=%d/%s\n",
1926                                         (int) td->pid, td->error, td->verror);
1927                         td_set_runstate(td, TD_REAPED);
1928                         todo--;
1929                 } else {
1930                         struct fio_file *f;
1931                         unsigned int j;
1932
1933                         /*
1934                          * for sharing to work, each job must always open
1935                          * its own files. so close them, if we opened them
1936                          * for creation
1937                          */
1938                         for_each_file(td, f, j) {
1939                                 if (fio_file_open(f))
1940                                         td_io_close_file(td, f);
1941                         }
1942                 }
1943         }
1944
1945         /* start idle threads before io threads start to run */
1946         fio_idle_prof_start();
1947
1948         set_genesis_time();
1949
1950         while (todo) {
1951                 struct thread_data *map[REAL_MAX_JOBS];
1952                 struct timeval this_start;
1953                 int this_jobs = 0, left;
1954
1955                 /*
1956                  * create threads (TD_NOT_CREATED -> TD_CREATED)
1957                  */
1958                 for_each_td(td, i) {
1959                         if (td->runstate != TD_NOT_CREATED)
1960                                 continue;
1961
1962                         /*
1963                          * never got a chance to start, killed by other
1964                          * thread for some reason
1965                          */
1966                         if (td->terminate) {
1967                                 todo--;
1968                                 continue;
1969                         }
1970
1971                         if (td->o.start_delay) {
1972                                 spent = utime_since_genesis();
1973
1974                                 if (td->o.start_delay > spent)
1975                                         continue;
1976                         }
1977
1978                         if (td->o.stonewall && (nr_started || nr_running)) {
1979                                 dprint(FD_PROCESS, "%s: stonewall wait\n",
1980                                                         td->o.name);
1981                                 break;
1982                         }
1983
1984                         init_disk_util(td);
1985
1986                         td->rusage_sem = fio_mutex_init(FIO_MUTEX_LOCKED);
1987                         td->update_rusage = 0;
1988
1989                         /*
1990                          * Set state to created. Thread will transition
1991                          * to TD_INITIALIZED when it's done setting up.
1992                          */
1993                         td_set_runstate(td, TD_CREATED);
1994                         map[this_jobs++] = td;
1995                         nr_started++;
1996
1997                         if (td->o.use_thread) {
1998                                 int ret;
1999
2000                                 dprint(FD_PROCESS, "will pthread_create\n");
2001                                 ret = pthread_create(&td->thread, NULL,
2002                                                         thread_main, td);
2003                                 if (ret) {
2004                                         log_err("pthread_create: %s\n",
2005                                                         strerror(ret));
2006                                         nr_started--;
2007                                         break;
2008                                 }
2009                                 ret = pthread_detach(td->thread);
2010                                 if (ret)
2011                                         log_err("pthread_detach: %s",
2012                                                         strerror(ret));
2013                         } else {
2014                                 pid_t pid;
2015                                 dprint(FD_PROCESS, "will fork\n");
2016                                 pid = fork();
2017                                 if (!pid) {
2018                                         int ret = fork_main(shm_id, i);
2019
2020                                         _exit(ret);
2021                                 } else if (i == fio_debug_jobno)
2022                                         *fio_debug_jobp = pid;
2023                         }
2024                         dprint(FD_MUTEX, "wait on startup_mutex\n");
2025                         if (fio_mutex_down_timeout(startup_mutex, 10)) {
2026                                 log_err("fio: job startup hung? exiting.\n");
2027                                 fio_terminate_threads(TERMINATE_ALL);
2028                                 fio_abort = 1;
2029                                 nr_started--;
2030                                 break;
2031                         }
2032                         dprint(FD_MUTEX, "done waiting on startup_mutex\n");
2033                 }
2034
2035                 /*
2036                  * Wait for the started threads to transition to
2037                  * TD_INITIALIZED.
2038                  */
2039                 fio_gettime(&this_start, NULL);
2040                 left = this_jobs;
2041                 while (left && !fio_abort) {
2042                         if (mtime_since_now(&this_start) > JOB_START_TIMEOUT)
2043                                 break;
2044
2045                         do_usleep(100000);
2046
2047                         for (i = 0; i < this_jobs; i++) {
2048                                 td = map[i];
2049                                 if (!td)
2050                                         continue;
2051                                 if (td->runstate == TD_INITIALIZED) {
2052                                         map[i] = NULL;
2053                                         left--;
2054                                 } else if (td->runstate >= TD_EXITED) {
2055                                         map[i] = NULL;
2056                                         left--;
2057                                         todo--;
2058                                         nr_running++; /* work-around... */
2059                                 }
2060                         }
2061                 }
2062
2063                 if (left) {
2064                         log_err("fio: %d job%s failed to start\n", left,
2065                                         left > 1 ? "s" : "");
2066                         for (i = 0; i < this_jobs; i++) {
2067                                 td = map[i];
2068                                 if (!td)
2069                                         continue;
2070                                 kill(td->pid, SIGTERM);
2071                         }
2072                         break;
2073                 }
2074
2075                 /*
2076                  * start created threads (TD_INITIALIZED -> TD_RUNNING).
2077                  */
2078                 for_each_td(td, i) {
2079                         if (td->runstate != TD_INITIALIZED)
2080                                 continue;
2081
2082                         if (in_ramp_time(td))
2083                                 td_set_runstate(td, TD_RAMP);
2084                         else
2085                                 td_set_runstate(td, TD_RUNNING);
2086                         nr_running++;
2087                         nr_started--;
2088                         m_rate += ddir_rw_sum(td->o.ratemin);
2089                         t_rate += ddir_rw_sum(td->o.rate);
2090                         todo--;
2091                         fio_mutex_up(td->mutex);
2092                 }
2093
2094                 reap_threads(&nr_running, &t_rate, &m_rate);
2095
2096                 if (todo)
2097                         do_usleep(100000);
2098         }
2099
2100         while (nr_running) {
2101                 reap_threads(&nr_running, &t_rate, &m_rate);
2102                 do_usleep(10000);
2103         }
2104
2105         fio_idle_prof_stop();
2106
2107         update_io_ticks();
2108 }
2109
2110 static void wait_for_helper_thread_exit(void)
2111 {
2112         void *ret;
2113
2114         helper_exit = 1;
2115         pthread_cond_signal(&helper_cond);
2116         pthread_join(helper_thread, &ret);
2117 }
2118
2119 static void free_disk_util(void)
2120 {
2121         disk_util_prune_entries();
2122
2123         pthread_cond_destroy(&helper_cond);
2124 }
2125
2126 static void *helper_thread_main(void *data)
2127 {
2128         int ret = 0;
2129
2130         fio_mutex_up(startup_mutex);
2131
2132         while (!ret) {
2133                 uint64_t sec = DISK_UTIL_MSEC / 1000;
2134                 uint64_t nsec = (DISK_UTIL_MSEC % 1000) * 1000000;
2135                 struct timespec ts;
2136                 struct timeval tv;
2137
2138                 gettimeofday(&tv, NULL);
2139                 ts.tv_sec = tv.tv_sec + sec;
2140                 ts.tv_nsec = (tv.tv_usec * 1000) + nsec;
2141
2142                 if (ts.tv_nsec >= 1000000000ULL) {
2143                         ts.tv_nsec -= 1000000000ULL;
2144                         ts.tv_sec++;
2145                 }
2146
2147                 pthread_cond_timedwait(&helper_cond, &helper_lock, &ts);
2148
2149                 ret = update_io_ticks();
2150
2151                 if (helper_do_stat) {
2152                         helper_do_stat = 0;
2153                         __show_running_run_stats();
2154                 }
2155
2156                 if (!is_backend)
2157                         print_thread_status();
2158         }
2159
2160         return NULL;
2161 }
2162
2163 static int create_helper_thread(void)
2164 {
2165         int ret;
2166
2167         setup_disk_util();
2168
2169         pthread_cond_init(&helper_cond, NULL);
2170         pthread_mutex_init(&helper_lock, NULL);
2171
2172         ret = pthread_create(&helper_thread, NULL, helper_thread_main, NULL);
2173         if (ret) {
2174                 log_err("Can't create helper thread: %s\n", strerror(ret));
2175                 return 1;
2176         }
2177
2178         dprint(FD_MUTEX, "wait on startup_mutex\n");
2179         fio_mutex_down(startup_mutex);
2180         dprint(FD_MUTEX, "done waiting on startup_mutex\n");
2181         return 0;
2182 }
2183
2184 int fio_backend(void)
2185 {
2186         struct thread_data *td;
2187         int i;
2188
2189         if (exec_profile) {
2190                 if (load_profile(exec_profile))
2191                         return 1;
2192                 free(exec_profile);
2193                 exec_profile = NULL;
2194         }
2195         if (!thread_number)
2196                 return 0;
2197
2198         if (write_bw_log) {
2199                 struct log_params p = {
2200                         .log_type = IO_LOG_TYPE_BW,
2201                 };
2202
2203                 setup_log(&agg_io_log[DDIR_READ], &p, "agg-read_bw.log");
2204                 setup_log(&agg_io_log[DDIR_WRITE], &p, "agg-write_bw.log");
2205                 setup_log(&agg_io_log[DDIR_TRIM], &p, "agg-trim_bw.log");
2206         }
2207
2208         startup_mutex = fio_mutex_init(FIO_MUTEX_LOCKED);
2209         if (startup_mutex == NULL)
2210                 return 1;
2211
2212         set_genesis_time();
2213         stat_init();
2214         create_helper_thread();
2215
2216         cgroup_list = smalloc(sizeof(*cgroup_list));
2217         INIT_FLIST_HEAD(cgroup_list);
2218
2219         run_threads();
2220
2221         wait_for_helper_thread_exit();
2222
2223         if (!fio_abort) {
2224                 __show_run_stats();
2225                 if (write_bw_log) {
2226                         for (i = 0; i < DDIR_RWDIR_CNT; i++) {
2227                                 struct io_log *log = agg_io_log[i];
2228
2229                                 flush_log(log);
2230                                 free_log(log);
2231                         }
2232                 }
2233         }
2234
2235         for_each_td(td, i) {
2236                 fio_options_free(td);
2237                 if (td->rusage_sem) {
2238                         fio_mutex_remove(td->rusage_sem);
2239                         td->rusage_sem = NULL;
2240                 }
2241         }
2242
2243         free_disk_util();
2244         cgroup_kill(cgroup_list);
2245         sfree(cgroup_list);
2246         sfree(cgroup_mnt);
2247
2248         fio_mutex_remove(startup_mutex);
2249         stat_exit();
2250         return exit_value;
2251 }