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 #include <math.h>
  39
  40 #include "fio.h"
  41 #ifndef FIO_NO_HAVE_SHM_H
  42 #include <sys/shm.h>
  43 #endif
  44 #include "hash.h"
  45 #include "smalloc.h"
  46 #include "verify.h"
  47 #include "trim.h"
  48 #include "diskutil.h"
  49 #include "cgroup.h"
  50 #include "profile.h"
  51 #include "lib/rand.h"
  52 #include "lib/memalign.h"
  53 #include "server.h"
  54 #include "lib/getrusage.h"
  55 #include "idletime.h"
  56 #include "err.h"
  57 #include "workqueue.h"
  58 #include "lib/mountcheck.h"
  59 #include "rate-submit.h"
  60 #include "helper_thread.h"
  61
  62 static struct fio_mutex *startup_mutex;
  63 static struct flist_head *cgroup_list;
  64 static char *cgroup_mnt;
  65 static int exit_value;
  66 static volatile int fio_abort;
  67 static unsigned int nr_process = 0;
  68 static unsigned int nr_thread = 0;
  69
  70 struct io_log *agg_io_log[DDIR_RWDIR_CNT];
  71
  72 int groupid = 0;
  73 unsigned int thread_number = 0;
  74 unsigned int stat_number = 0;
  75 int shm_id = 0;
  76 int temp_stall_ts;
  77 unsigned long done_secs = 0;
  78
  79 #define JOB_START_TIMEOUT       (5 * 1000)
  80
  81 static void sig_int(int sig)
  82 {
  83         if (threads) {
  84                 if (is_backend)
  85                         fio_server_got_signal(sig);
  86                 else {
  87                         log_info("\nfio: terminating on signal %d\n", sig);
  88                         log_info_flush();
  89                         exit_value = 128;
  90                 }
  91
  92                 fio_terminate_threads(TERMINATE_ALL);
  93         }
  94 }
  95
  96 void sig_show_status(int sig)
  97 {
  98         show_running_run_stats();
  99 }
 100
 101 static void set_sig_handlers(void)
 102 {
 103         struct sigaction act;
 104
 105         memset(&act, 0, sizeof(act));
 106         act.sa_handler = sig_int;
 107         act.sa_flags = SA_RESTART;
 108         sigaction(SIGINT, &act, NULL);
 109
 110         memset(&act, 0, sizeof(act));
 111         act.sa_handler = sig_int;
 112         act.sa_flags = SA_RESTART;
 113         sigaction(SIGTERM, &act, NULL);
 114
 115 /* Windows uses SIGBREAK as a quit signal from other applications */
 116 #ifdef WIN32
 117         memset(&act, 0, sizeof(act));
 118         act.sa_handler = sig_int;
 119         act.sa_flags = SA_RESTART;
 120         sigaction(SIGBREAK, &act, NULL);
 121 #endif
 122
 123         memset(&act, 0, sizeof(act));
 124         act.sa_handler = sig_show_status;
 125         act.sa_flags = SA_RESTART;
 126         sigaction(SIGUSR1, &act, NULL);
 127
 128         if (is_backend) {
 129                 memset(&act, 0, sizeof(act));
 130                 act.sa_handler = sig_int;
 131                 act.sa_flags = SA_RESTART;
 132                 sigaction(SIGPIPE, &act, NULL);
 133         }
 134 }
 135
 136 /*
 137  * Check if we are above the minimum rate given.
 138  */
 139 static bool __check_min_rate(struct thread_data *td, struct timeval *now,
 140                              enum fio_ddir ddir)
 141 {
 142         unsigned long long bytes = 0;
 143         unsigned long iops = 0;
 144         unsigned long spent;
 145         unsigned long rate;
 146         unsigned int ratemin = 0;
 147         unsigned int rate_iops = 0;
 148         unsigned int rate_iops_min = 0;
 149
 150         assert(ddir_rw(ddir));
 151
 152         if (!td->o.ratemin[ddir] && !td->o.rate_iops_min[ddir])
 153                 return false;
 154
 155         /*
 156          * allow a 2 second settle period in the beginning
 157          */
 158         if (mtime_since(&td->start, now) < 2000)
 159                 return false;
 160
 161         iops += td->this_io_blocks[ddir];
 162         bytes += td->this_io_bytes[ddir];
 163         ratemin += td->o.ratemin[ddir];
 164         rate_iops += td->o.rate_iops[ddir];
 165         rate_iops_min += td->o.rate_iops_min[ddir];
 166
 167         /*
 168          * if rate blocks is set, sample is running
 169          */
 170         if (td->rate_bytes[ddir] || td->rate_blocks[ddir]) {
 171                 spent = mtime_since(&td->lastrate[ddir], now);
 172                 if (spent < td->o.ratecycle)
 173                         return false;
 174
 175                 if (td->o.rate[ddir] || td->o.ratemin[ddir]) {
 176                         /*
 177                          * check bandwidth specified rate
 178                          */
 179                         if (bytes < td->rate_bytes[ddir]) {
 180                                 log_err("%s: rate_min=%uB/s not met, only transferred %lluB\n",
 181                                         td->o.name, ratemin, bytes);
 182                                 return true;
 183                         } else {
 184                                 if (spent)
 185                                         rate = ((bytes - td->rate_bytes[ddir]) * 1000) / spent;
 186                                 else
 187                                         rate = 0;
 188
 189                                 if (rate < ratemin ||
 190                                     bytes < td->rate_bytes[ddir]) {
 191                                         log_err("%s: rate_min=%uB/s not met, got %luB/s\n",
 192                                                 td->o.name, ratemin, rate);
 193                                         return true;
 194                                 }
 195                         }
 196                 } else {
 197                         /*
 198                          * checks iops specified rate
 199                          */
 200                         if (iops < rate_iops) {
 201                                 log_err("%s: rate_iops_min=%u not met, only performed %lu IOs\n",
 202                                                 td->o.name, rate_iops, iops);
 203                                 return true;
 204                         } else {
 205                                 if (spent)
 206                                         rate = ((iops - td->rate_blocks[ddir]) * 1000) / spent;
 207                                 else
 208                                         rate = 0;
 209
 210                                 if (rate < rate_iops_min ||
 211                                     iops < td->rate_blocks[ddir]) {
 212                                         log_err("%s: rate_iops_min=%u not met, got %lu IOPS\n",
 213                                                 td->o.name, rate_iops_min, rate);
 214                                         return true;
 215                                 }
 216                         }
 217                 }
 218         }
 219
 220         td->rate_bytes[ddir] = bytes;
 221         td->rate_blocks[ddir] = iops;
 222         memcpy(&td->lastrate[ddir], now, sizeof(*now));
 223         return false;
 224 }
 225
 226 static bool check_min_rate(struct thread_data *td, struct timeval *now)
 227 {
 228         bool ret = false;
 229
 230         if (td->bytes_done[DDIR_READ])
 231                 ret |= __check_min_rate(td, now, DDIR_READ);
 232         if (td->bytes_done[DDIR_WRITE])
 233                 ret |= __check_min_rate(td, now, DDIR_WRITE);
 234         if (td->bytes_done[DDIR_TRIM])
 235                 ret |= __check_min_rate(td, now, DDIR_TRIM);
 236
 237         return ret;
 238 }
 239
 240 /*
 241  * When job exits, we can cancel the in-flight IO if we are using async
 242  * io. Attempt to do so.
 243  */
 244 static void cleanup_pending_aio(struct thread_data *td)
 245 {
 246         int r;
 247
 248         /*
 249          * get immediately available events, if any
 250          */
 251         r = io_u_queued_complete(td, 0);
 252         if (r < 0)
 253                 return;
 254
 255         /*
 256          * now cancel remaining active events
 257          */
 258         if (td->io_ops->cancel) {
 259                 struct io_u *io_u;
 260                 int i;
 261
 262                 io_u_qiter(&td->io_u_all, io_u, i) {
 263                         if (io_u->flags & IO_U_F_FLIGHT) {
 264                                 r = td->io_ops->cancel(td, io_u);
 265                                 if (!r)
 266                                         put_io_u(td, io_u);
 267                         }
 268                 }
 269         }
 270
 271         if (td->cur_depth)
 272                 r = io_u_queued_complete(td, td->cur_depth);
 273 }
 274
 275 /*
 276  * Helper to handle the final sync of a file. Works just like the normal
 277  * io path, just does everything sync.
 278  */
 279 static bool fio_io_sync(struct thread_data *td, struct fio_file *f)
 280 {
 281         struct io_u *io_u = __get_io_u(td);
 282         int ret;
 283
 284         if (!io_u)
 285                 return true;
 286
 287         io_u->ddir = DDIR_SYNC;
 288         io_u->file = f;
 289
 290         if (td_io_prep(td, io_u)) {
 291                 put_io_u(td, io_u);
 292                 return true;
 293         }
 294
 295 requeue:
 296         ret = td_io_queue(td, io_u);
 297         if (ret < 0) {
 298                 td_verror(td, io_u->error, "td_io_queue");
 299                 put_io_u(td, io_u);
 300                 return true;
 301         } else if (ret == FIO_Q_QUEUED) {
 302                 if (td_io_commit(td))
 303                         return true;
 304                 if (io_u_queued_complete(td, 1) < 0)
 305                         return true;
 306         } else if (ret == FIO_Q_COMPLETED) {
 307                 if (io_u->error) {
 308                         td_verror(td, io_u->error, "td_io_queue");
 309                         return true;
 310                 }
 311
 312                 if (io_u_sync_complete(td, io_u) < 0)
 313                         return true;
 314         } else if (ret == FIO_Q_BUSY) {
 315                 if (td_io_commit(td))
 316                         return true;
 317                 goto requeue;
 318         }
 319
 320         return false;
 321 }
 322
 323 static int fio_file_fsync(struct thread_data *td, struct fio_file *f)
 324 {
 325         int ret;
 326
 327         if (fio_file_open(f))
 328                 return fio_io_sync(td, f);
 329
 330         if (td_io_open_file(td, f))
 331                 return 1;
 332
 333         ret = fio_io_sync(td, f);
 334         td_io_close_file(td, f);
 335         return ret;
 336 }
 337
 338 static inline void __update_tv_cache(struct thread_data *td)
 339 {
 340         fio_gettime(&td->tv_cache, NULL);
 341 }
 342
 343 static inline void update_tv_cache(struct thread_data *td)
 344 {
 345         if ((++td->tv_cache_nr & td->tv_cache_mask) == td->tv_cache_mask)
 346                 __update_tv_cache(td);
 347 }
 348
 349 static inline bool runtime_exceeded(struct thread_data *td, struct timeval *t)
 350 {
 351         if (in_ramp_time(td))
 352                 return false;
 353         if (!td->o.timeout)
 354                 return false;
 355         if (utime_since(&td->epoch, t) >= td->o.timeout)
 356                 return true;
 357
 358         return false;
 359 }
 360
 361 /*
 362  * We need to update the runtime consistently in ms, but keep a running
 363  * tally of the current elapsed time in microseconds for sub millisecond
 364  * updates.
 365  */
 366 static inline void update_runtime(struct thread_data *td,
 367                                   unsigned long long *elapsed_us,
 368                                   const enum fio_ddir ddir)
 369 {
 370         if (ddir == DDIR_WRITE && td_write(td) && td->o.verify_only)
 371                 return;
 372
 373         td->ts.runtime[ddir] -= (elapsed_us[ddir] + 999) / 1000;
 374         elapsed_us[ddir] += utime_since_now(&td->start);
 375         td->ts.runtime[ddir] += (elapsed_us[ddir] + 999) / 1000;
 376 }
 377
 378 static bool break_on_this_error(struct thread_data *td, enum fio_ddir ddir,
 379                                 int *retptr)
 380 {
 381         int ret = *retptr;
 382
 383         if (ret < 0 || td->error) {
 384                 int err = td->error;
 385                 enum error_type_bit eb;
 386
 387                 if (ret < 0)
 388                         err = -ret;
 389
 390                 eb = td_error_type(ddir, err);
 391                 if (!(td->o.continue_on_error & (1 << eb)))
 392                         return true;
 393
 394                 if (td_non_fatal_error(td, eb, err)) {
 395                         /*
 396                          * Continue with the I/Os in case of
 397                          * a non fatal error.
 398                          */
 399                         update_error_count(td, err);
 400                         td_clear_error(td);
 401                         *retptr = 0;
 402                         return false;
 403                 } else if (td->o.fill_device && err == ENOSPC) {
 404                         /*
 405                          * We expect to hit this error if
 406                          * fill_device option is set.
 407                          */
 408                         td_clear_error(td);
 409                         fio_mark_td_terminate(td);
 410                         return true;
 411                 } else {
 412                         /*
 413                          * Stop the I/O in case of a fatal
 414                          * error.
 415                          */
 416                         update_error_count(td, err);
 417                         return true;
 418                 }
 419         }
 420
 421         return false;
 422 }
 423
 424 static void check_update_rusage(struct thread_data *td)
 425 {
 426         if (td->update_rusage) {
 427                 td->update_rusage = 0;
 428                 update_rusage_stat(td);
 429                 fio_mutex_up(td->rusage_sem);
 430         }
 431 }
 432
 433 static int wait_for_completions(struct thread_data *td, struct timeval *time)
 434 {
 435         const int full = queue_full(td);
 436         int min_evts = 0;
 437         int ret;
 438
 439         if (td->flags & TD_F_REGROW_LOGS)
 440                 return io_u_quiesce(td);
 441
 442         /*
 443          * if the queue is full, we MUST reap at least 1 event
 444          */
 445         min_evts = min(td->o.iodepth_batch_complete_min, td->cur_depth);
 446         if ((full && !min_evts) || !td->o.iodepth_batch_complete_min)
 447                 min_evts = 1;
 448
 449         if (time && (__should_check_rate(td, DDIR_READ) ||
 450             __should_check_rate(td, DDIR_WRITE) ||
 451             __should_check_rate(td, DDIR_TRIM)))
 452                 fio_gettime(time, NULL);
 453
 454         do {
 455                 ret = io_u_queued_complete(td, min_evts);
 456                 if (ret < 0)
 457                         break;
 458         } while (full && (td->cur_depth > td->o.iodepth_low));
 459
 460         return ret;
 461 }
 462
 463 int io_queue_event(struct thread_data *td, struct io_u *io_u, int *ret,
 464                    enum fio_ddir ddir, uint64_t *bytes_issued, int from_verify,
 465                    struct timeval *comp_time)
 466 {
 467         int ret2;
 468
 469         switch (*ret) {
 470         case FIO_Q_COMPLETED:
 471                 if (io_u->error) {
 472                         *ret = -io_u->error;
 473                         clear_io_u(td, io_u);
 474                 } else if (io_u->resid) {
 475                         int bytes = io_u->xfer_buflen - io_u->resid;
 476                         struct fio_file *f = io_u->file;
 477
 478                         if (bytes_issued)
 479                                 *bytes_issued += bytes;
 480
 481                         if (!from_verify)
 482                                 trim_io_piece(td, io_u);
 483
 484                         /*
 485                          * zero read, fail
 486                          */
 487                         if (!bytes) {
 488                                 if (!from_verify)
 489                                         unlog_io_piece(td, io_u);
 490                                 td_verror(td, EIO, "full resid");
 491                                 put_io_u(td, io_u);
 492                                 break;
 493                         }
 494
 495                         io_u->xfer_buflen = io_u->resid;
 496                         io_u->xfer_buf += bytes;
 497                         io_u->offset += bytes;
 498
 499                         if (ddir_rw(io_u->ddir))
 500                                 td->ts.short_io_u[io_u->ddir]++;
 501
 502                         f = io_u->file;
 503                         if (io_u->offset == f->real_file_size)
 504                                 goto sync_done;
 505
 506                         requeue_io_u(td, &io_u);
 507                 } else {
 508 sync_done:
 509                         if (comp_time && (__should_check_rate(td, DDIR_READ) ||
 510                             __should_check_rate(td, DDIR_WRITE) ||
 511                             __should_check_rate(td, DDIR_TRIM)))
 512                                 fio_gettime(comp_time, NULL);
 513
 514                         *ret = io_u_sync_complete(td, io_u);
 515                         if (*ret < 0)
 516                                 break;
 517                 }
 518
 519                 if (td->flags & TD_F_REGROW_LOGS)
 520                         regrow_logs(td);
 521
 522                 /*
 523                  * when doing I/O (not when verifying),
 524                  * check for any errors that are to be ignored
 525                  */
 526                 if (!from_verify)
 527                         break;
 528
 529                 return 0;
 530         case FIO_Q_QUEUED:
 531                 /*
 532                  * if the engine doesn't have a commit hook,
 533                  * the io_u is really queued. if it does have such
 534                  * a hook, it has to call io_u_queued() itself.
 535                  */
 536                 if (td->io_ops->commit == NULL)
 537                         io_u_queued(td, io_u);
 538                 if (bytes_issued)
 539                         *bytes_issued += io_u->xfer_buflen;
 540                 break;
 541         case FIO_Q_BUSY:
 542                 if (!from_verify)
 543                         unlog_io_piece(td, io_u);
 544                 requeue_io_u(td, &io_u);
 545                 ret2 = td_io_commit(td);
 546                 if (ret2 < 0)
 547                         *ret = ret2;
 548                 break;
 549         default:
 550                 assert(*ret < 0);
 551                 td_verror(td, -(*ret), "td_io_queue");
 552                 break;
 553         }
 554
 555         if (break_on_this_error(td, ddir, ret))
 556                 return 1;
 557
 558         return 0;
 559 }
 560
 561 static inline bool io_in_polling(struct thread_data *td)
 562 {
 563         return !td->o.iodepth_batch_complete_min &&
 564                    !td->o.iodepth_batch_complete_max;
 565 }
 566 /*
 567  * Unlinks files from thread data fio_file structure
 568  */
 569 static int unlink_all_files(struct thread_data *td)
 570 {
 571         struct fio_file *f;
 572         unsigned int i;
 573         int ret = 0;
 574
 575         for_each_file(td, f, i) {
 576                 if (f->filetype != FIO_TYPE_FILE)
 577                         continue;
 578                 ret = td_io_unlink_file(td, f);
 579                 if (ret)
 580                         break;
 581         }
 582
 583         if (ret)
 584                 td_verror(td, ret, "unlink_all_files");
 585
 586         return ret;
 587 }
 588
 589 /*
 590  * The main verify engine. Runs over the writes we previously submitted,
 591  * reads the blocks back in, and checks the crc/md5 of the data.
 592  */
 593 static void do_verify(struct thread_data *td, uint64_t verify_bytes)
 594 {
 595         struct fio_file *f;
 596         struct io_u *io_u;
 597         int ret, min_events;
 598         unsigned int i;
 599
 600         dprint(FD_VERIFY, "starting loop\n");
 601
 602         /*
 603          * sync io first and invalidate cache, to make sure we really
 604          * read from disk.
 605          */
 606         for_each_file(td, f, i) {
 607                 if (!fio_file_open(f))
 608                         continue;
 609                 if (fio_io_sync(td, f))
 610                         break;
 611                 if (file_invalidate_cache(td, f))
 612                         break;
 613         }
 614
 615         check_update_rusage(td);
 616
 617         if (td->error)
 618                 return;
 619
 620         /*
 621          * verify_state needs to be reset before verification
 622          * proceeds so that expected random seeds match actual
 623          * random seeds in headers. The main loop will reset
 624          * all random number generators if randrepeat is set.
 625          */
 626         if (!td->o.rand_repeatable)
 627                 td_fill_verify_state_seed(td);
 628
 629         td_set_runstate(td, TD_VERIFYING);
 630
 631         io_u = NULL;
 632         while (!td->terminate) {
 633                 enum fio_ddir ddir;
 634                 int full;
 635
 636                 update_tv_cache(td);
 637                 check_update_rusage(td);
 638
 639                 if (runtime_exceeded(td, &td->tv_cache)) {
 640                         __update_tv_cache(td);
 641                         if (runtime_exceeded(td, &td->tv_cache)) {
 642                                 fio_mark_td_terminate(td);
 643                                 break;
 644                         }
 645                 }
 646
 647                 if (flow_threshold_exceeded(td))
 648                         continue;
 649
 650                 if (!td->o.experimental_verify) {
 651                         io_u = __get_io_u(td);
 652                         if (!io_u)
 653                                 break;
 654
 655                         if (get_next_verify(td, io_u)) {
 656                                 put_io_u(td, io_u);
 657                                 break;
 658                         }
 659
 660                         if (td_io_prep(td, io_u)) {
 661                                 put_io_u(td, io_u);
 662                                 break;
 663                         }
 664                 } else {
 665                         if (ddir_rw_sum(td->bytes_done) + td->o.rw_min_bs > verify_bytes)
 666                                 break;
 667
 668                         while ((io_u = get_io_u(td)) != NULL) {
 669                                 if (IS_ERR_OR_NULL(io_u)) {
 670                                         io_u = NULL;
 671                                         ret = FIO_Q_BUSY;
 672                                         goto reap;
 673                                 }
 674
 675                                 /*
 676                                  * We are only interested in the places where
 677                                  * we wrote or trimmed IOs. Turn those into
 678                                  * reads for verification purposes.
 679                                  */
 680                                 if (io_u->ddir == DDIR_READ) {
 681                                         /*
 682                                          * Pretend we issued it for rwmix
 683                                          * accounting
 684                                          */
 685                                         td->io_issues[DDIR_READ]++;
 686                                         put_io_u(td, io_u);
 687                                         continue;
 688                                 } else if (io_u->ddir == DDIR_TRIM) {
 689                                         io_u->ddir = DDIR_READ;
 690                                         io_u_set(td, io_u, IO_U_F_TRIMMED);
 691                                         break;
 692                                 } else if (io_u->ddir == DDIR_WRITE) {
 693                                         io_u->ddir = DDIR_READ;
 694                                         break;
 695                                 } else {
 696                                         put_io_u(td, io_u);
 697                                         continue;
 698                                 }
 699                         }
 700
 701                         if (!io_u)
 702                                 break;
 703                 }
 704
 705                 if (verify_state_should_stop(td, io_u)) {
 706                         put_io_u(td, io_u);
 707                         break;
 708                 }
 709
 710                 if (td->o.verify_async)
 711                         io_u->end_io = verify_io_u_async;
 712                 else
 713                         io_u->end_io = verify_io_u;
 714
 715                 ddir = io_u->ddir;
 716                 if (!td->o.disable_slat)
 717                         fio_gettime(&io_u->start_time, NULL);
 718
 719                 ret = td_io_queue(td, io_u);
 720
 721                 if (io_queue_event(td, io_u, &ret, ddir, NULL, 1, NULL))
 722                         break;
 723
 724                 /*
 725                  * if we can queue more, do so. but check if there are
 726                  * completed io_u's first. Note that we can get BUSY even
 727                  * without IO queued, if the system is resource starved.
 728                  */
 729 reap:
 730                 full = queue_full(td) || (ret == FIO_Q_BUSY && td->cur_depth);
 731                 if (full || io_in_polling(td))
 732                         ret = wait_for_completions(td, NULL);
 733
 734                 if (ret < 0)
 735                         break;
 736         }
 737
 738         check_update_rusage(td);
 739
 740         if (!td->error) {
 741                 min_events = td->cur_depth;
 742
 743                 if (min_events)
 744                         ret = io_u_queued_complete(td, min_events);
 745         } else
 746                 cleanup_pending_aio(td);
 747
 748         td_set_runstate(td, TD_RUNNING);
 749
 750         dprint(FD_VERIFY, "exiting loop\n");
 751 }
 752
 753 static bool exceeds_number_ios(struct thread_data *td)
 754 {
 755         unsigned long long number_ios;
 756
 757         if (!td->o.number_ios)
 758                 return false;
 759
 760         number_ios = ddir_rw_sum(td->io_blocks);
 761         number_ios += td->io_u_queued + td->io_u_in_flight;
 762
 763         return number_ios >= (td->o.number_ios * td->loops);
 764 }
 765
 766 static bool io_bytes_exceeded(struct thread_data *td, uint64_t *this_bytes)
 767 {
 768         unsigned long long bytes, limit;
 769
 770         if (td_rw(td))
 771                 bytes = this_bytes[DDIR_READ] + this_bytes[DDIR_WRITE];
 772         else if (td_write(td))
 773                 bytes = this_bytes[DDIR_WRITE];
 774         else if (td_read(td))
 775                 bytes = this_bytes[DDIR_READ];
 776         else
 777                 bytes = this_bytes[DDIR_TRIM];
 778
 779         if (td->o.io_limit)
 780                 limit = td->o.io_limit;
 781         else
 782                 limit = td->o.size;
 783
 784         limit *= td->loops;
 785         return bytes >= limit || exceeds_number_ios(td);
 786 }
 787
 788 static bool io_issue_bytes_exceeded(struct thread_data *td)
 789 {
 790         return io_bytes_exceeded(td, td->io_issue_bytes);
 791 }
 792
 793 static bool io_complete_bytes_exceeded(struct thread_data *td)
 794 {
 795         return io_bytes_exceeded(td, td->this_io_bytes);
 796 }
 797
 798 /*
 799  * used to calculate the next io time for rate control
 800  *
 801  */
 802 static long long usec_for_io(struct thread_data *td, enum fio_ddir ddir)
 803 {
 804         uint64_t secs, remainder, bps, bytes, iops;
 805
 806         assert(!(td->flags & TD_F_CHILD));
 807         bytes = td->rate_io_issue_bytes[ddir];
 808         bps = td->rate_bps[ddir];
 809
 810         if (td->o.rate_process == RATE_PROCESS_POISSON) {
 811                 uint64_t val;
 812                 iops = bps / td->o.bs[ddir];
 813                 val = (int64_t) (1000000 / iops) *
 814                                 -logf(__rand_0_1(&td->poisson_state));
 815                 if (val) {
 816                         dprint(FD_RATE, "poisson rate iops=%llu\n",
 817                                         (unsigned long long) 1000000 / val);
 818                 }
 819                 td->last_usec += val;
 820                 return td->last_usec;
 821         } else if (bps) {
 822                 secs = bytes / bps;
 823                 remainder = bytes % bps;
 824                 return remainder * 1000000 / bps + secs * 1000000;
 825         }
 826
 827         return 0;
 828 }
 829
 830 /*
 831  * Main IO worker function. It retrieves io_u's to process and queues
 832  * and reaps them, checking for rate and errors along the way.
 833  *
 834  * Returns number of bytes written and trimmed.
 835  */
 836 static void do_io(struct thread_data *td, uint64_t *bytes_done)
 837 {
 838         unsigned int i;
 839         int ret = 0;
 840         uint64_t total_bytes, bytes_issued = 0;
 841
 842         for (i = 0; i < DDIR_RWDIR_CNT; i++)
 843                 bytes_done[i] = td->bytes_done[i];
 844
 845         if (in_ramp_time(td))
 846                 td_set_runstate(td, TD_RAMP);
 847         else
 848                 td_set_runstate(td, TD_RUNNING);
 849
 850         lat_target_init(td);
 851
 852         total_bytes = td->o.size;
 853         /*
 854         * Allow random overwrite workloads to write up to io_limit
 855         * before starting verification phase as 'size' doesn't apply.
 856         */
 857         if (td_write(td) && td_random(td) && td->o.norandommap)
 858                 total_bytes = max(total_bytes, (uint64_t) td->o.io_limit);
 859         /*
 860          * If verify_backlog is enabled, we'll run the verify in this
 861          * handler as well. For that case, we may need up to twice the
 862          * amount of bytes.
 863          */
 864         if (td->o.verify != VERIFY_NONE &&
 865            (td_write(td) && td->o.verify_backlog))
 866                 total_bytes += td->o.size;
 867
 868         /* In trimwrite mode, each byte is trimmed and then written, so
 869          * allow total_bytes to be twice as big */
 870         if (td_trimwrite(td))
 871                 total_bytes += td->total_io_size;
 872
 873         while ((td->o.read_iolog_file && !flist_empty(&td->io_log_list)) ||
 874                 (!flist_empty(&td->trim_list)) || !io_issue_bytes_exceeded(td) ||
 875                 td->o.time_based) {
 876                 struct timeval comp_time;
 877                 struct io_u *io_u;
 878                 int full;
 879                 enum fio_ddir ddir;
 880
 881                 check_update_rusage(td);
 882
 883                 if (td->terminate || td->done)
 884                         break;
 885
 886                 update_tv_cache(td);
 887
 888                 if (runtime_exceeded(td, &td->tv_cache)) {
 889                         __update_tv_cache(td);
 890                         if (runtime_exceeded(td, &td->tv_cache)) {
 891                                 fio_mark_td_terminate(td);
 892                                 break;
 893                         }
 894                 }
 895
 896                 if (flow_threshold_exceeded(td))
 897                         continue;
 898
 899                 /*
 900                  * Break if we exceeded the bytes. The exception is time
 901                  * based runs, but we still need to break out of the loop
 902                  * for those to run verification, if enabled.
 903                  */
 904                 if (bytes_issued >= total_bytes &&
 905                     (!td->o.time_based ||
 906                      (td->o.time_based && td->o.verify != VERIFY_NONE)))
 907                         break;
 908
 909                 io_u = get_io_u(td);
 910                 if (IS_ERR_OR_NULL(io_u)) {
 911                         int err = PTR_ERR(io_u);
 912
 913                         io_u = NULL;
 914                         if (err == -EBUSY) {
 915                                 ret = FIO_Q_BUSY;
 916                                 goto reap;
 917                         }
 918                         if (td->o.latency_target)
 919                                 goto reap;
 920                         break;
 921                 }
 922
 923                 ddir = io_u->ddir;
 924
 925                 /*
 926                  * Add verification end_io handler if:
 927                  *      - Asked to verify (!td_rw(td))
 928                  *      - Or the io_u is from our verify list (mixed write/ver)
 929                  */
 930                 if (td->o.verify != VERIFY_NONE && io_u->ddir == DDIR_READ &&
 931                     ((io_u->flags & IO_U_F_VER_LIST) || !td_rw(td))) {
 932
 933                         if (!td->o.verify_pattern_bytes) {
 934                                 io_u->rand_seed = __rand(&td->verify_state);
 935                                 if (sizeof(int) != sizeof(long *))
 936                                         io_u->rand_seed *= __rand(&td->verify_state);
 937                         }
 938
 939                         if (verify_state_should_stop(td, io_u)) {
 940                                 put_io_u(td, io_u);
 941                                 break;
 942                         }
 943
 944                         if (td->o.verify_async)
 945                                 io_u->end_io = verify_io_u_async;
 946                         else
 947                                 io_u->end_io = verify_io_u;
 948                         td_set_runstate(td, TD_VERIFYING);
 949                 } else if (in_ramp_time(td))
 950                         td_set_runstate(td, TD_RAMP);
 951                 else
 952                         td_set_runstate(td, TD_RUNNING);
 953
 954                 /*
 955                  * Always log IO before it's issued, so we know the specific
 956                  * order of it. The logged unit will track when the IO has
 957                  * completed.
 958                  */
 959                 if (td_write(td) && io_u->ddir == DDIR_WRITE &&
 960                     td->o.do_verify &&
 961                     td->o.verify != VERIFY_NONE &&
 962                     !td->o.experimental_verify)
 963                         log_io_piece(td, io_u);
 964
 965                 if (td->o.io_submit_mode == IO_MODE_OFFLOAD) {
 966                         const unsigned long blen = io_u->xfer_buflen;
 967                         const enum fio_ddir ddir = acct_ddir(io_u);
 968
 969                         if (td->error)
 970                                 break;
 971
 972                         workqueue_enqueue(&td->io_wq, &io_u->work);
 973                         ret = FIO_Q_QUEUED;
 974
 975                         if (ddir_rw(ddir)) {
 976                                 td->io_issues[ddir]++;
 977                                 td->io_issue_bytes[ddir] += blen;
 978                                 td->rate_io_issue_bytes[ddir] += blen;
 979                         }
 980
 981                         if (should_check_rate(td))
 982                                 td->rate_next_io_time[ddir] = usec_for_io(td, ddir);
 983
 984                 } else {
 985                         ret = td_io_queue(td, io_u);
 986
 987                         if (should_check_rate(td))
 988                                 td->rate_next_io_time[ddir] = usec_for_io(td, ddir);
 989
 990                         if (io_queue_event(td, io_u, &ret, ddir, &bytes_issued, 0, &comp_time))
 991                                 break;
 992
 993                         /*
 994                          * See if we need to complete some commands. Note that
 995                          * we can get BUSY even without IO queued, if the
 996                          * system is resource starved.
 997                          */
 998 reap:
 999                         full = queue_full(td) ||
1000                                 (ret == FIO_Q_BUSY && td->cur_depth);
1001                         if (full || io_in_polling(td))
1002                                 ret = wait_for_completions(td, &comp_time);
1003                 }
1004                 if (ret < 0)
1005                         break;
1006                 if (!ddir_rw_sum(td->bytes_done) &&
1007                     !td_ioengine_flagged(td, FIO_NOIO))
1008                         continue;
1009
1010                 if (!in_ramp_time(td) && should_check_rate(td)) {
1011                         if (check_min_rate(td, &comp_time)) {
1012                                 if (exitall_on_terminate || td->o.exitall_error)
1013                                         fio_terminate_threads(td->groupid);
1014                                 td_verror(td, EIO, "check_min_rate");
1015                                 break;
1016                         }
1017                 }
1018                 if (!in_ramp_time(td) && td->o.latency_target)
1019                         lat_target_check(td);
1020
1021                 if (td->o.thinktime) {
1022                         unsigned long long b;
1023
1024                         b = ddir_rw_sum(td->io_blocks);
1025                         if (!(b % td->o.thinktime_blocks)) {
1026                                 int left;
1027
1028                                 io_u_quiesce(td);
1029
1030                                 if (td->o.thinktime_spin)
1031                                         usec_spin(td->o.thinktime_spin);
1032
1033                                 left = td->o.thinktime - td->o.thinktime_spin;
1034                                 if (left)
1035                                         usec_sleep(td, left);
1036                         }
1037                 }
1038         }
1039
1040         check_update_rusage(td);
1041
1042         if (td->trim_entries)
1043                 log_err("fio: %lu trim entries leaked?\n", td->trim_entries);
1044
1045         if (td->o.fill_device && td->error == ENOSPC) {
1046                 td->error = 0;
1047                 fio_mark_td_terminate(td);
1048         }
1049         if (!td->error) {
1050                 struct fio_file *f;
1051
1052                 if (td->o.io_submit_mode == IO_MODE_OFFLOAD) {
1053                         workqueue_flush(&td->io_wq);
1054                         i = 0;
1055                 } else
1056                         i = td->cur_depth;
1057
1058                 if (i) {
1059                         ret = io_u_queued_complete(td, i);
1060                         if (td->o.fill_device && td->error == ENOSPC)
1061                                 td->error = 0;
1062                 }
1063
1064                 if (should_fsync(td) && td->o.end_fsync) {
1065                         td_set_runstate(td, TD_FSYNCING);
1066
1067                         for_each_file(td, f, i) {
1068                                 if (!fio_file_fsync(td, f))
1069                                         continue;
1070
1071                                 log_err("fio: end_fsync failed for file %s\n",
1072                                                                 f->file_name);
1073                         }
1074                 }
1075         } else
1076                 cleanup_pending_aio(td);
1077
1078         /*
1079          * stop job if we failed doing any IO
1080          */
1081         if (!ddir_rw_sum(td->this_io_bytes))
1082                 td->done = 1;
1083
1084         for (i = 0; i < DDIR_RWDIR_CNT; i++)
1085                 bytes_done[i] = td->bytes_done[i] - bytes_done[i];
1086 }
1087
1088 static void free_file_completion_logging(struct thread_data *td)
1089 {
1090         struct fio_file *f;
1091         unsigned int i;
1092
1093         for_each_file(td, f, i) {
1094                 if (!f->last_write_comp)
1095                         break;
1096                 sfree(f->last_write_comp);
1097         }
1098 }
1099
1100 static int init_file_completion_logging(struct thread_data *td,
1101                                         unsigned int depth)
1102 {
1103         struct fio_file *f;
1104         unsigned int i;
1105
1106         if (td->o.verify == VERIFY_NONE || !td->o.verify_state_save)
1107                 return 0;
1108
1109         for_each_file(td, f, i) {
1110                 f->last_write_comp = scalloc(depth, sizeof(uint64_t));
1111                 if (!f->last_write_comp)
1112                         goto cleanup;
1113         }
1114
1115         return 0;
1116
1117 cleanup:
1118         free_file_completion_logging(td);
1119         log_err("fio: failed to alloc write comp data\n");
1120         return 1;
1121 }
1122
1123 static void cleanup_io_u(struct thread_data *td)
1124 {
1125         struct io_u *io_u;
1126
1127         while ((io_u = io_u_qpop(&td->io_u_freelist)) != NULL) {
1128
1129                 if (td->io_ops->io_u_free)
1130                         td->io_ops->io_u_free(td, io_u);
1131
1132                 fio_memfree(io_u, sizeof(*io_u));
1133         }
1134
1135         free_io_mem(td);
1136
1137         io_u_rexit(&td->io_u_requeues);
1138         io_u_qexit(&td->io_u_freelist);
1139         io_u_qexit(&td->io_u_all);
1140
1141         free_file_completion_logging(td);
1142 }
1143
1144 static int init_io_u(struct thread_data *td)
1145 {
1146         struct io_u *io_u;
1147         unsigned int max_bs, min_write;
1148         int cl_align, i, max_units;
1149         int data_xfer = 1, err;
1150         char *p;
1151
1152         max_units = td->o.iodepth;
1153         max_bs = td_max_bs(td);
1154         min_write = td->o.min_bs[DDIR_WRITE];
1155         td->orig_buffer_size = (unsigned long long) max_bs
1156                                         * (unsigned long long) max_units;
1157
1158         if (td_ioengine_flagged(td, FIO_NOIO) || !(td_read(td) || td_write(td)))
1159                 data_xfer = 0;
1160
1161         err = 0;
1162         err += io_u_rinit(&td->io_u_requeues, td->o.iodepth);
1163         err += io_u_qinit(&td->io_u_freelist, td->o.iodepth);
1164         err += io_u_qinit(&td->io_u_all, td->o.iodepth);
1165
1166         if (err) {
1167                 log_err("fio: failed setting up IO queues\n");
1168                 return 1;
1169         }
1170
1171         /*
1172          * if we may later need to do address alignment, then add any
1173          * possible adjustment here so that we don't cause a buffer
1174          * overflow later. this adjustment may be too much if we get
1175          * lucky and the allocator gives us an aligned address.
1176          */
1177         if (td->o.odirect || td->o.mem_align || td->o.oatomic ||
1178             td_ioengine_flagged(td, FIO_RAWIO))
1179                 td->orig_buffer_size += page_mask + td->o.mem_align;
1180
1181         if (td->o.mem_type == MEM_SHMHUGE || td->o.mem_type == MEM_MMAPHUGE) {
1182                 unsigned long bs;
1183
1184                 bs = td->orig_buffer_size + td->o.hugepage_size - 1;
1185                 td->orig_buffer_size = bs & ~(td->o.hugepage_size - 1);
1186         }
1187
1188         if (td->orig_buffer_size != (size_t) td->orig_buffer_size) {
1189                 log_err("fio: IO memory too large. Reduce max_bs or iodepth\n");
1190                 return 1;
1191         }
1192
1193         if (data_xfer && allocate_io_mem(td))
1194                 return 1;
1195
1196         if (td->o.odirect || td->o.mem_align || td->o.oatomic ||
1197             td_ioengine_flagged(td, FIO_RAWIO))
1198                 p = PTR_ALIGN(td->orig_buffer, page_mask) + td->o.mem_align;
1199         else
1200                 p = td->orig_buffer;
1201
1202         cl_align = os_cache_line_size();
1203
1204         for (i = 0; i < max_units; i++) {
1205                 void *ptr;
1206
1207                 if (td->terminate)
1208                         return 1;
1209
1210                 ptr = fio_memalign(cl_align, sizeof(*io_u));
1211                 if (!ptr) {
1212                         log_err("fio: unable to allocate aligned memory\n");
1213                         break;
1214                 }
1215
1216                 io_u = ptr;
1217                 memset(io_u, 0, sizeof(*io_u));
1218                 INIT_FLIST_HEAD(&io_u->verify_list);
1219                 dprint(FD_MEM, "io_u alloc %p, index %u\n", io_u, i);
1220
1221                 if (data_xfer) {
1222                         io_u->buf = p;
1223                         dprint(FD_MEM, "io_u %p, mem %p\n", io_u, io_u->buf);
1224
1225                         if (td_write(td))
1226                                 io_u_fill_buffer(td, io_u, min_write, max_bs);
1227                         if (td_write(td) && td->o.verify_pattern_bytes) {
1228                                 /*
1229                                  * Fill the buffer with the pattern if we are
1230                                  * going to be doing writes.
1231                                  */
1232                                 fill_verify_pattern(td, io_u->buf, max_bs, io_u, 0, 0);
1233                         }
1234                 }
1235
1236                 io_u->index = i;
1237                 io_u->flags = IO_U_F_FREE;
1238                 io_u_qpush(&td->io_u_freelist, io_u);
1239
1240                 /*
1241                  * io_u never leaves this stack, used for iteration of all
1242                  * io_u buffers.
1243                  */
1244                 io_u_qpush(&td->io_u_all, io_u);
1245
1246                 if (td->io_ops->io_u_init) {
1247                         int ret = td->io_ops->io_u_init(td, io_u);
1248
1249                         if (ret) {
1250                                 log_err("fio: failed to init engine data: %d\n", ret);
1251                                 return 1;
1252                         }
1253                 }
1254
1255                 p += max_bs;
1256         }
1257
1258         if (init_file_completion_logging(td, max_units))
1259                 return 1;
1260
1261         return 0;
1262 }
1263
1264 /*
1265  * This function is Linux specific.
1266  * FIO_HAVE_IOSCHED_SWITCH enabled currently means it's Linux.
1267  */
1268 static int switch_ioscheduler(struct thread_data *td)
1269 {
1270 #ifdef FIO_HAVE_IOSCHED_SWITCH
1271         char tmp[256], tmp2[128];
1272         FILE *f;
1273         int ret;
1274
1275         if (td_ioengine_flagged(td, FIO_DISKLESSIO))
1276                 return 0;
1277
1278         assert(td->files && td->files[0]);
1279         sprintf(tmp, "%s/queue/scheduler", td->files[0]->du->sysfs_root);
1280
1281         f = fopen(tmp, "r+");
1282         if (!f) {
1283                 if (errno == ENOENT) {
1284                         log_err("fio: os or kernel doesn't support IO scheduler"
1285                                 " switching\n");
1286                         return 0;
1287                 }
1288                 td_verror(td, errno, "fopen iosched");
1289                 return 1;
1290         }
1291
1292         /*
1293          * Set io scheduler.
1294          */
1295         ret = fwrite(td->o.ioscheduler, strlen(td->o.ioscheduler), 1, f);
1296         if (ferror(f) || ret != 1) {
1297                 td_verror(td, errno, "fwrite");
1298                 fclose(f);
1299                 return 1;
1300         }
1301
1302         rewind(f);
1303
1304         /*
1305          * Read back and check that the selected scheduler is now the default.
1306          */
1307         memset(tmp, 0, sizeof(tmp));
1308         ret = fread(tmp, sizeof(tmp), 1, f);
1309         if (ferror(f) || ret < 0) {
1310                 td_verror(td, errno, "fread");
1311                 fclose(f);
1312                 return 1;
1313         }
1314         /*
1315          * either a list of io schedulers or "none\n" is expected.
1316          */
1317         tmp[strlen(tmp) - 1] = '\0';
1318
1319         /*
1320          * Write to "none" entry doesn't fail, so check the result here.
1321          */
1322         if (!strcmp(tmp, "none")) {
1323                 log_err("fio: io scheduler is not tunable\n");
1324                 fclose(f);
1325                 return 0;
1326         }
1327
1328         sprintf(tmp2, "[%s]", td->o.ioscheduler);
1329         if (!strstr(tmp, tmp2)) {
1330                 log_err("fio: io scheduler %s not found\n", td->o.ioscheduler);
1331                 td_verror(td, EINVAL, "iosched_switch");
1332                 fclose(f);
1333                 return 1;
1334         }
1335
1336         fclose(f);
1337         return 0;
1338 #else
1339         return 0;
1340 #endif
1341 }
1342
1343 static bool keep_running(struct thread_data *td)
1344 {
1345         unsigned long long limit;
1346
1347         if (td->done)
1348                 return false;
1349         if (td->o.time_based)
1350                 return true;
1351         if (td->o.loops) {
1352                 td->o.loops--;
1353                 return true;
1354         }
1355         if (exceeds_number_ios(td))
1356                 return false;
1357
1358         if (td->o.io_limit)
1359                 limit = td->o.io_limit;
1360         else
1361                 limit = td->o.size;
1362
1363         if (limit != -1ULL && ddir_rw_sum(td->io_bytes) < limit) {
1364                 uint64_t diff;
1365
1366                 /*
1367                  * If the difference is less than the maximum IO size, we
1368                  * are done.
1369                  */
1370                 diff = limit - ddir_rw_sum(td->io_bytes);
1371                 if (diff < td_max_bs(td))
1372                         return false;
1373
1374                 if (fio_files_done(td) && !td->o.io_limit)
1375                         return false;
1376
1377                 return true;
1378         }
1379
1380         return false;
1381 }
1382
1383 static int exec_string(struct thread_options *o, const char *string, const char *mode)
1384 {
1385         size_t newlen = strlen(string) + strlen(o->name) + strlen(mode) + 9 + 1;
1386         int ret;
1387         char *str;
1388
1389         str = malloc(newlen);
1390         sprintf(str, "%s &> %s.%s.txt", string, o->name, mode);
1391
1392         log_info("%s : Saving output of %s in %s.%s.txt\n",o->name, mode, o->name, mode);
1393         ret = system(str);
1394         if (ret == -1)
1395                 log_err("fio: exec of cmd <%s> failed\n", str);
1396
1397         free(str);
1398         return ret;
1399 }
1400
1401 /*
1402  * Dry run to compute correct state of numberio for verification.
1403  */
1404 static uint64_t do_dry_run(struct thread_data *td)
1405 {
1406         td_set_runstate(td, TD_RUNNING);
1407
1408         while ((td->o.read_iolog_file && !flist_empty(&td->io_log_list)) ||
1409                 (!flist_empty(&td->trim_list)) || !io_complete_bytes_exceeded(td)) {
1410                 struct io_u *io_u;
1411                 int ret;
1412
1413                 if (td->terminate || td->done)
1414                         break;
1415
1416                 io_u = get_io_u(td);
1417                 if (IS_ERR_OR_NULL(io_u))
1418                         break;
1419
1420                 io_u_set(td, io_u, IO_U_F_FLIGHT);
1421                 io_u->error = 0;
1422                 io_u->resid = 0;
1423                 if (ddir_rw(acct_ddir(io_u)))
1424                         td->io_issues[acct_ddir(io_u)]++;
1425                 if (ddir_rw(io_u->ddir)) {
1426                         io_u_mark_depth(td, 1);
1427                         td->ts.total_io_u[io_u->ddir]++;
1428                 }
1429
1430                 if (td_write(td) && io_u->ddir == DDIR_WRITE &&
1431                     td->o.do_verify &&
1432                     td->o.verify != VERIFY_NONE &&
1433                     !td->o.experimental_verify)
1434                         log_io_piece(td, io_u);
1435
1436                 ret = io_u_sync_complete(td, io_u);
1437                 (void) ret;
1438         }
1439
1440         return td->bytes_done[DDIR_WRITE] + td->bytes_done[DDIR_TRIM];
1441 }
1442
1443 struct fork_data {
1444         struct thread_data *td;
1445         struct sk_out *sk_out;
1446 };
1447
1448 /*
1449  * Entry point for the thread based jobs. The process based jobs end up
1450  * here as well, after a little setup.
1451  */
1452 static void *thread_main(void *data)
1453 {
1454         struct fork_data *fd = data;
1455         unsigned long long elapsed_us[DDIR_RWDIR_CNT] = { 0, };
1456         struct thread_data *td = fd->td;
1457         struct thread_options *o = &td->o;
1458         struct sk_out *sk_out = fd->sk_out;
1459         int deadlock_loop_cnt;
1460         int clear_state;
1461         int ret;
1462
1463         sk_out_assign(sk_out);
1464         free(fd);
1465
1466         if (!o->use_thread) {
1467                 setsid();
1468                 td->pid = getpid();
1469         } else
1470                 td->pid = gettid();
1471
1472         fio_local_clock_init(o->use_thread);
1473
1474         dprint(FD_PROCESS, "jobs pid=%d started\n", (int) td->pid);
1475
1476         if (is_backend)
1477                 fio_server_send_start(td);
1478
1479         INIT_FLIST_HEAD(&td->io_log_list);
1480         INIT_FLIST_HEAD(&td->io_hist_list);
1481         INIT_FLIST_HEAD(&td->verify_list);
1482         INIT_FLIST_HEAD(&td->trim_list);
1483         INIT_FLIST_HEAD(&td->next_rand_list);
1484         td->io_hist_tree = RB_ROOT;
1485
1486         ret = mutex_cond_init_pshared(&td->io_u_lock, &td->free_cond);
1487         if (ret) {
1488                 td_verror(td, ret, "mutex_cond_init_pshared");
1489                 goto err;
1490         }
1491         ret = cond_init_pshared(&td->verify_cond);
1492         if (ret) {
1493                 td_verror(td, ret, "mutex_cond_pshared");
1494                 goto err;
1495         }
1496
1497         td_set_runstate(td, TD_INITIALIZED);
1498         dprint(FD_MUTEX, "up startup_mutex\n");
1499         fio_mutex_up(startup_mutex);
1500         dprint(FD_MUTEX, "wait on td->mutex\n");
1501         fio_mutex_down(td->mutex);
1502         dprint(FD_MUTEX, "done waiting on td->mutex\n");
1503
1504         /*
1505          * A new gid requires privilege, so we need to do this before setting
1506          * the uid.
1507          */
1508         if (o->gid != -1U && setgid(o->gid)) {
1509                 td_verror(td, errno, "setgid");
1510                 goto err;
1511         }
1512         if (o->uid != -1U && setuid(o->uid)) {
1513                 td_verror(td, errno, "setuid");
1514                 goto err;
1515         }
1516
1517         /*
1518          * Do this early, we don't want the compress threads to be limited
1519          * to the same CPUs as the IO workers. So do this before we set
1520          * any potential CPU affinity
1521          */
1522         if (iolog_compress_init(td, sk_out))
1523                 goto err;
1524
1525         /*
1526          * If we have a gettimeofday() thread, make sure we exclude that
1527          * thread from this job
1528          */
1529         if (o->gtod_cpu)
1530                 fio_cpu_clear(&o->cpumask, o->gtod_cpu);
1531
1532         /*
1533          * Set affinity first, in case it has an impact on the memory
1534          * allocations.
1535          */
1536         if (fio_option_is_set(o, cpumask)) {
1537                 if (o->cpus_allowed_policy == FIO_CPUS_SPLIT) {
1538                         ret = fio_cpus_split(&o->cpumask, td->thread_number - 1);
1539                         if (!ret) {
1540                                 log_err("fio: no CPUs set\n");
1541                                 log_err("fio: Try increasing number of available CPUs\n");
1542                                 td_verror(td, EINVAL, "cpus_split");
1543                                 goto err;
1544                         }
1545                 }
1546                 ret = fio_setaffinity(td->pid, o->cpumask);
1547                 if (ret == -1) {
1548                         td_verror(td, errno, "cpu_set_affinity");
1549                         goto err;
1550                 }
1551         }
1552
1553 #ifdef CONFIG_LIBNUMA
1554         /* numa node setup */
1555         if (fio_option_is_set(o, numa_cpunodes) ||
1556             fio_option_is_set(o, numa_memnodes)) {
1557                 struct bitmask *mask;
1558
1559                 if (numa_available() < 0) {
1560                         td_verror(td, errno, "Does not support NUMA API\n");
1561                         goto err;
1562                 }
1563
1564                 if (fio_option_is_set(o, numa_cpunodes)) {
1565                         mask = numa_parse_nodestring(o->numa_cpunodes);
1566                         ret = numa_run_on_node_mask(mask);
1567                         numa_free_nodemask(mask);
1568                         if (ret == -1) {
1569                                 td_verror(td, errno, \
1570                                         "numa_run_on_node_mask failed\n");
1571                                 goto err;
1572                         }
1573                 }
1574
1575                 if (fio_option_is_set(o, numa_memnodes)) {
1576                         mask = NULL;
1577                         if (o->numa_memnodes)
1578                                 mask = numa_parse_nodestring(o->numa_memnodes);
1579
1580                         switch (o->numa_mem_mode) {
1581                         case MPOL_INTERLEAVE:
1582                                 numa_set_interleave_mask(mask);
1583                                 break;
1584                         case MPOL_BIND:
1585                                 numa_set_membind(mask);
1586                                 break;
1587                         case MPOL_LOCAL:
1588                                 numa_set_localalloc();
1589                                 break;
1590                         case MPOL_PREFERRED:
1591                                 numa_set_preferred(o->numa_mem_prefer_node);
1592                                 break;
1593                         case MPOL_DEFAULT:
1594                         default:
1595                                 break;
1596                         }
1597
1598                         if (mask)
1599                                 numa_free_nodemask(mask);
1600
1601                 }
1602         }
1603 #endif
1604
1605         if (fio_pin_memory(td))
1606                 goto err;
1607
1608         /*
1609          * May alter parameters that init_io_u() will use, so we need to
1610          * do this first.
1611          */
1612         if (init_iolog(td))
1613                 goto err;
1614
1615         if (init_io_u(td))
1616                 goto err;
1617
1618         if (o->verify_async && verify_async_init(td))
1619                 goto err;
1620
1621         if (fio_option_is_set(o, ioprio) ||
1622             fio_option_is_set(o, ioprio_class)) {
1623                 ret = ioprio_set(IOPRIO_WHO_PROCESS, 0, o->ioprio_class, o->ioprio);
1624                 if (ret == -1) {
1625                         td_verror(td, errno, "ioprio_set");
1626                         goto err;
1627                 }
1628         }
1629
1630         if (o->cgroup && cgroup_setup(td, cgroup_list, &cgroup_mnt))
1631                 goto err;
1632
1633         errno = 0;
1634         if (nice(o->nice) == -1 && errno != 0) {
1635                 td_verror(td, errno, "nice");
1636                 goto err;
1637         }
1638
1639         if (o->ioscheduler && switch_ioscheduler(td))
1640                 goto err;
1641
1642         if (!o->create_serialize && setup_files(td))
1643                 goto err;
1644
1645         if (td_io_init(td))
1646                 goto err;
1647
1648         if (init_random_map(td))
1649                 goto err;
1650
1651         if (o->exec_prerun && exec_string(o, o->exec_prerun, (const char *)"prerun"))
1652                 goto err;
1653
1654         if (o->pre_read) {
1655                 if (pre_read_files(td) < 0)
1656                         goto err;
1657         }
1658
1659         fio_verify_init(td);
1660
1661         if (rate_submit_init(td, sk_out))
1662                 goto err;
1663
1664         set_epoch_time(td, o->log_unix_epoch);
1665         fio_getrusage(&td->ru_start);
1666         memcpy(&td->bw_sample_time, &td->epoch, sizeof(td->epoch));
1667         memcpy(&td->iops_sample_time, &td->epoch, sizeof(td->epoch));
1668         memcpy(&td->ss.prev_time, &td->epoch, sizeof(td->epoch));
1669
1670         if (o->ratemin[DDIR_READ] || o->ratemin[DDIR_WRITE] ||
1671                         o->ratemin[DDIR_TRIM]) {
1672                 memcpy(&td->lastrate[DDIR_READ], &td->bw_sample_time,
1673                                         sizeof(td->bw_sample_time));
1674                 memcpy(&td->lastrate[DDIR_WRITE], &td->bw_sample_time,
1675                                         sizeof(td->bw_sample_time));
1676                 memcpy(&td->lastrate[DDIR_TRIM], &td->bw_sample_time,
1677                                         sizeof(td->bw_sample_time));
1678         }
1679
1680         clear_state = 0;
1681         while (keep_running(td)) {
1682                 uint64_t verify_bytes;
1683
1684                 fio_gettime(&td->start, NULL);
1685                 memcpy(&td->tv_cache, &td->start, sizeof(td->start));
1686
1687                 if (clear_state) {
1688                         clear_io_state(td, 0);
1689
1690                         if (o->unlink_each_loop && unlink_all_files(td))
1691                                 break;
1692                 }
1693
1694                 prune_io_piece_log(td);
1695
1696                 if (td->o.verify_only && (td_write(td) || td_rw(td)))
1697                         verify_bytes = do_dry_run(td);
1698                 else {
1699                         uint64_t bytes_done[DDIR_RWDIR_CNT];
1700
1701                         do_io(td, bytes_done);
1702
1703                         if (!ddir_rw_sum(bytes_done)) {
1704                                 fio_mark_td_terminate(td);
1705                                 verify_bytes = 0;
1706                         } else {
1707                                 verify_bytes = bytes_done[DDIR_WRITE] +
1708                                                 bytes_done[DDIR_TRIM];
1709                         }
1710                 }
1711
1712                 /*
1713                  * If we took too long to shut down, the main thread could
1714                  * already consider us reaped/exited. If that happens, break
1715                  * out and clean up.
1716                  */
1717                 if (td->runstate >= TD_EXITED)
1718                         break;
1719
1720                 clear_state = 1;
1721
1722                 /*
1723                  * Make sure we've successfully updated the rusage stats
1724                  * before waiting on the stat mutex. Otherwise we could have
1725                  * the stat thread holding stat mutex and waiting for
1726                  * the rusage_sem, which would never get upped because
1727                  * this thread is waiting for the stat mutex.
1728                  */
1729                 deadlock_loop_cnt = 0;
1730                 do {
1731                         check_update_rusage(td);
1732                         if (!fio_mutex_down_trylock(stat_mutex))
1733                                 break;
1734                         usleep(1000);
1735                         if (deadlock_loop_cnt++ > 5000) {
1736                                 log_err("fio seems to be stuck grabbing stat_mutex, forcibly exiting\n");
1737                                 td->error = EDEADLK;
1738                                 goto err;
1739                         }
1740                 } while (1);
1741
1742                 if (td_read(td) && td->io_bytes[DDIR_READ])
1743                         update_runtime(td, elapsed_us, DDIR_READ);
1744                 if (td_write(td) && td->io_bytes[DDIR_WRITE])
1745                         update_runtime(td, elapsed_us, DDIR_WRITE);
1746                 if (td_trim(td) && td->io_bytes[DDIR_TRIM])
1747                         update_runtime(td, elapsed_us, DDIR_TRIM);
1748                 fio_gettime(&td->start, NULL);
1749                 fio_mutex_up(stat_mutex);
1750
1751                 if (td->error || td->terminate)
1752                         break;
1753
1754                 if (!o->do_verify ||
1755                     o->verify == VERIFY_NONE ||
1756                     td_ioengine_flagged(td, FIO_UNIDIR))
1757                         continue;
1758
1759                 clear_io_state(td, 0);
1760
1761                 fio_gettime(&td->start, NULL);
1762
1763                 do_verify(td, verify_bytes);
1764
1765                 /*
1766                  * See comment further up for why this is done here.
1767                  */
1768                 check_update_rusage(td);
1769
1770                 fio_mutex_down(stat_mutex);
1771                 update_runtime(td, elapsed_us, DDIR_READ);
1772                 fio_gettime(&td->start, NULL);
1773                 fio_mutex_up(stat_mutex);
1774
1775                 if (td->error || td->terminate)
1776                         break;
1777         }
1778
1779         td_set_runstate(td, TD_FINISHING);
1780
1781         update_rusage_stat(td);
1782         td->ts.total_run_time = mtime_since_now(&td->epoch);
1783         td->ts.io_bytes[DDIR_READ] = td->io_bytes[DDIR_READ];
1784         td->ts.io_bytes[DDIR_WRITE] = td->io_bytes[DDIR_WRITE];
1785         td->ts.io_bytes[DDIR_TRIM] = td->io_bytes[DDIR_TRIM];
1786
1787         if (td->o.verify_state_save && !(td->flags & TD_F_VSTATE_SAVED) &&
1788             (td->o.verify != VERIFY_NONE && td_write(td)))
1789                 verify_save_state(td->thread_number);
1790
1791         fio_unpin_memory(td);
1792
1793         td_writeout_logs(td, true);
1794
1795         iolog_compress_exit(td);
1796         rate_submit_exit(td);
1797
1798         if (o->exec_postrun)
1799                 exec_string(o, o->exec_postrun, (const char *)"postrun");
1800
1801         if (exitall_on_terminate || (o->exitall_error && td->error))
1802                 fio_terminate_threads(td->groupid);
1803
1804 err:
1805         if (td->error)
1806                 log_info("fio: pid=%d, err=%d/%s\n", (int) td->pid, td->error,
1807                                                         td->verror);
1808
1809         if (o->verify_async)
1810                 verify_async_exit(td);
1811
1812         close_and_free_files(td);
1813         cleanup_io_u(td);
1814         close_ioengine(td);
1815         cgroup_shutdown(td, &cgroup_mnt);
1816         verify_free_state(td);
1817
1818         if (td->zone_state_index) {
1819                 int i;
1820
1821                 for (i = 0; i < DDIR_RWDIR_CNT; i++)
1822                         free(td->zone_state_index[i]);
1823                 free(td->zone_state_index);
1824                 td->zone_state_index = NULL;
1825         }
1826
1827         if (fio_option_is_set(o, cpumask)) {
1828                 ret = fio_cpuset_exit(&o->cpumask);
1829                 if (ret)
1830                         td_verror(td, ret, "fio_cpuset_exit");
1831         }
1832
1833         /*
1834          * do this very late, it will log file closing as well
1835          */
1836         if (o->write_iolog_file)
1837                 write_iolog_close(td);
1838
1839         fio_mutex_remove(td->mutex);
1840         td->mutex = NULL;
1841
1842         td_set_runstate(td, TD_EXITED);
1843
1844         /*
1845          * Do this last after setting our runstate to exited, so we
1846          * know that the stat thread is signaled.
1847          */
1848         check_update_rusage(td);
1849
1850         sk_out_drop();
1851         return (void *) (uintptr_t) td->error;
1852 }
1853
1854 static void dump_td_info(struct thread_data *td)
1855 {
1856         log_err("fio: job '%s' (state=%d) hasn't exited in %lu seconds, it "
1857                 "appears to be stuck. Doing forceful exit of this job.\n",
1858                         td->o.name, td->runstate,
1859                         (unsigned long) time_since_now(&td->terminate_time));
1860 }
1861
1862 /*
1863  * Run over the job map and reap the threads that have exited, if any.
1864  */
1865 static void reap_threads(unsigned int *nr_running, uint64_t *t_rate,
1866                          uint64_t *m_rate)
1867 {
1868         struct thread_data *td;
1869         unsigned int cputhreads, realthreads, pending;
1870         int i, status, ret;
1871
1872         /*
1873          * reap exited threads (TD_EXITED -> TD_REAPED)
1874          */
1875         realthreads = pending = cputhreads = 0;
1876         for_each_td(td, i) {
1877                 int flags = 0;
1878
1879                 /*
1880                  * ->io_ops is NULL for a thread that has closed its
1881                  * io engine
1882                  */
1883                 if (td->io_ops && !strcmp(td->io_ops->name, "cpuio"))
1884                         cputhreads++;
1885                 else
1886                         realthreads++;
1887
1888                 if (!td->pid) {
1889                         pending++;
1890                         continue;
1891                 }
1892                 if (td->runstate == TD_REAPED)
1893                         continue;
1894                 if (td->o.use_thread) {
1895                         if (td->runstate == TD_EXITED) {
1896                                 td_set_runstate(td, TD_REAPED);
1897                                 goto reaped;
1898                         }
1899                         continue;
1900                 }
1901
1902                 flags = WNOHANG;
1903                 if (td->runstate == TD_EXITED)
1904                         flags = 0;
1905
1906                 /*
1907                  * check if someone quit or got killed in an unusual way
1908                  */
1909                 ret = waitpid(td->pid, &status, flags);
1910                 if (ret < 0) {
1911                         if (errno == ECHILD) {
1912                                 log_err("fio: pid=%d disappeared %d\n",
1913                                                 (int) td->pid, td->runstate);
1914                                 td->sig = ECHILD;
1915                                 td_set_runstate(td, TD_REAPED);
1916                                 goto reaped;
1917                         }
1918                         perror("waitpid");
1919                 } else if (ret == td->pid) {
1920                         if (WIFSIGNALED(status)) {
1921                                 int sig = WTERMSIG(status);
1922
1923                                 if (sig != SIGTERM && sig != SIGUSR2)
1924                                         log_err("fio: pid=%d, got signal=%d\n",
1925                                                         (int) td->pid, sig);
1926                                 td->sig = sig;
1927                                 td_set_runstate(td, TD_REAPED);
1928                                 goto reaped;
1929                         }
1930                         if (WIFEXITED(status)) {
1931                                 if (WEXITSTATUS(status) && !td->error)
1932                                         td->error = WEXITSTATUS(status);
1933
1934                                 td_set_runstate(td, TD_REAPED);
1935                                 goto reaped;
1936                         }
1937                 }
1938
1939                 /*
1940                  * If the job is stuck, do a forceful timeout of it and
1941                  * move on.
1942                  */
1943                 if (td->terminate &&
1944                     td->runstate < TD_FSYNCING &&
1945                     time_since_now(&td->terminate_time) >= FIO_REAP_TIMEOUT) {
1946                         dump_td_info(td);
1947                         td_set_runstate(td, TD_REAPED);
1948                         goto reaped;
1949                 }
1950
1951                 /*
1952                  * thread is not dead, continue
1953                  */
1954                 pending++;
1955                 continue;
1956 reaped:
1957                 (*nr_running)--;
1958                 (*m_rate) -= ddir_rw_sum(td->o.ratemin);
1959                 (*t_rate) -= ddir_rw_sum(td->o.rate);
1960                 if (!td->pid)
1961                         pending--;
1962
1963                 if (td->error)
1964                         exit_value++;
1965
1966                 done_secs += mtime_since_now(&td->epoch) / 1000;
1967                 profile_td_exit(td);
1968         }
1969
1970         if (*nr_running == cputhreads && !pending && realthreads)
1971                 fio_terminate_threads(TERMINATE_ALL);
1972 }
1973
1974 static bool __check_trigger_file(void)
1975 {
1976         struct stat sb;
1977
1978         if (!trigger_file)
1979                 return false;
1980
1981         if (stat(trigger_file, &sb))
1982                 return false;
1983
1984         if (unlink(trigger_file) < 0)
1985                 log_err("fio: failed to unlink %s: %s\n", trigger_file,
1986                                                         strerror(errno));
1987
1988         return true;
1989 }
1990
1991 static bool trigger_timedout(void)
1992 {
1993         if (trigger_timeout)
1994                 return time_since_genesis() >= trigger_timeout;
1995
1996         return false;
1997 }
1998
1999 void exec_trigger(const char *cmd)
2000 {
2001         int ret;
2002
2003         if (!cmd)
2004                 return;
2005
2006         ret = system(cmd);
2007         if (ret == -1)
2008                 log_err("fio: failed executing %s trigger\n", cmd);
2009 }
2010
2011 void check_trigger_file(void)
2012 {
2013         if (__check_trigger_file() || trigger_timedout()) {
2014                 if (nr_clients)
2015                         fio_clients_send_trigger(trigger_remote_cmd);
2016                 else {
2017                         verify_save_state(IO_LIST_ALL);
2018                         fio_terminate_threads(TERMINATE_ALL);
2019                         exec_trigger(trigger_cmd);
2020                 }
2021         }
2022 }
2023
2024 static int fio_verify_load_state(struct thread_data *td)
2025 {
2026         int ret;
2027
2028         if (!td->o.verify_state)
2029                 return 0;
2030
2031         if (is_backend) {
2032                 void *data;
2033
2034                 ret = fio_server_get_verify_state(td->o.name,
2035                                         td->thread_number - 1, &data);
2036                 if (!ret)
2037                         verify_assign_state(td, data);
2038         } else
2039                 ret = verify_load_state(td, "local");
2040
2041         return ret;
2042 }
2043
2044 static void do_usleep(unsigned int usecs)
2045 {
2046         check_for_running_stats();
2047         check_trigger_file();
2048         usleep(usecs);
2049 }
2050
2051 static bool check_mount_writes(struct thread_data *td)
2052 {
2053         struct fio_file *f;
2054         unsigned int i;
2055
2056         if (!td_write(td) || td->o.allow_mounted_write)
2057                 return false;
2058
2059         for_each_file(td, f, i) {
2060                 if (f->filetype != FIO_TYPE_BLOCK)
2061                         continue;
2062                 if (device_is_mounted(f->file_name))
2063                         goto mounted;
2064         }
2065
2066         return false;
2067 mounted:
2068         log_err("fio: %s appears mounted, and 'allow_mounted_write' isn't set. Aborting.\n", f->file_name);
2069         return true;
2070 }
2071
2072 static bool waitee_running(struct thread_data *me)
2073 {
2074         const char *waitee = me->o.wait_for;
2075         const char *self = me->o.name;
2076         struct thread_data *td;
2077         int i;
2078
2079         if (!waitee)
2080                 return false;
2081
2082         for_each_td(td, i) {
2083                 if (!strcmp(td->o.name, self) || strcmp(td->o.name, waitee))
2084                         continue;
2085
2086                 if (td->runstate < TD_EXITED) {
2087                         dprint(FD_PROCESS, "%s fenced by %s(%s)\n",
2088                                         self, td->o.name,
2089                                         runstate_to_name(td->runstate));
2090                         return true;
2091                 }
2092         }
2093
2094         dprint(FD_PROCESS, "%s: %s completed, can run\n", self, waitee);
2095         return false;
2096 }
2097
2098 /*
2099  * Main function for kicking off and reaping jobs, as needed.
2100  */
2101 static void run_threads(struct sk_out *sk_out)
2102 {
2103         struct thread_data *td;
2104         unsigned int i, todo, nr_running, nr_started;
2105         uint64_t m_rate, t_rate;
2106         uint64_t spent;
2107
2108         if (fio_gtod_offload && fio_start_gtod_thread())
2109                 return;
2110
2111         fio_idle_prof_init();
2112
2113         set_sig_handlers();
2114
2115         nr_thread = nr_process = 0;
2116         for_each_td(td, i) {
2117                 if (check_mount_writes(td))
2118                         return;
2119                 if (td->o.use_thread)
2120                         nr_thread++;
2121                 else
2122                         nr_process++;
2123         }
2124
2125         if (output_format & FIO_OUTPUT_NORMAL) {
2126                 log_info("Starting ");
2127                 if (nr_thread)
2128                         log_info("%d thread%s", nr_thread,
2129                                                 nr_thread > 1 ? "s" : "");
2130                 if (nr_process) {
2131                         if (nr_thread)
2132                                 log_info(" and ");
2133                         log_info("%d process%s", nr_process,
2134                                                 nr_process > 1 ? "es" : "");
2135                 }
2136                 log_info("\n");
2137                 log_info_flush();
2138         }
2139
2140         todo = thread_number;
2141         nr_running = 0;
2142         nr_started = 0;
2143         m_rate = t_rate = 0;
2144
2145         for_each_td(td, i) {
2146                 print_status_init(td->thread_number - 1);
2147
2148                 if (!td->o.create_serialize)
2149                         continue;
2150
2151                 if (fio_verify_load_state(td))
2152                         goto reap;
2153
2154                 /*
2155                  * do file setup here so it happens sequentially,
2156                  * we don't want X number of threads getting their
2157                  * client data interspersed on disk
2158                  */
2159                 if (setup_files(td)) {
2160 reap:
2161                         exit_value++;
2162                         if (td->error)
2163                                 log_err("fio: pid=%d, err=%d/%s\n",
2164                                         (int) td->pid, td->error, td->verror);
2165                         td_set_runstate(td, TD_REAPED);
2166                         todo--;
2167                 } else {
2168                         struct fio_file *f;
2169                         unsigned int j;
2170
2171                         /*
2172                          * for sharing to work, each job must always open
2173                          * its own files. so close them, if we opened them
2174                          * for creation
2175                          */
2176                         for_each_file(td, f, j) {
2177                                 if (fio_file_open(f))
2178                                         td_io_close_file(td, f);
2179                         }
2180                 }
2181         }
2182
2183         /* start idle threads before io threads start to run */
2184         fio_idle_prof_start();
2185
2186         set_genesis_time();
2187
2188         while (todo) {
2189                 struct thread_data *map[REAL_MAX_JOBS];
2190                 struct timeval this_start;
2191                 int this_jobs = 0, left;
2192                 struct fork_data *fd;
2193
2194                 /*
2195                  * create threads (TD_NOT_CREATED -> TD_CREATED)
2196                  */
2197                 for_each_td(td, i) {
2198                         if (td->runstate != TD_NOT_CREATED)
2199                                 continue;
2200
2201                         /*
2202                          * never got a chance to start, killed by other
2203                          * thread for some reason
2204                          */
2205                         if (td->terminate) {
2206                                 todo--;
2207                                 continue;
2208                         }
2209
2210                         if (td->o.start_delay) {
2211                                 spent = utime_since_genesis();
2212
2213                                 if (td->o.start_delay > spent)
2214                                         continue;
2215                         }
2216
2217                         if (td->o.stonewall && (nr_started || nr_running)) {
2218                                 dprint(FD_PROCESS, "%s: stonewall wait\n",
2219                                                         td->o.name);
2220                                 break;
2221                         }
2222
2223                         if (waitee_running(td)) {
2224                                 dprint(FD_PROCESS, "%s: waiting for %s\n",
2225                                                 td->o.name, td->o.wait_for);
2226                                 continue;
2227                         }
2228
2229                         init_disk_util(td);
2230
2231                         td->rusage_sem = fio_mutex_init(FIO_MUTEX_LOCKED);
2232                         td->update_rusage = 0;
2233
2234                         /*
2235                          * Set state to created. Thread will transition
2236                          * to TD_INITIALIZED when it's done setting up.
2237                          */
2238                         td_set_runstate(td, TD_CREATED);
2239                         map[this_jobs++] = td;
2240                         nr_started++;
2241
2242                         fd = calloc(1, sizeof(*fd));
2243                         fd->td = td;
2244                         fd->sk_out = sk_out;
2245
2246                         if (td->o.use_thread) {
2247                                 int ret;
2248
2249                                 dprint(FD_PROCESS, "will pthread_create\n");
2250                                 ret = pthread_create(&td->thread, NULL,
2251                                                         thread_main, fd);
2252                                 if (ret) {
2253                                         log_err("pthread_create: %s\n",
2254                                                         strerror(ret));
2255                                         free(fd);
2256                                         nr_started--;
2257                                         break;
2258                                 }
2259                                 ret = pthread_detach(td->thread);
2260                                 if (ret)
2261                                         log_err("pthread_detach: %s",
2262                                                         strerror(ret));
2263                         } else {
2264                                 pid_t pid;
2265                                 dprint(FD_PROCESS, "will fork\n");
2266                                 pid = fork();
2267                                 if (!pid) {
2268                                         int ret;
2269
2270                                         ret = (int)(uintptr_t)thread_main(fd);
2271                                         _exit(ret);
2272                                 } else if (i == fio_debug_jobno)
2273                                         *fio_debug_jobp = pid;
2274                         }
2275                         dprint(FD_MUTEX, "wait on startup_mutex\n");
2276                         if (fio_mutex_down_timeout(startup_mutex, 10000)) {
2277                                 log_err("fio: job startup hung? exiting.\n");
2278                                 fio_terminate_threads(TERMINATE_ALL);
2279                                 fio_abort = 1;
2280                                 nr_started--;
2281                                 break;
2282                         }
2283                         dprint(FD_MUTEX, "done waiting on startup_mutex\n");
2284                 }
2285
2286                 /*
2287                  * Wait for the started threads to transition to
2288                  * TD_INITIALIZED.
2289                  */
2290                 fio_gettime(&this_start, NULL);
2291                 left = this_jobs;
2292                 while (left && !fio_abort) {
2293                         if (mtime_since_now(&this_start) > JOB_START_TIMEOUT)
2294                                 break;
2295
2296                         do_usleep(100000);
2297
2298                         for (i = 0; i < this_jobs; i++) {
2299                                 td = map[i];
2300                                 if (!td)
2301                                         continue;
2302                                 if (td->runstate == TD_INITIALIZED) {
2303                                         map[i] = NULL;
2304                                         left--;
2305                                 } else if (td->runstate >= TD_EXITED) {
2306                                         map[i] = NULL;
2307                                         left--;
2308                                         todo--;
2309                                         nr_running++; /* work-around... */
2310                                 }
2311                         }
2312                 }
2313
2314                 if (left) {
2315                         log_err("fio: %d job%s failed to start\n", left,
2316                                         left > 1 ? "s" : "");
2317                         for (i = 0; i < this_jobs; i++) {
2318                                 td = map[i];
2319                                 if (!td)
2320                                         continue;
2321                                 kill(td->pid, SIGTERM);
2322                         }
2323                         break;
2324                 }
2325
2326                 /*
2327                  * start created threads (TD_INITIALIZED -> TD_RUNNING).
2328                  */
2329                 for_each_td(td, i) {
2330                         if (td->runstate != TD_INITIALIZED)
2331                                 continue;
2332
2333                         if (in_ramp_time(td))
2334                                 td_set_runstate(td, TD_RAMP);
2335                         else
2336                                 td_set_runstate(td, TD_RUNNING);
2337                         nr_running++;
2338                         nr_started--;
2339                         m_rate += ddir_rw_sum(td->o.ratemin);
2340                         t_rate += ddir_rw_sum(td->o.rate);
2341                         todo--;
2342                         fio_mutex_up(td->mutex);
2343                 }
2344
2345                 reap_threads(&nr_running, &t_rate, &m_rate);
2346
2347                 if (todo)
2348                         do_usleep(100000);
2349         }
2350
2351         while (nr_running) {
2352                 reap_threads(&nr_running, &t_rate, &m_rate);
2353                 do_usleep(10000);
2354         }
2355
2356         fio_idle_prof_stop();
2357
2358         update_io_ticks();
2359 }
2360
2361 static void free_disk_util(void)
2362 {
2363         disk_util_prune_entries();
2364         helper_thread_destroy();
2365 }
2366
2367 int fio_backend(struct sk_out *sk_out)
2368 {
2369         struct thread_data *td;
2370         int i;
2371
2372         if (exec_profile) {
2373                 if (load_profile(exec_profile))
2374                         return 1;
2375                 free(exec_profile);
2376                 exec_profile = NULL;
2377         }
2378         if (!thread_number)
2379                 return 0;
2380
2381         if (write_bw_log) {
2382                 struct log_params p = {
2383                         .log_type = IO_LOG_TYPE_BW,
2384                 };
2385
2386                 setup_log(&agg_io_log[DDIR_READ], &p, "agg-read_bw.log");
2387                 setup_log(&agg_io_log[DDIR_WRITE], &p, "agg-write_bw.log");
2388                 setup_log(&agg_io_log[DDIR_TRIM], &p, "agg-trim_bw.log");
2389         }
2390
2391         startup_mutex = fio_mutex_init(FIO_MUTEX_LOCKED);
2392         if (startup_mutex == NULL)
2393                 return 1;
2394
2395         set_genesis_time();
2396         stat_init();
2397         helper_thread_create(startup_mutex, sk_out);
2398
2399         cgroup_list = smalloc(sizeof(*cgroup_list));
2400         INIT_FLIST_HEAD(cgroup_list);
2401
2402         run_threads(sk_out);
2403
2404         helper_thread_exit();
2405
2406         if (!fio_abort) {
2407                 __show_run_stats();
2408                 if (write_bw_log) {
2409                         for (i = 0; i < DDIR_RWDIR_CNT; i++) {
2410                                 struct io_log *log = agg_io_log[i];
2411
2412                                 flush_log(log, false);
2413                                 free_log(log);
2414                         }
2415                 }
2416         }
2417
2418         for_each_td(td, i) {
2419                 if (td->ss.dur) {
2420                         if (td->ss.iops_data != NULL) {
2421                                 free(td->ss.iops_data);
2422                                 free(td->ss.bw_data);
2423                         }
2424                 }
2425                 fio_options_free(td);
2426                 if (td->rusage_sem) {
2427                         fio_mutex_remove(td->rusage_sem);
2428                         td->rusage_sem = NULL;
2429                 }
2430         }
2431
2432         free_disk_util();
2433         cgroup_kill(cgroup_list);
2434         sfree(cgroup_list);
2435         sfree(cgroup_mnt);
2436
2437         fio_mutex_remove(startup_mutex);
2438         stat_exit();
2439         return exit_value;
2440 }