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