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